void oskar_simulator_run(oskar_Simulator* h, int* status)
{
int i, num_threads = 1, num_vis_blocks;
if (*status) return;
/* Check the visibilities are going somewhere. */
if (!h->vis_name
#ifndef OSKAR_NO_MS
&& !h->ms_name
#endif
)
{
oskar_log_error(h->log, "No output file specified.");
#ifdef OSKAR_NO_MS
if (h->ms_name)
oskar_log_error(h->log,
"OSKAR was compiled without Measurement Set support.");
#endif
*status = OSKAR_ERR_FILE_IO;
return;
}
/* Initialise if required. */
oskar_simulator_check_init(h, status);
/* Get the number of visibility blocks to be processed. */
num_vis_blocks = oskar_simulator_num_vis_blocks(h);
/* Record memory usage. */
if (h->log && !*status)
{
oskar_log_section(h->log, 'M', "Initial memory usage");
#ifdef OSKAR_HAVE_CUDA
for (i = 0; i < h->num_gpus; ++i)
oskar_cuda_mem_log(h->log, 0, h->gpu_ids[i]);
#endif
system_mem_log(h->log);
oskar_log_section(h->log, 'M', "Starting simulation...");
}
/* Start simulation timer. */
oskar_timer_start(h->tmr_sim);
/*-----------------------------------------------------------------------
*-- START OF MULTITHREADED SIMULATION CODE -----------------------------
*-----------------------------------------------------------------------*/
/* Loop over blocks of observation time, running simulation and file
* writing one block at a time. Simulation and file output are overlapped
* by using double buffering, and a dedicated thread is used for file
* output.
*
* Thread 0 is used for file writes.
* Threads 1 to n (mapped to compute devices) do the simulation.
*
* Note that no write is launched on the first loop counter (as no
* data are ready yet) and no simulation is performed for the last loop
* counter (which corresponds to the last block + 1) as this iteration
* simply writes the last block.
*/
#ifdef _OPENMP
num_threads = h->num_devices + 1;
omp_set_num_threads(num_threads);
omp_set_nested(0);
#else
oskar_log_warning(h->log, "OpenMP not found: Using one compute device.");
#endif
oskar_simulator_reset_work_unit_index(h);
#pragma omp parallel
{
int b, thread_id = 0, device_id = 0;
/* Get host thread ID and device ID. */
#ifdef _OPENMP
thread_id = omp_get_thread_num();
device_id = thread_id - 1;
#endif
/* Loop over simulation time blocks (+1, for the last write). */
for (b = 0; b < num_vis_blocks + 1; ++b)
{
if ((thread_id > 0 || num_threads == 1) && b < num_vis_blocks)
oskar_simulator_run_block(h, b, device_id, status);
if (thread_id == 0 && b > 0)
{
oskar_VisBlock* block;
block = oskar_simulator_finalise_block(h, b - 1, status);
oskar_simulator_write_block(h, block, b - 1, status);
}
/* Barrier 1: Reset work unit index. */
#pragma omp barrier
if (thread_id == 0)
oskar_simulator_reset_work_unit_index(h);
/* Barrier 2: Synchronise before moving to the next block. */
#pragma omp barrier
if (thread_id == 0 && b < num_vis_blocks && h->log && !*status)
oskar_log_message(h->log, 'S', 0, "Block %*i/%i (%3.0f%%) "
"complete. Simulation time elapsed: %.3f s",
disp_width(num_vis_blocks), b+1, num_vis_blocks,
100.0 * (b+1) / (double)num_vis_blocks,
oskar_timer_elapsed(h->tmr_sim));
}
}
/*-----------------------------------------------------------------------
*-- END OF MULTITHREADED SIMULATION CODE -------------------------------
*-----------------------------------------------------------------------*/
/* Record memory usage. */
if (h->log && !*status)
{
oskar_log_section(h->log, 'M', "Final memory usage");
#ifdef OSKAR_HAVE_CUDA
for (i = 0; i < h->num_gpus; ++i)
oskar_cuda_mem_log(h->log, 0, h->gpu_ids[i]);
#endif
system_mem_log(h->log);
}
/* If there are sources in the simulation and the station beam is not
* normalised to 1.0 at the phase centre, the values of noise RMS
* may give a very unexpected S/N ratio!
