static void free_device_data(oskar_Simulator* h, int* status)
{
int i;
if (!h->d) return;
for (i = 0; i < h->num_devices; ++i)
{
DeviceData* d = &(h->d[i]);
if (!d) continue;
if (i < h->num_gpus)
oskar_device_set(h->gpu_ids[i], status);
oskar_timer_free(d->tmr_compute);
oskar_timer_free(d->tmr_copy);
oskar_timer_free(d->tmr_clip);
oskar_timer_free(d->tmr_E);
oskar_timer_free(d->tmr_K);
oskar_timer_free(d->tmr_join);
oskar_timer_free(d->tmr_correlate);
oskar_vis_block_free(d->vis_block_cpu[0], status);
oskar_vis_block_free(d->vis_block_cpu[1], status);
oskar_vis_block_free(d->vis_block, status);
oskar_mem_free(d->u, status);
oskar_mem_free(d->v, status);
oskar_mem_free(d->w, status);
oskar_sky_free(d->chunk, status);
oskar_sky_free(d->chunk_clip, status);
oskar_telescope_free(d->tel, status);
oskar_station_work_free(d->station_work, status);
oskar_jones_free(d->J, status);
oskar_jones_free(d->E, status);
oskar_jones_free(d->K, status);
oskar_jones_free(d->R, status);
memset(d, 0, sizeof(DeviceData));
}
}
void oskar_interferometer_free(oskar_Interferometer* h, int* status)
{
int i;
if (!h) return;
oskar_interferometer_reset_cache(h, status);
for (i = 0; i < h->num_gpus; ++i)
{
oskar_device_set(h->gpu_ids[i], status);
oskar_device_reset();
}
for (i = 0; i < h->num_sky_chunks; ++i)
oskar_sky_free(h->sky_chunks[i], status);
oskar_telescope_free(h->tel, status);
oskar_mem_free(h->temp, status);
oskar_timer_free(h->tmr_sim);
oskar_timer_free(h->tmr_write);
oskar_mutex_free(h->mutex);
oskar_barrier_free(h->barrier);
free(h->sky_chunks);
free(h->gpu_ids);
free(h->vis_name);
free(h->ms_name);
free(h->settings_path);
free(h->d);
free(h);
}
void oskar_simulator_set_gpus(oskar_Simulator* h, int num,
int* ids, int* status)
{
int i, num_gpus_avail;
if (*status) return;
free_device_data(h, status);
num_gpus_avail = oskar_device_count(status);
if (*status) return;
if (num < 0)
{
h->num_gpus = num_gpus_avail;
h->gpu_ids = (int*) realloc(h->gpu_ids, h->num_gpus * sizeof(int));
for (i = 0; i < h->num_gpus; ++i)
h->gpu_ids[i] = i;
}
else if (num > 0)
{
if (num > num_gpus_avail)
{
oskar_log_error(h->log, "More GPUs were requested than found.");
*status = OSKAR_ERR_COMPUTE_DEVICES;
return;
}
h->num_gpus = num;
h->gpu_ids = (int*) realloc(h->gpu_ids, h->num_gpus * sizeof(int));
for (i = 0; i < h->num_gpus; ++i)
h->gpu_ids[i] = ids[i];
}
else /* num == 0 */
{
free(h->gpu_ids);
h->gpu_ids = 0;
h->num_gpus = 0;
}
for (i = 0; i < h->num_gpus; ++i)
{
oskar_device_set(h->gpu_ids[i], status);
if (*status) return;
}
}
void oskar_imager_set_gpus(oskar_Imager* h, int num, const int* ids,
int* status)
{
int i, num_gpus_avail;
if (*status) return;
oskar_imager_free_device_data(h, status);
num_gpus_avail = oskar_device_count(status);
if (*status) return;
if (num < 0)
{
h->num_gpus = num_gpus_avail;
h->gpu_ids = (int*) realloc(h->gpu_ids, h->num_gpus * sizeof(int));
for (i = 0; i < h->num_gpus; ++i)
h->gpu_ids[i] = i;
}
else if (num > 0)
{
if (num > num_gpus_avail)
{
*status = OSKAR_ERR_COMPUTE_DEVICES;
return;
}
h->num_gpus = num;
h->gpu_ids = (int*) realloc(h->gpu_ids, h->num_gpus * sizeof(int));
for (i = 0; i < h->num_gpus; ++i)
h->gpu_ids[i] = ids[i];
