// host stub function void ops_par_loop_update_halo_kernel2_yvel_minus_4_top(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(67,"update_halo_kernel2_yvel_minus_4_top"); OPS_kernels[67].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]; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); if (xdim0 != xdim0_update_halo_kernel2_yvel_minus_4_top_h || ydim0 != ydim0_update_halo_kernel2_yvel_minus_4_top_h || xdim1 != xdim1_update_halo_kernel2_yvel_minus_4_top_h || ydim1 != ydim1_update_halo_kernel2_yvel_minus_4_top_h) { xdim0_update_halo_kernel2_yvel_minus_4_top = xdim0; xdim0_update_halo_kernel2_yvel_minus_4_top_h = xdim0; ydim0_update_halo_kernel2_yvel_minus_4_top = ydim0; ydim0_update_halo_kernel2_yvel_minus_4_top_h = ydim0; xdim1_update_halo_kernel2_yvel_minus_4_top = xdim1; xdim1_update_halo_kernel2_yvel_minus_4_top_h = xdim1; ydim1_update_halo_kernel2_yvel_minus_4_top = ydim1; ydim1_update_halo_kernel2_yvel_minus_4_top_h = ydim1; } 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]); 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_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[67].mpi_time += t1-t2; update_halo_kernel2_yvel_minus_4_top_c_wrapper( p_a0, p_a1, p_a2, x_size, y_size, z_size); ops_timers_core(&c2,&t2); OPS_kernels[67].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[67].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[67].transfer += ops_compute_transfer(dim, range, &arg1); }
// host stub function void ops_par_loop_update_halo_kernel3_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}; ops_timing_realloc(94,"update_halo_kernel3_minus_2_a"); OPS_kernels[94].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_kernel3_minus_2_a( 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[94].mpi_time += t1-t2; clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 0, sizeof(cl_mem), (void*) &arg0.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 1, sizeof(cl_mem), (void*) &arg1.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 2, sizeof(cl_mem), (void*) &arg2.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 3, sizeof(cl_int), (void*) &base0 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 4, sizeof(cl_int), (void*) &base1 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 5, sizeof(cl_int), (void*) &x_size )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 6, sizeof(cl_int), (void*) &y_size )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[94], 7, sizeof(cl_int), (void*) &z_size )); //call/enque opencl kernel wrapper function clSafeCall( clEnqueueNDRangeKernel(OPS_opencl_core.command_queue, OPS_opencl_core.kernel[94], 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[94].time += t2-t1; OPS_kernels[94].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[94].transfer += ops_compute_transfer(dim, range, &arg1); }
// host stub function void ops_par_loop_advec_cell_kernel4_xdir( 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 arg8, ops_arg arg9, ops_arg arg10) { // Timing double t1, t2, c1, c2; char *p_a[11]; int offs[11][3]; ops_arg args[11] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9, arg10}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 11, range, 111)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(111, "advec_cell_kernel4_xdir"); OPS_kernels[111].count++; ops_timers_core(&c2, &t2); } // 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]; #endif #ifdef OPS_DEBUG ops_register_args(args, "advec_cell_kernel4_xdir"); #endif int arg_idx[3]; int arg_idx_base[3]; #ifdef OPS_MPI if (compute_ranges(args, 11, 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 // OPS_MPI for (int n = 0; n < 3; n++) { arg_idx_base[n] = arg_idx[n]; } offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[0][1] = off3D(1, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[0][2] = off3D(2, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][1] - offs[0][0]; offs[1][0] = args[1].stencil->stride[0] * 1; // unit step in x dimension offs[1][1] = off3D(1, &start[0], &end[0], args[1].dat->size, args[1].stencil->stride) - offs[1][0]; offs[1][2] = off3D(2, &start[0], &end[0], args[1].dat->size, args[1].stencil->stride) - offs[1][1] - offs[1][0]; offs[2][0] = args[2].stencil->stride[0] * 1; // unit step in x dimension offs[2][1] = off3D(1, &start[0], &end[0], args[2].dat->size, args[2].stencil->stride) - offs[2][0]; offs[2][2] = off3D(2, &start[0], &end[0], args[2].dat->size, args[2].stencil->stride) - offs[2][1] - offs[2][0]; offs[3][0] = args[3].stencil->stride[0] * 1; // unit step in x dimension offs[3][1] = off3D(1, &start[0], &end[0], args[3].dat->size, args[3].stencil->stride) - offs[3][0]; offs[3][2] = off3D(2, &start[0], &end[0], args[3].dat->size, args[3].stencil->stride) - offs[3][1] - offs[3][0]; offs[4][0] = args[4].stencil->stride[0] * 1; // unit step in x dimension offs[4][1] = off3D(1, &start[0], &end[0], args[4].dat->size, args[4].stencil->stride) - offs[4][0]; offs[4][2] = off3D(2, &start[0], &end[0], args[4].dat->size, args[4].stencil->stride) - offs[4][1] - offs[4][0]; offs[5][0] = args[5].stencil->stride[0] * 1; // unit step in x dimension offs[5][1] = off3D(1, &start[0], &end[0], args[5].dat->size, args[5].stencil->stride) - offs[5][0]; offs[5][2] = off3D(2, &start[0], &end[0], args[5].dat->size, args[5].stencil->stride) - offs[5][1] - offs[5][0]; offs[6][0] = args[6].stencil->stride[0] * 1; // unit step in x dimension offs[6][1] = off3D(1, &start[0], &end[0], args[6].dat->size, args[6].stencil->stride) - offs[6][0]; offs[6][2] = off3D(2, &start[0], &end[0], args[6].dat->size, args[6].stencil->stride) - offs[6][1] - offs[6][0]; offs[7][0] = args[7].stencil->stride[0] * 1; // unit step in x dimension offs[7][1] = off3D(1, &start[0], &end[0], args[7].dat->size, args[7].stencil->stride) - offs[7][0]; offs[7][2] = off3D(2, &start[0], &end[0], args[7].dat->size, args[7].stencil->stride) - offs[7][1] - offs[7][0]; offs[8][0] = args[8].stencil->stride[0] * 1; // unit step in x dimension offs[8][1] = off3D(1, &start[0], &end[0], args[8].dat->size, args[8].stencil->stride) - offs[8][0]; offs[8][2] = off3D(2, &start[0], &end[0], args[8].dat->size, args[8].stencil->stride) - offs[8][1] - offs[8][0]; offs[9][0] = args[9].stencil->stride[0] * 1; // unit step in x dimension offs[9][1] = off3D(1, &start[0], &end[0], args[9].dat->size, args[9].stencil->stride) - offs[9][0]; offs[9][2] = off3D(2, &start[0], &end[0], args[9].dat->size, args[9].stencil->stride) - offs[9][1] - offs[9][0]; offs[10][0] = args[10].stencil->stride[0] * 1; // unit step in x dimension offs[10][1] = off3D(1, &start[0], &end[0], args[10].dat->size, args[10].stencil->stride) - offs[10][0]; offs[10][2] = off3D(2, &start[0], &end[0], args[10].dat->size, args[10].stencil->stride) - offs[10][1] - offs[10][0]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int off0_2 = offs[0][2]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int off1_2 = offs[1][2]; int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); int off2_0 = offs[2][0]; int off2_1 = offs[2][1]; int off2_2 = offs[2][2]; int dat2 = (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size); int off3_0 = offs[3][0]; int off3_1 = offs[3][1]; int off3_2 = offs[3][2]; int dat3 = (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size); int off4_0 = offs[4][0]; int off4_1 = offs[4][1]; int off4_2 = offs[4][2]; int dat4 = (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size); int off5_0 = offs[5][0]; int off5_1 = offs[5][1]; int off5_2 = offs[5][2]; int dat5 = (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size); int off6_0 = offs[6][0]; int off6_1 = offs[6][1]; int off6_2 = offs[6][2]; int dat6 = (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size); int off7_0 = offs[7][0]; int off7_1 = offs[7][1]; int off7_2 = offs[7][2]; int dat7 = (OPS_soa ? args[7].dat->type_size : args[7].dat->elem_size); int off8_0 = offs[8][0]; int off8_1 = offs[8][1]; int off8_2 = offs[8][2]; int dat8 = (OPS_soa ? args[8].dat->type_size : args[8].dat->elem_size); int off9_0 = offs[9][0]; int off9_1 = offs[9][1]; int off9_2 = offs[9][2]; int dat9 = (OPS_soa ? args[9].dat->type_size : args[9].dat->elem_size); int off10_0 = offs[10][0]; int off10_1 = offs[10][1]; int off10_2 = offs[10][2]; int dat10 = (OPS_soa ? args[10].dat->type_size : args[10].dat->elem_size); // set up initial pointers and exchange halos if necessary 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]; p_a[0] = (char *)args[0].data + base0; 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]; p_a[1] = (char *)args[1].data + base1; 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]; p_a[2] = (char *)args[2].data + base2; 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]; p_a[3] = (char *)args[3].data + base3; 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]; p_a[4] = (char *)args[4].data + base4; 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]; p_a[5] = (char *)args[5].data + base5; 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]; p_a[6] = (char *)args[6].data + base6; int base7 = args[7].dat->base_offset + (OPS_soa ? args[7].dat->type_size : args[7].dat->elem_size) * start[0] * args[7].stencil->stride[0]; base7 = base7 + (OPS_soa ? args[7].dat->type_size : args[7].dat->elem_size) * args[7].dat->size[0] * start[1] * args[7].stencil->stride[1]; base7 = base7 + (OPS_soa ? args[7].dat->type_size : args[7].dat->elem_size) * args[7].dat->size[0] * args[7].dat->size[1] * start[2] * args[7].stencil->stride[2]; p_a[7] = (char *)args[7].data + base7; int base8 = args[8].dat->base_offset + (OPS_soa ? args[8].dat->type_size : args[8].dat->elem_size) * start[0] * args[8].stencil->stride[0]; base8 = base8 + (OPS_soa ? args[8].dat->type_size : args[8].dat->elem_size) * args[8].dat->size[0] * start[1] * args[8].stencil->stride[1]; base8 = base8 + (OPS_soa ? args[8].dat->type_size : args[8].dat->elem_size) * args[8].dat->size[0] * args[8].dat->size[1] * start[2] * args[8].stencil->stride[2]; p_a[8] = (char *)args[8].data + base8; int base9 = args[9].dat->base_offset + (OPS_soa ? args[9].dat->type_size : args[9].dat->elem_size) * start[0] * args[9].stencil->stride[0]; base9 = base9 + (OPS_soa ? args[9].dat->type_size : args[9].dat->elem_size) * args[9].dat->size[0] * start[1] * args[9].stencil->stride[1]; base9 = base9 + (OPS_soa ? args[9].dat->type_size : args[9].dat->elem_size) * args[9].dat->size[0] * args[9].dat->size[1] * start[2] * args[9].stencil->stride[2]; p_a[9] = (char *)args[9].data + base9; int base10 = args[10].dat->base_offset + (OPS_soa ? args[10].dat->type_size : args[10].dat->elem_size) * start[0] * args[10].stencil->stride[0]; base10 = base10 + (OPS_soa ? args[10].dat->type_size : args[10].dat->elem_size) * args[10].dat->size[0] * start[1] * args[10].stencil->stride[1]; base10 = base10 + (OPS_soa ? args[10].dat->type_size : args[10].dat->elem_size) * args[10].dat->size[0] * args[10].dat->size[1] * start[2] * args[10].stencil->stride[2]; p_a[10] = (char *)args[10].data + base10; // 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]; xdim7 = args[7].dat->size[0]; ydim7 = args[7].dat->size[1]; xdim8 = args[8].dat->size[0]; ydim8 = args[8].dat->size[1]; xdim9 = args[9].dat->size[0]; ydim9 = args[9].dat->size[1]; xdim10 = args[10].dat->size[0]; ydim10 = args[10].dat->size[1]; // Halo Exchanges ops_H_D_exchanges_host(args, 11); ops_halo_exchanges(args, 11, range); ops_H_D_exchanges_host(args, 11); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[111].mpi_time += t1 - t2; } int n_x; for (int n_z = start[2]; n_z < end[2]; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { #pragma novector for (n_x = start[0]; n_x < start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x += SIMD_VEC) { // call kernel function, passing in pointers to data -vectorised #pragma simd for (int i = 0; i < SIMD_VEC; i++) { advec_cell_kernel4_xdir( (double *)p_a[0] + i * 1 * 1, (double *)p_a[1] + i * 1 * 1, (double *)p_a[2] + i * 1 * 1, (double *)p_a[3] + i * 1 * 1, (double *)p_a[4] + i * 1 * 1, (double *)p_a[5] + i * 1 * 1, (double *)p_a[6] + i * 1 * 1, (double *)p_a[7] + i * 1 * 1, (double *)p_a[8] + i * 1 * 1, (double *)p_a[9] + i * 1 * 1, (double *)p_a[10] + i * 1 * 1); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC; p_a[2] = p_a[2] + (dat2 * off2_0) * SIMD_VEC; p_a[3] = p_a[3] + (dat3 * off3_0) * SIMD_VEC; p_a[4] = p_a[4] + (dat4 * off4_0) * SIMD_VEC; p_a[5] = p_a[5] + (dat5 * off5_0) * SIMD_VEC; p_a[6] = p_a[6] + (dat6 * off6_0) * SIMD_VEC; p_a[7] = p_a[7] + (dat7 * off7_0) * SIMD_VEC; p_a[8] = p_a[8] + (dat8 * off8_0) * SIMD_VEC; p_a[9] = p_a[9] + (dat9 * off9_0) * SIMD_VEC; p_a[10] = p_a[10] + (dat10 * off10_0) * SIMD_VEC; } for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x < end[0]; n_x++) { // call kernel function, passing in pointers to data - remainder advec_cell_kernel4_xdir( (double *)p_a[0], (double *)p_a[1], (double *)p_a[2], (double *)p_a[3], (double *)p_a[4], (double *)p_a[5], (double *)p_a[6], (double *)p_a[7], (double *)p_a[8], (double *)p_a[9], (double *)p_a[10]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); p_a[1] = p_a[1] + (dat1 * off1_0); p_a[2] = p_a[2] + (dat2 * off2_0); p_a[3] = p_a[3] + (dat3 * off3_0); p_a[4] = p_a[4] + (dat4 * off4_0); p_a[5] = p_a[5] + (dat5 * off5_0); p_a[6] = p_a[6] + (dat6 * off6_0); p_a[7] = p_a[7] + (dat7 * off7_0); p_a[8] = p_a[8] + (dat8 * off8_0); p_a[9] = p_a[9] + (dat9 * off9_0); p_a[10] = p_a[10] + (dat10 * off10_0); } // shift pointers to data y direction p_a[0] = p_a[0] + (dat0 * off0_1); p_a[1] = p_a[1] + (dat1 * off1_1); p_a[2] = p_a[2] + (dat2 * off2_1); p_a[3] = p_a[3] + (dat3 * off3_1); p_a[4] = p_a[4] + (dat4 * off4_1); p_a[5] = p_a[5] + (dat5 * off5_1); p_a[6] = p_a[6] + (dat6 * off6_1); p_a[7] = p_a[7] + (dat7 * off7_1); p_a[8] = p_a[8] + (dat8 * off8_1); p_a[9] = p_a[9] + (dat9 * off9_1); p_a[10] = p_a[10] + (dat10 * off10_1); } // shift pointers to data z direction p_a[0] = p_a[0] + (dat0 * off0_2); p_a[1] = p_a[1] + (dat1 * off1_2); p_a[2] = p_a[2] + (dat2 * off2_2); p_a[3] = p_a[3] + (dat3 * off3_2); p_a[4] = p_a[4] + (dat4 * off4_2); p_a[5] = p_a[5] + (dat5 * off5_2); p_a[6] = p_a[6] + (dat6 * off6_2); p_a[7] = p_a[7] + (dat7 * off7_2); p_a[8] = p_a[8] + (dat8 * off8_2); p_a[9] = p_a[9] + (dat9 * off9_2); p_a[10] = p_a[10] + (dat10 * off10_2); } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[111].time += t2 - t1; } ops_set_dirtybit_host(args, 11); ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[1], range); ops_set_halo_dirtybit3(&args[6], range); ops_set_halo_dirtybit3(&args[7], range); ops_set_halo_dirtybit3(&args[8], range); ops_set_halo_dirtybit3(&args[9], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[111].mpi_time += t1 - t2; OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg6); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg7); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg8); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg9); OPS_kernels[111].transfer += ops_compute_transfer(dim, start, end, &arg10); } }
