diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_Basis.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_Basis.hpp
index 74b34efb6681..5779d95741e8 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_Basis.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_Basis.hpp
@@ -379,6 +379,61 @@ using HostBasisPtr = BasisPtr<typename Kokkos::HostSpace::device_type, OutputTyp
       }
     }
 
+
+    /** \brief  Return the size of the scratch space, in bytes, needed for using the team-level implementation of getValues.
+
+        Warning, <var>inputPoints</var> is only used to deduce the type of the points where to evaluate basis functions.
+        The rank of </var>inputPoints</var> and its size are not relevant, however, 
+        when using DFAD types, </var>inputPoints</var> cannot be empty, 
+        otherwise the size of the scracth space needed won't be deduced correctly.
+
+        \param  space             [in]  - inputPoints
+        \param  perTeamSpaceSize  [out] - size of the scratch space needed per team
+        \param  perThreadeSize    [out] - size of the scratch space beeded per thread
+    */
+    virtual
+    void getScratchSpaceSize(       ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                                const PointViewType inputPoints,
+                                const EOperator operatorType = OPERATOR_VALUE) const {
+      INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE( true, std::logic_error,
+                                    ">>> ERROR (Basis::getValuesScratchSpace): this method is not supported or should be overridden accordingly by derived classes.");
+    }
+
+
+    /** \brief  Team-level evaluation of basis functions on a <strong>reference cell</strong>.
+
+        Returns values of <var>operatorType</var> acting on basis functions for a set of
+        points in the <strong>reference cell</strong> for which the basis is defined.
+
+        The interface allow also to select basis functions associated to a particular entity. 
+        As an example, if <var>subcellDim==1</var> (edges) and <var>subcellOrdinal==0</var>, <var>outputValues</var> will contain all the basis functions associated with the first edge.
+        <var>outputValues</var> will contain all the cell basis functions when the default value (-1) is used for <var>subcellDim</var> and <var>subcellOrdinal</var>
+
+        \param  outputValues      [out] - variable rank array with the basis values
+        \param  inputPoints       [in]  - rank-2 array (P,D) with the evaluation points
+        \param  operatorType      [in]  - the operator acting on the basis functions
+        \param  teamMember        [in]  - team member of the Kokkos::TemaPolicy
+        \param  scratchStorage    [in]  - scratch space to use by each team
+        \param  subcellDim        [in]  - the dimension of the subcells, the default values of -1 returns basis functions associated to subcells of all dimensions
+        \param  subcellOrdinal    [in]  - the ordinal of the subcell, the default values of -1 returns basis functions associated to subcells of all ordinals
+
+        \remark This function is supposed to be called within a TeamPolicy kernel. 
+                The size of the required scratch space is determined by the getScratchSpaceSize function.
+    */
+    KOKKOS_INLINE_FUNCTION
+    virtual
+    void getValues(       OutputViewType /* outputValues */,
+                    const PointViewType  /* inputPoints */,
+                    const EOperator /* operatorType */,
+                    const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type& teamMember,
+                    const typename ExecutionSpace::scratch_memory_space &scratchStorage, 
+                    const ordinal_type subcellDim=-1,
+                    const ordinal_type subcellOrdinal=-1) const {
+      INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE( true, std::logic_error,
+                                    ">>> ERROR (Basis::getValues): this method is not supported or should be overridden accordingly by derived classes.");
+    }
+
     /** \brief  Evaluation of a FEM basis on a <strong>reference cell</strong>.
 
         Returns values of <var>operatorType</var> acting on FEM basis functions for a set of
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEM.hpp
index 299054557fca..72d0e9112c01 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEM.hpp
@@ -185,6 +185,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEMDef.hpp
index 7eff91667e1b..71ea78656fc1 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_I1_FEMDef.hpp
@@ -330,6 +330,57 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_HEX_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_HEX_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_HEX_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_HEX_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_HEX_I1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HCURL_HEX_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEM.hpp
index 1af120be9949..64327bb29c08 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEM.hpp
@@ -148,20 +148,21 @@ namespace Intrepid2 {
   class Basis_HCURL_HEX_In_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType, outputValueType, pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
 
     /** \brief  Constructor.
      */
     Basis_HCURL_HEX_In_FEM(const ordinal_type order,
                             const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
 
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -184,6 +185,23 @@ namespace Intrepid2 {
                                                 operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEMDef.hpp
index 1d18b7887096..182c05d721b0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_HEX_In_FEMDef.hpp
@@ -21,19 +21,19 @@ namespace Intrepid2 {
   // -------------------------------------------------------------------------------------
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HCURL_HEX_In_FEM::Serial<opType>::
+    Basis_HCURL_HEX_In_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinvLine,
-               const vinvViewType   vinvBubble) {
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinvLine,
+               const VinvViewType   vinvBubble) {
       const ordinal_type cardLine = vinvLine.extent(0);
       const ordinal_type cardBubble = vinvBubble.extent(0);
 
@@ -44,22 +44,22 @@ namespace Intrepid2 {
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
       const auto input_z = Kokkos::subview(input, Kokkos::ALL(), range_type(2,3));
 
-      const ordinal_type dim_s = get_dimension_scalar(work);
+      const ordinal_type dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data() + cardLine*npts*dim_s;
       auto ptr2 = work.data() + 2*cardLine*npts*dim_s;
       auto ptr3 = work.data() + 3*cardLine*npts*dim_s;   
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
 
-        viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType outputLine_A(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType outputLine_B(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
-        viewType outputBubble(Kokkos::view_wrap(ptr3, vcprop), cardBubble, npts);
+        ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType outputLine_A(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType outputLine_B(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+        ViewType outputBubble(Kokkos::view_wrap(ptr3, vcprop), cardBubble, npts);
         
         // tensor product
         ordinal_type idx = 0;
@@ -142,12 +142,12 @@ namespace Intrepid2 {
          auto ptr4 = work.data() + 4*cardLine*npts*dim_s;
          auto ptr5 = work.data() + 5*cardLine*npts*dim_s;
          
-        viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType outputLine_A(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType outputLine_B(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
-        viewType outputLine_DA(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts, 1);
-        viewType outputLine_DB(Kokkos::view_wrap(ptr4, vcprop), cardLine, npts, 1);
-        viewType outputBubble(Kokkos::view_wrap(ptr5, vcprop), cardBubble, npts);
+        ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType outputLine_A(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType outputLine_B(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+        ViewType outputLine_DA(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts, 1);
+        ViewType outputLine_DB(Kokkos::view_wrap(ptr4, vcprop), cardLine, npts, 1);
+        ViewType outputBubble(Kokkos::view_wrap(ptr5, vcprop), cardBubble, npts);
 
         // tensor product
         ordinal_type idx = 0;
@@ -588,6 +588,70 @@ namespace Intrepid2 {
     this->dofCoeffs_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoeffsHost);
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffsHost);
   }
-}
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_HEX_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    ordinal_type scalarWorkViewExtent = (operatorType == OPERATOR_VALUE) ? 
+                                      3*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0):
+                                      5*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0);
+    perThreadSpaceSize = scalarWorkViewExtent*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_HEX_In_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_HEX_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? 
+                                      3*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0):
+                                      5*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0);
+      ordinal_type sizePerPoint = scalarSizePerPoint*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HCURL_HEX_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+          case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HCURL_HEX_In_FEM::Serial<OPERATOR_CURL>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HCURL_HEX_In_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEM.hpp
index 15f266e2db91..24c4b26bf746 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEM.hpp
@@ -144,20 +144,21 @@ namespace Intrepid2 {
            typename pointValueType = double>
   class Basis_HCURL_QUAD_I1_FEM : public Basis<DeviceType,outputValueType, pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
 
 
     /** \brief  Constructor.
      */
     Basis_HCURL_QUAD_I1_FEM();
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
 
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -178,6 +179,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEMDef.hpp
index 548929fb74cc..8380a4665a05 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_I1_FEMDef.hpp
@@ -60,7 +60,7 @@ namespace Intrepid2 {
       default: {
         INTREPID2_TEST_FOR_ABORT( opType != OPERATOR_VALUE &&
                                   opType != OPERATOR_CURL,
-                                  ">>> ERROR: (Intrepid2::Basis_HGRAD_QUAD_C1_FEM::Serial::getValues) operator is not supported");
+                                  ">>> ERROR: (Intrepid2::Basis_HCURL_QUAD_I1_FEM::Serial::getValues) operator is not supported");
       }
       } //end switch
     }
@@ -219,7 +219,56 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_QUAD_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
 
-}// namespace Intrepid2
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_QUAD_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_QUAD_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
 
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_QUAD_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_QUAD_I1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HCURL_QUAD_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
+}// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEM.hpp
index 077f6de07afb..13d0c227d421 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEM.hpp
@@ -136,20 +136,21 @@ namespace Intrepid2 {
   class Basis_HCURL_QUAD_In_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType, outputValueType, pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
 
     /** \brief  Constructor.
      */
     Basis_HCURL_QUAD_In_FEM(const ordinal_type order,
                             const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
 
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -172,6 +173,23 @@ namespace Intrepid2 {
                                                 operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEMDef.hpp
index 13a732abb88d..b00248a51fc8 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_QUAD_In_FEMDef.hpp
@@ -21,19 +21,19 @@ namespace Intrepid2 {
   // -------------------------------------------------------------------------------------
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HCURL_QUAD_In_FEM::Serial<opType>::
+    Basis_HCURL_QUAD_In_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinvLine,
-               const vinvViewType   vinvBubble) {
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinvLine,
+               const VinvViewType   vinvBubble) {
       const ordinal_type cardLine = vinvLine.extent(0);
       const ordinal_type cardBubble = vinvBubble.extent(0);
 
@@ -43,19 +43,19 @@ namespace Intrepid2 {
       const auto input_x = Kokkos::subview(input, Kokkos::ALL(), range_type(0,1));
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
 
-      const int dim_s = get_dimension_scalar(work);
+      const int dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType outputBubble(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+        ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType outputBubble(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
 
         // tensor product
         ordinal_type idx = 0;
@@ -101,11 +101,11 @@ namespace Intrepid2 {
       case OPERATOR_CURL: {
         ordinal_type idx = 0;
         { // x - component
-          viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           // x bubble value
-          viewType output_x(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+          ViewType output_x(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
           // y line grad
-          viewType output_y(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
+          ViewType output_y(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
 
           Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
             getValues(output_x, input_x, workLine, vinvBubble);
@@ -120,11 +120,11 @@ namespace Intrepid2 {
                 output.access(idx,k) = -output_x.access(i,k)*output_y.access(j,k,0);
         }
         { // y - component
-          viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           // x line grad
-          viewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
+          ViewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
           // y bubble value
-          viewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+          ViewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
 
           Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
             getValues(output_y, input_y, workLine, vinvBubble);
@@ -386,6 +386,63 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffsHost);
   }
 
-}
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_QUAD_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = (2*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0))*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_QUAD_In_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_QUAD_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type sizePerPoint = (2*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0))*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HCURL_QUAD_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+          case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HCURL_QUAD_In_FEM::Serial<OPERATOR_CURL>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HCURL_QUAD_In_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEM.hpp
index 6d90318a4961..d293da0e38c0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEM.hpp
@@ -184,6 +184,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEMDef.hpp
index 9c3d2b2d1c23..4d38583f8e49 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_I1_FEMDef.hpp
@@ -255,6 +255,57 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_TET_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
 
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_TET_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_TET_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_TET_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_TET_I1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HCURL_TET_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEM.hpp
index cae49e5b09a5..ed253d57ec13 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEM.hpp
@@ -217,9 +217,26 @@ class Basis_HCURL_TET_In_FEM
                                         operatorType);
   }
 
-  virtual
-  void
-  getDofCoords( ScalarViewType dofCoords ) const override {
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+    virtual
+    void
+    getDofCoords( ScalarViewType dofCoords ) const override {
 #ifdef HAVE_INTREPID2_DEBUG
     // Verify rank of output array.
     INTREPID2_TEST_FOR_EXCEPTION( rank(dofCoords) != 2, std::invalid_argument,
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEMDef.hpp
index 8bb82254291f..56149a4a1820 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEMDef.hpp
@@ -26,18 +26,18 @@ namespace Intrepid2 {
 
 namespace Impl {
 
-template<EOperator opType>
+template<EOperator OpType>
 template<typename OutputViewType,
-typename inputViewType,
-typename workViewType,
-typename vinvViewType>
+typename InputViewType,
+typename WorkViewType,
+typename VinvViewType>
 KOKKOS_INLINE_FUNCTION
 void
-Basis_HCURL_TET_In_FEM::Serial<opType>::
+Basis_HCURL_TET_In_FEM::Serial<OpType>::
 getValues(       OutputViewType output,
-    const inputViewType  input,
-          workViewType   work,
-    const vinvViewType   coeffs ) {
+    const InputViewType  input,
+          WorkViewType   work,
+    const VinvViewType   coeffs ) {
 
   constexpr ordinal_type spaceDim = 3;
   const ordinal_type
@@ -54,17 +54,17 @@ getValues(       OutputViewType output,
     }
   }
 
-  typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-  auto vcprop = Kokkos::common_view_alloc_prop(work);
+  typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+  auto vcprop = Kokkos::common_view_alloc_prop(input);
   auto ptr = work.data();
 
-  switch (opType) {
+  switch (OpType) {
   case OPERATOR_VALUE: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-    workViewType dummyView;
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -76,9 +76,9 @@ getValues(       OutputViewType output,
     break;
   }
   case OPERATOR_CURL: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
     Serial<OPERATOR_GRAD>::getValues(phis, input, workView, order);
@@ -282,7 +282,7 @@ Basis_HCURL_TET_In_FEM( const ordinal_type order,
 #ifdef HAVE_INTREPID2_DEBUG
   ordinal_type num_nonzero_sv = 0;
   for (int i=0;i<cardVecPn;i++)
-    num_nonzero_sv += (S(i,0) > tolerence());
+    num_nonzero_sv += (S(i,0) > 10*tolerence());
 
   INTREPID2_TEST_FOR_EXCEPTION( num_nonzero_sv != card, std::invalid_argument,
       ">>> ERROR: (Intrepid2::Basis_HCURL_TET_In_FEM( order, pointType), Matrix V1 should have rank equal to the cardinality of HCURL space");
@@ -562,5 +562,64 @@ Basis_HCURL_TET_In_FEM( const ordinal_type order,
         posDfOrd);
   }
 }
+
+template<typename DT, typename OT, typename PT>
+void 
+Basis_HCURL_TET_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                  ordinal_type& perTeamSpaceSize,
+                                  ordinal_type& perThreadSpaceSize,
+                            const PointViewType inputPoints,
+                            const EOperator operatorType) const {
+  perTeamSpaceSize = 0;
+  ordinal_type scalarWorkViewExtent = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 7*this->basisCardinality_;
+  perThreadSpaceSize = scalarWorkViewExtent*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+}
+
+template<typename DT, typename OT, typename PT>
+KOKKOS_INLINE_FUNCTION
+void 
+Basis_HCURL_TET_In_FEM<DT,OT,PT>::getValues(       
+        OutputViewType outputValues,
+    const PointViewType  inputPoints,
+    const EOperator operatorType,
+    const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+    const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+    const ordinal_type subcellDim,
+    const ordinal_type subcellOrdinal) const {
+
+    INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+      ">>> ERROR: (Intrepid2::Basis_HCURL_TET_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+    const int numPoints = inputPoints.extent(0);
+    using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+    using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+    ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 7*this->basisCardinality_;
+    ordinal_type sizePerPoint = scalarSizePerPoint*get_dimension_scalar(inputPoints);
+    WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+    using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+    switch(operatorType) {
+      case OPERATOR_VALUE:
+        Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+          auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+          const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+          WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+          Impl::Basis_HCURL_TET_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->coeffs_ );
+        });
+        break;
+        case OPERATOR_CURL:
+        Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+          auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+          const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+          WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+          Impl::Basis_HCURL_TET_In_FEM::Serial<OPERATOR_CURL>::getValues( output, input, work, this->coeffs_ );
+        });
+        break;
+      default: {          
+        INTREPID2_TEST_FOR_ABORT( true,
+          ">>> ERROR (Basis_HCURL_TET_In_FEM): getValues not implemented for this operator");
+        }
+  }
+}
 } // namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEM.hpp
index 816b999560a6..109c96988649 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEM.hpp
@@ -187,6 +187,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEMDef.hpp
index 85e639ea8f10..813b764608db 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_I1_FEMDef.hpp
@@ -208,7 +208,56 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_TRI_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_TRI_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_TRI_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
 
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_TRI_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_TRI_I1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HCURL_TRI_!1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
-
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEM.hpp
index 3c34d125847a..a030f292fb50 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEM.hpp
@@ -209,9 +209,26 @@ class Basis_HCURL_TRI_In_FEM
       operatorType);
   }
 
-  virtual
-  void
-  getDofCoords( ScalarViewType dofCoords ) const override {
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+    virtual
+    void
+    getDofCoords( ScalarViewType dofCoords ) const override {
 #ifdef HAVE_INTREPID2_DEBUG
     // Verify rank of output array.
     INTREPID2_TEST_FOR_EXCEPTION( rank(dofCoords) != 2, std::invalid_argument,
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEMDef.hpp
index 7d10682a5e45..6cb65ab386de 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEMDef.hpp
@@ -25,18 +25,18 @@ namespace Intrepid2 {
 
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HCURL_TRI_In_FEM::Serial<opType>::
+    Basis_HCURL_TRI_In_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   coeffs ) {
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   coeffs ) {
 
       constexpr ordinal_type spaceDim = 2;
       const ordinal_type
@@ -53,17 +53,16 @@ namespace Intrepid2 {
         }
       }
 
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
       auto ptr = work.data();
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-        workViewType dummyView;
+        const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts), dummyView;
 
         Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-          Serial<opType>::getValues(phis, input, dummyView, order);
+          Serial<OpType>::getValues(phis, input, dummyView, order);
 
         for (ordinal_type i=0;i<card;++i)
           for (ordinal_type j=0;j<npts;++j)
@@ -75,9 +74,9 @@ namespace Intrepid2 {
         break;
       }
       case OPERATOR_CURL: {
-        const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-        ptr += card*npts*spaceDim*get_dimension_scalar(work);
-        const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+        const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+        ptr += card*npts*spaceDim*get_dimension_scalar(input);
+        const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
         Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
           Serial<OPERATOR_GRAD>::getValues(phis, input, workView, order);
@@ -452,5 +451,66 @@ namespace Intrepid2 {
                               posDfOrd);
     }
   }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_TRI_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    ordinal_type scalarWorkViewExtent = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 5*this->basisCardinality_;
+    perThreadSpaceSize = scalarWorkViewExtent*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_TRI_In_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_TRI_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 5*this->basisCardinality_;
+      ordinal_type sizePerPoint = scalarSizePerPoint*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HCURL_TRI_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->coeffs_ );
+          });
+          break;
+          case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HCURL_TRI_In_FEM::Serial<OPERATOR_CURL>::getValues( output, input, work, this->coeffs_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HCURL_TRI_In_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
 } // namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEM.hpp
index c7587cf3eec1..d2831d0ac47a 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEM.hpp
@@ -185,6 +185,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEMDef.hpp
index 59ad4da436e8..754355ffbd7d 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_WEDGE_I1_FEMDef.hpp
@@ -276,5 +276,57 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HCURL_WEDGE_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HCURL_WEDGE_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HCURL_WEDGE_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_WEDGE_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HCURL_WEDGE_I1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HCURL_WEDGE_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEM.hpp
index 1de1d7c654c7..66ab525b3aec 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEM.hpp
@@ -190,6 +190,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEMDef.hpp
index 79e9aaef60f8..b7e865178e64 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_I1_FEMDef.hpp
@@ -236,5 +236,57 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_HEX_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_HEX_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_HEX_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_HEX_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_HEX_I1_FEM::Serial<OPERATOR_DIV>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HDIV_HEX_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEM.hpp
index 4ed98a89967f..f563bd998237 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEM.hpp
@@ -138,20 +138,21 @@ namespace Intrepid2 {
   class Basis_HDIV_HEX_In_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType, outputValueType, pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
 
     /** \brief  Constructor.
      */
     Basis_HDIV_HEX_In_FEM(const ordinal_type order,
                           const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
 
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -174,6 +175,23 @@ namespace Intrepid2 {
                                                 operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
@@ -254,8 +272,6 @@ namespace Intrepid2 {
 
 }// namespace Intrepid2
 
-
-
 #include "Intrepid2_HDIV_HEX_In_FEMDef.hpp"
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEMDef.hpp
index 0bae2c8b1b3d..0d5d25113bdb 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_HEX_In_FEMDef.hpp
@@ -21,19 +21,19 @@ namespace Intrepid2 {
   // -------------------------------------------------------------------------------------
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HDIV_HEX_In_FEM::Serial<opType>::
+    Basis_HDIV_HEX_In_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinvLine,
-               const vinvViewType   vinvBubble) {
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinvLine,
+               const VinvViewType   vinvBubble) {
       const ordinal_type cardLine = vinvLine.extent(0);
       const ordinal_type cardBubble = vinvBubble.extent(0);
 
@@ -44,21 +44,21 @@ namespace Intrepid2 {
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
       const auto input_z = Kokkos::subview(input, Kokkos::ALL(), range_type(2,3));
 
-      const ordinal_type dim_s = get_dimension_scalar(work);
+      const ordinal_type dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       auto ptr3 = work.data()+(2*cardLine+cardBubble)*npts*dim_s;      
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType outputBubble_A(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
-        viewType outputBubble_B(Kokkos::view_wrap(ptr3, vcprop), cardBubble, npts);
+        ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType outputBubble_A(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+        ViewType outputBubble_B(Kokkos::view_wrap(ptr3, vcprop), cardBubble, npts);
         
         // tensor product
         ordinal_type idx = 0;
@@ -138,13 +138,13 @@ namespace Intrepid2 {
         break;
       }
       case OPERATOR_DIV: {      
-        viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
         // A line value
-        viewType outputBubble_A(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+        ViewType outputBubble_A(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
         // B line value
-        viewType outputBubble_B(Kokkos::view_wrap(ptr3, vcprop), cardBubble, npts);
+        ViewType outputBubble_B(Kokkos::view_wrap(ptr3, vcprop), cardBubble, npts);
         // Line grad
-        viewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
+        ViewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
         
         // tensor product
         ordinal_type idx = 0;
@@ -508,6 +508,64 @@ namespace Intrepid2 {
     this->dofCoeffs_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoeffsHost);
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffsHost);
   }
-}
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_HEX_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = (2*this->vinvLine_.extent(0)+2*this->vinvBubble_.extent(0))*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_HEX_In_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_HEX_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type sizePerPoint = (2*this->vinvLine_.extent(0)+2*this->vinvBubble_.extent(0))*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HDIV_HEX_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+          case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HDIV_HEX_In_FEM::Serial<OPERATOR_DIV>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HDIV_HEX_In_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEM.hpp
index fab13618142c..66c5843d4da0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEM.hpp
@@ -145,20 +145,21 @@ namespace Intrepid2 {
            typename pointValueType = double>
   class Basis_HDIV_QUAD_I1_FEM : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
 
 
     /** \brief  Constructor.
      */
     Basis_HDIV_QUAD_I1_FEM();
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
 
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -179,6 +180,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEMDef.hpp
index 1924df01978b..22eeef61929a 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_I1_FEMDef.hpp
@@ -213,10 +213,59 @@ namespace Intrepid2 {
 
     this->dofCoeffs_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoeffs);
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffs);
-
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_QUAD_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
 
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_QUAD_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_QUAD_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_QUAD_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_QUAD_I1_FEM::Serial<OPERATOR_DIV>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HDIV_QUAD_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEM.hpp
index 3db4472991f2..60da55f64220 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEM.hpp
@@ -135,20 +135,21 @@ namespace Intrepid2 {
   class Basis_HDIV_QUAD_In_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType, outputValueType, pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
 
     /** \brief  Constructor.
      */
     Basis_HDIV_QUAD_In_FEM(const ordinal_type order,
                             const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
 
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -170,6 +171,24 @@ namespace Intrepid2 {
                                                 this->vinvBubble_,
                                                 operatorType );
     }
+
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEMDef.hpp
index f6958e8152c2..ee5bdc9458c4 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_QUAD_In_FEMDef.hpp
@@ -21,19 +21,19 @@ namespace Intrepid2 {
   // -------------------------------------------------------------------------------------
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HDIV_QUAD_In_FEM::Serial<opType>::
+    Basis_HDIV_QUAD_In_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinvLine,
-               const vinvViewType   vinvBubble) {
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinvLine,
+               const VinvViewType   vinvBubble) {
       const ordinal_type cardLine = vinvLine.extent(0);
       const ordinal_type cardBubble = vinvBubble.extent(0);
 