}
else /* num == 0 */
{
free(h->gpu_ids);
h->gpu_ids = 0;
h->num_gpus = 0;
}
for (i = 0; i < h->num_gpus; ++i)
{
oskar_device_set(h->gpu_ids[i], status);
if (*status) return;
}
}
static void set_up_device_data(oskar_BeamPattern* h, int* status)
{
int i, beam_type, max_src, max_size, auto_power, cross_power, raw_data;
if (*status) return;
/* Get local variables. */
max_src = h->max_chunk_size;
max_size = h->num_active_stations * max_src;
beam_type = h->prec | OSKAR_COMPLEX;
if (h->pol_mode == OSKAR_POL_MODE_FULL)
beam_type |= OSKAR_MATRIX;
raw_data = h->ixr_txt || h->ixr_fits ||
h->voltage_raw_txt || h->voltage_amp_txt || h->voltage_phase_txt ||
h->voltage_amp_fits || h->voltage_phase_fits;
auto_power = h->auto_power_fits || h->auto_power_txt;
cross_power = h->cross_power_raw_txt ||
h->cross_power_amp_fits || h->cross_power_phase_fits ||
h->cross_power_amp_txt || h->cross_power_phase_txt;
/* Expand the number of devices to the number of selected GPUs,
* if required. */
if (h->num_devices < h->num_gpus)
oskar_beam_pattern_set_num_devices(h, h->num_gpus);
for (i = 0; i < h->num_devices; ++i)
{
int dev_loc, i_stokes;
DeviceData* d = &h->d[i];
if (*status) break;
/* Select the device. */
if (i < h->num_gpus)
{
oskar_device_set(h->gpu_ids[i], status);
dev_loc = OSKAR_GPU;
}
else
{
dev_loc = OSKAR_CPU;
}
/* Device memory. */
d->previous_chunk_index = -1;
if (!d->tel)
{
d->jones_data = oskar_mem_create(beam_type, dev_loc, max_size,
status);
d->x = oskar_mem_create(h->prec, dev_loc, 1 + max_src, status);
d->y = oskar_mem_create(h->prec, dev_loc, 1 + max_src, status);
d->z = oskar_mem_create(h->prec, dev_loc, 1 + max_src, status);
d->tel = oskar_telescope_create_copy(h->tel, dev_loc, status);
d->work = oskar_station_work_create(h->prec, dev_loc, status);
}
/* Host memory. */
if (!d->jones_data_cpu[0] && raw_data)
{
d->jones_data_cpu[0] = oskar_mem_create(beam_type, OSKAR_CPU,
max_size, status);
d->jones_data_cpu[1] = oskar_mem_create(beam_type, OSKAR_CPU,
max_size, status);
}
/* Auto-correlation beam output arrays. */
for (i_stokes = 0; i_stokes < 4; ++i_stokes)
{
if (!h->stokes[i_stokes]) continue;
if (!d->auto_power[i_stokes] && auto_power)
{
/* Device memory. */
d->auto_power[i_stokes] = oskar_mem_create(beam_type, dev_loc,
max_size, status);
/* Host memory. */
d->auto_power_cpu[i_stokes][0] = oskar_mem_create(
beam_type, OSKAR_CPU, max_size, status);
d->auto_power_cpu[i_stokes][1] = oskar_mem_create(
beam_type, OSKAR_CPU, max_size, status);
if (h->average_single_axis == 'T')
d->auto_power_time_avg[i_stokes] = oskar_mem_create(
beam_type, OSKAR_CPU, max_size, status);
if (h->average_single_axis == 'C')
d->auto_power_channel_avg[i_stokes] = oskar_mem_create(
beam_type, OSKAR_CPU, max_size, status);
if (h->average_time_and_channel)
d->auto_power_channel_and_time_avg[i_stokes] =
oskar_mem_create(beam_type, OSKAR_CPU,
max_size, status);
}
/* Cross-correlation beam output arrays. */
if (!d->cross_power[i_stokes] && cross_power)
{
if (h->num_active_stations < 2)
{
oskar_log_error(h->log, "Cannot create cross-power beam "
"using less than two active stations.");
*status = OSKAR_ERR_INVALID_ARGUMENT;
break;
}
/* Device memory. */
