Changeset 1113 for palm/trunk/SOURCE/exchange_horiz.f90

Timestamp:

Mar 10, 2013 2:48:14 AM (11 years ago)

Author:

raasch

Message:

GPU porting of boundary conditions and routine pres; index bug removec from radiation boundary condition

File:

• palm/trunk/SOURCE/exchange_horiz.f90 (modified) (3 diffs)

palm/trunk/SOURCE/exchange_horiz.f90

@@ -20 +20 @@
 ! Current revisions:
 ! -----------------
-!
+! GPU-porting for single-core (1PE) mode
 !
 ! Former revisions:
@@ -82 +82 @@
 
 
-    INTEGER ::  nbgp_local
+    INTEGER ::  i, j, k, nbgp_local
     REAL, DIMENSION(nzb:nzt+1,nys-nbgp_local:nyn+nbgp_local, &
                     nxl-nbgp_local:nxr+nbgp_local) ::  ar
@@ -199 +199 @@
 
 !
-!-- Lateral boundary conditions in the non-parallel case
+!-- Lateral boundary conditions in the non-parallel case.
+!-- Case dependent, because in GPU mode still not all arrays are on device. This
+!-- workaround has to be removed later. Also, since PGI compiler 12.5 has problems
+!-- with array syntax, explicit loops are used.
     IF ( bc_lr == 'cyclic' )  THEN
-        ar(:,:,nxl-nbgp_local:nxl-1) = ar(:,:,nxr-nbgp_local+1:nxr)
-        ar(:,:,nxr+1:nxr+nbgp_local) = ar(:,:,nxl:nxl+nbgp_local-1)
+       IF ( on_device )  THEN
+          !$acc kernels present( ar )
+          !$acc loop independent
+          DO  i = 0, nbgp_local-1
+             DO  j = nys-nbgp_local, nyn+nbgp_local
+                !$acc loop vector( 32 )
+                DO  k = nzb, nzt+1
+                   ar(k,j,nxl-nbgp_local+i) = ar(k,j,nxr-nbgp_local+1+i)
+                   ar(k,j,nxr+1+i)          = ar(k,j,nxl+i)
+                ENDDO
+             ENDDO
+          ENDDO
+          !$acc end kernels
+       ELSE
+          ar(:,:,nxl-nbgp_local:nxl-1) = ar(:,:,nxr-nbgp_local+1:nxr)
+          ar(:,:,nxr+1:nxr+nbgp_local) = ar(:,:,nxl:nxl+nbgp_local-1)
+       ENDIF
     ENDIF
 
     IF ( bc_ns == 'cyclic' )  THEN
-        ar(:,nys-nbgp_local:nys-1,:) = ar(:,nyn-nbgp_local+1:nyn,:)
-        ar(:,nyn+1:nyn+nbgp_local,:) = ar(:,nys:nys+nbgp_local-1,:)
+       IF ( on_device )  THEN
+          !$acc kernels present( ar )
+          !$acc loop
+          DO  i = nxl-nbgp_local, nxr+nbgp_local
+             !$acc loop independent
+             DO  j = 0, nbgp_local-1
+                !$acc loop vector( 32 )
+                DO  k = nzb, nzt+1
+                   ar(k,nys-nbgp_local+j,i) = ar(k,nyn-nbgp_local+1+j,i)
+                     ar(k,nyn+1+j,i)          = ar(k,nys+j,i)
+                ENDDO
+             ENDDO
+          ENDDO
+          !$acc end kernels
+       ELSE
+          ar(:,nys-nbgp_local:nys-1,:) = ar(:,nyn-nbgp_local+1:nyn,:)
+          ar(:,nyn+1:nyn+nbgp_local,:) = ar(:,nys:nys+nbgp_local-1,:)
+       ENDIF
     ENDIF