// host stub function void ops_par_loop_update_halo_kernel2_zvel_minus_2_back( 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, 57)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(57, "update_halo_kernel2_zvel_minus_2_back"); OPS_kernels[57].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_minus_2_back_h || ydim0 != ydim0_update_halo_kernel2_zvel_minus_2_back_h || xdim1 != xdim1_update_halo_kernel2_zvel_minus_2_back_h || ydim1 != ydim1_update_halo_kernel2_zvel_minus_2_back_h) { xdim0_update_halo_kernel2_zvel_minus_2_back = xdim0; xdim0_update_halo_kernel2_zvel_minus_2_back_h = xdim0; ydim0_update_halo_kernel2_zvel_minus_2_back = ydim0; ydim0_update_halo_kernel2_zvel_minus_2_back_h = ydim0; xdim1_update_halo_kernel2_zvel_minus_2_back = xdim1; xdim1_update_halo_kernel2_zvel_minus_2_back_h = xdim1; ydim1_update_halo_kernel2_zvel_minus_2_back = ydim1; ydim1_update_halo_kernel2_zvel_minus_2_back_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[57].mpi_time += t2 - t1; } update_halo_kernel2_zvel_minus_2_back_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[57].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[57].mpi_time += t2 - t1; OPS_kernels[57].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[57].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_update_halo_kernel2_yvel_plus_2_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; int offs[3][3]; ops_arg args[3] = {arg0, arg1, arg2}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 3, range, 43)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(43, "update_halo_kernel2_yvel_plus_2_right"); OPS_kernels[43].count++; ops_timers_core(&c1, &t1); } #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // compute locally allocated range for the sub-block int start[3]; int end[3]; int arg_idx[3]; #ifdef OPS_MPI 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]); if (end[n] < start[n]) end[n] = start[n]; } #else for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #endif #ifdef OPS_DEBUG ops_register_args(args, "update_halo_kernel2_yvel_plus_2_right"); #endif offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[0][1] = off3D(1, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[0][2] = off3D(2, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][1] - offs[0][0]; offs[1][0] = args[1].stencil->stride[0] * 1; // unit step in x dimension offs[1][1] = off3D(1, &start[0], &end[0], args[1].dat->size, args[1].stencil->stride) - offs[1][0]; offs[1][2] = off3D(2, &start[0], &end[0], args[1].dat->size, args[1].stencil->stride) - offs[1][1] - offs[1][0]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int off0_2 = offs[0][2]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int off1_2 = offs[1][2]; int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); // Halo Exchanges ops_H_D_exchanges_host(args, 3); ops_halo_exchanges(args, 3, range); ops_H_D_exchanges_host(args, 3); #ifdef _OPENMP int nthreads = omp_get_max_threads(); #else int nthreads = 1; #endif 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 (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[43].mpi_time += t2 - t1; } #pragma omp parallel for for (int thr = 0; thr < nthreads; thr++) { int z_size = end[2] - start[2]; char *p_a[3]; int start_i = start[2] + ((z_size - 1) / nthreads + 1) * thr; int finish_i = start[2] + MIN(((z_size - 1) / nthreads + 1) * (thr + 1), z_size); // get address per thread int start0 = start[0]; int start1 = start[1]; int start2 = start_i; // 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 = dat0 * 1 * (start0 * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0 + dat0 * args[0].dat->size[0] * (start1 * 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] * (start2 * args[0].stencil->stride[2] - args[0].dat->base[2] - d_m[2]); p_a[0] = (char *)args[0].data + base0; #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 = dat1 * 1 * (start0 * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); base1 = base1 + dat1 * args[1].dat->size[0] * (start1 * 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] * (start2 * args[1].stencil->stride[2] - args[1].dat->base[2] - d_m[2]); p_a[1] = (char *)args[1].data + base1; p_a[2] = (char *)args[2].data; for (int n_z = start_i; n_z < finish_i; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { for (int n_x = start[0]; n_x < start[0] + (end[0] - start[0]) / SIMD_VEC; n_x++) { // call kernel function, passing in pointers to data -vectorised #pragma simd for (int i = 0; i < SIMD_VEC; i++) { update_halo_kernel2_yvel_plus_2_right((double *)p_a[0] + i * 1 * 1, (double *)p_a[1] + i * 1 * 1, (int *)p_a[2]); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC; } for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x < end[0]; n_x++) { // call kernel function, passing in pointers to data - remainder update_halo_kernel2_yvel_plus_2_right( (double *)p_a[0], (double *)p_a[1], (int *)p_a[2]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); p_a[1] = p_a[1] + (dat1 * off1_0); } // shift pointers to data y direction p_a[0] = p_a[0] + (dat0 * off0_1); p_a[1] = p_a[1] + (dat1 * off1_1); } // shift pointers to data z direction p_a[0] = p_a[0] + (dat0 * off0_2); p_a[1] = p_a[1] + (dat1 * off1_2); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[43].time += t1 - t2; } ops_set_dirtybit_host(args, 3); 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[43].mpi_time += t2 - t1; OPS_kernels[43].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[43].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_advec_mom_kernel_post_pre_advec_y_execute( ops_kernel_descriptor *desc) { ops_block block = desc->block; int dim = desc->dim; int *range = desc->range; ops_arg arg0 = desc->args[0]; ops_arg arg1 = desc->args[1]; ops_arg arg2 = desc->args[2]; ops_arg arg3 = desc->args[3]; ops_arg arg4 = desc->args[4]; // Timing double t1, t2, c1, c2; ops_arg args[5] = {arg0, arg1, arg2, arg3, arg4}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 5, range, 132)) return; #endif if (OPS_diags > 1) { OPS_kernels[132].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[3]; int end[3]; for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #ifdef OPS_DEBUG ops_register_args(args, "advec_mom_kernel_post_pre_advec_y"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ node_mass_post = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ post_vol = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; const double *__restrict__ density1 = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; double *__restrict__ node_mass_pre = (double *)(args[3].data + base3); int base4 = args[4].dat->base_offset; const double *__restrict__ node_flux = (double *)(args[4].data + base4); // initialize global variable with the dimension of dats int xdim0_advec_mom_kernel_post_pre_advec_y = args[0].dat->size[0]; int ydim0_advec_mom_kernel_post_pre_advec_y = args[0].dat->size[1]; int xdim1_advec_mom_kernel_post_pre_advec_y = args[1].dat->size[0]; int ydim1_advec_mom_kernel_post_pre_advec_y = args[1].dat->size[1]; int xdim2_advec_mom_kernel_post_pre_advec_y = args[2].dat->size[0]; int ydim2_advec_mom_kernel_post_pre_advec_y = args[2].dat->size[1]; int xdim3_advec_mom_kernel_post_pre_advec_y = args[3].dat->size[0]; int ydim3_advec_mom_kernel_post_pre_advec_y = args[3].dat->size[1]; int xdim4_advec_mom_kernel_post_pre_advec_y = args[4].dat->size[0]; int ydim4_advec_mom_kernel_post_pre_advec_y = args[4].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[132].mpi_time += t1 - t2; } #pragma omp parallel for collapse(2) for (int n_z = start[2]; n_z < end[2]; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd aligned(node_mass_post, post_vol, density1, node_mass_pre, \ node_flux) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { node_mass_post[OPS_ACC0(0, 0, 0)] = 0.125 * (density1[OPS_ACC2(0, -1, 0)] * post_vol[OPS_ACC1(0, -1, 0)] + density1[OPS_ACC2(0, 0, 0)] * post_vol[OPS_ACC1(0, 0, 0)] + density1[OPS_ACC2(-1, -1, 0)] * post_vol[OPS_ACC1(-1, -1, 0)] + density1[OPS_ACC2(-1, 0, 0)] * post_vol[OPS_ACC1(-1, 0, 0)] + density1[OPS_ACC2(0, -1, -1)] * post_vol[OPS_ACC1(0, -1, -1)] + density1[OPS_ACC2(0, 0, -1)] * post_vol[OPS_ACC1(0, 0, -1)] + density1[OPS_ACC2(-1, -1, -1)] * post_vol[OPS_ACC1(-1, -1, -1)] + density1[OPS_ACC2(-1, 0, -1)] * post_vol[OPS_ACC1(-1, 0, -1)]); node_mass_pre[OPS_ACC3(0, 0, 0)] = node_mass_post[OPS_ACC0(0, 0, 0)] - node_flux[OPS_ACC4(0, -1, 0)] + node_flux[OPS_ACC4(0, 0, 0)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[132].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[132].mpi_time += t1 - t2; OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg4); } }
// host stub function void ops_par_loop_advec_mom_kernel_post_pre_advec_y( 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) { // Timing double t1, t2, c1, c2; ops_arg args[5] = {arg0, arg1, arg2, arg3, arg4}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 5, range, 132)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(132, "advec_mom_kernel_post_pre_advec_y"); OPS_kernels[132].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]; // build opencl kernel if not already built buildOpenCLKernels_advec_mom_kernel_post_pre_advec_y( xdim0, ydim0, xdim1, ydim1, xdim2, ydim2, xdim3, ydim3, xdim4, ydim4); // 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}; // 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]); ops_H_D_exchanges_device(args, 5); ops_halo_exchanges(args, 5, range); ops_H_D_exchanges_device(args, 5); if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[132].mpi_time += t2 - t1; } if (globalWorkSize[0] > 0 && globalWorkSize[1] > 0 && globalWorkSize[2] > 0) { clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 0, sizeof(cl_mem), (void *)&arg0.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 1, sizeof(cl_mem), (void *)&arg1.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 2, sizeof(cl_mem), (void *)&arg2.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 3, sizeof(cl_mem), (void *)&arg3.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 4, sizeof(cl_mem), (void *)&arg4.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 5, sizeof(cl_int), (void *)&base0)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 6, sizeof(cl_int), (void *)&base1)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 7, sizeof(cl_int), (void *)&base2)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 8, sizeof(cl_int), (void *)&base3)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 9, sizeof(cl_int), (void *)&base4)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 10, sizeof(cl_int), (void *)&x_size)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 11, sizeof(cl_int), (void *)&y_size)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[132], 12, sizeof(cl_int), (void *)&z_size)); // call/enque opencl kernel wrapper function clSafeCall(clEnqueueNDRangeKernel( OPS_opencl_core.command_queue, OPS_opencl_core.kernel[132], 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[132].time += t1 - t2; } ops_set_dirtybit_device(args, 5); ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[3], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[132].mpi_time += t2 - t1; OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[132].transfer += ops_compute_transfer(dim, start, end, &arg4); } }
// host stub function void ops_par_loop_update_halo_kernel5_minus_4_back_execute( ops_kernel_descriptor *desc) { ops_block block = desc->block; int dim = desc->dim; int *range = desc->range; ops_arg arg0 = desc->args[0]; ops_arg arg1 = desc->args[1]; ops_arg arg2 = desc->args[2]; // Timing double t1, t2, c1, c2; ops_arg args[3] = {arg0, arg1, arg2}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 3, range, 92)) return; #endif if (OPS_diags > 1) { OPS_kernels[92].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[3]; int end[3]; for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #ifdef OPS_DEBUG ops_register_args(args, "update_halo_kernel5_minus_4_back"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ vol_flux_z = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; double *__restrict__ mass_flux_z = (double *)(args[1].data + base1); const int *__restrict__ fields = (int *)args[2].data; // initialize global variable with the dimension of dats int xdim0_update_halo_kernel5_minus_4_back = args[0].dat->size[0]; int ydim0_update_halo_kernel5_minus_4_back = args[0].dat->size[1]; int xdim1_update_halo_kernel5_minus_4_back = args[1].dat->size[0]; int ydim1_update_halo_kernel5_minus_4_back = args[1].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[92].mpi_time += t1 - t2; } #pragma omp parallel for collapse(2) for (int n_z = start[2]; n_z < end[2]; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd aligned(vol_flux_z, mass_flux_z) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { if (fields[FIELD_VOL_FLUX_Z] == 1) vol_flux_z[OPS_ACC0(0, 0, 0)] = -vol_flux_z[OPS_ACC0(0, 0, 4)]; if (fields[FIELD_MASS_FLUX_Z] == 1) mass_flux_z[OPS_ACC1(0, 0, 0)] = -mass_flux_z[OPS_ACC1(0, 0, 4)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[92].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[92].mpi_time += t1 - t2; OPS_kernels[92].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[92].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_accelerate_kernel_execute(ops_kernel_descriptor *desc) { ops_block block = desc->block; int dim = desc->dim; int *range = desc->range; ops_arg arg0 = desc->args[0]; ops_arg arg1 = desc->args[1]; ops_arg arg2 = desc->args[2]; ops_arg arg3 = desc->args[3]; ops_arg arg4 = desc->args[4]; ops_arg arg5 = desc->args[5]; ops_arg arg6 = desc->args[6]; ops_arg arg7 = desc->args[7]; ops_arg arg8 = desc->args[8]; ops_arg arg9 = desc->args[9]; ops_arg arg10 = desc->args[10]; ops_arg arg11 = desc->args[11]; ops_arg arg12 = desc->args[12]; ops_arg arg13 = desc->args[13]; // Timing double t1, t2, c1, c2; ops_arg args[14] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9, arg10, arg11, arg12, arg13}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 14, range, 105)) return; #endif if (OPS_diags > 1) { OPS_kernels[105].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[3]; int end[3]; for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #ifdef OPS_DEBUG ops_register_args(args, "accelerate_kernel"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; const double *__restrict__ density0 = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ volume = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ stepbymass = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ xvel0 = (double *)(args[3].data + base3); int base4 = args[4].dat->base_offset; double *__restrict__ xvel1 = (double *)(args[4].data + base4); int base5 = args[5].dat->base_offset; const double *__restrict__ xarea = (double *)(args[5].data + base5); int base6 = args[6].dat->base_offset; const double *__restrict__ pressure = (double *)(args[6].data + base6); int base7 = args[7].dat->base_offset; const double *__restrict__ yvel0 = (double *)(args[7].data + base7); int base8 = args[8].dat->base_offset; double *__restrict__ yvel1 = (double *)(args[8].data + base8); int base9 = args[9].dat->base_offset; const double *__restrict__ yarea = (double *)(args[9].data + base9); int base10 = args[10].dat->base_offset; const double *__restrict__ viscosity = (double *)(args[10].data + base10); int base11 = args[11].dat->base_offset; const double *__restrict__ zvel0 = (double *)(args[11].data + base11); int base12 = args[12].dat->base_offset; double *__restrict__ zvel1 = (double *)(args[12].data + base12); int base13 = args[13].dat->base_offset; const double *__restrict__ zarea = (double *)(args[13].data + base13); // initialize global variable with the dimension of dats int xdim0_accelerate_kernel = args[0].dat->size[0]; int ydim0_accelerate_kernel = args[0].dat->size[1]; int xdim1_accelerate_kernel = args[1].dat->size[0]; int ydim1_accelerate_kernel = args[1].dat->size[1]; int xdim2_accelerate_kernel = args[2].dat->size[0]; int ydim2_accelerate_kernel = args[2].dat->size[1]; int xdim3_accelerate_kernel = args[3].dat->size[0]; int ydim3_accelerate_kernel = args[3].dat->size[1]; int xdim4_accelerate_kernel = args[4].dat->size[0]; int ydim4_accelerate_kernel = args[4].dat->size[1]; int xdim5_accelerate_kernel = args[5].dat->size[0]; int ydim5_accelerate_kernel = args[5].dat->size[1]; int xdim6_accelerate_kernel = args[6].dat->size[0]; int ydim6_accelerate_kernel = args[6].dat->size[1]; int xdim7_accelerate_kernel = args[7].dat->size[0]; int ydim7_accelerate_kernel = args[7].dat->size[1]; int xdim8_accelerate_kernel = args[8].dat->size[0]; int ydim8_accelerate_kernel = args[8].dat->size[1]; int xdim9_accelerate_kernel = args[9].dat->size[0]; int ydim9_accelerate_kernel = args[9].dat->size[1]; int xdim10_accelerate_kernel = args[10].dat->size[0]; int ydim10_accelerate_kernel = args[10].dat->size[1]; int xdim11_accelerate_kernel = args[11].dat->size[0]; int ydim11_accelerate_kernel = args[11].dat->size[1]; int xdim12_accelerate_kernel = args[12].dat->size[0]; int ydim12_accelerate_kernel = args[12].dat->size[1]; int xdim13_accelerate_kernel = args[13].dat->size[0]; int ydim13_accelerate_kernel = args[13].