* The alternative would be to scale the noise to match the station
* beam gain but that would require knowledge of the station beam
* amplitude at the phase centre for each time and channel. */
if (h->log && oskar_telescope_noise_enabled(h->tel) && !*status)
{
int have_sources, amp_calibrated;
have_sources = (h->num_sky_chunks > 0 &&
oskar_sky_num_sources(h->sky_chunks[0]) > 0);
amp_calibrated = oskar_station_normalise_final_beam(
oskar_telescope_station_const(h->tel, 0));
if (have_sources && !amp_calibrated)
{
const char* a = "WARNING: System noise added to visibilities";
const char* b = "without station beam normalisation enabled.";
const char* c = "This will give an invalid signal to noise ratio.";
oskar_log_line(h->log, 'W', ' '); oskar_log_line(h->log, 'W', '*');
oskar_log_message(h->log, 'W', -1, a);
oskar_log_message(h->log, 'W', -1, b);
oskar_log_message(h->log, 'W', -1, c);
oskar_log_line(h->log, 'W', '*'); oskar_log_line(h->log, 'W', ' ');
}
}
/* Record times and summarise output files. */
if (h->log && !*status)
{
size_t log_size = 0;
char* log_data;
oskar_log_set_value_width(h->log, 25);
record_timing(h);
oskar_log_section(h->log, 'M', "Simulation complete");
oskar_log_message(h->log, 'M', 0, "Output(s):");
if (h->vis_name)
oskar_log_value(h->log, 'M', 1,
"OSKAR binary file", "%s", h->vis_name);
if (h->ms_name)
oskar_log_value(h->log, 'M', 1,
"Measurement Set", "%s", h->ms_name);
/* Write simulation log to the output files. */
log_data = oskar_log_file_data(h->log, &log_size);
#ifndef OSKAR_NO_MS
if (h->ms)
oskar_ms_add_history(h->ms, "OSKAR_LOG", log_data, log_size);
#endif
if (h->vis)
oskar_binary_write(h->vis, OSKAR_CHAR, OSKAR_TAG_GROUP_RUN,
OSKAR_TAG_RUN_LOG, 0, log_size, log_data, status);
free(log_data);
}
/* Finalise. */
oskar_simulator_finalise(h, status);
}
void oskar_interferometer_run(oskar_Interferometer* h, int* status)
{
int i, num_threads;
oskar_Thread** threads = 0;
ThreadArgs* args = 0;
if (*status || !h) return;
/* Check the visibilities are going somewhere. */
if (!h->vis_name
#ifndef OSKAR_NO_MS
&& !h->ms_name
#endif
)
{
oskar_log_error(h->log, "No output file specified.");
#ifdef OSKAR_NO_MS
if (h->ms_name)
oskar_log_error(h->log,
"OSKAR was compiled without Measurement Set support.");
#endif
*status = OSKAR_ERR_FILE_IO;
return;
}
/* Initialise if required. */
oskar_interferometer_check_init(h, status);
/* Set up worker threads. */
num_threads = h->num_devices + 1;
oskar_barrier_set_num_threads(h->barrier, num_threads);
threads = (oskar_Thread**) calloc(num_threads, sizeof(oskar_Thread*));
args = (ThreadArgs*) calloc(num_threads, sizeof(ThreadArgs));
for (i = 0; i < num_threads; ++i)
{
args[i].h = h;
args[i].num_threads = num_threads;
args[i].thread_id = i;
}
/* Record memory usage. */
if (h->log && !*status)
{
oskar_log_section(h->log, 'M', "Initial memory usage");
#ifdef OSKAR_HAVE_CUDA
for (i = 0; i < h->num_gpus; ++i)
oskar_cuda_mem_log(h->log, 0, h->gpu_ids[i]);
#endif
system_mem_log(h->log);
oskar_log_section(h->log, 'M', "Starting simulation...");
}
/* Start simulation timer. */
oskar_timer_start(h->tmr_sim);
/* Set status code. */
h->status = *status;
/* Start the worker threads. */
oskar_interferometer_reset_work_unit_index(h);
for (i = 0; i < num_threads; ++i)
threads[i] = oskar_thread_create(run_blocks, (void*)&args[i], 0);
/* Wait for worker threads to finish. */
for (i = 0; i < num_threads; ++i)
{
oskar_thread_join(threads[i]);
oskar_thread_free(threads[i]);
}
free(threads);
free(args);
/* Get status code. */
*status = h->status;
/* Record memory usage. */
if (h->log && !*status)
{
oskar_log_section(h->log, 'M', "Final memory usage");
#ifdef OSKAR_HAVE_CUDA
for (i = 0; i < h->num_gpus; ++i)
oskar_cuda_mem_log(h->log, 0, h->gpu_ids[i]);
#endif
system_mem_log(h->log);
}
/* If there are sources in the simulation and the station beam is not
* normalised to 1.0 at the phase centre, the values of noise RMS
* may give a very unexpected S/N ratio!
* The alternative would be to scale the noise to match the station
* beam gain but that would require knowledge of the station beam
* amplitude at the phase centre for each time and channel. */
if (h->log && oskar_telescope_noise_enabled(h->tel) && !*status)
{
int have_sources, amp_calibrated;
have_sources = (h->num_sky_chunks > 0 &&
oskar_sky_num_sources(h->sky_chunks[0]) > 0);
amp_calibrated = oskar_station_normalise_final_beam(
oskar_telescope_station_const(h->tel, 0));
if (have_sources && !amp_calibrated)
{
const char* a = "WARNING: System noise added to visibilities";
const char* b = "without station beam normalisation enabled.";
const char* c = "This will give an invalid signal to noise ratio.";
oskar_log_line(h->log, 'W', ' '); oskar_log_line(h->log, 'W', '*');
oskar_log_message(h->log, 'W', -1, a);
oskar_log_message(h->log, 'W', -1, b);
oskar_log_message(h->log, 'W', -1, c);
oskar_log_line(h->log, 'W', '*'); oskar_log_line(h->log, 'W', ' ');
}
}
/* Record times and summarise output files. */
if (h->log && !*status)
{
size_t log_size = 0;
char* log_data;
oskar_log_set_value_width(h->log, 25);
record_timing(h);
oskar_log_section(h->log, 'M', "Simulation complete");
oskar_log_message(h->log, 'M', 0, "Output(s):");
if (h->vis_name)
oskar_log_value(h->log, 'M', 1,
"OSKAR binary file", "%s", h->vis_name);
if (h->ms_name)
oskar_log_value(h->log, 'M', 1,
"Measurement Set", "%s", h->ms_name);
/* Write simulation log to the output files. */
log_data = oskar_log_file_data(h->log, &log_size);
#ifndef OSKAR_NO_MS
if (h->ms)
oskar_ms_add_history(h->ms, "OSKAR_LOG", log_data, log_size);
#endif
if (h->vis)
oskar_binary_write(h->vis, OSKAR_CHAR, OSKAR_TAG_GROUP_RUN,
OSKAR_TAG_RUN_LOG, 0, log_size, log_data, status);
free(log_data);
}
/* Finalise. */
oskar_interferometer_finalise(h, status);
}
