// 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_update_halo_kernel1_fr2(char const *name, ops_block block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3, ops_arg arg4, ops_arg arg5, ops_arg arg6, ops_arg arg7) { //Timing double t1,t2,c1,c2; ops_timers_core(&c1,&t1); int offs[8][3]; ops_arg args[8] = { arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7}; ops_timing_realloc(51,"update_halo_kernel1_fr2"); OPS_kernels[51].count++; //compute 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_kernel1_fr2"); #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]; 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; #ifdef _OPENMP int nthreads = omp_get_max_threads( ); #else int nthreads = 1; #endif 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]; ops_H_D_exchanges_host(args, 8); //Halo Exchanges ops_halo_exchanges(args,8,range); ops_timers_core(&c2,&t2); OPS_kernels[51].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[8]; 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 //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]); 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 //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 * (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; #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 * (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]); base2 = base2+ dat2 * args[2].dat->size[0] * args[2].dat->size[1] * (start2 * 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 * (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]); base3 = base3+ dat3 * args[3].dat->size[0] * args[3].dat->size[1] * (start2 * 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 * (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]); base4 = base4+ dat4 * args[4].dat->size[0] * args[4].dat->size[1] * (start2 * 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 * (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]); base5 = base5+ dat5 * args[5].dat->size[0] * args[5].dat->size[1] * (start2 * 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 * (start0 * args[6].stencil->stride[0] - args[6].dat->base[0] - d_m[0]); base6 = base6+ dat6 * args[6].dat->size[0] * (start1 * 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] * (start2 * args[6].stencil->stride[2] - args[6].dat->base[2] - d_m[2]); p_a[6] = (char *)args[6].data + base6; p_a[7] = (char *)args[7].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_kernel1_fr2( (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, (int * )p_a[7] ); } //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; } 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_kernel1_fr2( (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], (int * )p_a[7] ); //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); } //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); } //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); } } ops_timers_core(&c1,&t1); OPS_kernels[51].time += t1-t2; 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_timers_core(&c2,&t2); OPS_kernels[51].mpi_time += t2-t1; OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg1); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg2); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg3); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg4); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg5); OPS_kernels[51].transfer += ops_compute_transfer(dim, range, &arg6); }
// host stub function void ops_par_loop_update_halo_kernel1_t1(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; char *p_a[8]; int offs[8][3]; ops_arg args[8] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 8, range, 14)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(14, "update_halo_kernel1_t1"); OPS_kernels[14].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, "update_halo_kernel1_t1"); #endif int arg_idx[3]; int arg_idx_base[3]; #ifdef OPS_MPI if (compute_ranges(args, 8, block, range, start, end, arg_idx) < 0) return; #else // OPS_MPI for (int n = 0; n < 3; n++) { start[n] = range[2 * n]; end[n] = range[2 * n + 1]; arg_idx[n] = start[n]; } #endif // 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]; 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); // 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; p_a[7] = args[7].data; // 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]; // Halo Exchanges ops_H_D_exchanges_host(args, 8); ops_halo_exchanges(args, 8, range); ops_H_D_exchanges_host(args, 8); if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[14].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++) { update_halo_kernel1_t1( (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, (int *)p_a[7]); } // 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; } 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_kernel1_t1((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], (int *)p_a[7]); // 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); } // 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); } // 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); } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[14].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); 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); OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[14].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_update_halo_kernel3_plus_4_a(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, 33)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(33, "update_halo_kernel3_plus_4_a"); OPS_kernels[33].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_kernel3_plus_4_a"); #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[33].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_kernel3_plus_4_a((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_kernel3_plus_4_a((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[33].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[33].mpi_time += t1 - t2; OPS_kernels[33].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[33].