@@ -43,20 +43,19 @@ namespace Intrepid2 {
       const auto input_x = Kokkos::subview(input, Kokkos::ALL(), range_type(0,1));
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
 
-      const int dim_s = get_dimension_scalar(work);
+      const ordinal_type dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       
-      
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType outputBubble(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+        ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType outputLine(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType outputBubble(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
 
         // tensor product
         ordinal_type idx = 0;
@@ -100,11 +99,11 @@ namespace Intrepid2 {
       case OPERATOR_DIV: {
         ordinal_type idx = 0;
         { // x - component
-          viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           // x bubble value
-          viewType output_x(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+          ViewType output_x(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
           // y line grad
-          viewType output_y(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
+          ViewType output_y(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
 
           Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
             getValues(output_x, input_x, workLine, vinvBubble);
@@ -119,11 +118,11 @@ namespace Intrepid2 {
                 output.access(idx,k) = output_x.access(i,k)*output_y.access(j,k,0);
         }
         { // y - component
-          viewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType workLine(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           // x line grad
-          viewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
+          ViewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts,1);
           // y bubble value
-          viewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
+          ViewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardBubble, npts);
 
           Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
             getValues(output_y, input_y, workLine, vinvBubble);
@@ -381,6 +380,63 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffsHost);
   }
 
-}
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_QUAD_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = (2*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0))*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_QUAD_In_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_QUAD_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type sizePerPoint = (2*this->vinvLine_.extent(0)+this->vinvBubble_.extent(0))*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HDIV_QUAD_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+          case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HDIV_QUAD_In_FEM::Serial<OPERATOR_DIV>::getValues( output, input, work, this->vinvLine_, this->vinvBubble_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HDIV_QUAD_In_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEM.hpp
index a5b72c79d540..40b58117d658 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEM.hpp
@@ -144,19 +144,21 @@ namespace Intrepid2 {
            typename pointValueType = double>
   class Basis_HDIV_TET_I1_FEM: public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
     Basis_HDIV_TET_I1_FEM();
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -177,6 +179,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEMDef.hpp
index febd2f87333c..7aac987e9439 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_I1_FEMDef.hpp
@@ -230,6 +230,58 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffs);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_TET_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_TET_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_TET_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_TET_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_TET_I1_FEM::Serial<OPERATOR_DIV>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HDIV_TET_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
+
 #endif
 
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEM.hpp
index c2f17ac83aef..9224322f9539 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEM.hpp
@@ -165,131 +165,144 @@ template<typename DeviceType = void,
     typename pointValueType = double>
 class Basis_HDIV_TET_In_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
-    public:
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost OrdinalTypeArray1DHost;
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost OrdinalTypeArray2DHost;
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost OrdinalTypeArray3DHost;
-
-  /** \brief  Constructor.
-   */
-  Basis_HDIV_TET_In_FEM(const ordinal_type order,
-      const EPointType   pointType = POINTTYPE_EQUISPACED);
-
-
-  using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-  using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-  using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::scalarType  scalarType;
-
-  using Basis<DeviceType,outputValueType,pointValueType>::getValues;
-
-  virtual
-  void
-  getValues( /* */ OutputViewType outputValues,
-      const PointViewType  inputPoints,
-      const EOperator operatorType = OPERATOR_VALUE) const override {
+  public:
+    using BasisBase = Basis<DeviceType, outputValueType, pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
+
+    /** \brief  Constructor.
+     */
+    Basis_HDIV_TET_In_FEM(const ordinal_type order,
+        const EPointType   pointType = POINTTYPE_EQUISPACED);
+
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
+    using scalarType = typename BasisBase::scalarType;
+    using BasisBase::getValues;
+
+    virtual
+    void
+    getValues( /* */ OutputViewType outputValues,
+        const PointViewType  inputPoints,
+        const EOperator operatorType = OPERATOR_VALUE) const override {
 #ifdef HAVE_INTREPID2_DEBUG
-    Intrepid2::getValues_HDIV_Args(outputValues,
+   Intrepid2::getValues_HDIV_Args(outputValues,
         inputPoints,
         operatorType,
         this->getBaseCellTopology(),
         this->getCardinality() );
 #endif
-constexpr ordinal_type numPtsPerEval = Parameters::MaxNumPtsPerBasisEval;
-Impl::Basis_HDIV_TET_In_FEM::
-getValues<DeviceType,numPtsPerEval>( outputValues,
-    inputPoints,
-    this->coeffs_,
-    operatorType);
-  }
-
-  virtual
-  void
-  getDofCoords( ScalarViewType dofCoords ) const override {
+    constexpr ordinal_type numPtsPerEval = Parameters::MaxNumPtsPerBasisEval;
+    Impl::Basis_HDIV_TET_In_FEM::
+    getValues<DeviceType,numPtsPerEval>( outputValues,
+        inputPoints,
+        this->coeffs_,
+        operatorType);
+     }
+
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+    virtual
+    void
+    getDofCoords( ScalarViewType dofCoords ) const override {
 #ifdef HAVE_INTREPID2_DEBUG
-    // Verify rank of output array.
-    INTREPID2_TEST_FOR_EXCEPTION( dofCoords.rank() != 2, std::invalid_argument,
-        ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoords) rank = 2 required for dofCoords array");
-    // Verify 0th dimension of output array.
-    INTREPID2_TEST_FOR_EXCEPTION( static_cast<ordinal_type>(dofCoords.extent(0)) != this->getCardinality(), std::invalid_argument,
-        ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoords) mismatch in number of dof and 0th dimension of dofCoords array");
-    // Verify 1st dimension of output array.
-    INTREPID2_TEST_FOR_EXCEPTION( dofCoords.extent(1) != this->getBaseCellTopology().getDimension(), std::invalid_argument,
-        ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoords) incorrect reference cell (1st) dimension in dofCoords array");
+      // Verify rank of output array.
+      INTREPID2_TEST_FOR_EXCEPTION( dofCoords.rank() != 2, std::invalid_argument,
+          ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoords) rank = 2 required for dofCoords array");
+      // Verify 0th dimension of output array.
+      INTREPID2_TEST_FOR_EXCEPTION( static_cast<ordinal_type>(dofCoords.extent(0)) != this->getCardinality(), std::invalid_argument,
+          ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoords) mismatch in number of dof and 0th dimension of dofCoords array");
+      // Verify 1st dimension of output array.
+      INTREPID2_TEST_FOR_EXCEPTION( dofCoords.extent(1) != this->getBaseCellTopology().getDimension(), std::invalid_argument,
+          ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoords) incorrect reference cell (1st) dimension in dofCoords array");
 #endif
-    Kokkos::deep_copy(dofCoords, this->dofCoords_);
-  }
+      Kokkos::deep_copy(dofCoords, this->dofCoords_);
+    }
 
-  virtual
-  void
-  getDofCoeffs( ScalarViewType dofCoeffs ) const override {
-#ifdef HAVE_INTREPID2_DEBUG
-    // Verify rank of output array.
-    INTREPID2_TEST_FOR_EXCEPTION( dofCoeffs.rank() != 2, std::invalid_argument,
-        ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoeffs) rank = 2 required for dofCoeffs array");
-    // Verify 0th dimension of output array.
-    INTREPID2_TEST_FOR_EXCEPTION( static_cast<ordinal_type>(dofCoeffs.extent(0)) != this->getCardinality(), std::invalid_argument,
+    virtual
+    void
+    getDofCoeffs( ScalarViewType dofCoeffs ) const override {
+  #ifdef HAVE_INTREPID2_DEBUG
+      // Verify rank of output array.
+      INTREPID2_TEST_FOR_EXCEPTION( dofCoeffs.rank() != 2, std::invalid_argument,
+          ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoeffs) rank = 2 required for dofCoeffs array");
+      // Verify 0th dimension of output array.
+      INTREPID2_TEST_FOR_EXCEPTION( static_cast<ordinal_type>(dofCoeffs.extent(0)) != this->getCardinality(), std::invalid_argument,
         ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoeffs) mismatch in number of dof and 0th dimension of dofCoeffs array");
-    // Verify 1st dimension of output array.
-    INTREPID2_TEST_FOR_EXCEPTION( dofCoeffs.extent(1) != this->getBaseCellTopology().getDimension(), std::invalid_argument,
+      // Verify 1st dimension of output array.
+      INTREPID2_TEST_FOR_EXCEPTION( dofCoeffs.extent(1) != this->getBaseCellTopology().getDimension(), std::invalid_argument,
         ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getDofCoeffs) incorrect reference cell (1st) dimension in dofCoeffs array");
 #endif
-    Kokkos::deep_copy(dofCoeffs, this->dofCoeffs_);
-  }
-
-  void
-  getExpansionCoeffs( ScalarViewType coeffs ) const {
-    // has to be same rank and dimensions
-    Kokkos::deep_copy(coeffs, this->coeffs_);
-  }
-
-  virtual
-  const char*
-  getName() const override {
-    return "Intrepid2_HDIV_TET_In_FEM";
-  }
-
-  virtual
-  bool
-  requireOrientation() const override {
-    return true;
-  }
-
-  /** \brief returns the basis associated to a subCell.
-
-      The bases of the subCell are the restriction to the subCell of the bases of the parent cell,
-      projected along normal to the subCell.
-
-      \param [in] subCellDim - dimension of subCell
-      \param [in] subCellOrd - position of the subCell among of the subCells having the same dimension
-      \return pointer to the subCell basis of dimension subCellDim and position subCellOrd
-   */
-  BasisPtr<DeviceType,outputValueType,pointValueType>
-    getSubCellRefBasis(const ordinal_type subCellDim, const ordinal_type subCellOrd) const override{
+      Kokkos::deep_copy(dofCoeffs, this->dofCoeffs_);
+    }
 
-    if(subCellDim == 2) {
-      return Teuchos::rcp(new
-          Basis_HVOL_TRI_Cn_FEM<DeviceType,outputValueType,pointValueType>
-          (this->basisDegree_-1, pointType_));
+    void
+    getExpansionCoeffs( ScalarViewType coeffs ) const {
+      // has to be same rank and dimensions
+      Kokkos::deep_copy(coeffs, this->coeffs_);
     }
-    INTREPID2_TEST_FOR_EXCEPTION(true,std::invalid_argument,"Input parameters out of bounds");
-  }
 
-  BasisPtr<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>
-  getHostBasis() const override{
-    return Teuchos::rcp(new Basis_HDIV_TET_In_FEM<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>(this->basisDegree_, pointType_));
-  }
-    private:
+    virtual
+    const char*
+    getName() const override {
+      return "Intrepid2_HDIV_TET_In_FEM";
+    }
 
-  /** \brief expansion coefficients of the nodal basis in terms of the
-        orthgonal one */
-  Kokkos::DynRankView<scalarType,DeviceType> coeffs_;
+    virtual
+    bool
+    requireOrientation() const override {
+      return true;
+    }
 
-  /** \brief type of lattice used for creating the DoF coordinates  */
-  EPointType pointType_;
+    /** \brief returns the basis associated to a subCell.
 
-};
+        The bases of the subCell are the restriction to the subCell of the bases of the parent cell,
+        projected along normal to the subCell.
+
+        \param [in] subCellDim - dimension of subCell
+        \param [in] subCellOrd - position of the subCell among of the subCells having the same dimension
+        \return pointer to the subCell basis of dimension subCellDim and position subCellOrd
+     */
+    BasisPtr<DeviceType,outputValueType,pointValueType>
+      getSubCellRefBasis(const ordinal_type subCellDim, const ordinal_type subCellOrd) const override{
+
+     if(subCellDim == 2) {
+        return Teuchos::rcp(new
+            Basis_HVOL_TRI_Cn_FEM<DeviceType,outputValueType,pointValueType>
+            (this->basisDegree_-1, pointType_));
+      }
+      INTREPID2_TEST_FOR_EXCEPTION(true,std::invalid_argument,"Input parameters out of bounds");
+    }
+
+    BasisPtr<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>
+    getHostBasis() const override{
+      return Teuchos::rcp(new Basis_HDIV_TET_In_FEM<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>(this->basisDegree_, pointType_));
+    }
+  private:
+
+    /** \brief expansion coefficients of the nodal basis in terms of the orthgonal one */
+    Kokkos::DynRankView<scalarType,DeviceType> coeffs_;
+  
+    /** \brief type of lattice used for creating the DoF coordinates  */
+    EPointType pointType_;
+  };
 
 }// namespace Intrepid2
 
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEMDef.hpp
index e8c97199198f..96e0e7cf2267 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TET_In_FEMDef.hpp
@@ -25,18 +25,18 @@ namespace Intrepid2 {
 
 namespace Impl {
 
-template<EOperator opType>
+template<EOperator OpType>
 template<typename OutputViewType,
-typename inputViewType,
-typename workViewType,
-typename vinvViewType>
+typename InputViewType,
+typename WorkViewType,
+typename VinvViewType>
 KOKKOS_INLINE_FUNCTION
 void
-Basis_HDIV_TET_In_FEM::Serial<opType>::
+Basis_HDIV_TET_In_FEM::Serial<OpType>::
 getValues(       OutputViewType output,
-    const inputViewType  input,
-          workViewType   work,
-    const vinvViewType   coeffs ) {
+    const InputViewType  input,
+          WorkViewType   work,
+    const VinvViewType   coeffs ) {
 
   constexpr ordinal_type spaceDim = 3;
   const ordinal_type
@@ -53,17 +53,17 @@ getValues(       OutputViewType output,
     }
   }
 
-  typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-  auto vcprop = Kokkos::common_view_alloc_prop(work);
+  typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+  auto vcprop = Kokkos::common_view_alloc_prop(input);
   auto ptr = work.data();
 
-  switch (opType) {
+  switch (OpType) {
   case OPERATOR_VALUE: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-    workViewType dummyView;
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -75,9 +75,9 @@ getValues(       OutputViewType output,
     break;
   }
   case OPERATOR_DIV: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
     Serial<OPERATOR_GRAD>::getValues(phis, input, workView, order);
@@ -104,10 +104,10 @@ typename inputPointValueType,  class ...inputPointProperties,
 typename vinvValueType,        class ...vinvProperties>
 void
 Basis_HDIV_TET_In_FEM::
-getValues( /* */ Kokkos::DynRankView<outputValueValueType,outputValueProperties...> outputValues,
-    const Kokkos::DynRankView<inputPointValueType, inputPointProperties...>  inputPoints,
-    const Kokkos::DynRankView<vinvValueType,       vinvProperties...>        coeffs,
-    const EOperator operatorType) {
+getValues(       Kokkos::DynRankView<outputValueValueType,outputValueProperties...> outputValues,
+           const Kokkos::DynRankView<inputPointValueType, inputPointProperties...>  inputPoints,
+           const Kokkos::DynRankView<vinvValueType,       vinvProperties...>        coeffs,
+           const EOperator operatorType) {
   typedef          Kokkos::DynRankView<outputValueValueType,outputValueProperties...>         outputValueViewType;
   typedef          Kokkos::DynRankView<inputPointValueType, inputPointProperties...>          inputPointViewType;
   typedef          Kokkos::DynRankView<vinvValueType,       vinvProperties...>                vinvViewType;
@@ -450,5 +450,64 @@ Basis_HDIV_TET_In_FEM( const ordinal_type order,
         posDfOrd);
   }
 }
+
+template<typename DT, typename OT, typename PT>
+void 
+Basis_HDIV_TET_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                  ordinal_type& perTeamSpaceSize,
+                                  ordinal_type& perThreadSpaceSize,
+                            const PointViewType inputPoints,
+                            const EOperator operatorType) const {
+  perTeamSpaceSize = 0;
+  ordinal_type scalarWorkViewExtent = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 7*this->basisCardinality_;
+  perThreadSpaceSize = scalarWorkViewExtent*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+}
+
+template<typename DT, typename OT, typename PT>
+KOKKOS_INLINE_FUNCTION
+void 
+Basis_HDIV_TET_In_FEM<DT,OT,PT>::getValues(       
+        OutputViewType outputValues,
+    const PointViewType  inputPoints,
+    const EOperator operatorType,
+    const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+    const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+    const ordinal_type subcellDim,
+    const ordinal_type subcellOrdinal) const {
+
+    INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+      ">>> ERROR: (Intrepid2::Basis_HDIV_TET_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+    const int numPoints = inputPoints.extent(0);
+    using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+    using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+    ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 7*this->basisCardinality_;
+    ordinal_type sizePerPoint = scalarSizePerPoint*get_dimension_scalar(inputPoints);
+    WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+    using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+    switch(operatorType) {
+      case OPERATOR_VALUE:
+        Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+          auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+          const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+          WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+          Impl::Basis_HDIV_TET_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->coeffs_ );
+        });
+        break;
+        case OPERATOR_DIV:
+        Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+          auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+          const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+          WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+          Impl::Basis_HDIV_TET_In_FEM::Serial<OPERATOR_DIV>::getValues( output, input, work, this->coeffs_ );
+        });
+        break;
+      default: {          
+        INTREPID2_TEST_FOR_ABORT( true,
+          ">>> ERROR (Basis_HDIV_TET_In_FEM): getValues not implemented for this operator");
+        }
+  }
+}
 } // namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEM.hpp
index 6f6596778567..8315ac027c92 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEM.hpp
@@ -142,19 +142,21 @@ namespace Intrepid2 {
            typename pointValueType = double>
   class Basis_HDIV_TRI_I1_FEM: public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
     Basis_HDIV_TRI_I1_FEM();
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -175,6 +177,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEMDef.hpp
index a7c13864b3a2..ef00a009fdc7 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_I1_FEMDef.hpp
@@ -72,6 +72,7 @@ namespace Intrepid2 {
       typedef          Kokkos::DynRankView<inputPointValueType, inputPointProperties...>          inputPointViewType;
       typedef typename ExecSpace<typename inputPointViewType::execution_space,typename DT::execution_space>::ExecSpaceType ExecSpaceType;
 
+
       // Number of evaluation points = dim 0 of inputPoints
       const auto loopSize = inputPoints.extent(0);
       Kokkos::RangePolicy<ExecSpaceType,Kokkos::Schedule<Kokkos::Static> > policy(0, loopSize);
@@ -215,5 +216,56 @@ namespace Intrepid2 {
 
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_TRI_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_TRI_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_TRI_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_TRI_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_TRI_I1_FEM::Serial<OPERATOR_DIV>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HDIV_TRI_!1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEM.hpp
index 71fd4f8ae99d..1ca324c938ef 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEM.hpp
@@ -169,31 +169,30 @@ template<typename DeviceType = void,
     typename pointValueType = double>
 class Basis_HDIV_TRI_In_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
-    public:
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost OrdinalTypeArray1DHost;
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost OrdinalTypeArray2DHost;
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost OrdinalTypeArray3DHost;
-
-  /** \brief  Constructor.
-   */
-  Basis_HDIV_TRI_In_FEM(const ordinal_type order,
+  public:
+    using BasisBase = Basis<DeviceType, outputValueType, pointValueType>;
+    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
+    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
+    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
+
+    /** \brief  Constructor.
+     */
+    Basis_HDIV_TRI_In_FEM(const ordinal_type order,
       const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-  using HostBasis = Basis_HDIV_TRI_In_FEM<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>;
+    using HostBasis = Basis_HDIV_TRI_In_FEM<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>;
       
-  using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-  using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-  using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-  typedef typename Basis<DeviceType,outputValueType,pointValueType>::scalarType  scalarType;
-
-  using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using OutputViewType = typename BasisBase::OutputViewType;
+    using PointViewType  = typename BasisBase::PointViewType;
+    using ScalarViewType = typename BasisBase::ScalarViewType;
+    using scalarType = typename BasisBase::scalarType;
+    using BasisBase::getValues;
 
   virtual
   void
-  getValues( /* */ OutputViewType outputValues,
-      const PointViewType  inputPoints,
-      const EOperator operatorType = OPERATOR_VALUE) const override {
+  getValues(       OutputViewType outputValues,
+             const PointViewType  inputPoints,
+             const EOperator operatorType = OPERATOR_VALUE) const override {
 #ifdef HAVE_INTREPID2_DEBUG
     Intrepid2::getValues_HDIV_Args(outputValues,
         inputPoints,
@@ -209,9 +208,26 @@ class Basis_HDIV_TRI_In_FEM
         operatorType);
   }
 
-  virtual
-  void
-  getDofCoords( ScalarViewType dofCoords ) const override {
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+    virtual
+    void
+    getDofCoords( ScalarViewType dofCoords ) const override {
 #ifdef HAVE_INTREPID2_DEBUG
     // Verify rank of output array.
     INTREPID2_TEST_FOR_EXCEPTION( dofCoords.rank() != 2, std::invalid_argument,
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEMDef.hpp
index e412989caa56..89c86f5274e9 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_TRI_In_FEMDef.hpp
@@ -24,18 +24,18 @@ namespace Intrepid2 {
 // -------------------------------------------------------------------------------------
 namespace Impl {
 
-template<EOperator opType>
+template<EOperator OpType>
 template<typename OutputViewType,
-typename inputViewType,
-typename workViewType,
-typename vinvViewType>
+typename InputViewType,
+typename WorkViewType,
+typename VinvViewType>
 KOKKOS_INLINE_FUNCTION
 void
-Basis_HDIV_TRI_In_FEM::Serial<opType>::
-getValues( /* */ OutputViewType output,
-    const inputViewType  input,
-    /* */ workViewType   work,
-    const vinvViewType   coeffs ) {
+Basis_HDIV_TRI_In_FEM::Serial<OpType>::
+getValues(      OutputViewType  output,
+            const InputViewType input,
+                WorkViewType    work,
+            const VinvViewType  coeffs ) {
 
   constexpr ordinal_type spaceDim = 2;
   const ordinal_type
@@ -52,17 +52,17 @@ getValues( /* */ OutputViewType output,
     }
   }
 
-  typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-  auto vcprop = Kokkos::common_view_alloc_prop(work);
+  typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+  auto vcprop = Kokkos::common_view_alloc_prop(input);
   auto ptr = work.data();
 
-  switch (opType) {
+  switch (OpType) {
   case OPERATOR_VALUE: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-    workViewType dummyView;
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -74,9 +74,9 @@ getValues( /* */ OutputViewType output,
     break;
   }
   case OPERATOR_DIV: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
     ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
     Serial<OPERATOR_GRAD>::getValues(phis, input, workView, order);
@@ -447,5 +447,65 @@ Basis_HDIV_TRI_In_FEM( const ordinal_type order,
         posDfOrd);
   }
 }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_TRI_In_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    ordinal_type scalarWorkViewExtent = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 5*this->basisCardinality_;
+    perThreadSpaceSize = scalarWorkViewExtent*get_dimension_scalar(inputPoints)*sizeof(scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_TRI_In_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_TRI_In_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using WorkViewType = Kokkos::DynRankView< scalarType, typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 5*this->basisCardinality_;
+      ordinal_type sizePerPoint = scalarSizePerPoint*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HDIV_TRI_In_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->coeffs_ );
+          });
+          break;
+          case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HDIV_TRI_In_FEM::Serial<OPERATOR_DIV>::getValues( output, input, work, this->coeffs_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HDIV_TRI_In_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
 } // namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEM.hpp
index d30c7d1ee4a3..56a257b9ce54 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEM.hpp
@@ -159,6 +159,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEMDef.hpp
index 79dabbc6bfa6..ceefba8ac59e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HDIV_WEDGE_I1_FEMDef.hpp
@@ -183,7 +183,60 @@ namespace Intrepid2 {
 
     this->dofCoeffs_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoeffs);
     Kokkos::deep_copy(this->dofCoeffs_, dofCoeffs);
+
+  }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HDIV_WEDGE_I1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
   }
 
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HDIV_WEDGE_I1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HDIV_WEDGE_I1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_WEDGE_I1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_DIV:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HDIV_WEDGE_I1_FEM::Serial<OPERATOR_DIV>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HDIV_WEDGE_I1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEM.hpp
index 2ea99d81a411..59dfdfff0451 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEM.hpp
@@ -177,6 +177,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEMDef.hpp
index 96a5fcf39429..3c2813faa81e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C1_FEMDef.hpp
@@ -413,6 +413,55 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_HEX_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_HEX_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_HEX_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_HEX_C1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_HEX_C1_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
 }// namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEM.hpp
index 9823b1fe23bd..1af419c4300e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEM.hpp
@@ -245,6 +245,23 @@ namespace Intrepid2 {
                                 operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEMDef.hpp
index 22e1a54b74b4..c5a5ec102087 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_C2_FEMDef.hpp
@@ -1626,8 +1626,57 @@ namespace Intrepid2 {
 
     this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoords);
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
+  }
 
