Adding CUDA support final part

DATA/Par_file

@@ -72,11 +72,11 @@ enreg_surf_same_vertical        = .true.         # receivers inside the medium o
 # display parameters
 NSTEP_BETWEEN_OUTPUT_INFO       = 100            # every how many time steps we display information about the simulation (costly, do not use a very small value)
 ####
-NSTEP_BETWEEN_OUTPUT_IMAGES     = 100            # every how many time steps we draw JPEG or PostScript pictures of the simulation (costly, do not use a very small value)
+NSTEP_BETWEEN_OUTPUT_IMAGES     = 30            # every how many time steps we draw JPEG or PostScript pictures of the simulation (costly, do not use a very small value)
 cutsnaps                        = 1.             # minimum amplitude kept in % for the JPEG and PostScript snapshots; amplitudes below that are muted
 #### for JPEG color images ####
 output_color_image              = .true.         # output JPEG color image of the results every NSTEP_BETWEEN_OUTPUT_IMAGES time steps or not
-imagetype_JPEG                  = 2              # display 1=displ_Ux 2=displ_Uz 3=displ_norm 4=veloc_Vx 5=veloc_Vz 6=veloc_norm 7=accel_Ax 8=accel_Az 9=accel_norm 10=pressure
+imagetype_JPEG                  = 10              # display 1=displ_Ux 2=displ_Uz 3=displ_norm 4=veloc_Vx 5=veloc_Vz 6=veloc_norm 7=accel_Ax 8=accel_Az 9=accel_norm 10=pressure
 factor_subsample_image          = 1.0d0          # (double precision) factor to subsample color images output by the code (useful for very large models)
 USE_CONSTANT_MAX_AMPLITUDE      = .false.        # by default the code normalizes each image independently to its maximum; use this option to use the global maximum below instead
 CONSTANT_MAX_AMPLITUDE_TO_USE   = 1.17d4         # constant maximum amplitude to use for all color images if the above USE_CONSTANT_MAX_AMPLITUDE option is true
@@ -85,7 +85,7 @@ DRAW_SOURCES_AND_RECEIVERS      = .true.         # display sources as orange cro
 DRAW_WATER_IN_BLUE              = .true.         # display acoustic layers as constant blue in JPEG images, because they likely correspond to water in the case of ocean acoustics or in the case of offshore oil industry experiments (if off, display them as greyscale, as for elastic or poroelastic elements, for instance for acoustic-only oil industry models of solid media)
 USE_SNAPSHOT_NUMBER_IN_FILENAME = .false.        # use snapshot number in the file name of JPEG color snapshots instead of the time step (for instance to create movies in an easier way later)
 #### for PostScript snapshots ####
-output_postscript_snapshot      = .true.         # output Postscript snapshot of the results every NSTEP_BETWEEN_OUTPUT_IMAGES time steps or not
+output_postscript_snapshot      = .false.         # output Postscript snapshot of the results every NSTEP_BETWEEN_OUTPUT_IMAGES time steps or not
 imagetype_postscript            = 1              # display 1=displ vector 2=veloc vector 3=accel vector; small arrows are displayed for the vectors
 meshvect                        = .true.         # display mesh on PostScript plots or not
 modelvect                       = .false.        # display velocity model on PostScript plots or not

src/cuda/assemble_MPI_scalar_cuda.cu

@@ -0,0 +1,294 @@
+/*
+!========================================================================
+!
+!                   S P E C F E M 2 D  Version 7 . 0
+!                   --------------------------------
+!
+!     Main historical authors: Dimitri Komatitsch and Jeroen Tromp
+!                        Princeton University, USA
+!                and CNRS / University of Marseille, France
+!                 (there are currently many more authors!)
+! (c) Princeton University and CNRS / University of Marseille, April 2014
+!
+! This software is a computer program whose purpose is to solve
+! the two-dimensional viscoelastic anisotropic or poroelastic wave equation
+! using a spectral-element method (SEM).
+!
+! This software is governed by the CeCILL license under French law and
+! abiding by the rules of distribution of free software. You can use,
+! modify and/or redistribute the software under the terms of the CeCILL
+! license as circulated by CEA, CNRS and Inria at the following URL
+! "http://www.cecill.info".
+!