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_advec_cell_kernel2_zdir(char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3) { // Timing double t1, t2, c1, c2; int offs[4][3]; ops_arg args[4] = {arg0, arg1, arg2, arg3}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 4, range, 118)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(118, "advec_cell_kernel2_zdir"); OPS_kernels[118].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, "advec_cell_kernel2_zdir"); #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]; 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); // Halo Exchanges ops_H_D_exchanges_host(args, 4); ops_halo_exchanges(args, 4, range); ops_H_D_exchanges_host(args, 4); #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]; xdim2 = args[2].dat->size[0]; ydim2 = args[2].dat->size[1]; xdim3 = args[3].dat->size[0]; ydim3 = args[3].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[118].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[4]; 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; #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]); base2 = base2 + dat2 * args[2].dat->size[0] * args[2].dat->size[1] * (start2 * 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 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]); base3 = base3 + dat3 * args[3].dat->size[0] * args[3].dat->size[1] * (start2 * args[3].stencil->stride[2] - args[3].dat->base[2] - d_m[2]); p_a[3] = (char *)args[3].data + base3; 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++) { advec_cell_kernel2_zdir((double *)p_a[0] + i * 1 * 1, (double *)p_a[1] + i * 1 * 1, (const double *)p_a[2] + i * 1 * 1, (const double *)p_a[3] + 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; } 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_kernel2_zdir((double *)p_a[0], (double *)p_a[1], (const double *)p_a[2], (const double *)p_a[3]); // 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); } // 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); } // 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); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[118].time += t1 - t2; } ops_set_dirtybit_host(args, 4); 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[118].mpi_time += t2 - t1; OPS_kernels[118].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[118].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[118].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[118].transfer += ops_compute_transfer(dim, start, end, &arg3); } }
// host stub function void ops_par_loop_initialise_chunk_kernel_z_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, 5)) return; #endif if (OPS_diags > 1) { OPS_kernels[5].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, "initialise_chunk_kernel_z"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ vertexz = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const int *__restrict__ zz = (int *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ vertexdz = (double *)(args[2].data + base2); // initialize global variable with the dimension of dats int xdim0_initialise_chunk_kernel_z = args[0].dat->size[0]; int ydim0_initialise_chunk_kernel_z = args[0].dat->size[1]; int xdim1_initialise_chunk_kernel_z = args[1].dat->size[0]; int ydim1_initialise_chunk_kernel_z = args[1].dat->size[1]; int xdim2_initialise_chunk_kernel_z = args[2].dat->size[0]; int ydim2_initialise_chunk_kernel_z = args[2].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[5].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(vertexz, zz, vertexdz) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { int z_min = field.z_min - 2; double min_z, d_z; d_z = (grid.zmax - grid.zmin) / (double)grid.z_cells; min_z = grid.zmin + d_z * field.back; vertexz[OPS_ACC0(0, 0, 0)] = min_z + d_z * (zz[OPS_ACC1(0, 0, 0)] - z_min); vertexdz[OPS_ACC2(0, 0, 0)] = (double)d_z; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[5].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[5].mpi_time += t1 - t2; OPS_kernels[5].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[5].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[5].transfer += ops_compute_transfer(dim, start, end, &arg2); } }
// host stub function void ops_par_loop_tea_leaf_init_zero2_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]; // Timing double t1, t2, c1, c2; ops_arg args[2] = {arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 2, range, 16)) return; #endif if (OPS_diags > 1) { OPS_kernels[16].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, "tea_leaf_init_zero2_kernel"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ p = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; double *__restrict__ z = (double *)(args[1].data + base1); // initialize global variable with the dimension of dats int xdim0_tea_leaf_init_zero2_kernel = args[0].dat->size[0]; int xdim1_tea_leaf_init_zero2_kernel = args[1].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[16].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(p, z) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { p[OPS_ACC0(0, 0)] = 0.0; z[OPS_ACC1(0, 0)] = 0.0; } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[16].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[16].mpi_time += t1 - t2; OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[16].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_advec_mom_kernel_y2(char const *name, ops_block block, int dim, int* range, ops_arg arg0, ops_arg arg1, ops_arg arg2, ops_arg arg3) { char *p_a[4]; int offs[4][2]; ops_arg args[4] = { arg0, arg1, arg2, arg3}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args,4,range,18)) return; #endif ops_timing_realloc(18,"advec_mom_kernel_y2"); OPS_kernels[18].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, "advec_mom_kernel_y2"); #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]; //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int dat0 = args[0].dat->elem_size; int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int dat1 = args[1].dat->elem_size; int off2_0 = offs[2][0]; int off2_1 = offs[2][1]; int dat2 = args[2].dat->elem_size; int off3_0 = offs[3][0]; int off3_1 = offs[3][1]; int dat3 = args[3].