d->cross_power[i_stokes] = oskar_mem_create(
beam_type, dev_loc, max_src, status);
/* Host memory. */
d->cross_power_cpu[i_stokes][0] = oskar_mem_create(
beam_type, OSKAR_CPU, max_src, status);
d->cross_power_cpu[i_stokes][1] = oskar_mem_create(
beam_type, OSKAR_CPU, max_src, status);
if (h->average_single_axis == 'T')
d->cross_power_time_avg[i_stokes] = oskar_mem_create(
beam_type, OSKAR_CPU, max_src, status);
if (h->average_single_axis == 'C')
d->cross_power_channel_avg[i_stokes] = oskar_mem_create(
beam_type, OSKAR_CPU, max_src, status);
if (h->average_time_and_channel)
d->cross_power_channel_and_time_avg[i_stokes] =
oskar_mem_create(beam_type, OSKAR_CPU,
max_src, status);
}
if (d->auto_power[i_stokes])
oskar_mem_clear_contents(d->auto_power[i_stokes], status);
if (d->cross_power[i_stokes])
oskar_mem_clear_contents(d->cross_power[i_stokes], status);
}
/* Timers. */
if (!d->tmr_compute)
d->tmr_compute = oskar_timer_create(OSKAR_TIMER_NATIVE);
}
}
void oskar_simulator_run_block(oskar_Simulator* h, int block_index,
int device_id, int* status)
{
double obs_start_mjd, dt_dump_days;
int i_active, time_index_start, time_index_end;
int num_channels, num_times_block, total_chunks, total_times;
DeviceData* d;
if (*status) return;
/* Check that initialisation has happened. We can't initialise here,
* as we're already multi-threaded at this point. */
if (!h->header)
{
*status = OSKAR_ERR_MEMORY_NOT_ALLOCATED;
oskar_log_error(h->log, "Simulator not initalised. "
"Call oskar_simulator_check_init() first.");
return;
}
#ifdef _OPENMP
if (!h->coords_only)
{
/* Disable any nested parallelism. */
omp_set_num_threads(1);
omp_set_nested(0);
}
#endif
/* Set the GPU to use. (Supposed to be a very low-overhead call.) */
if (device_id >= 0 && device_id < h->num_gpus)
oskar_device_set(h->gpu_ids[device_id], status);
/* Clear the visibility block. */
i_active = block_index % 2; /* Index of the active buffer. */
d = &(h->d[device_id]);
oskar_timer_resume(d->tmr_compute);
oskar_vis_block_clear(d->vis_block, status);
/* Set the visibility block meta-data. */
total_chunks = h->num_sky_chunks;
num_channels = h->num_channels;
total_times = h->num_time_steps;
obs_start_mjd = h->time_start_mjd_utc;
dt_dump_days = h->time_inc_sec / 86400.0;
time_index_start = block_index * h->max_times_per_block;
time_index_end = time_index_start + h->max_times_per_block - 1;
if (time_index_end >= total_times)
time_index_end = total_times - 1;
num_times_block = 1 + time_index_end - time_index_start;
/* Set the number of active times in the block. */
oskar_vis_block_set_num_times(d->vis_block, num_times_block, status);
oskar_vis_block_set_start_time_index(d->vis_block, time_index_start);
/* Go though all possible work units in the block. A work unit is defined
* as the simulation for one time and one sky chunk. */
while (!h->coords_only)
{
oskar_Sky* sky;
int i_work_unit, i_chunk, i_time, i_channel, sim_time_idx;
oskar_mutex_lock(h->mutex);
i_work_unit = (h->work_unit_index)++;
oskar_mutex_unlock(h->mutex);
if ((i_work_unit >= num_times_block * total_chunks) || *status) break;
/* Convert slice index to chunk/time index. */
i_chunk = i_work_unit / num_times_block;
