Changeset 1015 for palm/trunk/SOURCE/palm.f90

Timestamp:

Sep 27, 2012 9:23:24 AM (12 years ago)

Author:

raasch

Message:

Starting with changes required for GPU optimization. OpenACC statements for using NVIDIA GPUs added.
Adjustment of mixing length to the Prandtl mixing length at first grid point above ground removed.
mask array is set zero for ghost boundaries

File:

• palm/trunk/SOURCE/palm.f90 (modified) (6 diffs)

palm/trunk/SOURCE/palm.f90

@@ -4 +4 @@
 ! Current revisions:
 ! -----------------
-! 
+! OpenACC statements added + code changes required for GPU optimization
 !
 ! Former revisions:
@@ -77 +77 @@
     USE statistics
 
+#if defined( __openacc )
+    USE OPENACC
+#endif
+
     IMPLICIT NONE
 
@@ -84 +88 @@
     CHARACTER (LEN=1) ::  cdum
     INTEGER           ::  i, run_description_header_i(80)
+#if defined( __openacc )
+    REAL, DIMENSION(100) ::  acc_dum
+#endif
 
     version = 'PALM 3.8a'
@@ -102 +109 @@
 #endif
 
+#if defined( __openacc )
+!
+!-- Get the number of accelerator boards per node and assign the MPI processes
+!-- to these boards
+    num_acc_per_node  = ACC_GET_NUM_DEVICES( ACC_DEVICE_NVIDIA )
+    acc_rank = MOD( myid, num_acc_per_node )
+    CALL ACC_SET_DEVICE_NUM ( acc_rank, ACC_DEVICE_NVIDIA )
+!
+!-- Test output (to be removed later)
+    WRITE (*,'(A,I4,A,I3,A,I3,A,I3)') '*** Connect MPI-Task ', myid,' to CPU ',&
+                                      acc_rank, ' Devices: ', num_acc_per_node,&
+                                      ' connected to:',                        &
+                                      ACC_GET_DEVICE_NUM( ACC_DEVICE_NVIDIA )
+#endif
+!
+!-- Ensure that OpenACC first attaches the GPU devices by copying a dummy data
+!-- region
+    !$acc data copyin( acc_dum )
+
 !
 !-- Initialize measuring of the CPU-time remaining to the run
@@ -177 +203 @@
     ENDIF
 
+!
+!-- Declare and initialize variables in the accelerator memory with their
+!-- host values
+    !$acc  data copyin( diss, e, e_p, kh, km, pt, pt_p, q, ql, tend, te_m, tpt_m, tu_m, tv_m, tw_m, u, u_p, v, vpt, v_p, w, w_p )          &
+    !$acc       copyin( ddzu, ddzw, dd2zu, l_grid, l_wall, ptdf_x, ptdf_y, pt_init, rdf, rdf_sc, ug, vg, zu, zw )   &
+    !$acc       copyin( hom, qs, qsws, qswst, rif, rif_wall, shf, ts, tswst, us, usws, uswst, vsws, vswst, z0, z0h )      &
+    !$acc       copyin( fxm, fxp, fym, fyp, fwxm, fwxp, fwym, fwyp, nzb_diff_s_inner, nzb_diff_s_outer, nzb_diff_u )       &
+    !$acc       copyin( nzb_diff_v, nzb_s_inner, nzb_s_outer, nzb_u_inner )    &
+    !$acc       copyin( nzb_u_outer, nzb_v_inner, nzb_v_outer, nzb_w_inner )   &
+    !$acc       copyin( nzb_w_outer, wall_heatflux, wall_e_x, wall_e_y, wall_u, wall_v, wall_w_x, wall_w_y, wall_flags_0 )
 !
 !-- Integration of the model equations using timestep-scheme
@@ -242 +278 @@
 
 !
+!-- Close the OpenACC dummy data region
+    !$acc end data
+    !$acc end data
+
+!
 !-- Take final CPU-time for CPU-time analysis
     CALL cpu_log( log_point(1), 'total', 'stop' )