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]); 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]); 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]); 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]); p_a[3] = (char *)args[3].data + base3; ops_H_D_exchanges_host(args, 4); ops_halo_exchanges(args,4,range); ops_H_D_exchanges_host(args, 4); ops_timers_core(&c1,&t1); OPS_kernels[18].mpi_time += t1-t2; xdim0 = args[0].dat->size[0]*args[0].dat->dim; xdim1 = args[1].dat->size[0]*args[1].dat->dim; xdim2 = args[2].dat->size[0]*args[2].dat->dim; xdim3 = args[3].dat->size[0]*args[3].dat->dim; 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++ ){ advec_mom_kernel_y2( (double *)p_a[0]+ i*1, (double *)p_a[1]+ i*1, (double *)p_a[2]+ i*1, (double *)p_a[3]+ 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; } 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_mom_kernel_y2( (double *)p_a[0], (double *)p_a[1], (double *)p_a[2], (double *)p_a[3] ); //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); } //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); } ops_timers_core(&c2,&t2); OPS_kernels[18].time += t2-t1; ops_set_dirtybit_host(args, 4); ops_set_halo_dirtybit3(&args[0],range); ops_set_halo_dirtybit3(&args[1],range); //Update kernel record OPS_kernels[18].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[18].transfer += ops_compute_transfer(dim, range, &arg1); OPS_kernels[18].transfer += ops_compute_transfer(dim, range, &arg2); OPS_kernels[18].transfer += ops_compute_transfer(dim, range, &arg3); }
// host stub function void ops_par_loop_ideal_gas_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]; // Timing double t1, t2, c1, c2; ops_arg args[4] = {arg0, arg1, arg2, arg3}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 4, range, 8)) return; #endif if (OPS_diags > 1) { OPS_kernels[8].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, "ideal_gas_kernel"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; const double *__restrict__ density = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ energy = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ pressure = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; double *__restrict__ soundspeed = (double *)(args[3].data + base3); // initialize global variable with the dimension of dats int xdim0_ideal_gas_kernel = args[0].dat->size[0]; int xdim1_ideal_gas_kernel = args[1].dat->size[0]; int xdim2_ideal_gas_kernel = args[2].dat->size[0]; int xdim3_ideal_gas_kernel = args[3].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[8].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(density, energy, pressure, soundspeed) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double sound_speed_squared, v, pressurebyenergy, pressurebyvolume; v = 1.0 / density[OPS_ACC0(0, 0)]; pressure[OPS_ACC2(0, 0)] = (1.4 - 1.0) * density[OPS_ACC0(0, 0)] * energy[OPS_ACC1(0, 0)]; pressurebyenergy = (1.4 - 1.0) * density[OPS_ACC0(0, 0)]; pressurebyvolume = -1 * density[OPS_ACC0(0, 0)] * pressure[OPS_ACC2(0, 0)]; sound_speed_squared = v * v * (pressure[OPS_ACC2(0, 0)] * pressurebyenergy - pressurebyvolume); soundspeed[OPS_ACC3(0, 0)] = sqrt(sound_speed_squared); } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[8].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[8].mpi_time += t1 - t2; OPS_kernels[8].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[8].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[8].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[8].transfer += ops_compute_transfer(dim, start, end, &arg3); } }
// host stub function void ops_par_loop_advec_mom_kernel_post_pre_advec_z_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, 136)) return; #endif if (OPS_diags > 1) { OPS_kernels[136].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_z"); #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_z = args[0].dat->size[0]; int ydim0_advec_mom_kernel_post_pre_advec_z = args[0].dat->size[1]; int xdim1_advec_mom_kernel_post_pre_advec_z = args[1].dat->size[0]; int ydim1_advec_mom_kernel_post_pre_advec_z = args[1].dat->size[1]; int xdim2_advec_mom_kernel_post_pre_advec_z = args[2].dat->size[0]; int ydim2_advec_mom_kernel_post_pre_advec_z = args[2].dat->size[1]; int xdim3_advec_mom_kernel_post_pre_advec_z = args[3].dat->size[0]; int ydim3_advec_mom_kernel_post_pre_advec_z = args[3].dat->size[1]; int xdim4_advec_mom_kernel_post_pre_advec_z = args[4].dat->size[0]; int ydim4_advec_mom_kernel_post_pre_advec_z = args[4].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[136].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, 0, -1)] + node_flux[OPS_ACC4(0, 0, 0)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[136].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[136].mpi_time += t1 - t2; OPS_kernels[136].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[136].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[136].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[136].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[136].transfer += ops_compute_transfer(dim, start, end, &arg4); } }
// host stub function void ops_par_loop_advec_mom_kernel1_z_nonvector_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, 137)) return; #endif if (OPS_diags > 1) { OPS_kernels[137].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_kernel1_z_nonvector"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; const double *__restrict__ node_flux = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ node_mass_pre = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ mom_flux = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ celldz = (double *)(args[3].data + base3); int base4 = args[4].dat->base_offset; const double *__restrict__ vel1 = (double *)(args[4].data + base4); // initialize global variable with the dimension of dats int xdim0_advec_mom_kernel1_z_nonvector = args[0].dat->size[0]; int ydim0_advec_mom_kernel1_z_nonvector = args[0].dat->size[1]; int xdim1_advec_mom_kernel1_z_nonvector = args[1].dat->size[0]; int ydim1_advec_mom_kernel1_z_nonvector = args[1].dat->size[1]; int xdim2_advec_mom_kernel1_z_nonvector = args[2].dat->size[0]; int ydim2_advec_mom_kernel1_z_nonvector = args[2].dat->size[1]; int xdim3_advec_mom_kernel1_z_nonvector = args[3].dat->size[0]; int ydim3_advec_mom_kernel1_z_nonvector = args[3].dat->size[1]; int xdim4_advec_mom_kernel1_z_nonvector = args[4].dat->size[0]; int ydim4_advec_mom_kernel1_z_nonvector = args[4].