Changeset 3690 for palm/trunk/SOURCE/tridia_solver_mod.f90

Timestamp:

Jan 22, 2019 10:56:42 PM (4 years ago)

Author:

knoop

Message:

Enabled OpenACC usage without using the cudaFFT library.
Added respective palmtest build configuration and testcase.

File:

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

palm/trunk/SOURCE/tridia_solver_mod.f90

@@ -101 +101 @@
 ! the beginning, i.e. routine split is called within tridia_init.
 !
+
+#define __acc_fft_device ( defined( _OPENACC ) && ( defined ( __cuda_fft ) ) )
+
 !
 ! Description:
@@ -173 +176 @@
        CALL split
 
+#if __acc_fft_device
        !$ACC ENTER DATA &
        !$ACC COPYIN(ddzuw(0:nz-1,1:3)) &
        !$ACC COPYIN(tri(nxl_z:nxr_z,nys_z:nyn_z,0:nz-1,1:2))
+#endif
 
     END SUBROUTINE tridia_init
@@ -297 +302 @@
 
           REAL(wp), DIMENSION(nxl_z:nxr_z,nys_z:nyn_z,0:nz-1)   ::  ar1 !< 
+#if __acc_fft_device
           !$ACC DECLARE CREATE(ar1)
+#endif
 
 !
 !--       Forward substitution
+#if __acc_fft_device
           !$ACC PARALLEL PRESENT(ar, ar1, tri) PRIVATE(i,j,k)
+#endif
           DO  k = 0, nz - 1
+#if __acc_fft_device
              !$ACC LOOP COLLAPSE(2)
+#endif
              DO  j = nys_z, nyn_z
                 DO  i = nxl_z, nxr_z
@@ -316 +327 @@
              ENDDO
           ENDDO
+#if __acc_fft_device
           !$ACC END PARALLEL
+#endif
 
 !
@@ -323 +336 @@
 !--       by zero appearing if the pressure bc is set to neumann at the top of
 !--       the model domain.
+#if __acc_fft_device
           !$ACC PARALLEL PRESENT(ar, ar1, ddzuw, tri) PRIVATE(i,j,k)
+#endif
           DO  k = nz-1, 0, -1
+#if __acc_fft_device
              !$ACC LOOP COLLAPSE(2)
+#endif
              DO  j = nys_z, nyn_z
                 DO  i = nxl_z, nxr_z
@@ -338 +355 @@
              ENDDO
           ENDDO
+#if __acc_fft_device
           !$ACC END PARALLEL
+#endif
 
 !
@@ -346 +365 @@
           IF ( ibc_p_b == 1  .AND.  ibc_p_t == 1 )  THEN
              IF ( nys_z == 0  .AND.  nxl_z == 0 )  THEN
+#if __acc_fft_device
                 !$ACC PARALLEL LOOP PRESENT(ar)
+#endif
                 DO  k = 1, nz
                    ar(nxl_z,nys_z,k) = 0.0_wp