diff --git a/batched/dense/unit_test/Test_Batched_SerialGemm.hpp b/batched/dense/unit_test/Test_Batched_SerialGemm.hpp
index 144bb2251e..0b2ed4a162 100644
--- a/batched/dense/unit_test/Test_Batched_SerialGemm.hpp
+++ b/batched/dense/unit_test/Test_Batched_SerialGemm.hpp
@@ -25,6 +25,7 @@
 #include "KokkosBatched_Gemm_Serial_Impl.hpp"
 
 #include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestVanilla.hpp"
 
 using namespace KokkosBatched;
 
diff --git a/batched/dense/unit_test/Test_Batched_SerialSVD.hpp b/batched/dense/unit_test/Test_Batched_SerialSVD.hpp
index a30c9dcc89..586b69dc7f 100644
--- a/batched/dense/unit_test/Test_Batched_SerialSVD.hpp
+++ b/batched/dense/unit_test/Test_Batched_SerialSVD.hpp
@@ -18,6 +18,8 @@
 #include "KokkosBatched_SVD_Decl.hpp"             //For testing overall kernel
 #include "KokkosBatched_SVD_Serial_Internal.hpp"  //For unit testing individual components
 #include "KokkosBatched_SetIdentity_Decl.hpp"
+#include "KokkosKernels_TestMatrixUtils.hpp"
+#include "KokkosKernels_TestVanilla.hpp"
 
 namespace Test {
 template <typename Scalar>
diff --git a/blas/unit_test/Test_Blas2_gemv.hpp b/blas/unit_test/Test_Blas2_gemv.hpp
index d70935c2ac..be6e4a6d51 100644
--- a/blas/unit_test/Test_Blas2_gemv.hpp
+++ b/blas/unit_test/Test_Blas2_gemv.hpp
@@ -19,6 +19,7 @@
 #include <KokkosBlas2_gemv.hpp>
 #include <KokkosBlas1_dot.hpp>
 #include <KokkosKernels_TestUtils.hpp>
+#include "KokkosKernels_TestVanilla.hpp"
 
 namespace Test {
 template <class ExecutionSpace, class ViewTypeA, class ViewTypeX, class ViewTypeY, class Device>
diff --git a/perf_test/CMakeLists.txt b/perf_test/CMakeLists.txt
index 28271dfb0d..8d740966f0 100644
--- a/perf_test/CMakeLists.txt
+++ b/perf_test/CMakeLists.txt
@@ -11,21 +11,6 @@ if (KokkosKernels_ENABLE_PERFTESTS)
     #build correctly with or without MPI, but only run them with a single
     #MPI process.
 
-    SET(GTEST_SOURCE_DIR ${PACKAGE_SOURCE_DIR}/tpls/gtest)
-
-    KOKKOSKERNELS_ADD_TEST_LIBRARY(
-                                kokkoskernelsperf_gtest
-                                HEADERS ${GTEST_SOURCE_DIR}/gtest/gtest.h
-                                SOURCES ${GTEST_SOURCE_DIR}/gtest/gtest-all.cc
-                                    )
-    #Disables pthreads, this is a problem for serial builds in Trilinos & Sierra if it's enabled.
-
-    TARGET_COMPILE_DEFINITIONS(kokkoskernelsperf_gtest PUBLIC "-DGTEST_HAS_PTHREAD=0")
-    TARGET_INCLUDE_DIRECTORIES(kokkoskernelsperf_gtest PUBLIC $<BUILD_INTERFACE:${GTEST_SOURCE_DIR}>)
-
-    #Gtest minimally requires C++ 11
-    TARGET_COMPILE_FEATURES(kokkoskernelsperf_gtest PUBLIC cxx_std_11)
-
     KOKKOSKERNELS_INCLUDE_DIRECTORIES(sparse)
 
     if(KokkosKernels_ENABLE_TESTS_AND_PERFSUITE)
diff --git a/perf_test/KokkosKernels_perf_test_utilities.hpp b/perf_test/KokkosKernels_perf_test_utilities.hpp
index ec767c68f7..6d11e021a9 100644
--- a/perf_test/KokkosKernels_perf_test_utilities.hpp
+++ b/perf_test/KokkosKernels_perf_test_utilities.hpp
@@ -20,7 +20,7 @@
 #ifndef KOKKOSKERNELS_PERF_TEST_UTILITIES_HPP
 #define KOKKOSKERNELS_PERF_TEST_UTILITIES_HPP
 
-#include "KokkosKernels_TestUtils.hpp"  // for string_compare_no_case
+#include "KokkosKernels_TestStringUtils.hpp"  // for string_compare_no_case
 
 // Namepsace that defines common utilities
 // for performance tests
diff --git a/perf_test/blas/blas1/KokkosBlas_dot_mv_perf_test.cpp b/perf_test/blas/blas1/KokkosBlas_dot_mv_perf_test.cpp
index 23d53ba106..49c7a76210 100644
--- a/perf_test/blas/blas1/KokkosBlas_dot_mv_perf_test.cpp
+++ b/perf_test/blas/blas1/KokkosBlas_dot_mv_perf_test.cpp
@@ -14,10 +14,12 @@
 //
 //@HEADER
 
+#include <iostream>
+
 #include <Kokkos_Core.hpp>
 #include <src/KokkosBlas1_dot.hpp>
 #include <Kokkos_Random.hpp>
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 struct Params {
   int use_cuda    = 0;
diff --git a/perf_test/blas/blas1/KokkosBlas_dot_perf_test.cpp b/perf_test/blas/blas1/KokkosBlas_dot_perf_test.cpp
index 8e6ebf985c..36f11733ce 100644
--- a/perf_test/blas/blas1/KokkosBlas_dot_perf_test.cpp
+++ b/perf_test/blas/blas1/KokkosBlas_dot_perf_test.cpp
@@ -14,12 +14,14 @@
 //
 //@HEADER
 
+#include <iostream>
+
 #include <Kokkos_Core.hpp>
 #include <Kokkos_Random.hpp>
 
 // For RPS implementation
 #include "KokkosBlas_dot_perf_test.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 struct Params {
   int use_cuda    = 0;
diff --git a/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test.cpp b/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test.cpp
index 16dadd34b6..ca30ed50ac 100644
--- a/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test.cpp
+++ b/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test.cpp
@@ -14,10 +14,12 @@
 //
 //@HEADER
 