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[105].mpi_time += t1 - t2; } #pragma omp parallel for collapse(2) for (int n_z = start[2]; n_z < end[2]; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd aligned(density0, volume, stepbymass, xvel0, xvel1, xarea, \ pressure, yvel0, yvel1, yarea, viscosity, zvel0, \ zvel1, zarea) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double nodal_mass = 0.0; nodal_mass = (density0[OPS_ACC0(-1, -1, 0)] * volume[OPS_ACC1(-1, -1, 0)] + density0[OPS_ACC0(0, -1, 0)] * volume[OPS_ACC1(0, -1, 0)] + density0[OPS_ACC0(0, 0, 0)] * volume[OPS_ACC1(0, 0, 0)] + density0[OPS_ACC0(-1, 0, 0)] * volume[OPS_ACC1(-1, 0, 0)] + density0[OPS_ACC0(-1, -1, -1)] * volume[OPS_ACC1(-1, -1, -1)] + density0[OPS_ACC0(0, -1, -1)] * volume[OPS_ACC1(0, -1, -1)] + density0[OPS_ACC0(0, 0, -1)] * volume[OPS_ACC1(0, 0, -1)] + density0[OPS_ACC0(-1, 0, -1)] * volume[OPS_ACC1(-1, 0, -1)]) * 0.125; stepbymass[OPS_ACC2(0, 0, 0)] = 0.25 * dt / nodal_mass; xvel1[OPS_ACC4(0, 0, 0)] = xvel0[OPS_ACC3(0, 0, 0)] - stepbymass[OPS_ACC2(0, 0, 0)] * (xarea[OPS_ACC5(0, 0, 0)] * (pressure[OPS_ACC6(0, 0, 0)] - pressure[OPS_ACC6(-1, 0, 0)]) + xarea[OPS_ACC5(0, -1, 0)] * (pressure[OPS_ACC6(0, -1, 0)] - pressure[OPS_ACC6(-1, -1, 0)]) + xarea[OPS_ACC5(0, 0, -1)] * (pressure[OPS_ACC6(0, 0, -1)] - pressure[OPS_ACC6(-1, 0, -1)]) + xarea[OPS_ACC5(0, -1, -1)] * (pressure[OPS_ACC6(0, -1, -1)] - pressure[OPS_ACC6(-1, -1, -1)])); yvel1[OPS_ACC8(0, 0, 0)] = yvel0[OPS_ACC7(0, 0, 0)] - stepbymass[OPS_ACC2(0, 0, 0)] * (yarea[OPS_ACC9(0, 0, 0)] * (pressure[OPS_ACC6(0, 0, 0)] - pressure[OPS_ACC6(0, -1, 0)]) + yarea[OPS_ACC9(-1, 0, 0)] * (pressure[OPS_ACC6(-1, 0, 0)] - pressure[OPS_ACC6(-1, -1, 0)]) + yarea[OPS_ACC9(0, 0, -1)] * (pressure[OPS_ACC6(0, 0, -1)] - pressure[OPS_ACC6(0, -1, -1)]) + yarea[OPS_ACC9(-1, 0, -1)] * (pressure[OPS_ACC6(-1, 0, -1)] - pressure[OPS_ACC6(-1, -1, -1)])); zvel1[OPS_ACC12(0, 0, 0)] = zvel0[OPS_ACC11(0, 0, 0)] - stepbymass[OPS_ACC2(0, 0, 0)] * (zarea[OPS_ACC13(0, 0, 0)] * (pressure[OPS_ACC6(0, 0, 0)] - pressure[OPS_ACC6(0, 0, -1)]) + zarea[OPS_ACC13(0, -1, 0)] * (pressure[OPS_ACC6(0, -1, 0)] - pressure[OPS_ACC6(0, -1, -1)]) + zarea[OPS_ACC13(-1, 0, 0)] * (pressure[OPS_ACC6(-1, 0, 0)] - pressure[OPS_ACC6(-1, 0, -1)]) + zarea[OPS_ACC13(-1, -1, 0)] * (pressure[OPS_ACC6(-1, -1, 0)] - pressure[OPS_ACC6(-1, -1, -1)])); xvel1[OPS_ACC4(0, 0, 0)] = xvel1[OPS_ACC4(0, 0, 0)] - stepbymass[OPS_ACC2(0, 0, 0)] * (xarea[OPS_ACC5(0, 0, 0)] * (viscosity[OPS_ACC10(0, 0, 0)] - viscosity[OPS_ACC10(-1, 0, 0)]) + xarea[OPS_ACC5(0, -1, 0)] * (viscosity[OPS_ACC10(0, -1, 0)] - viscosity[OPS_ACC10(-1, -1, 0)]) + xarea[OPS_ACC5(0, 0, -1)] * (viscosity[OPS_ACC10(0, 0, -1)] - viscosity[OPS_ACC10(-1, 0, -1)]) + xarea[OPS_ACC5(0, -1, -1)] * (viscosity[OPS_ACC10(0, -1, -1)] - viscosity[OPS_ACC10(-1, -1, -1)])); yvel1[OPS_ACC8(0, 0, 0)] = yvel1[OPS_ACC8(0, 0, 0)] - stepbymass[OPS_ACC2(0, 0, 0)] * (yarea[OPS_ACC9(0, 0, 0)] * (viscosity[OPS_ACC10(0, 0, 0)] - viscosity[OPS_ACC10(0, -1, 0)]) + yarea[OPS_ACC9(-1, 0, 0)] * (viscosity[OPS_ACC10(-1, 0, 0)] - viscosity[OPS_ACC10(-1, -1, 0)]) + yarea[OPS_ACC9(0, 0, -1)] * (viscosity[OPS_ACC10(0, 0, -1)] - viscosity[OPS_ACC10(0, -1, -1)]) + yarea[OPS_ACC9(-1, 0, -1)] * (viscosity[OPS_ACC10(-1, 0, -1)] - viscosity[OPS_ACC10(-1, -1, -1)])); zvel1[OPS_ACC12(0, 0, 0)] = zvel1[OPS_ACC12(0, 0, 0)] - stepbymass[OPS_ACC2(0, 0, 0)] * (zarea[OPS_ACC13(0, 0, 0)] * (viscosity[OPS_ACC10(0, 0, 0)] - viscosity[OPS_ACC10(0, 0, -1)]) + zarea[OPS_ACC13(0, -1, 0)] * (viscosity[OPS_ACC10(0, -1, 0)] - viscosity[OPS_ACC10(0, -1, -1)]) + zarea[OPS_ACC13(-1, 0, 0)] * (viscosity[OPS_ACC10(-1, 0, 0)] - viscosity[OPS_ACC10(-1, 0, -1)]) + zarea[OPS_ACC13(-1, -1, 0)] * (viscosity[OPS_ACC10(-1, -1, 0)] - viscosity[OPS_ACC10(-1, -1, -1)])); } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[105].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[105].mpi_time += t1 - t2; OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg6); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg7); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg8); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg9); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg10); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg11); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg12); OPS_kernels[105].transfer += ops_compute_transfer(dim, start, end, &arg13); } }
// host stub function void ops_par_loop_reset_field_kernel1(char const *name, ops_block Block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3) { ops_arg args[4] = { arg0, arg1, arg2, arg3}; ops_timing_realloc(3,"reset_field_kernel1"); OPS_kernels[3].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]; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); if (xdim0 != xdim0_reset_field_kernel1_h || ydim0 != ydim0_reset_field_kernel1_h || xdim1 != xdim1_reset_field_kernel1_h || ydim1 != ydim1_reset_field_kernel1_h || xdim2 != xdim2_reset_field_kernel1_h || ydim2 != ydim2_reset_field_kernel1_h || xdim3 != xdim3_reset_field_kernel1_h || ydim3 != ydim3_reset_field_kernel1_h) { xdim0_reset_field_kernel1 = xdim0; xdim0_reset_field_kernel1_h = xdim0; ydim0_reset_field_kernel1 = ydim0; ydim0_reset_field_kernel1_h = ydim0; xdim1_reset_field_kernel1 = xdim1; xdim1_reset_field_kernel1_h = xdim1; ydim1_reset_field_kernel1 = ydim1; ydim1_reset_field_kernel1_h = ydim1; xdim2_reset_field_kernel1 = xdim2; xdim2_reset_field_kernel1_h = xdim2; ydim2_reset_field_kernel1 = ydim2; ydim2_reset_field_kernel1_h = ydim2; xdim3_reset_field_kernel1 = xdim3; xdim3_reset_field_kernel1_h = xdim3; ydim3_reset_field_kernel1 = ydim3; ydim3_reset_field_kernel1_h = ydim3; } 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; //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_GPU ops_H_D_exchanges_device(args, 4); #else ops_H_D_exchanges_host(args, 4); #endif ops_halo_exchanges(args,4,range); ops_timers_core(&c1,&t1); OPS_kernels[3].mpi_time += t1-t2; reset_field_kernel1_c_wrapper( p_a0, p_a1, p_a2, p_a3, x_size, y_size, z_size); ops_timers_core(&c2,&t2); OPS_kernels[3].time += t2-t1; #ifdef OPS_GPU ops_set_dirtybit_device(args, 4); #else ops_set_dirtybit_host(args, 4); #endif ops_set_halo_dirtybit3(&args[0],range); ops_set_halo_dirtybit3(&args[2],range); //Update kernel record OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg1); OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg2); OPS_kernels[3].transfer += ops_compute_transfer(dim, range, &arg3); }
// host stub function void ops_par_loop_update_halo_kernel2_xvel_minus_4_right(char const *name, ops_block block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2) { char *p_a[3]; int offs[3][3]; ops_arg args[3] = { arg0, arg1, arg2}; ops_timing_realloc(59,"update_halo_kernel2_xvel_minus_4_right"); OPS_kernels[59].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 #ifdef OPS_DEBUG ops_register_args(args, "update_halo_kernel2_xvel_minus_4_right"); #endif offs[0][0] = args[0].stencil->stride[0]*1; //unit step in x dimension offs[0][1] = off3D(1, &start[0], &end[0],args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[0][2] = off3D(2, &start[0], &end[0],args[0].dat->size, args[0].stencil->stride) - offs[0][1] - offs[0][0]; offs[1][0] = args[1].stencil->stride[0]*1; //unit step in x dimension offs[1][1] = off3D(1, &start[0], &end[0],args[1].dat->size, args[1].stencil->stride) - offs[1][0]; offs[1][2] = off3D(2, &start[0], &end[0],args[1].dat->size, args[1].stencil->stride) - offs[1][1] - offs[1][0]; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int off0_2 = offs[0][2]; int dat0 = args[0].dat->elem_size; int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int off1_2 = offs[1][2]; int dat1 = args[1].dat->elem_size; //set up initial pointers and exchange halos if necessary 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]); p_a[0] = (char *)args[0].data + base0; #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]); p_a[1] = (char *)args[1].data + base1; p_a[2] = args[2].data; ops_H_D_exchanges_host(args, 3); ops_halo_exchanges(args,3,range); ops_H_D_exchanges_host(args, 3); ops_timers_core(&c1,&t1); OPS_kernels[59].mpi_time += t1-t2; 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]; int n_x; for ( int n_z=start[2]; n_z<end[2]; n_z++ ){ for ( int n_y=start[1]; n_y<end[1]; n_y++ ){ #pragma novector for( n_x=start[0]; n_x<start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x+=SIMD_VEC ) { //call kernel function, passing in pointers to data -vectorised #pragma simd for ( int i=0; i<SIMD_VEC; i++ ){ update_halo_kernel2_xvel_minus_4_right( (double *)p_a[0]+ i*1, (double *)p_a[1]+ i*1, (int *)p_a[2] ); } //shift pointers to data x direction p_a[0]= p_a[0] + (dat0 * off0_0)*SIMD_VEC; p_a[1]= p_a[1] + (dat1 * off1_0)*SIMD_VEC; } for ( int n_x=start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x<end[0]; n_x++ ){ //call kernel function, passing in pointers to data - remainder update_halo_kernel2_xvel_minus_4_right( (double *)p_a[0], (double *)p_a[1], (int *)p_a[2] ); //shift pointers to data x direction p_a[0]= p_a[0] + (dat0 * off0_0); p_a[1]= p_a[1] + (dat1 * off1_0); } //shift pointers to data y direction p_a[0]= p_a[0] + (dat0 * off0_1); p_a[1]= p_a[1] + (dat1 * off1_1); } //shift pointers to data z direction p_a[0]= p_a[0] + (dat0 * off0_2); p_a[1]= p_a[1] + (dat1 * off1_2); } ops_timers_core(&c2,&t2); OPS_kernels[59].time += t2-t1; ops_set_dirtybit_host(args, 3); ops_set_halo_dirtybit3(&args[0],range); ops_set_halo_dirtybit3(&args[1],range); //Update kernel record OPS_kernels[59].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[59].transfer += ops_compute_transfer(dim, range, &arg1); }
// host stub function void ops_par_loop_tea_leaf_common_init_kernel(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) { // Timing double t1, t2, c1, c2; ops_arg args[7] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 7, range, 36)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(36, "tea_leaf_common_init_kernel"); OPS_kernels[36].count++; ops_timers_core(&c1, &t1); } // 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]; #endif // OPS_MPI int arg_idx[2]; int arg_idx_base[2]; #ifdef OPS_MPI if (compute_ranges(args, 7, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 2; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif for (int n = 0; n < 2; 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; // 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]; #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]; #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]; #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]; #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]; #ifdef OPS_GPU double *p_a4 = (double *)((char *)args[4].data_d + base4); #else double *p_a4 = (double *)((char *)args[4].data + base4); #endif double *p_a5 = (double *)args[5].data; double *p_a6 = (double *)args[6].data; int x_size = MAX(0, end[0] - start[0]); int y_size = MAX(0, end[1] - start[1]); // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; xdim2 = args[2].dat->size[0]; xdim3 = args[3].dat->size[0]; xdim4 = args[4].dat->size[0]; if (xdim0 != xdim0_tea_leaf_common_init_kernel_h || xdim1 != xdim1_tea_leaf_common_init_kernel_h || xdim2 != xdim2_tea_leaf_common_init_kernel_h || xdim3 != xdim3_tea_leaf_common_init_kernel_h || xdim4 != xdim4_tea_leaf_common_init_kernel_h) { xdim0_tea_leaf_common_init_kernel = xdim0; xdim0_tea_leaf_common_init_kernel_h = xdim0; xdim1_tea_leaf_common_init_kernel = xdim1; xdim1_tea_leaf_common_init_kernel_h = xdim1; xdim2_tea_leaf_common_init_kernel = xdim2; xdim2_tea_leaf_common_init_kernel_h = xdim2; xdim3_tea_leaf_common_init_kernel = xdim3; xdim3_tea_leaf_common_init_kernel_h = xdim3; xdim4_tea_leaf_common_init_kernel = xdim4; xdim4_tea_leaf_common_init_kernel_h = xdim4; } // Halo Exchanges #ifdef OPS_GPU ops_H_D_exchanges_device(args, 7); #else ops_H_D_exchanges_host(args, 7); #endif ops_halo_exchanges(args, 7, range); #ifdef OPS_GPU ops_H_D_exchanges_device(args, 7); #else ops_H_D_exchanges_host(args, 7); #endif if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[36].mpi_time += t2 - t1; } tea_leaf_common_init_kernel_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, *p_a5, *p_a6, x_size, y_size); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[36].time += t1 - t2; } #ifdef OPS_GPU ops_set_dirtybit_device(args, 7); #else ops_set_dirtybit_host(args, 7); #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[36].mpi_time += t2 - t1; OPS_kernels[36].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[36].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[36].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[36].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[36].transfer += ops_compute_transfer(dim, start, end, &arg4); } }
void ops_timers(double * cpu, double * et) { ops_timers_core(cpu,et); }
int main(int argc, char **argv) { /**-------------------------- Initialisation --------------------------**/ // OPS initialisation ops_init(argc,argv,6); int logical_size_x = 200; int logical_size_y = 200; int ngrid_x = 1; int ngrid_y = 1; int n_iter = 10000; dx = 0.01; dy = 0.01; ops_decl_const("dx",1,"double",&dx); ops_decl_const("dy",1,"double",&dy); //declare blocks ops_block *blocks = (ops_block *)malloc(ngrid_x*ngrid_y*sizeof(ops_block*)); char buf[50]; for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { sprintf(buf,"block %d,%d",i,j); blocks[i+ngrid_x*j] = ops_decl_block(2,buf); } } //declare stencils int s2D_00[] = {0,0}; ops_stencil S2D_00 = ops_decl_stencil( 2, 1, s2D_00, "00"); int s2D_00_P10_M10_0P1_0M1[] = {0,0, 1,0, -1,0, 0,1, 0,-1}; ops_stencil S2D_00_P10_M10_0P1_0M1 = ops_decl_stencil( 2, 5, s2D_00_P10_M10_0P1_0M1, "00:10:-10:01:0-1"); ops_reduction red_err = ops_decl_reduction_handle(sizeof(double), "double", "err"); //declare datasets int d_p[2] = {1,1}; //max halo depths for the dat in the possitive direction int d_m[2] = {-1,-1}; //max halo depths for the dat in the negative direction int base[2] = {0,0}; int uniform_size[2] = {(logical_size_x-1)/ngrid_x+1,(logical_size_y-1)/ngrid_y+1}; double* temp = NULL; ops_dat *coordx = (ops_dat *)malloc(ngrid_x*ngrid_y*sizeof(ops_dat*)); ops_dat *coordy = (ops_dat *)malloc(ngrid_x*ngrid_y*sizeof(ops_dat*)); ops_dat *u = (ops_dat *)malloc(ngrid_x*ngrid_y*sizeof(ops_dat*)); ops_dat *u2 = (ops_dat *)malloc(ngrid_x*ngrid_y*sizeof(ops_dat*)); ops_dat *f = (ops_dat *)malloc(ngrid_x*ngrid_y*sizeof(ops_dat*)); ops_dat *ref = (ops_dat *)malloc(ngrid_x*ngrid_y*sizeof(ops_dat*)); int *sizes = (int*)malloc(2*ngrid_x*ngrid_y*sizeof(int)); int *disps = (int*)malloc(2*ngrid_x*ngrid_y*sizeof(int)); for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { int size[2] = {uniform_size[0], uniform_size[1]}; if ((i+1)*size[0]>logical_size_x) size[0] = logical_size_x - i*size[0]; if ((j+1)*size[1]>logical_size_y) size[1] = logical_size_y - j*size[1]; sprintf(buf,"coordx %d,%d",i,j); coordx[i+ngrid_x*j] = ops_decl_dat(blocks[i+ngrid_x*j], 1, size, base, d_m, d_p, temp, "double", buf); sprintf(buf,"coordy %d,%d",i,j); coordy[i+ngrid_x*j] = ops_decl_dat(blocks[i+ngrid_x*j], 1, size, base, d_m, d_p, temp, "double", buf); sprintf(buf,"u %d,%d",i,j); u[i+ngrid_x*j] = ops_decl_dat(blocks[i+ngrid_x*j], 1, size, base, d_m, d_p, temp, "double", buf); sprintf(buf,"u2 %d,%d",i,j); u2[i+ngrid_x*j] = ops_decl_dat(blocks[i+ngrid_x*j], 1, size, base, d_m, d_p, temp, "double", buf); sprintf(buf,"f %d,%d",i,j); f[i+ngrid_x*j] = ops_decl_dat(blocks[i+ngrid_x*j], 1, size, base, d_m, d_p, temp, "double", buf); sprintf(buf,"ref %d,%d",i,j); ref[i+ngrid_x*j] = ops_decl_dat(blocks[i+ngrid_x*j], 1, size, base, d_m, d_p, temp, "double", buf); sizes[2*(i+ngrid_x*j)] = size[0]; sizes[2*(i+ngrid_x*j)+1] = size[1]; disps[2*(i+ngrid_x*j)] = i*uniform_size[0]; disps[2*(i+ngrid_x*j)+1] = j*uniform_size[1]; } } ops_halo *halos = (ops_halo *)malloc(2*(ngrid_x*(ngrid_y-1)+(ngrid_x-1)*ngrid_y)*sizeof(ops_halo *)); int off = 0; for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { if (i > 0) { int halo_iter[] = {1,sizes[2*(i+ngrid_x*j)+1]}; int base_from[] = {sizes[2*(i-1+ngrid_x*j)]-1,0}; int base_to[] = {-1,0}; int dir[] = {1,2}; halos[off++] = ops_decl_halo(u[i-1+ngrid_x*j], u[i+ngrid_x*j], halo_iter, base_from, base_to, dir, dir); base_from[0] = 0; base_to[0] = sizes[2*(i+ngrid_x*j)]; halos[off++] = ops_decl_halo(u[i+ngrid_x*j], u[i-1+ngrid_x*j], halo_iter, base_from, base_to, dir, dir); } if (j > 0) { int halo_iter[] = {sizes[2*(i+ngrid_x*j)],1}; int base_from[] = {0,sizes[2*(i+ngrid_x*(j-1))+1]-1}; int base_to[] = {0,-1}; int dir[] = {1,2}; halos[off++] = ops_decl_halo(u[i+ngrid_x*(j-1)], u[i+ngrid_x*j], halo_iter, base_from, base_to, dir, dir); base_from[1] = 0; base_to[1] = sizes[2*(i+ngrid_x*j)+1]; halos[off++] = ops_decl_halo(u[i+ngrid_x*j], u[i+ngrid_x*(j-1)], halo_iter, base_from, base_to, dir, dir); } } } if (off != 2*(ngrid_x*(ngrid_y-1)+(ngrid_x-1)*ngrid_y)) printf("Something is not right\n"); ops_halo_group u_halos = ops_decl_halo_group(off,halos); ops_partition(""); ops_checkpointing_init("check.h5", 5.0); /**-------------------------- Computations --------------------------**/ double ct0, ct1, et0, et1; ops_timers_core(&ct0, &et0); //populate forcing, reference solution and boundary conditions for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { int iter_range[] = {-1,sizes[2*(i+ngrid_x*j)]+1,-1,sizes[2*(i+ngrid_x*j)+1]+1}; ops_par_loop(poisson_kernel_populate, "poisson_kernel_populate", blocks[i+ngrid_x*j], 2, iter_range, ops_arg_gbl(&disps[2*(i+ngrid_x*j)], 1, "int", OPS_READ), ops_arg_gbl(&disps[2*(i+ngrid_x*j)+1], 1, "int", OPS_READ), ops_arg_idx(), ops_arg_dat(u[i+ngrid_x*j], S2D_00, "double", OPS_WRITE), ops_arg_dat(f[i+ngrid_x*j], S2D_00, "double", OPS_WRITE), ops_arg_dat(ref[i+ngrid_x*j], S2D_00, "double", OPS_WRITE)); } } //initial guess 0 for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { int iter_range[] = {0,sizes[2*(i+ngrid_x*j)],0,sizes[2*(i+ngrid_x*j)+1]}; ops_par_loop(poisson_kernel_initialguess, "poisson_kernel_initialguess", blocks[i+ngrid_x*j], 2, iter_range, ops_arg_dat(u[i+ngrid_x*j], S2D_00, "double", OPS_WRITE)); } } for (int iter = 0; iter < n_iter; iter++) { ops_halo_transfer(u_halos); for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { int iter_range[] = {0,sizes[2*(i+ngrid_x*j)],0,sizes[2*(i+ngrid_x*j)+1]}; ops_par_loop(poisson_kernel_stencil, "poisson_kernel_stencil", blocks[i+ngrid_x*j], 2, iter_range, ops_arg_dat(u[i+ngrid_x*j], S2D_00_P10_M10_0P1_0M1, "double", OPS_READ), ops_arg_dat(f[i+ngrid_x*j], S2D_00, "double", OPS_READ), ops_arg_dat(u2[i+ngrid_x*j], S2D_00, "double", OPS_WRITE)); } } for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { int iter_range[] = {0,sizes[2*(i+ngrid_x*j)],0,sizes[2*(i+ngrid_x*j)+1]}; ops_par_loop(poisson_kernel_update, "poisson_kernel_update", blocks[i+ngrid_x*j], 2, iter_range, ops_arg_dat(u2[i+ngrid_x*j], S2D_00, "double", OPS_READ), ops_arg_dat(u[i+ngrid_x*j] , S2D_00, "double", OPS_WRITE)); } } } double err = 0.0; for (int j = 0; j < ngrid_y; j++) { for (int i = 0; i < ngrid_x; i++) { int iter_range[] = {0,sizes[2*(i+ngrid_x*j)],0,sizes[2*(i+ngrid_x*j)+1]}; ops_par_loop(poisson_kernel_error, "poisson_kernel_error", blocks[i+ngrid_x*j], 2, iter_range, ops_arg_dat(u[i+ngrid_x*j], S2D_00, "double", OPS_READ), ops_arg_dat(ref[i+ngrid_x*j] , S2D_00, "double", OPS_READ), ops_arg_reduce(red_err, 1, "double", OPS_INC)); } } ops_reduction_result(red_err,&err); ops_printf("Total error: %g\n",err); ops_timers_core(&ct1, &et1); ops_timing_output(); ops_printf("\nTotal Wall time %lf\n",et1-et0); ops_exit(); }
