ACC: use data regions rather than hard-coded parallel gang loops

ecmwf-ifs · Oct 7, 2024 · dc27eef · dc27eef
1 parent e5e34b1
commit dc27eef
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 6 deletions.
diff --git a/src/ecwam/cireduce_loki_gpu.F90 b/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
 

diff --git a/src/ecwam/outbs_loki_gpu.F90 b/src/ecwam/outbs_loki_gpu.F90
@@ -95,7 +95,8 @@ SUBROUTINE OUTBS_LOKI_GPU (MIJ, FL1, XLLWS,                            &
 !$loki update_device
 
       CALL GSTATS(1502,0)
-!$acc parallel loop gang default(present) copyin(NPROMA_WAM) copyout(BOUT) vector_length(NPROMA_WAM)
+!$acc data present(MIJ,WVPRPT,WVENVI,INTFLDS,FF_NOW,NEMO2WAM) copyout(BOUT)
+
       DO ICHNK = 1, NCHNK
         CALL OUTBLOCK(1, NPROMA_WAM, MIJ(:,ICHNK),                        &
      &                FL1(:,:,:,ICHNK), XLLWS(:,:,:,ICHNK),               &
@@ -117,7 +118,8 @@ SUBROUTINE OUTBS_LOKI_GPU (MIJ, FL1, XLLWS,                            &
      &                NEMO2WAM%NEMOVCUR(:, ICHNK), &
      &                BOUT(:,:,ICHNK))
       ENDDO
-!$acc end parallel loop
+
+!$acc end data
       CALL GSTATS(1502,1)
 
 !     PRINT OUT NORMS

diff --git a/src/ecwam/wamintgr_loki_gpu.F90 b/src/ecwam/wamintgr_loki_gpu.F90
@@ -155,7 +155,8 @@ SUBROUTINE WAMINTGR_LOKI_GPU(CDTPRA, CDATE, CDATEWH, CDTIMP, CDTIMPNEXT,  &
       IF (LHOOK) CALL DR_HOOK('IMPLSCH',0,ZHOOK_HANDLE_IMPLSCH)
       TIME0=-WAM_USER_CLOCK()
 
-!$acc parallel loop gang vector_length(NPROMA_WAM) default(present) copyin(NPROMA_WAM)
+!$acc data present(VARS_4D,WVPRPT,WVENVI,FF_NOW,WAM2NEMO,INTFLDS,MIJ) 
+
       DO ICHNK=1,NCHNK
          CALL IMPLSCH (1, NPROMA_WAM, VARS_4D%FL1(:,:,:,ICHNK),   &
  &                     WVPRPT%WAVNUM(:,:,ICHNK), WVPRPT%CGROUP(:,:,ICHNK), WVPRPT%CIWA(:,:,ICHNK), &
@@ -179,7 +180,8 @@ SUBROUTINE WAMINTGR_LOKI_GPU(CDTPRA, CDATE, CDATEWH, CDTIMP, CDTIMPNEXT,  &
  &                     INTFLDS%PHIEPS(:,ICHNK), INTFLDS%PHIAW(:,ICHNK), &
  &                     MIJ%PTR(:,ICHNK), VARS_4D%XLLWS(:,:,:,ICHNK) )
       END DO
-!$acc end parallel loop
+
+!$acc end data
 
       TIME_PHYS = TIME_PHYS + (TIME0+WAM_USER_CLOCK())*1.E-06
       IF (LHOOK) CALL DR_HOOK('IMPLSCH',1,ZHOOK_HANDLE_IMPLSCH)