+#include <iostream>
+
 #include <Kokkos_Core.hpp>
 #include <KokkosBlas1_team_dot.hpp>
 #include <Kokkos_Random.hpp>
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 struct Params {
   int use_cuda    = 0;
diff --git a/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test_benchmark.cpp b/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test_benchmark.cpp
index 776c291ad0..ab0750e5b6 100644
--- a/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test_benchmark.cpp
+++ b/perf_test/blas/blas1/KokkosBlas_team_dot_perf_test_benchmark.cpp
@@ -42,10 +42,12 @@
 //@HEADER
 */
 
+#include <iostream>
+
 #include <Kokkos_Core.hpp>
 #include <KokkosBlas1_team_dot.hpp>
 #include <Kokkos_Random.hpp>
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 #include <benchmark/benchmark.h>
 
diff --git a/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test.cpp b/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test.cpp
index 924b63ecee..90fb065741 100644
--- a/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test.cpp
+++ b/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test.cpp
@@ -16,7 +16,7 @@
 
 #include "KokkosBlas2_gemv.hpp"
 #include <Kokkos_Random.hpp>
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 struct Params {
   int use_cuda    = 0;
diff --git a/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test_benchmark.cpp b/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test_benchmark.cpp
index cdae173ce0..2ff22b7078 100644
--- a/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test_benchmark.cpp
+++ b/perf_test/blas/blas2/KokkosBlas2_gemv_perf_test_benchmark.cpp
@@ -19,7 +19,7 @@
 
 #include "KokkosBlas2_gemv.hpp"
 
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include <Benchmark_Context.hpp>
diff --git a/perf_test/blas/blas2/KokkosBlas2_ger_perf_test_benchmark.cpp b/perf_test/blas/blas2/KokkosBlas2_ger_perf_test_benchmark.cpp
index 8efd9920b1..2feeb2c2ff 100644
--- a/perf_test/blas/blas2/KokkosBlas2_ger_perf_test_benchmark.cpp
+++ b/perf_test/blas/blas2/KokkosBlas2_ger_perf_test_benchmark.cpp
@@ -21,7 +21,7 @@
 #include "KokkosBlas2_ger.hpp"
 #include <typeinfo>
 
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include <Benchmark_Context.hpp>
diff --git a/perf_test/blas/blas3/CMakeLists.txt b/perf_test/blas/blas3/CMakeLists.txt
index 80c9d25c1c..4dbf0ae5dd 100644
--- a/perf_test/blas/blas3/CMakeLists.txt
+++ b/perf_test/blas/blas3/CMakeLists.txt
@@ -4,7 +4,6 @@ KOKKOSKERNELS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 KOKKOSKERNELS_ADD_EXECUTABLE(
     KokkosBlas3_perf_test
     SOURCES KokkosBlas3_perf_test.cpp
-    TESTONLYLIBS kokkoskernelsperf_gtest
 )
 
 KOKKOSKERNELS_ADD_EXECUTABLE(
diff --git a/perf_test/blas/blas3/KokkosBlas3_gemm_perf_test.hpp b/perf_test/blas/blas3/KokkosBlas3_gemm_perf_test.hpp
index 7e3b0061d8..65a5cdc549 100644
--- a/perf_test/blas/blas3/KokkosBlas3_gemm_perf_test.hpp
+++ b/perf_test/blas/blas3/KokkosBlas3_gemm_perf_test.hpp
@@ -25,11 +25,12 @@
 
 #include <KokkosBlas3_gemm.hpp>
 
+#include "Kokkos_ArithTraits.hpp"
 #include "KokkosBatched_HostLevel_Gemm.hpp"
 #include "KokkosBatched_Gemm_Decl.hpp"
 #include "KokkosBatched_Util.hpp"
-#include "gtest/gtest.h"  // EXPECT_NEAR
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
+#include "KokkosKernels_TestVanilla.hpp"
 
 #include <chrono>
 
@@ -1368,7 +1369,8 @@ static inline bool __gemm_print_compare_failure(ViewType h_expected, ViewType h_
  */
 template <class ScalarType, class LayoutType>
 static inline bool __gemm_do_compare(view_type_3d expected, view_type_3d actual) {
-  double epsilon = Test::epsilon<ScalarType>::value * 1e3;
+  double epsilon = Kokkos::ArithTraits<ScalarType>::eps() * 1e3;
+
   STATUS;
 
   typename view_type_3d::HostMirror h_expected = Kokkos::create_mirror_view(expected);
diff --git a/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test.cpp b/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test.cpp
index aead1cb46f..a9f8db5be4 100644
--- a/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test.cpp
+++ b/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test.cpp
@@ -16,7 +16,7 @@
 
 #include "KokkosBlas3_gemm.hpp"
 #include <Kokkos_Random.hpp>
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 struct Params {
   int use_cuda    = 0;
diff --git a/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test_benchmark.cpp b/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test_benchmark.cpp
index eee2eea53c..6f0737180e 100644
--- a/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test_benchmark.cpp
+++ b/perf_test/blas/blas3/KokkosBlas3_gemm_standalone_perf_test_benchmark.cpp
@@ -16,7 +16,7 @@
 
 #include "KokkosBlas3_gemm.hpp"
 #include <Kokkos_Random.hpp>
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 #include "Benchmark_Context.hpp"
 #include <benchmark/benchmark.h>
diff --git a/perf_test/graph/KokkosGraph_color.cpp b/perf_test/graph/KokkosGraph_color.cpp
index 548abd0052..0c3706bdfe 100644
--- a/perf_test/graph/KokkosGraph_color.cpp
+++ b/perf_test/graph/KokkosGraph_color.cpp
@@ -26,7 +26,7 @@
 #include "KokkosSparse_CrsMatrix.hpp"
 #include "KokkosKernels_TestParameters.hpp"
 #include "KokkosGraph_Distance1Color.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosSparse_IOUtils.hpp"
 
 void print_options(std::ostream &os, const char *app_name, unsigned int indent = 0) {
diff --git a/perf_test/graph/KokkosGraph_color_d2.cpp b/perf_test/graph/KokkosGraph_color_d2.cpp
index 0811f7c5b4..04242f31f0 100644
--- a/perf_test/graph/KokkosGraph_color_d2.cpp
+++ b/perf_test/graph/KokkosGraph_color_d2.cpp
@@ -36,7 +36,7 @@
 #include <KokkosKernels_TestParameters.hpp>
 #include <KokkosGraph_Distance2Color.hpp>
 #include "KokkosKernels_default_types.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosSparse_IOUtils.hpp"
 