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[137].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_flux, node_mass_pre, mom_flux, celldz, vel1) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double sigma, wind, width; double vdiffuw, vdiffdw, auw, adw, limiter; int upwind, donor, downwind, dif; double advec_vel_temp; if ((node_flux[OPS_ACC0(0, 0, 0)]) < 0.0) { upwind = 2; donor = 1; downwind = 0; dif = donor; } else { upwind = -1; donor = 0; downwind = 1; dif = upwind; } sigma = fabs(node_flux[OPS_ACC0(0, 0, 0)]) / node_mass_pre[OPS_ACC1(0, 0, donor)]; width = celldz[OPS_ACC3(0, 0, 0)]; vdiffuw = vel1[OPS_ACC4(0, 0, donor)] - vel1[OPS_ACC4(0, 0, upwind)]; vdiffdw = vel1[OPS_ACC4(0, 0, downwind)] - vel1[OPS_ACC4(0, 0, donor)]; limiter = 0.0; if (vdiffuw * vdiffdw > 0.0) { auw = fabs(vdiffuw); adw = fabs(vdiffdw); wind = 1.0; if (vdiffdw <= 0.0) wind = -1.0; limiter = wind * MIN(width * ((2.0 - sigma) * adw / width + (1.0 + sigma) * auw / celldz[OPS_ACC3(0, 0, dif)]) / 6.0, MIN(auw, adw)); } advec_vel_temp = vel1[OPS_ACC4(0, 0, donor)] + (1.0 - sigma) * limiter; mom_flux[OPS_ACC2(0, 0, 0)] = advec_vel_temp * node_flux[OPS_ACC0(0, 0, 0)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[137].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[137].mpi_time += t1 - t2; OPS_kernels[137].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[137].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[137].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[137].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[137].transfer += ops_compute_transfer(dim, start, end, &arg4); } }
// host stub function void ops_par_loop_initialise_chunk_kernel_cellx_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, 12)) return; #endif if (OPS_diags > 1) { OPS_kernels[12].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, "initialise_chunk_kernel_cellx"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; const double *__restrict__ vertexx = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; double *__restrict__ cellx = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ celldx = (double *)(args[2].data + base2); // initialize global variable with the dimension of dats int xdim0_initialise_chunk_kernel_cellx = args[0].dat->size[0]; int xdim1_initialise_chunk_kernel_cellx = args[1].dat->size[0]; int xdim2_initialise_chunk_kernel_cellx = args[2].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[12].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(vertexx, cellx, celldx) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double d_x; d_x = (grid.xmax - grid.xmin) / (double)grid.x_cells; cellx[OPS_ACC1(0, 0)] = 0.5 * (vertexx[OPS_ACC0(0, 0)] + vertexx[OPS_ACC0(1, 0)]); celldx[OPS_ACC2(0, 0)] = d_x; } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[12].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[12].mpi_time += t1 - t2; OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[12].transfer += ops_compute_transfer(dim, start, end, &arg2); } }
// 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_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_test_kernel(char const *name, ops_block block, int dim, int *range, ops_arg arg0, ops_arg arg1) { // Timing double t1, t2, c1, c2; char *p_a[2]; int offs[2][1]; ops_arg args[2] = {arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 2, range, 14)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(14, "test_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, "test_kernel"); #endif 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 // 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 int off0_0 = offs[0][0]; int dat0 = (OPS_soa ? args[0].dat->type_size : args[0].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; #ifdef OPS_MPI p_a[1] = ((ops_reduction)args[1].data)->data + ((ops_reduction)args[1].data)->size * block->index; #else p_a[1] = ((ops_reduction)args[1].data)->data; #endif // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; // Halo Exchanges ops_H_D_exchanges_host(args, 2); ops_halo_exchanges(args, 2, range); ops_H_D_exchanges_host(args, 2); 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++) { test_kernel((double *)p_a[0] + i * 1 * 1, (double *)p_a[1]); } // 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 test_kernel((double *)p_a[0], (double *)p_a[1]); // shift pointers to data x direction p_a[0] = p_a[0] + (dat0 * off0_0); } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[14].time += t2 - t1; } ops_set_dirtybit_host(args, 2); 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); } }
// host stub function void ops_par_loop_tea_leaf_norm2_kernel(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][2]; ops_arg args[2] = {arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 2, range, 39)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(39, "tea_leaf_norm2_kernel"); OPS_kernels[39].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, "tea_leaf_norm2_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 = (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] = ZERO_double; } } xdim0 = args[0].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[39].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; // 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; p_a[1] = (char *)arg1h; 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++) { tea_leaf_norm2_kernel((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 tea_leaf_norm2_kernel((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); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[39].time += t1 - t2; } // combine reduction data for (int thr = 0; thr < nthreads; thr++) { for (int d = 0; d < 1; d++) { 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[39].mpi_time += t2 - t1; OPS_kernels[39].transfer += ops_compute_transfer(dim, start, end, &arg0); } }
