gpu-aware MPI improved data clauses

ecmwf-ifs · Aug 24, 2023 · ab807d2 · ab807d2
1 parent 2aac87e
commit ab807d2
Show file tree

Hide file tree

Showing 4 changed files with 31 additions and 34 deletions.
diff --git a/src/trans/gpu/internal/trgtol_mod.F90 b/src/trans/gpu/internal/trgtol_mod.F90
@@ -453,10 +453,10 @@ SUBROUTINE TRGTOL_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
   ZCOMBUFS(:,:) = 0.
   ZCOMBUFR(:,:) = 0.
 #ifdef ACCGPU
-  !$ACC END KERNELS
+    !$ACC END KERNELS
 #endif
 #ifdef OMPGPU
-  !$OMP END TARGET
+    !$OMP END TARGET
 #endif
 
   CALL GSTATS(1805,1)
@@ -553,9 +553,10 @@ SUBROUTINE TRGTOL_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
     !$ACC&        IGP2PARS,IGP3ALEVS,IGP3APARS)
 #elif defined(_CRAYFTN)
     !$ACC PARALLEL LOOP COLLAPSE(3) DEFAULT(NONE) PRIVATE(IPOS,IFIRST,ILAST,IFLD,JK) &
-    !$ACC& PRESENT(NGPBLKS,IFLDS,JK_MAX,IGPTRSEND,MYSETW,INDOFF,MYPROC,IGPTROFF,IFLDOFF, &
-    !$ACC&        LLUV,LLGP2,LLGP3A,LLGP3B,PGLAT,KINDEX,PGPUV,PGP2,PGP3A,PGP3B,IUVLEVS, &
-    !$ACC&        IUVPARS,IGP2PARS,IGP3ALEVS,IGP3APARS,IGP3BLEVS,IGP3BPARS)
+    !$ACC  COPYIN(NGPBLKS,IFLDS,JK_MAX,MYSETW,MYPROC) &
+    !$ACC& PRESENT(IGPTRSEND,INDOFF,IGPTROFF,IFLDOFF, &
+    !$ACC&        LLUV,LLGP2,LLGP3A,LLGP3B,PGLAT,KINDEX,PGPUV,PGP2,PGP3A,PGP3B, &
+    !$ACC&        IGP2PARS,IGP3ALEVS,IGP3APARS) !!,IUVLEVS,IUVPARS,IGP3BLEVS,IGP3BPARS)
 #endif
 #endif
     DO JBLK=1,NGPBLKS
@@ -637,10 +638,10 @@ SUBROUTINE TRGTOL_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
       !$ACC& PRESENT(IGPTRSEND,LLUV,LLGP2,LLGP3A,LLGP3B,ZCOMBUFS,KPTRGP,PGP,PGPUV,PGP2,PGP3A,PGP3B, &
       !$ACC&        IGP2PARS,IGP3ALEVS,IGP3APARS)
 #elif defined(_CRAYFTN)
-      !$ACC PARALLEL LOOP COLLAPSE(3) DEFAULT(NONE) PRIVATE(JK,JI,IFLDT,IFIRST,ILAST) COPYIN(INS,JK_MAX,IJPOS,IFLDA) &
-      !$ACC& PRESENT(ISEND_FLD_END,NGPBLKS,ISETW,IPOS,IGPTRSEND,LLINDER,LLPGPONLY, &
-      !$ACC&        LLUV,LLGP2,LLGP3A,LLGP3B,ZCOMBUFS,KPTRGP,PGP,PGPUV,PGP2,PGP3A,PGP3B,IUVLEVS, &
-      !$ACC&        IUVPARS,IGP2PARS,IGP3ALEVS,IGP3APARS,IGP3BLEVS,IGP3BPARS)
+      !$ACC PARALLEL LOOP COLLAPSE(3) DEFAULT(NONE) PRIVATE(JK,JI,IFLDT,IFIRST,ILAST) &
+      !$ACC  COPYIN(NGPBLKS,ISEND_FLD_END,INS,JK_MAX,IJPOS,IFLDA,IPOS,ISETW,LLINDER,LLPGPONLY) &
+      !$ACC& PRESENT(IGPTRSEND,LLUV,LLGP2,LLGP3A,LLGP3B,ZCOMBUFS,KPTRGP,PGP,PGPUV,PGP2,PGP3A,PGP3B, &
+      !$ACC&        IGP2PARS,IGP3ALEVS,IGP3APARS) !!,IUVLEVS,IUVPARS,IGP3BLEVS,IGP3BPARS)
 #endif
 #endif
       DO JJ=1,ISEND_FLD_END

diff --git a/src/trans/gpu/internal/trltog_mod.F90 b/src/trans/gpu/internal/trltog_mod.F90
@@ -444,10 +444,10 @@ SUBROUTINE TRLTOG_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
   ZCOMBUFS(:,:) = 0.
   ZCOMBUFR(:,:) = 0.
 #ifdef ACCGPU
-  !$ACC END KERNELS
+    !$ACC END KERNELS
 #endif
 #ifdef OMPGPU
-  !$OMP END TARGET
+    !$OMP END TARGET
 #endif
 