+  template<bool serendipity, typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_HEX_DEG2_FEM<serendipity,DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
   }
 
+  template<bool serendipity, typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_HEX_DEG2_FEM<serendipity,DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_HEX_DEG2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            using SerialValue = typename Impl::Basis_HGRAD_HEX_DEG2_FEM<serendipity>::template Serial<OPERATOR_VALUE>;
+            SerialValue::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            using SerialGrad = typename Impl::Basis_HGRAD_HEX_DEG2_FEM<serendipity>::template Serial<OPERATOR_GRAD>;
+            SerialGrad::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEM.hpp
index 4bcefb52d286..987206b0a1e0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEM.hpp
@@ -184,6 +184,23 @@ namespace Intrepid2 {
                                             operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
 
     virtual
     void
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEMDef.hpp
index ac50cde72fb3..36139dfb95f4 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_HEX_Cn_FEMDef.hpp
@@ -44,14 +44,14 @@ namespace Intrepid2 {
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
       const auto input_z = Kokkos::subview(input, Kokkos::ALL(), range_type(2,3));
 
-      const ordinal_type dim_s = get_dimension_scalar(work);
+      const ordinal_type dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       auto ptr3 = work.data()+3*cardLine*npts*dim_s;      
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename inputViewType::value_type, typename workViewType::memory_space> viewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
 
       switch (opType) {
       case OPERATOR_VALUE: {
@@ -382,7 +382,64 @@ namespace Intrepid2 {
     this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoordsHost);
     Kokkos::deep_copy(this->dofCoords_, dofCoordsHost);
   }
-  
+
+  template<typename DT, typename OT, typename PT>  
+  void 
+  Basis_HGRAD_HEX_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    (void) operatorType; //avoid warning for unused variable
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 4*this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_HEX_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_HEX_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+    
+    const int numPoints = inputPoints.extent(0);
+    using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+    using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+    ordinal_type sizePerPoint = 4*this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+    WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+    using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+    switch(operatorType) {
+      case OPERATOR_VALUE:
+        Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+          auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+          const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+          WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+          Impl::Basis_HGRAD_HEX_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_ );
+        });
+        break;
+      case OPERATOR_GRAD:
+        Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+          auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+          const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+          WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+          Impl::Basis_HGRAD_HEX_Cn_FEM::Serial<OPERATOR_GRAD>::getValues( output, input, work, this->vinv_ );
+        });
+        break;
+      default: {
+        INTREPID2_TEST_FOR_ABORT( true,
+          ">>> ERROR (Basis_HGRAD_TET_Cn_FEM): getValues not implemented for this operator");
+      }
+    }      
+  }  
 }// namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEM.hpp
index c07fdfd71f04..71ee77e7d55a 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEM.hpp
@@ -161,6 +161,23 @@ namespace Intrepid2 {
                               operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEMDef.hpp
index 4572b5b5c899..05291e830a3d 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C1_FEMDef.hpp
@@ -177,6 +177,55 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
-}
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_LINE_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_LINE_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_LINE_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_LINE_C1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_LINE_C1_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
+}// namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEM.hpp
index 595cc815e9be..4e141fc5ad41 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEM.hpp
@@ -161,6 +161,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEMDef.hpp
index c08bd3692a27..cb3157028b0e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_C2_FEMDef.hpp
@@ -173,6 +173,55 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
-}
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_LINE_C2_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_LINE_C2_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_LINE_C2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_LINE_C2_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_LINE_C2_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
+}// namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEM.hpp
index c5e0e2308545..ac2bfdbcae0b 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEM.hpp
@@ -192,6 +192,23 @@ namespace Intrepid2 {
                                             operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEMDef.hpp
index fc116eb32ebe..6140c13821cb 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_LINE_Cn_FEMDef.hpp
@@ -23,16 +23,16 @@ namespace Intrepid2 {
     
     template<EOperator opType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
     Basis_HGRAD_LINE_Cn_FEM::Serial<opType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinv,
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv,
                const ordinal_type   operatorDn ) {    
       ordinal_type opDn = operatorDn;
 
@@ -42,12 +42,12 @@ namespace Intrepid2 {
       const ordinal_type order = card - 1;
       const double alpha = 0.0, beta = 0.0;
 
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
       
       switch (opType) {
       case OPERATOR_VALUE: {
-        viewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts);     
+        ViewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts);     
 
         Impl::Basis_HGRAD_LINE_Cn_FEM_JACOBI::
           Serial<opType>::getValues(phis, input, order, alpha, beta);
@@ -75,7 +75,7 @@ namespace Intrepid2 {
       case OPERATOR_Dn: {
         // dkcard is always 1 for 1D element
         const ordinal_type dkcard = 1;
-        viewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts, dkcard);     
+        ViewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts, dkcard);     
         Impl::Basis_HGRAD_LINE_Cn_FEM_JACOBI::
           Serial<opType>::getValues(phis, input, order, alpha, beta, opDn);
 
@@ -326,22 +326,64 @@ namespace Intrepid2 {
                               posDfOrd);
     }  
   }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_LINE_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_LINE_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_LINE_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type sizePerPoint = this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_ );
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_GRAD>::getValues( output, input, work, this->vinv_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HGRAD_LINE_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
   
 }// namespace Intrepid2
 
 #endif
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEM.hpp
index 36463fc8baec..7f00c2a90a2d 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEM.hpp
@@ -163,6 +163,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEMDef.hpp
index e25212499fbd..078ad0e97178 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_C1_FEMDef.hpp
@@ -256,6 +256,54 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
-}
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_PYR_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
 
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_PYR_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_PYR_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_PYR_C1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_PYR_C1_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
+}// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEM.hpp
index aad5bae97d00..81a09cc56c91 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEM.hpp
@@ -180,6 +180,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEMDef.hpp
index 3247c5aaa0b7..a6ce124d1c2b 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_PYR_I2_FEMDef.hpp
@@ -371,6 +371,54 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
-}
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_PYR_I2_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
 
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_PYR_I2_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_PYR_I2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_PYR_I2_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_PYR_I2_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
+}// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEM.hpp
index 17af8c1c8685..e6b9d7b10c44 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEM.hpp
@@ -167,6 +167,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEMDef.hpp
index 39504493f063..246a92ac44d5 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C1_FEMDef.hpp
@@ -245,5 +245,63 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_QUAD_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_QUAD_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_QUAD_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_QUAD_C1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_QUAD_C1_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_QUAD_C1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HGRAD_QUAD_C1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEM.hpp
index d82127eb1ead..038214825966 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEM.hpp
@@ -194,6 +194,23 @@ namespace Intrepid2 {
                                 operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEMDef.hpp
index 4ec411a0ffc1..36ae4378091c 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_C2_FEMDef.hpp
@@ -557,5 +557,66 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);   
   }
 
+  template<bool serendipity, typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_QUAD_DEG2_FEM<serendipity,DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<bool serendipity, typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_QUAD_DEG2_FEM<serendipity,DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_QUAD_DEG2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            using SerialValue = typename Impl::Basis_HGRAD_QUAD_DEG2_FEM<serendipity>::template Serial<OPERATOR_VALUE>;
+            SerialValue::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );            
+            using SerialGrad = typename Impl::Basis_HGRAD_QUAD_DEG2_FEM<serendipity>::template Serial<OPERATOR_GRAD>;
+            SerialGrad::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            using SerialCurl = typename Impl::Basis_HGRAD_QUAD_DEG2_FEM<serendipity>::template Serial<OPERATOR_CURL>;
+            SerialCurl::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HGRAD_QUAD_DEG2_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEM.hpp
index d07b46cafb28..00a206c12ec8 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEM.hpp
@@ -181,6 +181,23 @@ namespace Intrepid2 {
                                             operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEMDef.hpp
index 582b465a6c57..9e232352285e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_QUAD_Cn_FEMDef.hpp
@@ -21,18 +21,18 @@ namespace Intrepid2 {
   // -------------------------------------------------------------------------------------
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HGRAD_QUAD_Cn_FEM::Serial<opType>::
+    Basis_HGRAD_QUAD_Cn_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinv,
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv,
                const ordinal_type   operatorDn ) {
       ordinal_type opDn = operatorDn;
       
@@ -43,19 +43,19 @@ namespace Intrepid2 {
       const auto input_x = Kokkos::subview(input, Kokkos::ALL(), range_type(0,1));
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
 
-      const int dim_s = get_dimension_scalar(work);
+      const int dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
       
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+        ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
         
         Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
           getValues(output_x, input_x, work_line, vinv);
@@ -73,29 +73,29 @@ namespace Intrepid2 {
       }
       case OPERATOR_CURL: {
         for (auto l=0;l<2;++l) {
-          viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           
-          viewType output_x, output_y;
+          ViewType output_x, output_y;
           
-          typename workViewType::value_type s = 0.0;
+          typename WorkViewType::value_type s = 0.0;
           if (l) {
             // l = 1
-            output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
+            output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
               getValues(output_x, input_x, work_line, vinv, 1);                           
 
-            output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+            output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
               getValues(output_y, input_y, work_line, vinv);                           
 
             s = -1.0;
           } else {
             // l = 0
-            output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+            output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
               getValues(output_x, input_x, work_line, vinv);                           
 
-            output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
+            output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
               getValues(output_y, input_y, work_line, vinv, 1);                           
 
@@ -122,33 +122,33 @@ namespace Intrepid2 {
       case OPERATOR_D8:
       case OPERATOR_D9:
       case OPERATOR_D10:
-        opDn = getOperatorOrder(opType);
+        opDn = getOperatorOrder(OpType);
       case OPERATOR_Dn: {
         const auto dkcard = opDn + 1;
         for (auto l=0;l<dkcard;++l) {
-          viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           
-          viewType output_x, output_y;
+          ViewType output_x, output_y;
           
           const auto mult_x = opDn - l;
           const auto mult_y = l;
           
           if (mult_x) {
-            output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
+            output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
               getValues(output_x, input_x, work_line, vinv, mult_x);                           
           } else {
-            output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+            output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
               getValues(output_x, input_x, work_line, vinv);                           
           }
 
           if (mult_y) {
-            output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
+            output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
               getValues(output_y, input_y, work_line, vinv, mult_y);                           
           } else {
-            output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+            output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
             Impl::Basis_HGRAD_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
               getValues(output_y, input_y, work_line, vinv);                           
           }
@@ -357,7 +357,72 @@ namespace Intrepid2 {
     this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoordsHost);
     Kokkos::deep_copy(this->dofCoords_, dofCoordsHost);
   }
-  
-}// namespace Intrepid2
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_QUAD_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 3*this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_QUAD_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_QUAD_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type sizePerPoint = 3*this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_QUAD_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_ );
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_QUAD_Cn_FEM::Serial<OPERATOR_GRAD>::getValues( output, input, work, this->vinv_ );
+          });
+          break;
+          case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_QUAD_Cn_FEM::Serial<OPERATOR_CURL>::getValues( output, input, work, this->vinv_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HGRAD_QUAD_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEM.hpp
index a2bd5cce665e..d8ed43bad7c7 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEM.hpp
@@ -164,6 +164,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEMDef.hpp
index 65d632b1b578..c8e0cc996c65 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C1_FEMDef.hpp
@@ -202,5 +202,54 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TET_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TET_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TET_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TET_C1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TET_C1_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEM.hpp
index 3b544a29b8b9..703eef86b224 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEM.hpp
@@ -183,6 +183,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEMDef.hpp
index 36aa3d7df7b8..7f4f39634cd0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_C2_FEMDef.hpp
@@ -323,5 +323,54 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TET_C2_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TET_C2_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TET_C2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TET_C2_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TET_C2_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEM.hpp
index 2bd2814a2d6e..6cb9802be376 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEM.hpp
@@ -203,12 +203,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
-    /** \brief  Returns spatial locations (coordinates) of degrees of freedom on a
-        <strong>reference Tetrahedron</strong>.
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
 
-        \param  DofCoords      [out] - array with the coordinates of degrees of freedom,
-        dimensioned (F,D)
-    */
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEMDef.hpp
index ca7c10d67005..09888eddc924 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_COMP12_FEMDef.hpp
@@ -408,6 +408,55 @@ namespace Intrepid2 {
     this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoords);
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
-}
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TET_COMP12_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TET_COMP12_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TET_COMP12_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TET_COMP12_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TET_COMP12_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
+}// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEM.hpp
index 39b8900291f4..d23d95594bdb 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEM.hpp
@@ -62,53 +62,57 @@ namespace Intrepid2 {
       /**
          \brief See Intrepid2::Basis_HGRAD_TET_Cn_FEM
       */
-      template<EOperator opType>
+      template<EOperator OpType>
       struct Serial {
-        template<typename outputValueViewType,
-                 typename inputPointViewType,
-                 typename workViewType,
-                 typename vinvViewType>
+        template<typename OutputValueViewType,
+                 typename InputPointViewType,
+                 typename WorkViewType,
+                 typename VinvViewType>
         KOKKOS_INLINE_FUNCTION
         static void
-        getValues(       outputValueViewType outputValues,
-                         const inputPointViewType  inputPoints,
-                         workViewType        work,
-                         const vinvViewType        vinv );
+        getValues(       OutputValueViewType outputValues,
+                         const InputPointViewType  inputPoints,
+                         WorkViewType              work,
+                         const VinvViewType        vinv,
+                         const ordinal_type        order);
       };
 
       template<typename DeviceType, ordinal_type numPtsPerEval,
-               typename outputValueValueType, class ...outputValueProperties,
-               typename inputPointValueType,  class ...inputPointProperties,
-               typename vinvValueType,        class ...vinvProperties>
+               typename OutputValueValueType, class ...OutputValueProperties,
+               typename InputPointValueType,  class ...InputPointProperties,
+               typename VinvValueType,        class ...VinvProperties>
       static void
       getValues( const typename DeviceType::execution_space& space,
-                       Kokkos::DynRankView<outputValueValueType,outputValueProperties...> outputValues,
-                 const Kokkos::DynRankView<inputPointValueType, inputPointProperties...>  inputPoints,
-                 const Kokkos::DynRankView<vinvValueType,       vinvProperties...>        vinv,
+                       Kokkos::DynRankView<OutputValueValueType,OutputValueProperties...> outputValues,
+                 const Kokkos::DynRankView<InputPointValueType, InputPointProperties...>  inputPoints,
+                 const Kokkos::DynRankView<VinvValueType,       VinvProperties...>        vinv,
+                 const ordinal_type order,
                  const EOperator operatorType);
 
       /**
          \brief See Intrepid2::Basis_HGRAD_TET_Cn_FEM
       */
-      template<typename outputValueViewType,
-               typename inputPointViewType,
-               typename vinvViewType,
-               typename workViewType,
-               EOperator opType,
+      template<typename OutputValueViewType,
+               typename InputPointViewType,
+               typename VinvViewType,
+               typename WorkViewType,
+               EOperator OpType,
                ordinal_type numPtsEval>
       struct Functor {
-        outputValueViewType _outputValues;
-        const inputPointViewType  _inputPoints;
-        const vinvViewType        _vinv;
-        workViewType        _work;
+        OutputValueViewType _outputValues;
+        const InputPointViewType  _inputPoints;
+        const VinvViewType        _vinv;
+        WorkViewType              _work;
+        const ordinal_type        _order;
 
         KOKKOS_INLINE_FUNCTION
-        Functor(       outputValueViewType outputValues_,
-                       inputPointViewType  inputPoints_,
-                       vinvViewType        vinv_,
-                       workViewType        work_)
+        Functor(       OutputValueViewType outputValues_,
+                       InputPointViewType  inputPoints_,
+                       VinvViewType        vinv_,
+                       WorkViewType        work_,
+                       ordinal_type        order_)
           : _outputValues(outputValues_), _inputPoints(inputPoints_),
-            _vinv(vinv_), _work(work_) {}
+            _vinv(vinv_), _work(work_), _order(order_) {}
 
         KOKKOS_INLINE_FUNCTION
         void operator()(const size_type iter) const {
@@ -118,15 +122,15 @@ namespace Intrepid2 {
           const auto ptRange = Kokkos::pair<ordinal_type,ordinal_type>(ptBegin, ptEnd);
           const auto input   = Kokkos::subview( _inputPoints, ptRange, Kokkos::ALL() );
 
-          typename workViewType::pointer_type ptr = _work.data() + _work.extent(0)*ptBegin*get_dimension_scalar(_work);
+          typename WorkViewType::pointer_type ptr = _work.data() + _work.extent(0)*ptBegin*get_dimension_scalar(_work);
 
           auto vcprop = Kokkos::common_view_alloc_prop(_work);
-          workViewType  work(Kokkos::view_wrap(ptr,vcprop), (ptEnd-ptBegin)*_work.extent(0));
+          WorkViewType  work(Kokkos::view_wrap(ptr,vcprop), (ptEnd-ptBegin)*_work.extent(0));
 
-          switch (opType) {
+          switch (OpType) {
           case OPERATOR_VALUE : {
             auto output = Kokkos::subview( _outputValues, Kokkos::ALL(), ptRange );
-            Serial<opType>::getValues( output, input, work, _vinv );
+            Serial<OpType>::getValues( output, input, work, _vinv, _order );
             break;
           }
           case OPERATOR_GRAD :
@@ -135,7 +139,7 @@ namespace Intrepid2 {
             //case OPERATOR_D3 :
             {
               auto output = Kokkos::subview( _outputValues, Kokkos::ALL(), ptRange, Kokkos::ALL() );
-              Serial<opType>::getValues( output, input, work, _vinv );
+              Serial<OpType>::getValues( output, input, work, _vinv, _order );
               break;
             }
           default: {
@@ -204,9 +208,29 @@ namespace Intrepid2 {
                                             outputValues,
                                             inputPoints,
                                             this->vinv_,
+                                            this->basisDegree_,
                                             operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEMDef.hpp
index c2c5aaf6cbe3..1c8715525bc0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TET_Cn_FEMDef.hpp
@@ -24,44 +24,36 @@ namespace Intrepid2 {
 // -------------------------------------------------------------------------------------
 namespace Impl {
 
-template<EOperator opType>
+template<EOperator OpType>
 template<typename OutputViewType,
-typename inputViewType,
-typename workViewType,
-typename vinvViewType>
+typename InputViewType,
+typename WorkViewType,
+typename VinvViewType>
 KOKKOS_INLINE_FUNCTION
 void
-Basis_HGRAD_TET_Cn_FEM::Serial<opType>::
+Basis_HGRAD_TET_Cn_FEM::Serial<OpType>::
 getValues(       OutputViewType output,
-    const inputViewType  input,
-    workViewType   work,
-    const vinvViewType   vinv ) {
+    const InputViewType  input,
+    WorkViewType         work,
+    const VinvViewType   vinv,
+    const ordinal_type   order ) {
 
   constexpr ordinal_type spaceDim = 3;
   const ordinal_type
   card = vinv.extent(0),
   npts = input.extent(0);
 
-  // compute order
-  ordinal_type order = 0;
-  for (ordinal_type p=0;p<=Parameters::MaxOrder;++p) {
-    if (card == Intrepid2::getPnCardinality<spaceDim>(p)) {
-      order = p;
-      break;
-    }
-  }
-
-  typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-  auto vcprop = Kokkos::common_view_alloc_prop(work);
+  typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+  auto vcprop = Kokkos::common_view_alloc_prop(input);
   auto ptr = work.data();
 
-  switch (opType) {
+  switch (OpType) {
   case OPERATOR_VALUE: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-    viewType dummyView;
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j) {
@@ -73,11 +65,11 @@ getValues(       OutputViewType output,
   }
   case OPERATOR_GRAD:
   case OPERATOR_D1: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, workView, order);
+    Serial<OpType>::getValues(phis, input, workView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -97,12 +89,12 @@ getValues(       OutputViewType output,
   case OPERATOR_D8:
   case OPERATOR_D9:
   case OPERATOR_D10: {
-    const ordinal_type dkcard = getDkCardinality<opType,spaceDim>(); //(orDn + 1);
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
-    viewType dummyView;
+    const ordinal_type dkcard = getDkCardinality<OpType,spaceDim>(); //(orDn + 1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -131,6 +123,7 @@ getValues(
           Kokkos::DynRankView<outputValueValueType,outputValueProperties...> outputValues,
     const Kokkos::DynRankView<inputPointValueType, inputPointProperties...>  inputPoints,
     const Kokkos::DynRankView<vinvValueType,       vinvProperties...>        vinv,
+    const ordinal_type order,
     const EOperator operatorType) {
   typedef          Kokkos::DynRankView<outputValueValueType,outputValueProperties...>         outputValueViewType;
   typedef          Kokkos::DynRankView<inputPointValueType, inputPointProperties...>          inputPointViewType;
@@ -156,7 +149,7 @@ getValues(
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TET_Cn_FEM::getValues::work", vcprop), cardinality, inputPoints.extent(0));
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_VALUE,numPtsPerEval> FunctorType;
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
   case OPERATOR_GRAD:
@@ -164,23 +157,16 @@ getValues(
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TET_Cn_FEM::getValues::work", vcprop), cardinality*(2*spaceDim+1), inputPoints.extent(0));
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_D1,numPtsPerEval> FunctorType;
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
   case OPERATOR_D2: {
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_D2,numPtsPerEval> FunctorType;
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TET_Cn_FEM::getValues::work", vcprop), cardinality*outputValues.extent(2), inputPoints.extent(0));
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
-  /*  case OPERATOR_D3: {
-        typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType
-            OPERATOR_D3,numPtsPerEval> FunctorType;
-        workViewType  work(Kokkos::view_alloc("Basis_HGRAD_TET_Cn_FEM::getValues::work", vcprop), cardinality, inputPoints.extent(0), outputValues.extent(2));
-        Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
-        break;
-      }*/
   default: {
     INTREPID2_TEST_FOR_EXCEPTION( true , std::invalid_argument,
         ">>> ERROR (Basis_HGRAD_TET_Cn_FEM): Operator type not implemented" );
@@ -431,5 +417,65 @@ Basis_HGRAD_TET_Cn_FEM( const ordinal_type order,
         posDfOrd);
   }
 }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TET_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = getWorkSizePerPoint(operatorType)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TET_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+      
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TET_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      constexpr ordinal_type spaceDim = 3;
+      auto sizePerPoint = (operatorType==OPERATOR_VALUE) ? 
+                          this->vinv_.extent(0)*get_dimension_scalar(inputPoints) : 
+                          (2*spaceDim+1)*this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_TET_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_TET_Cn_FEM::Serial<OPERATOR_GRAD>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HGRAD_TET_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
 } // namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEM.hpp
index 46349310b210..81439bdf3f50 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEM.hpp
@@ -162,6 +162,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEMDef.hpp
index e771ae7a3ee0..bc926788c290 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C1_FEMDef.hpp
@@ -204,5 +204,63 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TRI_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TRI_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TRI_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TRI_C1_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TRI_C1_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TRI_C1_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HGRAD_TRI_C1_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEM.hpp
index 9eb45f9c2716..627fa113720e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEM.hpp
@@ -173,6 +173,23 @@ namespace Intrepid2 {
                               operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEMDef.hpp
index 87bb96e2bbe0..86df77f41a27 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_C2_FEMDef.hpp
@@ -263,5 +263,63 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TRI_C2_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TRI_C2_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TRI_C2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TRI_C2_FEM::Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TRI_C2_FEM::Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_TRI_C2_FEM::Serial<OPERATOR_CURL>::getValues( output, input);
+          });
+          break;
+        default: {
+          INTREPID2_TEST_FOR_ABORT( true, ">>> ERROR: (Intrepid2::Basis_HGRAD_TRI_C2_FEM::getValues), Operator Type not supported.");
+        }
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM.hpp
index c8bc97c3fb76..17ada895efe0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM.hpp
@@ -60,53 +60,57 @@ namespace Intrepid2 {
          work is a rank 1 view having the same value_type of inputPoints
          and having size equal to getWorkSizePerPoint()*inputPoints.extent(0);
       */
-      template<EOperator opType>
+      template<EOperator OpType>
       struct Serial {
-        template<typename outputValueViewType,
-                 typename inputPointViewType,
-                 typename workViewType,
-                 typename vinvViewType>
+        template<typename OutputValueViewType,
+                 typename InputPointViewType,
+                 typename WorkViewType,
+                 typename VinvViewType>
         KOKKOS_INLINE_FUNCTION
         static void
-        getValues(       outputValueViewType outputValues,
-                         const inputPointViewType  inputPoints,
-                         workViewType        work,
-                         const vinvViewType        vinv );
+        getValues(       OutputValueViewType outputValues,
+                         const InputPointViewType  inputPoints,
+                         WorkViewType              work,
+                         const VinvViewType        vinv,
+                         const ordinal_type        order);
       };
 
       template<typename DeviceType, ordinal_type numPtsPerEval,
-               typename outputValueValueType, class ...outputValueProperties,
-               typename inputPointValueType,  class ...inputPointProperties,
-               typename vinvValueType,        class ...vinvProperties>
+               typename OutputValueValueType, class ...OutputValueProperties,
+               typename InputPointValueType,  class ...InputPointProperties,
+               typename VinvValueType,        class ...VinvProperties>
       static void
-      getValues(const typename DeviceType::execution_space& space,
-                      Kokkos::DynRankView<outputValueValueType,outputValueProperties...> outputValues,
-                const Kokkos::DynRankView<inputPointValueType, inputPointProperties...>  inputPoints,
-                const Kokkos::DynRankView<vinvValueType,       vinvProperties...>        vinv,
-                const EOperator operatorType);
+      getValues( const typename DeviceType::execution_space& space,
+                       Kokkos::DynRankView<OutputValueValueType,OutputValueProperties...> outputValues,
+                 const Kokkos::DynRankView<InputPointValueType, InputPointProperties...>  inputPoints,
+                 const Kokkos::DynRankView<VinvValueType,       VinvProperties...>        vinv,
+                 const ordinal_type order,
+                 const EOperator operatorType);
 