i_time = i_work_unit - i_chunk * num_times_block;
sim_time_idx = time_index_start + i_time;
/* Copy sky chunk to device only if different from the previous one. */
if (i_chunk != d->previous_chunk_index)
{
oskar_timer_resume(d->tmr_copy);
oskar_sky_copy(d->chunk, h->sky_chunks[i_chunk], status);
oskar_timer_pause(d->tmr_copy);
}
sky = h->apply_horizon_clip ? d->chunk_clip : d->chunk;
/* Apply horizon clip if required. */
if (h->apply_horizon_clip)
{
double gast, mjd;
mjd = obs_start_mjd + dt_dump_days * (sim_time_idx + 0.5);
gast = oskar_convert_mjd_to_gast_fast(mjd);
oskar_timer_resume(d->tmr_clip);
oskar_sky_horizon_clip(d->chunk_clip, d->chunk, d->tel, gast,
d->station_work, status);
oskar_timer_pause(d->tmr_clip);
}
/* Simulate all baselines for all channels for this time and chunk. */
for (i_channel = 0; i_channel < num_channels; ++i_channel)
{
if (*status) break;
if (h->log)
{
oskar_mutex_lock(h->mutex);
oskar_log_message(h->log, 'S', 1, "Time %*i/%i, "
"Chunk %*i/%i, Channel %*i/%i [Device %i, %i sources]",
disp_width(total_times), sim_time_idx + 1, total_times,
disp_width(total_chunks), i_chunk + 1, total_chunks,
disp_width(num_channels), i_channel + 1, num_channels,
device_id, oskar_sky_num_sources(sky));
oskar_mutex_unlock(h->mutex);
}
sim_baselines(h, d, sky, i_channel, i_time, sim_time_idx, status);
}
d->previous_chunk_index = i_chunk;
}
/* Copy the visibility block to host memory. */
oskar_timer_resume(d->tmr_copy);
oskar_vis_block_copy(d->vis_block_cpu[i_active], d->vis_block, status);
oskar_timer_pause(d->tmr_copy);
oskar_timer_pause(d->tmr_compute);
}
static void set_up_device_data(oskar_Simulator* h, int* status)
{
int i, dev_loc, complx, vistype, num_stations, num_src;
if (*status) return;
/* Get local variables. */
num_stations = oskar_telescope_num_stations(h->tel);
num_src = h->max_sources_per_chunk;
complx = (h->prec) | OSKAR_COMPLEX;
vistype = complx;
if (oskar_telescope_pol_mode(h->tel) == OSKAR_POL_MODE_FULL)
vistype |= OSKAR_MATRIX;
/* Expand the number of devices to the number of selected GPUs,
* if required. */
if (h->num_devices < h->num_gpus)
oskar_simulator_set_num_devices(h, h->num_gpus);
for (i = 0; i < h->num_devices; ++i)
{
DeviceData* d = &h->d[i];
d->previous_chunk_index = -1;
/* Select the device. */
if (i < h->num_gpus)
{
oskar_device_set(h->gpu_ids[i], status);
dev_loc = OSKAR_GPU;
}
else
{
dev_loc = OSKAR_CPU;
}
/* Timers. */
if (!d->tmr_compute)
{
d->tmr_compute = oskar_timer_create(OSKAR_TIMER_NATIVE);
d->tmr_copy = oskar_timer_create(OSKAR_TIMER_NATIVE);
d->tmr_clip = oskar_timer_create(OSKAR_TIMER_NATIVE);
d->tmr_E = oskar_timer_create(OSKAR_TIMER_NATIVE);
d->tmr_K = oskar_timer_create(OSKAR_TIMER_NATIVE);
d->tmr_join = oskar_timer_create(OSKAR_TIMER_NATIVE);
d->tmr_correlate = oskar_timer_create(OSKAR_TIMER_NATIVE);
}
/* Visibility blocks. */
if (!d->vis_block)
{
d->vis_block = oskar_vis_block_create_from_header(dev_loc,
h->header, status);
d->vis_block_cpu[0] = oskar_vis_block_create_from_header(OSKAR_CPU,
h->header, status);
d->vis_block_cpu[1] = oskar_vis_block_create_from_header(OSKAR_CPU,
h->header, status);
}
oskar_vis_block_clear(d->vis_block, status);
