// host stub function void ops_par_loop_update_halo_kernel1_t2(char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3, ops_arg arg4, ops_arg arg5, ops_arg arg6, ops_arg arg7) { // Timing double t1, t2, c1, c2; ops_arg args[8] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 8, range, 13)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(13, "update_halo_kernel1_t2"); OPS_kernels[13].count++; ops_timers_core(&c1, &t1); } // compute localy allocated range for the sub-block int start[3]; int end[3]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // OPS_MPI int arg_idx[3]; int arg_idx_base[3]; #ifdef OPS_MPI if (compute_ranges(args, 8, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif for (int n = 0; n < 3; n++) { arg_idx_base[n] = arg_idx[n]; } int dat0 = args[0].dat->elem_size; int dat1 = args[1].dat->elem_size; int dat2 = args[2].dat->elem_size; int dat3 = args[3].dat->elem_size; int dat4 = args[4].dat->elem_size; int dat5 = args[5].dat->elem_size; int dat6 = args[6].dat->elem_size; int *arg7h = (int *)arg7.data; // Upload large globals #ifdef OPS_GPU int consts_bytes = 0; consts_bytes += ROUND_UP(NUM_FIELDS * sizeof(int)); reallocConstArrays(consts_bytes); consts_bytes = 0; args[7].data = OPS_consts_h + consts_bytes; args[7].data_d = OPS_consts_d + consts_bytes; for (int d = 0; d < NUM_FIELDS; d++) ((int *)args[7].data)[d] = arg7h[d]; consts_bytes += ROUND_UP(NUM_FIELDS * sizeof(int)); mvConstArraysToDevice(consts_bytes); #endif // OPS_GPU // set up initial pointers int base0 = args[0].dat->base_offset + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * start[0] * args[0].stencil->stride[0]; base0 = base0 + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * args[0].dat->size[0] * start[1] * args[0].stencil->stride[1]; base0 = base0 + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * args[0].dat->size[0] * args[0].dat->size[1] * start[2] * args[0].stencil->stride[2]; #ifdef OPS_GPU double *p_a0 = (double *)((char *)args[0].data_d + base0); #else double *p_a0 = (double *)((char *)args[0].data + base0); #endif int base1 = args[1].dat->base_offset + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * start[0] * args[1].stencil->stride[0]; base1 = base1 + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * args[1].dat->size[0] * start[1] * args[1].stencil->stride[1]; base1 = base1 + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * args[1].dat->size[0] * args[1].dat->size[1] * start[2] * args[1].stencil->stride[2]; #ifdef OPS_GPU double *p_a1 = (double *)((char *)args[1].data_d + base1); #else double *p_a1 = (double *)((char *)args[1].data + base1); #endif int base2 = args[2].dat->base_offset + (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size) * start[0] * args[2].stencil->stride[0]; base2 = base2 + (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size) * args[2].dat->size[0] * start[1] * args[2].stencil->stride[1]; base2 = base2 + (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size) * args[2].dat->size[0] * args[2].dat->size[1] * start[2] * args[2].stencil->stride[2]; #ifdef OPS_GPU double *p_a2 = (double *)((char *)args[2].data_d + base2); #else double *p_a2 = (double *)((char *)args[2].data + base2); #endif int base3 = args[3].dat->base_offset + (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size) * start[0] * args[3].stencil->stride[0]; base3 = base3 + (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size) * args[3].dat->size[0] * start[1] * args[3].stencil->stride[1]; base3 = base3 + (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size) * args[3].dat->size[0] * args[3].dat->size[1] * start[2] * args[3].stencil->stride[2]; #ifdef OPS_GPU double *p_a3 = (double *)((char *)args[3].data_d + base3); #else double *p_a3 = (double *)((char *)args[3].data + base3); #endif int base4 = args[4].dat->base_offset + (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size) * start[0] * args[4].stencil->stride[0]; base4 = base4 + (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size) * args[4].dat->size[0] * start[1] * args[4].stencil->stride[1]; base4 = base4 + (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size) * args[4].dat->size[0] * args[4].dat->size[1] * start[2] * args[4].stencil->stride[2]; #ifdef OPS_GPU double *p_a4 = (double *)((char *)args[4].data_d + base4); #else double *p_a4 = (double *)((char *)args[4].data + base4); #endif int base5 = args[5].dat->base_offset + (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size) * start[0] * args[5].stencil->stride[0]; base5 = base5 + (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size) * args[5].dat->size[0] * start[1] * args[5].stencil->stride[1]; base5 = base5 + (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size) * args[5].dat->size[0] * args[5].dat->size[1] * start[2] * args[5].stencil->stride[2]; #ifdef OPS_GPU double *p_a5 = (double *)((char *)args[5].data_d + base5); #else double *p_a5 = (double *)((char *)args[5].data + base5); #endif int base6 = args[6].dat->base_offset + (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size) * start[0] * args[6].stencil->stride[0]; base6 = base6 + (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size) * args[6].dat->size[0] * start[1] * args[6].stencil->stride[1]; base6 = base6 + (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size) * args[6].dat->size[0] * args[6].dat->size[1] * start[2] * args[6].stencil->stride[2]; #ifdef OPS_GPU double *p_a6 = (double *)((char *)args[6].data_d + base6); #else double *p_a6 = (double *)((char *)args[6].data + base6); #endif #ifdef OPS_GPU int *p_a7 = (int *)args[7].data_d; #else int *p_a7 = arg7h; #endif int x_size = MAX(0, end[0] - start[0]); int y_size = MAX(0, end[1] - start[1]); int z_size = MAX(0, end[2] - start[2]); // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; ydim0 = args[0].dat->size[1]; xdim1 = args[1].dat->size[0]; ydim1 = args[1].dat->size[1]; xdim2 = args[2].dat->size[0]; ydim2 = args[2].dat->size[1]; xdim3 = args[3].dat->size[0]; ydim3 = args[3].dat->size[1]; xdim4 = args[4].dat->size[0]; ydim4 = args[4].dat->size[1]; xdim5 = args[5].dat->size[0]; ydim5 = args[5].dat->size[1]; xdim6 = args[6].dat->size[0]; ydim6 = args[6].dat->size[1]; if (xdim0 != xdim0_update_halo_kernel1_t2_h || ydim0 != ydim0_update_halo_kernel1_t2_h || xdim1 != xdim1_update_halo_kernel1_t2_h || ydim1 != ydim1_update_halo_kernel1_t2_h || xdim2 != xdim2_update_halo_kernel1_t2_h || ydim2 != ydim2_update_halo_kernel1_t2_h || xdim3 != xdim3_update_halo_kernel1_t2_h || ydim3 != ydim3_update_halo_kernel1_t2_h || xdim4 != xdim4_update_halo_kernel1_t2_h || ydim4 != ydim4_update_halo_kernel1_t2_h || xdim5 != xdim5_update_halo_kernel1_t2_h || ydim5 != ydim5_update_halo_kernel1_t2_h || xdim6 != xdim6_update_halo_kernel1_t2_h || ydim6 != ydim6_update_halo_kernel1_t2_h) { xdim0_update_halo_kernel1_t2 = xdim0; xdim0_update_halo_kernel1_t2_h = xdim0; ydim0_update_halo_kernel1_t2 = ydim0; ydim0_update_halo_kernel1_t2_h = ydim0; xdim1_update_halo_kernel1_t2 = xdim1; xdim1_update_halo_kernel1_t2_h = xdim1; ydim1_update_halo_kernel1_t2 = ydim1; ydim1_update_halo_kernel1_t2_h = ydim1; xdim2_update_halo_kernel1_t2 = xdim2; xdim2_update_halo_kernel1_t2_h = xdim2; ydim2_update_halo_kernel1_t2 = ydim2; ydim2_update_halo_kernel1_t2_h = ydim2; xdim3_update_halo_kernel1_t2 = xdim3; xdim3_update_halo_kernel1_t2_h = xdim3; ydim3_update_halo_kernel1_t2 = ydim3; ydim3_update_halo_kernel1_t2_h = ydim3; xdim4_update_halo_kernel1_t2 = xdim4; xdim4_update_halo_kernel1_t2_h = xdim4; ydim4_update_halo_kernel1_t2 = ydim4; ydim4_update_halo_kernel1_t2_h = ydim4; xdim5_update_halo_kernel1_t2 = xdim5; xdim5_update_halo_kernel1_t2_h = xdim5; ydim5_update_halo_kernel1_t2 = ydim5; ydim5_update_halo_kernel1_t2_h = ydim5; xdim6_update_halo_kernel1_t2 = xdim6; xdim6_update_halo_kernel1_t2_h = xdim6; ydim6_update_halo_kernel1_t2 = ydim6; ydim6_update_halo_kernel1_t2_h = ydim6; } // Halo Exchanges #ifdef OPS_GPU ops_H_D_exchanges_device(args, 8); #else ops_H_D_exchanges_host(args, 8); #endif ops_halo_exchanges(args, 8, range); #ifdef OPS_GPU ops_H_D_exchanges_device(args, 8); #else ops_H_D_exchanges_host(args, 8); #endif if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[13].mpi_time += t2 - t1; } update_halo_kernel1_t2_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, p_a5, p_a6, p_a7, x_size, y_size, z_size); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[13].time += t1 - t2; } #ifdef OPS_GPU ops_set_dirtybit_device(args, 8); #else ops_set_dirtybit_host(args, 8); #endif ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[1], range); ops_set_halo_dirtybit3(&args[2], range); ops_set_halo_dirtybit3(&args[3], range); ops_set_halo_dirtybit3(&args[4], range); ops_set_halo_dirtybit3(&args[5], range); ops_set_halo_dirtybit3(&args[6], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[13].mpi_time += t2 - t1; OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[13].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_update_halo_kernel5_plus_2_right(char const *name, ops_block block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2) { ops_arg args[3] = { arg0, arg1, arg2}; ops_timing_realloc(120,"update_halo_kernel5_plus_2_right"); OPS_kernels[120].count++; //compute locally allocated range for the sub-block int start[3]; int end[3]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; if (!sb->owned) return; for ( int n=0; n<3; n++ ){ start[n] = sb->decomp_disp[n];end[n] = sb->decomp_disp[n]+sb->decomp_size[n]; if (start[n] >= range[2*n]) { start[n] = 0; } else { start[n] = range[2*n] - start[n]; } if (sb->id_m[n]==MPI_PROC_NULL && range[2*n] < 0) start[n] = range[2*n]; if (end[n] >= range[2*n+1]) { end[n] = range[2*n+1] - sb->decomp_disp[n]; } else { end[n] = sb->decomp_size[n]; } if (sb->id_p[n]==MPI_PROC_NULL && (range[2*n+1] > sb->decomp_disp[n]+sb->decomp_size[n])) end[n] += (range[2*n+1]-sb->decomp_disp[n]-sb->decomp_size[n]); } #else //OPS_MPI for ( int n=0; n<3; n++ ){ start[n] = range[2*n];end[n] = range[2*n+1]; } #endif //OPS_MPI int x_size = MAX(0,end[0]-start[0]); int y_size = MAX(0,end[1]-start[1]); int z_size = MAX(0,end[2]-start[2]); int xdim0 = args[0].dat->size[0]*args[0].dat->dim; int ydim0 = args[0].dat->size[1]; int