// host stub function void ops_par_loop_calc_dt_kernel_min(char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1) { // Timing double t1, t2, c1, c2; int offs[2][3]; ops_arg args[2] = {arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 2, range, 99)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(99, "calc_dt_kernel_min"); OPS_kernels[99].count++; ops_timers_core(&c1, &t1); } #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // compute locally allocated range for the sub-block int start[3]; int end[3]; int arg_idx[3]; #ifdef OPS_MPI 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]); if (end[n] < start[n]) end[n] = start[n]; } #else for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #endif #ifdef OPS_DEBUG ops_register_args(args, "calc_dt_kernel_min"); #endif offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[0][1] = off3D(1, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[0][2] = off3D(2, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][1] - offs[0][0]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int off0_2 = offs[0][2]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); #ifdef OPS_MPI double *arg1h = (double *)(((ops_reduction)args[1].data)->data + ((ops_reduction)args[1].data)->size * block->index); #else double *arg1h = (double *)(((ops_reduction)args[1].data)->data); #endif // Halo Exchanges ops_H_D_exchanges_host(args, 2); ops_halo_exchanges(args, 2, range); ops_H_D_exchanges_host(args, 2); #ifdef _OPENMP int nthreads = omp_get_max_threads(); #else int nthreads = 1; #endif // allocate and initialise arrays for global reduction // assumes a max of MAX_REDUCT_THREADS threads with a cacche line size of 64 // bytes double arg_gbl1[MAX(1, 64) * MAX_REDUCT_THREADS]; for (int thr = 0; thr < nthreads; thr++) { for (int d = 0; d < 1; d++) { arg_gbl1[d + 64 * thr] = INFINITY_double; } } xdim0 = args[0].dat->size[0]; ydim0 = args[0].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[99].mpi_time += t2 - t1; } #pragma omp parallel for for (int thr = 0; thr < nthreads; thr++) { int z_size = end[2] - start[2]; char *p_a[2]; int start_i = start[2] + ((z_size - 1) / nthreads + 1) * thr; int finish_i = start[2] + MIN(((z_size - 1) / nthreads + 1) * (thr + 1), z_size); // get address per thread int start0 = start[0]; int start1 = start[1]; int start2 = start_i; // 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 = dat0 * 1 * (start0 * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0 + dat0 * args[0].dat->size[0] * (start1 * 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] * (start2 * args[0].stencil->stride[2] - args[0].dat->base[2] - d_m[2]); p_a[0] = (char *)args[0].data + base0; p_a[1] = (char *)arg1h; for (int n_z = start_i; n_z < finish_i; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { for (int n_x = start[0]; n_x < start[0] + (end[0] - start[0]) / SIMD_VEC; n_x++) { // call kernel function, passing in pointers to data -vectorised for (int i = 0; i < SIMD_VEC; i++) { calc_dt_kernel_min((const double *)p_a[0] + i * 1 * 1, &arg_gbl1[64 * thr]); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; } for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x < end[0]; n_x++) { // call kernel function, passing in pointers to data - remainder calc_dt_kernel_min((const double *)p_a[0], &arg_gbl1[64 * thr]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); } // shift pointers to data y direction p_a[0] = p_a[0] + (dat0 * off0_1); } // shift pointers to data z direction p_a[0] = p_a[0] + (dat0 * off0_2); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[99].time += t1 - t2; } // combine reduction data for (int thr = 0; thr < nthreads; thr++) { for (int d = 0; d < 1; d++) { arg1h[d] = MIN(arg1h[d], arg_gbl1[64 * thr + d]); } } ops_set_dirtybit_host(args, 2); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[99].mpi_time += t2 - t1; OPS_kernels[99].transfer += ops_compute_transfer(dim, start, end, &arg0); } }
// host stub function void ops_par_loop_update_halo_kernel1_b1(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, 10)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(10, "update_halo_kernel1_b1"); OPS_kernels[10].count++; ops_timers_core(&c1, &t1); } // 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]; #endif // OPS_MPI int arg_idx[2]; int arg_idx_base[2]; #ifdef OPS_MPI if (compute_ranges(args, 8, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 2; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif for (int n = 0; n < 2; 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]; #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]; #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]; #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]; #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]; #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]; #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]; #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]); // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; xdim2 = args[2].dat->size[0]; xdim3 = args[3].dat->size[0]; xdim4 = args[4].dat->size[0]; xdim5 = args[5].dat->size[0]; xdim6 = args[6].dat->size[0]; if (xdim0 != xdim0_update_halo_kernel1_b1_h || xdim1 != xdim1_update_halo_kernel1_b1_h || xdim2 != xdim2_update_halo_kernel1_b1_h || xdim3 != xdim3_update_halo_kernel1_b1_h || xdim4 != xdim4_update_halo_kernel1_b1_h || xdim5 != xdim5_update_halo_kernel1_b1_h || xdim6 != xdim6_update_halo_kernel1_b1_h) { xdim0_update_halo_kernel1_b1 = xdim0; xdim0_update_halo_kernel1_b1_h = xdim0; xdim1_update_halo_kernel1_b1 = xdim1; xdim1_update_halo_kernel1_b1_h = xdim1; xdim2_update_halo_kernel1_b1 = xdim2; xdim2_update_halo_kernel1_b1_h = xdim2; xdim3_update_halo_kernel1_b1 = xdim3; xdim3_update_halo_kernel1_b1_h = xdim3; xdim4_update_halo_kernel1_b1 = xdim4; xdim4_update_halo_kernel1_b1_h = xdim4; xdim5_update_halo_kernel1_b1 = xdim5; xdim5_update_halo_kernel1_b1_h = xdim5; xdim6_update_halo_kernel1_b1 = xdim6; xdim6_update_halo_kernel1_b1_h = xdim6; } // 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[10].mpi_time += t2 - t1; } update_halo_kernel1_b1_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, p_a5, p_a6, p_a7, x_size, y_size); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[10].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[10].mpi_time += t2 - t1; OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[10].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_field_summary_kernel(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 arg8, ops_arg arg9, ops_arg arg10) { // Timing double t1, t2, c1, c2; int offs[11][2]; ops_arg args[11] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9, arg10}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 11, range, 49)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(49, "field_summary_kernel"); OPS_kernels[49].count++; ops_timers_core(&c1, &t1); } #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // compute locally allocated range for the sub-block int start[2]; int end[2]; int arg_idx[2]; #ifdef OPS_MPI 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]); if (end[n] < start[n]) end[n] = start[n]; } #else for (int n = 0; n < 2; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #endif #ifdef OPS_DEBUG ops_register_args(args, "field_summary_kernel"); #endif offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[0][1] = off2D(1, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[1][0] = args[1].stencil->stride[0] * 1; // unit step in x dimension offs[1][1] = off2D(1, &start[0], &end[0], args[1].dat->size, args[1].stencil->stride) - offs[1][0]; offs[2][0] = args[2].stencil->stride[0] * 1; // unit step in x dimension offs[2][1] = off2D(1, &start[0], &end[0], args[2].dat->size, args[2].stencil->stride) - offs[2][0]; offs[3][0] = args[3].stencil->stride[0] * 1; // unit step in x dimension offs[3][1] = off2D(1, &start[0], &end[0], args[3].dat->size, args[3].stencil->stride) - offs[3][0]; offs[4][0] = args[4].stencil->stride[0] * 1; // unit step in x dimension offs[4][1] = off2D(1, &start[0], &end[0], args[4].dat->size, args[4].stencil->stride) - offs[4][0]; offs[5][0] = args[5].stencil->stride[0] * 1; // unit step in x dimension offs[5][1] = off2D(1, &start[0], &end[0], args[5].dat->size, args[5].stencil->stride) - offs[5][0]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); int off2_0 = offs[2][0]; int off2_1 = offs[2][1]; int dat2 = (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size); int off3_0 = offs[3][0]; int off3_1 = offs[3][1]; int dat3 = (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size); int off4_0 = offs[4][0]; int off4_1 = offs[4][1]; int dat4 = (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size); int off5_0 = offs[5][0]; int off5_1 = offs[5][1]; int dat5 = (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size); #ifdef OPS_MPI double *arg6h = (double *)(((ops_reduction)args[6].data)->data + ((ops_reduction)args[6].data)->size * block->index); #else double *arg6h = (double *)(((ops_reduction)args[6].data)->data); #endif #ifdef OPS_MPI double *arg7h = (double *)(((ops_reduction)args[7].data)->data + ((ops_reduction)args[7].data)->size * block->index); #else double *arg7h = (double *)(((ops_reduction)args[7].data)->data); #endif #ifdef OPS_MPI double *arg8h = (double *)(((ops_reduction)args[8].data)->data + ((ops_reduction)args[8].data)->size * block->index); #else double *arg8h = (double *)(((ops_reduction)args[8].data)->data); #endif #ifdef OPS_MPI double *arg9h = (double *)(((ops_reduction)args[9].data)->data + ((ops_reduction)args[9].data)->size * block->index); #else double *arg9h = (double *)(((ops_reduction)args[9].data)->data); #endif #ifdef OPS_MPI double *arg10h = (double *)(((ops_reduction)args[10].data)->data + ((ops_reduction)args[10].data)->size * block->index); #else double *arg10h = (double *)(((ops_reduction)args[10].data)->data); #endif // Halo Exchanges ops_H_D_exchanges_host(args, 11); ops_halo_exchanges(args, 11, range); ops_H_D_exchanges_host(args, 11); #ifdef _OPENMP int nthreads = omp_get_max_threads(); #else int nthreads = 1; #endif // allocate and initialise arrays for global reduction // assumes a max of MAX_REDUCT_THREADS threads with a cacche line size of 64 // bytes double arg_gbl6[MAX(1, 64) * MAX_REDUCT_THREADS]; double arg_gbl7[MAX(1, 64) * MAX_REDUCT_THREADS]; double arg_gbl8[MAX(1, 64) * MAX_REDUCT_THREADS]; double arg_gbl9[MAX(1, 64) * MAX_REDUCT_THREADS]; double arg_gbl10[MAX(1, 64) * MAX_REDUCT_THREADS]; for (int thr = 0; thr < nthreads; thr++) { for (int d = 0; d < 1; d++) { arg_gbl6[d + 64 * thr] = ZERO_double; } for (int d = 0; d < 1; d++) { arg_gbl7[d + 64 * thr] = ZERO_double; } for (int d = 0; d < 1; d++) { arg_gbl8[d + 64 * thr] = ZERO_double; } for (int d = 0; d < 1; d++) { arg_gbl9[d + 64 * thr] = ZERO_double; } for (int d = 0; d < 1; d++) { arg_gbl10[d + 64 * thr] = ZERO_double; } } xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; xdim2 = args[2].dat->size[0]; xdim3 = args[3].dat->size[0]; xdim4 = args[4].dat->size[0]; xdim5 = args[5].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[49].mpi_time += t2 - t1; } #pragma omp parallel for for (int thr = 0; thr < nthreads; thr++) { int y_size = end[1] - start[1]; char *p_a[11]; int start_i = start[1] + ((y_size - 1) / nthreads + 1) * thr; int finish_i = start[1] + MIN(((y_size - 1) / nthreads + 1) * (thr + 1), y_size); // get address per thread int start0 = start[0]; int start1 = start_i; // 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 = dat0 * 1 * (start0 * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0 + dat0 * args[0].dat->size[0] * (start1 * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]); p_a[0] = (char *)args[0].data + base0; #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 = dat1 * 1 * (start0 * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); base1 = base1 + dat1 * args[1].dat->size[0] * (start1 * args[1].stencil->stride[1] - args[1].dat->base[1] - d_m[1]); p_a[1] = (char *)args[1].data + base1; #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 = dat2 * 1 * (start0 * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]); base2 = base2 + dat2 * args[2].dat->size[0] * (start1 * args[2].stencil->stride[1] - args[2].dat->base[1] - d_m[1]); p_a[2] = (char *)args[2].data + base2; #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 = dat3 * 1 * (start0 * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]); base3 = base3 + dat3 * args[3].dat->size[0] * (start1 * args[3].stencil->stride[1] - args[3].dat->base[1] - d_m[1]); p_a[3] = (char *)args[3].data + base3; #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 = dat4 * 1 * (start0 * args[4].stencil->stride[0] - args[4].dat->base[0] - d_m[0]); base4 = base4 + dat4 * args[4].dat->size[0] * (start1 * args[4].stencil->stride[1] - args[4].dat->base[1] - d_m[1]); p_a[4] = (char *)args[4].data + base4; #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 = dat5 * 1 * (start0 * args[5].stencil->stride[0] - args[5].dat->base[0] - d_m[0]); base5 = base5 + dat5 * args[5].dat->size[0] * (start1 * args[5].stencil->stride[1] - args[5].dat->base[1] - d_m[1]); p_a[5] = (char *)args[5].data + base5; p_a[6] = (char *)arg6h; p_a[7] = (char *)arg7h; p_a[8] = (char *)arg8h; p_a[9] = (char *)arg9h; p_a[10] = (char *)arg10h; for (int n_y = start_i; n_y < finish_i; n_y++) { for (int n_x = start[0]; n_x < start[0] + (end[0] - start[0]) / SIMD_VEC; n_x++) { // call kernel function, passing in pointers to data -vectorised for (int i = 0; i < SIMD_VEC; i++) { field_summary_kernel((const double *)p_a[0] + i * 1 * 1, (const double *)p_a[1] + i * 1 * 1, (const double *)p_a[2] + i * 1 * 1, (const double *)p_a[3] + i * 1 * 1, (const double *)p_a[4] + i * 1 * 1, (const double *)p_a[5] + i * 1 * 1, &arg_gbl6[64 * thr], &arg_gbl7[64 * thr], &arg_gbl8[64 * thr], &arg_gbl9[64 * thr], &arg_gbl10[64 * thr]); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC; p_a[2] = p_a[2] + (dat2 * off2_0) * SIMD_VEC; p_a[3] = p_a[3] + (dat3 * off3_0) * SIMD_VEC; p_a[4] = p_a[4] + (dat4 * off4_0) * SIMD_VEC; p_a[5] = p_a[5] + (dat5 * off5_0) * SIMD_VEC; } for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x < end[0]; n_x++) { // call kernel function, passing in pointers to data - remainder field_summary_kernel((const double *)p_a[0], (const double *)p_a[1], (const double *)p_a[2], (const double *)p_a[3], (const double *)p_a[4], (const double *)p_a[5], &arg_gbl6[64 * thr], &arg_gbl7[64 * thr], &arg_gbl8[64 * thr], &arg_gbl9[64 * thr], &arg_gbl10[64 * thr]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); p_a[1] = p_a[1] + (dat1 * off1_0); p_a[2] = p_a[2] + (dat2 * off2_0); p_a[3] = p_a[3] + (dat3 * off3_0); p_a[4] = p_a[4] + (dat4 * off4_0); p_a[5] = p_a[5] + (dat5 * off5_0); } // shift pointers to data y direction p_a[0] = p_a[0] + (dat0 * off0_1); p_a[1] = p_a[1] + (dat1 * off1_1); p_a[2] = p_a[2] + (dat2 * off2_1); p_a[3] = p_a[3] + (dat3 * off3_1); p_a[4] = p_a[4] + (dat4 * off4_1); p_a[5] = p_a[5] + (dat5 * off5_1); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[49].time += t1 - t2; } // combine reduction data for (int thr = 0; thr < nthreads; thr++) { for (int d = 0; d < 1; d++) { arg6h[d] += arg_gbl6[64 * thr + d]; } for (int d = 0; d < 1; d++) { arg7h[d] += arg_gbl7[64 * thr + d]; } for (int d = 0; d < 1; d++) { arg8h[d] += arg_gbl8[64 * thr + d]; } for (int d = 0; d < 1; d++) { arg9h[d] += arg_gbl9[64 * thr + d]; } for (int d = 0; d < 1; d++) { arg10h[d] += arg_gbl10[64 * thr + d]; } } ops_set_dirtybit_host(args, 11); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[49].mpi_time += t2 - t1; OPS_kernels[49].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[49].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[49].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[49].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[49].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[49].transfer += ops_compute_transfer(dim, start, end, &arg5); } }
