[cig-commits] r19070 - in seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src: cuda specfem3D

[rietmann at geodynamics.org]

Author: rietmann
Date: 2011-10-14 06:47:03 -0700 (Fri, 14 Oct 2011)
New Revision: 19070

Modified:
   seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/cuda/specfem3D_gpu_cuda_method_stubs.c
   seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/specfem3D/compute_forces_elastic_Dev_openmp.f90
Log:
quick fix to openmp and added GPU_MODE msg to function stubs

Modified: seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/cuda/specfem3D_gpu_cuda_method_stubs.c
===================================================================
--- seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/cuda/specfem3D_gpu_cuda_method_stubs.c	2011-10-14 08:09:47 UTC (rev 19069)
+++ seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/cuda/specfem3D_gpu_cuda_method_stubs.c	2011-10-14 13:47:03 UTC (rev 19070)
@@ -333,7 +333,11 @@
                                         int* h_ispec_selected_rec,
                                         int* nrec_f,
                                         int* nrec_local_f,
-                                        int* SIMULATION_TYPE){}
+                                        int* SIMULATION_TYPE)
+{
+  fprintf(stderr,"ERROR: GPU_MODE enabled without GPU/CUDA Support. To enable GPU support, reconfigure with --with-cuda flag.\n");
+  exit(1);
+}
 
 void FC_FUNC_(prepare_adjoint_sim2_or_3_constants_device,
               PREPARE_ADJOINT_SIM2_OR_3_CONSTANTS_DEVICE)(

Modified: seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/specfem3D/compute_forces_elastic_Dev_openmp.f90
===================================================================
--- seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/specfem3D/compute_forces_elastic_Dev_openmp.f90	2011-10-14 08:09:47 UTC (rev 19069)
+++ seismo/3D/SPECFEM3D/branches/SPECFEM3D_SUNFLOWER/src/specfem3D/compute_forces_elastic_Dev_openmp.f90	2011-10-14 13:47:03 UTC (rev 19070)
@@ -131,7 +131,7 @@
        tempx1,tempx2,tempx3,tempy1,tempy2,tempy3,tempz1,tempz2,tempz3  
 
   ! real(kind=CUSTOM_REAL), dimension(NDIM,NGLOB_AB,NUM_THREADS) :: accel_omp
-  real(kind=CUSTOM_REAL), dimension(:,:,:),allocatable :: accel_omp
+  ! real(kind=CUSTOM_REAL), dimension(:,:,:),allocatable :: accel_omp
   ! local attenuation parameters
   real(kind=CUSTOM_REAL), dimension(NGLLX,NGLLY,NGLLZ) :: epsilondev_xx_loc, &
        epsilondev_yy_loc, epsilondev_xy_loc, epsilondev_xz_loc, epsilondev_yz_loc
@@ -177,7 +177,7 @@
   NUM_THREADS = OMP_GET_MAX_THREADS()
   
   
-  allocate(accel_omp(NDIM,NGLOB_AB,NUM_THREADS))
+  ! allocate(accel_omp(NDIM,NGLOB_AB,NUM_THREADS))
   
   ! allocate local arrays
   allocate(dummyx_loc(NGLLX,NGLLY,NGLLZ,NUM_THREADS))
@@ -206,7 +206,10 @@
     num_elements = nspec_outer_elastic
   else
     num_elements = nspec_inner_elastic
-  endif  
+ endif
+ ! "start" timer
+ start_time = omp_get_wtime()
+  
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(&
   !$OMP R_xx_val1,R_yy_val1,R_xx_val2,R_yy_val2,R_xx_val3,R_yy_val3,&
   !$OMP factor_loc,alphaval_loc,betaval_loc,gammaval_loc,&
@@ -229,9 +232,8 @@
   thread_id = OMP_get_thread_num()+1
   ! thread_id = 1
 
-  ! "start" timer
-  start_time = omp_get_wtime()
   
+  
   ! accel_omp(:,:,thread_id) = 0.0
   !$OMP DO
   do ispec_p = 1,num_elements