 using namespace KokkosGraph;
diff --git a/perf_test/graph/KokkosGraph_mis_d2.cpp b/perf_test/graph/KokkosGraph_mis_d2.cpp
index 3d08db0a9f..952f28e64c 100644
--- a/perf_test/graph/KokkosGraph_mis_d2.cpp
+++ b/perf_test/graph/KokkosGraph_mis_d2.cpp
@@ -37,7 +37,7 @@
 #include "KokkosSparse_spadd.hpp"
 #include "KokkosGraph_MIS2.hpp"
 #include "KokkosKernels_default_types.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosSparse_IOUtils.hpp"
 
 using namespace KokkosGraph;
diff --git a/perf_test/graph/KokkosGraph_triangle.cpp b/perf_test/graph/KokkosGraph_triangle.cpp
index cd5f40c376..84676607d9 100644
--- a/perf_test/graph/KokkosGraph_triangle.cpp
+++ b/perf_test/graph/KokkosGraph_triangle.cpp
@@ -18,7 +18,7 @@
 #include "KokkosGraph_Triangle.hpp"
 #include "KokkosSparse_CrsMatrix.hpp"
 #include "KokkosSparse_IOUtils.hpp"  //for read_kokkos_crst_graph
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_TestParameters.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
diff --git a/perf_test/lapack/KokkosLapack_SVD_benchmark.cpp b/perf_test/lapack/KokkosLapack_SVD_benchmark.cpp
index 8336b3b737..55aa70b9cf 100644
--- a/perf_test/lapack/KokkosLapack_SVD_benchmark.cpp
+++ b/perf_test/lapack/KokkosLapack_SVD_benchmark.cpp
@@ -14,9 +14,11 @@
 //
 //@HEADER
 
-#include "KokkosLapack_svd.hpp"
+#include <Kokkos_Random.hpp>
 
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_IOUtils.hpp"  // getRandomBounds
+#include "KokkosLapack_svd.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include <benchmark/benchmark.h>
@@ -72,7 +74,7 @@ void run_svd_benchmark(benchmark::State& state, const svd_parameters& svd_params
 
   // Initialize A with random numbers
   double randStart = 0, randEnd = 0;
-  Test::getRandomBounds(10.0, randStart, randEnd);
+  KokkosKernels::Impl::getRandomBounds(10.0, randStart, randEnd);
   Kokkos::fill_random(A, rand_pool, randStart, randEnd);
 
   for (auto _ : state) {
diff --git a/perf_test/ode/KokkosODE_BDF.cpp b/perf_test/ode/KokkosODE_BDF.cpp
index 60cc5e01ff..0b1ecdf276 100644
--- a/perf_test/ode/KokkosODE_BDF.cpp
+++ b/perf_test/ode/KokkosODE_BDF.cpp
@@ -16,7 +16,7 @@
 
 #include "KokkosODE_BDF.hpp"
 
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include <benchmark/benchmark.h>
diff --git a/perf_test/ode/KokkosODE_RK.cpp b/perf_test/ode/KokkosODE_RK.cpp
index 83635d7af6..2c52cfbb5d 100644
--- a/perf_test/ode/KokkosODE_RK.cpp
+++ b/perf_test/ode/KokkosODE_RK.cpp
@@ -16,7 +16,7 @@
 
 #include "KokkosODE_RungeKutta.hpp"
 
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include <benchmark/benchmark.h>
diff --git a/perf_test/sparse/CMakeLists.txt b/perf_test/sparse/CMakeLists.txt
index 514ef0ed82..70b5928efb 100644
--- a/perf_test/sparse/CMakeLists.txt
+++ b/perf_test/sparse/CMakeLists.txt
@@ -103,7 +103,6 @@ KOKKOSKERNELS_ADD_EXECUTABLE(
 KOKKOSKERNELS_ADD_EXECUTABLE(
         sparse_gs
         SOURCES KokkosSparse_gs.cpp
-        TESTONLYLIBS kokkoskernelsperf_gtest
 )
 
 KOKKOSKERNELS_ADD_EXECUTABLE(
diff --git a/perf_test/sparse/KokkosSparse_block_pcg.cpp b/perf_test/sparse/KokkosSparse_block_pcg.cpp
index f1f9684dd6..91a599ad71 100644
--- a/perf_test/sparse/KokkosSparse_block_pcg.cpp
+++ b/perf_test/sparse/KokkosSparse_block_pcg.cpp
@@ -23,7 +23,7 @@
 #include "KokkosKernels_Utils.hpp"
 #include "KokkosSparse_IOUtils.hpp"
 
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 #define MAXVAL 1
 
diff --git a/perf_test/sparse/KokkosSparse_gs.cpp b/perf_test/sparse/KokkosSparse_gs.cpp
index 9c968c92cd..0ed984a8f8 100644
--- a/perf_test/sparse/KokkosSparse_gs.cpp
+++ b/perf_test/sparse/KokkosSparse_gs.cpp
@@ -14,10 +14,17 @@
 //
 //@HEADER
 
+#include <iostream>
+#include <random>
+#include <vector>
+#include <string>
+#include <unordered_set>
+
 #include <Kokkos_Core.hpp>
 #include <Kokkos_Timer.hpp>
+
 #include <KokkosKernels_Handle.hpp>
-#include <KokkosKernels_TestUtils.hpp>
+#include <KokkosKernels_TestStringUtils.hpp>
 #include <KokkosSparse_gauss_seidel.hpp>
 #include <KokkosSparse_spmv.hpp>
 #include <KokkosKernels_IOUtils.hpp>
@@ -25,11 +32,7 @@
 #include <KokkosKernels_config.h>
 #include "KokkosKernels_default_types.hpp"
 #include "KokkosSparse_IOUtils.hpp"
-#include <iostream>
-#include <random>
-#include <vector>
-#include <string>
-#include <unordered_set>
+#include "KokkosKernels_TestMatrixUtils.hpp"
 