xdim1 = args[1].dat->size[0]*args[1].dat->dim; int ydim1 = args[1].dat->size[1]; //build opencl kernel if not already built buildOpenCLKernels_update_halo_kernel5_plus_2_right( xdim0,ydim0,xdim1,ydim1); //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); //set up OpenCL thread blocks size_t globalWorkSize[3] = {((x_size-1)/OPS_block_size_x+ 1)*OPS_block_size_x, ((y_size-1)/OPS_block_size_y + 1)*OPS_block_size_y, MAX(1,end[2]-start[2])}; size_t localWorkSize[3] = {OPS_block_size_x,OPS_block_size_y,1}; int *arg2h = (int *)arg2.data; int consts_bytes = 0; consts_bytes += ROUND_UP(NUM_FIELDS*sizeof(int)); reallocConstArrays(consts_bytes); consts_bytes = 0; arg2.data = OPS_consts_h + consts_bytes; arg2.data_d = OPS_consts_d + consts_bytes; for (int d=0; d<NUM_FIELDS; d++) ((int *)arg2.data)[d] = arg2h[d]; consts_bytes += ROUND_UP(NUM_FIELDS*sizeof(int)); mvConstArraysToDevice(consts_bytes); int dat0 = args[0].dat->elem_size; int dat1 = args[1].dat->elem_size; //set up initial pointers int d_m[OPS_MAX_DIM]; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d] + OPS_sub_dat_list[args[0].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d]; #endif //OPS_MPI int base0 = 1 * (start[0] * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0 + args[0].dat->size[0] * (start[1] * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]); base0 = base0 + args[0].dat->size[0] * args[0].dat->size[1] * (start[2] * args[0].stencil->stride[2] - args[0].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d] + OPS_sub_dat_list[args[1].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d]; #endif //OPS_MPI int base1 = 1 * (start[0] * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); base1 = base1 + args[1].dat->size[0] * (start[1] * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]); base1 = base1 + args[1].dat->size[0] * args[1].dat->size[1] * (start[2] * args[1].stencil->stride[2] - args[1].dat->base[2] - d_m[2]); ops_H_D_exchanges_device(args, 3); ops_halo_exchanges(args,3,range); ops_H_D_exchanges_device(args, 3); ops_timers_core(&c1,&t1); OPS_kernels[120].mpi_time += t1-t2; clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 0, sizeof(cl_mem), (void*) &arg0.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 1, sizeof(cl_mem), (void*) &arg1.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 2, sizeof(cl_mem), (void*) &arg2.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 3, sizeof(cl_int), (void*) &base0 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 4, sizeof(cl_int), (void*) &base1 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 5, sizeof(cl_int), (void*) &x_size )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 6, sizeof(cl_int), (void*) &y_size )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[120], 7, sizeof(cl_int), (void*) &z_size )); //call/enque opencl kernel wrapper function clSafeCall( clEnqueueNDRangeKernel(OPS_opencl_core.command_queue, OPS_opencl_core.kernel[120], 3, NULL, globalWorkSize, localWorkSize, 0, NULL, NULL) ); if (OPS_diags>1) { clSafeCall( clFinish(OPS_opencl_core.command_queue) ); } ops_set_dirtybit_device(args, 3); ops_set_halo_dirtybit3(&args[0],range); ops_set_halo_dirtybit3(&args[1],range); //Update kernel record ops_timers_core(&c2,&t2); OPS_kernels[120].time += t2-t1; OPS_kernels[120].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[120].transfer += ops_compute_transfer(dim, range, &arg1); }
// host stub function void ops_par_loop_update_halo_kernel1_b2(char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3, ops_arg arg4, ops_arg arg5, ops_arg arg6, ops_arg arg7) { // Timing double t1, t2, c1, c2; ops_arg args[8] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 8, range, 12)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(12, "update_halo_kernel1_b2"); OPS_kernels[12].count++; ops_timers_core(&c1, &t1); } // compute locally allocated range for the sub-block int start[3]; int end[3]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; if (!sb->owned) return; for (int n = 0; n < 3; n++) { start[n] = sb->decomp_disp[n]; end[n] = sb->decomp_disp[n] + sb->decomp_size[n]; if (start[n] >= range[2 * n]) { start[n] = 0; } else { start[n] = range[2 * n] - start[n]; } if (sb->id_m[n] == MPI_PROC_NULL && range[2 * n] < 0) start[n] = range[2 * n]; if (end[n] >= range[2 * n + 1]) { end[n] = range[2 * n + 1] - sb->decomp_disp[n]; } else { end[n] = sb->decomp_size[n]; } if (sb->id_p[n] == MPI_PROC_NULL && (range[2 * n + 1] > sb->decomp_disp[n] + sb->decomp_size[n])) end[n] += (range[2 * n + 1] - sb->decomp_disp[n] - sb->decomp_size[n]); } #else for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #endif int x_size = MAX(0, end[0] - start[0]); int y_size = MAX(0, end[1] - start[1]); int z_size = MAX(0, end[2] - start[2]); int xdim0 = args[0].dat->size[0]; int ydim0 = args[0].dat->size[1]; int xdim1 = args[1].dat->size[0]; int ydim1 = args[1].dat->size[1]; int xdim2 = args[2].dat->size[0]; int ydim2 = args[2].dat->size[1]; int xdim3 = args[3].dat->size[0]; int ydim3 = args[3].dat->size[1]; int xdim4 = args[4].dat->size[0]; int ydim4 = args[4].dat->size[1]; int xdim5 = args[5].dat->size[0]; int ydim5 = args[5].dat->size[1]; int xdim6 = args[6].dat->size[0]; int ydim6 = args[6].dat->size[1]; // build opencl kernel if not already built buildOpenCLKernels_update_halo_kernel1_b2(xdim0, ydim0, xdim1, ydim1, xdim2, ydim2, xdim3, ydim3, xdim4, ydim4, xdim5, ydim5, xdim6, ydim6); // set up OpenCL thread blocks size_t globalWorkSize[3] = { ((x_size - 1) / OPS_block_size_x + 1) * OPS_block_size_x, ((y_size - 1) / OPS_block_size_y + 1) * OPS_block_size_y, ((z_size - 1) / OPS_block_size_z + 1) * OPS_block_size_z}; size_t localWorkSize[3] = {OPS_block_size_x, OPS_block_size_y, OPS_block_size_z}; int *arg7h = (int *)arg7.data; int consts_bytes = 0; consts_bytes += ROUND_UP(NUM_FIELDS * sizeof(int)); reallocConstArrays(consts_bytes); consts_bytes = 0; arg7.data = OPS_consts_h + consts_bytes; arg7.data_d = OPS_consts_d + consts_bytes; for (int d = 0; d < NUM_FIELDS; d++) ((int *)arg7.data)[d] = arg7h[d]; consts_bytes += ROUND_UP(NUM_FIELDS * sizeof(int)); mvConstArraysToDevice(consts_bytes); // set up initial pointers int d_m[OPS_MAX_DIM]; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d] + OPS_sub_dat_list[args[0].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d]; #endif int base0 = 1 * 1 * (start[0] * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0 + args[0].dat->size[0] * 1 * (start[1] * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]); base0 = base0 + args[0].dat->size[0] * 1 * args[0].dat->size[1] * 1 * (start[2] * args[0].stencil->stride[2] - args[0].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d] + OPS_sub_dat_list[args[1].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d]; #endif int base1 = 1 * 1 * (start[0] * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); base1 = base1 + args[1].dat->size[0] * 1 * (start[1] * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]); base1 = base1 + args[1].dat->size[0] * 1 * args[1].dat->size[1] * 1 * (start[2] * args[1].stencil->stride[2] - args[1].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d] + OPS_sub_dat_list[args[2].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d]; #endif int base2 = 1 * 1 * (start[0] * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]); base2 = base2 + args[2].dat->size[0] * 1 * (start[1] * args[2].stencil->stride[1] - args[2].dat->base[1] - d_m[1]); base2 = base2 + args[2].dat->size[0] * 1 * args[2].dat->size[1] * 1 * (start[2] * args[2].stencil->stride[2] - args[2].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d] + OPS_sub_dat_list[args[3].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d]; #endif int base3 = 1 * 1 * (start[0] * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]); base3 = base3 + args[3].dat->size[0] * 1 * (start[1] * args[3].stencil->stride[1] - args[3].dat->base[1] - d_m[1]); base3 = base3 + args[3].dat->size[0] * 1 * args[3].dat->size[1] * 1 * (start[2] * args[3].stencil->stride[2] - args[3].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[4].dat->d_m[d] + OPS_sub_dat_list[args[4].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[4].dat->d_m[d]; #endif int base4 = 1 * 1 * (start[0] * args[4].stencil->stride[0] - args[4].dat->base[0] - d_m[0]); base4 = base4 + args[4].dat->size[0] * 1 * (start[1] * args[4].stencil->stride[1] - args[4].dat->base[1] - d_m[1]); base4 = base4 + args[4].dat->size[0] * 1 * args[4].dat->size[1] * 1 * (start[2] * args[4].stencil->stride[2] - args[4].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[5].dat->d_m[d] + OPS_sub_dat_list[args[5].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[5].dat->d_m[d]; #endif int base5 = 1 * 1 * (start[0] * args[5].stencil->stride[0] - args[5].dat->base[0] - d_m[0]); base5 = base5 + args[5].dat->size[0] * 1 * (start[1] * args[5].stencil->stride[1] - args[5].dat->base[1] - d_m[1]); base5 = base5 + args[5].dat->size[0] * 1 * args[5].dat->size[1] * 1 * (start[2] * args[5].stencil->stride[2] - args[5].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[6].dat->d_m[d] + OPS_sub_dat_list[args[6].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[6].dat->d_m[d]; #endif int base6 = 1 * 1 * (start[0] * args[6].stencil->stride[0] - args[6].dat->base[0] - d_m[0]); base6 = base6 + args[6].dat->size[0] * 1 * (start[1] * args[6].stencil->stride[1] - args[6].dat->base[1] - d_m[1]); base6 = base6 + args[6].dat->size[0] * 1 * args[6].dat->size[1] * 1 * (start[2] * args[6].stencil->stride[2] - args[6].dat->base[2] - d_m[2]); ops_H_D_exchanges_device(args, 8); ops_halo_exchanges(args, 8, range); ops_H_D_exchanges_device(args, 8); if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[12].mpi_time += t2 - t1; } if (globalWorkSize[0] > 0 && globalWorkSize[1] > 0 && globalWorkSize[2] > 0) { clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 0, sizeof(cl_mem), (void *)&arg0.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 1, sizeof(cl_mem), (void *)&arg1.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 2, sizeof(cl_mem), (void *)&arg2.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 3, sizeof(cl_mem), (void *)&arg3.