// host stub function void ops_par_loop_calupwindeff_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; int offs[7][1]; ops_arg args[7] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 7, range, 11)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(11, "calupwindeff_kernel"); OPS_kernels[11].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, "calupwindeff_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 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); // Halo Exchanges ops_H_D_exchanges_host(args, 7); ops_halo_exchanges(args, 7, range); ops_H_D_exchanges_host(args, 7); #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]; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[11].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[7]; 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; 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++) { calupwindeff_kernel((const double *)p_a[0] + i * 1 * 3, (const double *)p_a[1] + i * 1 * 3, (const double *)p_a[2] + i * 1 * 3, (const double *)p_a[3] + i * 1 * 3, (const double *)p_a[4] + i * 1 * 3, (const double *)p_a[5] + i * 1 * 9, (double *)p_a[6] + i * 1 * 3); } // 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; } 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 calupwindeff_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], (double *)p_a[6]); // 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); } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[11].time += t1 - t2; } ops_set_dirtybit_host(args, 7); ops_set_halo_dirtybit3(&args[6], range); if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c2, &t2); OPS_kernels[11].mpi_time += t2 - t1; OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[11].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_viscosity_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]; // 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, 50)) return; #endif if (OPS_diags > 1) { OPS_kernels[50].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, "viscosity_kernel"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; const double *__restrict__ xvel0 = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ yvel0 = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; const double *__restrict__ celldx = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ celldy = (double *)(args[3].data + base3); int base4 = args[4].dat->base_offset; const double *__restrict__ pressure = (double *)(args[4].data + base4); int base5 = args[5].dat->base_offset; const double *__restrict__ density0 = (double *)(args[5].data + base5); int base6 = args[6].dat->base_offset; double *__restrict__ viscosity = (double *)(args[6].data + base6); // initialize global variable with the dimension of dats int xdim0_viscosity_kernel = args[0].dat->size[0]; int xdim1_viscosity_kernel = args[1].dat->size[0]; int xdim2_viscosity_kernel = args[2].dat->size[0]; int xdim3_viscosity_kernel = args[3].dat->size[0]; int xdim4_viscosity_kernel = args[4].dat->size[0]; int xdim5_viscosity_kernel = args[5].dat->size[0]; int xdim6_viscosity_kernel = args[6].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[50].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(xvel0, yvel0, celldx, celldy, pressure, density0, \ viscosity) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double ugrad, vgrad, grad2, pgradx, pgrady, pgradx2, pgrady2, grad, ygrad, xgrad, div, strain2, limiter, pgrad; ugrad = (xvel0[OPS_ACC0(1, 0)] + xvel0[OPS_ACC0(1, 1)]) - (xvel0[OPS_ACC0(0, 0)] + xvel0[OPS_ACC0(0, 1)]); vgrad = (yvel0[OPS_ACC1(0, 1)] + yvel0[OPS_ACC1(1, 1)]) - (yvel0[OPS_ACC1(0, 0)] + yvel0[OPS_ACC1(1, 0)]); div = (celldx[OPS_ACC2(0, 0)]) * (ugrad) + (celldy[OPS_ACC3(0, 0)]) * (vgrad); strain2 = 0.5 * (xvel0[OPS_ACC0(0, 1)] + xvel0[OPS_ACC0(1, 1)] - xvel0[OPS_ACC0(0, 0)] - xvel0[OPS_ACC0(1, 0)]) / (celldy[OPS_ACC3(0, 0)]) + 0.5 * (yvel0[OPS_ACC1(1, 0)] + yvel0[OPS_ACC1(1, 1)] - yvel0[OPS_ACC1(0, 0)] - yvel0[OPS_ACC1(0, 1)]) / (celldx[OPS_ACC2(0, 0)]); pgradx = (pressure[OPS_ACC4(1, 0)] - pressure[OPS_ACC4(-1, 0)]) / (celldx[OPS_ACC2(0, 0)] + celldx[OPS_ACC2(1, 0)]); pgrady = (pressure[OPS_ACC4(0, 1)] - pressure[OPS_ACC4(0, -1)]) / (celldy[OPS_ACC3(0, 0)] + celldy[OPS_ACC3(0, 1)]); pgradx2 = pgradx * pgradx; pgrady2 = pgrady * pgrady; limiter = ((0.5 * (ugrad) / celldx[OPS_ACC2(0, 0)]) * pgradx2 + (0.5 * (vgrad) / celldy[OPS_ACC3(0, 0)]) * pgrady2 + strain2 * pgradx * pgrady) / MAX(pgradx2 + pgrady2, 1.0e-16); if ((limiter > 0.0) || (div >= 0.0)) { viscosity[OPS_ACC6(0, 0)] = 0.0; } else { pgradx = SIGN(MAX(1.0e-16, fabs(pgradx)), pgradx); pgrady = SIGN(MAX(1.0e-16, fabs(pgrady)), pgrady); pgrad = sqrt(pgradx * pgradx + pgrady * pgrady); xgrad = fabs(celldx[OPS_ACC2(0, 0)] * pgrad / pgradx); ygrad = fabs(celldy[OPS_ACC3(0, 0)] * pgrad / pgrady); grad = MIN(xgrad, ygrad); grad2 = grad * grad; viscosity[OPS_ACC6(0, 0)] = 2.0 * (density0[OPS_ACC5(0, 0)]) * grad2 * limiter * limiter; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[50].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[50].mpi_time += t1 - t2; OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[50].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_update_halo_kernel1_b2_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, 9)) return; #endif if (OPS_diags > 1) { OPS_kernels[9].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, "update_halo_kernel1_b2"); #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_b2 = args[0].dat->size[0]; int xdim1_update_halo_kernel1_b2 = args[1].dat->size[0]; int xdim2_update_halo_kernel1_b2 = args[2].dat->size[0]; int xdim3_update_halo_kernel1_b2 = args[3].dat->size[0]; int xdim4_update_halo_kernel1_b2 = args[4].dat->size[0]; int xdim5_update_halo_kernel1_b2 = args[5].dat->size[0]; int xdim6_update_halo_kernel1_b2 = args[6].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[9].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(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)] = density0[OPS_ACC0(0, 3)]; if (fields[FIELD_DENSITY1] == 1) density1[OPS_ACC1(0, 0)] = density1[OPS_ACC1(0, 3)]; if (fields[FIELD_ENERGY0] == 1) energy0[OPS_ACC2(0, 0)] = energy0[OPS_ACC2(0, 3)]; if (fields[FIELD_ENERGY1] == 1) energy1[OPS_ACC3(0, 0)] = energy1[OPS_ACC3(0, 3)]; if (fields[FIELD_PRESSURE] == 1) pressure[OPS_ACC4(0, 0)] = pressure[OPS_ACC4(0, 3)]; if (fields[FIELD_VISCOSITY] == 1) viscosity[OPS_ACC5(0, 0)] = viscosity[OPS_ACC5(0, 3)]; if (fields[FIELD_SOUNDSPEED] == 1) soundspeed[OPS_ACC6(0, 0)] = soundspeed[OPS_ACC6(0, 3)]; } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[9].