// host stub function void ops_par_loop_calc_dt_kernel_print(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 args[7] = { arg0, arg1, arg2, arg3, arg4, arg5, arg6}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args,7,range,30)) return; #endif ops_timing_realloc(30,"calc_dt_kernel_print"); OPS_kernels[30].count++; //compute locally 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]); int xdim0 = args[0].dat->size[0]*args[0].dat->dim; int xdim1 = args[1].dat->size[0]*args[1].dat->dim; int xdim2 = args[2].dat->size[0]*args[2].dat->dim; int xdim3 = args[3].dat->size[0]*args[3].dat->dim; int xdim4 = args[4].dat->size[0]*args[4].dat->dim; int xdim5 = args[5].dat->size[0]*args[5].dat->dim; //build opencl kernel if not already built buildOpenCLKernels_calc_dt_kernel_print( xdim0,xdim1,xdim2,xdim3,xdim4,xdim5); //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, 1}; size_t localWorkSize[3] = {OPS_block_size_x,OPS_block_size_y,1}; #ifdef OPS_MPI double *arg6h = (double *)(((ops_reduction)args[6].data)->data + ((ops_reduction)args[6].data)->size * block->index); #else //OPS_MPI double *arg6h = (double *)(((ops_reduction)args[6].data)->data); #endif //OPS_MPI int nblocks = ((x_size-1)/OPS_block_size_x+ 1)*((y_size-1)/OPS_block_size_y + 1); int maxblocks = nblocks; int reduct_bytes = 0; reduct_bytes += ROUND_UP(maxblocks*12*sizeof(double)); reallocReductArrays(reduct_bytes); reduct_bytes = 0; int r_bytes6 = reduct_bytes/sizeof(double); arg6.data = OPS_reduct_h + reduct_bytes; arg6.data_d = OPS_reduct_d;// + reduct_bytes; for (int b=0; b<maxblocks; b++) for (int d=0; d<12; d++) ((double *)arg6.data)[d+b*12] = ZERO_double; reduct_bytes += ROUND_UP(maxblocks*12*sizeof(double)); mvReductArraysToDevice(reduct_bytes); 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; //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]); #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]); #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 = 1 * (start[0] * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]); base2 = base2 + args[2].dat->size[0] * (start[1] * args[2].stencil->stride[1] - args[2].dat->base[1] - d_m[1]); #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 = 1 * (start[0] * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]); base3 = base3 + args[3].dat->size[0] * (start[1] * args[3].stencil->stride[1] - args[3].dat->base[1] - d_m[1]); #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 = 1 * (start[0] * args[4].stencil->stride[0] - args[4].dat->base[0] - d_m[0]); base4 = base4 + args[4].dat->size[0] * (start[1] * args[4].stencil->stride[1] - args[4].dat->base[1] - d_m[1]); #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 = 1 * (start[0] * args[5].stencil->stride[0] - args[5].dat->base[0] - d_m[0]); base5 = base5 + args[5].dat->size[0] * (start[1] * args[5].stencil->stride[1] - args[5].dat->base[1] - d_m[1]); ops_H_D_exchanges_device(args, 7); ops_halo_exchanges(args,7,range); ops_H_D_exchanges_device(args, 7); ops_timers_core(&c1,&t1); OPS_kernels[30].mpi_time += t1-t2; int nthread = OPS_block_size_x*OPS_block_size_y; clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 0, sizeof(cl_mem), (void*) &arg0.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 1, sizeof(cl_mem), (void*) &arg1.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 2, sizeof(cl_mem), (void*) &arg2.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 3, sizeof(cl_mem), (void*) &arg3.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 4, sizeof(cl_mem), (void*) &arg4.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 5, sizeof(cl_mem), (void*) &arg5.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 6, sizeof(cl_mem), (void*) &arg6.data_d )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 7, nthread*sizeof(double), NULL)); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 8, sizeof(cl_int), (void*) &r_bytes6 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 9, sizeof(cl_int), (void*) &base0 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 10, sizeof(cl_int), (void*) &base1 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 11, sizeof(cl_int), (void*) &base2 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 12, sizeof(cl_int), (void*) &base3 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 13, sizeof(cl_int), (void*) &base4 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 14, sizeof(cl_int), (void*) &base5 )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 15, sizeof(cl_int), (void*) &x_size )); clSafeCall( clSetKernelArg(OPS_opencl_core.kernel[30], 16, sizeof(cl_int), (void*) &y_size )); //call/enque opencl kernel wrapper function clSafeCall( clEnqueueNDRangeKernel(OPS_opencl_core.command_queue, OPS_opencl_core.kernel[30], 3, NULL, globalWorkSize, localWorkSize, 0, NULL, NULL) ); if (OPS_diags>1) { clSafeCall( clFinish(OPS_opencl_core.command_queue) ); } mvReductArraysToHost(reduct_bytes); for ( int b=0; b<maxblocks; b++ ){ for ( int d=0; d<12; d++ ){ arg6h[d] = arg6h[d] + ((double *)arg6.data)[d+b*12]; } } arg6.data = (char *)arg6h; ops_set_dirtybit_device(args, 7); //Update kernel record ops_timers_core(&c2,&t2); OPS_kernels[30].time += t2-t1; OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg1); OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg2); OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg3); OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg4); OPS_kernels[30].transfer += ops_compute_transfer(dim, range, &arg5); }
// host stub function void ops_par_loop_update_halo_kernel4_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, 42)) return; #endif ops_timing_realloc(42, "update_halo_kernel4_minus_2_a"); OPS_kernels[42].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 for (int n = 0; n < 2; 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]); xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; // Timing double t1, t2, c1, c2; ops_timers_core(&c2, &t2); if (xdim0 != xdim0_update_halo_kernel4_minus_2_a_h || xdim1 != xdim1_update_halo_kernel4_minus_2_a_h) { xdim0_update_halo_kernel4_minus_2_a = xdim0; xdim0_update_halo_kernel4_minus_2_a_h = xdim0; xdim1_update_halo_kernel4_minus_2_a = xdim1; xdim1_update_halo_kernel4_minus_2_a_h = xdim1; } int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); int *arg2h = (int *)arg2.data; // 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 = 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]); double *p_a0 = (double *)((char *)args[0].data + base0); #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 = 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]); double *p_a1 = (double *)((char *)args[1].data + base1); int *p_a2 = arg2h; ops_H_D_exchanges_host(args, 3); ops_halo_exchanges(args, 3, range); ops_timers_core(&c1, &t1); OPS_kernels[42].mpi_time += t1 - t2; update_halo_kernel4_minus_2_a_c_wrapper(p_a0, p_a1, p_a2, x_size, y_size); ops_timers_core(&c2, &t2); OPS_kernels[42].time += t2 - t1; ops_set_dirtybit_host(args, 3); ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[1], range); // Update kernel record OPS_kernels[42].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[42].transfer += ops_compute_transfer(dim, start, end, &arg1); }
// host stub function void ops_par_loop_update_halo_kernel1_r1(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}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 8, range, 16)) return; #endif ops_timing_realloc(16, "update_halo_kernel1_r1"); OPS_kernels[16].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 for (int n = 0; n < 2; 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]); xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; xdim2 = args[2].dat->size[0]; xdim3 = args[3].dat->size[0]; xdim4 = args[4].dat->size[0]; xdim5 = args[5].dat->size[0]; xdim6 = args[6].dat->size[0]; // Timing double t1, t2, c1, c2; ops_timers_core(&c2, &t2); if (xdim0 != xdim0_update_halo_kernel1_r1_h || xdim1 != xdim1_update_halo_kernel1_r1_h || xdim2 != xdim2_update_halo_kernel1_r1_h || xdim3 != xdim3_update_halo_kernel1_r1_h || xdim4 != xdim4_update_halo_kernel1_r1_h || xdim5 != xdim5_update_halo_kernel1_r1_h || xdim6 != xdim6_update_halo_kernel1_r1_h) { xdim0_update_halo_kernel1_r1 = xdim0; xdim0_update_halo_kernel1_r1_h = xdim0; xdim1_update_halo_kernel1_r1 = xdim1; xdim1_update_halo_kernel1_r1_h = xdim1; xdim2_update_halo_kernel1_r1 = xdim2; xdim2_update_halo_kernel1_r1_h = xdim2; xdim3_update_halo_kernel1_r1 = xdim3; xdim3_update_halo_kernel1_r1_h = xdim3; xdim4_update_halo_kernel1_r1 = xdim4; xdim4_update_halo_kernel1_r1_h = xdim4; xdim5_update_halo_kernel1_r1 = xdim5; xdim5_update_halo_kernel1_r1_h = xdim5; xdim6_update_halo_kernel1_r1 = xdim6; xdim6_update_halo_kernel1_r1_h = xdim6; } int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); int dat2 = (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size); int dat3 = (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size); int dat4 = (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size); int dat5 = (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size); int dat6 = (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size); int *arg7h = (int *)arg7.data; // 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 = 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]); double *p_a0 = (double *)((char *)args[0].data + base0); #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 = 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]); double *p_a1 = (double *)((char *)args[1].data + base1); #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 = 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]); double *p_a2 = (double *)((char *)args[2].data + base2); #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 = 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]); double *p_a3 = (double *)((char *)args[3].data + base3); #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 = 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]); double *p_a4 = (double *)((char *)args[4].data + base4); #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 = 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]); double *p_a5 = (double *)((char *)args[5].data + base5); #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 = 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]); double *p_a6 = (double *)((char *)args[6].data + base6); int *p_a7 = arg7h; ops_H_D_exchanges_host(args, 8); ops_halo_exchanges(args, 8, range); ops_timers_core(&c1, &t1); OPS_kernels[16].mpi_time += t1 - t2; update_halo_kernel1_r1_c_wrapper(p_a0, p_a1, p_a2, p_a3, p_a4, p_a5, p_a6, p_a7, x_size, y_size); ops_timers_core(&c2, &t2); OPS_kernels[16].time += t2 - t1; ops_set_dirtybit_host(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); // Update kernel record OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg6); }
// host stub function void ops_par_loop_PdV_kernel_predict(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 arg8, ops_arg arg9, ops_arg arg10, ops_arg arg11, ops_arg arg12, ops_arg arg13) { char *p_a[14]; int offs[14][3]; ops_arg args[14] = { arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9, arg10, arg11, arg12, arg13}; ops_timing_realloc(5,"PdV_kernel_predict"); OPS_kernels[5].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 #ifdef OPS_DEBUG ops_register_args(args, "PdV_kernel_predict"); #endif offs[0][0] = args[0].stencil->stride[0]*1; //unit step in x dimension offs[0][1] = off3D(1, &start[0], &end[0],args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[0][2] = off3D(2, &start[0], &end[0],args[0].dat->size, args[0].stencil->stride) - offs[0][1] - offs[0][0]; offs[1][0] = args[1].stencil->stride[0]*1; //unit step in x dimension offs[1][1] = off3D(1, &start[0], &end[0],args[1].dat->size, args[1].stencil->stride) - offs[1][0]; offs[1][2] = off3D(2, &start[0], &end[0],args[1].dat->size, args[1].stencil->stride) - offs[1][1] - offs[1][0]; offs[2][0] = args[2].stencil->stride[0]*1; //unit step in x dimension offs[2][1] = off3D(1, &start[0], &end[0],args[2].dat->size, args[2].stencil->stride) - offs[2][0]; offs[2][2] = off3D(2, &start[0], &end[0],args[2].dat->size, args[2].stencil->stride) - offs[2][1] - offs[2][0]; offs[3][0] = args[3].stencil->stride[0]*1; //unit step in x dimension offs[3][1] = off3D(1, &start[0], &end[0],args[3].dat->size, args[3].stencil->stride) - offs[3][0]; offs[3][2] = off3D(2, &start[0], &end[0],args[3].dat->size, args[3].stencil->stride) - offs[3][1] - offs[3][0]; offs[4][0] = args[4].stencil->stride[0]*1; //unit step in x dimension offs[4][1] = off3D(1, &start[0], &end[0],args[4].dat->size, args[4].stencil->stride) - offs[4][0]; offs[4][2] = off3D(2, &start[0], &end[0],args[4].dat->size, args[4].stencil->stride) - offs[4][1] - offs[4][0]; offs[5][0] = args[5].stencil->stride[0]*1; //unit step in x dimension offs[5][1] = off3D(1, &start[0], &end[0],args[5].dat->size, args[5].stencil->stride) - offs[5][0]; offs[5][2] = off3D(2, &start[0], &end[0],args[5].dat->size, args[5].stencil->stride) - offs[5][1] - offs[5][0]; offs[6][0] = args[6].stencil->stride[0]*1; //unit step in x dimension offs[6][1] = off3D(1, &start[0], &end[0],args[6].dat->size, args[6].stencil->stride) - offs[6][0]; offs[6][2] = off3D(2, &start[0], &end[0],args[6].dat->size, args[6].stencil->stride) - offs[6][1] - offs[6][0]; offs[7][0] = args[7].stencil->stride[0]*1; //unit step in x dimension offs[7][1] = off3D(1, &start[0], &end[0],args[7].dat->size, args[7].stencil->stride) - offs[7][0]; offs[7][2] = off3D(2, &start[0], &end[0],args[7].dat->size, args[7].stencil->stride) - offs[7][1] - offs[7][0]; offs[8][0] = args[8].stencil->stride[0]*1; //unit step in x dimension offs[8][1] = off3D(1, &start[0], &end[0],args[8].dat->size, args[8].stencil->stride) - offs[8][0]; offs[8][2] = off3D(2, &start[0], &end[0],args[8].dat->size, args[8].stencil->stride) - offs[8][1] - offs[8][0]; offs[9][0] = args[9].stencil->stride[0]*1; //unit step in x dimension offs[9][1] = off3D(1, &start[0], &end[0],args[9].dat->size, args[9].stencil->stride) - offs[9][0]; offs[9][2] = off3D(2, &start[0], &end[0],args[9].dat->size, args[9].stencil->stride) - offs[9][1] - offs[9][0]; offs[10][0] = args[10].stencil->stride[0]*1; //unit step in x dimension offs[10][1] = off3D(1, &start[0], &end[0],args[10].dat->size, args[10].stencil->stride) - offs[10][0]; offs[10][2] = off3D(2, &start[0], &end[0],args[10].dat->size, args[10].stencil->stride) - offs[10][1] - offs[10][0]; offs[11][0] = args[11].stencil->stride[0]*1; //unit step in x dimension offs[11][1] = off3D(1, &start[0], &end[0],args[11].dat->size, args[11].stencil->stride) - offs[11][0]; offs[11][2] = off3D(2, &start[0], &end[0],args[11].dat->size, args[11].stencil->stride) - offs[11][1] - offs[11][0]; offs[12][0] = args[12].stencil->stride[0]*1; //unit step in x dimension offs[12][1] = off3D(1, &start[0], &end[0],args[12].dat->size, args[12].stencil->stride) - offs[12][0]; offs[12][2] = off3D(2, &start[0], &end[0],args[12].dat->size, args[12].stencil->stride) - offs[12][1] - offs[12][0]; offs[13][0] = args[13].stencil->stride[0]*1; //unit step in x dimension