+! As a counterpart to the access to the source code and rights to copy,
+! modify and redistribute granted by the license, users are provided only
+! with a limited warranty and the software's author, the holder of the
+! economic rights, and the successive licensors have only limited
+! liability.
+!
+! In this respect, the user's attention is drawn to the risks associated
+! with loading, using, modifying and/or developing or reproducing the
+! software by the user in light of its specific status of free software,
+! that may mean that it is complicated to manipulate, and that also
+! therefore means that it is reserved for developers and experienced
+! professionals having in-depth computer knowledge. Users are therefore
+! encouraged to load and test the software's suitability as regards their
+! requirements in conditions enabling the security of their systems and/or
+! data to be ensured and, more generally, to use and operate it in the
+! same conditions as regards security.
+!
+! The full text of the license is available in file "LICENSE".
+!
+!========================================================================
+
+*/
+
+#include <stdio.h>
+#include <cuda.h>
+#include <cublas.h>
+
+#include <sys/time.h>
+#include <sys/resource.h>
+
+#include "config.h"
+#include "mesh_constants_cuda.h"
+
+
+/* ----------------------------------------------------------------------------------------------- */
+
+// ASSEMBLY - mpi data transfer between CPU-GPU
+
+/* ----------------------------------------------------------------------------------------------- */
+
+// prepares a device array with with all inter-element edge-nodes -- this
+// is followed by a memcpy and MPI operations
+__global__ void prepare_boundary_potential_on_device(realw* d_potential_dot_dot_acoustic,
+                                                     realw* d_send_potential_dot_dot_buffer,
+                                                     const int ninterface_ac,
+                                                     const int max_nibool_interfaces_ext_mesh,
+                                                     const int* d_nibool_interfaces_ext_mesh,
+                                                     const int* d_ibool_interfaces_ext_mesh,
+                                                     const int* inum_inter_acoustic) {
+
+  int id = threadIdx.x + blockIdx.x*blockDim.x + blockIdx.y*gridDim.x*blockDim.x;
+  int ientry,iglob,num_int;
+
+  for(int iinterface=0; iinterface < ninterface_ac; iinterface++) {
+
+   num_int=inum_inter_acoustic[iinterface]-1;
+
+    if(id<d_nibool_interfaces_ext_mesh[num_int]) {
+
+      // entry in interface array
+      ientry = id + max_nibool_interfaces_ext_mesh*num_int;
+      // global index in wavefield
+      iglob = d_ibool_interfaces_ext_mesh[ientry] - 1;
+
+      d_send_potential_dot_dot_buffer[ientry] = d_potential_dot_dot_acoustic[iglob];
+    }
+  }
+
+}
+
+
+/* ----------------------------------------------------------------------------------------------- */
+
+// prepares and transfers the inter-element edge-nodes to the host to be MPI'd
+extern "C"
+void FC_FUNC_(transfer_boun_pot_from_device,
+              TRANSFER_BOUN_POT_FROM_DEVICE)(long* Mesh_pointer,
+                                             realw* send_potential_dot_dot_buffer,
+                                             const int* FORWARD_OR_ADJOINT){
+
+TRACE("transfer_boun_pot_from_device");
+
+  Mesh* mp = (Mesh*)(*Mesh_pointer); //get mesh pointer out of fortran integer container
+
+  // checks if anything to do
+  if( mp->size_mpi_buffer_potential > 0 ){
+
+    int blocksize = BLOCKSIZE_TRANSFER;
+    int size_padded = ((int)ceil(((double)(mp->max_nibool_interfaces_ext_mesh))/((double)blocksize)))*blocksize;
+
+    int num_blocks_x, num_blocks_y;
+    get_blocks_xy(size_padded/blocksize,&num_blocks_x,&num_blocks_y);
+
+    dim3 grid(num_blocks_x,num_blocks_y);
+    dim3 threads(blocksize,1,1);
+
+    if(*FORWARD_OR_ADJOINT == 1) {
+
+     prepare_boundary_potential_on_device<<<grid,threads,0,mp->compute_stream>>>(mp->d_potential_dot_dot_acoustic,
+                                                                                   mp->d_send_potential_dot_dot_buffer,
+                                                                                   mp->ninterface_acoustic,
+                                                                                   mp->max_nibool_interfaces_ext_mesh,
+                                                                                   mp->d_nibool_interfaces_ext_mesh,