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[9].mpi_time += t1 - t2; OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_left_bndcon(char const *name, ops_block block, int dim, int* range, ops_arg arg0, ops_arg arg1) { //Timing double t1,t2,c1,c2; char *p_a[2]; int offs[2][2]; ops_arg args[2] = { arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args,2,range,2)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(2,"left_bndcon"); OPS_kernels[2].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]; 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 #ifdef OPS_DEBUG ops_register_args(args, "left_bndcon"); #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 arg_idx[2]; #ifdef OPS_MPI arg_idx[0] = sb->decomp_disp[0]+start[0]; arg_idx[1] = sb->decomp_disp[1]+start[1]; #else arg_idx[0] = start[0]; arg_idx[1] = start[1]; #endif 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); //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; p_a[1] = (char *)arg_idx; //initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; //Halo Exchanges ops_H_D_exchanges_host(args, 2); ops_halo_exchanges(args,2,range); ops_H_D_exchanges_host(args, 2); if (OPS_diags > 1) { ops_timers_core(&c1,&t1); OPS_kernels[2].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 for ( int i=0; i<SIMD_VEC; i++ ){ left_bndcon( (double *)p_a[0]+ i*1*1, (int *)p_a[1] ); 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 left_bndcon( (double *)p_a[0], (int *)p_a[1] ); //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]+start[0]; #else arg_idx[0] = start[0]; #endif arg_idx[1]++; } if (OPS_diags > 1) { ops_timers_core(&c2,&t2); OPS_kernels[2].time += t2-t1; } ops_set_dirtybit_host(args, 2); ops_set_halo_dirtybit3(&args[0],range); if (OPS_diags > 1) { //Update kernel record ops_timers_core(&c1,&t1); OPS_kernels[2].mpi_time += t1-t2; OPS_kernels[2].transfer += ops_compute_transfer(dim, start, end, &arg0); } }
// host stub function void ops_par_loop_tea_leaf_cg_calc_ur_r_reduce_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]; // Timing double t1, t2, c1, c2; ops_arg args[4] = {arg0, arg1, arg2, arg3}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 4, range, 21)) return; #endif if (OPS_diags > 1) { OPS_kernels[21].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, "tea_leaf_cg_calc_ur_r_reduce_kernel"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ r = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ w = (double *)(args[1].data + base1); const double *__restrict__ alpha = (double *)args[2].data; #ifdef OPS_MPI double *__restrict__ p_a3 = (double *)(((ops_reduction)args[3].data)->data + ((ops_reduction)args[3].data)->size * block->index); #else // OPS_MPI double *__restrict__ p_a3 = (double *)((ops_reduction)args[3].data)->data; #endif // OPS_MPI // initialize global variable with the dimension of dats int xdim0_tea_leaf_cg_calc_ur_r_reduce_kernel = args[0].dat->size[0]; int xdim1_tea_leaf_cg_calc_ur_r_reduce_kernel = args[1].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[21].mpi_time += t1 - t2; } double p_a3_0 = p_a3[0]; #pragma omp parallel for reduction(+ : p_a3_0) for (int n_y = start[1]; n_y < end[1]; n_y++) { #ifdef intel #pragma loop_count(10000) #pragma omp simd reduction(+ : p_a3_0) aligned(r, w) #else #pragma simd reduction(+ : p_a3_0) #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double *rnn = &p_a3_0; r[OPS_ACC0(0, 0)] = r[OPS_ACC0(0, 0)] - (*alpha) * w[OPS_ACC1(0, 0)]; *rnn = *rnn + r[OPS_ACC0(0, 0)] * r[OPS_ACC0(0, 0)]; } } p_a3[0] = p_a3_0; if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[21].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[21].mpi_time += t1 - t2; OPS_kernels[21].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[21].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_poisson_kernel_populate(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; int offs[6][2]; ops_arg args[6] = {arg0, arg1, arg2, arg3, arg4, arg5}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args, 6, range, 0)) return; #endif if (OPS_diags > 1) { ops_timing_realloc(0, "poisson_kernel_populate"); OPS_kernels[0].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, "poisson_kernel_populate"); #endif 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 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); // Halo Exchanges ops_H_D_exchanges_host(args, 6); ops_halo_exchanges(args, 6, range); ops_H_D_exchanges_host(args, 6); #ifdef _OPENMP int nthreads = omp_get_max_threads(); #else int nthreads = 1; #endif 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[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[6]; 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 arg_idx[0] = start0; arg_idx[1] = start1; #endif // set up initial pointers int d_m[OPS_MAX_DIM]; p_a[0] = (char *)args[0].data; p_a[1] = (char *)args[1].data; p_a[2] = (char *)arg_idx; #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; 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++) { poisson_kernel_populate((int *)p_a[0], (int *)p_a[1], arg_idx, (double *)p_a[3] + i * 1 * 1, (double *)p_a[4] + i * 1 * 1, (double *)p_a[5] + i * 1 * 1); arg_idx[0]++; } // shift pointers to data x direction 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 poisson_kernel_populate((int *)p_a[0], (int *)p_a[1], arg_idx, (double *)p_a[3], (double *)p_a[4], (double *)p_a[5]); // shift pointers to data x direction 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); arg_idx[0]++; } // shift pointers to data y direction 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); #ifdef OPS_MPI arg_idx[0] = sb->decomp_disp[0] + start0; #else arg_idx[0] = start0; #endif arg_idx[1]++; } } if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[0].time += t1 - t2; } ops_set_dirtybit_host(args, 6); 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[0].mpi_time += t2 - t1; OPS_kernels[0].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[0].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[0].transfer += ops_compute_transfer(dim, start, end, &arg5); } }