offs[13][1] = off3D(1, &start[0], &end[0],args[13].dat->size, args[13].stencil->stride) - offs[13][0]; offs[13][2] = off3D(2, &start[0], &end[0],args[13].dat->size, args[13].stencil->stride) - offs[13][1] - offs[13][0]; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int off0_2 = offs[0][2]; int dat0 = args[0].dat->elem_size; int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int off1_2 = offs[1][2]; int dat1 = args[1].dat->elem_size; int off2_0 = offs[2][0]; int off2_1 = offs[2][1]; int off2_2 = offs[2][2]; int dat2 = args[2].dat->elem_size; int off3_0 = offs[3][0]; int off3_1 = offs[3][1]; int off3_2 = offs[3][2]; int dat3 = args[3].dat->elem_size; int off4_0 = offs[4][0]; int off4_1 = offs[4][1]; int off4_2 = offs[4][2]; int dat4 = args[4].dat->elem_size; int off5_0 = offs[5][0]; int off5_1 = offs[5][1]; int off5_2 = offs[5][2]; int dat5 = args[5].dat->elem_size; int off6_0 = offs[6][0]; int off6_1 = offs[6][1]; int off6_2 = offs[6][2]; int dat6 = args[6].dat->elem_size; int off7_0 = offs[7][0]; int off7_1 = offs[7][1]; int off7_2 = offs[7][2]; int dat7 = args[7].dat->elem_size; int off8_0 = offs[8][0]; int off8_1 = offs[8][1]; int off8_2 = offs[8][2]; int dat8 = args[8].dat->elem_size; int off9_0 = offs[9][0]; int off9_1 = offs[9][1]; int off9_2 = offs[9][2]; int dat9 = args[9].dat->elem_size; int off10_0 = offs[10][0]; int off10_1 = offs[10][1]; int off10_2 = offs[10][2]; int dat10 = args[10].dat->elem_size; int off11_0 = offs[11][0]; int off11_1 = offs[11][1]; int off11_2 = offs[11][2]; int dat11 = args[11].dat->elem_size; int off12_0 = offs[12][0]; int off12_1 = offs[12][1]; int off12_2 = offs[12][2]; int dat12 = args[12].dat->elem_size; int off13_0 = offs[13][0]; int off13_1 = offs[13][1]; int off13_2 = offs[13][2]; int dat13 = args[13].dat->elem_size; //set up initial pointers and exchange halos if necessary 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]); p_a[0] = (char *)args[0].data + base0; #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]); p_a[1] = (char *)args[1].data + base1; #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]); p_a[2] = (char *)args[2].data + base2; #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]); p_a[3] = (char *)args[3].data + base3; #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]); p_a[4] = (char *)args[4].data + base4; #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]); p_a[5] = (char *)args[5].data + base5; #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]); p_a[6] = (char *)args[6].data + base6; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[7].dat->d_m[d] + OPS_sub_dat_list[args[7].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[7].dat->d_m[d]; #endif //OPS_MPI int base7 = dat7 * 1 * (start[0] * args[7].stencil->stride[0] - args[7].dat->base[0] - d_m[0]); base7 = base7+ dat7 * args[7].dat->size[0] * (start[1] * args[7].stencil->stride[1] - args[7].dat->base[1] - d_m[1]); base7 = base7+ dat7 * args[7].dat->size[0] * args[7].dat->size[1] * (start[2] * args[7].stencil->stride[2] - args[7].dat->base[2] - d_m[2]); p_a[7] = (char *)args[7].data + base7; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[8].dat->d_m[d] + OPS_sub_dat_list[args[8].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[8].dat->d_m[d]; #endif //OPS_MPI int base8 = dat8 * 1 * (start[0] * args[8].stencil->stride[0] - args[8].dat->base[0] - d_m[0]); base8 = base8+ dat8 * args[8].dat->size[0] * (start[1] * args[8].stencil->stride[1] - args[8].dat->base[1] - d_m[1]); base8 = base8+ dat8 * args[8].dat->size[0] * args[8].dat->size[1] * (start[2] * args[8].stencil->stride[2] - args[8].dat->base[2] - d_m[2]); p_a[8] = (char *)args[8].data + base8; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[9].dat->d_m[d] + OPS_sub_dat_list[args[9].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[9].dat->d_m[d]; #endif //OPS_MPI int base9 = dat9 * 1 * (start[0] * args[9].stencil->stride[0] - args[9].dat->base[0] - d_m[0]); base9 = base9+ dat9 * args[9].dat->size[0] * (start[1] * args[9].stencil->stride[1] - args[9].dat->base[1] - d_m[1]); base9 = base9+ dat9 * args[9].dat->size[0] * args[9].dat->size[1] * (start[2] * args[9].stencil->stride[2] - args[9].dat->base[2] - d_m[2]); p_a[9] = (char *)args[9].data + base9; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[10].dat->d_m[d] + OPS_sub_dat_list[args[10].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[10].dat->d_m[d]; #endif //OPS_MPI int base10 = dat10 * 1 * (start[0] * args[10].stencil->stride[0] - args[10].dat->base[0] - d_m[0]); base10 = base10+ dat10 * args[10].dat->size[0] * (start[1] * args[10].stencil->stride[1] - args[10].dat->base[1] - d_m[1]); base10 = base10+ dat10 * args[10].dat->size[0] * args[10].dat->size[1] * (start[2] * args[10].stencil->stride[2] - args[10].dat->base[2] - d_m[2]); p_a[10] = (char *)args[10].data + base10; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[11].dat->d_m[d] + OPS_sub_dat_list[args[11].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[11].dat->d_m[d]; #endif //OPS_MPI int base11 = dat11 * 1 * (start[0] * args[11].stencil->stride[0] - args[11].dat->base[0] - d_m[0]); base11 = base11+ dat11 * args[11].dat->size[0] * (start[1] * args[11].stencil->stride[1] - args[11].dat->base[1] - d_m[1]); base11 = base11+ dat11 * args[11].dat->size[0] * args[11].dat->size[1] * (start[2] * args[11].stencil->stride[2] - args[11].dat->base[2] - d_m[2]); p_a[11] = (char *)args[11].data + base11; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[12].dat->d_m[d] + OPS_sub_dat_list[args[12].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[12].dat->d_m[d]; #endif //OPS_MPI int base12 = dat12 * 1 * (start[0] * args[12].stencil->stride[0] - args[12].dat->base[0] - d_m[0]); base12 = base12+ dat12 * args[12].dat->size[0] * (start[1] * args[12].stencil->stride[1] - args[12].dat->base[1] - d_m[1]); base12 = base12+ dat12 * args[12].dat->size[0] * args[12].dat->size[1] * (start[2] * args[12].stencil->stride[2] - args[12].dat->base[2] - d_m[2]); p_a[12] = (char *)args[12].data + base12; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[13].dat->d_m[d] + OPS_sub_dat_list[args[13].dat->index]->d_im[d]; #else //OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[13].dat->d_m[d]; #endif //OPS_MPI int base13 = dat13 * 1 * (start[0] * args[13].stencil->stride[0] - args[13].dat->base[0] - d_m[0]); base13 = base13+ dat13 * args[13].dat->size[0] * (start[1] * args[13].stencil->stride[1] - args[13].dat->base[1] - d_m[1]); base13 = base13+ dat13 * args[13].dat->size[0] * args[13].dat->size[1] * (start[2] * args[13].stencil->stride[2] - args[13].dat->base[2] - d_m[2]); p_a[13] = (char *)args[13].data + base13; ops_H_D_exchanges_host(args, 14); ops_halo_exchanges(args,14,range); ops_H_D_exchanges_host(args, 14); ops_timers_core(&c1,&t1); OPS_kernels[5].mpi_time += t1-t2; 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]; xdim7 = args[7].dat->size[0]*args[7].dat->dim; ydim7 = args[7].dat->size[1]; xdim8 = args[8].dat->size[0]*args[8].dat->dim; ydim8 = args[8].dat->size[1]; xdim9 = args[9].dat->size[0]*args[9].dat->dim; ydim9 = args[9].dat->size[1]; xdim10 = args[10].dat->size[0]*args[10].dat->dim; ydim10 = args[10].dat->size[1]; xdim11 = args[11].dat->size[0]*args[11].dat->dim; ydim11 = args[11].dat->size[1]; xdim12 = args[12].dat->size[0]*args[12].dat->dim; ydim12 = args[12].dat->size[1]; xdim13 = args[13].dat->size[0]*args[13].dat->dim; ydim13 = args[13].dat->size[1]; int n_x; for ( int n_z=start[2]; n_z<end[2]; n_z++ ){ for ( int n_y=start[1]; n_y<end[1]; n_y++ ){ #pragma novector for( n_x=start[0]; n_x<start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x+=SIMD_VEC ) { //call kernel function, passing in pointers to data -vectorised #pragma simd for ( int i=0; i<SIMD_VEC; i++ ){ PdV_kernel_predict( (double *)p_a[0]+ i*1, (double *)p_a[1]+ i*1, (double *)p_a[2]+ i*1, (double *)p_a[3]+ i*1, (double *)p_a[4]+ i*1, (double *)p_a[5]+ i*1, (double *)p_a[6]+ i*1, (double *)p_a[7]+ i*1, (double *)p_a[8]+ i*1, (double *)p_a[9]+ i*1, (double *)p_a[10]+ i*1, (double *)p_a[11]+ i*1, (double *)p_a[12]+ i*1, (double *)p_a[13]+ i*1 ); } //shift pointers to data x direction p_a[0]= p_a[0] + (dat0 * off0_0)*SIMD_VEC; p_a[1]= p_a[1] + (dat1 * off1_0)*SIMD_VEC; p_a[2]= p_a[2] + (dat2 * off2_0)*SIMD_VEC; p_a[3]= p_a[3] + (dat3 * off3_0)*SIMD_VEC; p_a[4]= p_a[4] + (dat4 * off4_0)*SIMD_VEC; p_a[5]= p_a[5] + (dat5 * off5_0)*SIMD_VEC; p_a[6]= p_a[6] + (dat6 * off6_0)*SIMD_VEC; p_a[7]= p_a[7] + (dat7 * off7_0)*SIMD_VEC; p_a[8]= p_a[8] + (dat8 * off8_0)*SIMD_VEC; p_a[9]= p_a[9] + (dat9 * off9_0)*SIMD_VEC; p_a[10]= p_a[10] + (dat10 * off10_0)*SIMD_VEC; p_a[11]= p_a[11] + (dat11 * off11_0)*SIMD_VEC; p_a[12]= p_a[12] + (dat12 * off12_0)*SIMD_VEC; p_a[13]= p_a[13] + (dat13 * off13_0)*SIMD_VEC; } for ( int n_x=start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x<end[0]; n_x++ ){ //call kernel function, passing in pointers to data - remainder PdV_kernel_predict( (double *)p_a[0], (double *)p_a[1], (double *)p_a[2], (double *)p_a[3], (double *)p_a[4], (double *)p_a[5], (double *)p_a[6], (double *)p_a[7], (double *)p_a[8], (double *)p_a[9], (double *)p_a[10], (double *)p_a[11], (double *)p_a[12], (double *)p_a[13] ); //shift pointers to data x direction p_a[0]= p_a[0] + (dat0 * off0_0); p_a[1]= p_a[1] + (dat1 * off1_0); p_a[2]= p_a[2] + (dat2 * off2_0); p_a[3]= p_a[3] + (dat3 * off3_0); p_a[4]= p_a[4] + (dat4 * off4_0); p_a[5]= p_a[5] + (dat5 * off5_0); p_a[6]= p_a[6] + (dat6 * off6_0); p_a[7]= p_a[7] + (dat7 * off7_0); p_a[8]= p_a[8] + (dat8 * off8_0); p_a[9]= p_a[9] + (dat9 * off9_0); p_a[10]= p_a[10] + (dat10 * off10_0); p_a[11]= p_a[11] + (dat11 * off11_0); p_a[12]= p_a[12] + (dat12 * off12_0); p_a[13]= p_a[13] + (dat13 * off13_0); } //shift pointers to data y direction p_a[0]= p_a[0] + (dat0 * off0_1); p_a[1]= p_a[1] + (dat1 * off1_1); p_a[2]= p_a[2] + (dat2 * off2_1); p_a[3]= p_a[3] + (dat3 * off3_1); p_a[4]= p_a[4] + (dat4 * off4_1); p_a[5]= p_a[5] + (dat5 * off5_1); p_a[6]= p_a[6] + (dat6 * off6_1); p_a[7]= p_a[7] + (dat7 * off7_1); p_a[8]= p_a[8] + (dat8 * off8_1); p_a[9]= p_a[9] + (dat9 * off9_1); p_a[10]= p_a[10] + (dat10 * off10_1); p_a[11]= p_a[11] + (dat11 * off11_1); p_a[12]= p_a[12] + (dat12 * off12_1); p_a[13]= p_a[13] + (dat13 * off13_1); } //shift pointers to data z direction p_a[0]= p_a[0] + (dat0 * off0_2); p_a[1]= p_a[1] + (dat1 * off1_2); p_a[2]= p_a[2] + (dat2 * off2_2); p_a[3]= p_a[3] + (dat3 * off3_2); p_a[4]= p_a[4] + (dat4 * off4_2); p_a[5]= p_a[5] + (dat5 * off5_2); p_a[6]= p_a[6] + (dat6 * off6_2); p_a[7]= p_a[7] + (dat7 * off7_2); p_a[8]= p_a[8] + (dat8 * off8_2); p_a[9]= p_a[9] + (dat9 * off9_2); p_a[10]= p_a[10] + (dat10 * off10_2); p_a[11]= p_a[11] + (dat11 * off11_2); p_a[12]= p_a[12] + (dat12 * off12_2); p_a[13]= p_a[13] + (dat13 * off13_2); } ops_timers_core(&c2,&t2); OPS_kernels[5].time += t2-t1; ops_set_dirtybit_host(args, 14); ops_set_halo_dirtybit3(&args[4],range); ops_set_halo_dirtybit3(&args[8],range); ops_set_halo_dirtybit3(&args[11],range); //Update kernel record OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg1); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg2); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg3); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg4); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg5); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg6); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg7); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg8); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg9); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg10); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg11); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg12); OPS_kernels[5].transfer += ops_compute_transfer(dim, range, &arg13); }
// host stub function void ops_par_loop_reset_field_kernel1_execute(ops_kernel_descriptor *desc) { ops_block block = desc->block; int dim = desc->dim; int *range = desc->range; ops_arg arg0 = desc->args[0]; ops_arg arg1 = desc->args[1]; ops_arg arg2 = desc->args[2]; ops_arg arg3 = desc->args[3]; // Timing double t1, t2, c1, c2; ops_arg args[4] = {arg0, arg1, arg2, arg3}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 4, range, 139)) return; #endif if (OPS_diags > 1) { OPS_kernels[139].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[3]; int end[3]; for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #ifdef OPS_DEBUG ops_register_args(args, "reset_field_kernel1"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ density0 = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ density1 = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ energy0 = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ energy1 = (double *)(args[3].data + base3); // initialize global variable with the dimension of dats int xdim0_reset_field_kernel1 = args[0].dat->size[0]; int ydim0_reset_field_kernel1 = args[0].dat->size[1]; int xdim1_reset_field_kernel1 = args[1].dat->size[0]; int ydim1_reset_field_kernel1 = args[1].dat->size[1]; int xdim2_reset_field_kernel1 = args[2].dat->size[0]; int ydim2_reset_field_kernel1 = args[2].dat->size[1]; int xdim3_reset_field_kernel1 = args[3].dat->size[0]; int ydim3_reset_field_kernel1 = args[3].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[139].mpi_time += t1 - t2; } #pragma omp parallel for collapse(2) for (int n_z = start[2]; n_z < end[2]; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd aligned(density0, density1, energy0, energy1) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { density0[OPS_ACC0(0, 0, 0)] = density1[OPS_ACC1(0, 0, 0)]; energy0[OPS_ACC2(0, 0, 0)] = energy1[OPS_ACC3(0, 0, 0)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[139].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[139].mpi_time += t1 - t2; OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[139].transfer += ops_compute_transfer(dim, start, end, &arg3); } }
// host stub function void ops_par_loop_flux_calc_kernelx_execute(ops_kernel_descriptor *desc) { ops_block block = desc->block; int dim = desc->dim; int *range = desc->range; ops_arg arg0 = desc->args[0]; ops_arg arg1 = desc->args[1]; ops_arg arg2 = desc->args[2]; ops_arg arg3 = desc->args[3]; // Timing double t1, t2, c1, c2; ops_arg args[4] = {arg0, arg1, arg2, arg3}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 4, range, 59)) return; #endif if (OPS_diags > 1) { OPS_kernels[59].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[2]; int end[2]; for (int n = 0; n < 2; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #ifdef OPS_DEBUG ops_register_args(args, "flux_calc_kernelx"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ vol_flux_x = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ xarea = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; const double *__restrict__ xvel0 = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ xvel1 = (double *)(args[3].data + base3); // initialize global variable with the dimension of dats int xdim0_flux_calc_kernelx = args[0].dat->size[0]; int xdim1_flux_calc_kernelx = args[1].dat->size[0]; int xdim2_flux_calc_kernelx = args[2].dat->size[0]; int xdim3_flux_calc_kernelx = args[3].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[59].mpi_time += t1 - t2; } #pragma omp parallel for for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd aligned(vol_flux_x, xarea, xvel0, xvel1) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { vol_flux_x[OPS_ACC0(0, 0)] = 0.25 * dt * (xarea[OPS_ACC1(0, 0)]) * ((xvel0[OPS_ACC2(0, 0)]) + (xvel0[OPS_ACC2(0, 1)]) + (xvel1[OPS_ACC3(0, 0)]) + (xvel1[OPS_ACC3(0, 1)])); } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[59].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[59].mpi_time += t1 - t2; OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[59].transfer += ops_compute_transfer(dim, start, end, &arg3); } }