 #ifdef ACCGPU
@@ -506,7 +506,7 @@ SUBROUTINE TRLTOG_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
       !$OMP TARGET TEAMS DISTRIBUTE PARALLEL DO DEFAULT(NONE) PRIVATE(IPOS,IFIRST,ILAST,IFLD,JK) &
       !$OMP&     MAP(TO:IGPTROFF,IFLDOFF,JK_MAX) COLLAPSE(3) &
       !$OMP&     SHARED(NGPBLKS,IFLDS,IGPTRSEND,MYSETW,INDOFF,MYPROC, &
-      !$OMP&            IGPTROFF,LLINDER,KPTRGP,PGP,PGLAT,KINDEX,IFLDOFF,JK_MAX)
+      !$OMP&            LLINDER,KPTRGP,PGP,PGLAT,KINDEX)
 #endif
 #ifdef ACCGPU
       !$ACC PARALLEL LOOP DEFAULT(NONE) PRIVATE(IPOS,IFIRST,ILAST,IFLD,JK) &
@@ -550,11 +550,10 @@ SUBROUTINE TRLTOG_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
       !$ACC&            PGLAT,LLGP3A,PGP3A,LLGP3B,PGP3B,KINDEX)
 #elif defined(_CRAYFTN)
       !$ACC PARALLEL LOOP DEFAULT(NONE) PRIVATE(IPOS,IFIRST,ILAST,IFLD,JK) &
-      !$ACC&     COPYIN(IGPTROFF,IFLDOFF,JK_MAX) COLLAPSE(3) &
-      !$ACC&     PRESENT(NGPBLKS,IFLDS,IGPTRSEND,MYSETW,INDOFF,MYPROC, &
-      !$ACC&            LLUV,PGPUV,IUVLEVS,IUVPARS,LLGP2,PGP2,IGP2PARS, &
-      !$ACC&            PGLAT,LLGP3A,PGP3A,IGP3ALEVS,IGP3APARS,LLGP3B,PGP3B,&
-      !$ACC&            IGP3BLEVS,IGP3BPARS,KINDEX)
+      !$ACC&     COPYIN(NGPBLKS,IGPTROFF,IFLDOFF,JK_MAX,IFLDS,MYPROC,MYSETW) COLLAPSE(3) &
+      !$ACC&     PRESENT(IGPTRSEND,INDOFF,LLUV,PGPUV,LLGP2,PGP2,IGP2PARS, &
+      !$ACC&            PGLAT,LLGP3A,PGP3A,IGP3ALEVS,IGP3APARS,LLGP3B,PGP3B,KINDEX) !!,&
+      !!!$ACC&            IUVLEVS,IUVPARS,IGP3BLEVS,IGP3BPARS)
 #endif
 #endif
       DO JBLK=1,NGPBLKS
@@ -608,8 +607,8 @@ SUBROUTINE TRLTOG_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
       !$ACC PARALLEL LOOP DEFAULT(NONE) PRIVATE(II) COPYIN(ILEN) COLLAPSE(2) &
       !$ACC&  PRESENT(KINDEX,INDOFF,ZCOMBUFS,PGLAT)
 #elif defined(_CRAYFTN)
-      !$ACC PARALLEL LOOP DEFAULT(NONE) PRIVATE(II) COPYIN(ILEN) COLLAPSE(2) &
-      !$ACC&  PRESENT(KF_FS,KINDEX,INDOFF,ISEND,ZCOMBUFS,INS,PGLAT)
+      !$ACC PARALLEL LOOP DEFAULT(NONE) PRIVATE(II) COPYIN(ILEN,INS,ISEND,KF_FS) COLLAPSE(2) &
+      !$ACC&  PRESENT(KINDEX,INDOFF,ZCOMBUFS,PGLAT)
 #endif
 #endif
       DO JL=1,ILEN
@@ -756,11 +755,11 @@ SUBROUTINE TRLTOG_CUDAAWARE(PGLAT,KF_FS,KF_GP,KF_SCALARS_G,KVSET,KPTRGP,&
       !$ACC&            LLGP2,PGP2,IGP2PARS,LLGP3A,PGP3A,LLGP3B,PGP3B)
 #elif defined(_CRAYFTN)
       !$ACC PARALLEL LOOP DEFAULT(NONE) PRIVATE(IFIRST,ILAST,JI,JK,IFLDT) &
-      !$ACC&     COPYIN(INR,KF_FS,IPOS,JPOS,IFLD,IFLDA,JK_MAX,IRECV_FLD_END) COLLAPSE(3) &
-      !$ACC&     PRESENT(NGPBLKS,IGPTRSEND,ISETW, &
-      !$ACC&            LLINDER,PGP,KPTRGP,ZCOMBUFR,LLPGPONLY,LLUV,PGPUV,IUVLEVS, &
-      !$ACC&            IUVPARS,LLGP2,PGP2,IGP2PARS,LLGP3A,PGP3A,IGP3ALEVS,IGP3APARS, &
-      !$ACC&            LLGP3B,PGP3B,IGP3BLEVS,IGP3BPARS)
+      !$ACC&     COPYIN(NGPBLKS,INR,KF_FS,IPOS,JPOS,IFLD,IFLDA,JK_MAX,IRECV_FLD_END,ISETW,LLINDER,LLPGPONLY) COLLAPSE(3) &
+      !$ACC&     PRESENT(IGPTRSEND,PGP,KPTRGP,ZCOMBUFR,LLUV,PGPUV, &
+      !$ACC&            LLGP2,PGP2,IGP2PARS,LLGP3A,PGP3A,IGP3ALEVS,IGP3APARS, &
+      !$ACC&            LLGP3B,PGP3B) !,
+      !!$ACC&           IUVLEVS,IUVPARS,IGP3BLEVS,IGP3BPARS)
 #endif
 #endif
       DO JBLK=1,NGPBLKS

