Run the entire advection loop on device

CMakeLists.txt

@@ -135,7 +135,17 @@ ecbuild_add_option( FEATURE ACC
                     REQUIRED_PACKAGES "OpenACC COMPONENTS Fortran"
                     CONDITION HAVE_LOKI AND NOT LOKI_MODE MATCHES "idem|idem-stack" )
 
-### CUDA-aware MPI
+### CUDA
+include(CheckLanguage)
+check_language(CUDA)
+ecbuild_add_option( FEATURE CUDA
+    DESCRIPTION "CUDA" DEFAULT OFF
+    CONDITION CMAKE_CUDA_COMPILER AND HAVE_ACC )
+if( HAVE_CUDA )
+    enable_language( CUDA )
+endif()
+
+### GPU-aware MPI
 ecbuild_add_option( FEATURE GPU_AWARE_MPI
                     DEFAULT OFF
                     DESCRIPTION "Enable GPU-aware MPI"

README.md

@@ -232,16 +232,15 @@ Building
 The recommended option for building the GPU enabled ecWAM is to use the provided bundle, and pass the
 `--with-loki --with-acc` options. Different Loki transformations can also be chosen at build-time via the following 
 bundle option: `--loki-mode=<trafo>`. Direct GPU-to-GPU MPI communications can be enabled by passing the 
-`--with-gpu-aware-mpi` option.
+`--with-gpu-aware-mpi` option. CPU to GPU data transfers can be accelerated (via pinning of host-side allocations)
+by building with the `--with-cuda` option.
 
 The ecwam-bundle also provides appropriate arch files for the nvhpc suite on the ECMWF ATOS system.
 
 Running
 -------
 No extra run-time options are needed to run the GPU enabled ecWam. Please note that this means that if ecWam is built
 using the `--with-loki` and `--with-acc` bundle arguments, it will necessarily be offloaded for GPU execution.
-For multi-GPU runs, the number of GPUs maps to the number of MPI ranks. Thus multiple GPUs can be requested by
-launching with multiple MPI ranks. The mapping of MPI ranks to GPUs assumes at most 4 GPUs per host node.
 
 Environment variables
 ---------------------

cmake/ecwam_expand_drv_types.cmake

@@ -18,7 +18,7 @@ macro( ecwam_expand_drv_types )
       list(APPEND FYPP_ARGS -DPARKIND1_SINGLE_NEMO)
    endif()
 
-   if( HAVE_ACC )
+   if( HAVE_LOKI AND NOT LOKI_MODE MATCHES "idem|idem-stack" )
       list(APPEND FYPP_ARGS -DWAM_GPU)
    endif()
 

package/bundle/bundle.yml

@@ -75,6 +75,11 @@ options :
         cmake : >
             ENABLE_ACC=ON
 
+    - with-cuda :
+        help  : Enable FIELD_API CUDA backend
+        cmake : >
+            ENABLE_CUDA=ON
+
     - without-loki-install :
         help  : Skip installation of Loki (Requires Loki to be on the PATH)
         cmake : >

src/ecwam/CMakeLists.txt

@@ -269,7 +269,6 @@ list( APPEND ecwam_srcs
   w_mode_st.F90
   w_pdf.F90
   w_pmax.F90
-  wam_init_gpu_mod.F90
   wam_multio_mod.F90
   wam_nproma.F90
   wam_sorti.F90
@@ -363,6 +362,7 @@ if(HAVE_LOKI AND NOT LOKI_MODE MATCHES "idem|idem-stack")
   list(APPEND ecwam_srcs wamintgr_loki_gpu.F90)
   list(REMOVE_ITEM ecwam_srcs wamintgr.F90)
   list(APPEND ecwam_srcs cireduce_loki_gpu.F90)
+  list(APPEND ecwam_srcs outbs_loki_gpu.F90)
 endif()
 
 # expand derived-types using src/ecwam/yowdrvtype_config.yml
@@ -445,7 +445,7 @@ ecbuild_add_library(
                      $<${HAVE_ACC}:OpenACC::OpenACC_Fortran>
     PUBLIC_INCLUDES  $<INSTALL_INTERFACE:include>
     PRIVATE_INCLUDES ${CMAKE_CURRENT_SOURCE_DIR}
-    PRIVATE_DEFINITIONS ${ECWAM_PRIVATE_DEFINITIONS}
+    PRIVATE_DEFINITIONS ${ECWAM_PRIVATE_DEFINITIONS} $<${HAVE_CUDA}:_CUDA>
     PUBLIC_DEFINITIONS ${ECWAM_DEFINITIONS}
 )
 
@@ -460,6 +460,10 @@ if( HAVE_ACC AND CMAKE_Fortran_COMPILER_ID MATCHES "PGI|NVHPC")
    target_compile_options( ${ecwam} PRIVATE "-gpu=maxregcount:128" )
 endif()
 