// host stub function void ops_par_loop_update_halo_kernel4_minus_4_a_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, 72)) return; #endif if (OPS_diags > 1) { OPS_kernels[72].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_kernel4_minus_4_a"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ vol_flux_y = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; double *__restrict__ mass_flux_y = (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_kernel4_minus_4_a = args[0].dat->size[0]; int ydim0_update_halo_kernel4_minus_4_a = args[0].dat->size[1]; int xdim1_update_halo_kernel4_minus_4_a = args[1].dat->size[0]; int ydim1_update_halo_kernel4_minus_4_a = args[1].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[72].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_y, mass_flux_y) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { if (fields[FIELD_VOL_FLUX_Y] == 1) vol_flux_y[OPS_ACC0(0, 0, 0)] = -(vol_flux_y[OPS_ACC0(0, 4, 0)]); if (fields[FIELD_MASS_FLUX_Y] == 1) mass_flux_y[OPS_ACC1(0, 0, 0)] = -(mass_flux_y[OPS_ACC1(0, 4, 0)]); } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[72].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[72].mpi_time += t1 - t2; OPS_kernels[72].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[72].transfer += ops_compute_transfer(dim, start, end, &arg1); } }
// host stub function void ops_par_loop_initialise_chunk_kernel_volume_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]; // 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, 9)) return; #endif if (OPS_diags > 1) { OPS_kernels[9].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, "initialise_chunk_kernel_volume"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ volume = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; const double *__restrict__ celldy = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; double *__restrict__ xarea = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ celldx = (double *)(args[3].data + base3); int base4 = args[4].dat->base_offset; double *__restrict__ yarea = (double *)(args[4].data + base4); int base5 = args[5].dat->base_offset; const double *__restrict__ celldz = (double *)(args[5].data + base5); int base6 = args[6].dat->base_offset; double *__restrict__ zarea = (double *)(args[6].data + base6); // initialize global variable with the dimension of dats int xdim0_initialise_chunk_kernel_volume = args[0].dat->size[0]; int ydim0_initialise_chunk_kernel_volume = args[0].dat->size[1]; int xdim1_initialise_chunk_kernel_volume = args[1].dat->size[0]; int ydim1_initialise_chunk_kernel_volume = args[1].dat->size[1]; int xdim2_initialise_chunk_kernel_volume = args[2].dat->size[0]; int ydim2_initialise_chunk_kernel_volume = args[2].dat->size[1]; int xdim3_initialise_chunk_kernel_volume = args[3].dat->size[0]; int ydim3_initialise_chunk_kernel_volume = args[3].dat->size[1]; int xdim4_initialise_chunk_kernel_volume = args[4].dat->size[0]; int ydim4_initialise_chunk_kernel_volume = args[4].dat->size[1]; int xdim5_initialise_chunk_kernel_volume = args[5].dat->size[0]; int ydim5_initialise_chunk_kernel_volume = args[5].dat->size[1]; int xdim6_initialise_chunk_kernel_volume = args[6].dat->size[0]; int ydim6_initialise_chunk_kernel_volume = args[6].dat->size[1]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[9].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(volume, celldy, xarea, celldx, yarea, celldz, zarea) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { double d_x, d_y, d_z; d_x = (grid.xmax - grid.xmin) / (double)grid.x_cells; d_y = (grid.ymax - grid.ymin) / (double)grid.y_cells; d_z = (grid.zmax - grid.zmin) / (double)grid.z_cells; volume[OPS_ACC0(0, 0, 0)] = d_x * d_y * d_z; xarea[OPS_ACC2(0, 0, 0)] = celldy[OPS_ACC1(0, 0, 0)] * celldz[OPS_ACC5(0, 0, 0)]; yarea[OPS_ACC4(0, 0, 0)] = celldx[OPS_ACC3(0, 0, 0)] * celldz[OPS_ACC5(0, 0, 0)]; zarea[OPS_ACC6(0, 0, 0)] = celldx[OPS_ACC3(0, 0, 0)] * celldy[OPS_ACC1(0, 0, 0)]; } } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[9].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[9].mpi_time += t1 - t2; OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg3); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg4); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg5); OPS_kernels[9].transfer += ops_compute_transfer(dim, start, end, &arg6); } }
// host stub function void ops_par_loop_initialise_chunk_kernel_zz(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][3]; ops_arg args[2] = { arg0, arg1}; ops_timing_realloc(132,"initialise_chunk_kernel_zz"); OPS_kernels[132].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, "initialise_chunk_kernel_zz"); #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 = 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; ydim0 = args[0].dat->size[1]; ops_H_D_exchanges_host(args, 2); //Halo Exchanges ops_halo_exchanges(args,2,range); ops_timers_core(&c2,&t2); OPS_kernels[132].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; int arg_idx[3]; #ifdef OPS_MPI arg_idx[0] = sb->decomp_disp[0]+start0; arg_idx[1] = sb->decomp_disp[1]+start1; arg_idx[2] = sb->decomp_disp[2]+start2; #else //OPS_MPI arg_idx[0] = start0; arg_idx[1] = start1; arg_idx[2] = start2; #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]); 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 *)arg_idx; 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++ ){ initialise_chunk_kernel_zz( (int * )p_a[0]+ i*0, 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 initialise_chunk_kernel_zz( (int * )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]++; } //shift pointers to data z direction p_a[0]= p_a[0] + (dat0 * off0_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 arg_idx[2]++; } } ops_timers_core(&c1,&t1); OPS_kernels[132].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[132].mpi_time += t2-t1; OPS_kernels[132].transfer += ops_compute_transfer(dim, range, &arg0); }
