void op_par_loop_res(char const *name, op_set set,
op_arg arg0,
op_arg arg1,
op_arg arg2,
op_arg arg3 ){
float *arg3h = (float *)arg3.data;
int nargs = 4;
op_arg args[4] = {arg0,arg1,arg2,arg3};
int ninds = 2;
int inds[4] = {-1,0,1,-1};
if (OP_diags>2) {
printf(" kernel routine with indirection: res \n");
}
// get plan
#ifdef OP_PART_SIZE_0
int part_size = OP_PART_SIZE_0;
#else
int part_size = OP_part_size;
#endif
op_plan *Plan = op_plan_get(name,set,part_size,nargs,args,ninds,inds);
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set number of threads
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
// execute plan
int block_offset = 0;
for (int col=0; col < Plan->ncolors; col++) {
int nblocks = Plan->ncolblk[col];
#pragma omp parallel for
for (int blockIdx=0; blockIdx<nblocks; blockIdx++)
op_x86_res( blockIdx,
(float *)arg1.data, Plan->ind_maps[0],
(float *)arg2.data, Plan->ind_maps[1],
(float *)arg0.data,
Plan->loc_maps[1],
Plan->loc_maps[2],
(float *)arg3.data,
Plan->ind_sizes,
Plan->ind_offs,
block_offset,
Plan->blkmap,
Plan->offset,
Plan->nelems,
Plan->nthrcol,
Plan->thrcol);
block_offset += nblocks;
}
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[0].name = name;
OP_kernels[0].count += 1;
OP_kernels[0].time += wall_t2 - wall_t1;
OP_kernels[0].transfer += Plan->transfer;
OP_kernels[0].transfer2 += Plan->transfer2;
}
void op_par_loop_res(char const *name, op_set set,
op_arg arg0,
op_arg arg1,
op_arg arg2,
op_arg arg3 ){
float *arg3h = (float *)arg3.data;
int nargs = 4;
op_arg args[4] = {arg0,arg1,arg2,arg3};
int ninds = 2;
int inds[4] = {-1,0,1,-1};
if (OP_diags>2) {
printf(" kernel routine with indirection: res \n");
}
// get plan
#ifdef OP_PART_SIZE_0
int part_size = OP_PART_SIZE_0;
#else
int part_size = OP_part_size;
#endif
int set_size = op_mpi_halo_exchanges(set, nargs, args);
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers_core(&cpu_t1, &wall_t1);
if (set->size >0) {
op_plan *Plan = op_plan_get(name,set,part_size,nargs,args,ninds,inds);
// execute plan
int block_offset = 0;
for (int col=0; col < Plan->ncolors; col++) {
if (col==Plan->ncolors_core) op_mpi_wait_all(nargs, args);
int nblocks = Plan->ncolblk[col];
#pragma omp parallel for
for (int blockIdx=0; blockIdx<nblocks; blockIdx++)
op_x86_res( blockIdx,
(float *)arg1.data,
(float *)arg2.data,
Plan->ind_map,
Plan->loc_map,
(float *)arg0.data,
(float *)arg3.data,
Plan->ind_sizes,
Plan->ind_offs,
block_offset,
Plan->blkmap,
Plan->offset,
Plan->nelems,
Plan->nthrcol,
Plan->thrcol,
set_size);
block_offset += nblocks;
}
op_timing_realloc(0);
OP_kernels[0].transfer += Plan->transfer;
OP_kernels[0].transfer2 += Plan->transfer2;
}
// combine reduction data
op_mpi_set_dirtybit(nargs, args);
// update kernel record
op_timers_core(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[0].name = name;
OP_kernels[0].count += 1;
OP_kernels[0].time += wall_t2 - wall_t1;
}