+if( HAVE_CUDA )
+  target_link_options( ${ecwam} PUBLIC "-cuda;-gpu=pinned" )
+endif()
+
 ecwam_target_compile_definitions_FILENAME( ${ecwam} )
 
 ### The file mubuf.F90, which is only used for "preproc" is sensitive to optimisations

src/ecwam/aki.F90

@@ -52,6 +52,7 @@ REAL(KIND=JWRB) FUNCTION AKI(OM,BETA)
 ! ----------------------------------------------------------------------
 
       IMPLICIT NONE
+!$loki routine seq
 
       REAL(KIND=JWRB), INTENT(IN) :: OM, BETA
 

src/ecwam/cal_second_order_spec.F90

@@ -52,8 +52,8 @@ SUBROUTINE CAL_SECOND_ORDER_SPEC(KIJS, KIJL, F1, WAVNUM, DEPTH, SIG)
       USE YOWPARAM, ONLY : NANG, NFRE
       USE YOWPCONS, ONLY : G, PI, ZPI
       USE YOWSHAL , ONLY : NDEPTH, DEPTHA, DEPTHD
-      USE YOWTABL , ONLY : MR, XMR, MA, XMA, NFREH, NANGH, NMAX,         &
-     &                     OMEGA, DFDTH, THH, DELTHH, IM_P, IM_M,        &
+      USE YOWTABL , ONLY : MR, XMR, MA, XMA, NFREH, NANGH,          &
+     &                     OMEGA, DFDTH, THH, DELTHH, IM_P, IM_M,   &
      &                     TA, TB, TC_QL, TT_4M, TT_4P
       USE YOWTEST , ONLY : IU06
 
@@ -66,20 +66,20 @@ SUBROUTINE CAL_SECOND_ORDER_SPEC(KIJS, KIJL, F1, WAVNUM, DEPTH, SIG)
 #include "secspom.intfb.h"
 
       INTEGER(KIND=JWIM), INTENT(IN) :: KIJS, KIJL
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL,NANG,NFRE), INTENT(INOUT) :: F1
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL,NFRE), INTENT(IN) :: WAVNUM 
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL), INTENT(IN) :: DEPTH
+      REAL(KIND=JWRB), DIMENSION(KIJL,NANG,NFRE), INTENT(INOUT) :: F1
+      REAL(KIND=JWRB), DIMENSION(KIJL,NFRE), INTENT(IN) :: WAVNUM 
+      REAL(KIND=JWRB), DIMENSION(KIJL), INTENT(IN) :: DEPTH
       REAL(KIND=JWRB), INTENT(IN) :: SIG
 
       INTEGER(KIND=JWIM) :: IJ,M,K,K0,M0,MP,KP,MM,KM,KL,KLL,ML
 
       REAL(KIND=JWRB) :: FRAC,CO1,DEL,DELF,D1,D2,D3,D4,C1
       REAL(KIND=JWRB) :: C2,XM,XK,OMSTART,AREA,SUM,SUM1,SUM3,GAM_B_J,ZFAC
       REAL(KIND=JPHOOK) :: ZHOOK_HANDLE
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL) :: EMEAN, FMEAN
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL) :: F1MEAN, AKMEAN, XKMEAN, EMAXL  
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL,NANG,NFRE) :: F3
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL,NANGH,NFREH) :: PF1, PF3
+      REAL(KIND=JWRB), DIMENSION(KIJL) :: EMEAN, FMEAN
+      REAL(KIND=JWRB), DIMENSION(KIJL) :: F1MEAN, AKMEAN, XKMEAN, EMAXL  
+      REAL(KIND=JWRB), DIMENSION(KIJL,NANG,NFRE) :: F3
+      REAL(KIND=JWRB), DIMENSION(KIJL,NANGH,NFREH) :: PF1, PF3
 
 !-----------------------------------------------------------------------
 
@@ -104,8 +104,8 @@ SUBROUTINE CAL_SECOND_ORDER_SPEC(KIJS, KIJL, F1, WAVNUM, DEPTH, SIG)
 !***     1.11 NO INTERPOLATION.
 !        ----------------------
 
-         CALL SECSPOM(F1,F3,KIJS,KIJL,NFRE,NANG,NMAX,NDEPTH,DEPTHA,       &
-     &                DEPTHD,OMSTART,FRAC,MR,DFDTH,OMEGA,DEPTH,         &
+         CALL SECSPOM(F1,F3,KIJS,KIJL,NFRE,NANG,NDEPTH,DEPTHA,       &
+     &                DEPTHD,OMSTART,FRAC,MR,DFDTH,OMEGA,DEPTH,      &
      &                AKMEAN,TA,TB,TC_QL,TT_4M,TT_4P,IM_P,IM_M)
          DO M=1,NFRE
             DO K=1,NANG
@@ -138,8 +138,8 @@ SUBROUTINE CAL_SECOND_ORDER_SPEC(KIJS, KIJL, F1, WAVNUM, DEPTH, SIG)
 !***     1.13 DETERMINE SECOND-ORDER SPEC
 !        --------------------------------
 
