/**
* Tests creation, getting info from and destruction of
* profiler objects, and their relationship with context, device and
* queue wrapper objects.
* */
static void create_add_destroy_test() {
/* Test variables. */
CCLErr* err = NULL;
CCLBuffer* buf1 = NULL;
CCLBuffer* buf2 = NULL;
CCLProf* prof = NULL;
CCLContext* ctx = NULL;
CCLDevice* d = NULL;
CCLQueue* cq1 = NULL;
CCLQueue* cq2 = NULL;
CCLEvent* evt = NULL;
CCLEventWaitList ewl = NULL;
size_t buf_size = 8 * sizeof(cl_short);
cl_short hbuf[8] = {1, 2, 3, 4, 5, 6, 7, 8};
cl_ulong duration, eff_duration;
double time_elapsed;
/* Create a new profile object. */
prof = ccl_prof_new();
/* Get a context and a device. */
ctx = ccl_test_context_new(&err);
g_assert_no_error(err);
d = ccl_context_get_device(ctx, 0, &err);
g_assert_no_error(err);
/* Create two command queue wrappers. */
cq1 = ccl_queue_new(ctx, d, CL_QUEUE_PROFILING_ENABLE, &err);
g_assert_no_error(err);
cq2 = ccl_queue_new(ctx, d, CL_QUEUE_PROFILING_ENABLE, &err);
g_assert_no_error(err);
/* Create device buffers. */
buf1 = ccl_buffer_new(ctx, CL_MEM_READ_ONLY, buf_size, NULL, &err);
g_assert_no_error(err);
buf2 = ccl_buffer_new(ctx, CL_MEM_READ_WRITE, buf_size, NULL, &err);
g_assert_no_error(err);
/* Start profile object timer. */
ccl_prof_start(prof);
/* Transfer data to buffer. */
evt = ccl_buffer_enqueue_write(
buf1, cq1, CL_FALSE, 0, buf_size, hbuf, NULL, &err);
g_assert_no_error(err);
/* Transfer data from one buffer to another. */
evt = ccl_buffer_enqueue_copy(buf1, buf2, cq2, 0, 0, buf_size,
ccl_ewl(&ewl, evt, NULL), &err);
g_assert_no_error(err);
/* Wait for copy. */
ccl_event_wait(ccl_ewl(&ewl, evt, NULL), &err);
g_assert_no_error(err);
/* Stop profile object timer. */
ccl_prof_stop(prof);
/* Add both queues to profile object. */
ccl_prof_add_queue(prof, "A Queue", cq1);
ccl_prof_add_queue(prof, "Another Queue", cq2);
/* Process queues. */
ccl_prof_calc(prof, &err);
g_assert_no_error(err);
/* Request some profiling information. */
time_elapsed = ccl_prof_time_elapsed(prof);
duration = ccl_prof_get_duration(prof);
eff_duration = ccl_prof_get_eff_duration(prof);
g_debug("Profiling time elapsed: %lf", time_elapsed);
g_debug("Profiling duration: %d", (cl_int) duration);
g_debug("Profiling eff. duration: %d", (cl_int) eff_duration);
/* Destroy buffers. */
ccl_buffer_destroy(buf1);
ccl_buffer_destroy(buf2);
/* Unref cq1, which should not be destroyed because it is held
* by the profile object. */
ccl_queue_destroy(cq1);
/* Destroy the profile object, which will also destroy cq1. cq2
* will me merely unrefed and must still be explicitly destroyed. */
ccl_prof_destroy(prof);
/* Destroy cq2. */
ccl_queue_destroy(cq2);
/* Destroy the context. */
ccl_context_destroy(ctx);
/* Confirm that memory allocated by wrappers has been properly
* freed. */
g_assert(ccl_wrapper_memcheck());
}
/**
* Main program.
*
* @param argc Number of command line arguments.
* @param argv Vector of command line arguments.
* @return @link clo_error_codes::CLO_SUCCESS @endlink if program
* terminates successfully, or another value of #clo_error_codes if an
* error occurs.