       /**
          \brief See Intrepid2::Basis_HGRAD_TRI_Cn_FEM
       */
-      template<typename outputValueViewType,
-               typename inputPointViewType,
-               typename vinvViewType,
-               typename workViewType,
-               EOperator opType,
+      template<typename OutputValueViewType,
+               typename InputPointViewType,
+               typename VinvViewType,
+               typename WorkViewType,
+               EOperator OpType,
                ordinal_type numPtsEval>
       struct Functor {
-        outputValueViewType _outputValues;
-        const inputPointViewType  _inputPoints;
-        const vinvViewType        _vinv;
-        workViewType        _work;
+        OutputValueViewType _outputValues;
+        const InputPointViewType  _inputPoints;
+        const VinvViewType        _vinv;
+        WorkViewType              _work;
+        const ordinal_type        _order;
 
         KOKKOS_INLINE_FUNCTION
-        Functor(       outputValueViewType outputValues_,
-                       inputPointViewType  inputPoints_,
-                       vinvViewType        vinv_,
-                       workViewType        work_)
+        Functor(       OutputValueViewType outputValues_,
+                       InputPointViewType  inputPoints_,
+                       VinvViewType        vinv_,
+                       WorkViewType        work_,
+                       ordinal_type        order_)
           : _outputValues(outputValues_), _inputPoints(inputPoints_),
-            _vinv(vinv_), _work(work_) {}
+            _vinv(vinv_), _work(work_), _order(order_) {}
 
         KOKKOS_INLINE_FUNCTION
         void operator()(const size_type iter) const {
@@ -116,22 +120,22 @@ namespace Intrepid2 {
           const auto ptRange = Kokkos::pair<ordinal_type,ordinal_type>(ptBegin, ptEnd);
           const auto input   = Kokkos::subview( _inputPoints, ptRange, Kokkos::ALL() );
 
-          typename workViewType::pointer_type ptr = _work.data() + _work.extent(0)*ptBegin*get_dimension_scalar(_work);
+          typename WorkViewType::pointer_type ptr = _work.data() + _work.extent(0)*ptBegin*get_dimension_scalar(_work);
 
           auto vcprop = Kokkos::common_view_alloc_prop(_work);
-          workViewType  work(Kokkos::view_wrap(ptr,vcprop), (ptEnd-ptBegin)*_work.extent(0));
+          WorkViewType  work(Kokkos::view_wrap(ptr,vcprop), (ptEnd-ptBegin)*_work.extent(0));
 
-          switch (opType) {
+          switch (OpType) {
           case OPERATOR_VALUE : {
             auto output = Kokkos::subview( _outputValues, Kokkos::ALL(), ptRange );
-            Serial<opType>::getValues( output, input, work, _vinv );
+            Serial<OpType>::getValues( output, input, work, _vinv, _order );
             break;
           }
           case OPERATOR_CURL:
           case OPERATOR_D1:
           case OPERATOR_D2:  {
             auto output = Kokkos::subview( _outputValues, Kokkos::ALL(), ptRange, Kokkos::ALL() );
-            Serial<opType>::getValues( output, input, work, _vinv );
+            Serial<OpType>::getValues( output, input, work, _vinv, _order );
             break;
           }
           default: {
@@ -200,9 +204,29 @@ namespace Intrepid2 {
                                             outputValues,
                                             inputPoints,
                                             this->vinv_,
+                                            this->basisDegree_,
                                             operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEMDef.hpp
index 681148713a06..c7b7a40cfa7b 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEMDef.hpp
@@ -23,44 +23,36 @@ namespace Intrepid2 {
 // -------------------------------------------------------------------------------------
 namespace Impl {
 
-template<EOperator opType>
+template<EOperator OpType>
 template<typename OutputViewType,
-typename inputViewType,
-typename workViewType,
-typename vinvViewType>
+typename InputViewType,
+typename WorkViewType,
+typename VinvViewType>
 KOKKOS_INLINE_FUNCTION
 void
-Basis_HGRAD_TRI_Cn_FEM::Serial<opType>::
+Basis_HGRAD_TRI_Cn_FEM::Serial<OpType>::
 getValues(       OutputViewType output,
-    const inputViewType  input,
-    workViewType   work,
-    const vinvViewType   vinv ) {
+    const InputViewType  input,
+    WorkViewType         work,
+    const VinvViewType   vinv,
+    const ordinal_type   order ) {
 
   constexpr ordinal_type spaceDim = 2;
   const ordinal_type
   card = vinv.extent(0),
   npts = input.extent(0);
 
-  // compute order
-  ordinal_type order = 0;
-  for (ordinal_type p=0;p<=Parameters::MaxOrder;++p) {
-    if (card == Intrepid2::getPnCardinality<spaceDim>(p) ) {
-      order = p;
-      break;
-    }
-  }
-
-  typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-  auto vcprop = Kokkos::common_view_alloc_prop(work);
+  typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+  auto vcprop = Kokkos::common_view_alloc_prop(input);
   auto ptr = work.data();
 
-  switch (opType) {
+  switch (OpType) {
   case OPERATOR_VALUE: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-    viewType dummyView;
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j) {
@@ -72,11 +64,11 @@ getValues(       OutputViewType output,
   }
   case OPERATOR_GRAD:
   case OPERATOR_D1: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, workView, order);
+    Serial<OpType>::getValues(phis, input, workView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -96,12 +88,12 @@ getValues(       OutputViewType output,
   case OPERATOR_D8:
   case OPERATOR_D9:
   case OPERATOR_D10: {
-    const ordinal_type dkcard = getDkCardinality<opType,spaceDim>(); //(orDn + 1);
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
-    viewType dummyView;
+    const ordinal_type dkcard = getDkCardinality<OpType,spaceDim>(); //(orDn + 1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -113,9 +105,9 @@ getValues(       OutputViewType output,
     break;
   }
   case OPERATOR_CURL: { // only works in 2d. first component is -d/dy, second is d/dx
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
@@ -150,6 +142,7 @@ getValues(
           Kokkos::DynRankView<outputValueValueType,outputValueProperties...> outputValues,
     const Kokkos::DynRankView<inputPointValueType, inputPointProperties...>  inputPoints,
     const Kokkos::DynRankView<vinvValueType,       vinvProperties...>        vinv,
+    const ordinal_type order,
     const EOperator operatorType) {
   typedef          Kokkos::DynRankView<outputValueValueType,outputValueProperties...>         outputValueViewType;
   typedef          Kokkos::DynRankView<inputPointValueType, inputPointProperties...>          inputPointViewType;
@@ -175,7 +168,7 @@ getValues(
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TRI_Cn_FEM::getValues::work", vcprop), cardinality, inputPoints.extent(0));
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_VALUE,numPtsPerEval> FunctorType;
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
   case OPERATOR_GRAD:
@@ -183,30 +176,23 @@ getValues(
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TRI_Cn_FEM::getValues::work", vcprop), cardinality*(2*spaceDim+1), inputPoints.extent(0));
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_D1,numPtsPerEval> FunctorType;
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
   case OPERATOR_CURL: {
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TRI_Cn_FEM::getValues::work", vcprop), cardinality*(2*spaceDim+1), inputPoints.extent(0));
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_CURL,numPtsPerEval> FunctorType;
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
   case OPERATOR_D2: {
     typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType,
         OPERATOR_D2,numPtsPerEval> FunctorType;
     workViewType  work(Kokkos::view_alloc(space, "Basis_HGRAD_TRI_Cn_FEM::getValues::work", vcprop), cardinality*outputValues.extent(2), inputPoints.extent(0));
-    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
+    Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work, order) );
     break;
   }
-  /*  case OPERATOR_D3: {
-        typedef Functor<outputValueViewType,inputPointViewType,vinvViewType, workViewType
-            OPERATOR_D3,numPtsPerEval> FunctorType;
-        workViewType  work(Kokkos::view_alloc("Basis_HGRAD_TRI_Cn_FEM::getValues::work", vcprop), cardinality, inputPoints.extent(0), outputValues.extent(2));
-        Kokkos::parallel_for( policy, FunctorType(outputValues, inputPoints, vinv, work) );
-        break;
-      }*/
   default: {
     INTREPID2_TEST_FOR_EXCEPTION( true , std::invalid_argument,
         ">>> ERROR (Basis_HGRAD_TRI_Cn_FEM): Operator type not implemented" );
@@ -242,7 +228,7 @@ Basis_HGRAD_TRI_Cn_FEM( const ordinal_type order,
   PointTools::getLattice( dofCoords,
       cellTopo,
       order, offset,
-      pointType_ );
+      this->pointType_ );
 
   this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoords);
   Kokkos::deep_copy(this->dofCoords_, dofCoords);
@@ -384,5 +370,74 @@ Basis_HGRAD_TRI_Cn_FEM( const ordinal_type order,
         posDfOrd);
   }
 }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_TRI_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = getWorkSizePerPoint(operatorType)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_TRI_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+      
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_TRI_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      constexpr ordinal_type spaceDim = 2;
+      auto sizePerPoint = (operatorType==OPERATOR_VALUE) ? 
+                          this->vinv_.extent(0)*get_dimension_scalar(inputPoints) : 
+                          (2*spaceDim+1)*this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_TRI_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_TRI_Cn_FEM::Serial<OPERATOR_GRAD>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+          case OPERATOR_CURL:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type(pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt,pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HGRAD_TRI_Cn_FEM::Serial<OPERATOR_CURL>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HGRAD_TRI_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
 } // namespace Intrepid2
+
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM_ORTHDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM_ORTHDef.hpp
index 6f5e5abb1dd0..a2b4271518ba 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM_ORTHDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEM_ORTHDef.hpp
@@ -198,69 +198,8 @@ void OrthPolynomialTri<OutputViewType,inputViewType,workViewType,hasDeriv,n>::ge
     const inputViewType  /* input */,
     workViewType   /* work */,
     const ordinal_type   /* order */ ) {
-#if 0   //#ifdef HAVE_INTREPID2_SACADO
-
-constexpr ordinal_type spaceDim = 2;
-constexpr ordinal_type maxCard = Intrepid2::getPnCardinality<spaceDim, Parameters::MaxOrder>();
-
-typedef typename OutputViewType::value_type value_type;
-typedef Sacado::Fad::SFad<value_type,spaceDim> fad_type;
-
-const ordinal_type
-npts = input.extent(0),
-card = output.extent(0);
-
-// use stack buffer
-fad_type inBuf[Parameters::MaxNumPtsPerBasisEval][spaceDim],
-outBuf[maxCard][Parameters::MaxNumPtsPerBasisEval][n];
-
-typedef typename inputViewType::memory_space memory_space;
-typedef typename Kokkos::View<fad_type***, memory_space> outViewType;
-typedef typename Kokkos::View<fad_type**, memory_space> inViewType;
-auto vcprop = Kokkos::common_view_alloc_prop(input);
-
-inViewType in(Kokkos::view_wrap((value_type*)&inBuf[0][0], vcprop), npts, spaceDim);
-outViewType out(Kokkos::view_wrap((value_type*)&outBuf[0][0][0], vcprop), card, npts, n);
-
-for (ordinal_type i=0;i<npts;++i)
-  for (ordinal_type j=0;j<spaceDim;++j) {
-    in(i,j) = input(i,j);
-    in(i,j).diff(j,spaceDim);
-  }
-
-typedef typename Kokkos::DynRankView<fad_type, memory_space> outViewType_;
-outViewType_ workView;
-if (n==2) {
-  //char outBuf[bufSize*sizeof(typename inViewType::value_type)];
-  fad_type outBuf[maxCard][Parameters::MaxNumPtsPerBasisEval][spaceDim+1];
-  auto vcprop = Kokkos::common_view_alloc_prop(in);
-  workView = outViewType_( Kokkos::view_wrap((value_type*)&outBuf[0][0][0], vcprop), card, npts, spaceDim+1);
-}
-OrthPolynomialTri<outViewType,inViewType,outViewType_,hasDeriv,n-1>::generate(out, in, workView, order);
-
-for (ordinal_type i=0;i<card;++i)
-  for (ordinal_type j=0;j<npts;++j) {
-    for (ordinal_type i_dx = 0; i_dx <= n; ++i_dx) {
-      ordinal_type i_dy =  n-i_dx;
-      ordinal_type i_Dn = i_dy;
-      if(i_dx > 0) {
-        //n=2:  (f_x)_x, (f_y)_x
-        //n=3:  (f_xx)_x, (f_xy)_x, (f_yy)_x
-        ordinal_type i_Dnm1 = i_dy;
-        output.access(i,j,i_Dn) = out(i,j,i_Dnm1).dx(0);
-      }
-      else {
-        //n=2:  (f_y)_y, (f_z)_y
-        //n=3:  (f_yy)_y
-        ordinal_type i_Dnm1 = i_dy-1;
-        output.access(i,j,i_Dn) = out(i,j,i_Dnm1).dx(1);
-      }
-    }
-  }
-#else
 INTREPID2_TEST_FOR_ABORT( true,
     ">>> ERROR: (Intrepid2::Basis_HGRAD_TRI_Cn_FEM_ORTH::OrthPolynomialTri) Computing of second and higher-order derivatives is not currently supported");
-#endif
 }
 
 
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEM.hpp
index 15daedfbfe49..5b8b73634bb4 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEM.hpp
@@ -166,6 +166,23 @@ namespace Intrepid2 {
                                   operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEMDef.hpp
index 9d2c461edca2..8d76318a49e0 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C1_FEMDef.hpp
@@ -245,5 +245,54 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_WEDGE_C1_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_WEDGE_C1_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_WEDGE_C1_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_WEDGE_C1_FEM::template Serial<OPERATOR_VALUE>::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            Impl::Basis_HGRAD_WEDGE_C1_FEM::template Serial<OPERATOR_GRAD>::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+  
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEM.hpp
index d4cb38e7ca55..c952afcf0e6e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEM.hpp
@@ -215,6 +215,23 @@ namespace Intrepid2 {
                                     operatorType );;
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEMDef.hpp
index 363d21ad19ea..9f5327d94187 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_WEDGE_C2_FEMDef.hpp
@@ -30,12 +30,13 @@ namespace Intrepid2 {
     Basis_HGRAD_WEDGE_DEG2_FEM<serendipity>::Serial<opType>::
     getValues(       OutputViewType output,
                const inputViewType input ) {
+      typedef typename inputViewType::value_type value_type;
       switch (opType) {
       case OPERATOR_VALUE: {
-        const auto x = input(0);
-        const auto y = input(1);
-        const auto z = input(2);
-        const auto w = 1.0 - x - y;
+        const value_type x = input(0);
+        const value_type y = input(1);
+        const value_type z = input(2);
+        const value_type w = 1.0 - x - y;
 
         // output is a rank-1 array with dimensions (basisCardinality_)
         if constexpr (!serendipity) {
@@ -80,9 +81,9 @@ namespace Intrepid2 {
         break;
       }
       case OPERATOR_GRAD: {
-        const auto x = input(0);
-        const auto y = input(1);
-        const auto z = input(2);
+        const value_type x = input(0);
+        const value_type y = input(1);
+        const value_type z = input(2);
 
         if constexpr (!serendipity) {
           output.access(0, 0) = ((-3 + 4*x + 4*y)*(-1 + z)*z)/2.;
@@ -158,7 +159,7 @@ namespace Intrepid2 {
           output.access(17, 1) =  4*(-1 + x + 2*y)*(-1 + z*z);
           output.access(17, 2) =  8*y*(-1 + x + y)*z;
         } else {
-          const auto w = 1.0 - x - y;
+          const value_type w = 1.0 - x - y;
       
           output.access(0, 0) = -(2.0*w - 1.0 - 0.5*z)*(1.0 - z);
           output.access(0, 1) = -(2.0*w - 1.0 - 0.5*z)*(1.0 - z);
@@ -223,9 +224,9 @@ namespace Intrepid2 {
         break;
       }
       case OPERATOR_D2: {
-        const auto x = input(0);
-        const auto y = input(1);
-        const auto z = input(2);
+        const value_type x = input(0);
+        const value_type y = input(1);
+        const value_type z = input(2);
 
         if constexpr (!serendipity) {
           output.access(0, 0) =  2.*(-1. + z)*z;
@@ -356,7 +357,7 @@ namespace Intrepid2 {
 
         } else { //serendipity element
         
-          const auto w = 1.0 - x - y;
+          const value_type w = 1.0 - x - y;
           output.access(0, 0) =  2.0*(1.0 - z);     
           output.access(0, 1) =  2.0*(1.0 - z);     
           output.access(0, 2) =  2.0*w - 0.5 - z;
@@ -466,9 +467,9 @@ namespace Intrepid2 {
       }
       case OPERATOR_D3: {
         if constexpr (!serendipity) {
-          const auto x = input(0);
-          const auto y = input(1);
-          const auto z = input(2);
+          const value_type x = input(0);
+          const value_type y = input(1);
+          const value_type z = input(2);
 
           output.access(0, 0) =  0.;
           output.access(0, 1) =  0.;
@@ -1082,5 +1083,56 @@ namespace Intrepid2 {
     Kokkos::deep_copy(this->dofCoords_, dofCoords);
   }
 
+  template<bool serendipity, typename DT, typename OT, typename PT>
+  void 
+  Basis_HGRAD_WEDGE_DEG2_FEM<serendipity,DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 0;
+  }
+
+  template<bool serendipity, typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HGRAD_WEDGE_DEG2_FEM<serendipity,DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim <= 0) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HGRAD_WEDGE_DEG2_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      (void) scratchStorage; //avoid unused variable warning
+
+      const int numPoints = inputPoints.extent(0);
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            using SerialValue = typename Impl::Basis_HGRAD_WEDGE_DEG2_FEM<serendipity>::template Serial<OPERATOR_VALUE>;
+            SerialValue::getValues( output, input);
+          });
+          break;
+        case OPERATOR_GRAD:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), pt, Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 pt, Kokkos::ALL() );
+            using SerialGrad = typename Impl::Basis_HGRAD_WEDGE_DEG2_FEM<serendipity>::template Serial<OPERATOR_GRAD>;
+            SerialGrad::getValues( output, input);
+          });
+          break;
+        default: {}
+    }
+  }
+
 }// namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEM.hpp
index 96cf0a64405b..388eb9ccdd1b 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEM.hpp
@@ -137,20 +137,22 @@ namespace Intrepid2 {
   class Basis_HVOL_HEX_Cn_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
     Basis_HVOL_HEX_Cn_FEM(const ordinal_type order,
                             const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -172,6 +174,23 @@ namespace Intrepid2 {
                                                 operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
 
     virtual
     void
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEMDef.hpp
index 652df8ee2689..617eeb9cad84 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_HEX_Cn_FEMDef.hpp
@@ -9,7 +9,7 @@
 
 /** \file   Intrepid2_HVOL_HEX_Cn_FEMDef.hpp
     \brief  Definition file for FEM basis functions of degree n for H(vol) functions on HEX cells
-    \author Created by M. Perego, based on the Intrepid2::HGRAD_HEX_Cn_FEM class
+    \author Created by M. Perego, based on the Intrepid2::HVOL_HEX_Cn_FEM class
 */
 
 #ifndef __INTREPID2_HVOL_HEX_CN_FEMDEF_HPP__
@@ -20,18 +20,18 @@ namespace Intrepid2 {
   // -------------------------------------------------------------------------------------
   namespace Impl {
     
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HVOL_HEX_Cn_FEM::Serial<opType>::
+    Basis_HVOL_HEX_Cn_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinv,
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv,
                const ordinal_type   operatorDn ) {
       ordinal_type opDn = operatorDn;
 
@@ -43,21 +43,21 @@ namespace Intrepid2 {
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
       const auto input_z = Kokkos::subview(input, Kokkos::ALL(), range_type(2,3));
 
-      const ordinal_type dim_s = get_dimension_scalar(work);
+      const ordinal_type dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       auto ptr3 = work.data()+3*cardLine*npts*dim_s;      
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
-        viewType output_z(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts);
+        ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+        ViewType output_z(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts);
         
         Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
           getValues(output_x, input_x, work_line, vinv);
@@ -88,7 +88,7 @@ namespace Intrepid2 {
       case OPERATOR_D8:
       case OPERATOR_D9:
       case OPERATOR_D10:
-        opDn = getOperatorOrder(opType);
+        opDn = getOperatorOrder(OpType);
       case OPERATOR_Dn: {
         const ordinal_type dkcard = opDn + 1;
 
@@ -105,35 +105,35 @@ namespace Intrepid2 {
             if (mult_x < 0) {
               // pass
             } else {              
-              viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+              ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
               decltype(work_line)  output_x, output_y, output_z;
                   
               if (mult_x) {
-                output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
+                output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
                 Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
                   getValues(output_x, input_x, work_line, vinv, mult_x);
               } else {
-                output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+                output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
                 Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
                   getValues(output_x, input_x, work_line, vinv);
               }
               
               if (mult_y) {
-                output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
+                output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
                 Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
                   getValues(output_y, input_y, work_line, vinv, mult_y);
               } else {
-                output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+                output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
                 Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
                   getValues(output_y, input_y, work_line, vinv);
               }
 
               if (mult_z) {
-                output_z = viewType(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts, 1);
+                output_z = ViewType(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts, 1);
                 Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
                   getValues(output_z, input_z, work_line, vinv, mult_z);
               } else {
-                output_z = viewType(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts);
+                output_z = ViewType(Kokkos::view_wrap(ptr3, vcprop), cardLine, npts);
                 Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
                   getValues(output_z, input_z, work_line, vinv);
               }
@@ -316,7 +316,55 @@ namespace Intrepid2 {
     this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoordsHost);
     Kokkos::deep_copy(this->dofCoords_, dofCoordsHost);
   }
-  
-}// namespace Intrepid2
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HVOL_HEX_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 4*this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HVOL_HEX_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+      
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HVOL_HEX_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      auto sizePerPoint = 4*this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HVOL_HEX_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HVOL_HEX_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEM.hpp
index 380438f33bb4..0be4ce27fba8 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEM.hpp
@@ -145,15 +145,16 @@ namespace Intrepid2 {
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
     using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+
     using HostBasis = Basis_HVOL_LINE_Cn_FEM<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>;
-    
-    using OrdinalTypeArray1DHost = typename BasisBase::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename BasisBase::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename BasisBase::OrdinalTypeArray3DHost;
-    
-    using OutputViewType = typename BasisBase::OutputViewType;
-    using PointViewType  = typename BasisBase::PointViewType ;
-    using ScalarViewType = typename BasisBase::ScalarViewType;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
@@ -182,6 +183,23 @@ namespace Intrepid2 {
                                                 operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPointsconst,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEMDef.hpp
index 3d742d4a30a4..dc8f25d3cd7e 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_LINE_Cn_FEMDef.hpp
@@ -9,7 +9,7 @@
 
 /** \file   Intrepid2_HVOL_LINE_Cn_FEMDef.hpp
     \brief  Definition file for FEM basis functions of degree n for H(vol) functions on LINE.
-    \author Created by M. Perego, based on the Intrepid2::HGRAD_LINE_Cn_FEM class
+    \author Created by M. Perego, based on the Intrepid2::HVOL_LINE_Cn_FEM class
 */
 
