Changeset 1482 for palm/trunk/SOURCE/fft_xy.f90

Timestamp:

Oct 18, 2014 12:34:45 PM (10 years ago)

Author:

raasch

Message:

adjustments for using CUDA-aware MPI

File:

• palm/trunk/SOURCE/fft_xy.f90 (modified) (7 diffs)

palm/trunk/SOURCE/fft_xy.f90

@@ -20 +20 @@
 ! Current revisions:
 ! -----------------
-! 
+! cudafft workaround for data declaration of ar_tmp because of PGI 14.1 bug
 ! 
 ! Former revisions:
@@ -404 +404 @@
        COMPLEX(dp), DIMENSION(0:(nx+1)/2,nys_x:nyn_x,nzb_x:nzt_x) ::           &
           ar_tmp  !:
-       !$acc declare create( ar_tmp )
+       ! following does not work for PGI 14.1 -> to be removed later
+       ! !$acc declare create( ar_tmp )
 #endif
 
@@ -711 +712 @@
 #elif defined( __cuda_fft )
 
+          !$acc data create( ar_tmp )
           IF ( forward_fft )  THEN
 
@@ -757 +759 @@
 
           ENDIF
+          !$acc end data
 
 #else
@@ -1054 +1057 @@
        COMPLEX(dp), DIMENSION(0:(ny+1)/2,nxl_y:nxr_y,nzb_y:nzt_y) ::           &
           ar_tmp  !:
+       ! following does not work for PGI 14.1 -> to be removed later
        !$acc declare create( ar_tmp )
 #endif
@@ -1334 +1338 @@
 #elif defined( __cuda_fft )
 
+          !$acc data create( ar_tmp )
           IF ( forward_fft )  THEN
 
@@ -1380 +1385 @@
 
           ENDIF
+          !$acc end data
 
 #else