Philippe 22/01/2024: OpenACC: bl89: better performance (with NVHPC 23.11)

src/MNH/bl89.f90

@@ -334,12 +334,13 @@ ZUSRBL89 = 1./ZBL89EXP
 !
 IF( KRR > 0 ) THEN
 !$acc kernels
-  ZSUM(:,:) = 0.
-  DO JRR=1,KRR
-    ZSUM(:,:) = ZSUM(:,:)+ZRM(:,:,JRR)
+  ZSUM(:,:) = ZRM(:,:,1)
+  DO JRR = 2, KRR
+    !$mnh_do_concurrent ( JI=1:IIU,JJ=1:IJU )
+    ZSUM(JI,JJ) = ZSUM(JI,JJ) + ZRM(JI,JJ,JRR)
+    !$mnh_end_do()
   ENDDO
-  ZVPT(:,:)=ZTHM(:,:) * ( 1. + ZRVORD*ZRM(:,:,1) )  &
-                           / ( 1. + ZSUM(:,:) )
+  ZVPT(:,:)=ZTHM(:,:) * ( 1. + ZRVORD*ZRM(:,:,1) ) / ( 1. + ZSUM(:,:) )
 !$acc end kernels
 ELSE
 !$acc kernels
@@ -379,11 +380,13 @@ end do
 ZHLVPT(:,IKTB:IKTE)= 0.5 * ( ZVPT(:,IKTB:IKTE)+ZVPT(:,IKTB-KKL:IKTE-KKL) )
 ZHLVPT(:,KKU)= 0.5 * ( ZVPT(:,KKU)+ZVPT(:,KKU-KKL) )
 ZHLVPT(:,KKA)    =         ZVPT(:,KKA)
+!$acc end kernels
 !-------------------------------------------------------------------------------
 !
 !*       3.  loop on model levels
 !            --------------------
 DO JK=IKTB,IKTE
+!$acc kernels
 !
 !-------------------------------------------------------------------------------
 !
@@ -394,11 +397,11 @@ DO JK=IKTB,IKTE
   ZINTE(:)=ZTKEM(:,JK)
   ZLWORK(:)=0.
   ZTESTM=1.
+!$acc loop seq
   DO JKK=JK,IKB,-KKL  
     IF(ZTESTM > 0.) THEN
       ZTESTM=0.
-      DO J1D=1,IIU*IJU
-        
+      !$mnh_do_concurrent( J1D = 1 : IIU*IJU )
        ZTEST0=0.5+SIGN(0.5,ZINTE(J1D))
         
         !--------- SHEAR + STABILITY -----------
@@ -421,7 +424,7 @@ DO JK=IKTB,IKTE
         
 
         !--------- SHEAR + STABILITY ----------- 
-    ZLWORK2 = (ZG_O_THVREF(J1D,JK) *(ZVPT(J1D,JKK) - ZVPT(J1D,JK))  & 
+        ZLWORK2 = (ZG_O_THVREF(J1D,JK) *(ZVPT(J1D,JKK) - ZVPT(J1D,JK))  &
           -XRM17*ZSHEAR(J1D,JKK)*ZSQRT_TKE(J1D,JK) &
 #ifndef MNH_BITREP
           + SQRT(ABS( (XRM17*ZSHEAR(J1D,JKK)*ZSQRT_TKE(J1D,JK) &
@@ -436,8 +439,7 @@ DO JK=IKTB,IKTE
 
         ZLWORK(J1D)=ZLWORK(J1D)+ZTEST0*(ZTEST*ZLWORK1+(1-ZTEST)*ZLWORK2)
         ZINTE(J1D) = ZINTE(J1D) - ZPOTE
-
-      END DO
+      !$mnh_end_do()
     ENDIF
   END DO
 !-------------------------------------------------------------------------------
@@ -456,10 +458,11 @@ DO JK=IKTB,IKTE
   ZLWORK(:)=0.
   ZTESTM=1.
 !
+!$acc loop seq
   DO JKK=JK+KKL,IKE,KKL 
     IF(ZTESTM > 0.) THEN
       ZTESTM=0.
-      DO J1D=1,IIU*IJU
+      !$mnh_do_concurrent( J1D = 1 : IIU*IJU )
         ZTEST0=0.5+SIGN(0.5,ZINTE(J1D))
 
         !-------- ORIGINAL -------------        
@@ -508,7 +511,7 @@ DO JK=IKTB,IKTE
 
         ZLWORK(J1D)=ZLWORK(J1D)+ZTEST0*(ZTEST*ZLWORK1+(1-ZTEST)*ZLWORK2)
         ZINTE(J1D) = ZINTE(J1D) - ZPOTE
-      END DO 
+      !$mnh_end_do()
     ENDIF
   END DO
 !
@@ -516,7 +519,7 @@ DO JK=IKTB,IKTE
 !
 !*       7.  final mixing length
 !
-  DO J1D=1,IIU*IJU
+  !$mnh_do_concurrent( J1D = 1 : IIU*IJU )
     ZLWORK1=MAX(ZLMDN(J1D,JK),1.E-10_MNHREAL)
     ZLWORK2=MAX(ZLWORK(J1D),1.E-10_MNHREAL)
     ZPOTE = ZLWORK1 / ZLWORK2
@@ -527,7 +530,7 @@ DO JK=IKTB,IKTE
     ZLWORK2=1.d0 + br_pow( ZPOTE, ZBL89EXP )
     ZLM(J1D,JK) = ZLWORK1 * br_pow( 2. / ZLWORK2, ZUSRBL89 )
 #endif
-  END DO
+  !$mnh_end_do()
 
   ZLM(:,JK)=MAX(ZLM(:,JK),XLINI)
 !
@@ -535,6 +538,7 @@ DO JK=IKTB,IKTE
 !*       8.  end of the loop on the vertical levels
 !            --------------------------------------
 !
+!$acc end kernels
 END DO
 !
 !-------------------------------------------------------------------------------
@@ -542,6 +546,7 @@ END DO
 !*       9.  boundaries
 !            ----------
 !
+!$acc kernels
 ZLM(:,KKA)=ZLM(:,IKB)
 ZLM(:,IKE)=ZLM(:,IKE-KKL)
 ZLM(:,KKU)=ZLM(:,IKE-KKL)