void initialise_chunk_kernel_zz_c_wrapper( int *p_a0, int *p_a1, int arg_idx0, int arg_idx1, int arg_idx2, int x_size, int y_size, int z_size) { #ifdef OPS_GPU #pragma acc parallel deviceptr(p_a0) #pragma acc loop #endif for ( int n_z=0; n_z<z_size; n_z++ ){ #ifdef OPS_GPU #pragma acc loop #endif for ( int n_y=0; n_y<y_size; n_y++ ){ #ifdef OPS_GPU #pragma acc loop #endif for ( int n_x=0; n_x<x_size; n_x++ ){ int arg_idx[] = {arg_idx0+n_x, arg_idx1+n_y, arg_idx2+n_z}; initialise_chunk_kernel_zz( p_a0 + n_x*0 + n_y*xdim0_initialise_chunk_kernel_zz*0 + n_z*xdim0_initialise_chunk_kernel_zz*ydim0_initialise_chunk_kernel_zz*1, arg_idx ); } } } }
// 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_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; char *p_a[2]; int offs[2][3]; 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, "initialise_chunk_kernel_zz"); OPS_kernels[2].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, "initialise_chunk_kernel_zz"); #endif int arg_idx[3]; int arg_idx_base[3]; #ifdef OPS_MPI if (compute_ranges(args, 2, 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]; 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); // 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; p_a[1] = (char *)arg_idx; // initialize global variable with the dimension of dats xdim0 = args[0].dat->size[0]; ydim0 = args[0].dat->size[1]; // 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_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 for (int i = 0; i < SIMD_VEC; i++) { initialise_chunk_kernel_zz((int *)p_a[0] + i * 0 * 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 initialise_chunk_kernel_zz((int *)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); arg_idx[0] = arg_idx_base[0]; arg_idx[1]++; } // shift pointers to data z direction p_a[0] = p_a[0] + (dat0 * off0_2); arg_idx[0] = arg_idx_base[0]; arg_idx[1] = arg_idx_base[1]; arg_idx[2]++; } 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); } }