 using std::cout;
 using std::string;
diff --git a/perf_test/sparse/KokkosSparse_mdf.cpp b/perf_test/sparse/KokkosSparse_mdf.cpp
index 4db45ce55f..0cd33f2eca 100644
--- a/perf_test/sparse/KokkosSparse_mdf.cpp
+++ b/perf_test/sparse/KokkosSparse_mdf.cpp
@@ -19,7 +19,7 @@
 #include "KokkosKernels_Handle.hpp"
 #include "KokkosSparse_IOUtils.hpp"
 #include "KokkosSparse_Utils_cusparse.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include "KokkosSparse_mdf.hpp"
diff --git a/perf_test/sparse/KokkosSparse_pcg.cpp b/perf_test/sparse/KokkosSparse_pcg.cpp
index e1d469d6aa..00466ae6c9 100644
--- a/perf_test/sparse/KokkosSparse_pcg.cpp
+++ b/perf_test/sparse/KokkosSparse_pcg.cpp
@@ -20,7 +20,7 @@
 #include "KokkosKernels_Utils.hpp"
 #include "KokkosKernels_IOUtils.hpp"
 #include "KokkosKernels_default_types.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosSparse_IOUtils.hpp"
 #include <iostream>
 
diff --git a/perf_test/sparse/KokkosSparse_spadd.cpp b/perf_test/sparse/KokkosSparse_spadd.cpp
index d2b41608db..063c151812 100644
--- a/perf_test/sparse/KokkosSparse_spadd.cpp
+++ b/perf_test/sparse/KokkosSparse_spadd.cpp
@@ -20,7 +20,7 @@
 #include "KokkosSparse_IOUtils.hpp"
 #include "KokkosSparse_Utils_cusparse.hpp"
 #include "KokkosSparse_Utils_mkl.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #include "KokkosSparse_spadd.hpp"
diff --git a/perf_test/sparse/KokkosSparse_spgemm.cpp b/perf_test/sparse/KokkosSparse_spgemm.cpp
index 0dd13f6804..2c63127c2d 100644
--- a/perf_test/sparse/KokkosSparse_spgemm.cpp
+++ b/perf_test/sparse/KokkosSparse_spgemm.cpp
@@ -23,7 +23,7 @@
 #include "KokkosBlas1_nrminf.hpp"
 #include "KokkosBlas1_axpby.hpp"
 #include "KokkosKernels_TestParameters.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 #define TRANSPOSEFIRST false
diff --git a/perf_test/sparse/KokkosSparse_spgemm_jacobi.cpp b/perf_test/sparse/KokkosSparse_spgemm_jacobi.cpp
index 3b66889fab..a5646603b7 100644
--- a/perf_test/sparse/KokkosSparse_spgemm_jacobi.cpp
+++ b/perf_test/sparse/KokkosSparse_spgemm_jacobi.cpp
@@ -18,7 +18,7 @@
 #include "KokkosKernels_default_types.hpp"
 #include "KokkosKernels_IOUtils.hpp"
 #include "KokkosSparse_run_spgemm_jacobi.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 
 void print_options() {
   std::cerr << "Options\n" << std::endl;
diff --git a/perf_test/sparse/KokkosSparse_spmv_benchmark.cpp b/perf_test/sparse/KokkosSparse_spmv_benchmark.cpp
index 5ecf2f7248..6a2f92165b 100644
--- a/perf_test/sparse/KokkosSparse_spmv_benchmark.cpp
+++ b/perf_test/sparse/KokkosSparse_spmv_benchmark.cpp
@@ -21,7 +21,7 @@
 #include <KokkosKernels_IOUtils.hpp>
 #include <KokkosSparse_IOUtils.hpp>
 #include "KokkosKernels_default_types.hpp"
-#include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestStringUtils.hpp"
 #include "KokkosKernels_perf_test_utilities.hpp"
 
 // Headers for benchmark library
diff --git a/sparse/unit_test/Test_Sparse_gauss_seidel.hpp b/sparse/unit_test/Test_Sparse_gauss_seidel.hpp
index ebfd2a6a18..3704b63fbd 100644
--- a/sparse/unit_test/Test_Sparse_gauss_seidel.hpp
+++ b/sparse/unit_test/Test_Sparse_gauss_seidel.hpp
@@ -14,6 +14,13 @@
 //
 //@HEADER
 
+#include <cstdlib>
+#include <iostream>
+#include <complex>
+#include <map>
+#include <random>
+#include <vector>
+
 #include <gtest/gtest.h>
 
 #include <Kokkos_Core.hpp>
@@ -25,17 +32,13 @@
 #include <KokkosBlas1_dot.hpp>
 #include <KokkosBlas1_axpby.hpp>
 #include <KokkosBlas1_nrm2.hpp>
-#include <cstdlib>
-#include <iostream>
-#include <complex>
-#include <map>
-#include <random>
-#include <vector>
+
 #include "KokkosSparse_gauss_seidel.hpp"
 #include "KokkosSparse_partitioning_impl.hpp"
 #include "KokkosSparse_sor_sequential_impl.hpp"
 #include "KokkosSparse_SortCrs.hpp"
 #include "KokkosKernels_TestUtils.hpp"
+#include "KokkosKernels_TestMatrixUtils.hpp"
 #include "Test_Sparse_Utils.hpp"
 
