static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
const S *ego = (const S *) ego_;
P *pln;
const problem_rdft *p;
iodim *d;
INT rs, cs, b, n;
static const plan_adt padt = {
X(rdft_solve), X(null_awake), print, destroy
};
UNUSED(plnr);
if (ego->bufferedp) {
if (!applicable_buf(ego_, p_))
return (plan *)0;
} else {
if (!applicable(ego_, p_))
return (plan *)0;
}
p = (const problem_rdft *) p_;
if (R2HC_KINDP(p->kind[0])) {
rs = p->sz->dims[0].is; cs = p->sz->dims[0].os;
pln = MKPLAN_RDFT(P, &padt,
ego->bufferedp ? apply_buf_r2hc : apply_r2hc);
} else {
rs = p->sz->dims[0].os; cs = p->sz->dims[0].is;
pln = MKPLAN_RDFT(P, &padt,
ego->bufferedp ? apply_buf_hc2r : apply_hc2r);
}
d = p->sz->dims;
n = d[0].n;
pln->k = ego->k;
pln->n = n;
pln->rs0 = rs;
pln->rs = X(mkstride)(n, 2 * rs);
pln->csr = X(mkstride)(n, cs);
pln->csi = X(mkstride)(n, -cs);
pln->ioffset = ioffset(p->kind[0], n, cs);
b = compute_batchsize(n);
pln->brs = X(mkstride)(n, 2 * b);
pln->bcsr = X(mkstride)(n, b);
pln->bcsi = X(mkstride)(n, -b);
pln->bioffset = ioffset(p->kind[0], n, b);
X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
pln->slv = ego;
X(ops_zero)(&pln->super.super.ops);
X(ops_madd2)(pln->vl / ego->desc->genus->vl,
&ego->desc->ops,
&pln->super.super.ops);
if (ego->bufferedp)
pln->super.super.ops.other += 2 * n * pln->vl;
pln->super.super.could_prune_now_p = !ego->bufferedp;
return &(pln->super.super);
}
extern "C" void do_calc(void)
{ double g = 9.80;
double sigma = 0.95;
int icount, jcount;
if (cycle_reorder == ZORDER || cycle_reorder == HILBERT_SORT) {
do_comparison_calc = 1;
do_sync = 0;
do_gpu_sync = 1;
}
size_t ncells = mesh->ncells;
cl_mem &dev_H = state->dev_H;
cl_mem &dev_U = state->dev_U;
cl_mem &dev_V = state->dev_V;
cl_mem &dev_celltype = mesh->dev_celltype;
cl_mem &dev_i = mesh->dev_i;
cl_mem &dev_j = mesh->dev_j;
cl_mem &dev_level = mesh->dev_level;
cl_mem &dev_mpot = state->dev_mpot;
vector<int> mpot;
size_t old_ncells = ncells;
size_t new_ncells = 0;
size_t new_ncells_gpu = 0;
double H_sum = -1.0;
double deltaT = 0.0;
cl_command_queue command_queue = ezcl_get_command_queue();
// Main loop.
for (int nburst = 0; nburst < outputInterval && ncycle < niter; nburst++, ncycle++) {
// To reduce drift in solution
if (do_sync) {
ezcl_enqueue_read_buffer(command_queue, dev_H, CL_FALSE, 0, ncells*sizeof(cl_state_t), (void *)&state->H[0], NULL);
ezcl_enqueue_read_buffer(command_queue, dev_U, CL_FALSE, 0, ncells*sizeof(cl_state_t), (void *)&state->U[0], NULL);
ezcl_enqueue_read_buffer(command_queue, dev_V, CL_TRUE, 0, ncells*sizeof(cl_state_t), (void *)&state->V[0], NULL);
}
// Define basic domain decomposition parameters for GPU.
old_ncells = ncells;
// Calculate the real time step for the current discrete time step.
double deltaT_cpu = state->set_timestep(g, sigma);
double deltaT_gpu = state->gpu_set_timestep(sigma);
#ifdef XXX
// Compare time step values and pass deltaT in to the kernel.
if (do_comparison_calc)
{ if (fabs(deltaT_gpu - deltaT_cpu) > .000001)
{ printf("Error with deltaT calc --- cpu %lf gpu %lf\n",deltaT_cpu,deltaT_gpu); } }
#endif
deltaT = (do_gpu_calc) ? deltaT_gpu : deltaT_cpu;
simTime += deltaT;
if (mesh->nlft == NULL) mesh->calc_neighbors();
if (mesh->dev_nlft == NULL) mesh->gpu_calc_neighbors();
if (do_comparison_calc) {
mesh->compare_neighbors_gpu_global_to_cpu_global();
}
mesh->partition_measure();
// Currently not working -- may need to be earlier?