* */
int main(int argc, char **argv)
{
/* Status var aux */
int status;
/* Context object for command line argument parsing. */
GOptionContext *context = NULL;
/* Test data structures. */
cl_uchar* host_data = NULL;
size_t bytes;
cl_ulong total_time;
FILE *outfile = NULL;
CloType clotype_elem;
/* Sorter object. */
CloSort* sorter = NULL;
/* cf4ocl wrappers. */
CCLQueue* cq_exec = NULL;
CCLQueue* cq_comm = NULL;
CCLContext* ctx = NULL;
CCLDevice* dev = NULL;
/* Profiler object. */
CCLProf* prof;
/* Host-based random number generator (mersenne twister) */
GRand* rng_host = NULL;
/* Error management object. */
GError *err = NULL;
/* Sorting benchmarks. */
cl_ulong** benchmarks = NULL;
/* Parse command line options. */
context = g_option_context_new (" - " CLO_SORT_DESCRIPTION);
g_option_context_add_main_entries(context, entries, NULL);
g_option_context_parse(context, &argc, &argv, &err);
g_if_err_goto(err, error_handler);
clotype_elem = clo_type_by_name(
type != NULL ? type : CLO_SORT_BENCHMARK_TYPE, &err);
g_if_err_goto(err, error_handler);
if (algorithm == NULL) algorithm = g_strdup(CLO_SORT_BENCHMARK_ALGORITHM);
if (alg_options == NULL) alg_options = g_strdup(CLO_SORT_BENCHMARK_ALG_OPTS);
/* Determine size in bytes of each element to sort. */
bytes = clo_type_sizeof(clotype_elem);
/* Initialize random number generator. */
rng_host = g_rand_new_with_seed(rng_seed);
/* Get the context wrapper and the chosen device. */
ctx = ccl_context_new_from_menu_full(&dev_idx, &err);
g_if_err_goto(err, error_handler);
dev = ccl_context_get_device(ctx, 0, &err);
g_if_err_goto(err, error_handler);
/* Get sorter object. */
sorter = clo_sort_new(
algorithm, alg_options, ctx, &clotype_elem, NULL, NULL, NULL,
compiler_opts, &err);
g_if_err_goto(err, error_handler);
/* Create command queues. */
cq_exec = ccl_queue_new(ctx, dev, CL_QUEUE_PROFILING_ENABLE, &err);
g_if_err_goto(err, error_handler);
cq_comm = ccl_queue_new(ctx, dev, 0, &err);
g_if_err_goto(err, error_handler);
/* Create benchmarks table. */
benchmarks = g_new(cl_ulong*, maxpo2);
for (unsigned int i = 0; i < maxpo2; i++)
benchmarks[i] = g_new0(cl_ulong, runs);
/* Print options. */
printf("\n =========================== Selected options ============================\n\n");
printf(" Random number generator seed: %u\n", rng_seed);
printf(" Maximum local worksize (0 is auto-select): %d\n", (int) lws);
printf(" Type of elements to sort: %s\n", clo_type_get_name(clotype_elem));
printf(" Number of runs: %d\n", runs);
printf(" Compiler Options: %s\n", compiler_opts);
/* Create host buffer. */
host_data = g_slice_alloc(bytes * (1 << maxpo2));
/* Perform test. */
for (unsigned int N = 4; N <= maxpo2; N++) {
unsigned int num_elems = 1 << N;
gboolean sorted_ok = TRUE;
for (unsigned int r = 0; r < runs; r++) {
/* Initialize host buffer. */
for (unsigned int i = 0; i < num_elems; i++) {
clo_bench_rand(
rng_host, clotype_elem, host_data + bytes * i);
}
/* Perform sort. */
clo_sort_with_host_data(sorter, cq_exec, cq_comm,
host_data, host_data, num_elems, lws, &err);
g_if_err_goto(err, error_handler);
/* Perform profiling. */
prof = ccl_prof_new();
ccl_prof_add_queue(prof, "q_exec", cq_exec);
ccl_prof_calc(prof, &err);
g_if_err_goto(err, error_handler);
/* Save duration to benchmarks. */
benchmarks[N - 1][r] = ccl_prof_get_duration(prof);
ccl_prof_destroy(prof);
/* Check if sorting was well performed. */
sorted_ok = TRUE;
/* Wait on host thread for data transfer queue to finish... */
ccl_queue_finish(cq_comm, &err);
g_if_err_goto(err, error_handler);
/* Start check. */
for (unsigned int i = 0; i < num_elems - 1; i++) {
/* Perform comparison. */
if (clo_bench_compare(clotype_elem, host_data + bytes*i,
host_data + bytes*(i + 1)) > 0) {
sorted_ok = FALSE;
break;
}
}
}
/* Print info. */
total_time = 0;
for (unsigned int i = 0; i < runs; i++)
total_time += benchmarks[N - 1][i];
printf(" - 2^%d: %lf Mkeys/s %s\n", N,
(1e-6 * num_elems * runs) / (total_time * 1e-9),
sorted_ok ? "" : "(sort did not work)");
}
/* Save benchmarks to file, if filename was given as cli option. */
if (out) {
outfile = fopen(out, "w");
for (unsigned int i = 0; i < maxpo2; i++) {
fprintf(outfile, "%d", i);
for (unsigned int j = 0; j < runs; j++) {
fprintf(outfile, "\t%lu", (unsigned long)benchmarks[i][j]);
}
fprintf(outfile, "\n");
}
fclose(outfile);
}
/* If we get here, everything went Ok. */
status = CLO_SUCCESS;
g_assert(err == NULL);
goto cleanup;
error_handler:
/* Handle error. */
g_assert(err != NULL);
fprintf(stderr, "Error: %s\n", err->message);
g_error_free(err);
cleanup:
/* Free sorter object. */
if (sorter) clo_sort_destroy(sorter);
/* Free command line options. */
if (context) g_option_context_free(context);
if (algorithm) g_free(algorithm);
if (alg_options) g_free(alg_options);
if (compiler_opts) g_free(compiler_opts);
if (out) g_free(out);
/* Free host-based random number generator. */
if (rng_host) g_rand_free(rng_host);
/* Free OpenCL wrappers. */
if (cq_exec) ccl_queue_destroy(cq_exec);
if (cq_comm) ccl_queue_destroy(cq_comm);
if (ctx) ccl_context_destroy(ctx);
/* Free host resources */
if (host_data) g_slice_free1(bytes * (1 << maxpo2), host_data);
/* Free benchmarks. */
if (benchmarks) {
for (unsigned int i = 0; i < maxpo2; i++)
if (benchmarks[i]) g_free(benchmarks[i]);
g_free(benchmarks);
}
/* Bye. */
return status;
}