 // #ifndef kokkos_complex_double
diff --git a/test_common/KokkosKernels_TestMatrixUtils.hpp b/test_common/KokkosKernels_TestMatrixUtils.hpp
new file mode 100644
index 0000000000..998ccdba90
--- /dev/null
+++ b/test_common/KokkosKernels_TestMatrixUtils.hpp
@@ -0,0 +1,89 @@
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 4.0
+//       Copyright (2022) National Technology & Engineering
+//               Solutions of Sandia, LLC (NTESS).
+//
+// Under the terms of Contract DE-NA0003525 with NTESS,
+// the U.S. Government retains certain rights in this software.
+//
+// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions.
+// See https://kokkos.org/LICENSE for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//@HEADER
+
+#ifndef KOKKOSKERNELS_TESTMATRIXUTILS_HPP
+#define KOKKOSKERNELS_TESTMATRIXUTILS_HPP
+
+#include <random>
+
+#include "KokkosKernels_Utils.hpp"
+#include "KokkosKernels_IOUtils.hpp"
+#include "Kokkos_ArithTraits.hpp"
+#include "KokkosBatched_Vector.hpp"
+// Make this include-able from all subdirectories
+
+namespace Test {
+
+template <typename scalar_t, typename lno_t, typename size_type, typename device, typename crsMat_t>
+crsMat_t symmetrize(crsMat_t A) {
+  typedef typename crsMat_t::StaticCrsGraphType graph_t;
+  typedef typename crsMat_t::values_type::non_const_type scalar_view_t;
+  typedef typename graph_t::row_map_type::non_const_type lno_view_t;
+  typedef typename graph_t::entries_type::non_const_type lno_nnz_view_t;
+  auto host_rowmap  = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A.graph.row_map);
+  auto host_entries = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A.graph.entries);
+  auto host_values  = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A.values);
+  lno_t numRows     = A.numRows();
+  // symmetrize as input_mat + input_mat^T, to still have a diagonally dominant
+  // matrix
+  typedef std::map<lno_t, scalar_t> Row;
+  std::vector<Row> symRows(numRows);
+  for (lno_t r = 0; r < numRows; r++) {
+    auto& row = symRows[r];
+    for (size_type i = host_rowmap(r); i < host_rowmap(r + 1); i++) {
+      lno_t c   = host_entries(i);
+      auto& col = symRows[c];
+      auto it   = row.find(c);
+      if (it == row.end())
+        row[c] = host_values(i);
+      else
+        row[c] += host_values(i);
+      it = col.find(r);
+      if (it == col.end())
+        col[r] = host_values(i);
+      else
+        col[r] += host_values(i);
+    }
+  }
+  // Count entries
+  Kokkos::View<size_type*, Kokkos::LayoutLeft, Kokkos::HostSpace> new_host_rowmap("Rowmap", numRows + 1);
+  size_t accum = 0;
+  for (lno_t r = 0; r <= numRows; r++) {
+    new_host_rowmap(r) = accum;
+    if (r < numRows) accum += symRows[r].size();
+  }
+  // Allocate new entries/values
+  Kokkos::View<lno_t*, Kokkos::LayoutLeft, Kokkos::HostSpace> new_host_entries("Entries", accum);
+  Kokkos::View<scalar_t*, Kokkos::LayoutLeft, Kokkos::HostSpace> new_host_values("Values", accum);
+  for (lno_t r = 0; r < numRows; r++) {
+    auto rowIt = symRows[r].begin();
+    for (size_type i = new_host_rowmap(r); i < new_host_rowmap(r + 1); i++) {
+      new_host_entries(i) = rowIt->first;
+      new_host_values(i)  = rowIt->second;
+      rowIt++;
+    }
+  }
+  lno_view_t new_rowmap("Rowmap", numRows + 1);
+  lno_nnz_view_t new_entries("Entries", accum);
+  scalar_view_t new_values("Values", accum);
+  Kokkos::deep_copy(new_rowmap, new_host_rowmap);
+  Kokkos::deep_copy(new_entries, new_host_entries);
+  Kokkos::deep_copy(new_values, new_host_values);
+  return crsMat_t("SymA", numRows, numRows, accum, new_values, new_rowmap, new_entries);
+}
+
+}  // namespace Test
+#endif
diff --git a/test_common/KokkosKernels_TestStringUtils.hpp b/test_common/KokkosKernels_TestStringUtils.hpp
new file mode 100644
index 0000000000..fec1b09361
--- /dev/null
+++ b/test_common/KokkosKernels_TestStringUtils.hpp
@@ -0,0 +1,38 @@
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 4.0
+//       Copyright (2022) National Technology & Engineering
+//               Solutions of Sandia, LLC (NTESS).
+//
+// Under the terms of Contract DE-NA0003525 with NTESS,
+// the U.S. Government retains certain rights in this software.
+//
+// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions.
+// See https://kokkos.org/LICENSE for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//@HEADER
+
+#ifndef KOKKOSKERNELS_TESTSTRINGUTILS_HPP
+#define KOKKOSKERNELS_TESTSTRINGUTILS_HPP
+
+#include <string>
+#include <cstring>
+
+namespace Test {
+
+inline int string_compare_no_case(const char* str1, const char* str2) {
+  std::string str1_s(str1);
+  std::string str2_s(str2);
+  for (size_t i = 0; i < str1_s.size(); i++) str1_s[i] = std::tolower(str1_s[i]);
+  for (size_t i = 0; i < str2_s.size(); i++) str2_s[i] = std::tolower(str2_s[i]);
+  return std::strcmp(str1_s.c_str(), str2_s.c_str());
+}
+
+inline int string_compare_no_case(const std::string& str1, const std::string& str2) {
+  return string_compare_no_case(str1.c_str(), str2.c_str());
+}
+
+}  // namespace Test
+#endif
diff --git a/test_common/KokkosKernels_TestUtils.hpp b/test_common/KokkosKernels_TestUtils.hpp
index c829278867..1d33bc8b75 100644
--- a/test_common/KokkosKernels_TestUtils.hpp
+++ b/test_common/KokkosKernels_TestUtils.hpp
@@ -212,153 +212,6 @@ using halfScalarType = Kokkos::Experimental::half_t;
 using bhalfScalarType = Kokkos::Experimental::bhalf_t;
 #endif  // KOKKOS_BHALF_T_IS_FLOAT
 