 #ifndef __INTREPID2_HVOL_LINE_CN_FEM_DEF_HPP__
@@ -22,16 +22,16 @@ namespace Intrepid2 {
     
     template<EOperator opType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
     Basis_HVOL_LINE_Cn_FEM::Serial<opType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinv,
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv,
                const ordinal_type   operatorDn ) {    
       ordinal_type opDn = operatorDn;
 
@@ -41,12 +41,12 @@ namespace Intrepid2 {
       const ordinal_type order = card - 1;
       const double alpha = 0.0, beta = 0.0;
 
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
       
       switch (opType) {
       case OPERATOR_VALUE: {
-        viewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts);     
+        ViewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts);     
 
         Impl::Basis_HGRAD_LINE_Cn_FEM_JACOBI::
           Serial<opType>::getValues(phis, input, order, alpha, beta);
@@ -74,7 +74,7 @@ namespace Intrepid2 {
       case OPERATOR_Dn: {
         // dkcard is always 1 for 1D element
         const ordinal_type dkcard = 1;
-        viewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts, dkcard);     
+        ViewType phis(Kokkos::view_wrap(work.data(), vcprop), card, npts, dkcard);     
         Impl::Basis_HGRAD_LINE_Cn_FEM_JACOBI::
           Serial<opType>::getValues(phis, input, order, alpha, beta, opDn);
 
@@ -289,22 +289,56 @@ namespace Intrepid2 {
                               posDfOrd);
     }  
   }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HVOL_LINE_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HVOL_LINE_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HVOL_LINE_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      ordinal_type sizePerPoint = this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_ );
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HVOL_LINE_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
   
 }// namespace Intrepid2
 
 #endif
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEM.hpp
index 496522a6278c..6329c3cb30ce 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEM.hpp
@@ -132,20 +132,22 @@ namespace Intrepid2 {
   class Basis_HVOL_QUAD_Cn_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
     Basis_HVOL_QUAD_Cn_FEM(const ordinal_type order,
                             const EPointType   pointType = POINTTYPE_EQUISPACED);
 
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -167,6 +169,24 @@ namespace Intrepid2 {
                                                 operatorType );
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEMDef.hpp
index 2a9e2678b771..f492b6a65f7c 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_QUAD_Cn_FEMDef.hpp
@@ -19,22 +19,22 @@ namespace Intrepid2 {
 
   // -------------------------------------------------------------------------------------
   namespace Impl {
-
-    template<EOperator opType>
+    
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HVOL_QUAD_Cn_FEM::Serial<opType>::
+    Basis_HVOL_QUAD_Cn_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinv,
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv,
                const ordinal_type   operatorDn ) {
       ordinal_type opDn = operatorDn;
-      
+
       const ordinal_type cardLine = vinv.extent(0);
       const ordinal_type npts = input.extent(0);
 
@@ -42,19 +42,19 @@ namespace Intrepid2 {
       const auto input_x = Kokkos::subview(input, Kokkos::ALL(), range_type(0,1));
       const auto input_y = Kokkos::subview(input, Kokkos::ALL(), range_type(1,2));
 
-      const int dim_s = get_dimension_scalar(work);
+      const ordinal_type dim_s = get_dimension_scalar(input);
       auto ptr0 = work.data();
       auto ptr1 = work.data()+cardLine*npts*dim_s;
       auto ptr2 = work.data()+2*cardLine*npts*dim_s;
       
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
-      
-      switch (opType) {
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
+
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
-        viewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
-        viewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+        ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+        ViewType output_x(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+        ViewType output_y(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
         
         Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
           getValues(output_x, input_x, work_line, vinv);
@@ -81,33 +81,33 @@ namespace Intrepid2 {
       case OPERATOR_D8:
       case OPERATOR_D9:
       case OPERATOR_D10:
-        opDn = getOperatorOrder(opType);
+        opDn = getOperatorOrder(OpType);
       case OPERATOR_Dn: {
         const auto dkcard = opDn + 1;
         for (auto l=0;l<dkcard;++l) {
-          viewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
+          ViewType work_line(Kokkos::view_wrap(ptr0, vcprop), cardLine, npts);
           
-          viewType output_x, output_y;
+          ViewType output_x, output_y;
           
           const auto mult_x = opDn - l;
           const auto mult_y = l;
           
           if (mult_x) {
-            output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
+            output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts, 1);
             Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
               getValues(output_x, input_x, work_line, vinv, mult_x);                           
           } else {
-            output_x = viewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
+            output_x = ViewType(Kokkos::view_wrap(ptr1, vcprop), cardLine, npts);
             Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
               getValues(output_x, input_x, work_line, vinv);                           
           }
 
           if (mult_y) {
-            output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
+            output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts, 1);
             Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_Dn>::
               getValues(output_y, input_y, work_line, vinv, mult_y);                           
           } else {
-            output_y = viewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
+            output_y = ViewType(Kokkos::view_wrap(ptr2, vcprop), cardLine, npts);
             Impl::Basis_HVOL_LINE_Cn_FEM::Serial<OPERATOR_VALUE>::
               getValues(output_y, input_y, work_line, vinv);                           
           }
@@ -282,7 +282,55 @@ namespace Intrepid2 {
     this->dofCoords_ = Kokkos::create_mirror_view(typename DT::memory_space(), dofCoordsHost);
     Kokkos::deep_copy(this->dofCoords_, dofCoordsHost);
   }
-  
-}
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HVOL_QUAD_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = 3*this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HVOL_QUAD_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+      
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HVOL_QUAD_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      auto sizePerPoint = 3*this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HVOL_QUAD_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_, this->basisDegree_);
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HVOL_QUAD_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
+} // namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEM.hpp
index d47afbf7724f..8f9010f619b8 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEM.hpp
@@ -156,23 +156,23 @@ namespace Intrepid2 {
   class Basis_HVOL_TET_Cn_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
     Basis_HVOL_TET_Cn_FEM(const ordinal_type order,
                            const EPointType   pointType = POINTTYPE_EQUISPACED);
-
-
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-    typedef typename Basis<DeviceType,outputValueType,pointValueType>::scalarType  scalarType;
-
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    
+    using scalarType = typename BasisBase::scalarType;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -194,6 +194,24 @@ namespace Intrepid2 {
                                                 operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEMDef.hpp
index a0945a008159..7927a1e124f6 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TET_Cn_FEMDef.hpp
@@ -23,18 +23,18 @@ namespace Intrepid2 {
 
   namespace Impl {
 
-    template<EOperator opType>
+    template<EOperator OpType>
     template<typename OutputViewType,
-             typename inputViewType,
-             typename workViewType,
-             typename vinvViewType>
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
     KOKKOS_INLINE_FUNCTION
     void
-    Basis_HVOL_TET_Cn_FEM::Serial<opType>::
+    Basis_HVOL_TET_Cn_FEM::Serial<OpType>::
     getValues(       OutputViewType output,
-               const inputViewType  input,
-                     workViewType   work,
-               const vinvViewType   vinv ) {
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv ) {
 
       constexpr ordinal_type spaceDim = 3;
       const ordinal_type
@@ -50,17 +50,17 @@ namespace Intrepid2 {
         }
       }
 
-      typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-      auto vcprop = Kokkos::common_view_alloc_prop(work);
+      typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+      auto vcprop = Kokkos::common_view_alloc_prop(input);
       auto ptr = work.data();
 
-      switch (opType) {
+      switch (OpType) {
       case OPERATOR_VALUE: {
-        const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-        workViewType dummyView;
+        const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+        ViewType dummyView;
 
         Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-          Serial<opType>::getValues(phis, input, dummyView, order);
+          Serial<OpType>::getValues(phis, input, dummyView, order);
 
         for (ordinal_type i=0;i<card;++i)
           for (ordinal_type j=0;j<npts;++j) {
@@ -72,11 +72,11 @@ namespace Intrepid2 {
       }
       case OPERATOR_GRAD:
       case OPERATOR_D1: {
-        const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-        ptr += card*npts*spaceDim*get_dimension_scalar(work);
-        const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+        const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+        ptr += card*npts*spaceDim*get_dimension_scalar(input);
+        const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
         Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-          Serial<opType>::getValues(phis, input, workView, order);
+          Serial<OpType>::getValues(phis, input, workView, order);
 
         for (ordinal_type i=0;i<card;++i)
           for (ordinal_type j=0;j<npts;++j)
@@ -96,12 +96,13 @@ namespace Intrepid2 {
       case OPERATOR_D8:
       case OPERATOR_D9:
       case OPERATOR_D10: {
-        const ordinal_type dkcard = getDkCardinality<opType,spaceDim>(); //(orDn + 1);
-        const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
-        workViewType dummyView;
+        const ordinal_type dkcard = getDkCardinality<OpType,spaceDim>(); //(orDn + 1);
+        const 
+        ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
+        ViewType dummyView;
 
         Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
-          Serial<opType>::getValues(phis, input, dummyView, order);
+          Serial<OpType>::getValues(phis, input, dummyView, order);
 
         for (ordinal_type i=0;i<card;++i)
           for (ordinal_type j=0;j<npts;++j)
@@ -309,5 +310,54 @@ namespace Intrepid2 {
                               posDfOrd);
     }
   }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HVOL_TET_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HVOL_TET_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+      
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HVOL_TET_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      auto sizePerPoint = this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HVOL_TET_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_);
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HVOL_TET_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
 } // namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEM.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEM.hpp
index 43a2161c9050..ff20e7426957 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEM.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEM.hpp
@@ -151,25 +151,24 @@ namespace Intrepid2 {
   class Basis_HVOL_TRI_Cn_FEM
     : public Basis<DeviceType,outputValueType,pointValueType> {
   public:
+    using BasisBase = Basis<DeviceType,outputValueType,pointValueType>;
     using HostBasis = Basis_HVOL_TRI_Cn_FEM<typename Kokkos::HostSpace::device_type,outputValueType,pointValueType>;
-      
-    using OrdinalTypeArray1DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray1DHost;
-    using OrdinalTypeArray2DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray2DHost;
-    using OrdinalTypeArray3DHost = typename Basis<DeviceType,outputValueType,pointValueType>::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OrdinalTypeArray1DHost;
+    using typename BasisBase::OrdinalTypeArray2DHost;
+    using typename BasisBase::OrdinalTypeArray3DHost;
+
+    using typename BasisBase::OutputViewType;
+    using typename BasisBase::PointViewType ;
+    using typename BasisBase::ScalarViewType;
 
     /** \brief  Constructor.
      */
     Basis_HVOL_TRI_Cn_FEM(const ordinal_type order,
                            const EPointType   pointType = POINTTYPE_EQUISPACED);
-
-
-    using OutputViewType = typename Basis<DeviceType,outputValueType,pointValueType>::OutputViewType;
-    using PointViewType  = typename Basis<DeviceType,outputValueType,pointValueType>::PointViewType;
-    using ScalarViewType = typename Basis<DeviceType,outputValueType,pointValueType>::ScalarViewType;
-
-    typedef typename Basis<DeviceType,outputValueType,pointValueType>::scalarType  scalarType;
-
-    using Basis<DeviceType,outputValueType,pointValueType>::getValues;
+    
+    using scalarType = typename BasisBase::scalarType;
+    using BasisBase::getValues;
 
     virtual
     void
@@ -191,6 +190,24 @@ namespace Intrepid2 {
                                                 operatorType);
     }
 
+    virtual void 
+    getScratchSpaceSize(      ordinal_type& perTeamSpaceSize,
+                              ordinal_type& perThreadSpaceSize,
+                        const PointViewType inputPoints,
+                        const EOperator operatorType = OPERATOR_VALUE) const override;
+
+    KOKKOS_INLINE_FUNCTION
+    virtual void 
+    getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DeviceType::execution_space>::member_type& team_member,
+      const typename DeviceType::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim = -1,
+      const ordinal_type subcellOrdinal = -1) const override;
+
+
     virtual
     void
     getDofCoords( ScalarViewType dofCoords ) const override {
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEMDef.hpp
index f870940f506b..aa6f54065ff6 100644
--- a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEMDef.hpp
+++ b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HVOL_TRI_Cn_FEMDef.hpp
@@ -22,18 +22,18 @@ namespace Intrepid2 {
 // -------------------------------------------------------------------------------------
 namespace Impl {
 
-template<EOperator opType>
-template<typename OutputViewType,
-typename inputViewType,
-typename workViewType,
-typename vinvViewType>
-KOKKOS_INLINE_FUNCTION
-void
-Basis_HVOL_TRI_Cn_FEM::Serial<opType>::
-getValues(       OutputViewType output,
-    const inputViewType  input,
-    workViewType   work,
-    const vinvViewType   vinv ) {
+    template<EOperator OpType>
+    template<typename OutputViewType,
+             typename InputViewType,
+             typename WorkViewType,
+             typename VinvViewType>
+    KOKKOS_INLINE_FUNCTION
+    void
+    Basis_HVOL_TRI_Cn_FEM::Serial<OpType>::
+    getValues(       OutputViewType output,
+               const InputViewType  input,
+                     WorkViewType   work,
+               const VinvViewType   vinv ) {
 
   constexpr ordinal_type spaceDim = 2;
   const ordinal_type
@@ -49,17 +49,17 @@ getValues(       OutputViewType output,
     }
   }
 
-  typedef typename Kokkos::DynRankView<typename workViewType::value_type, typename workViewType::memory_space> viewType;
-  auto vcprop = Kokkos::common_view_alloc_prop(work);
+  typedef typename Kokkos::DynRankView<typename InputViewType::value_type, typename WorkViewType::memory_space> ViewType;
+  auto vcprop = Kokkos::common_view_alloc_prop(input);
   auto ptr = work.data();
 
-  switch (opType) {
+  switch (OpType) {
   case OPERATOR_VALUE: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
-    workViewType dummyView;
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts);
+      ViewType dummyView;
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j) {
@@ -71,11 +71,11 @@ getValues(       OutputViewType output,
   }
   case OPERATOR_GRAD:
   case OPERATOR_D1: {
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
-    const viewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
+    const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, workView, order);
+    Serial<OpType>::getValues(phis, input, workView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -95,12 +95,12 @@ getValues(       OutputViewType output,
   case OPERATOR_D8:
   case OPERATOR_D9:
   case OPERATOR_D10: {
-    const ordinal_type dkcard = getDkCardinality<opType,spaceDim>(); //(orDn + 1);
-    const viewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
-    workViewType dummyView;
+    const ordinal_type dkcard = getDkCardinality<OpType,spaceDim>(); //(orDn + 1);
+    const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, dkcard);
+    ViewType dummyView;
 
     Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::
-    Serial<opType>::getValues(phis, input, dummyView, order);
+    Serial<OpType>::getValues(phis, input, dummyView, order);
 
     for (ordinal_type i=0;i<card;++i)
       for (ordinal_type j=0;j<npts;++j)
@@ -302,5 +302,54 @@ Basis_HVOL_TRI_Cn_FEM( const ordinal_type order,
         posDfOrd);
   }
 }
+
+  template<typename DT, typename OT, typename PT>
+  void 
+  Basis_HVOL_TRI_Cn_FEM<DT,OT,PT>::getScratchSpaceSize(       
+                                    ordinal_type& perTeamSpaceSize,
+                                    ordinal_type& perThreadSpaceSize,
+                              const PointViewType inputPoints,
+                              const EOperator operatorType) const {
+    perTeamSpaceSize = 0;
+    perThreadSpaceSize = this->vinv_.extent(0)*get_dimension_scalar(inputPoints)*sizeof(typename BasisBase::scalarType);
+  }
+
+  template<typename DT, typename OT, typename PT>
+  KOKKOS_INLINE_FUNCTION
+  void 
+  Basis_HVOL_TRI_Cn_FEM<DT,OT,PT>::getValues(       
+          OutputViewType outputValues,
+      const PointViewType  inputPoints,
+      const EOperator operatorType,
+      const typename Kokkos::TeamPolicy<typename DT::execution_space>::member_type& team_member,
+      const typename DT::execution_space::scratch_memory_space & scratchStorage, 
+      const ordinal_type subcellDim,
+      const ordinal_type subcellOrdinal) const {
+      
+      INTREPID2_TEST_FOR_ABORT( !((subcellDim == -1) && (subcellOrdinal == -1)),
+        ">>> ERROR: (Intrepid2::Basis_HVOL_TRI_Cn_FEM::getValues), The capability of selecting subsets of basis functions has not been implemented yet.");
+
+      const int numPoints = inputPoints.extent(0);
+      using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
+      using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
+      auto sizePerPoint = this->vinv_.extent(0)*get_dimension_scalar(inputPoints);
+      WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
+      using range_type = Kokkos::pair<ordinal_type,ordinal_type>;
+      switch(operatorType) {
+        case OPERATOR_VALUE:
+          Kokkos::parallel_for (Kokkos::TeamThreadRange (team_member, numPoints), [=] (ordinal_type& pt) {
+            auto       output = Kokkos::subview( outputValues, Kokkos::ALL(), range_type  (pt,pt+1), Kokkos::ALL() );
+            const auto input  = Kokkos::subview( inputPoints,                 range_type(pt, pt+1), Kokkos::ALL() );
+            WorkViewType  work(workView.data() + sizePerPoint*team_member.team_rank(), sizePerPoint);
+            Impl::Basis_HVOL_TRI_Cn_FEM::Serial<OPERATOR_VALUE>::getValues( output, input, work, this->vinv_);
+          });
+          break;
+        default: {          
+          INTREPID2_TEST_FOR_ABORT( true,
+            ">>> ERROR (Basis_HVOL_TRI_Cn_FEM): getValues not implemented for this operator");
+          }
+    }
+  }
+
 } // namespace Intrepid2
 #endif
diff --git a/packages/intrepid2/src/Shared/Intrepid2_PolylibDef.hpp b/packages/intrepid2/src/Shared/Intrepid2_PolylibDef.hpp
index 26d71b652d17..b60f5fe09297 100644
--- a/packages/intrepid2/src/Shared/Intrepid2_PolylibDef.hpp
+++ b/packages/intrepid2/src/Shared/Intrepid2_PolylibDef.hpp
@@ -591,57 +591,50 @@ namespace Intrepid2 {
         for (ordinal_type i = 0; i < np; ++i) 
           polyd(i) = 0.5*(alpha + beta + two);
     } else {
-      double a1, a2, a3, a4;
-      const double apb = alpha + beta;
+      INTREPID2_TEST_FOR_ABORT(polyd.data() && !polyd.data() , 
+                               ">>> ERROR (Polylib::Serial::JacobiPolynomial): polyi view needed to compute polyd view.");
+      if(!polyi.data()) return;
 
-      typename polyiViewType::value_type
-        poly[MaxPolylibPoint]={}, polyn1[MaxPolylibPoint]={}, polyn2[MaxPolylibPoint]={};
+      constexpr ordinal_type maxOrder = 2*MaxPolylibPoint-1;
 
-      if (polyi.data()) 
-        for (ordinal_type i=0;i<np;++i)
-          poly[i] = polyi(i);
+      INTREPID2_TEST_FOR_ABORT(maxOrder < n, 
+          ">>> ERROR (Polylib::Serial::JacobiPolynomial): Requested order exceeds maxOrder .");
+      
 
-      for (ordinal_type i = 0; i < np; ++i) {
-        polyn2[i] = one;
-        polyn1[i] = 0.5*(alpha - beta + (alpha + beta + two)*z(i));
-      }
+      double a2[maxOrder]={}, a3[maxOrder]={}, a4[maxOrder]={};
+      double ad1(0.0), ad2(0.0), ad3(0.0);
+      const double apb = alpha + beta;
+      const double amb = alpha - beta;
+
+      typename polyiViewType::value_type polyn1 = polyi(0), polyn2 = polyi(0);
 
       for (auto k = 2; k <= n; ++k) {
-        a1 =  two*k*(k + apb)*(two*k + apb - two);
-        a2 = (two*k + apb - one)*(alpha*alpha - beta*beta);
-        a3 = (two*k + apb - two)*(two*k + apb - one)*(two*k + apb);
-        a4 =  two*(k + alpha - one)*(k + beta - one)*(two*k + apb);
-
-        a2 /= a1;
-        a3 /= a1;
-        a4 /= a1;
-
-        for (ordinal_type i = 0; i < np; ++i) {
-          poly  [i] = (a2 + a3*z(i))*polyn1[i] - a4*polyn2[i];
-          polyn2[i] = polyn1[i];
-          polyn1[i] = poly  [i];
-        }
+        double a1 =  two*k*(k + apb)*(two*k + apb - two);
+        a2[k] = (two*k + apb - one)*(apb*amb)/a1;
+        a3[k] = (two*k + apb - two)*(two*k + apb - one)*(two*k + apb)/a1;
+        a4[k] =  two*(k + alpha - one)*(k + beta - one)*(two*k + apb)/a1;
       }
 
       if (polyd.data()) {
-        a1 = n*(alpha - beta);
-        a2 = n*(two*n + alpha + beta);
-        a3 = two*(n + alpha)*(n + beta);
-        a4 = (two*n + alpha + beta);
-        a1 /= a4;
-        a2 /= a4;
-        a3 /= a4;
-
-        // note polyn2 points to polyn1 at end of poly iterations
-        for (ordinal_type i = 0; i < np; ++i) {
-          polyd(i)  = (a1- a2*z(i))*poly[i] + a3*polyn2[i];
+        double ad4 = (two*n + alpha + beta);
+        ad1 = n*(alpha - beta)/ad4;
+        ad2 = n*(two*n + alpha + beta)/ad4;
+        ad3 = two*(n + alpha)*(n + beta)/ad4;
+      }
+
+      for (ordinal_type i = 0; i < np; ++i) {
+        polyn2 = one;
+        polyn1 = 0.5*(amb + (apb + two)*z(i));
+        for (auto k = 2; k <= n; ++k) {
+          polyi(i) = (a2[k] + a3[k]*z(i))*polyn1 - a4[k]*polyn2;
+          polyn2 = polyn1;
+          polyn1 = polyi(i);
+        }
+        if (polyd.data()) {
+          polyd(i)  = (ad1- ad2*z(i))*polyi(i) + ad3*polyn2;
           polyd(i) /= (one - z(i)*z(i));
         }
       }
-
-      if (polyi.data()) 
-        for (ordinal_type i=0;i<np;++i)
-          polyi(i) = poly[i];
     }
   }
 
diff --git a/packages/intrepid2/src/Shared/Intrepid2_Utils.hpp b/packages/intrepid2/src/Shared/Intrepid2_Utils.hpp
index 45c5f09816d1..107ab788644f 100644
--- a/packages/intrepid2/src/Shared/Intrepid2_Utils.hpp
+++ b/packages/intrepid2/src/Shared/Intrepid2_Utils.hpp
@@ -805,6 +805,32 @@ namespace Intrepid2 {
   {
     return (std::is_standard_layout<typename ViewType::value_type>::value && std::is_trivial<typename ViewType::value_type>::value) ? 0 : get_dimension_scalar(view);
   }
+
+  /// Struct for deleting device instantiation
+  template<typename Device>
+  struct DeviceDeleter {
+    template<typename T>
+    void operator()(T* ptr) {
+      Kokkos::parallel_for(Kokkos::RangePolicy<typename Device::execution_space>(0,1),
+                           KOKKOS_LAMBDA (const int i) { ptr->~T(); });
+      typename Device::execution_space().fence();
+      Kokkos::kokkos_free<typename Device::memory_space>(ptr);
+    }
+  };
+
+  /// Function for creating a vtable on device (requires copy ctor for
+  /// derived object). Allocates device memory and must be called from
+  /// host.
+  template<typename Device,typename Derived>
+  std::unique_ptr<Derived,DeviceDeleter<Device>>
+  copy_virtual_class_to_device(const Derived& host_source)
+  {
+    auto* p = static_cast<Derived*>(Kokkos::kokkos_malloc<typename Device::memory_space>(sizeof(Derived)));
+    Kokkos::parallel_for(Kokkos::RangePolicy<typename Device::execution_space>(0,1),
+                         KOKKOS_LAMBDA (const int i) {new (p) Derived(host_source); });
+    typename Device::execution_space().fence();
+    return std::unique_ptr<Derived,DeviceDeleter<Device>>(p);
+  }
 } // end namespace Intrepid2
 