oskar_vis_block_clear(d->vis_block_cpu[0], status);
oskar_vis_block_clear(d->vis_block_cpu[1], status);
/* Device scratch memory. */
if (!d->tel)
{
d->u = oskar_mem_create(h->prec, dev_loc, num_stations, status);
d->v = oskar_mem_create(h->prec, dev_loc, num_stations, status);
d->w = oskar_mem_create(h->prec, dev_loc, num_stations, status);
d->chunk = oskar_sky_create(h->prec, dev_loc, num_src, status);
d->chunk_clip = oskar_sky_create(h->prec, dev_loc, num_src, status);
d->tel = oskar_telescope_create_copy(h->tel, dev_loc, status);
d->J = oskar_jones_create(vistype, dev_loc, num_stations, num_src,
status);
d->R = oskar_type_is_matrix(vistype) ? oskar_jones_create(vistype,
dev_loc, num_stations, num_src, status) : 0;
d->E = oskar_jones_create(vistype, dev_loc, num_stations, num_src,
status);
d->K = oskar_jones_create(complx, dev_loc, num_stations, num_src,
status);
d->Z = 0;
d->station_work = oskar_station_work_create(h->prec, dev_loc,
status);
}
}
}
void oskar_imager_finalise_plane(oskar_Imager* h, oskar_Mem* plane,
double plane_norm, int* status)
{
int size, num_cells;
DeviceData* d;
if (*status) return;
/* Apply normalisation. */
if (plane_norm > 0.0 || plane_norm < 0.0)
oskar_mem_scale_real(plane, 1.0 / plane_norm, status);
if (h->algorithm == OSKAR_ALGORITHM_DFT_2D ||
h->algorithm == OSKAR_ALGORITHM_DFT_3D)
return;
/* Check plane is complex type, as plane must be gridded visibilities. */
if (!oskar_mem_is_complex(plane))
{
*status = OSKAR_ERR_TYPE_MISMATCH;
return;
}
/* Make image using FFT and apply grid correction. */
size = h->grid_size;
num_cells = size * size;
d = &h->d[0];
if (oskar_mem_precision(plane) == OSKAR_DOUBLE)
{
oskar_fftphase_cd(size, size, oskar_mem_double(plane, status));
if (h->fft_on_gpu)
{
#ifdef OSKAR_HAVE_CUDA
oskar_device_set(h->cuda_device_ids[0], status);
oskar_mem_copy(d->plane_gpu, plane, status);
cufftExecZ2Z(h->cufft_plan, oskar_mem_void(d->plane_gpu),
oskar_mem_void(d->plane_gpu), CUFFT_FORWARD);
oskar_mem_copy(plane, d->plane_gpu, status);
#else
*status = OSKAR_ERR_CUDA_NOT_AVAILABLE;
#endif
}
else
{
oskar_fftpack_cfft2f(size, size, size,
oskar_mem_double(plane, status),
oskar_mem_double(h->fftpack_wsave, status),
oskar_mem_double(h->fftpack_work, status));
oskar_mem_scale_real(plane, (double)num_cells, status);
}
oskar_fftphase_cd(size, size, oskar_mem_double(plane, status));
oskar_grid_correction_d(size, oskar_mem_double(h->corr_func, status),
oskar_mem_double(plane, status));
}
else
{
oskar_fftphase_cf(size, size, oskar_mem_float(plane, status));
if (h->fft_on_gpu)
{
#ifdef OSKAR_HAVE_CUDA
oskar_device_set(h->cuda_device_ids[0], status);
oskar_mem_copy(d->plane_gpu, plane, status);
cufftExecC2C(h->cufft_plan, oskar_mem_void(d->plane_gpu),
oskar_mem_void(d->plane_gpu), CUFFT_FORWARD);
oskar_mem_copy(plane, d->plane_gpu, status);
#else
*status = OSKAR_ERR_CUDA_NOT_AVAILABLE;
#endif
}
else
{
oskar_fftpack_cfft2f_f(size, size, size,
oskar_mem_float(plane, status),
oskar_mem_float(h->fftpack_wsave, status),
oskar_mem_float(h->fftpack_work, status));
oskar_mem_scale_real(plane, (double)num_cells, status);
}
oskar_fftphase_cf(size, size, oskar_mem_float(plane, status));
oskar_grid_correction_f(size, oskar_mem_double(h->corr_func, status),