-template <class ViewTypeA, class ViewTypeB, class ViewTypeC, class ExecutionSpace>
-struct SharedVanillaGEMM {
-  bool A_t, B_t, A_c, B_c;
-  int C_rows, C_cols, A_cols;
-  ViewTypeA A;
-  ViewTypeB B;
-  ViewTypeC C;
-
-  typedef typename ViewTypeA::value_type ScalarA;
-  typedef typename ViewTypeB::value_type ScalarB;
-  typedef typename ViewTypeC::value_type ScalarC;
-  typedef Kokkos::View<ScalarA*, Kokkos::LayoutStride, typename ViewTypeA::device_type> SubviewTypeA;
-  typedef Kokkos::View<ScalarB*, Kokkos::LayoutStride, typename ViewTypeB::device_type> SubviewTypeB;
-  typedef Kokkos::ArithTraits<ScalarC> APT;
-  typedef typename APT::mag_type mag_type;
-  ScalarA alpha;
-  ScalarC beta;
-
-  KOKKOS_INLINE_FUNCTION
-  void operator()(const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type& team) const {
-    Kokkos::parallel_for(Kokkos::TeamThreadRange(team, C_rows), [&](const int& i) {
-      // Give each kokkos thread a vector of A
-      SubviewTypeA a_vec;
-      if (A_t)
-        a_vec = Kokkos::subview(A, Kokkos::ALL(), i);
-      else
-        a_vec = Kokkos::subview(A, i, Kokkos::ALL());
-
-      // Have all vector lanes perform the dot product
-      Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, C_cols), [&](const int& j) {
-        SubviewTypeB b_vec;
-        if (B_t)
-          b_vec = Kokkos::subview(B, j, Kokkos::ALL());
-        else
-          b_vec = Kokkos::subview(B, Kokkos::ALL(), j);
-        ScalarC ab = ScalarC(0);
-        for (int k = 0; k < A_cols; k++) {
-          auto a = A_c ? APT::conj(a_vec(k)) : a_vec(k);
-          auto b = B_c ? APT::conj(b_vec(k)) : b_vec(k);
-          ab += a * b;
-        }
-        C(i, j) = beta * C(i, j) + alpha * ab;
-      });
-    });
-  }
-};
-// C(i,:,:) = alpha * (A(i,:,:) * B(i,:,:)) + beta * C(i,:,:)
-template <class ViewTypeA, class ViewTypeB, class ViewTypeC, class ExecutionSpace>
-struct Functor_BatchedVanillaGEMM {
-  bool A_t, B_t, A_c, B_c, batch_size_last_dim = false;
-  ViewTypeA A;
-  ViewTypeB B;
-  ViewTypeC C;
-
-  using ScalarA      = typename ViewTypeA::value_type;
-  using ScalarB      = typename ViewTypeB::value_type;
-  using ScalarC      = typename ViewTypeC::value_type;
-  using SubviewTypeA = typename Kokkos::View<ScalarA**, Kokkos::LayoutStride, typename ViewTypeA::device_type>;
-  using SubviewTypeB = typename Kokkos::View<ScalarB**, Kokkos::LayoutStride, typename ViewTypeA::device_type>;
-  using SubviewTypeC = typename Kokkos::View<ScalarC**, Kokkos::LayoutStride, typename ViewTypeA::device_type>;
-
-  ScalarA alpha;
-  ScalarC beta;
-
-  KOKKOS_INLINE_FUNCTION
-  void operator()(const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type& team) const {
-    int i = team.league_rank();
-    SubviewTypeA _A;
-    SubviewTypeB _B;
-    SubviewTypeC _C;
-
-    if (batch_size_last_dim) {
-      _A = Kokkos::subview(A, Kokkos::ALL(), Kokkos::ALL(), i);
-      _B = Kokkos::subview(B, Kokkos::ALL(), Kokkos::ALL(), i);
-      _C = Kokkos::subview(C, Kokkos::ALL(), Kokkos::ALL(), i);
-    } else {
-      _A = Kokkos::subview(A, i, Kokkos::ALL(), Kokkos::ALL());
-      _B = Kokkos::subview(B, i, Kokkos::ALL(), Kokkos::ALL());
-      _C = Kokkos::subview(C, i, Kokkos::ALL(), Kokkos::ALL());
-    }
-    struct SharedVanillaGEMM<SubviewTypeA, SubviewTypeB, SubviewTypeC, ExecutionSpace> vgemm;
-    vgemm.A_t    = A_t;
-    vgemm.B_t    = B_t;
-    vgemm.A_c    = A_c;
-    vgemm.B_c    = B_c;
-    vgemm.C_rows = batch_size_last_dim ? C.extent(0) : C.extent(1);
-    vgemm.C_cols = batch_size_last_dim ? C.extent(1) : C.extent(2);
-    vgemm.A_cols = batch_size_last_dim ? (A_t ? A.extent(0) : A.extent(1)) : (A_t ? A.extent(1) : A.extent(2));
-    vgemm.A      = _A;
-    vgemm.B      = _B;
-    vgemm.C      = _C;
-    vgemm.alpha  = alpha;
-    vgemm.beta   = beta;
-    vgemm(team);
-  }
-
-  inline void run() {
-    Kokkos::parallel_for(
-        "Test::VanillaGEMM",
-        Kokkos::TeamPolicy<ExecutionSpace>(batch_size_last_dim ? C.extent(2) : C.extent(0), Kokkos::AUTO,
-                                           KokkosKernels::Impl::kk_get_max_vector_size<ExecutionSpace>()),
-        *this);
-  }
-};
-
-// Compute C := alpha * AB + beta * C
-template <class ViewTypeA, class ViewTypeB, class ViewTypeC>
-void vanillaGEMM(typename ViewTypeC::non_const_value_type alpha, const ViewTypeA& A, const ViewTypeB& B,
-                 typename ViewTypeC::non_const_value_type beta, const ViewTypeC& C) {
-  using value_type = typename ViewTypeC::non_const_value_type;
-  using KAT        = Kokkos::ArithTraits<value_type>;
-  int m            = A.extent(0);
-  int k            = A.extent(1);
-  int n            = B.extent(1);
-  for (int i = 0; i < m; i++) {
-    for (int j = 0; j < n; j++) {
-      value_type sum = KAT::zero();
-      for (int ii = 0; ii < k; ii++) {
-        sum += A(i, ii) * B(ii, j);
-      }
-      C(i, j) = alpha * sum + beta * C(i, j);
-    }
-  }
-}
-
-template <class AlphaType, class ViewTypeA, class ViewTypeX, class BetaType, class ViewTypeY>
-KOKKOS_INLINE_FUNCTION void vanillaGEMV(char mode, AlphaType alpha, const ViewTypeA& A, const ViewTypeX& x,
-                                        BetaType beta, const ViewTypeY& y) {
-  using ScalarY         = typename ViewTypeY::non_const_value_type;
-  using KAT_A           = Kokkos::ArithTraits<typename ViewTypeA::non_const_value_type>;
-  const bool transposed = mode == 'T' || mode == 'C';
-  const bool conjugated = mode == 'C';
-  const bool has_beta   = beta != Kokkos::ArithTraits<BetaType>::zero();
-  int M                 = A.extent(transposed ? 1 : 0);
-  int N                 = A.extent(transposed ? 0 : 1);
-  for (int i = 0; i < M; i++) {
-    ScalarY y_i{};
-    if (has_beta) y_i = beta * y(i);
-    for (int j = 0; j < N; j++) {
-      const auto a   = transposed ? A(j, i) : A(i, j);
-      const auto Aij = conjugated ? KAT_A::conj(a) : a;
-      y_i += alpha * Aij * x(j);
-    }
-    y(i) = y_i;
-  }
-}
-
 template <class T>
 class epsilon {
  public:
@@ -367,64 +220,6 @@ class epsilon {
 
 using KokkosKernels::Impl::getRandomBounds;
 