// host stub function void ops_par_loop_PdV_kernel_nopredict( 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 arg8, ops_arg arg9, ops_arg arg10, ops_arg arg11, ops_arg arg12, ops_arg arg13, ops_arg arg14, ops_arg arg15, ops_arg arg16) { // Timing double t1, t2, c1, c2; ops_arg args[17] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9, arg10, arg11, arg12, arg13, arg14, arg15, arg16}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 17, range, 103)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(103, "PdV_kernel_nopredict"); OPS_kernels[103].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]; int xdim7 = args[7].dat->size[0]; int ydim7 = args[7].dat->size[1]; int xdim8 = args[8].dat->size[0]; int ydim8 = args[8].dat->size[1]; int xdim9 = args[9].dat->size[0]; int ydim9 = args[9].dat->size[1]; int xdim10 = args[10].dat->size[0]; int ydim10 = args[10].dat->size[1]; int xdim11 = args[11].dat->size[0]; int ydim11 = args[11].dat->size[1]; int xdim12 = args[12].dat->size[0]; int ydim12 = args[12].dat->size[1]; int xdim13 = args[13].dat->size[0]; int ydim13 = args[13].dat->size[1]; int xdim14 = args[14].dat->size[0]; int ydim14 = args[14].dat->size[1]; int xdim15 = args[15].dat->size[0]; int ydim15 = args[15].dat->size[1]; int xdim16 = args[16].dat->size[0]; int ydim16 = args[16].dat->size[1]; // build opencl kernel if not already built buildOpenCLKernels_PdV_kernel_nopredict( xdim0, ydim0, xdim1, ydim1, xdim2, ydim2, xdim3, ydim3, xdim4, ydim4, xdim5, ydim5, xdim6, ydim6, xdim7, ydim7, xdim8, ydim8, xdim9, ydim9, xdim10, ydim10, xdim11, ydim11, xdim12, ydim12, xdim13, ydim13, xdim14, ydim14, xdim15, ydim15, xdim16, ydim16); // 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}; // 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]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[7].dat->d_m[d] + OPS_sub_dat_list[args[7].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[7].dat->d_m[d]; #endif int base7 = 1 * 1 * (start[0] * args[7].stencil->stride[0] - args[7].dat->base[0] - d_m[0]); base7 = base7 + args[7].dat->size[0] * 1 * (start[1] * args[7].stencil->stride[1] - args[7].dat->base[1] - d_m[1]); base7 = base7 + args[7].dat->size[0] * 1 * args[7].dat->size[1] * 1 * (start[2] * args[7].stencil->stride[2] - args[7].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[8].dat->d_m[d] + OPS_sub_dat_list[args[8].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[8].dat->d_m[d]; #endif int base8 = 1 * 1 * (start[0] * args[8].stencil->stride[0] - args[8].dat->base[0] - d_m[0]); base8 = base8 + args[8].dat->size[0] * 1 * (start[1] * args[8].stencil->stride[1] - args[8].dat->base[1] - d_m[1]); base8 = base8 + args[8].dat->size[0] * 1 * args[8].dat->size[1] * 1 * (start[2] * args[8].stencil->stride[2] - args[8].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[9].dat->d_m[d] + OPS_sub_dat_list[args[9].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[9].dat->d_m[d]; #endif int base9 = 1 * 1 * (start[0] * args[9].stencil->stride[0] - args[9].dat->base[0] - d_m[0]); base9 = base9 + args[9].dat->size[0] * 1 * (start[1] * args[9].stencil->stride[1] - args[9].dat->base[1] - d_m[1]); base9 = base9 + args[9].dat->size[0] * 1 * args[9].dat->size[1] * 1 * (start[2] * args[9].stencil->stride[2] - args[9].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[10].dat->d_m[d] + OPS_sub_dat_list[args[10].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[10].dat->d_m[d]; #endif int base10 = 1 * 1 * (start[0] * args[10].stencil->stride[0] - args[10].dat->base[0] - d_m[0]); base10 = base10 + args[10].dat->size[0] * 1 * (start[1] * args[10].stencil->stride[1] - args[10].dat->base[1] - d_m[1]); base10 = base10 + args[10].dat->size[0] * 1 * args[10].dat->size[1] * 1 * (start[2] * args[10].stencil->stride[2] - args[10].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[11].dat->d_m[d] + OPS_sub_dat_list[args[11].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[11].dat->d_m[d]; #endif int base11 = 1 * 1 * (start[0] * args[11].stencil->stride[0] - args[11].dat->base[0] - d_m[0]); base11 = base11 + args[11].dat->size[0] * 1 * (start[1] * args[11].stencil->stride[1] - args[11].dat->base[1] - d_m[1]); base11 = base11 + args[11].dat->size[0] * 1 * args[11].dat->size[1] * 1 * (start[2] * args[11].stencil->stride[2] - args[11].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[12].dat->d_m[d] + OPS_sub_dat_list[args[12].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[12].dat->d_m[d]; #endif int base12 = 1 * 1 * (start[0] * args[12].stencil->stride[0] - args[12].dat->base[0] - d_m[0]); base12 = base12 + args[12].dat->size[0] * 1 * (start[1] * args[12].stencil->stride[1] - args[12].dat->base[1] - d_m[1]); base12 = base12 + args[12].dat->size[0] * 1 * args[12].dat->size[1] * 1 * (start[2] * args[12].stencil->stride[2] - args[12].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[13].dat->d_m[d] + OPS_sub_dat_list[args[13].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[13].dat->d_m[d]; #endif int base13 = 1 * 1 * (start[0] * args[13].stencil->stride[0] - args[13].dat->base[0] - d_m[0]); base13 = base13 + args[13].dat->size[0] * 1 * (start[1] * args[13].stencil->stride[1] - args[13].dat->base[1] - d_m[1]); base13 = base13 + args[13].dat->size[0] * 1 * args[13].dat->size[1] * 1 * (start[2] * args[13].stencil->stride[2] - args[13].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[14].dat->d_m[d] + OPS_sub_dat_list[args[14].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[14].dat->d_m[d]; #endif int base14 = 1 * 1 * (start[0] * args[14].stencil->stride[0] - args[14].dat->base[0] - d_m[0]); base14 = base14 + args[14].dat->size[0] * 1 * (start[1] * args[14].stencil->stride[1] - args[14].dat->base[1] - d_m[1]); base14 = base14 + args[14].dat->size[0] * 1 * args[14].dat->size[1] * 1 * (start[2] * args[14].stencil->stride[2] - args[14].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[15].dat->d_m[d] + OPS_sub_dat_list[args[15].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[15].dat->d_m[d]; #endif int base15 = 1 * 1 * (start[0] * args[15].stencil->stride[0] - args[15].dat->base[0] - d_m[0]); base15 = base15 + args[15].dat->size[0] * 1 * (start[1] * args[15].stencil->stride[1] - args[15].dat->base[1] - d_m[1]); base15 = base15 + args[15].dat->size[0] * 1 * args[15].dat->size[1] * 1 * (start[2] * args[15].stencil->stride[2] - args[15].dat->base[2] - d_m[2]); #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[16].dat->d_m[d] + OPS_sub_dat_list[args[16].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[16].dat->d_m[d]; #endif int base16 = 1 * 1 * (start[0] * args[16].stencil->stride[0] - args[16].dat->base[0] - d_m[0]); base16 = base16 + args[16].dat->size[0] * 1 * (start[1] * args[16].stencil->stride[1] - args[16].dat->base[1] - d_m[1]); base16 = base16 + args[16].dat->size[0] * 1 * args[16].dat->size[1] * 1 * (start[2] * args[16].stencil->stride[2] - args[16].dat->base[2] - d_m[2]); ops_H_D_exchanges_device(args, 17); ops_halo_exchanges(args, 17, range); ops_H_D_exchanges_device(args, 17); if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[103].mpi_time += t2 - t1; } if (globalWorkSize[0] > 0 && globalWorkSize[1] > 0 && globalWorkSize[2] > 0) { clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 0, sizeof(cl_mem), (void *)&arg0.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 1, sizeof(cl_mem), (void *)&arg1.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 2, sizeof(cl_mem), (void *)&arg2.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 3, sizeof(cl_mem), (void *)&arg3.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 4, sizeof(cl_mem), (void *)&arg4.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 5, sizeof(cl_mem), (void *)&arg5.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 6, sizeof(cl_mem), (void *)&arg6.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 7, sizeof(cl_mem), (void *)&arg7.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 8, sizeof(cl_mem), (void *)&arg8.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 9, sizeof(cl_mem), (void *)&arg9.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 10, sizeof(cl_mem), (void *)&arg10.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 11, sizeof(cl_mem), (void *)&arg11.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 12, sizeof(cl_mem), (void *)&arg12.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 13, sizeof(cl_mem), (void *)&arg13.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 14, sizeof(cl_mem), (void *)&arg14.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 15, sizeof(cl_mem), (void *)&arg15.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 16, sizeof(cl_mem), (void *)&arg16.data_d)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 17, sizeof(cl_double), (void *)&dt)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 18, sizeof(cl_int), (void *)&base0)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 19, sizeof(cl_int), (void *)&base1)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 20, sizeof(cl_int), (void *)&base2)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 21, sizeof(cl_int), (void *)&base3)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 22, sizeof(cl_int), (void *)&base4)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 23, sizeof(cl_int), (void *)&base5)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 24, sizeof(cl_int), (void *)&base6)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 25, sizeof(cl_int), (void *)&base7)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 26, sizeof(cl_int), (void *)&base8)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 27, sizeof(cl_int), (void *)&base9)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 28, sizeof(cl_int), (void *)&base10)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 29, sizeof(cl_int), (void *)&base11)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 30, sizeof(cl_int), (void *)&base12)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 31, sizeof(cl_int), (void *)&base13)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 32, sizeof(cl_int), (void *)&base14)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 33, sizeof(cl_int), (void *)&base15)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 34, sizeof(cl_int), (void *)&base16)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 35, sizeof(cl_int), (void *)&x_size)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 36, sizeof(cl_int), (void *)&y_size)); clSafeCall(clSetKernelArg(OPS_opencl_core.kernel[103], 37, sizeof(cl_int), (void *)&z_size)); // call/enque opencl kernel wrapper function clSafeCall(clEnqueueNDRangeKernel( OPS_opencl_core.command_queue, OPS_opencl_core.kernel[103], 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[103].time += t1 - t2; } ops_set_dirtybit_device(args, 17); ops_set_halo_dirtybit3(&args[6], range); ops_set_halo_dirtybit3(&args[10], range); ops_set_halo_dirtybit3(&args[13], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[103].mpi_time += t2 - t1; OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg6); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg7); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg8); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg9); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg10); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg11); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg12); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg13); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg14); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg15); OPS_kernels[103].transfer += ops_compute_transfer(dim, start, end, &arg16); } }
// host stub function void ops_par_loop_update_halo_kernel4_minus_4_b(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; char *p_a[3]; int offs[3][2]; ops_arg args[3] = {arg0, arg1, arg2}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 3, range, 43)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(43, "update_halo_kernel4_minus_4_b"); OPS_kernels[43].count++; ops_timers_core(&c2, &t2); } // compute locally 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]; #endif #ifdef OPS_DEBUG ops_register_args(args, "update_halo_kernel4_minus_4_b"); #endif int arg_idx[2]; int arg_idx_base[2]; #ifdef OPS_MPI if (compute_ranges(args, 3, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 2; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif // OPS_MPI for (int n = 0; n < 2; n++) { arg_idx_base[n] = arg_idx[n]; } offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[0][1] = off2D(1, &start[0], &end[0], args[0].dat->size, args[0].stencil->stride) - offs[0][0]; offs[1][0] = args[1].stencil->stride[0] * 1; // unit step in x dimension offs[1][1] = off2D(1, &start[0], &end[0], args[1].dat->size, args[1].stencil->stride) - offs[1][0]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); // set up initial pointers and exchange halos if necessary 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]; p_a[0] = (char *)args[0].data + base0; 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]; p_a[1] = (char *)args[1].data + base1; p_a[2] = args[2].data; // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; // Halo Exchanges ops_H_D_exchanges_host(args, 3); ops_halo_exchanges(args, 3, range); ops_H_D_exchanges_host(args, 3); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[43].mpi_time += t1 - t2; } int n_x; for (int n_y = start[1]; n_y < end[1]; n_y++) { #pragma novector for (n_x = start[0]; n_x < start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x += SIMD_VEC) { // call kernel function, passing in pointers to data -vectorised #pragma simd for (int i = 0; i < SIMD_VEC; i++) { update_halo_kernel4_minus_4_b((double *)p_a[0] + i * 1 * 1, (double *)p_a[1] + i * 1 * 1, (int *)p_a[2]); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC; } for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x < end[0]; n_x++) { // call kernel function, passing in pointers to data - remainder update_halo_kernel4_minus_4_b((double *)p_a[0], (double *)p_a[1], (int *)p_a[2]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); p_a[1] = p_a[1] + (dat1 * off1_0); } // shift pointers to data y direction p_a[0] = p_a[0] + (dat0 * off0_1); p_a[1] = p_a[1] + (dat1 * off1_1); } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[43].time += t2 - t1; } ops_set_dirtybit_host(args, 3); ops_set_halo_dirtybit3(&args[0], range); ops_set_halo_dirtybit3(&args[1], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[43].mpi_time += t1 - t2; OPS_kernels[43].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[43].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_updateRK3_kernel(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 arg8, ops_arg arg9, ops_arg arg10) { // Timing double t1, t2, c1, c2; int offs[11][1]; ops_arg args[11] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9, arg10}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 11, range, 6)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(6, "updateRK3_kernel"); OPS_kernels[6].count++; ops_timers_core(&c1, &t1); } #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // compute locally allocated range for the sub-block int start[1]; int end[1]; int arg_idx[1]; #ifdef OPS_MPI if (!sb->owned) return; for (int n = 0; n < 1; 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]); if (end[n] < start[n]) end[n] = start[n]; } #else for (int n = 0; n < 1; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #endif #ifdef OPS_DEBUG ops_register_args(args, "updateRK3_kernel"); #endif offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[1][0] = args[1].stencil->stride[0] * 1; // unit step in x dimension offs[2][0] = args[2].stencil->stride[0] * 1; // unit step in x dimension offs[3][0] = args[3].stencil->stride[0] * 1; // unit step in x dimension offs[4][0] = args[4].stencil->stride[0] * 1; // unit step in x dimension offs[5][0] = args[5].stencil->stride[0] * 1; // unit step in x dimension offs[6][0] = args[6].stencil->stride[0] * 1; // unit step in x dimension offs[7][0] = args[7].stencil->stride[0] * 1; // unit step in x dimension offs[8][0] = args[8].stencil->stride[0] * 1; // unit step in x dimension int off0_0 = offs[0][0]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int off1_0 = offs[1][0]; int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); int off2_0 = offs[2][0]; int dat2 = (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size); int off3_0 = offs[3][0]; int dat3 = (OPS_soa ? args[3].dat->type_size : args[3].dat->elem_size); int off4_0 = offs[4][0]; int dat4 = (OPS_soa ? args[4].dat->type_size : args[4].dat->elem_size); int off5_0 = offs[5][0]; int dat5 = (OPS_soa ? args[5].dat->type_size : args[5].dat->elem_size); int off6_0 = offs[6][0]; int dat6 = (OPS_soa ? args[6].dat->type_size : args[6].dat->elem_size); int off7_0 = offs[7][0]; int dat7 = (OPS_soa ? args[7].dat->type_size : args[7].dat->elem_size); int off8_0 = offs[8][0]; int dat8 = (OPS_soa ? args[8].dat->type_size : args[8].dat->elem_size); // Halo Exchanges ops_H_D_exchanges_host(args, 11); ops_halo_exchanges(args, 11, range); ops_H_D_exchanges_host(args, 11); #ifdef _OPENMP int nthreads = omp_get_max_threads(); #else int nthreads = 1; #endif xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; xdim2 = args[2].dat->size[0]; xdim3 = args[3].dat->size[0]; xdim4 = args[4].dat->size[0]; xdim5 = args[5].dat->size[0]; xdim6 = args[6].dat->size[0]; xdim7 = args[7].dat->size[0]; xdim8 = args[8].