-         CALL SECSPOM(PF1,PF3,KIJS,KIJL,NFREH,NANGH,NMAX,NDEPTH,DEPTHA,   &
-     &                DEPTHD,OMSTART,FRAC,MR,DFDTH,OMEGA,DEPTH,           &
+         CALL SECSPOM(PF1,PF3,KIJS,KIJL,NFREH,NANGH,NDEPTH,DEPTHA,   &
+     &                DEPTHD,OMSTART,FRAC,MR,DFDTH,OMEGA,DEPTH,      &
      &                AKMEAN,TA,TB,TC_QL,TT_4M,TT_4P,IM_P,IM_M)
 
 !***     2.24 INTERPOLATE TOWARDS HIGH-RES GRID

src/ecwam/cireduce_loki_gpu.F90

@@ -99,12 +99,14 @@ SUBROUTINE CIREDUCE_LOKI_GPU (WVPRPT, FF_NOW)
 ENDIF
           CALL GSTATS(1493,0)
 !         DETERMINE THE WAVE ATTENUATION FACTOR
-!$acc parallel loop gang present(FF_NOW, WVPRPT) vector_length(NPROMA_WAM)
+!$acc data present(FF_NOW, WVPRPT)
+
           DO ICHNK = 1, NCHNK
             CALL CIWAF(1, NPROMA_WAM, WVPRPT%CGROUP(:,:,ICHNK), FF_NOW%CICOVER(:,ICHNK), &
 &                      FF_NOW%CITHICK(:,ICHNK), WVPRPT%CIWA(:,:,ICHNK))
           ENDDO
-!$acc end parallel loop
+
+!$acc end data
           CALL GSTATS(1493,1)
         ENDIF
 

src/ecwam/ctcor.F90

@@ -49,16 +49,16 @@ SUBROUTINE CTCOR (KIJS, KIJL, F, CTR)
       IMPLICIT NONE
 
       INTEGER(KIND=JWIM), INTENT(IN) :: KIJS, KIJL
-      REAL(KIND=JWRB), INTENT(IN) :: F(KIJS:KIJL,NANG,NFRE)
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL), INTENT(OUT) :: CTR 
+      REAL(KIND=JWRB), INTENT(IN) :: F(KIJL,NANG,NFRE)
+      REAL(KIND=JWRB), DIMENSION(KIJL), INTENT(OUT) :: CTR 
 
 
       INTEGER(KIND=JWIM) :: IJ, K, M
       REAL(KIND=JWRB) :: FR1M1, ZARG, ZAMP
       REAL(KIND=JPHOOK) :: ZHOOK_HANDLE
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL) :: EM, ZT1 
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL) :: ZRHO, ZLAM
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL, NFRE) :: TEMP
+      REAL(KIND=JWRB), DIMENSION(KIJL) :: EM, ZT1 
+      REAL(KIND=JWRB), DIMENSION(KIJL) :: ZRHO, ZLAM
+      REAL(KIND=JWRB), DIMENSION(KIJL, NFRE) :: TEMP
 
 ! ----------------------------------------------------------------------
 

src/ecwam/dominant_period.F90

@@ -57,14 +57,14 @@ SUBROUTINE DOMINANT_PERIOD (KIJS, KIJL, FL1, DP)
 
 
       INTEGER(KIND=JWIM), INTENT(IN) :: KIJS, KIJL
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL,NANG,NFRE), INTENT(IN) :: FL1
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL), INTENT(OUT) :: DP
+      REAL(KIND=JWRB), DIMENSION(KIJL,NANG,NFRE), INTENT(IN) :: FL1
+      REAL(KIND=JWRB), DIMENSION(KIJL), INTENT(OUT) :: DP
 
       REAL(KIND=JWRB), PARAMETER :: FLTHRS = 0.1_JWRB
 
       INTEGER(KIND=JWIM) :: IJ, K, M
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL) :: TEMP, EM, FCROP
-      REAL(KIND=JWRB), DIMENSION(KIJS:KIJL,NFRE) :: F1D4
+      REAL(KIND=JWRB), DIMENSION(KIJL) :: TEMP, EM, FCROP
+      REAL(KIND=JWRB), DIMENSION(KIJL,NFRE) :: F1D4
 
       REAL(KIND=JPHOOK) :: ZHOOK_HANDLE