-template <typename scalar_t, typename lno_t, typename size_type, typename device, typename crsMat_t>
-crsMat_t symmetrize(crsMat_t A) {
-  typedef typename crsMat_t::StaticCrsGraphType graph_t;
-  typedef typename crsMat_t::values_type::non_const_type scalar_view_t;
-  typedef typename graph_t::row_map_type::non_const_type lno_view_t;
-  typedef typename graph_t::entries_type::non_const_type lno_nnz_view_t;
-  auto host_rowmap  = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A.graph.row_map);
-  auto host_entries = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A.graph.entries);
-  auto host_values  = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A.values);
-  lno_t numRows     = A.numRows();
-  // symmetrize as input_mat + input_mat^T, to still have a diagonally dominant
-  // matrix
-  typedef std::map<lno_t, scalar_t> Row;
-  std::vector<Row> symRows(numRows);
-  for (lno_t r = 0; r < numRows; r++) {
-    auto& row = symRows[r];
-    for (size_type i = host_rowmap(r); i < host_rowmap(r + 1); i++) {
-      lno_t c   = host_entries(i);
-      auto& col = symRows[c];
-      auto it   = row.find(c);
-      if (it == row.end())
-        row[c] = host_values(i);
-      else
-        row[c] += host_values(i);
-      it = col.find(r);
-      if (it == col.end())
-        col[r] = host_values(i);
-      else
-        col[r] += host_values(i);
-    }
-  }
-  // Count entries
-  Kokkos::View<size_type*, Kokkos::LayoutLeft, Kokkos::HostSpace> new_host_rowmap("Rowmap", numRows + 1);
-  size_t accum = 0;
-  for (lno_t r = 0; r <= numRows; r++) {
-    new_host_rowmap(r) = accum;
-    if (r < numRows) accum += symRows[r].size();
-  }
-  // Allocate new entries/values
-  Kokkos::View<lno_t*, Kokkos::LayoutLeft, Kokkos::HostSpace> new_host_entries("Entries", accum);
-  Kokkos::View<scalar_t*, Kokkos::LayoutLeft, Kokkos::HostSpace> new_host_values("Values", accum);
-  for (lno_t r = 0; r < numRows; r++) {
-    auto rowIt = symRows[r].begin();
-    for (size_type i = new_host_rowmap(r); i < new_host_rowmap(r + 1); i++) {
-      new_host_entries(i) = rowIt->first;
-      new_host_values(i)  = rowIt->second;
-      rowIt++;
-    }
-  }
-  lno_view_t new_rowmap("Rowmap", numRows + 1);
-  lno_nnz_view_t new_entries("Entries", accum);
-  scalar_view_t new_values("Values", accum);
-  Kokkos::deep_copy(new_rowmap, new_host_rowmap);
-  Kokkos::deep_copy(new_entries, new_host_entries);
-  Kokkos::deep_copy(new_values, new_host_values);
-  return crsMat_t("SymA", numRows, numRows, accum, new_values, new_rowmap, new_entries);
-}
-
 // create_random_x_vector and create_random_y_vector can be used together to
 // generate a random linear system Ax = y.
 template <typename vec_t>
@@ -485,17 +280,6 @@ std::string value_type_name<Kokkos::complex<double>>() {
   return "::ComplexDouble";
 }
 