//if (do_cpu_calc && ! mesh->have_boundary) {
// state->add_boundary_cells(mesh);
//}
// Apply BCs is currently done as first part of gpu_finite_difference and so comparison won't work here
// Execute main kernel
state->calc_finite_difference(deltaT);
state->gpu_calc_finite_difference(deltaT);
if (do_comparison_calc) {
// Need to compare dev_H to H, etc
state->compare_state_gpu_global_to_cpu_global("finite difference",ncycle,ncells);
}
if (ezcl_get_compute_device() == COMPUTE_DEVICE_ATI) {
fflush(stdout);
exit(0);
}
// Size of arrays gets reduced to just the real cells in this call for have_boundary = 0
state->remove_boundary_cells();
mpot.resize(ncells);
new_ncells = state->calc_refine_potential(mpot, icount, jcount);
//printf("DEBUG cpu icount %d jcount %d new_ncells %d\n",icount,jcount,new_ncells);
new_ncells_gpu = state->gpu_calc_refine_potential(icount, jcount);
if (do_comparison_calc) {
if (new_ncells != new_ncells_gpu) {
printf("ERROR -- new_ncells cpu %lu not equal to new_ncells gpu %lu\n",new_ncells,new_ncells_gpu);
exit(0);
}
// Need to compare dev_mpot to mpot
if (dev_mpot) {
mesh->compare_mpot_gpu_global_to_cpu_global(&mpot[0], dev_mpot);
}
}
// Sync up cpu array with gpu version to reduce differences due to minor numerical differences
// otherwise cell count will diverge causing code problems and crashes
if (dev_mpot) {
if (do_sync) {
ezcl_enqueue_read_buffer(command_queue, dev_mpot, CL_TRUE, 0, ncells*sizeof(cl_int), &mpot[0], NULL);
}
if (do_gpu_sync) {
ezcl_enqueue_write_buffer(command_queue, dev_mpot, CL_TRUE, 0, ncells*sizeof(cl_int), &mpot[0], NULL);
}
}
if (do_comparison_calc) {
// This compares ioffset for each block in the calculation
if (dev_mpot) {
mesh->compare_ioffset_gpu_global_to_cpu_global(old_ncells, &mpot[0]);
}
}
if (do_gpu_sync) {
if (dev_mpot) {
size_t local_work_size = MIN(old_ncells, TILE_SIZE);
size_t global_work_size = ((old_ncells+local_work_size - 1) /local_work_size) * local_work_size;
//size_t block_size = (ncells + TILE_SIZE - 1) / TILE_SIZE; // For on-device global reduction kernel.
size_t block_size = global_work_size/local_work_size;
vector<int> ioffset(block_size);
int mtotal = 0;
for (int ig=0; ig<(int)(old_ncells+TILE_SIZE-1)/TILE_SIZE; ig++){
int mcount = 0;
for (int ic=ig*TILE_SIZE; ic<(ig+1)*TILE_SIZE; ic++){
if (ic >= (int)old_ncells) break;
if (mesh->celltype[ic] == REAL_CELL) {
mcount += mpot[ic] ? 4 : 1;
} else {
mcount += mpot[ic] ? 2 : 1;
}
}
ioffset[ig] = mtotal;
mtotal += mcount;
}
ezcl_enqueue_write_buffer(command_queue, mesh->dev_ioffset, CL_TRUE, 0, block_size*sizeof(cl_int), &ioffset[0], NULL);
}
}
if (do_comparison_calc) {
new_ncells = new_ncells_gpu;
}
//int add_ncells = new_ncells - old_ncells;
state->rezone_all(icount, jcount, mpot);
// Clear does not delete mpot, so have to swap with an empty vector to get
// it to delete the mpot memory. This is all to avoid valgrind from showing
// it as a reachable memory leak
//mpot.clear();
vector<int>().swap(mpot);
// Resize the mesh, inserting cells where refinement is necessary.
if (dev_mpot) state->gpu_rezone_all(icount, jcount, localStencil);
ncells = new_ncells;
mesh->ncells = new_ncells;
//ezcl_device_memory_remove(dev_ioffset);
if (do_comparison_calc) {
state->compare_state_gpu_global_to_cpu_global("rezone all",ncycle,ncells);
mesh->compare_indices_gpu_global_to_cpu_global();
}
//if (do_gpu_calc) {
// int bcount = mesh->gpu_count_BCs();
//}
mesh->proc.resize(ncells);
if (icount || jcount) {
if (cycle_reorder == ZORDER || cycle_reorder == HILBERT_SORT) {
mesh->calc_spatial_coordinates(0);
}
vector<int> index(ncells);
mesh->partition_cells(numpe, index, cycle_reorder);
//state->state_reorder(index);
if (do_gpu_sync) {
ezcl_enqueue_write_buffer(command_queue, dev_celltype, CL_FALSE, 0, ncells*sizeof(cl_int), (void *)&mesh->celltype[0], NULL);
ezcl_enqueue_write_buffer(command_queue, dev_i, CL_FALSE, 0, ncells*sizeof(cl_int), (void *)&mesh->i[0], NULL);