diff --git a/src/trans/gpu/internal/trltom_mod.F90 b/src/trans/gpu/internal/trltom_mod.F90
@@ -166,7 +166,8 @@ END FUNCTION ALLTOALLV_CUDAIPC
         !$ACC KERNELS ASYNC(1) DEFAULT(NONE) PRESENT(PFBUF,PFBUF_IN)
 #elif defined(_CRAYFTN)
         !$ACC KERNELS ASYNC(1) DEFAULT(NONE) &
-        !$ACC& PRESENT(PFBUF,PFBUF_IN,FROM_RECV,TO_RECV,FROM_SEND,TO_SEND)
+        !$ACC& COPYIN(FROM_RECV,TO_RECV,FROM_SEND,TO_SEND) &
+        !$ACC& PRESENT(PFBUF,PFBUF_IN)
 #endif
 #endif
         PFBUF(FROM_RECV:TO_RECV) = PFBUF_IN(FROM_SEND:TO_SEND)

diff --git a/src/trans/gpu/internal/trmtol_mod.F90 b/src/trans/gpu/internal/trmtol_mod.F90
@@ -142,8 +142,8 @@ SUBROUTINE TRMTOL_CUDAAWARE(PFBUF_IN,PFBUF,KFIELD)
 #if defined(__NVCOMPILER) || defined(__PGI)
       !$ACC KERNELS ASYNC(1) DEFAULT(NONE) PRESENT(PFBUF,PFBUF_IN)
 #elif defined(_CRAYFTN)
-      !$ACC KERNELS ASYNC(1) DEFAULT(NONE) &
-      !$ACC& PRESENT(PFBUF,PFBUF_IN,FROM_RECV,TO_RECV,FROM_SEND,TO_SEND)
+      !$ACC KERNELS ASYNC(1) DEFAULT(NONE) COPYIN(FROM_RECV,TO_RECV,FROM_SEND,TO_SEND) &
+      !$ACC& PRESENT(PFBUF,PFBUF_IN)
 #endif
 #endif
       PFBUF(FROM_RECV:TO_RECV) = PFBUF_IN(FROM_SEND:TO_SEND)
@@ -159,15 +159,11 @@ SUBROUTINE TRMTOL_CUDAAWARE(PFBUF_IN,PFBUF,KFIELD)
   CALL GSTATS(410,0)
 
 #ifdef ACCGPU
-  !$ACC DATA PRESENT(PFBUF_IN, PFBUF)
+  !$ACC DATA PRESENT(PFBUF_IN,PFBUF) COPYIN(ILENS,ILENR,IOFFS,IOFFR)
+  !$ACC HOST_DATA USE_DEVICE(PFBUF_IN,PFBUF)
 #endif
 #ifdef OMPGPU
   !$OMP TARGET DATA MAP(PRESENT,ALLOC:PFBUF_IN, PFBUF)
-#endif
-#ifdef ACCGPU
-  !$ACC HOST_DATA USE_DEVICE(PFBUF_IN, PFBUF)
-#endif
-#ifdef OMPGPU
   !$OMP TARGET DATA USE_DEVICE_PTR(PFBUF_IN, PFBUF)
 #endif
   CALL GSTATS(807,0)
@@ -202,7 +198,7 @@ SUBROUTINE TRMTOL_CUDAAWARE(PFBUF_IN,PFBUF,KFIELD)
   !$OMP& SHARED(ISTA,ILEN,PFBUF,PFBUF_IN)
 #endif
 #ifdef ACCGPU
-  !$ACC PARALLEL LOOP DEFAULT(NONE) PRESENT(ISTA,ILEN,PFBUF,PFBUF_IN)
+  !$ACC PARALLEL LOOP DEFAULT(NONE) COPYIN(ISTA,ILEN) PRESENT(PFBUF,PFBUF_IN)
 #endif
   DO J=ISTA,ISTA+ILEN-1
     PFBUF(J) = PFBUF_IN(J)