Changeset 953 for codes/icosagcm/trunk/src/vertical

Timestamp:

07/15/19 12:29:31 (5 years ago)

Author:

adurocher

Message:

trunk : GPU implementation with OpenACC ( merge from glcp.idris.fr )

File:

• codes/icosagcm/trunk/src/vertical/disvert.f90 (modified) (6 diffs)

codes/icosagcm/trunk/src/vertical/disvert.f90

@@ -1 +1 @@
 MODULE disvert_mod
   USE icosa
+  USE abort_mod
   REAL(rstd), SAVE, POINTER :: ap(:)
 !$OMP THREADPRIVATE(ap) 
@@ -146 +147 @@
     END IF
 
+    !$acc enter data copyin(ap(:), bp(:), mass_dak(:), mass_dbk(:)) async
+    
   END SUBROUTINE init_disvert  
   
-  SUBROUTINE compute_rhodz(comp, ps, rhodz)
+  SUBROUTINE compute_rhodz(comp, ps, rhodz, ondevice)
     USE icosa
     USE omp_para
@@ -154 +157 @@
     REAL(rstd), INTENT(IN) :: ps(iim*jjm)
     REAL(rstd), INTENT(INOUT) :: rhodz(iim*jjm,llm)
+    LOGICAL, INTENT(IN), OPTIONAL :: ondevice ! .TRUE. compute on device, .FALSE. stay on host
     REAL(rstd) :: m, err
     INTEGER :: l,i,j,ij,dd
+    LOGICAL :: ondevice_
+
     err=0.
+
+    ! by default, do not compute on device
+    ondevice_ = .FALSE.
+    IF (PRESENT(ondevice)) ondevice_ = ondevice
 
     IF(comp) THEN
@@ -164 +174 @@
     END IF
 
+    !$acc data present(ps(:), rhodz(:,:), ap(:), bp(:)) async if (ondevice_)
+
     IF(comp) THEN
-    DO l = ll_begin, ll_end
-       DO j=jj_begin-dd,jj_end+dd
-          DO i=ii_begin-dd,ii_end+dd
-             ij=(j-1)*iim+i
-             m = ( ap(l) - ap(l+1) + (bp(l)-bp(l+1))*ps(ij) )/g 
-                rhodz(ij,l) = m
-          ENDDO
-        ENDDO
-      ENDDO
-             ELSE
+      !$acc parallel loop collapse(3) async if (ondevice_)
       DO l = ll_begin, ll_end
         DO j=jj_begin-dd,jj_end+dd
@@ -180 +183 @@
             ij=(j-1)*iim+i
             m = ( ap(l) - ap(l+1) + (bp(l)-bp(l+1))*ps(ij) )/g
-                err = MAX(err,abs(m-rhodz(ij,l)))
+            rhodz(ij,l) = m
           ENDDO
-       ENDDO
-    ENDDO
-
-       IF(err>1e-10) THEN
+        ENDDO
+      ENDDO
+    ELSE
+      !$acc parallel loop reduction(max:err) collapse(3) async if (ondevice_)
+      DO l = ll_begin, ll_end
+        DO j=jj_begin-dd,jj_end+dd
+          DO i=ii_begin-dd,ii_end+dd
+            ij=(j-1)*iim+i
+            m = ( ap(l) - ap(l+1) + (bp(l)-bp(l+1))*ps(ij) )/g
+            err = MAX(err,abs(m-rhodz(ij,l)))
+          ENDDO
+        ENDDO
+      ENDDO
+
+      !$acc wait
+      IF(err>1e-10) THEN
           PRINT *, 'Discrepancy between ps and rhodz detected', err
           STOP
@@ -191 +206 @@
     ENDIF
 
+    !$acc end data
   END SUBROUTINE compute_rhodz
 
- 
+
   SUBROUTINE write_apbp
   USE icosa