 #endif
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/CMakeLists.txt
index f080139e2292..f0c9e31cd911 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_HEX_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..cd4d6dabdf43
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_HEX_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_HEX_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..7f4e2807360d
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_I1_FEM/test_02.hpp
@@ -0,0 +1,187 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_HEX_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_HEX_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_HEX_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_HEX_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HCURL_HEX_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and curls on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0)<< ", " << outputCurlsA_Host(ic,i,j,1) << ", " << outputCurlsA_Host(ic,i,j,2) << "]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " << outputCurlsB_Host(i,j,1)<< ", " << outputCurlsB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/CMakeLists.txt
index b682181c9d5b..2e5f6844ed27 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -83,3 +86,75 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_HEX_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..41f15e65574e
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_HEX_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_HEX_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..e9ecf8ca65a4
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/test_02.hpp
@@ -0,0 +1,203 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_HEX_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_HEX_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_HEX_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_HEX_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HCURL_HEX_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };   
+
+              // avoid using a team size larger than needed, to reduce allocated scrach space memory
+              ordinal_type team_size = teamPolicy.team_size_recommended(functor, Kokkos::ParallelForTag());
+              *outStream << "Max Recommended team size: " << team_size << ", Requested team size: " << npts <<std::endl;
+              team_size = std::min(team_size, npts);
+              teamPolicy = Kokkos::TeamPolicy<typename DeviceType::execution_space>(ncells, team_size,vectorSize);
+
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute curls
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+              };     
+              
+              // avoid using a team size larger than needed, to reduce allocated scrach space memory
+              ordinal_type team_size = teamPolicy.team_size_recommended(functor, Kokkos::ParallelForTag());
+              *outStream << "Max Recommended team size: " << team_size << ", Requested team size: " << npts <<std::endl;
+              team_size = std::min(team_size, npts);
+              teamPolicy = Kokkos::TeamPolicy<typename DeviceType::execution_space>(ncells, team_size,vectorSize);         
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+          
+          *outStream << "Order: " << order << ": Comparing values and curls on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare curls
+            const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+            const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", curls A: [" << outputCurlsA_Host(ic,i,j,0)<< ", " << outputCurlsA_Host(ic,i,j,1) << ", " << outputCurlsA_Host(ic,i,j,2) << "]"
+                              << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " << outputCurlsB_Host(i,j,1)<< ", " << outputCurlsB_Host(i,j,2) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/CMakeLists.txt
index 716000daf9b3..89117d0742fb 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,75 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_QUAD_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..01d5359f6b02
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_QUAD_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_QUAD_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..ecef9a26d9d4
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_I1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_QUAD_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_QUAD_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_QUAD_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_QUAD_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HCURL_QUAD_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and curls on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = std::abs(outputCurlsB_Host(i,j) - outputCurlsA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j)<<"]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/CMakeLists.txt
index c831e83ec896..2f44c158238f 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HCURL_QUAD_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_QUAD_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HCURL_QUAD_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..daa3176be226
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_QUAD_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_QUAD_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..2ae8438a11ac
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_QUAD_In_FEM/test_02.hpp
@@ -0,0 +1,189 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_QUAD_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_QUAD_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_QUAD_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_QUAD_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HCURL_QUAD_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute curls
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+          
+          *outStream << "Order: " << order << ": Comparing values and curls on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare curls
+            const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+            const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                  diff = std::abs(outputCurlsB_Host(i,j) - outputCurlsA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", curls A: [" << outputCurlsA_Host(ic,i,j)<<"]"
+                              << ", curls B: [" << outputCurlsB_Host(i,j) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/CMakeLists.txt
index 46e84774b70d..234b8e2d6fed 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_TET_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..8fadc4a2c865
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_TET_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_TET_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..9c112664aff0
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_I1_FEM/test_02.hpp
@@ -0,0 +1,187 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_TET_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_TET_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_TET_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_TET_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HCURL_TET_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and curls on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << ", " << outputValuesA_Host(ic,i,j,2) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << ", " << outputValuesB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0)<< ", " << outputCurlsA_Host(ic,i,j,1) << ", " << outputCurlsA_Host(ic,i,j,2) << "]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " << outputCurlsB_Host(i,j,1)<< ", " << outputCurlsB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/CMakeLists.txt
index 46e4453c0d57..c40f3503ccf9 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HCURL_TET_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_TET_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HCURL_TET_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..278f9326b54c
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_TET_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_TET_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..c06ec5688c39
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/test_02.hpp
@@ -0,0 +1,204 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_TET_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_TET_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_TET_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_TET_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 7;
+
+      try { 
+        for (int order=1;order <= maxOrder;++order) {
+          using BasisType = Basis_HCURL_TET_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };     
+              
+              // avoid using a team size larger than needed, to reduce allocated scrach space memory
+              ordinal_type team_size = teamPolicy.team_size_recommended(functor, Kokkos::ParallelForTag());
+              *outStream << "Max Recommended team size: " << team_size << ", Requested team size: " << npts <<std::endl;
+              team_size = std::min(team_size, npts);
+              teamPolicy = Kokkos::TeamPolicy<typename DeviceType::execution_space>(ncells, team_size,vectorSize);
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute curls
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+              };      
+
+              // avoid using a team size larger than needed, to reduce allocated scrach space memory
+              ordinal_type team_size = teamPolicy.team_size_recommended(functor, Kokkos::ParallelForTag());
+              *outStream << "Max Recommended team size: " << team_size << ", Requested team size: " << npts <<std::endl;
+              team_size = std::min(team_size, npts);
+              teamPolicy = Kokkos::TeamPolicy<typename DeviceType::execution_space>(ncells, team_size,vectorSize);        
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+          
+          *outStream << "Order: " << order << ": Comparing values and curls on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();  
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare curls
+            const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+            const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>(); 
+            for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", curls A: [" << outputCurlsA_Host(ic,i,j,0)<< ", " << outputCurlsA_Host(ic,i,j,1) << ", " << outputCurlsA_Host(ic,i,j,2) << "]"
+                              << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " << outputCurlsB_Host(i,j,1)<< ", " << outputCurlsB_Host(i,j,2) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/CMakeLists.txt
index 2bf7bfdee691..b87adda0a338 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,75 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_TRI_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..2fe8396db2d1
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_TRI_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_TRI_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..3b255303dd1f
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_I1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_TRI_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego 
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_TRI_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_TRI_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_TRI_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HCURL_TRI_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and curls on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = std::abs(outputCurlsB_Host(i,j) - outputCurlsA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j)<<"]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/CMakeLists.txt
index 49f5b786efc9..ec30d2154004 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HCURL_TRI_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_TRI_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HCURL_TRI_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..509fff60809f
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_TRI_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_TRI_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..aa19ce2114e0
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TRI_In_FEM/test_02.hpp
@@ -0,0 +1,189 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_TRI_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_TRI_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_TRI_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_TRI_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HCURL_TRI_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute curls
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+          
+          *outStream << "Order: " << order << ": Comparing values and curls on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare curls
+            const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+            const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                  diff = std::abs(outputCurlsB_Host(i,j) - outputCurlsA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", curls A: [" << outputCurlsA_Host(ic,i,j)<<"]"
+                              << ", curls B: [" << outputCurlsB_Host(i,j) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/CMakeLists.txt
index ba0496748a48..cb1ebf7b3de1 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HCURL_WEDGE_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..1b963155651c
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HCURL_WEDGE_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HCURL_WEDGE_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..de75f4cf2d72
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_WEDGE_I1_FEM/test_02.hpp
@@ -0,0 +1,188 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HCURL_WEDGE_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HCURL_WEDGE_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HCURL_WEDGE_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HCURL_WEDGE_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HCURL_WEDGE_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and curls for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and curls for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and curls on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0)<< ", " << outputCurlsA_Host(ic,i,j,1) << ", " << outputCurlsA_Host(ic,i,j,2) << "]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " << outputCurlsB_Host(i,j,1)<< ", " << outputCurlsB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/CMakeLists.txt
index fd4d688a591d..3fb8fc747f9c 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/CMakeLists.txt
@@ -1,13 +1,18 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
-
+  
 IF (HAVE_INTREPID2_SACADO) 
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DOUBLE")
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_HEX_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..ab24cfec247d
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_HEX_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_HEX_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..fb05ad186945
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_I1_FEM/test_02.hpp
@@ -0,0 +1,186 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_HEX_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_HEX_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_HEX_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_HEX_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HDIV_HEX_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute divergences
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+        
+        *outStream << "Comparing values and divergences on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << ", " << outputValuesA_Host(ic,i,j,2) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << ", " << outputValuesB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare divergences
+          const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+          const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                            << ", divergence B: " << outputDivergencesB_Host(i,j)
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/CMakeLists.txt
index a0e677500751..7a81181c8403 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HDIV_HEX_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_HEX_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HDIV_HEX_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..71b715c78833
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_HEX_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_HEX_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..61c3d844f5dd
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_HEX_In_FEM/test_02.hpp
@@ -0,0 +1,190 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_HEX_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_HEX_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_HEX_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_HEX_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HDIV_HEX_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute divergences
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+          
+          *outStream << "Order: " << order << ": Comparing values and divergences on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << ", " << outputValuesA_Host(ic,i,j,2) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << ", " << outputValuesB_Host(i,j,2) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare divergences
+            const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+            const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                  diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                              << ", divergence B: " << outputDivergencesB_Host(i,j)
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/CMakeLists.txt
index 5900fa72e32a..b21760f88ec4 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/CMakeLists.txt
@@ -1,13 +1,18 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
-
+  
 IF (HAVE_INTREPID2_SACADO) 
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DOUBLE")
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_QUAD_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..5e1eb4c2dc79
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_QUAD_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_QUAD_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..a811df7230c7
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_I1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_QUAD_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_QUAD_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_QUAD_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_QUAD_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HDIV_QUAD_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute divergences
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+        
+        *outStream << "Comparing values and divergences on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare divergences
+          const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+          const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                            << ", divergence B: " << outputDivergencesB_Host(i,j)
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/CMakeLists.txt
index 59d15e7c716c..cdc2989d6036 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HDIV_QUAD_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_QUAD_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HDIV_QUAD_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..328d40fda920
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_QUAD_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_QUAD_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..529007a97787
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_QUAD_In_FEM/test_02.hpp
@@ -0,0 +1,190 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_QUAD_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_QUAD_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_QUAD_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_QUAD_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HDIV_QUAD_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+
+
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute divergences
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+          
+          *outStream << "Order: " << order << ": Comparing values and divergences on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare divergences
+            const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+            const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                  diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                              << ", divergence B: " << outputDivergencesB_Host(i,j)
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/CMakeLists.txt
index ea49b4cde715..6b8c89a459f9 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/CMakeLists.txt
@@ -1,13 +1,18 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
-
+  
 IF (HAVE_INTREPID2_SACADO) 
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DOUBLE")
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_TET_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..72f12aed7a9b
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_TET_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_TET_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..2da333826892
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_I1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_TET_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_TET_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_TET_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_TET_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HDIV_TET_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute divergences
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+        
+        *outStream << "Comparing values and divergences on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << ", " << outputValuesA_Host(ic,i,j,2) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << ", " << outputValuesB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare divergences
+          const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+          const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                            << ", divergence B: " << outputDivergencesB_Host(i,j)
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/CMakeLists.txt
index 264ce7c056ff..f4a2093e0e4f 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HDIV_TET_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_TET_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HDIV_TET_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..c08e06044acf
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_TET_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_TET_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..6186dd1fb4bc
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TET_In_FEM/test_02.hpp
@@ -0,0 +1,190 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_TET_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_TET_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_TET_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_TET_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 7;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HDIV_TET_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute divergences
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+          
+          *outStream << "Order: " << order << ": Comparing values and divergences on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << ", " << outputValuesA_Host(ic,i,j,2) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << ", " << outputValuesB_Host(i,j,2) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare divergences
+            const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+            const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+            
+            OutValueType diff = 0;
+            //Note, a PR release-debug serial build shows substantially higher errors (possibly due to operation rearrangements). Need to investigate more.
+            auto tol = 1e6*epsilon<double>();
+            for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                  diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                              << ", divergence B: " << outputDivergencesB_Host(i,j)
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/CMakeLists.txt
index 4f47ee20c141..581f594311e0 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/CMakeLists.txt
@@ -1,13 +1,18 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
-
+  
 IF (HAVE_INTREPID2_SACADO) 
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DOUBLE")
 #  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,76 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_TRI_I1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..99b3fb273163
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_TRI_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_TRI_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..15b4152a781a
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_I1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_TRI_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_TRI_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_TRI_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_TRI_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HDIV_TRI_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute divergences
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+        
+        *outStream << "Comparing values and divergences on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare divergences
+          const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+          const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                            << ", divergence B: " << outputDivergencesB_Host(i,j)
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/CMakeLists.txt
index 4f11a0b1e70c..f06b5f1bb859 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HDIV_TRI_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_TRI_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HDIV_TRI_In_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..060c322dc641
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_TRI_In_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_TRI_In_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/test_02.hpp
new file mode 100644
index 000000000000..1d3c940090d7
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_TRI_In_FEM/test_02.hpp
@@ -0,0 +1,189 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_TRI_In_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_TRI_In_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_TRI_In_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_TRI_In_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HDIV_TRI_In_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute divergences
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+          
+          *outStream << "Order: " << order << ": Comparing values and divergences on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << "]"
+                              << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare divergences
+            const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+            const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                  diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                              << ", divergence B: " << outputDivergencesB_Host(i,j)
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/CMakeLists.txt
index 67fcf71311f7..d801c634869e 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,3 +71,75 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HDIV_WEDGE_In_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..16906a746c00
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HDIV_WEDGE_I1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HDIV_WEDGE_I1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/test_02.hpp
new file mode 100644
index 000000000000..c991769a4852
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HDIV_WEDGE_I1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HDIV_WEDGE_I1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HDIV_WEDGE_I1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HDIV_WEDGE_I1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HDIV_WEDGE_I1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HDIV_WEDGE_I1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputDivergencesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and divergences for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute divergences
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto divergencesACell = Kokkos::subview(outputDivergencesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(divergencesACell, inputPoints, OPERATOR_DIV, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_DIV);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and divergences for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputDivergencesB, inputPoints, OPERATOR_DIV);
+        
+        *outStream << "Comparing values and divergences on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputValuesB_Host(i,j,d) - outputValuesA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: [" << outputValuesA_Host(ic,i,j,0) << ", " << outputValuesA_Host(ic,i,j,1) << ", " << outputValuesA_Host(ic,i,j,2) << "]"
+                            << ", val B: [" << outputValuesB_Host(i,j,0) << ", " << outputValuesB_Host(i,j,1) << ", " << outputValuesB_Host(i,j,2) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare divergences
+          const auto outputDivergencesA_Host = Kokkos::create_mirror_view(outputDivergencesA); Kokkos::deep_copy(outputDivergencesA_Host, outputDivergencesA);
+          const auto outputDivergencesB_Host = Kokkos::create_mirror_view(outputDivergencesB); Kokkos::deep_copy(outputDivergencesB_Host, outputDivergencesB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputDivergencesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputDivergencesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputDivergencesA_Host.extent(2);++j) {
+                diff = std::abs(outputDivergencesB_Host(i,j) - outputDivergencesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", divergence A: " << outputDivergencesA_Host(ic,i,j)
+                            << ", divergence B: " << outputDivergencesB_Host(i,j)
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/CMakeLists.txt
index 88da0999c2ab..4982bd4f8dff 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/CMakeLists.txt
@@ -3,6 +3,7 @@ TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -67,6 +68,7 @@ SET(Intrepid2_TEST_ETI_FILE "test_01")
 
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
+
 IF(Kokkos_ENABLE_CUDA)
   LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
   LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
@@ -106,3 +108,75 @@ IF (${ETI_DEVICE_COUNT} GREATER_EQUAL 0)
     ENDFOREACH()
   ENDFOREACH()
 ENDIF()
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_03")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_HEX_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_HEX_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/eti/test_03_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/eti/test_03_ETI.in
new file mode 100644
index 000000000000..a88bf31183c7
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/eti/test_03_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_03.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_HEX_C1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_03.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_HEX_C1_FEM_Test03<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/test_03.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/test_03.hpp
new file mode 100644
index 000000000000..9d326a80cf33
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C1_FEM/test_03.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_03.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_HEX_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_HEX_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_HEX_C1_FEM_Test03(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_HEX_C1_FEM, Test 3", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_HEX_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/CMakeLists.txt
index a9bfc7f38abb..29e3244386c9 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/CMakeLists.txt
@@ -1,8 +1,16 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "")
+LIST(APPEND Intrepid2_TEST_ETI_FILE 
+  "test_01"
+  "test_01_Serendipity")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,12 +25,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "")
-LIST(APPEND Intrepid2_TEST_ETI_FILE 
-  "test_01"
-  "test_01_Serendipity")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_HEX_C2_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_HEX_C2_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..f10b05aa223e
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_HEX_C2_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_HEX_C2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/test_02.hpp
new file mode 100644
index 000000000000..a29875462280
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_C2_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_HEX_C2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_HEX_C2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_HEX_C2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_HEX_C2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_HEX_C2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/CMakeLists.txt
index 793d773f707f..ba86fece89df 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/CMakeLists.txt
@@ -7,6 +7,7 @@ SET(Intrepid2_TEST_ETI_FILE "test_01")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -94,11 +95,18 @@ SET(Intrepid2_TEST_ETI_FILE "test_02")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
 
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_01_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_01_ETI.in
index 487708632660..25426631d6de 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_01_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_01_ETI.in
@@ -28,9 +28,8 @@ constexpr int num_deriv = 10;
 #define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 23)
-constexpr int num_deriv = 9;
-constexpr int max_deriv = 10;
-#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, max_deriv+1)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 20)
 constexpr int num_deriv = 2;
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_02_ETI.in
index 6a200d58b21d..d314677fd1db 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_02_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/eti/test_02_ETI.in
@@ -7,22 +7,45 @@
 // *****************************************************************************
 // @HEADER
 
-/** \file test_01.cpp
-    \brief  Unit test of serial interface Intrepid2::Basis_HGRAD_QUAD_Cn_FEM.
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_HEX_Cn_FEM team-level getValues.
     \author Kyungjoo Kim
 */
 
 #include "Kokkos_Core.hpp"
 
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
 #include "test_02.hpp"
 
 int main(int argc, char *argv[]) {
 
+  const bool verbose = (argc-1) > 0;
   Kokkos::initialize();
-  {
-    const bool verbose = (argc-1) > 0;
-    Intrepid2::Test::HGRAD_HEX_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
-  }
+
+  Intrepid2::Test::HGRAD_HEX_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
   Kokkos::finalize();
   return 0;
 }
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/test_02.hpp
index b98955113bde..e392f1540447 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/test_02.hpp
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_HEX_Cn_FEM/test_02.hpp
@@ -7,9 +7,9 @@
 // *****************************************************************************
 // @HEADER
 
-/** \file test_01.hpp
+/** \file test_02.hpp
     \brief  Unit tests for the Intrepid2::HGRAD_HEX_Cn_FEM class.
-    \author Created by P. Bochev, D. Ridzal, K. Peterson, Kyungjoo Kim
+    \author Created by Kyungjoo Kim, Mauro Perego
  */
 
 
@@ -23,100 +23,151 @@
 #include "Intrepid2_Utils.hpp"
 
 #include "Intrepid2_HGRAD_HEX_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
 
 namespace Intrepid2 {
 
   namespace Test {
 
-    // This code provides an example to use serial interface of high order elements
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
     template<typename OutValueType, typename PointValueType, typename DeviceType>
     int HGRAD_HEX_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_HEX_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
       using DeviceSpaceType = typename DeviceType::execution_space;
       Kokkos::print_configuration(std::cout, false);
 
       int errorFlag = 0;
-
+      constexpr int maxOrder = 9;
       try { 
-        // for higher orders in certain environments, this test can take a while to run in ctest.  So we reduce the number of points as we go beyond 10th order. Also, @mperego is replacing this test, so for now we just restrict to the 10 orders we supported until recently.
-        for (int order=1;order<10;++order) {
-          Basis_HGRAD_HEX_Cn_FEM<DeviceType,OutValueType,PointValueType> basis(order);
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HGRAD_HEX_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
           
-          // problem setup 
-          //   let's say we want to evaluate 1000 points in parallel. output values are stored in outputValuesA and B.
-          //   A is compuated via serial interface and B is computed with top-level interface.
-          const int npts = 1000, ndim = 3;
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesA("outputValuesA", basis.getCardinality(), npts);
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesB("outputValuesB", basis.getCardinality(), npts);
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
           
-          Kokkos::View<PointValueType**,DeviceType> inputPointsViewToUseRandom("inputPoints", npts, ndim);
-          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (inputPointsViewToUseRandom.data(),  npts, ndim);
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
           
-          // random values between (-1,1) x (-1,1)
+          // random values between (0,1)
           Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
           Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
           
-          // compute setup
-          //   we need vinv and workspace
-          const auto vinv = basis.getVandermondeInverse();
-          
-          // worksize 
-          //   workspace per thread is required for serial interface. 
-          //   parallel_for with range policy would be good to use stack workspace 
-          //   as team policy only can create shared memory 
-          //   this part would be tricky as the max size should be determined at compile time
-          //   let's think about this and find out the best practice. for now I use the following.
-          constexpr int worksize = (Parameters::MaxOrder+1)*4;
-          
-          // if you use team policy, worksize can be gathered from the basis object and use 
-          // kokkos shmem_size APIs to create workspace per team or per thread.
-          //const auto worksize_for_teampolicy = basis.getWorksizePerPoint(OPERATOR_VALUE);
-          
-          // extract point range to be evaluated in each thread
-          typedef Kokkos::pair<int,int> range_type;
-          
-          // parallel execution with serial interface
-          Kokkos::RangePolicy<DeviceSpaceType> policy(0, npts);
-          Kokkos::parallel_for(policy, KOKKOS_LAMBDA(int i) {
-              // we evaluate a single point 
-              const range_type pointRange = range_type(i,i+1);
-              
-              // out (# dofs, # pts), input (# pts, # dims)
-              auto output = Kokkos::subview(outputValuesA, Kokkos::ALL(), pointRange);
-              auto input  = Kokkos::subview(inputPoints,   pointRange, Kokkos::ALL());
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // wrap static workspace with a view; serial interface has a template view interface.
-              // either view or dynrankview with a right size is okay.
-              OutValueType workbuf[worksize];
-              Kokkos::View<OutValueType*,Kokkos::AnonymousSpace> work(&workbuf[0], worksize);
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute gradients
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // evaluate basis using serial interface
-              Impl::Basis_HGRAD_HEX_Cn_FEM
-                ::Serial<OPERATOR_VALUE>::getValues(output, input, work, vinv);
-            });
-          
-          // evaluation using high level interface
-          basis.getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
-          
-          // compare 
-          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
-          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
-          
-          double sum = 0, diff = 0;
-          for (size_t i=0;i<outputValuesA_Host.extent(0);++i)
-            for (size_t j=0;j<outputValuesA_Host.extent(1);++j) {
-              sum += std::abs(outputValuesB_Host(i,j));
-              diff += std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(i,j));
-              if (verbose && (diff/sum > 1.0e-9)) {
-                std::cout << " order = " << order
-                          << " i = " << i << " j = " << j 
-                          << " val A = " << outputValuesA_Host(i,j) 
-                          << " val B = " << outputValuesB_Host(i,j) 
-                          << " diff  = " << (outputValuesA_Host(i,j) - outputValuesB_Host(i,j)) 
-                          << std::endl;
-              }
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
             }
-          if (diff/sum > 1.0e-9) {
-            errorFlag = -1;
+          }
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+          
+          *outStream << "Order: " << order << ": Comparing values and gradients on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare grads
+            const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+            const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                              << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
           }
         }
       } catch (std::exception &err) {
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/CMakeLists.txt
index 940d4ad3ebb4..186c0369d09e 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  # MESSAGE(STATUS "Generating TEST HGRAD_LINE_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_LINE_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..69bb74d6a746
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_LINE_C1_FEM team-level get Values.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_LINE_C1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/test_02.hpp
new file mode 100644
index 000000000000..c0d1db740ce9
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C1_FEM/test_02.hpp
@@ -0,0 +1,185 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_LINE_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_LINE_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_LINE_C1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_LINE_C1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_LINE_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        // problem setup 
+        //   let's say we want to evaluate 1000 points in parallel. output values are stored in outputValuesA and B.
+        //   A is compuated via serial interface and B is computed with top-level interface.
+        const int ncells = 5, npts = 10, ndim = 1;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << "]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0)  <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/CMakeLists.txt
index f26e93eb35f6..47fc663ffd89 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_LINE_C2_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_LINE_C2_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..76e7d225aa79
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_LINE_C2_FEM team-level get Values.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_LINE_C2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/test_02.hpp
new file mode 100644
index 000000000000..7c40e6e00dd1
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_C2_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_LINE_C2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_LINE_C2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_LINE_C2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_LINE_C2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_LINE_C2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 1;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << "]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/CMakeLists.txt
index 40be3eb7ddf0..088e2285ac8f 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_LINE_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_LINE_Cn_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_01_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_01_ETI.in
index 070fba1f3916..b662965e7493 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_01_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_01_ETI.in
@@ -28,10 +28,8 @@ constexpr int num_deriv = 10;
 #define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 23)
-///  Mauro, the master branch uses this derivative dimension which sounds dummy
-constexpr int num_deriv = 0; //9;
-constexpr int max_deriv = 1; //10;  
-#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, max_deriv+1)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 20)
 constexpr int num_deriv = 2;
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..2ff629694b47
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_LINE_Cn_FEM team-level get Values.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_LINE_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/test_02.hpp
new file mode 100644
index 000000000000..23dafa935f7e
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_LINE_Cn_FEM/test_02.hpp
@@ -0,0 +1,188 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_LINE_Cn_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_LINE_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.s
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_LINE_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_LINE_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        for (int order=1;order<=Parameters::MaxOrder;++order) {
+          using BasisType = Basis_HGRAD_LINE_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 1;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute gradients
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+          
+          *outStream << "Order: " << order << ": Comparing values and gradients on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare grads
+            const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+            const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << "]"
+                              << ", grads B: [" << outputGradsB_Host(i,j,0) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/CMakeLists.txt
index aac1913c1e91..60b3eaa1ed20 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_PYR_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_PYR_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..ac8e1393df20
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_PYR_C1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_PYR_C1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/test_02.hpp
new file mode 100644
index 000000000000..ae1ba8b9b47d
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_C1_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_PYR_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_PYR_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_PYR_C1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_PYR_C1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_PYR_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/CMakeLists.txt
index fdbf58124c2e..813910ef9f3d 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,7 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
-
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -68,6 +71,79 @@ FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   ENDFOREACH()
 ENDFOREACH()
 
