Apply "IFS-2867 post receives before sends" to GPU code path

src/trans/gpu/internal/sutrle_mod.F90

@@ -52,9 +52,9 @@ SUBROUTINE SUTRLE(PNM,KGL,KLOOP)
 !     ------------------------------------------------------------------
 
 
-USE PARKIND1        ,ONLY : JPRD, JPIM
-USE MPL_MODULE      ,ONLY : MPL_ALLREDUCE, MPL_RECV, MPL_SEND, MPL_BARRIER, MPL_WAIT, &
- &                          JP_NON_BLOCKING_STANDARD
+USE PARKIND1  ,ONLY : JPRD, JPIM
+USE MPL_MODULE  ,ONLY : MPL_ALLREDUCE, MPL_RECV, MPL_SEND, MPL_BARRIER, MPL_WAIT, &
+     &                  JP_NON_BLOCKING_STANDARD
 
 USE TPM_GEN         ,ONLY : NOUT
 USE TPM_DIM         ,ONLY : R
@@ -128,21 +128,6 @@ SUBROUTINE SUTRLE(PNM,KGL,KLOOP)
   CALL GSTATS(1141,1)
 ENDIF
 
-ISREQ = 0
-DO JROC=1,NPRTRV-1
-  ISEND = MYSETV-JROC
-  IF (ISEND <= 0)     ISEND = ISEND+NPRTRV
-  ISENDSET = ISEND
-  CALL SET2PE(ISEND,0,0,MYSETW,ISENDSET)
-  ISREQ = ISREQ+1
-  CALL GSTATS(801,0)
-  CALL MPL_SEND(ZSNDBUFV(1:ISENDSIZE),KDEST=NPRCIDS(ISEND), &
-   &KMP_TYPE=JP_NON_BLOCKING_STANDARD,KREQUEST=ISENDREQ(ISREQ),&
-   & KTAG=ITAG,CDSTRING='SUTRLE:')
-  CALL GSTATS(801,1)
-ENDDO
-
-
 IRREQ=0
 DO JROC=1,NPRTRV-1
   IRECV = MYSETV+JROC
@@ -157,6 +142,20 @@ SUBROUTINE SUTRLE(PNM,KGL,KLOOP)
   CALL GSTATS(801,1)
 ENDDO
 
+ISREQ = 0
+DO JROC=1,NPRTRV-1
+  ISEND = MYSETV-JROC
+  IF (ISEND <= 0)     ISEND = ISEND+NPRTRV
+  ISENDSET = ISEND
+  CALL SET2PE(ISEND,0,0,MYSETW,ISENDSET)
+  ISREQ = ISREQ+1
+  CALL GSTATS(801,0)
+  CALL MPL_SEND(ZSNDBUFV(1:ISENDSIZE),KDEST=NPRCIDS(ISEND), &
+   &KMP_TYPE=JP_NON_BLOCKING_STANDARD,KREQUEST=ISENDREQ(ISREQ),&
+   & KTAG=ITAG,CDSTRING='SUTRLE:')
+  CALL GSTATS(801,1)
+ENDDO
+
 IF(ISREQ > 0) THEN
   CALL GSTATS(801,0)
   CALL MPL_WAIT(KREQUEST=ISENDREQ(1:ISREQ), &
@@ -276,22 +275,6 @@ SUBROUTINE SUTRLE(PNM,KGL,KLOOP)
 !$OMP END PARALLEL DO
 CALL GSTATS(1141,1)
 
-ISREQ = 0
-DO JROC=1,NPRTRW-1
-  ISEND = MYSETW-JROC
-  IF (ISEND <= 0)     ISEND = ISEND+NPRTRW
-  ISENDSET = ISEND
-  CALL SET2PE(ISEND,0,0,ISENDSET,MYSETV)
-  ISENDSIZE = IPOSW(ISENDSET)
-  ISREQ = ISREQ+1
-  CALL GSTATS(801,0)
-  CALL MPL_SEND(ZSNDBUFW(1:ISENDSIZE,ISENDSET),KDEST=NPRCIDS(ISEND), &
-   & KMP_TYPE=JP_NON_BLOCKING_STANDARD,KREQUEST=ISENDREQ(ISREQ),&
-   & KTAG=ITAG,CDSTRING='SUTRLE:')
-  CALL GSTATS(801,1)
-ENDDO
-
-
 IRREQ = 0
 DO JROC=1,NPRTRW-1
 
@@ -309,6 +292,22 @@ SUBROUTINE SUTRLE(PNM,KGL,KLOOP)
   CALL GSTATS(801,1)
 ENDDO
 
+ISREQ = 0
+DO JROC=1,NPRTRW-1
+  ISEND = MYSETW-JROC
+  IF (ISEND <= 0)     ISEND = ISEND+NPRTRW
+  ISENDSET = ISEND
+  CALL SET2PE(ISEND,0,0,ISENDSET,MYSETV)
+  ISENDSIZE = IPOSW(ISENDSET)
+  ISREQ = ISREQ+1
+  CALL GSTATS(801,0)
+  CALL MPL_SEND(ZSNDBUFW(1:ISENDSIZE,ISENDSET),KDEST=NPRCIDS(ISEND), &
+   & KMP_TYPE=JP_NON_BLOCKING_STANDARD,KREQUEST=ISENDREQ(ISREQ),&
+   & KTAG=ITAG,CDSTRING='SUTRLE:')
+  CALL GSTATS(801,1)
+ENDDO
+
+
 IF(ISREQ > 0) THEN
   CALL GSTATS(801,0)
   CALL MPL_WAIT(KREQUEST=ISENDREQ(1:ISREQ), &