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[6].mpi_time += t2 - t1; } #pragma omp parallel for for (int thr = 0; thr < nthreads; thr++) { int x_size = end[0] - start[0]; char *p_a[11]; int start_i = start[0] + ((x_size - 1) / nthreads + 1) * thr; int finish_i = start[0] + MIN(((x_size - 1) / nthreads + 1) * (thr + 1), x_size); // get address per thread int start0 = start_i; // 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 = dat0 * 1 * (start0 * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); p_a[0] = (char *)args[0].data + base0; #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 = dat1 * 1 * (start0 * args[1].stencil->stride[0] - args[1].dat->base[0] - d_m[0]); p_a[1] = (char *)args[1].data + base1; #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 = dat2 * 1 * (start0 * args[2].stencil->stride[0] - args[2].dat->base[0] - d_m[0]); p_a[2] = (char *)args[2].data + base2; #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 = dat3 * 1 * (start0 * args[3].stencil->stride[0] - args[3].dat->base[0] - d_m[0]); p_a[3] = (char *)args[3].data + base3; #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 = dat4 * 1 * (start0 * args[4].stencil->stride[0] - args[4].dat->base[0] - d_m[0]); p_a[4] = (char *)args[4].data + base4; #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 = dat5 * 1 * (start0 * args[5].stencil->stride[0] - args[5].dat->base[0] - d_m[0]); p_a[5] = (char *)args[5].data + base5; #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 = dat6 * 1 * (start0 * args[6].stencil->stride[0] - args[6].dat->base[0] - d_m[0]); p_a[6] = (char *)args[6].data + base6; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[7].dat->d_m[d] + OPS_sub_dat_list[args[7].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[7].dat->d_m[d]; #endif int base7 = dat7 * 1 * (start0 * args[7].stencil->stride[0] - args[7].dat->base[0] - d_m[0]); p_a[7] = (char *)args[7].data + base7; #ifdef OPS_MPI for (int d = 0; d < dim; d++) d_m[d] = args[8].dat->d_m[d] + OPS_sub_dat_list[args[8].dat->index]->d_im[d]; #else for (int d = 0; d < dim; d++) d_m[d] = args[8].dat->d_m[d]; #endif int base8 = dat8 * 1 * (start0 * args[8].stencil->stride[0] - args[8].dat->base[0] - d_m[0]); p_a[8] = (char *)args[8].data + base8; p_a[9] = (char *)args[9].data; p_a[10] = (char *)args[10].data; for (int n_x = start_i; n_x < start_i + (finish_i - start_i) / SIMD_VEC; n_x++) { // call kernel function, passing in pointers to data -vectorised #pragma simd for (int i = 0; i < SIMD_VEC; i++) { updateRK3_kernel( (double *)p_a[0] + i * 1 * 1, (double *)p_a[1] + i * 1 * 1, (double *)p_a[2] + i * 1 * 1, (double *)p_a[3] + i * 1 * 1, (double *)p_a[4] + i * 1 * 1, (double *)p_a[5] + i * 1 * 1, (const double *)p_a[6] + i * 1 * 1, (const double *)p_a[7] + i * 1 * 1, (const double *)p_a[8] + i * 1 * 1, (double *)p_a[9], (double *)p_a[10]); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC; p_a[2] = p_a[2] + (dat2 * off2_0) * SIMD_VEC; p_a[3] = p_a[3] + (dat3 * off3_0) * SIMD_VEC; p_a[4] = p_a[4] + (dat4 * off4_0) * SIMD_VEC; p_a[5] = p_a[5] + (dat5 * off5_0) * SIMD_VEC; p_a[6] = p_a[6] + (dat6 * off6_0) * SIMD_VEC; p_a[7] = p_a[7] + (dat7 * off7_0) * SIMD_VEC; p_a[8] = p_a[8] + (dat8 * off8_0) * SIMD_VEC; } for (int n_x = start_i + ((finish_i - start_i) / SIMD_VEC) * SIMD_VEC; n_x < finish_i; n_x++) { // call kernel function, passing in pointers to data - remainder updateRK3_kernel((double *)p_a[0], (double *)p_a[1], (double *)p_a[2], (double *)p_a[3], (double *)p_a[4], (double *)p_a[5], (const double *)p_a[6], (const double *)p_a[7], (const double *)p_a[8], (double *)p_a[9], (double *)p_a[10]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); p_a[1] = p_a[1] + (dat1 * off1_0); p_a[2] = p_a[2] + (dat2 * off2_0); p_a[3] = p_a[3] + (dat3 * off3_0); p_a[4] = p_a[4] + (dat4 * off4_0); p_a[5] = p_a[5] + (dat5 * off5_0); p_a[6] = p_a[6] + (dat6 * off6_0); p_a[7] = p_a[7] + (dat7 * off7_0); p_a[8] = p_a[8] + (dat8 * off8_0); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[6].time += t1 - t2; } ops_set_dirtybit_host(args, 11); 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); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[6].mpi_time += t2 - t1; OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg6); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg7); OPS_kernels[6].transfer += ops_compute_transfer(dim, start, end, &arg8); } }
// host stub function void ops_par_loop_gridgen_kernel(char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1) { // Timing double t1, t2, c1, c2; ops_arg args[2] = {arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 2, range, 0)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(0, "gridgen_kernel"); OPS_kernels[0].count++; ops_timers_core(&c1, &t1); } // compute localy allocated range for the sub-block int start[1]; int end[1]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif // OPS_MPI int arg_idx[1]; int arg_idx_base[1]; #ifdef OPS_MPI if (compute_ranges(args, 2, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 1; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif for (int n = 0; n < 1; n++) { arg_idx_base[n] = arg_idx[n]; } int dat0 = args[0].dat->elem_size; // 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]; #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 *p_a1 = NULL; int x_size = MAX(0, end[0] - start[0]); // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; if (xdim0 != xdim0_gridgen_kernel_h) { xdim0_gridgen_kernel = xdim0; xdim0_gridgen_kernel_h = xdim0; } // Halo Exchanges #ifdef OPS_GPU ops_H_D_exchanges_device(args, 2); #else ops_H_D_exchanges_host(args, 2); #endif ops_halo_exchanges(args, 2, range); #ifdef OPS_GPU ops_H_D_exchanges_device(args, 2); #else ops_H_D_exchanges_host(args, 2); #endif if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[0].mpi_time += t2 - t1; } gridgen_kernel_c_wrapper(p_a0, p_a1, arg_idx[0], x_size); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[0].time += t1 - t2; } #ifdef OPS_GPU ops_set_dirtybit_device(args, 2); #else ops_set_dirtybit_host(args, 2); #endif ops_set_halo_dirtybit3(&args[0], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[0].mpi_time += t2 - t1; OPS_kernels[0].transfer += ops_compute_transfer(dim, start, end, &arg0); } }
// host stub function void ops_par_loop_update_halo_kernel1_t1_execute(ops_kernel_descriptor *desc) { ops_block block = desc->block; int dim = desc->dim; int *range = desc->range; ops_arg arg0 = desc->args[0]; ops_arg arg1 = desc->args[1]; ops_arg arg2 = desc->args[2]; ops_arg arg3 = desc->args[3]; ops_arg arg4 = desc->args[4]; ops_arg arg5 = desc->args[5]; ops_arg arg6 = desc->args[6]; ops_arg arg7 = desc->args[7]; // 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, 15)) return; #endif if (OPS_diags > 1) { OPS_kernels[15].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[3]; int end[3]; for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; } #ifdef OPS_DEBUG ops_register_args(args, "update_halo_kernel1_t1"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ density0 = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; double *__restrict__ density1 = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ energy0 = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; double *__restrict__ energy1 = (double *)(args[3].data + base3); int base4 = args[4].dat->base_offset; double *__restrict__ pressure = (double *)(args[4].data + base4); int base5 = args[5].dat->base_offset; double *__restrict__ viscosity = (double *)(args[5].data + base5); int base6 = args[6].dat->base_offset; double *__restrict__ soundspeed = (double *)(args[6].data + base6); const int *__restrict__ fields = (int *)args[7].data; // initialize global variable with the dimension of dats int xdim0_update_halo_kernel1_t1 = args[0].dat->size[0]; int ydim0_update_halo_kernel1_t1 = args[0].dat->size[1]; int xdim1_update_halo_kernel1_t1 = args[1].dat->size[0]; int ydim1_update_halo_kernel1_t1 = args[1].dat->size[1]; int xdim2_update_halo_kernel1_t1 = args[2].dat->size[0]; int ydim2_update_halo_kernel1_t1 = args[2].dat->size[1]; int xdim3_update_halo_kernel1_t1 = args[3].dat->size[0]; int ydim3_update_halo_kernel1_t1 = args[3].dat->size[1]; int xdim4_update_halo_kernel1_t1 = args[4].dat->size[0]; int ydim4_update_halo_kernel1_t1 = args[4].dat->size[1]; int xdim5_update_halo_kernel1_t1 = args[5].dat->size[0]; int ydim5_update_halo_kernel1_t1 = args[5].dat->size[1]; int xdim6_update_halo_kernel1_t1 = args[6].dat->size[0]; int ydim6_update_halo_kernel1_t1 = args[6].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[15].mpi_time += t1 - t2; } #pragma omp parallel for collapse(2) for (int n_z = start[2]; n_z < end[2]; n_z++) { for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd aligned(density0, density1, energy0, energy1, pressure, \ viscosity, soundspeed) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { if (fields[FIELD_DENSITY0] == 1) density0[OPS_ACC0(0, 0, 0)] = density0[OPS_ACC0(0, -1, 0)]; if (fields[FIELD_DENSITY1] == 1) density1[OPS_ACC1(0, 0, 0)] = density1[OPS_ACC1(0, -1, 0)]; if (fields[FIELD_ENERGY0] == 1) energy0[OPS_ACC2(0, 0, 0)] = energy0[OPS_ACC2(0, -1, 0)]; if (fields[FIELD_ENERGY1] == 1) energy1[OPS_ACC3(0, 0, 0)] = energy1[OPS_ACC3(0, -1, 0)]; if (fields[FIELD_PRESSURE] == 1) pressure[OPS_ACC4(0, 0, 0)] = pressure[OPS_ACC4(0, -1, 0)]; if (fields[FIELD_VISCOSITY] == 1) viscosity[OPS_ACC5(0, 0, 0)] = viscosity[OPS_ACC5(0, -1, 0)]; if (fields[FIELD_SOUNDSPEED] == 1) soundspeed[OPS_ACC6(0, 0, 0)] = soundspeed[OPS_ACC6(0, -1, 0)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[15].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[15].mpi_time += t1 - t2; OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[15].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_mblock_populate_kernel(char const *name, ops_block block, int dim, int* range, ops_arg arg0, ops_arg arg1) { //Timing double t1,t2,c1,c2; ops_timers_core(&c1,&t1); int offs[2][2]; ops_arg args[2] = { arg0, arg1}; ops_timing_realloc(0,"mblock_populate_kernel"); OPS_kernels[0].count++; //compute locally 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 #ifdef OPS_DEBUG ops_register_args(args, "mblock_populate_kernel"); #endif offs[0][0] = args[0].stencil->stride[0]*1; //unit step in x dimension offs[0][1] = off2D(1, &start[0], &end[0],args[0].dat->size, args[0].stencil->stride) - offs[0][0]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int dat0 = args[0].dat->elem_size; #ifdef _OPENMP int nthreads = omp_get_max_threads( ); #else int nthreads = 1; #endif xdim0 = args[0].dat->size[0]*args[0].dat->dim; ops_H_D_exchanges_host(args, 2); //Halo Exchanges ops_halo_exchanges(args,2,range); ops_timers_core(&c2,&t2); OPS_kernels[0].mpi_time += t2-t1; #pragma omp parallel for for ( int thr=0; thr<nthreads; thr++ ){ int y_size = end[1]-start[1]; char *p_a[2]; int start_i = start[1] + ((y_size-1)/nthreads+1)*thr; int finish_i = start[1] + MIN(((y_size-1)/nthreads+1)*(thr+1),y_size); //get address per thread int start0 = start[0]; int start1 = start_i; int arg_idx[2]; #ifdef OPS_MPI arg_idx[0] = sb->decomp_disp[0]+start0; arg_idx[1] = sb->decomp_disp[1]+start1; #else //OPS_MPI arg_idx[0] = start0; arg_idx[1] = start1; #endif //OPS_MPI //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 * (start0 * args[0].stencil->stride[0] - args[0].dat->base[0] - d_m[0]); base0 = base0+ dat0 * args[0].dat->size[0] * (start1 * args[0].stencil->stride[1] - args[0].dat->base[1] - d_m[1]); p_a[0] = (char *)args[0].data + base0; p_a[1] = (char *)arg_idx; for ( int n_y=start_i; n_y<finish_i; n_y++ ){ for ( int n_x=start[0]; n_x<start[0]+(end[0]-start[0])/SIMD_VEC; n_x++ ){ //call kernel function, passing in pointers to data -vectorised for ( int i=0; i<SIMD_VEC; i++ ){ mblock_populate_kernel( (double * )p_a[0]+ i*1, arg_idx ); arg_idx[0]++; } //shift pointers to data x direction p_a[0]= p_a[0] + (dat0 * off0_0)*SIMD_VEC; } for ( int n_x=start[0]+((end[0]-start[0])/SIMD_VEC)*SIMD_VEC; n_x<end[0]; n_x++ ){ //call kernel function, passing in pointers to data - remainder mblock_populate_kernel( (double * )p_a[0], arg_idx ); //shift pointers to data x direction p_a[0]= p_a[0] + (dat0 * off0_0); arg_idx[0]++; } //shift pointers to data y direction p_a[0]= p_a[0] + (dat0 * off0_1); #ifdef OPS_MPI arg_idx[0] = sb->decomp_disp[0]+start0; #else //OPS_MPI arg_idx[0] = start0; #endif //OPS_MPI arg_idx[1]++; } } ops_timers_core(&c1,&t1); OPS_kernels[0].time += t1-t2; ops_set_dirtybit_host(args, 2); ops_set_halo_dirtybit3(&args[0],range); //Update kernel record ops_timers_core(&c2,&t2); OPS_kernels[0].mpi_time += t2-t1; OPS_kernels[0].transfer += ops_compute_transfer(dim, range, &arg0); }
// host stub function void ops_par_loop_checkop_kernel(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) { // Timing double t1, t2, c1, c2; char *p_a[6]; int offs[6][1]; ops_arg args[6] = {arg0, arg1, arg2, arg3, arg4, arg5}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 6, range, 14)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(14, "checkop_kernel"); OPS_kernels[14].count++; ops_timers_core(&c2, &t2); } // compute locally allocated range for the sub-block int start[1]; int end[1]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; #endif #ifdef OPS_DEBUG ops_register_args(args, "checkop_kernel"); #endif int arg_idx[1]; int arg_idx_base[1]; #ifdef OPS_MPI if (compute_ranges(args, 6, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 1; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif // OPS_MPI for (int n = 0; n < 1; n++) { arg_idx_base[n] = arg_idx[n]; } offs[0][0] = args[0].stencil->stride[0] * 1; // unit step in x dimension offs[1][0] = args[1].stencil->stride[0] * 1; // unit step in x dimension offs[2][0] = args[2].stencil->stride[0] * 1; // unit step in x dimension int off0_0 = offs[0][0]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].dat->elem_size); int off1_0 = offs[1][0]; int dat1 = (OPS_soa ? args[1].dat->type_size : args[1].dat->elem_size); int off2_0 = offs[2][0]; int dat2 = (OPS_soa ? args[2].dat->type_size : args[2].dat->elem_size); // set up initial pointers and exchange halos if necessary 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]; p_a[0] = (char *)args[0].data + base0; 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]; p_a[1] = (char *)args[1].data + base1; 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]; p_a[2] = (char *)args[2].data + base2; #ifdef OPS_MPI p_a[3] = ((ops_reduction)args[3].data)->data + ((ops_reduction)args[3].data)->size * block->index; #else p_a[3] = ((ops_reduction)args[3].data)->data; #endif #ifdef OPS_MPI p_a[4] = ((ops_reduction)args[4].data)->data + ((ops_reduction)args[4].data)->size * block->index; #else p_a[4] = ((ops_reduction)args[4].data)->data; #endif #ifdef OPS_MPI p_a[5] = ((ops_reduction)args[5].data)->data + ((ops_reduction)args[5].data)->size * block->index; #else p_a[5] = ((ops_reduction)args[5].data)->data; #endif // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; xdim1 = args[1].dat->size[0]; xdim2 = args[2].dat->size[0]; // Halo Exchanges ops_H_D_exchanges_host(args, 6); ops_halo_exchanges(args, 6, range); ops_H_D_exchanges_host(args, 6); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[14].mpi_time += t1 - t2; } int n_x; #pragma novector for (n_x = start[0]; n_x < start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x += SIMD_VEC) { // call kernel function, passing in pointers to data -vectorised for (int i = 0; i < SIMD_VEC; i++) { checkop_kernel((double *)p_a[0] + i * 1 * 1, (double *)p_a[1] + i * 1 * 1, (double *)p_a[2] + i * 1 * 1, (double *)p_a[3], (double *)p_a[4], (int *)p_a[5]); } // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0) * SIMD_VEC; p_a[1] = p_a[1] + (dat1 * off1_0) * SIMD_VEC; p_a[2] = p_a[2] + (dat2 * off2_0) * SIMD_VEC; } for (int n_x = start[0] + ((end[0] - start[0]) / SIMD_VEC) * SIMD_VEC; n_x < end[0]; n_x++) { // call kernel function, passing in pointers to data - remainder checkop_kernel((double *)p_a[0], (double *)p_a[1], (double *)p_a[2], (double *)p_a[3], (double *)p_a[4], (int *)p_a[5]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); p_a[1] = p_a[1] + (dat1 * off1_0); p_a[2] = p_a[2] + (dat2 * off2_0); } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[14].time += t2 - t1; } ops_set_dirtybit_host(args, 6); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[14].mpi_time += t1 - t2; OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg2); } }