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_TET_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_TET_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
 TRIBITS_COPY_FILES_TO_BINARY_DIR(HGRAD_PYR_I2TestDataCopy
   SOURCE_FILES
     PYR_I2_D2Vals.dat
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..b1bc686c303d
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_PYR_I2_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_PYR_I2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/test_02.hpp
new file mode 100644
index 000000000000..39b7903b384a
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_PYR_I2_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_PYR_I2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_PYR_I2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_PYR_I2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_PYR_I2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_PYR_I2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/CMakeLists.txt
index ef324d6c681f..593042946b82 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_QUAD_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_QUAD_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..7650cb60968c
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of serial interface Intrepid2::Basis_HGRAD_QUAD_C1_FEM.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_QUAD_C1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/test_02.hpp
new file mode 100644
index 000000000000..2dba47d6f022
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C1_FEM/test_02.hpp
@@ -0,0 +1,228 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_QUAD_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_QUAD_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_QUAD_C1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_QUAD_C1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_QUAD_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << "]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        {
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0) << ", " << outputCurlsA_Host(ic,i,j,1) <<"]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " <<  outputCurlsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/CMakeLists.txt
index 6d92bb337ac4..14d863a19fc7 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/CMakeLists.txt
@@ -1,8 +1,16 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "")
+LIST(APPEND Intrepid2_TEST_ETI_FILE 
+  "test_01"
+  "test_01_Serendipity")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,12 +25,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "")
-LIST(APPEND Intrepid2_TEST_ETI_FILE 
-  "test_01"
-  "test_01_Serendipity")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_QUAD_C2_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_QUAD_C2_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..3dac2095b0b6
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of serial interface Intrepid2::Basis_HGRAD_QUAD_C2_FEM.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_QUAD_C2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/test_02.hpp
new file mode 100644
index 000000000000..cf2ba0043d7b
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_C2_FEM/test_02.hpp
@@ -0,0 +1,228 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_QUAD_C2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_QUAD_C2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+  // and compares the results with those obtained using the classic getValues function.
+  namespace Test {
+
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_QUAD_C2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_QUAD_C2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_QUAD_C2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << "]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        {
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0) << ", " << outputCurlsA_Host(ic,i,j,1) <<"]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " <<  outputCurlsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/CMakeLists.txt
index 9a6190ea4405..50f38bf22177 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/CMakeLists.txt
@@ -7,6 +7,7 @@ SET(Intrepid2_TEST_ETI_FILE "test_01")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -94,11 +95,18 @@ SET(Intrepid2_TEST_ETI_FILE "test_02")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
 
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_01_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_01_ETI.in
index 46bd4b13ccfb..0cae06ee9e31 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_01_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_01_ETI.in
@@ -28,9 +28,8 @@ constexpr int num_deriv = 10;
 #define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 23)
-constexpr int num_deriv = 9;
-constexpr int max_deriv = 10;
-#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, max_deriv+1)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 20)
 constexpr int num_deriv = 2;
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_02_ETI.in
index ccb60ba60798..07adf2c5e888 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_02_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/eti/test_02_ETI.in
@@ -8,21 +8,44 @@
 // @HEADER
 
 /** \file test_01.cpp
-    \brief  Unit test of serial interface Intrepid2::Basis_HGRAD_QUAD_Cn_FEM.
+    \brief  Unit test of Intrepid2::Basis_HGRAD_QUAD_Cn_FEM team-level getValues.
     \author Kyungjoo Kim
 */
 
 #include "Kokkos_Core.hpp"
 
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
 #include "test_02.hpp"
 
 int main(int argc, char *argv[]) {
 
+  const bool verbose = (argc-1) > 0;
   Kokkos::initialize();
-  {
-    const bool verbose = (argc-1) > 0;
-    Intrepid2::Test::HGRAD_QUAD_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
-  }
+
+  Intrepid2::Test::HGRAD_QUAD_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
   Kokkos::finalize();
   return 0;
 }
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/test_02.hpp
index 36a858dec901..2e2bdf715470 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/test_02.hpp
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_QUAD_Cn_FEM/test_02.hpp
@@ -7,9 +7,9 @@
 // *****************************************************************************
 // @HEADER
 
-/** \file test_01.hpp
+/** \file test_02.hpp
     \brief  Unit tests for the Intrepid2::HGRAD_QUAD_Cn_FEM class.
-    \author Created by P. Bochev, D. Ridzal, K. Peterson, Kyungjoo Kim
+    \author Created by Kyungjoo Kim, Mauro Perego
  */
 
 
@@ -23,98 +23,196 @@
 #include "Intrepid2_Utils.hpp"
 
 #include "Intrepid2_HGRAD_QUAD_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
 
 namespace Intrepid2 {
 
   namespace Test {
 
-    // This code provides an example to use serial interface of high order elements
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
     template<typename OutValueType, typename PointValueType, typename DeviceType>
     int HGRAD_QUAD_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_QUAD_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
       Kokkos::print_configuration(std::cout, false);
 
       int errorFlag = 0;
-
+      constexpr int maxOrder = 9;
       try { 
-        for (int order=1;order<Parameters::MaxOrder;++order) {
-          Basis_HGRAD_QUAD_Cn_FEM<DeviceType,OutValueType,PointValueType> basis(order);
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HGRAD_QUAD_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
           
-          // problem setup 
-          //   let's say we want to evaluate 1000 points in parallel. output values are stored in outputValuesA and B.
-          //   A is compuated via serial interface and B is computed with top-level interface.
-          const int npts = 1000, ndim = 2;
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesA("outputValuesA", basis.getCardinality(), npts);
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesB("outputValuesB", basis.getCardinality(), npts);
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
           
-          Kokkos::View<PointValueType**,DeviceType> inputPointsViewToUseRandom("inputPoints", npts, ndim);
-          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (inputPointsViewToUseRandom.data(),  npts, ndim);
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
           
-          // random values between (-1,1) x (-1,1)
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
           Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
           Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
           
-          // compute setup
-          //   we need vinv and workspace
-          const auto vinv = basis.getVandermondeInverse();
-          
-          // worksize 
-          //   workspace per thread is required for serial interface. 
-          //   parallel_for with range policy would be good to use stack workspace 
-          //   as team policy only can create shared memory 
-          //   this part would be tricky as the max size should be determined at compile time
-          //   let's think about this and find out the best practice. for now I use the following.
-          constexpr int worksize = (Parameters::MaxOrder+1)*3;
-          
-          // if you use team policy, worksize can be gathered from the basis object and use 
-          // kokkos shmem_size APIs to create workspace per team or per thread.
-          //const auto worksize_for_teampolicy = basis.getWorksizePerPoint(OPERATOR_VALUE);
-          
-          // extract point range to be evaluated in each thread
-          typedef Kokkos::pair<int,int> range_type;
-          
-          // parallel execution with serial interface
-          Kokkos::RangePolicy<typename DeviceType::execution_space> policy(0, npts);
-          Kokkos::parallel_for(policy, KOKKOS_LAMBDA(int i) {
-              // we evaluate a single point 
-              const range_type pointRange = range_type(i,i+1);
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // out (# dofs, # pts), input (# pts, # dims)
-              auto output = Kokkos::subview(outputValuesA, Kokkos::ALL(), pointRange);
-              auto input  = Kokkos::subview(inputPoints,   pointRange, Kokkos::ALL());
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute gradients
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // wrap static workspace with a view; serial interface has a template view interface.
-              // either view or dynrankview with a right size is okay.
-              OutValueType workbuf[worksize];
-              Kokkos::View<OutValueType*,Kokkos::AnonymousSpace> work(&workbuf[0], worksize);
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute curls
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // evaluate basis using serial interface
-              Impl::Basis_HGRAD_QUAD_Cn_FEM
-                ::Serial<OPERATOR_VALUE>::getValues(output, input, work, vinv);
-            });
-          
-          // evaluation using high level interface
-          basis.getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
-          
-          // compare 
-          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
-          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
-          
-          double sum = 0, diff = 0;
-          for (size_t i=0;i<outputValuesA_Host.extent(0);++i)
-            for (size_t j=0;j<outputValuesA_Host.extent(1);++j) {
-              sum += std::abs(outputValuesB_Host(i,j));
-              diff += std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(i,j));
-              if (verbose) {
-                std::cout << " order = " << order
-                          << " i = " << i << " j = " << j 
-                          << " val A = " << outputValuesA_Host(i,j) 
-                          << " val B = " << outputValuesB_Host(i,j) 
-                          << " diff  = " << (outputValuesA_Host(i,j) - outputValuesB_Host(i,j)) 
-                          << std::endl;
-              }
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
             }
-          if (diff/sum > 1.0e-9) {
-            errorFlag = -1;
+          }
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+          basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+          
+          *outStream << "Order: " << order << ": Comparing values and gradients on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare grads
+            const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+            const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << "]"
+                              << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare curls
+            const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+            const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", curls A: [" << outputCurlsA_Host(ic,i,j,0) << ", " << outputCurlsA_Host(ic,i,j,1) <<"]"
+                              << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " <<  outputCurlsB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
           }
         }
       } catch (std::exception &err) {
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/CMakeLists.txt
index 37135caa841f..325ab37afd81 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_TET_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_TET_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..7fab72655932
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_TET_C1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_TET_C1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/test_02.hpp
new file mode 100644
index 000000000000..48d5b3eb9e2f
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C1_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_TET_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_TET_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_TET_C1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TET_C1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_TET_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/CMakeLists.txt
index 2d0041218982..cac75a8505b4 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_TET_C2_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_TET_C2_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..b01c59418753
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_TET_C2_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_TET_C2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/test_02.hpp
new file mode 100644
index 000000000000..5b788cc85328
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_C2_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_TET_C2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_TET_C2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+  // and compares the results with those obtained using the classic getValues function.
+  namespace Test {
+
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_TET_C2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TET_C2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_TET_C2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/CMakeLists.txt
index 09fddb77ac02..3bc181264b2b 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_TET_COMP12_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_TET_COMP12_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..c7a155d2f8a9
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_TET_COMP12_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_TET_COMP12_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/test_02.hpp
new file mode 100644
index 000000000000..9120a9bf8b53
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_COMP12_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_TET_COMP12_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_TET_COMP12_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_TET_COMP12_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TET_COMP12_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_TET_COMP12_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/CMakeLists.txt
index cb2c34d9f2e6..b669b1b2ba1f 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/CMakeLists.txt
@@ -7,6 +7,7 @@ SET(Intrepid2_TEST_ETI_FILE "test_01")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -94,11 +95,18 @@ SET(Intrepid2_TEST_ETI_FILE "test_02")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
 
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_01_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_01_ETI.in
index 74f1bccc00db..c997523b3120 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_01_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_01_ETI.in
@@ -28,10 +28,8 @@ constexpr int num_deriv = 10;
 #define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 23)
-///  Mauro, the master branch uses this derivative dimension which sounds dummy
-constexpr int num_deriv = 0; //9;
-constexpr int max_deriv = 1; //10;
-#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, max_deriv+1)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 20)
 constexpr int num_deriv = 2;
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_02_ETI.in
index c26586d323cf..a6b3263eb7c0 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_02_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/eti/test_02_ETI.in
@@ -8,21 +8,44 @@
 // @HEADER
 
 /** \file test_01.cpp
-    \brief  Unit test of serial interface Intrepid2::Basis_HGRAD_TET_Cn_FEM.
+    \brief  Unit test of Intrepid2::Basis_HGRAD_TET_Cn_FEM team-level getValues.
     \author Kyungjoo Kim
 */
 
 #include "Kokkos_Core.hpp"
 
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
 #include "test_02.hpp"
 
 int main(int argc, char *argv[]) {
 
+  const bool verbose = (argc-1) > 0;
   Kokkos::initialize();
-  {
-    const bool verbose = (argc-1) > 0;
-    Intrepid2::Test::HGRAD_TET_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
-  }
+
+  Intrepid2::Test::HGRAD_TET_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
   Kokkos::finalize();
   return 0;
 }
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/test_02.hpp
index 4f6c6c3a33e0..a2cd75fc812e 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/test_02.hpp
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TET_Cn_FEM/test_02.hpp
@@ -7,9 +7,9 @@
 // *****************************************************************************
 // @HEADER
 
-/** \file test_01.hpp
+/** \file test_02.hpp
     \brief  Unit tests for the Intrepid2::HGRAD_TET_Cn_FEM class.
-    \author Created by P. Bochev, D. Ridzal, K. Peterson, Kyungjoo Kim
+    \author Created by Kyungjoo Kim, Mauro Perego
  */
 
 
@@ -23,99 +23,152 @@
 #include "Intrepid2_Utils.hpp"
 
 #include "Intrepid2_HGRAD_TET_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
 
 namespace Intrepid2 {
 
   namespace Test {
 
-    // This code provides an example to use serial interface of high order elements
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
     template<typename OutValueType, typename PointValueType, typename DeviceType>
     int HGRAD_TET_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TET_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
       using DeviceSpaceType = typename DeviceType::execution_space;
       Kokkos::print_configuration(std::cout, false);
 
       int errorFlag = 0;
-
+      constexpr int maxOrder = 7;
       try { 
-        for (int order=1;order<10;++order) {
-          Basis_HGRAD_TET_Cn_FEM<DeviceType,OutValueType,PointValueType> basis(order);
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HGRAD_TET_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
           
-          // problem setup 
-          //   let's say we want to evaluate 1000 points in parallel. output values are stored in outputValuesA and B.
-          //   A is compuated via serial interface and B is computed with top-level interface.
-          const int npts = 1000, ndim = 3;
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesA("outputValuesA", basis.getCardinality(), npts);
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesB("outputValuesB", basis.getCardinality(), npts);
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
           
-          Kokkos::View<PointValueType**,DeviceType> inputPointsViewToUseRandom("inputPoints", npts, ndim);
-          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (inputPointsViewToUseRandom.data(),  npts, ndim);
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
           
-          // random values between (-1,1) x (-1,1)
+          // random values between (0,1)
           Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
           Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
           
-          // compute setup
-          //   we need vinv and workspace
-          const auto vinv = basis.getVandermondeInverse();
-          
-          // worksize 
-          //   workspace per thread is required for serial interface. 
-          //   parallel_for with range policy would be good to use stack workspace 
-          //   as team policy only can create shared memory 
-          //   this part would be tricky as the max size should be determined at compile time
-          //   let's think about this and find out the best practice. for now I use the following.
-          constexpr int worksize = (Parameters::MaxOrder+1)*(Parameters::MaxOrder+1)*(Parameters::MaxOrder+1);
-          
-          // if you use team policy, worksize can be gathered from the basis object and use 
-          // kokkos shmem_size APIs to create workspace per team or per thread.
-          //const auto worksize_for_teampolicy = basis.getWorksizePerPoint(OPERATOR_VALUE);
-          
-          // extract point range to be evaluated in each thread
-          typedef Kokkos::pair<int,int> range_type;
-          
-          // parallel execution with serial interface
-          Kokkos::RangePolicy<DeviceSpaceType> policy(0, npts);
-          Kokkos::parallel_for(policy, KOKKOS_LAMBDA(int i) {
-              // we evaluate a single point 
-              const range_type pointRange = range_type(i,i+1);
-              
-              // out (# dofs, # pts), input (# pts, # dims)
-              auto output = Kokkos::subview(outputValuesA, Kokkos::ALL(), pointRange);
-              auto input  = Kokkos::subview(inputPoints,   pointRange, Kokkos::ALL());
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // wrap static workspace with a view; serial interface has a template view interface.
-              // either view or dynrankview with a right size is okay.
-              OutValueType workbuf[worksize];
-              Kokkos::View<OutValueType*,Kokkos::AnonymousSpace> work(&workbuf[0], worksize);
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute gradients
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // evaluate basis using serial interface
-              Impl::Basis_HGRAD_TET_Cn_FEM
-                ::Serial<OPERATOR_VALUE>::getValues(output, input, work, vinv);
-            });
-          
-          // evaluation using high level interface
-          basis.getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
-          
-          // compare 
-          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
-          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
-          
-          double sum = 0, diff = 0;
-          for (size_t i=0;i<outputValuesA_Host.extent(0);++i)
-            for (size_t j=0;j<outputValuesA_Host.extent(1);++j) {
-              sum += std::abs(outputValuesB_Host(i,j));
-              diff += std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(i,j));
-              if (verbose) {
-                std::cout << " order = " << order
-                          << " i = " << i << " j = " << j 
-                          << " val A = " << outputValuesA_Host(i,j) 
-                          << " val B = " << outputValuesB_Host(i,j) 
-                          << " diff  = " << (outputValuesA_Host(i,j) - outputValuesB_Host(i,j)) 
-                          << std::endl;
-              }
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
             }
-          if (diff/sum > 1.0e-9) {
-            errorFlag = -1;
+          }
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+          
+          *outStream << "Order: " << order << ": Comparing values and gradients on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare grads
+            const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+            const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+            
+            OutValueType diff = 0;
+            //Note, a PR release-debug serial build shows substantially higher errors (possibly due to operation rearrangements). Need to investigate more.
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                              << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
           }
         }
       } catch (std::exception &err) {
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/CMakeLists.txt
index eb726da6bb26..c5f307a89f52 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_TRI_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_TRI_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..1a918203d7cc
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_TRI_C1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_TRI_C1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/test_02.hpp
new file mode 100644
index 000000000000..928394ded0a4
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C1_FEM/test_02.hpp
@@ -0,0 +1,228 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_TRI_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_TRI_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_TRI_C1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TRI_C1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_TRI_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << "]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        {
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0) << ", " << outputCurlsA_Host(ic,i,j,1) <<"]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " <<  outputCurlsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/CMakeLists.txt
index 21c4f220d58a..ae831c937e39 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_TRI_C2_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_TRI_C2_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..cd49ca800b02
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_TRI_C2_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_TRI_C2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/test_02.hpp
new file mode 100644
index 000000000000..60b8c49aa454
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_C2_FEM/test_02.hpp
@@ -0,0 +1,228 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_TRI_C2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_TRI_C2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_TRI_C2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TRI_C2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_TRI_C2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 2;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute curls
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << "]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        {
+          // compare curls
+          const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+          const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", curls A: [" << outputCurlsA_Host(ic,i,j,0) << ", " << outputCurlsA_Host(ic,i,j,1) <<"]"
+                            << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " <<  outputCurlsB_Host(i,j,1) << "]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/CMakeLists.txt
index 28b96612c334..4855e54c85a1 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/CMakeLists.txt
@@ -7,6 +7,7 @@ SET(Intrepid2_TEST_ETI_FILE "test_01")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -94,11 +95,18 @@ SET(Intrepid2_TEST_ETI_FILE "test_02")
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
 LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
 
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_01_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_01_ETI.in
index 513cf7cb37bc..eaaead469fb6 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_01_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_01_ETI.in
@@ -14,7 +14,6 @@
 
 #include "Kokkos_Core.hpp"
 
-
 #define ETI_SACADO @ETI_SACADO@
 #if (ETI_SACADO != 0)  /// SACADO
 #include "Kokkos_ViewFactory.hpp"
@@ -29,16 +28,15 @@ constexpr int num_deriv = 10;
 #define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 23)
-///  Mauro, the master branch uses this derivative dimension which sounds dummy
-constexpr int num_deriv = 0; //9;
-constexpr int max_deriv = 1; //10;
-#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, max_deriv+1)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #elif (ETI_SACADO == 20)
 constexpr int num_deriv = 2;
 #define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
 #define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
 #endif
+
 #include "test_01.hpp"
 
 int main(int argc, char *argv[]) {
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_02_ETI.in
index c78997d6ea0c..a5343a485d3a 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_02_ETI.in
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/eti/test_02_ETI.in
@@ -14,15 +14,38 @@
 
 #include "Kokkos_Core.hpp"
 
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
 #include "test_02.hpp"
 
 int main(int argc, char *argv[]) {
 
+  const bool verbose = (argc-1) > 0;
   Kokkos::initialize();
-  {
-    const bool verbose = (argc-1) > 0;
-    Intrepid2::Test::HGRAD_TRI_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
-  }
+
+  Intrepid2::Test::HGRAD_TRI_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
   Kokkos::finalize();
   return 0;
 }
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/test_02.hpp
index 80d75c9bf099..a82178b45f9a 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/test_02.hpp
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_TRI_Cn_FEM/test_02.hpp
@@ -7,9 +7,9 @@
 // *****************************************************************************
 // @HEADER
 
-/** \file test_01.hpp
+/** \file test_02.hpp
     \brief  Unit tests for the Intrepid2::HGRAD_TRI_Cn_FEM class.
-    \author Created by P. Bochev, D. Ridzal, K. Peterson, Kyungjoo Kim
+    \author Created by Kyungjoo Kim, Mauro Perego
  */
 
 
@@ -23,99 +23,196 @@
 #include "Intrepid2_Utils.hpp"
 
 #include "Intrepid2_HGRAD_TRI_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
 
 namespace Intrepid2 {
 
   namespace Test {
 
-    // This code provides an example to use serial interface of high order elements
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
     template<typename OutValueType, typename PointValueType, typename DeviceType>
     int HGRAD_TRI_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_TRI_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
       using DeviceSpaceType = typename DeviceType::execution_space;
       Kokkos::print_configuration(std::cout, false);
 
       int errorFlag = 0;
-
+      constexpr int maxOrder = 9;
       try { 
-        for (int order=1;order<Parameters::MaxOrder;++order) {
-          Basis_HGRAD_TRI_Cn_FEM<DeviceType,OutValueType,PointValueType> basis(order);
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HGRAD_TRI_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
           
-          // problem setup 
-          //   let's say we want to evaluate 1000 points in parallel. output values are stored in outputValuesA and B.
-          //   A is compuated via serial interface and B is computed with top-level interface.
-          const int npts = 1000, ndim = 2;
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesA("outputValuesA", basis.getCardinality(), npts);
-          Kokkos::DynRankView<OutValueType,DeviceType> outputValuesB("outputValuesB", basis.getCardinality(), npts);
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsA, ncells, basisPtr->getCardinality(), npts, ndim);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputCurlsB, basisPtr->getCardinality(), npts, ndim);
+
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
           
-          Kokkos::View<PointValueType**,DeviceType> inputPointsViewToUseRandom("inputPoints", npts, ndim);
-          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (inputPointsViewToUseRandom.data(),  npts, ndim);
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
           
-          // random values between (-1,1) x (-1,1)
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
           Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
           Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
           
