// 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_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_update_halo_kernel1_r2(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][2];
ops_arg args[7] = {arg0, arg1, arg2, arg3, arg4, arg5, arg6};
#ifdef CHECKPOINTING
if (!ops_checkpointing_before(args, 7, range, 55))
return;
#endif
if (OPS_diags > 1) {
ops_timing_realloc(55, "update_halo_kernel1_r2");
OPS_kernels[55].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, "update_halo_kernel1_r2");
#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);
// 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];
if (OPS_diags > 1) {
ops_timers_core(&c2, &t2);
OPS_kernels[55].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[7];
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 *)args[6].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_kernel1_r2(
(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,
(int *)p_a[6]);
}
// 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
update_halo_kernel1_r2((double *)p_a[0], (double *)p_a[1],
(double *)p_a[2], (double *)p_a[3],
(double *)p_a[4], (double *)p_a[5],
(int *)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);
}
// 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[55].time += t1 - t2;
}
ops_set_dirtybit_host(args, 7);
ops_set_halo_dirtybit3(&args[0], range);
ops_set_halo_dirtybit3(&args[1], range);
ops_set_halo_dirtybit3(&args[2], range);
ops_set_halo_dirtybit3(&args[3], range);
ops_set_halo_dirtybit3(&args[4], range);
ops_set_halo_dirtybit3(&args[5], range);
if (OPS_diags > 1) {
// Update kernel record
ops_timers_core(&c2, &t2);
OPS_kernels[55].mpi_time += t2 - t1;
OPS_kernels[55].transfer += ops_compute_transfer(dim, start, end, &arg0);
OPS_kernels[55].transfer += ops_compute_transfer(dim, start, end, &arg1);
OPS_kernels[55].transfer += ops_compute_transfer(dim, start, end, &arg2);
OPS_kernels[55].transfer += ops_compute_transfer(dim, start, end, &arg3);
OPS_kernels[55].transfer += ops_compute_transfer(dim, start, end, &arg4);
OPS_kernels[55].transfer += ops_compute_transfer(dim, start, end, &arg5);
}
}
// 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_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_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_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_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;
ops_timers_core(&c1,&t1);
int offs[6][2];
ops_arg args[6] = { arg0, arg1, arg2, arg3, arg4, arg5};
ops_timing_realloc(0,"poisson_kernel_populate");
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, "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 = args[3].dat->elem_size;
int off4_0 = offs[4][0];
int off4_1 = offs[4][1];
int dat4 = args[4].dat->elem_size;
int off5_0 = offs[5][0];
int off5_1 = offs[5][1];
int dat5 = args[5].dat->elem_size;
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
xdim3 = args[3].dat->size[0]*args[3].dat->dim;
xdim4 = args[4].dat->size[0]*args[4].dat->dim;
xdim5 = args[5].dat->size[0]*args[5].dat->dim;
ops_H_D_exchanges_host(args, 6);
//Halo Exchanges
ops_halo_exchanges(args,6,range);
ops_timers_core(&c2,&t2);
OPS_kernels[0].mpi_time += t2-t1;
#pragma omp parallel for
for ( int thr=0; thr<nthreads; thr++ ){
int y_size = end[1]-start[1];
char *p_a[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 //OPS_MPI
arg_idx[0] = start0;
arg_idx[1] = start1;
#endif //OPS_MPI
//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 //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]);
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]);
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]);
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, (double * )p_a[4]+ i*1, (double * )p_a[5]+ i*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 //OPS_MPI
arg_idx[0] = start0;
#endif //OPS_MPI
arg_idx[1]++;
}
}
ops_timers_core(&c1,&t1);
OPS_kernels[0].time += t1-t2;
ops_set_dirtybit_host(args, 6);
ops_set_halo_dirtybit3(&args[3],range);
ops_set_halo_dirtybit3(&args[4],range);
ops_set_halo_dirtybit3(&args[5],range);
//Update kernel record
ops_timers_core(&c2,&t2);
OPS_kernels[0].mpi_time += t2-t1;
OPS_kernels[0].transfer += ops_compute_transfer(dim, range, &arg3);
OPS_kernels[0].transfer += ops_compute_transfer(dim, range, &arg4);
OPS_kernels[0].transfer += ops_compute_transfer(dim, range, &arg5);
}
// 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);
}
}