+                                                                                   mp->d_ibool_interfaces_ext_mesh,
+                                                                                   mp->d_inum_interfaces_acoustic);
+
+      // synchronizes
+      //synchronize_cuda();
+      // explicitly waits until previous compute stream finishes
+      // (cudaMemcpy implicitly synchronizes all other cuda operations)
+      print_CUDA_error_if_any(cudaStreamSynchronize(mp->compute_stream),95);
+
+      print_CUDA_error_if_any(cudaMemcpy(send_potential_dot_dot_buffer,mp->d_send_potential_dot_dot_buffer,
+                                         mp->size_mpi_buffer_potential*sizeof(realw),cudaMemcpyDeviceToHost),98000);
+    }
+    else if(*FORWARD_OR_ADJOINT == 3) {
+      // backward/reconstructed wavefield buffer
+      prepare_boundary_potential_on_device<<<grid,threads,0,mp->compute_stream>>>(mp->d_b_potential_dot_dot_acoustic,
+                                                                                   mp->d_b_send_potential_dot_dot_buffer,
+                                                                                   mp->ninterface_acoustic,
+                                                                                   mp->max_nibool_interfaces_ext_mesh,
+                                                                                   mp->d_nibool_interfaces_ext_mesh,
+                                                                                   mp->d_ibool_interfaces_ext_mesh,
+                                                                                   mp->d_inum_interfaces_acoustic);
+
+      // synchronizes
+      //synchronize_cuda();
+      // explicitly waits until previous compute stream finishes
+      // (cudaMemcpy implicitly synchronizes all other cuda operations)
+      cudaStreamSynchronize(mp->compute_stream);
+
+      print_CUDA_error_if_any(cudaMemcpy(send_potential_dot_dot_buffer,mp->d_b_send_potential_dot_dot_buffer,
+                                         mp->size_mpi_buffer_potential*sizeof(realw),cudaMemcpyDeviceToHost),98001);
+    }
+  }
+
+#ifdef ENABLE_VERY_SLOW_ERROR_CHECKING
+  exit_on_cuda_error("after prepare_boundary_potential_on_device");
+#endif
+
+
+  // finish timing of kernel+memcpy
+  // cudaEventRecord( stop, 0 );
+  // cudaEventSynchronize( stop );
+  // cudaEventElapsedTime( &time, start, stop );
+  // cudaEventDestroy( start );
+  // cudaEventDestroy( stop );
+  // printf("boundary xfer d->h Time: %f ms\n",time);
+#ifdef ENABLE_VERY_SLOW_ERROR_CHECKING
+  exit_on_cuda_error("transfer_boun_pot_from_device");
+#endif
+}
+
+/* ----------------------------------------------------------------------------------------------- */
+
+// Assembly
+
+/* ----------------------------------------------------------------------------------------------- */
+
+
+__global__ void assemble_boundary_potential_on_device(realw* d_potential_dot_dot_acoustic,
+                                                      realw* d_send_potential_dot_dot_buffer,
+                                                      const int ninterface_ac,
+                                                      const int max_nibool_interfaces_ext_mesh,
+                                                      const int* d_nibool_interfaces_ext_mesh,
+                                                      const int* d_ibool_interfaces_ext_mesh,
+                                                      const int* inum_inter_acoustic) {
+
+  int id = threadIdx.x + blockIdx.x*blockDim.x + blockIdx.y*gridDim.x*blockDim.x;
+  int ientry,iglob,num_int;
+
+  for( int iinterface=0; iinterface < ninterface_ac; iinterface++) {
+   num_int=inum_inter_acoustic[iinterface]-1;
+
+
+    if(id<d_nibool_interfaces_ext_mesh[num_int]) {
+
+      // entry in interface array
+      ientry = id + max_nibool_interfaces_ext_mesh*num_int;
+      // global index in wavefield
+      iglob = d_ibool_interfaces_ext_mesh[ientry] - 1;
+
+      // for testing atomic operations against not atomic operations (0.1ms vs. 0.04 ms)
+      // d_potential_dot_dot_acoustic[3*(d_ibool_interfaces_ext_mesh[id+max_nibool_interfaces_ext_mesh*iinterface]-1)] +=
+      // d_send_potential_dot_dot_buffer[3*(id + max_nibool_interfaces_ext_mesh*iinterface)];
+    atomicAdd(&d_potential_dot_dot_acoustic[iglob],d_send_potential_dot_dot_buffer[ientry]);
+    }
+  }
+  // ! This step is done via previous function transfer_and_assemble...