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 4, sizeof(cl_mem), (void *)&arg4.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 5, sizeof(cl_mem), (void *)&arg5.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 6, sizeof(cl_mem), (void *)&arg6.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 7, sizeof(cl_mem), (void *)&arg7.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 8, sizeof(cl_int), (void *)&base0)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 9, sizeof(cl_int), (void *)&base1)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 10, sizeof(cl_int), (void *)&base2)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 11, sizeof(cl_int), (void *)&base3)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 12, sizeof(cl_int), (void *)&base4)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 13, sizeof(cl_int), (void *)&base5)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 14, sizeof(cl_int), (void *)&base6)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 15, sizeof(cl_int), (void *)&x_size)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 16, sizeof(cl_int), (void *)&y_size)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[12], 17, sizeof(cl_int), (void *)&z_size)); // call/enque opencl kernel wrapper function clSafeCall(clEnqueueNDRangeKernel( OPS_opencl_core.command_queue, OPS_opencl_core.kernel[12], 3, NULL, globalWorkSize, localWorkSize, 0, NULL, NULL)); } if (OPS_diags > 1) { clSafeCall(clFinish(OPS_opencl_core.command_queue)); } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[12].time += t1 - t2; } ops_set_dirtybit_device(args, 8); ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[1], range); ops_set_halo_dirtybit3(&args[2], range); ops_set_halo_dirtybit3(&args[3], range); ops_set_halo_dirtybit3(&args[4], range); ops_set_halo_dirtybit3(&args[5], range); ops_set_halo_dirtybit3(&args[6], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[12].mpi_time += t2 - t1; OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_update_halo_kernel2_xvel_minus_2_a(char const *name, ops_block Block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2) { ops_arg args[3] = { arg0, arg1, arg2}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args,3,range,56)) return; #endif ops_timing_realloc(56,"update_halo_kernel2_xvel_minus_2_a"); OPS_kernels[56].count++; //compute localy allocated range for the sub-block int start[2]; int end[2]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; if (!sb->owned) return; for ( int n=0; n<2; n++ ){ start[n] = sb->decomp_disp[n];end[n] = sb->decomp_disp[n]+sb->decomp_size[n]; if (start[n] >= range[2*n]) { start[n] = 0; } else { start[n] = range[2*n] - start[n]; } if (sb->id_m[n]==MPI_PROC_NULL && range[2*n] < 0) start[n] = range[2*n]; if (end[n] >= range[2*n+1]) { end[n] = range[2*n+1] - sb->decomp_disp[n]; } else { end[n] = sb->decomp_size[n]; } if (sb->id_p[n]==MPI_PROC_NULL && (range[2*n+1] > sb->decomp_disp[n]+sb->decomp_size[n])) end[n] += (range[2*n+1]-sb->decomp_disp[n]-sb->decomp_size[n]); } #else //OPS_MPI for ( int n=0; n<2; n++ ){ start[n] = range[2*n];end[n] = range[2*n+1]; } #endif //OPS_MPI int x_size = MAX(0,end[0]-start[0]); int y_size = MAX(0,end[1]-start[1]); xdim0 = args[0].dat->size[0]*args[0].dat->dim; xdim1 = args[1].dat->size[0]*args[1].dat->dim; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); if (xdim0 != xdim0_update_halo_kernel2_xvel_minus_2_a_h || xdim1 != xdim1_update_halo_kernel2_xvel_minus_2_a_h) { xdim0_update_halo_kernel2_xvel_minus_2_a = xdim0; xdim0_update_halo_kernel2_xvel_minus_2_a_h = xdim0; xdim1_update_halo_kernel2_xvel_minus_2_a = xdim1; xdim1_update_halo_kernel2_xvel_minus_2_a_h = xdim1; } int dat0 = args[0].dat->elem_size; int dat1 = args[1].dat->elem_size; int *arg2h = (int *)arg2.data; //Upload large globals int consts_bytes = 0; consts_bytes += ROUND_UP(NUM_FIELDS*sizeof(int)); reallocConstArrays(consts_bytes); consts_bytes = 0; args[2].data = OPS_consts_h + consts_bytes; args[2].data_d = OPS_consts_d + consts_bytes; for (int d=0; d<NUM_FIELDS; d++) ((int *)args[2].data)[d] = arg2h[d]; consts_bytes += ROUND_UP(NUM_FIELDS*sizeof(int)); mvConstArraysToDevice(consts_bytes); //set up initial pointers int d_m[OPS_MAX_DIM]; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d] + OPS_sub_dat_list[args[0].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d]; #endif //OPS_MPI int base0 = dat0 * 1 * (start[0] * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0+ dat0 * args[0].dat->size[0] * (start[1] * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]); #ifdef OPS_GPU double *p_a0 = (double *)((char *)args[0].data_d + base0); #else double *p_a0 = (double *)((char *)args[0].data + base0); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d] + OPS_sub_dat_list[args[1].