Changeset 4719 for palm/trunk/SOURCE/pres.f90

Timestamp:

Oct 1, 2020 11:28:33 AM (4 years ago)

Author:

pavelkrc

Message:

Bugfix of previous commit (div_old discrepancy)

File:

• palm/trunk/SOURCE/pres.f90 (modified) (12 diffs)

palm/trunk/SOURCE/pres.f90

@@ -25 +25 @@
 ! -----------------
 ! $Id$
+! Bugfix of previous commit (div_old discrepancy)
+!
+! 4717 2020-09-30 22:27:40Z pavelkrc
 ! Fixes and optimizations of OpenMP parallelization, formatting of OpenMP
 ! directives (J. Resler)
@@ -174 +177 @@
     REAL(wp) ::  ddt_3d            !<
     REAL(wp) ::  d_weight_pres     !<
+    REAL(wp) ::  localsum          !<
     REAL(wp) ::  threadsum         !<
     REAL(wp) ::  weight_pres_l     !<
@@ -386 +390 @@
     ENDIF
 
-    !$OMP PARALLEL DO PRIVATE (i,j,k) SCHEDULE( STATIC )
+    localsum  = 0.0_wp
+    threadsum = 0.0_wp
+
+    !$OMP PARALLEL PRIVATE (i,j,k)
+    !$OMP DO SCHEDULE( STATIC )
     !$ACC PARALLEL LOOP COLLAPSE(3) PRIVATE(i, j, k) &
     !$ACC PRESENT(u, v, w, rho_air, rho_air_zw, ddzw, wall_flags_total_0) &
@@ -401 +409 @@
        ENDDO
     ENDDO
+    !$OMP END PARALLEL
 
 !
@@ -407 +416 @@
     IF ( intermediate_timestep_count == intermediate_timestep_count_max  .OR.                      &
          intermediate_timestep_count == 0 )  THEN
-       threadsum = 0.0_wp
-       !$OMP PARALLEL DO PRIVATE (i,j,k) REDUCTION(+:threadsum) SCHEDULE( STATIC )
+       !$OMP PARALLEL PRIVATE (i,j,k) FIRSTPRIVATE(threadsum) REDUCTION(+:localsum)
+       !$OMP DO SCHEDULE( STATIC )
        !$ACC PARALLEL LOOP COLLAPSE(3) PRIVATE(i,j,k) &
        !$ACC REDUCTION(+:threadsum) COPY(threadsum) &
@@ -419 +428 @@
           ENDDO
        ENDDO
-       threadsum = threadsum + threadsum * dt_3d * weight_pres_l
+       localsum = localsum + threadsum * dt_3d * weight_pres_l
+       !$OMP END PARALLEL
     ENDIF
 
@@ -426 +436 @@
     IF ( intermediate_timestep_count == intermediate_timestep_count_max  .OR.                      &
          intermediate_timestep_count == 0 )  THEN
-       sums_divold_l(0:statistic_regions) = threadsum
+       sums_divold_l(0:statistic_regions) = localsum
     ENDIF
 
@@ -665 +675 @@
     IF ( conserve_volume_flow  .AND.  bc_lr_cyc  .AND.  bc_ns_cyc  .AND.  nxr == nx )  THEN
 
-       !$OMP PARALLEL DO PRIVATE (j,k) REDUCTION (volume_flow_l(1))
+       !$OMP PARALLEL DO PRIVATE (j,k) REDUCTION (+:volume_flow_l(1))
        DO  j = nys, nyn
           DO  k = nzb+1, nzt
@@ -677 +687 @@
     IF ( conserve_volume_flow  .AND.  bc_ns_cyc  .AND.  bc_lr_cyc  .AND. nyn == ny )  THEN
 
-       !$OMP PARALLEL DO PRIVATE (i,k) REDUCTION (volume_flow_l(2))
+       !$OMP PARALLEL DO PRIVATE (i,k) REDUCTION (+:volume_flow_l(2))
        DO  i = nxl, nxr
           DO  k = nzb+1, nzt
@@ -736 +746 @@
        IF ( topography /= 'flat' )  d = 0.0_wp
 
-       !$OMP PARALLEL DO PRIVATE (i,j,k) SCHEDULE( STATIC )
+       localsum  = 0.0_wp
+       threadsum = 0.0_wp
+
+       !$OMP PARALLEL PRIVATE (i,j,k) FIRSTPRIVATE(threadsum) REDUCTION(+:localsum)
+       !$OMP DO SCHEDULE( STATIC )
        !$ACC PARALLEL LOOP COLLAPSE(3) PRIVATE(i, j, k) &
        !$ACC PRESENT(u, v, w, rho_air, rho_air_zw, ddzw, wall_flags_total_0) &
@@ -753 +767 @@
 !
 !--    Compute possible PE-sum of divergences for flow_statistics
-       threadsum = 0.0_wp
-       !$OMP PARALLEL DO PRIVATE (i,j,k) REDUCTION(+:threadsum) SCHEDULE( STATIC )
+       !$OMP DO SCHEDULE( STATIC )
        !$ACC PARALLEL LOOP COLLAPSE(3) PRIVATE(i, j, k) &
        !$ACC REDUCTION(+:threadsum) COPY(threadsum) &
@@ -766 +779 @@
        ENDDO
 
+       localsum = localsum + threadsum
+       !$OMP END PARALLEL
+
 !
 !--    For completeness, set the divergence sum of all statistic regions to those of the total
 !--    domain
-       sums_divnew_l(0:statistic_regions) = threadsum
+       sums_divnew_l(0:statistic_regions) = localsum
 
        CALL cpu_log( log_point_s(1), 'divergence', 'stop' )