// host stub function void ops_par_loop_right_bndcon_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]; //Timing double t1,t2,c1,c2; ops_arg args[2] = { arg0, arg1}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args,2,range,3)) return; #endif if (OPS_diags > 1) { OPS_kernels[3].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, "right_bndcon"); #endif int arg_idx[2]; #ifdef OPS_MPI sub_block_list sb = OPS_sub_block_list[block->index]; arg_idx[0] = sb->decomp_disp[0]; arg_idx[1] = sb->decomp_disp[1]; #else //OPS_MPI arg_idx[0] = 0; arg_idx[1] = 0; #endif //OPS_MPI //set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double * __restrict__ A = (double *)(args[0].data + base0); //initialize global variable with the dimension of dats int xdim0_right_bndcon = args[0].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1,&t1); OPS_kernels[3].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(A) #else #pragma simd #endif for ( int n_x=start[0]; n_x<end[0]; n_x++ ){ int idx[] = {arg_idx[0]+n_x, arg_idx[1]+n_y}; A[OPS_ACC0(0,0)] = sin(pi * (idx[1]+1) / (jmax+1))*exp(-pi); } } if (OPS_diags > 1) { ops_timers_core(&c2,&t2); OPS_kernels[3].time += t2-t1; } if (OPS_diags > 1) { //Update kernel record ops_timers_core(&c1,&t1); OPS_kernels[3].mpi_time += t1-t2; OPS_kernels[3].transfer += ops_compute_transfer(dim, start, end, &arg0); } }
// host stub function void ops_par_loop_update_halo_kernel4_plus_2_a(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_timers_core(&c1,&t1); int offs[3][2]; ops_arg args[3] = { arg0, arg1, arg2}; #ifdef CHECKPOINTING if (!ops_checkpointing_before(args,3,range,80)) return; #endif ops_timing_realloc(80,"update_halo_kernel4_plus_2_a"); OPS_kernels[80].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, "update_halo_kernel4_plus_2_a"); #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]; int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int dat0 = args[0].dat->elem_size; int off1_0 = offs[1][0]; int off1_1 = offs[1][1]; int dat1 = args[1].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; xdim1 = args[1].dat->size[0]*args[1].dat->dim; ops_H_D_exchanges_host(args, 3); //Halo Exchanges ops_halo_exchanges(args,3,range); ops_timers_core(&c2,&t2); OPS_kernels[80].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[3]; 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 //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; #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 * (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; p_a[2] = (char *)args[2].data; 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 #pragma simd for ( int i=0; i<SIMD_VEC; i++ ){ update_halo_kernel4_plus_2_a( (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_kernel4_plus_2_a( (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); } } ops_timers_core(&c1,&t1); OPS_kernels[80].time += t1-t2; ops_set_dirtybit_host(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[80].mpi_time += t2-t1; OPS_kernels[80].transfer += ops_compute_transfer(dim, range, &arg0); OPS_kernels[80].transfer += ops_compute_transfer(dim, range, &arg1); }
// host stub function void ops_par_loop_advec_cell_kernel2_ydir_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, 66)) return; #endif if (OPS_diags > 1) { OPS_kernels[66].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, "advec_cell_kernel2_ydir"); #endif // set up initial pointers and exchange halos if necessary int base0 = args[0].dat->base_offset; double *__restrict__ pre_vol = (double *)(args[0].data + base0); int base1 = args[1].dat->base_offset; double *__restrict__ post_vol = (double *)(args[1].data + base1); int base2 = args[2].dat->base_offset; const double *__restrict__ volume = (double *)(args[2].data + base2); int base3 = args[3].dat->base_offset; const double *__restrict__ vol_flux_y = (double *)(args[3].data + base3); // initialize global variable with the dimension of dats int xdim0_advec_cell_kernel2_ydir = args[0].dat->size[0]; int xdim1_advec_cell_kernel2_ydir = args[1].dat->size[0]; int xdim2_advec_cell_kernel2_ydir = args[2].dat->size[0]; int xdim3_advec_cell_kernel2_ydir = args[3].dat->size[0]; if (OPS_diags > 1) { ops_timers_core(&c1, &t1); OPS_kernels[66].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(pre_vol, post_vol, volume, vol_flux_y) #else #pragma simd #endif for (int n_x = start[0]; n_x < end[0]; n_x++) { pre_vol[OPS_ACC0(0, 0)] = volume[OPS_ACC2(0, 0)] + vol_flux_y[OPS_ACC3(0, 1)] - vol_flux_y[OPS_ACC3(0, 0)]; post_vol[OPS_ACC1(0, 0)] = volume[OPS_ACC2(0, 0)]; } } if (OPS_diags > 1) { ops_timers_core(&c2, &t2); OPS_kernels[66].time += t2 - t1; } if (OPS_diags > 1) { // Update kernel record ops_timers_core(&c1, &t1); OPS_kernels[66].mpi_time += t1 - t2; OPS_kernels[66].transfer += ops_compute_transfer(dim, start, end, &arg0); OPS_kernels[66].transfer += ops_compute_transfer(dim, start, end, &arg1); OPS_kernels[66].transfer += ops_compute_transfer(dim, start, end, &arg2); OPS_kernels[66].transfer += ops_compute_transfer(dim, start, end, &arg3); } }
// 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) { char *p_a[2]; 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 arg_idx[2]; #ifdef OPS_MPI arg_idx[0] = sb->decomp_disp[0]+start[0]; arg_idx[1] = sb->decomp_disp[1]+start[1]; #else //OPS_MPI arg_idx[0] = start[0]; arg_idx[1] = start[1]; #endif //OPS_MPI //Timing double t1,t2,c1,c2; ops_timers_core(&c2,&t2); int off0_0 = offs[0][0]; int off0_1 = offs[0][1]; int dat0 = args[0].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]); p_a[0] = (char *)args[0].data + base0; p_a[1] = (char *)arg_idx; ops_H_D_exchanges_host(args, 2); ops_halo_exchanges(args,2,range); ops_timers_core(&c1,&t1); OPS_kernels[0].mpi_time += t1-t2; xdim0 = args[0].dat->size[0]*args[0].dat->dim; 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 for ( int i=0; i<SIMD_VEC; i++ ){ mblock_populate_kernel( (double *)p_a[0]+ i*1, (int *)p_a[1] ); 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], (int *)p_a[1] ); //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]+start[0]; #else //OPS_MPI arg_idx[0] = start[0]; #endif //OPS_MPI arg_idx[1]++; } ops_timers_core(&c2,&t2); OPS_kernels[0].time += t2-t1; ops_set_dirtybit_host(args, 2); ops_set_halo_dirtybit3(&args[0],range); //Update kernel record OPS_kernels[0].transfer += ops_compute_transfer(dim, range, &arg0); }
// 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); } }