ezcl_enqueue_write_buffer(command_queue, dev_j, CL_FALSE, 0, ncells*sizeof(cl_int), (void *)&mesh->j[0], NULL);
ezcl_enqueue_write_buffer(command_queue, dev_level, CL_TRUE, 0, ncells*sizeof(cl_int), (void *)&mesh->level[0], NULL);
}
}
if (do_gpu_sync) {
ezcl_enqueue_write_buffer(command_queue, dev_H, CL_FALSE, 0, ncells*sizeof(cl_state_t), (void *)&state->H[0], NULL);
ezcl_enqueue_write_buffer(command_queue, dev_U, CL_FALSE, 0, ncells*sizeof(cl_state_t), (void *)&state->U[0], NULL);
ezcl_enqueue_write_buffer(command_queue, dev_V, CL_TRUE, 0, ncells*sizeof(cl_state_t), (void *)&state->V[0], NULL);
}
} // End burst loop
H_sum = state->mass_sum(enhanced_precision_sum);
if (isnan(H_sum)) {
printf("Got a NAN on cycle %d\n",ncycle);
exit(-1);
}
if (do_comparison_calc) {
double total_mass = state->gpu_mass_sum(enhanced_precision_sum);
if (fabs(total_mass - H_sum) > CONSERVATION_EPS) printf("Error: mass sum gpu %f cpu %f\n", total_mass, H_sum);/***/
}
printf("Iteration %3d timestep %lf Sim Time %lf cells %ld Mass Sum %14.12lg Mass Change %12.6lg\n",
ncycle, deltaT, simTime, ncells, H_sum, H_sum - H_sum_initial);
#ifdef HAVE_GRAPHICS
if (do_cpu_calc){
mesh->calc_spatial_coordinates(0);
}
if (do_gpu_calc){
cl_mem dev_x = ezcl_malloc(NULL, const_cast<char *>("dev_x"), &ncells, sizeof(cl_spatial_t), CL_MEM_READ_WRITE, 0);
cl_mem dev_dx = ezcl_malloc(NULL, const_cast<char *>("dev_dx"), &ncells, sizeof(cl_spatial_t), CL_MEM_READ_WRITE, 0);
cl_mem dev_y = ezcl_malloc(NULL, const_cast<char *>("dev_y"), &ncells, sizeof(cl_spatial_t), CL_MEM_READ_WRITE, 0);
cl_mem dev_dy = ezcl_malloc(NULL, const_cast<char *>("dev_dy"), &ncells, sizeof(cl_spatial_t), CL_MEM_READ_WRITE, 0);
mesh->gpu_calc_spatial_coordinates(dev_x, dev_dx, dev_y, dev_dy);
if (do_comparison_calc){
#ifdef FULL_PRECISION
mesh->compare_coordinates_gpu_global_to_cpu_global_double(dev_x, dev_dx, dev_y, dev_dy, dev_H, &state->H[0]);
#else
mesh->compare_coordinates_gpu_global_to_cpu_global_float(dev_x, dev_dx, dev_y, dev_dy, dev_H, &state->H[0]);
#endif
}
ezcl_device_memory_remove(dev_x);
ezcl_device_memory_remove(dev_dx);
ezcl_device_memory_remove(dev_y);
ezcl_device_memory_remove(dev_dy);
}
set_mysize(ncells);
set_viewmode(view_mode);
set_cell_coordinates(&mesh->x[0], &mesh->dx[0], &mesh->y[0], &mesh->dy[0]);
set_cell_data(&state->H[0]);
set_cell_proc(&mesh->proc[0]);
set_circle_radius(circle_radius);
draw_scene();
#endif
// Output final results and timing information.
if (ncycle >= niter) {
//free_display();
// Get overall program timing.
double elapsed_time = cpu_timer_stop(tstart);
state->output_timing_info(do_cpu_calc, do_gpu_calc, elapsed_time);
mesh->print_partition_measure();
mesh->print_calc_neighbor_type();
mesh->print_partition_type();
printf("CPU: rezone frequency \t %8.4f\tpercent\n", (double)mesh->get_cpu_rezone_count()/(double)ncycle*100.0 );
printf("CPU: calc neigh frequency \t %8.4f\tpercent\n", (double)mesh->get_cpu_calc_neigh_count()/(double)ncycle*100.0 );
if (mesh->get_cpu_rezone_count() > 0) {
printf("CPU: refine_smooth_iter per rezone \t %8.4f\t\n", (double)mesh->get_cpu_refine_smooth_count()/(double)mesh->get_cpu_rezone_count() );
}
printf("GPU: rezone frequency \t %8.4f\tpercent\n", (double)mesh->get_gpu_rezone_count()/(double)ncycle*100.0 );
printf("GPU: calc neigh frequency \t %8.4f\tpercent\n", (double)mesh->get_gpu_calc_neigh_count()/(double)ncycle*100.0 );
if (mesh->get_gpu_rezone_count() > 0) {
printf("GPU: refine_smooth_iter per rezone \t %8.4f\t\n", (double)mesh->get_gpu_refine_smooth_count()/(double)mesh->get_gpu_rezone_count() );
}
if (mesh->dev_nlft != NULL){
ezcl_device_memory_remove(mesh->dev_nlft);
ezcl_device_memory_remove(mesh->dev_nrht);
ezcl_device_memory_remove(mesh->dev_nbot);
ezcl_device_memory_remove(mesh->dev_ntop);
}
mesh->terminate();
state->terminate();
ezcl_terminate();
ezcl_mem_walk_all();
exit(0);
} // Complete final output.
}