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d]; #endif //OPS_MPI int base1 = dat1 * 1 * (start[0] * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); base1 = base1+ dat1 * args[1].dat->size[0] * (start[1] * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]); #ifdef OPS_GPU double *p_a1 = (double *)((char *)args[1].data_d + base1); #else double *p_a1 = (double *)((char *)args[1].data + base1); #endif #ifdef OPS_GPU int *p_a2 = (int *)args[2].data_d; #else int *p_a2 = arg2h; #endif #ifdef OPS_GPU ops_H_D_exchanges_device(args, 3); #else ops_H_D_exchanges_host(args, 3); #endif ops_halo_exchanges(args,3,range); ops_timers_core(&c1,&t1); OPS_kernels[56].mpi_time += t1-t2; update_halo_kernel2_xvel_minus_2_a_c_wrapper( p_a0, p_a1, p_a2, x_size, y_size); ops_timers_core(&c2,&t2); OPS_kernels[56].time += t2-t1; #ifdef OPS_GPU ops_set_dirtybit_device(args, 3); #else ops_set_dirtybit_host(args, 3); #endif ops_set_halo_dirtybit3(&args[0],range); ops_set_halo_dirtybit3(&args[1],range); //Update kernel record OPS_kernels[56].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[56].transfer += ops_compute_transfer(dim, range, &arg1); }
// host stub function void ops_par_loop_update_halo_kernel1_fr2(char const *name, ops_block Block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3, ops_arg arg4, ops_arg arg5, ops_arg arg6, ops_arg arg7) { ops_arg args[8] = { arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7}; ops_timing_realloc(51,"update_halo_kernel1_fr2"); OPS_kernels[51].count++; //compute localy allocated range for the sub-block int start[3]; int end[3]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; if (!sb->owned) return; for ( int n=0; n<3; n++ ){ start[n] = sb->decomp_disp[n];end[n] = sb->decomp_disp[n]+sb->decomp_size[n]; if (start[n] >= range[2*n]) { start[n] = 0; } else { start[n] = range[2*n] - start[n]; } if (sb->id_m[n]==MPI_PROC_NULL && range[2*n] < 0) start[n] = range[2*n]; if (end[n] >= range[2*n+1]) { end[n] = range[2*n+1] - sb->decomp_disp[n]; } else { end[n] = sb->decomp_size[n]; } if (sb->id_p[n]==MPI_PROC_NULL && (range[2*n+1] > sb->decomp_disp[n]+sb->decomp_size[n])) end[n] += (range[2*n+1]-sb->decomp_disp[n]-sb->decomp_size[n]); } #else //OPS_MPI for ( int n=0; n<3; n++ ){ start[n] = range[2*n];end[n] = range[2*n+1]; } #endif //OPS_MPI int x_size = MAX(0,end[0]-start[0]); int y_size = MAX(0,end[1]-start[1]); int z_size = MAX(0,end[2]-start[2]); xdim0 = args[0].dat->size[0]*args[0].dat->dim; ydim0 = args[0].dat->size[1]; xdim1 = args[1].dat->size[0]*args[1].dat->dim; ydim1 = args[1].dat->size[1]; xdim2 = args[2].dat->size[0]*args[2].dat->dim; ydim2 = args[2].dat->size[1]; xdim3 = args[3].dat->size[0]*args[3].dat->dim; ydim3 = args[3].dat->size[1]; xdim4 = args[4].dat->size[0]*args[4].dat->dim; ydim4 = args[4].dat->size[1]; xdim5 = args[5].dat->size[0]*args[5].dat->dim; ydim5 = args[5].dat->size[1]; xdim6 = args[6].dat->size[0]*args[6].dat->dim; ydim6 = args[6].dat->size[1]; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); if (xdim0 != xdim0_update_halo_kernel1_fr2_h || ydim0 != ydim0_update_halo_kernel1_fr2_h || xdim1 != xdim1_update_halo_kernel1_fr2_h || ydim1 != ydim1_update_halo_kernel1_fr2_h || xdim2 != xdim2_update_halo_kernel1_fr2_h || ydim2 != ydim2_update_halo_kernel1_fr2_h || xdim3 != xdim3_update_halo_kernel1_fr2_h || ydim3 != ydim3_update_halo_kernel1_fr2_h || xdim4 != xdim4_update_halo_kernel1_fr2_h || ydim4 != ydim4_update_halo_kernel1_fr2_h || xdim5 != xdim5_update_halo_kernel1_fr2_h || ydim5 != ydim5_update_halo_kernel1_fr2_h || xdim6 != xdim6_update_halo_kernel1_fr2_h || ydim6 != ydim6_update_halo_kernel1_fr2_h) { xdim0_update_halo_kernel1_fr2 = xdim0; xdim0_update_halo_kernel1_fr2_h = xdim0; ydim0_update_halo_kernel1_fr2 = ydim0; ydim0_update_halo_kernel1_fr2_h = ydim0; xdim1_update_halo_kernel1_fr2 = xdim1; xdim1_update_halo_kernel1_fr2_h = xdim1; ydim1_update_halo_kernel1_fr2 = ydim1; ydim1_update_halo_kernel1_fr2_h = ydim1; xdim2_update_halo_kernel1_fr2 = xdim2; xdim2_update_halo_kernel1_fr2_h = xdim2; ydim2_update_halo_kernel1_fr2 = ydim2; ydim2_update_halo_kernel1_fr2_h = ydim2; xdim3_update_halo_kernel1_fr2 = xdim3; xdim3_update_halo_kernel1_fr2_h = xdim3; ydim3_update_halo_kernel1_fr2 = ydim3; ydim3_update_halo_kernel1_fr2_h = ydim3; xdim4_update_halo_kernel1_fr2 = xdim4; xdim4_update_halo_kernel1_fr2_h = xdim4; ydim4_update_halo_kernel1_fr2 = ydim4; ydim4_update_halo_kernel1_fr2_h = ydim4; xdim5_update_halo_kernel1_fr2 = xdim5; xdim5_update_halo_kernel1_fr2_h = xdim5; ydim5_update_halo_kernel1_fr2 = ydim5; ydim5_update_halo_kernel1_fr2_h = ydim5; xdim6_update_halo_kernel1_fr2 = xdim6; xdim6_update_halo_kernel1_fr2_h = xdim6; ydim6_update_halo_kernel1_fr2 = ydim6; ydim6_update_halo_kernel1_fr2_h = ydim6; } int dat0 = args[0].dat->elem_size; int dat1 = args[1].dat->elem_size; int dat2 = args[2].dat->elem_size; int dat3 = args[3].dat->elem_size; int dat4 = args[4].dat->elem_size; int dat5 = args[5].dat->elem_size; int dat6 = args[6].dat->elem_size; int *arg7h = (int *)arg7.data; //Upload large globals int consts_bytes = 0; consts_bytes += ROUND_UP(NUM_FIELDS*sizeof(int)); reallocConstArrays(consts_bytes); consts_bytes = 0; args[7].data = OPS_consts_h + consts_bytes; args[7].data_d = OPS_consts_d + consts_bytes; for (int d=0; d<NUM_FIELDS; d++) ((int *)args[7].data)[d] = arg7h[d]; consts_bytes += ROUND_UP(NUM_FIELDS*sizeof(int)); mvConstArraysToDevice(consts_bytes); //set up initial pointers int d_m[OPS_MAX_DIM]; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d] + OPS_sub_dat_list[args[0].