Changeset 1257 for palm/trunk/SOURCE/fft_xy.f90

Timestamp:

Nov 8, 2013 3:18:40 PM (10 years ago)

Author:

raasch

Message:

New:
---

openACC porting of timestep calculation
(modules, timestep, time_integration)

Changed:

---

openACC loop directives and vector clauses removed (because they do not give any performance improvement with PGI
compiler versions > 13.6)
(advec_ws, buoyancy, coriolis, diffusion_e, diffusion_s, diffusion_u, diffusion_v, diffusion_w, diffusivities, exchange_horiz, fft_xy, pres, production_e, transpose, tridia_solver, wall_fluxes)

openACC loop independent clauses added
(boundary_conds, prandtl_fluxes, pres)

openACC declare create statements moved after FORTRAN declaration statement
(diffusion_u, diffusion_v, diffusion_w, fft_xy, poisfft, production_e, tridia_solver)

openACC end parallel replaced by end parallel loop
(flow_statistics, pres)

openACC "kernels do" replaced by "kernels loop"
(prandtl_fluxes)

output format for theta* changed to avoid output of *
(run_control)

Errors:

---

bugfix for calculation of advective timestep (old version may cause wrong timesteps in case of
vertixcally stretched grids)
Attention: standard run-control output has changed!
(timestep)

File:

• palm/trunk/SOURCE/fft_xy.f90 (modified) (9 diffs)

palm/trunk/SOURCE/fft_xy.f90

@@ -20 +20 @@
 ! Current revisions:
 ! -----------------
-!
+! openacc loop and loop vector clauses removed, declare create moved after
+! the FORTRAN declaration statement
 !
 ! Former revisions:
@@ -328 +329 @@
        REAL, DIMENSION(6*(nx+1)) ::  work2
 #elif defined( __cuda_fft )
+       COMPLEX(dpk), DIMENSION(0:(nx+1)/2,nys_x:nyn_x,nzb_x:nzt_x) ::  ar_tmp
        !$acc declare create( ar_tmp )
-       COMPLEX(dpk), DIMENSION(0:(nx+1)/2,nys_x:nyn_x,nzb_x:nzt_x) ::  ar_tmp
 #endif
        REAL, DIMENSION(0:nx,nys_x:nyn_x), OPTIONAL   ::  ar_2d
@@ -634 +635 @@
 
              !$acc kernels
-             !$acc loop
              DO  k = nzb_x, nzt_x
                 DO  j = nys_x, nyn_x
 
-                   !$acc loop vector( 32 )
                    DO  i = 0, (nx+1)/2
                       ar(i,j,k)      = REAL( ar_tmp(i,j,k) )  * dnx
                    ENDDO
 
-                   !$acc loop vector( 32 )
                    DO  i = 1, (nx+1)/2 - 1
                       ar(nx+1-i,j,k) = AIMAG( ar_tmp(i,j,k) ) * dnx
@@ -657 +655 @@
              !$acc data present( ar )
              !$acc kernels
-             !$acc loop
              DO  k = nzb_x, nzt_x
                 DO  j = nys_x, nyn_x
@@ -663 +660 @@
                    ar_tmp(0,j,k) = CMPLX( ar(0,j,k), 0.0 )
 
-                   !$acc loop vector( 32 )
                    DO  i = 1, (nx+1)/2 - 1
                       ar_tmp(i,j,k) = CMPLX( ar(i,j,k), ar(nx+1-i,j,k) )
@@ -957 +953 @@
        REAL, DIMENSION(6*(ny+1)) ::  work2
 #elif defined( __cuda_fft )
+       COMPLEX(dpk), DIMENSION(0:(ny+1)/2,nxl_y:nxr_y,nzb_y:nzt_y) ::  ar_tmp
        !$acc declare create( ar_tmp )
-       COMPLEX(dpk), DIMENSION(0:(ny+1)/2,nxl_y:nxr_y,nzb_y:nzt_y) ::  ar_tmp
 #endif
        REAL, DIMENSION(0:ny,nxl_y_l:nxr_y_l,nzb_y:nzt_y) ::  ar
@@ -1236 +1232 @@
 
              !$acc kernels
-             !$acc loop
              DO  k = nzb_y, nzt_y
                 DO  i = nxl_y, nxr_y
 
-                   !$acc loop vector( 32 )
                    DO  j = 0, (ny+1)/2
                       ar(j,i,k)      = REAL( ar_tmp(j,i,k) )  * dny
                    ENDDO
 
-                   !$acc loop vector( 32 )
                    DO  j = 1, (ny+1)/2 - 1
                       ar(ny+1-j,i,k) = AIMAG( ar_tmp(j,i,k) ) * dny
@@ -1259 +1252 @@
              !$acc data present( ar )
              !$acc kernels
-             !$acc loop
              DO  k = nzb_y, nzt_y
                 DO  i = nxl_y, nxr_y
@@ -1265 +1257 @@
                    ar_tmp(0,i,k) = CMPLX( ar(0,i,k), 0.0 )
 
-                   !$acc loop vector( 32 )
                    DO  j = 1, (ny+1)/2 - 1
                       ar_tmp(j,i,k) = CMPLX( ar(j,i,k), ar(ny+1-j,i,k) )