void advec_mom_kernel_y2_c_wrapper(double *p_a0, double *p_a1, double *p_a2, double *p_a3, int x_size, int y_size) { #ifdef OPS_GPU #pragma acc parallel deviceptr(p_a0, p_a1, p_a2, p_a3) #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++) { advec_mom_kernel_y2( p_a0 + n_x * 1 * 1 + n_y * xdim0_advec_mom_kernel_y2 * 1 * 1, p_a1 + n_x * 1 * 1 + n_y * xdim1_advec_mom_kernel_y2 * 1 * 1, p_a2 + n_x * 1 * 1 + n_y * xdim2_advec_mom_kernel_y2 * 1 * 1, p_a3 + n_x * 1 * 1 + n_y * xdim3_advec_mom_kernel_y2 * 1 * 1); } } }
// 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); }