-          // compute setup
-          //   we need vinv and workspace
-          const auto vinv = basis.getVandermondeInverse();
-          
-          // worksize 
-          //   workspace per thread is required for serial interface. 
-          //   parallel_for with range policy would be good to use stack workspace 
-          //   as team policy only can create shared memory 
-          //   this part would be tricky as the max size should be determined at compile time
-          //   let's think about this and find out the best practice. for now I use the following.
-          constexpr int worksize = (Parameters::MaxOrder+1)*(Parameters::MaxOrder+1);
-          
-          // if you use team policy, worksize can be gathered from the basis object and use 
-          // kokkos shmem_size APIs to create workspace per team or per thread.
-          //const auto worksize_for_teampolicy = basis.getWorksizePerPoint(OPERATOR_VALUE);
-          
-          // extract point range to be evaluated in each thread
-          typedef Kokkos::pair<int,int> range_type;
-          
-          // parallel execution with serial interface
-          Kokkos::RangePolicy<DeviceSpaceType> policy(0, npts);
-          Kokkos::parallel_for(policy, KOKKOS_LAMBDA(int i) {
-              // we evaluate a single point 
-              const range_type pointRange = range_type(i,i+1);
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // out (# dofs, # pts), input (# pts, # dims)
-              auto output = Kokkos::subview(outputValuesA, Kokkos::ALL(), pointRange);
-              auto input  = Kokkos::subview(inputPoints,   pointRange, Kokkos::ALL());
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute gradients
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // wrap static workspace with a view; serial interface has a template view interface.
-              // either view or dynrankview with a right size is okay.
-              OutValueType workbuf[worksize];
-              Kokkos::View<OutValueType*,Kokkos::AnonymousSpace> work(&workbuf[0], worksize);
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+
+            { //compute curls
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto curlsACell = Kokkos::subview(outputCurlsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(curlsACell, inputPoints, OPERATOR_CURL, team_member, team_member.team_scratch(scratch_space_level));
+              };              
               
-              // evaluate basis using serial interface
-              Impl::Basis_HGRAD_TRI_Cn_FEM
-                ::Serial<OPERATOR_VALUE>::getValues(output, input, work, vinv);
-            });
-          
-          // evaluation using high level interface
-          basis.getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
-          
-          // compare 
-          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
-          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
-          
-          double sum = 0, diff = 0;
-          for (size_t i=0;i<outputValuesA_Host.extent(0);++i)
-            for (size_t j=0;j<outputValuesA_Host.extent(1);++j) {
-              sum += std::abs(outputValuesB_Host(i,j));
-              diff += std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(i,j));
-              if (verbose) {
-                std::cout << " order = " << order
-                          << " i = " << i << " j = " << j 
-                          << " val A = " << outputValuesA_Host(i,j) 
-                          << " val B = " << outputValuesB_Host(i,j) 
-                          << " diff  = " << (outputValuesA_Host(i,j) - outputValuesB_Host(i,j)) 
-                          << std::endl;
-              }
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_CURL);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
             }
-          if (diff/sum > 1.0e-9) {
-            errorFlag = -1;
+          }
+
+          *outStream << "Order: " << order << ": Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+          basisPtr->getValues(outputCurlsB, inputPoints, OPERATOR_CURL);
+          
+          *outStream << "Order: " << order << ": Comparing values and gradients on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare grads
+            const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+            const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << "]"
+                              << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+
+          { 
+            // compare curls
+            const auto outputCurlsA_Host = Kokkos::create_mirror_view(outputCurlsA); Kokkos::deep_copy(outputCurlsA_Host, outputCurlsA);
+            const auto outputCurlsB_Host = Kokkos::create_mirror_view(outputCurlsB); Kokkos::deep_copy(outputCurlsB_Host, outputCurlsB);
+            
+            OutValueType diff = 0;
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputCurlsA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputCurlsA_Host.extent(1);++i)
+                for (size_t j=0;j<outputCurlsA_Host.extent(2);++j) {
+                  diff = 0;
+                  for (int d=0;d<ndim;++d)
+                    diff += std::abs(outputCurlsB_Host(i,j,d) - outputCurlsA_Host(ic,i,j,d));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", curls A: [" << outputCurlsA_Host(ic,i,j,0) << ", " << outputCurlsA_Host(ic,i,j,1) <<"]"
+                              << ", curls B: [" << outputCurlsB_Host(i,j,0) << ", " <<  outputCurlsB_Host(i,j,1) << "]"
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
           }
         }
       } catch (std::exception &err) {
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/CMakeLists.txt
index a32463e45988..4b1a7626d4ff 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,9 +22,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_WEDGE_C1_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_WEDGE_C1_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..759d2a05ad2a
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_WEDGE_C1_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_WEDGE_C1_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/test_02.hpp
new file mode 100644
index 000000000000..33f9bb85137d
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C1_FEM/test_02.hpp
@@ -0,0 +1,184 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_WEDGE_C1_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_WEDGE_C1_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_WEDGE_C1_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_WEDGE_C1_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_WEDGE_C1_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/CMakeLists.txt
index 759b1bc6bc22..f82d6a4f7fc4 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/CMakeLists.txt
@@ -1,8 +1,16 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "")
+LIST(APPEND Intrepid2_TEST_ETI_FILE 
+  "test_01"
+  "test_01_Serendipity")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double")
@@ -17,12 +25,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
-# Host test
-SET(Intrepid2_TEST_ETI_FILE "")
-LIST(APPEND Intrepid2_TEST_ETI_FILE 
-  "test_01"
-  "test_01_Serendipity")
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HGRAD_WEDGE_C2_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HGRAD_WEDGE_C2_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
 
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_02")
+
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -83,4 +159,3 @@ TRIBITS_COPY_FILES_TO_BINARY_DIR(HGRAD_WEDGE_C2TestDataCopy
   SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/testdata"
   DEST_DIR "${CMAKE_CURRENT_BINARY_DIR}/testdata"
 )
-
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..86de2ff60329
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_01.cpp
+    \brief  Unit test of Intrepid2::Basis_HGRAD_WEDGE_C2_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HGRAD_WEDGE_C2_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/test_02.hpp
new file mode 100644
index 000000000000..35f0e18c3dbe
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HGRAD_WEDGE_C2_FEM/test_02.hpp
@@ -0,0 +1,182 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HGRAD_WEDGE_C2_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HGRAD_WEDGE_C2_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HGRAD_WEDGE_C2_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HGRAD_WEDGE_C2_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      try { 
+        using BasisType = Basis_HGRAD_WEDGE_C2_FEM<DeviceType,OutValueType,PointValueType>;
+        auto basisPtr = Teuchos::rcp(new BasisType());
+        
+        const int ncells = 5, npts = 10, ndim = 3;
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsA, ncells, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputGradsB, basisPtr->getCardinality(), npts, ndim);
+        Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+        
+        using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+        
+        Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+        auto vcprop = Kokkos::common_view_alloc_prop(point);
+        Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+        
+        // random values between (0,1)
+        Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+        Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+        
+
+        *outStream << "Computing values and gradients for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+        { // evaluation using parallel loop over cell
+          auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+          auto basisRawPtr_device = basisPtr_device.get();
+
+          int scratch_space_level =1;
+          const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+          Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+          { //compute values
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+
+          { //compute gradients
+            auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                auto gradsACell = Kokkos::subview(outputGradsA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                basisRawPtr_device->getValues(gradsACell, inputPoints, OPERATOR_GRAD, team_member, team_member.team_scratch(scratch_space_level));
+            };              
+            
+            //Get the required size of the scratch space per team and per thread.
+            int perThreadSpaceSize(0), perTeamSpaceSize(0);
+            basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_GRAD);
+            teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+            Kokkos::parallel_for (teamPolicy,functor);
+          }
+        }
+
+        *outStream << "Computing values and gradients for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+        // evaluation using high level interface (on one cell)
+        basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+        basisPtr->getValues(outputGradsB, inputPoints, OPERATOR_GRAD);
+        
+        *outStream << "Comparing values and gradients on host" <<std::endl;
+        { // compare values
+          const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+          const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+          
+          OutValueType diff = 0; 
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+              for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", val A: " << outputValuesA_Host(ic,i,j) 
+                            << ", val B: " << outputValuesB_Host(i,j) 
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+
+        { 
+          // compare grads
+          const auto outputGradsA_Host = Kokkos::create_mirror_view(outputGradsA); Kokkos::deep_copy(outputGradsA_Host, outputGradsA);
+          const auto outputGradsB_Host = Kokkos::create_mirror_view(outputGradsB); Kokkos::deep_copy(outputGradsB_Host, outputGradsB);
+          
+          OutValueType diff = 0;
+          auto tol = epsilon<double>();
+          for (size_t ic=0;ic<outputGradsA_Host.extent(0);++ic)
+            for (size_t i=0;i<outputGradsA_Host.extent(1);++i)
+              for (size_t j=0;j<outputGradsA_Host.extent(2);++j) {
+                diff = 0;
+                for (int d=0;d<ndim;++d)
+                  diff += std::abs(outputGradsB_Host(i,j,d) - outputGradsA_Host(ic,i,j,d));
+                if (diff > tol) {
+                  ++errorFlag;
+                  std::cout << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                            << ", grads A: [" << outputGradsA_Host(ic,i,j,0) << ", " << outputGradsA_Host(ic,i,j,1) << ", " <<  outputGradsA_Host(ic,i,j,2) <<"]"
+                            << ", grads B: [" << outputGradsB_Host(i,j,0) << ", " <<  outputGradsB_Host(i,j,1) << ", " << outputGradsB_Host(i,j,2) <<"]"
+                            << ", |diff|: " << diff
+                            << ", tol: " << tol
+                            << std::endl;
+                }
+              }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/CMakeLists.txt
index ebf2144cc03f..fa03caf7d02b 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HVOL_HEX_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HVOL_HEX_Cn_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HVOL_HEX_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..766fea986506
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HVOL_HEX_Cn_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HVOL_HEX_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/test_02.hpp
new file mode 100644
index 000000000000..115a371bc82a
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_HEX_Cn_FEM/test_02.hpp
@@ -0,0 +1,144 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HVOL_HEX_Cn_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HVOL_HEX_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HVOL_HEX_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HVOL_HEX_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HVOL_HEX_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          
+          *outStream << "Order: " << order << ": Comparing values on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/CMakeLists.txt
index f5caa97dcaf1..24d663415a74 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HVOL_LINE_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HVOL_LINE_Cn_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HVOL_LINE_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..a6f42f8ba7b8
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HVOL_LINE_Cn_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HVOL_LINE_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/test_02.hpp
new file mode 100644
index 000000000000..c71f401eda49
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_LINE_Cn_FEM/test_02.hpp
@@ -0,0 +1,144 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HVOL_LINE_Cn_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HVOL_LINE_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HVOL_LINE_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HVOL_LINE_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HVOL_LINE_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 1;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          
+          *outStream << "Order: " << order << ": Comparing values on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/CMakeLists.txt
index 769157316641..aedb132dcc82 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HVOL_QUAD_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HVOL_QUAD_Cn_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HVOL_QUAD_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..1c01cd896135
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HVOL_QUAD_Cn_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HVOL_QUAD_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/test_02.hpp
new file mode 100644
index 000000000000..044f8fad53a0
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_QUAD_Cn_FEM/test_02.hpp
@@ -0,0 +1,144 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HVOL_QUAD_Cn_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HVOL_QUAD_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HVOL_QUAD_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HVOL_QUAD_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HVOL_QUAD_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          
+          *outStream << "Order: " << order << ": Comparing values on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/CMakeLists.txt
index 0f61f9a7cfff..7dfea7c9986c 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HVOL_TET_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HVOL_TET_Cn_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HVOL_TET_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..d15050227457
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HVOL_TET_Cn_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HVOL_TET_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/test_02.hpp
new file mode 100644
index 000000000000..d27c666355c1
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TET_Cn_FEM/test_02.hpp
@@ -0,0 +1,144 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HVOL_TET_Cn_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HVOL_TET_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HVOL_TET_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HVOL_TET_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HVOL_TET_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 3;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          
+          *outStream << "Order: " << order << ": Comparing values on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/CMakeLists.txt b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/CMakeLists.txt
index 700140bb8337..49398919d584 100644
--- a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/CMakeLists.txt
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/CMakeLists.txt
@@ -1,8 +1,13 @@
 TRIBITS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 
+
+# test
+SET(Intrepid2_TEST_ETI_FILE "test_01")
+
 # value types
 SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
 SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
 
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
 LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
@@ -32,9 +37,80 @@ ENDIF()
 LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
 MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
 
+# device
+SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
+SET(Intrepid2_TEST_ETI_DEVICE "")
+IF(Kokkos_ENABLE_SERIAL)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "Serial")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Serial,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_OPENMP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "OpenMP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::OpenMP,Kokkos::HostSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_CUDA)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "CUDA")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::Cuda,Kokkos::CudaSpace>")
+ENDIF()
+IF(Kokkos_ENABLE_HIP)
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE_NAME "HIP")
+  LIST(APPEND Intrepid2_TEST_ETI_DEVICE      "Kokkos::Device<Kokkos::HIP,Kokkos::HIPSpace>")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_DEVICE_NAME ETI_DEVICE_COUNT)
+MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
+
+FOREACH(I RANGE ${ETI_DEVICE_COUNT})
+  LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
+  LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
+  #MESSAGE(STATUS "Generating TEST HVOL_TRI_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
+  FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
+    LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
+    LIST(GET Intrepid2_TEST_ETI_SACADO         ${J} ETI_SACADO)
+    FOREACH(ETI_FILE IN LISTS Intrepid2_TEST_ETI_FILE)
+      SET(ETI_NAME "${ETI_FILE}_${ETI_DEVICE_NAME}_${ETI_VALUETYPE_NAME}")
+      MESSAGE(STATUS "Generating TEST: HVOL_TRI_Cn_FEM ${ETI_NAME}.cpp")
+      CONFIGURE_FILE(eti/${ETI_FILE}_ETI.in ${ETI_NAME}.cpp)
+
+      TRIBITS_ADD_EXECUTABLE_AND_TEST(
+        ${ETI_NAME}
+        SOURCES ${ETI_NAME}.cpp 
+        ARGS PrintItAll
+        NUM_MPI_PROCS 1
+        PASS_REGULAR_EXPRESSION "TEST PASSED"
+        ADD_DIR_TO_NAME
+      )
+
+    ENDFOREACH()
+  ENDFOREACH()
+ENDFOREACH()
+
+
+
+
 # test
-SET(Intrepid2_TEST_ETI_FILE "test_01")
+SET(Intrepid2_TEST_ETI_FILE "test_02")
 
+# value types
+SET(Intrepid2_TEST_ETI_VALUETYPE_NAME "")
+SET(Intrepid2_TEST_ETI_VALUETYPE "")
+SET(Intrepid2_TEST_ETI_SACADO "")
+
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DOUBLE_DOUBLE")
+LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "double,double")
+LIST(APPEND Intrepid2_TEST_ETI_SACADO         "0")
+
+IF (HAVE_INTREPID2_SACADO) 
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE_NAME "DFAD_DFAD")
+  LIST(APPEND Intrepid2_TEST_ETI_VALUETYPE      "Sacado::Fad::DFad<double>,Sacado::Fad::DFad<double> ")
+  LIST(APPEND Intrepid2_TEST_ETI_SACADO         "33")
+ENDIF()
+
+LIST(LENGTH Intrepid2_TEST_ETI_VALUETYPE_NAME ETI_VALUETYPE_COUNT)
+MATH(EXPR ETI_VALUETYPE_COUNT "${ETI_VALUETYPE_COUNT}-1")
+
+# device
 SET(Intrepid2_TEST_ETI_DEVICE_NAME "")
 SET(Intrepid2_TEST_ETI_DEVICE "")
 IF(Kokkos_ENABLE_SERIAL)
@@ -60,7 +136,6 @@ MATH(EXPR ETI_DEVICE_COUNT "${ETI_DEVICE_COUNT}-1")
 FOREACH(I RANGE ${ETI_DEVICE_COUNT})
   LIST(GET Intrepid2_TEST_ETI_DEVICE_NAME ${I} ETI_DEVICE_NAME)
   LIST(GET Intrepid2_TEST_ETI_DEVICE      ${I} ETI_DEVICE)
-  #MESSAGE(STATUS "Generating TEST HVOL_TRI_Cn_FEM for ${ETI_DEVICE_NAME} with ${ETI_DEVICE}")
   FOREACH(J RANGE ${ETI_VALUETYPE_COUNT})
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE_NAME ${J} ETI_VALUETYPE_NAME)
     LIST(GET Intrepid2_TEST_ETI_VALUETYPE      ${J} ETI_VALUETYPE)
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/eti/test_02_ETI.in b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/eti/test_02_ETI.in
new file mode 100644
index 000000000000..4b98bc03263a
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/eti/test_02_ETI.in
@@ -0,0 +1,52 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.cpp
+    \brief  Unit test of Intrepid2::Basis_HVOL_TRI_Cn_FEM team-level getValues.
+    \author Kyungjoo Kim
+*/
+
+#include "Kokkos_Core.hpp"
+
+#define ETI_SACADO @ETI_SACADO@
+#if (ETI_SACADO != 0)  /// SACADO
+#include "Kokkos_ViewFactory.hpp"
+#include "Sacado.hpp"
+#endif
+
+#if (ETI_SACADO == 0)  /// double double
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#elif (ETI_SACADO == 11 /* SFAD SFAD */ || ETI_SACADO == 33 /* DFAD DFAD */)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 23)
+constexpr int num_deriv = 3;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#elif (ETI_SACADO == 20)
+constexpr int num_deriv = 2;
+#define ConstructWithLabelOutView(obj, ...) obj(#obj, __VA_ARGS__, num_deriv+1)
+#define ConstructWithLabelPointView(obj, ...) obj(#obj, __VA_ARGS__)
+#endif
+
+#include "test_02.hpp"
+
+int main(int argc, char *argv[]) {
+
+  const bool verbose = (argc-1) > 0;
+  Kokkos::initialize();
+
+  Intrepid2::Test::HVOL_TRI_Cn_FEM_Test02<@ETI_VALUETYPE@,@ETI_DEVICE@>(verbose);
+
+  Kokkos::finalize();
+  return 0;
+}
+
diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/test_02.hpp
new file mode 100644
index 000000000000..aaefaa951c33
--- /dev/null
+++ b/packages/intrepid2/unit-test/Discretization/Basis/HVOL_TRI_Cn_FEM/test_02.hpp
@@ -0,0 +1,145 @@
+// @HEADER
+// *****************************************************************************
+//                           Intrepid2 Package
+//
+// Copyright 2007 NTESS and the Intrepid2 contributors.
+// SPDX-License-Identifier: BSD-3-Clause
+// *****************************************************************************
+// @HEADER
+
+/** \file test_02.hpp
+    \brief  Unit tests for the Intrepid2::HVOL_TRI_Cn_FEM class.
+    \author Created by Kyungjoo Kim, Mauro Perego
+ */
+
+
+#include "Intrepid2_config.h"
+#include "Kokkos_Random.hpp"
+#ifdef HAVE_INTREPID2_DEBUG
+#define INTREPID2_TEST_FOR_DEBUG_ABORT_OVERRIDE_TO_CONTINUE
+#endif
+
+#include "Intrepid2_Types.hpp"
+#include "Intrepid2_Utils.hpp"
+
+#include "Intrepid2_HVOL_TRI_Cn_FEM.hpp"
+#include "packages/intrepid2/unit-test/Discretization/Basis/Setup.hpp"
+
+namespace Intrepid2 {
+
+  namespace Test {
+
+    // This test evaluates the basis functions at a set of points on a batch of cells using the team-level getValues,
+    // and compares the results with those obtained using the classic getValues function.
+    template<typename OutValueType, typename PointValueType, typename DeviceType>
+    int HVOL_TRI_Cn_FEM_Test02(const bool verbose) {
+
+      //! Setup test output stream.
+      Teuchos::RCP<std::ostream> outStream = setup_output_stream<DeviceType>(
+        verbose, "HVOL_TRI_Cn_FEM, Test 2", {}
+      );
+
+      *outStream
+      << "\n"
+      << "===============================================================================\n"
+      << "| Testing Team-level Implemntation of getValues                               |\n"
+      << "===============================================================================\n";
+
+      using DeviceSpaceType = typename DeviceType::execution_space;
+      Kokkos::print_configuration(std::cout, false);
+
+      int errorFlag = 0;
+
+      constexpr int maxOrder = 9;
+      try { 
+        for (int order=1;order<=maxOrder;++order) {
+          using BasisType = Basis_HVOL_TRI_Cn_FEM<DeviceType,OutValueType,PointValueType>;
+          auto basisPtr = Teuchos::rcp(new BasisType(order));
+          
+          const int ncells = 5, npts = 10, ndim = 2;
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesA, ncells, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<OutValueType,DeviceType> ConstructWithLabelOutView(outputValuesB, basisPtr->getCardinality(), npts);
+          Kokkos::DynRankView<PointValueType,DeviceType> ConstructWithLabelPointView(point, 1);
+          
+          using ScalarType = typename ScalarTraits<PointValueType>::scalar_type;
+          
+          Kokkos::View<ScalarType*,DeviceType> inputPointsViewToUseRandom("inputPoints", npts*ndim*get_dimension_scalar(point));
+          auto vcprop = Kokkos::common_view_alloc_prop(point);
+          Kokkos::DynRankView<PointValueType,DeviceType> inputPoints (Kokkos::view_wrap(inputPointsViewToUseRandom.data(), vcprop),  npts, ndim);
+          
+          // random values between (0,1)
+          Kokkos::Random_XorShift64_Pool<DeviceType> random(13718);
+          Kokkos::fill_random(inputPointsViewToUseRandom, random, 1.0);
+          
+
+          *outStream << "Order: " << order << ": Computing values for " << ncells << " cells and " << npts << " points using team-level getValues function" <<std::endl;
+
+          { // evaluation using parallel loop over cell
+            auto basisPtr_device = copy_virtual_class_to_device<DeviceType,BasisType>(*basisPtr);
+            auto basisRawPtr_device = basisPtr_device.get();
+
+            int scratch_space_level =1;
+            const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
+            Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+
+            { //compute values
+              auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
+                  auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
+                  basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
+              };              
+              
+              //Get the required size of the scratch space per team and per thread.
+              int perThreadSpaceSize(0), perTeamSpaceSize(0);
+              basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);
+              teamPolicy.set_scratch_size(scratch_space_level, Kokkos::PerTeam(perTeamSpaceSize), Kokkos::PerThread(perThreadSpaceSize));
+
+              Kokkos::parallel_for (teamPolicy,functor);
+            }
+          }
+
+          *outStream << "Order: " << order << ": Computing values for " << npts << " points using high-level getValues function" <<std::endl;
+
+
+          // evaluation using high level interface (on one cell)
+          basisPtr->getValues(outputValuesB, inputPoints, OPERATOR_VALUE);
+          
+          *outStream << "Order: " << order << ": Comparing values on host" <<std::endl;
+          { // compare values
+            const auto outputValuesA_Host = Kokkos::create_mirror_view(outputValuesA); Kokkos::deep_copy(outputValuesA_Host, outputValuesA);
+            const auto outputValuesB_Host = Kokkos::create_mirror_view(outputValuesB); Kokkos::deep_copy(outputValuesB_Host, outputValuesB);
+            
+            OutValueType diff = 0; 
+            auto tol = epsilon<double>();
+            for (size_t ic=0;ic<outputValuesA_Host.extent(0);++ic)
+              for (size_t i=0;i<outputValuesA_Host.extent(1);++i)
+                for (size_t j=0;j<outputValuesA_Host.extent(2);++j) {
+                  diff = std::abs(outputValuesB_Host(i,j) - outputValuesA_Host(ic,i,j));
+                  if (diff > tol) {
+                    ++errorFlag;
+                    std::cout << " order: " << order
+                              << ", ic: " << ic << ", i: " << i << ", j: " << j 
+                              << ", val A: " << outputValuesA_Host(ic,i,j) 
+                              << ", val B: " << outputValuesB_Host(i,j) 
+                              << ", |diff|: " << diff
+                              << ", tol: " << tol
+                              << std::endl;
+                  }
+                }
+          }
+        }
+      } catch (std::exception &err) {
+        std::cout << "UNEXPECTED ERROR !!! ----------------------------------------------------------\n";
+        std::cout << err.what() << '\n';
+        std::cout << "-------------------------------------------------------------------------------" << "\n\n";
+        errorFlag = -1000;
+      };
+      
+      if (errorFlag != 0)
+        std::cout << "End Result: TEST FAILED\n";
+      else
+        std::cout << "End Result: TEST PASSED\n";
+      
+      return errorFlag;
+    }
+  }
+}