oskar_mem_float(plane, status));
}
}
static void* run_blocks(void* arg)
{
oskar_Imager* h;
oskar_Mem *plane, *uu, *vv, *ww = 0, *amp, *weight, *block, *l, *m, *n;
size_t max_size;
const size_t smallest = 1024, largest = 65536;
int dev_loc = OSKAR_CPU, *status;
/* Get thread function arguments. */
h = ((ThreadArgs*)arg)->h;
const int thread_id = ((ThreadArgs*)arg)->thread_id;
const int num_vis = ((ThreadArgs*)arg)->num_vis;
plane = ((ThreadArgs*)arg)->plane;
status = &(h->status);
/* Set the device used by the thread. */
if (thread_id < h->num_gpus)
{
dev_loc = h->dev_loc;
oskar_device_set(h->dev_loc, h->gpu_ids[thread_id], status);
}
/* Copy visibility data to device. */
uu = oskar_mem_create_copy(((ThreadArgs*)arg)->uu, dev_loc, status);
vv = oskar_mem_create_copy(((ThreadArgs*)arg)->vv, dev_loc, status);
amp = oskar_mem_create_copy(((ThreadArgs*)arg)->amp, dev_loc, status);
weight = oskar_mem_create_copy(((ThreadArgs*)arg)->weight, dev_loc, status);
if (h->algorithm == OSKAR_ALGORITHM_DFT_3D)
ww = oskar_mem_create_copy(((ThreadArgs*)arg)->ww, dev_loc, status);
#ifdef _OPENMP
/* Disable nested parallelism. */
omp_set_nested(0);
omp_set_num_threads(1);
#endif
/* Calculate the maximum pixel block size, and number of blocks. */
const size_t num_pixels = (size_t)h->image_size * (size_t)h->image_size;
max_size = num_pixels / h->num_devices;
max_size = ((max_size + smallest - 1) / smallest) * smallest;
if (max_size > largest) max_size = largest;
if (max_size < smallest) max_size = smallest;
const int num_blocks = (int) ((num_pixels + max_size - 1) / max_size);
/* Allocate device memory for pixel block data. */
block = oskar_mem_create(h->imager_prec, dev_loc, 0, status);
l = oskar_mem_create(h->imager_prec, dev_loc, max_size, status);
m = oskar_mem_create(h->imager_prec, dev_loc, max_size, status);
n = oskar_mem_create(h->imager_prec, dev_loc, max_size, status);
/* Loop until all blocks are done. */
for (;;)
{
size_t block_size;
/* Get a unique block index. */
oskar_mutex_lock(h->mutex);
const int i_block = (h->i_block)++;
oskar_mutex_unlock(h->mutex);
if ((i_block >= num_blocks) || *status) break;
/* Calculate the block size. */
const size_t block_start = i_block * max_size;
block_size = num_pixels - block_start;
if (block_size > max_size) block_size = max_size;
/* Copy the (l,m,n) positions for the block. */
oskar_mem_copy_contents(l, h->l, 0, block_start, block_size, status);
oskar_mem_copy_contents(m, h->m, 0, block_start, block_size, status);
if (h->algorithm == OSKAR_ALGORITHM_DFT_3D)
oskar_mem_copy_contents(n, h->n, 0, block_start,
block_size, status);
/* Run DFT for the block. */
oskar_dft_c2r(num_vis, 2.0 * M_PI, uu, vv, ww, amp, weight,
(int) block_size, l, m, n, block, status);
/* Add data to existing pixels. */
oskar_mem_add(plane, plane, block,
block_start, block_start, 0, block_size, status);
}
/* Free memory. */
oskar_mem_free(uu, status);
oskar_mem_free(vv, status);
oskar_mem_free(ww, status);
oskar_mem_free(amp, status);
oskar_mem_free(weight, status);
oskar_mem_free(block, status);
oskar_mem_free(l, status);
oskar_mem_free(m, status);
oskar_mem_free(n, status);
return 0;
}