-int string_compare_no_case(const char* str1, const char* str2) {
-  std::string str1_s(str1);
-  std::string str2_s(str2);
-  for (size_t i = 0; i < str1_s.size(); i++) str1_s[i] = std::tolower(str1_s[i]);
-  for (size_t i = 0; i < str2_s.size(); i++) str2_s[i] = std::tolower(str2_s[i]);
-  return strcmp(str1_s.c_str(), str2_s.c_str());
-}
-
-int string_compare_no_case(const std::string& str1, const std::string& str2) {
-  return string_compare_no_case(str1.c_str(), str2.c_str());
-}
 /// /brief Coo matrix class for testing purposes.
 /// \tparam ScalarType
 /// \tparam LayoutType
diff --git a/test_common/KokkosKernels_TestVanilla.hpp b/test_common/KokkosKernels_TestVanilla.hpp
new file mode 100644
index 0000000000..43ab7a3cf0
--- /dev/null
+++ b/test_common/KokkosKernels_TestVanilla.hpp
@@ -0,0 +1,177 @@
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 4.0
+//       Copyright (2022) National Technology & Engineering
+//               Solutions of Sandia, LLC (NTESS).
+//
+// Under the terms of Contract DE-NA0003525 with NTESS,
+// the U.S. Government retains certain rights in this software.
+//
+// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions.
+// See https://kokkos.org/LICENSE for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//@HEADER
+
+#ifndef KOKKOSKERNELS_TESTVANILLA_HPP
+#define KOKKOSKERNELS_TESTVANILLA_HPP
+
+#include <random>
+
+#include "KokkosKernels_Utils.hpp"
+#include "KokkosKernels_IOUtils.hpp"
+#include "Kokkos_ArithTraits.hpp"
+#include "KokkosBatched_Vector.hpp"
+
+namespace Test {
+
+template <class ViewTypeA, class ViewTypeB, class ViewTypeC, class ExecutionSpace>
+struct SharedVanillaGEMM {
+  bool A_t, B_t, A_c, B_c;
+  int C_rows, C_cols, A_cols;
+  ViewTypeA A;
+  ViewTypeB B;
+  ViewTypeC C;
+
+  typedef typename ViewTypeA::value_type ScalarA;
+  typedef typename ViewTypeB::value_type ScalarB;
+  typedef typename ViewTypeC::value_type ScalarC;
+  typedef Kokkos::View<ScalarA*, Kokkos::LayoutStride, typename ViewTypeA::device_type> SubviewTypeA;
+  typedef Kokkos::View<ScalarB*, Kokkos::LayoutStride, typename ViewTypeB::device_type> SubviewTypeB;
+  typedef Kokkos::ArithTraits<ScalarC> APT;
+  typedef typename APT::mag_type mag_type;
+  ScalarA alpha;
+  ScalarC beta;
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()(const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type& team) const {
+    Kokkos::parallel_for(Kokkos::TeamThreadRange(team, C_rows), [&](const int& i) {
+      // Give each kokkos thread a vector of A
+      SubviewTypeA a_vec;
+      if (A_t)
+        a_vec = Kokkos::subview(A, Kokkos::ALL(), i);
+      else
+        a_vec = Kokkos::subview(A, i, Kokkos::ALL());
+
+      // Have all vector lanes perform the dot product
+      Kokkos::parallel_for(Kokkos::ThreadVectorRange(team, C_cols), [&](const int& j) {
+        SubviewTypeB b_vec;
+        if (B_t)
+          b_vec = Kokkos::subview(B, j, Kokkos::ALL());
+        else
+          b_vec = Kokkos::subview(B, Kokkos::ALL(), j);
+        ScalarC ab = ScalarC(0);
+        for (int k = 0; k < A_cols; k++) {
+          auto a = A_c ? APT::conj(a_vec(k)) : a_vec(k);
+          auto b = B_c ? APT::conj(b_vec(k)) : b_vec(k);
+          ab += a * b;
+        }
+        C(i, j) = beta * C(i, j) + alpha * ab;
+      });
+    });
+  }
+};
+// C(i,:,:) = alpha * (A(i,:,:) * B(i,:,:)) + beta * C(i,:,:)
+template <class ViewTypeA, class ViewTypeB, class ViewTypeC, class ExecutionSpace>
+struct Functor_BatchedVanillaGEMM {
+  bool A_t, B_t, A_c, B_c, batch_size_last_dim = false;
+  ViewTypeA A;
+  ViewTypeB B;
+  ViewTypeC C;
+
+  using ScalarA      = typename ViewTypeA::value_type;
+  using ScalarB      = typename ViewTypeB::value_type;
+  using ScalarC      = typename ViewTypeC::value_type;
+  using SubviewTypeA = typename Kokkos::View<ScalarA**, Kokkos::LayoutStride, typename ViewTypeA::device_type>;
+  using SubviewTypeB = typename Kokkos::View<ScalarB**, Kokkos::LayoutStride, typename ViewTypeA::device_type>;
+  using SubviewTypeC = typename Kokkos::View<ScalarC**, Kokkos::LayoutStride, typename ViewTypeA::device_type>;
+
+  ScalarA alpha;
+  ScalarC beta;
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()(const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type& team) const {
+    int i = team.league_rank();
+    SubviewTypeA _A;
+    SubviewTypeB _B;
+    SubviewTypeC _C;
+
+    if (batch_size_last_dim) {
+      _A = Kokkos::subview(A, Kokkos::ALL(), Kokkos::ALL(), i);
+      _B = Kokkos::subview(B, Kokkos::ALL(), Kokkos::ALL(), i);
+      _C = Kokkos::subview(C, Kokkos::ALL(), Kokkos::ALL(), i);
+    } else {
+      _A = Kokkos::subview(A, i, Kokkos::ALL(), Kokkos::ALL());
+      _B = Kokkos::subview(B, i, Kokkos::ALL(), Kokkos::ALL());
+      _C = Kokkos::subview(C, i, Kokkos::ALL(), Kokkos::ALL());
+    }
+    struct SharedVanillaGEMM<SubviewTypeA, SubviewTypeB, SubviewTypeC, ExecutionSpace> vgemm;
+    vgemm.A_t    = A_t;
+    vgemm.B_t    = B_t;
+    vgemm.A_c    = A_c;
+    vgemm.B_c    = B_c;
+    vgemm.C_rows = batch_size_last_dim ? C.extent(0) : C.extent(1);
+    vgemm.C_cols = batch_size_last_dim ? C.extent(1) : C.extent(2);
+    vgemm.A_cols = batch_size_last_dim ? (A_t ? A.extent(0) : A.extent(1)) : (A_t ? A.extent(1) : A.extent(2));
+    vgemm.A      = _A;
+    vgemm.B      = _B;
+    vgemm.C      = _C;
+    vgemm.alpha  = alpha;
+    vgemm.beta   = beta;
+    vgemm(team);
+  }
+
+  inline void run() {
+    Kokkos::parallel_for(
+        "Test::VanillaGEMM",
+        Kokkos::TeamPolicy<ExecutionSpace>(batch_size_last_dim ? C.extent(2) : C.extent(0), Kokkos::AUTO,
+                                           KokkosKernels::Impl::kk_get_max_vector_size<ExecutionSpace>()),
+        *this);
+  }
+};
+
+// Compute C := alpha * AB + beta * C
+template <class ViewTypeA, class ViewTypeB, class ViewTypeC>
+void vanillaGEMM(typename ViewTypeC::non_const_value_type alpha, const ViewTypeA& A, const ViewTypeB& B,
+                 typename ViewTypeC::non_const_value_type beta, const ViewTypeC& C) {
+  using value_type = typename ViewTypeC::non_const_value_type;
+  using KAT        = Kokkos::ArithTraits<value_type>;
+  int m            = A.extent(0);
+  int k            = A.extent(1);
+  int n            = B.extent(1);
+  for (int i = 0; i < m; i++) {
+    for (int j = 0; j < n; j++) {
+      value_type sum = KAT::zero();
+      for (int ii = 0; ii < k; ii++) {
+        sum += A(i, ii) * B(ii, j);
+      }
+      C(i, j) = alpha * sum + beta * C(i, j);
+    }
+  }
+}
+
+template <class AlphaType, class ViewTypeA, class ViewTypeX, class BetaType, class ViewTypeY>
+KOKKOS_INLINE_FUNCTION void vanillaGEMV(char mode, AlphaType alpha, const ViewTypeA& A, const ViewTypeX& x,
+                                        BetaType beta, const ViewTypeY& y) {
+  using ScalarY         = typename ViewTypeY::non_const_value_type;
+  using KAT_A           = Kokkos::ArithTraits<typename ViewTypeA::non_const_value_type>;
+  const bool transposed = mode == 'T' || mode == 'C';
+  const bool conjugated = mode == 'C';
+  const bool has_beta   = beta != Kokkos::ArithTraits<BetaType>::zero();
+  int M                 = A.extent(transposed ? 1 : 0);
+  int N                 = A.extent(transposed ? 0 : 1);
+  for (int i = 0; i < M; i++) {
+    ScalarY y_i{};
+    if (has_beta) y_i = beta * y(i);
+    for (int j = 0; j < N; j++) {
+      const auto a   = transposed ? A(j, i) : A(i, j);
+      const auto Aij = conjugated ? KAT_A::conj(a) : a;
+      y_i += alpha * Aij * x(j);
+    }
+    y(i) = y_i;
+  }
+}
+
+}  // namespace Test
+#endif