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[0].dat->d_m[d]; #endif //OPS_MPI int base0 = dat0 * 1 * (start[0] * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0+ dat0 * args[0].dat->size[0] * (start[1] * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]); base0 = base0+ dat0 * args[0].dat->size[0] * args[0].dat->size[1] * (start[2] * args[0].stencil->stride[2] - args[0].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a0 = (double *)((char *)args[0].data_d + base0); #else double *p_a0 = (double *)((char *)args[0].data + base0); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d] + OPS_sub_dat_list[args[1].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[1].dat->d_m[d]; #endif //OPS_MPI int base1 = dat1 * 1 * (start[0] * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); base1 = base1+ dat1 * args[1].dat->size[0] * (start[1] * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]); base1 = base1+ dat1 * args[1].dat->size[0] * args[1].dat->size[1] * (start[2] * args[1].stencil->stride[2] - args[1].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a1 = (double *)((char *)args[1].data_d + base1); #else double *p_a1 = (double *)((char *)args[1].data + base1); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d] + OPS_sub_dat_list[args[2].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[2].dat->d_m[d]; #endif //OPS_MPI int base2 = dat2 * 1 * (start[0] * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]); base2 = base2+ dat2 * args[2].dat->size[0] * (start[1] * args[2].stencil->stride[1] - args[2].dat->base[1] - d_m[1]); base2 = base2+ dat2 * args[2].dat->size[0] * args[2].dat->size[1] * (start[2] * args[2].stencil->stride[2] - args[2].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a2 = (double *)((char *)args[2].data_d + base2); #else double *p_a2 = (double *)((char *)args[2].data + base2); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d] + OPS_sub_dat_list[args[3].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[3].dat->d_m[d]; #endif //OPS_MPI int base3 = dat3 * 1 * (start[0] * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]); base3 = base3+ dat3 * args[3].dat->size[0] * (start[1] * args[3].stencil->stride[1] - args[3].dat->base[1] - d_m[1]); base3 = base3+ dat3 * args[3].dat->size[0] * args[3].dat->size[1] * (start[2] * args[3].stencil->stride[2] - args[3].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a3 = (double *)((char *)args[3].data_d + base3); #else double *p_a3 = (double *)((char *)args[3].data + base3); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[4].dat->d_m[d] + OPS_sub_dat_list[args[4].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[4].dat->d_m[d]; #endif //OPS_MPI int base4 = dat4 * 1 * (start[0] * args[4].stencil->stride[0] - args[4].dat->base[0] - d_m[0]); base4 = base4+ dat4 * args[4].dat->size[0] * (start[1] * args[4].stencil->stride[1] - args[4].dat->base[1] - d_m[1]); base4 = base4+ dat4 * args[4].dat->size[0] * args[4].dat->size[1] * (start[2] * args[4].stencil->stride[2] - args[4].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a4 = (double *)((char *)args[4].data_d + base4); #else double *p_a4 = (double *)((char *)args[4].data + base4); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[5].dat->d_m[d] + OPS_sub_dat_list[args[5].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[5].dat->d_m[d]; #endif //OPS_MPI int base5 = dat5 * 1 * (start[0] * args[5].stencil->stride[0] - args[5].dat->base[0] - d_m[0]); base5 = base5+ dat5 * args[5].dat->size[0] * (start[1] * args[5].stencil->stride[1] - args[5].dat->base[1] - d_m[1]); base5 = base5+ dat5 * args[5].dat->size[0] * args[5].dat->size[1] * (start[2] * args[5].stencil->stride[2] - args[5].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a5 = (double *)((char *)args[5].data_d + base5); #else double *p_a5 = (double *)((char *)args[5].data + base5); #endif #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[6].dat->d_m[d] + OPS_sub_dat_list[args[6].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[6].dat->d_m[d]; #endif //OPS_MPI int base6 = dat6 * 1 * (start[0] * args[6].stencil->stride[0] - args[6].dat->base[0] - d_m[0]); base6 = base6+ dat6 * args[6].dat->size[0] * (start[1] * args[6].stencil->stride[1] - args[6].dat->base[1] - d_m[1]); base6 = base6+ dat6 * args[6].dat->size[0] * args[6].dat->size[1] * (start[2] * args[6].stencil->stride[2] - args[6].dat->base[2] - d_m[2]); #ifdef OPS_GPU double *p_a6 = (double *)((char *)args[6].data_d + base6); #else double *p_a6 = (double *)((char *)args[6].data + base6); #endif #ifdef OPS_GPU int *p_a7 = (int *)args[7].data_d; #else int *p_a7 = arg7h; #endif #ifdef OPS_GPU ops_H_D_exchanges_device(args, 8); #else ops_H_D_exchanges_host(args, 8); #endif ops_halo_exchanges(args,8,range); ops_timers_core(&c1,&t1); OPS_kernels[51].mpi_time += t1-t2; update_halo_kernel1_fr2_c_wrapper( p_a0, p_a1, p_a2, p_a3, p_a4, p_a5, p_a6, p_a7, x_size, y_size, z_size); ops_timers_core(&c2,&t2); OPS_kernels[51].time += t2-t1; #ifdef OPS_GPU ops_set_dirtybit_device(args, 8); #else ops_set_dirtybit_host(args, 8); #endif ops_set_halo_dirtybit3(&args[0],range); ops_set_halo_dirtybit3(&args[1],range); ops_set_halo_dirtybit3(&args[2],range); ops_set_halo_dirtybit3(&args[3],range); ops_set_halo_dirtybit3(&args[4],range); ops_set_halo_dirtybit3(&args[5],range); ops_set_halo_dirtybit3(&args[6],range); //Update kernel record OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg1); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg2); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg3); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg4); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg5); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg6); }