+  // ! do iinterface = 1, num_interfaces_ext_mesh
+  // !   do ipoin = 1, nibool_interfaces_ext_mesh(iinterface)
+  // !     array_val(:,ibool_interfaces_ext_mesh(ipoin,iinterface)) = &
+  // !          array_val(:,ibool_interfaces_ext_mesh(ipoin,iinterface)) + buffer_recv_vector_ext_mesh(:,ipoin,iinterface)
+  // !   enddo
+  // ! enddo
+}
+
+
+/* ----------------------------------------------------------------------------------------------- */
+
+extern "C"
+void FC_FUNC_(transfer_asmbl_pot_to_device,
+              TRANSFER_ASMBL_POT_TO_DEVICE)(long* Mesh_pointer,
+                                            realw* buffer_recv_scalar_ext_mesh,
+                                            const int* FORWARD_OR_ADJOINT) {
+
+TRACE("transfer_asmbl_pot_to_device");
+
+  Mesh* mp = (Mesh*)(*Mesh_pointer); //get mesh pointer out of fortran integer container
+
+  // Cuda timing
+  //cudaEvent_t start, stop;
+  //start_timing_cuda(&start,&stop);
+
+  // checks if anything to do
+  if( mp->size_mpi_buffer_potential > 0 ){
+
+
+    // assembles on GPU
+    int blocksize = BLOCKSIZE_TRANSFER;
+    int size_padded = ((int)ceil(((double)mp->max_nibool_interfaces_ext_mesh)/((double)blocksize)))*blocksize;
+
+    int num_blocks_x, num_blocks_y;
+    get_blocks_xy(size_padded/blocksize,&num_blocks_x,&num_blocks_y);
+
+    dim3 grid(num_blocks_x,num_blocks_y);
+    dim3 threads(blocksize,1,1);
+
+
+    // synchronizes
+    synchronize_cuda();
+
+    if(*FORWARD_OR_ADJOINT == 1) {
+      // copies buffer onto GPU
+      print_CUDA_error_if_any(cudaMemcpy(mp->d_send_potential_dot_dot_buffer, buffer_recv_scalar_ext_mesh,
+                                         mp->size_mpi_buffer_potential*sizeof(realw), cudaMemcpyHostToDevice),98010);
+
+      //assemble forward field
+      assemble_boundary_potential_on_device<<<grid,threads,0,mp->compute_stream>>>(mp->d_potential_dot_dot_acoustic,
+                                                                                    mp->d_send_potential_dot_dot_buffer,
+                                                                                    mp->ninterface_acoustic,
+                                                                                    mp->max_nibool_interfaces_ext_mesh,
+                                                                                    mp->d_nibool_interfaces_ext_mesh,
+                                                                                    mp->d_ibool_interfaces_ext_mesh,
+                                                                                   mp->d_inum_interfaces_acoustic);
+
+
+    }
+    else if(*FORWARD_OR_ADJOINT == 3) {
+      // copies buffer onto GPU
+      print_CUDA_error_if_any(cudaMemcpy(mp->d_b_send_potential_dot_dot_buffer, buffer_recv_scalar_ext_mesh,
+                                         mp->size_mpi_buffer_potential*sizeof(realw), cudaMemcpyHostToDevice),98011);
+
+      //assemble reconstructed/backward field
+      assemble_boundary_potential_on_device<<<grid,threads,0,mp->compute_stream>>>(mp->d_b_potential_dot_dot_acoustic,
+                                                                                    mp->d_b_send_potential_dot_dot_buffer,
+                                                                                    mp->ninterface_acoustic,
+                                                                                    mp->max_nibool_interfaces_ext_mesh,
+                                                                                    mp->d_nibool_interfaces_ext_mesh,
+                                                                                    mp->d_ibool_interfaces_ext_mesh,
+                                                                                   mp->d_inum_interfaces_acoustic);
+    }
+  }
+
+  // Cuda timing
+  //stop_timing_cuda(&start,&stop,"assemble_boundary_potential_on_device");
+
+#ifdef ENABLE_VERY_SLOW_ERROR_CHECKING
+  exit_on_cuda_error("transfer_asmbl_pot_to_device");
+#endif
+}
+