// host stub function void ops_par_loop_update_halo_kernel2_zvel_plus_4_right( char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1, ops_arg arg2) { // Timing double t1, t2, c1, c2; ops_arg args[3] = {arg0, arg1, arg2}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 3, range, 53)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(53, "update_halo_kernel2_zvel_plus_4_right"); OPS_kernels[53].count++; ops_timers_core(&c1, &t1); } // compute localy allocated range for the sub-block int start[3]; int end[3]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // OPS_MPI int arg_idx[3]; int arg_idx_base[3]; #ifdef OPS_MPI if (compute_ranges(args, 3, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif for (int n = 0; n < 3; n++) { arg_idx_base[n] = arg_idx[n]; } int dat0 = args[0].dat->elem_size; int dat1 = args[1].dat->elem_size; int *arg2h = (int *)arg2.data; // Upload large globals #ifdef OPS_GPU int consts_bytes = 0; consts_bytes += ROUND_UP(NUM_FIELDS * sizeof(int)); reallocConstArrays(consts_bytes); consts_bytes = 0; args[2].data = OPS_consts_h + consts_bytes; args[2].data_d = OPS_consts_d + consts_bytes; for (int d = 0; d < NUM_FIELDS; d++) ((int *)args[2].data)[d] = arg2h[d]; consts_bytes += ROUND_UP(NUM_FIELDS * sizeof(int)); mvConstArraysToDevice(consts_bytes); #endif // OPS_GPU // set up initial pointers int base0 = args[0].dat->base_offset + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * start[0] * args[0].stencil->stride[0]; base0 = base0 + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * args[0].dat->size[0] * start[1] * args[0].stencil->stride[1]; base0 = base0 + (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size) * args[0].dat->size[0] * args[0].dat->size[1] * start[2] * args[0].stencil->stride[2]; #ifdef OPS_GPU double *p_a0 = (double *)((char *)args[0].data_d + base0); #else double *p_a0 = (double *)((char *)args[0].data + base0); #endif int base1 = args[1].dat->base_offset + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * start[0] * args[1].stencil->stride[0]; base1 = base1 + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * args[1].dat->size[0] * start[1] * args[1].stencil->stride[1]; base1 = base1 + (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size) * args[1].dat->size[0] * args[1].dat->size[1] * start[2] * args[1].stencil->stride[2]; #ifdef OPS_GPU double *p_a1 = (double *)((char *)args[1].data_d + base1); #else double *p_a1 = (double *)((char *)args[1].data + base1); #endif #ifdef OPS_GPU int *p_a2 = (int *)args[2].data_d; #else int *p_a2 = arg2h; #endif int x_size = MAX(0, end[0] - start[0]); int y_size = MAX(0, end[1] - start[1]); int z_size = MAX(0, end[2] - start[2]); // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; ydim0 = args[0].dat->size[1]; xdim1 = args[1].dat->size[0]; ydim1 = args[1].dat->size[1]; if (xdim0 != xdim0_update_halo_kernel2_zvel_plus_4_right_h || ydim0 != ydim0_update_halo_kernel2_zvel_plus_4_right_h || xdim1 != xdim1_update_halo_kernel2_zvel_plus_4_right_h || ydim1 != ydim1_update_halo_kernel2_zvel_plus_4_right_h) { xdim0_update_halo_kernel2_zvel_plus_4_right = xdim0; xdim0_update_halo_kernel2_zvel_plus_4_right_h = xdim0; ydim0_update_halo_kernel2_zvel_plus_4_right = ydim0; ydim0_update_halo_kernel2_zvel_plus_4_right_h = ydim0; xdim1_update_halo_kernel2_zvel_plus_4_right = xdim1; xdim1_update_halo_kernel2_zvel_plus_4_right_h = xdim1; ydim1_update_halo_kernel2_zvel_plus_4_right = ydim1; ydim1_update_halo_kernel2_zvel_plus_4_right_h = ydim1; } // Halo Exchanges #ifdef OPS_GPU ops_H_D_exchanges_device(args, 3); #else ops_H_D_exchanges_host(args, 3); #endif ops_halo_exchanges(args, 3, range); #ifdef OPS_GPU ops_H_D_exchanges_device(args, 3); #else ops_H_D_exchanges_host(args, 3); #endif if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[53].mpi_time += t2 - t1; } update_halo_kernel2_zvel_plus_4_right_c_wrapper(p_a0, p_a1, p_a2, x_size, y_size, z_size); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[53].time += t1 - t2; } #ifdef OPS_GPU ops_set_dirtybit_device(args, 3); #else ops_set_dirtybit_host(args, 3); #endif ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[1], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[53].mpi_time += t2 - t1; OPS_kernels[53].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[53].transfer += ops_compute_transfer(dim, start, end, &arg1); } }