/**
* 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());
}
/**
* @internal
*
* @brief Tests basic read/write operations from/to buffer objects.
* */
static void read_write_test() {
/* Test variables. */
CCLContext * ctx = NULL;
CCLDevice * d = NULL;
CCLBuffer * b = NULL;
CCLQueue * q;
cl_uint h_in[CCL_TEST_BUFFER_SIZE];
cl_uint h_out[CCL_TEST_BUFFER_SIZE];
size_t buf_size = sizeof(cl_uint) * CCL_TEST_BUFFER_SIZE;
CCLErr * err = NULL;
/* Create a host array, put some stuff in it. */
for (guint i = 0; i < CCL_TEST_BUFFER_SIZE; ++i)
h_in[i] = g_test_rand_int();
/* Get the test context with the pre-defined device. */
ctx = ccl_test_context_new(&err);
g_assert_no_error(err);
/* Get first device in context. */
d = ccl_context_get_device(ctx, 0, &err);
g_assert_no_error(err);
/* Create a command queue. */
q = ccl_queue_new(ctx, d, 0, &err);
g_assert_no_error(err);
/* Create regular buffer and write data from the host buffer. */
b = ccl_buffer_new(
ctx, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, buf_size, h_in, &err);
g_assert_no_error(err);
/* Read data back to host. */
ccl_buffer_enqueue_read(
b, q, CL_TRUE, 0, buf_size, (void *) h_out, NULL, &err);
g_assert_no_error(err);
/* Check data is OK. */
for (guint i = 0; i < CCL_TEST_BUFFER_SIZE; ++i)
g_assert_cmpuint(h_in[i], ==, h_out[i]);
/* Set some other data in host array. */
for (guint i = 0; i < CCL_TEST_BUFFER_SIZE; ++i)
h_in[i] = g_test_rand_int();
/* Write it explicitly to buffer. */
ccl_buffer_enqueue_write(
b, q, CL_TRUE, 0, buf_size, (void *) h_in, NULL, &err);
g_assert_no_error(err);
/* Read new data to host. */
ccl_buffer_enqueue_read(
b, q, CL_TRUE, 0, buf_size, (void *) h_out, NULL, &err);
g_assert_no_error(err);
/* Check data is OK. */
for (guint i = 0; i < CCL_TEST_BUFFER_SIZE; ++i)
g_assert_cmpuint(h_in[i], ==, h_out[i]);
/* Free stuff. */
ccl_buffer_destroy(b);
ccl_queue_destroy(q);
ccl_context_destroy(ctx);
/* Confirm that memory allocated by wrappers has been properly
* freed. */
g_assert(ccl_wrapper_memcheck());
}
/**
* Canonical example main function.
* */
int main(int argc, char** argv) {
/* Number of elements in buffer. */
size_t buf_n = DEF_BUF_N;
/* Device selected specified in the command line. */
int dev_idx = -1;
/* Program return value. */
int ret_val;
/* Check if a device was specified in the command line. */
if (argc >= 2) {
dev_idx = atoi(argv[1]);
}
/* Check if a new buffer size was specified in the command line. */
if (argc >= 3) {
buf_n = atoi(argv[2]);
}
/* Wrappers. */
CCLContext* ctx = NULL;
CCLProgram* prg = NULL;
CCLDevice* dev = NULL;
CCLQueue* queue = NULL;
CCLKernel* krnl = NULL;
CCLBuffer* a_dev;
CCLBuffer* b_dev;
CCLBuffer* c_dev;
CCLEvent* evt_write1;
CCLEvent* evt_write2;
CCLEvent* evt_exec;
CCLEventWaitList ewl = NULL;
/* Profiler. */
CCLProf* prof;
/* Global and local worksizes. */
size_t gws = 0;
size_t lws = 0;
/* Host buffers. */
cl_uint* a_host = NULL;
cl_uint* b_host = NULL;
cl_uint* c_host = NULL;
cl_uint d_host;
/* Error reporting object. */
CCLErr* err = NULL;
/* Check results flag. */
cl_bool check_result;
/* Create a context with device selected from menu. */
ctx = ccl_context_new_from_menu_full(&dev_idx, &err);
HANDLE_ERROR(err);
/* Get the selected device. */
dev = ccl_context_get_device(ctx, 0, &err);
HANDLE_ERROR(err);
/* Create a new program from kernel source. */
prg = ccl_program_new_from_source(ctx, KERNEL_SRC, &err);
HANDLE_ERROR(err);
/* Build program. */
ccl_program_build(prg, NULL, &err);
HANDLE_ERROR(err);
/* Create a command queue. */
queue = ccl_queue_new(ctx, dev, CL_QUEUE_PROFILING_ENABLE, &err);
HANDLE_ERROR(err);
/* Get kernel object. */
krnl = ccl_program_get_kernel(prg, KERNEL_NAME, &err);
HANDLE_ERROR(err);
/* Get worksizes. */
lws = ccl_kernel_suggest_worksizes(krnl, dev, 1, &buf_n, &gws, &lws, &err);
HANDLE_ERROR(err);
/* Show worksizes. */
printf("\n");
printf(" * Global worksize: %d\n", (int) gws);
printf(" * Local worksize : %d\n", (int) lws);
/* Initialize host buffers. */
a_host = (cl_uint*) malloc(sizeof(cl_uint) * buf_n);
b_host = (cl_uint*) malloc(sizeof(cl_uint) * buf_n);
c_host = (cl_uint*) malloc(sizeof(cl_uint) * buf_n);
/* Fill host buffers. */
for (cl_uint i = 0; i < buf_n; ++i) {
a_host[i] = i;
b_host[i] = buf_n - i;
}
d_host = buf_n / 4;
/* Create device buffers. */
a_dev = ccl_buffer_new(ctx, CL_MEM_READ_ONLY,
buf_n * sizeof(cl_uint), NULL, &err);
HANDLE_ERROR(err);
b_dev = ccl_buffer_new(ctx, CL_MEM_READ_ONLY,
buf_n * sizeof(cl_uint), NULL, &err);
HANDLE_ERROR(err);
c_dev = ccl_buffer_new(ctx, CL_MEM_WRITE_ONLY,
buf_n * sizeof(cl_uint), NULL, &err);
HANDLE_ERROR(err);
/* Copy host data to device buffers without waiting for transfer
* to terminate before continuing host program. */
evt_write1 = ccl_buffer_enqueue_write(a_dev, queue, CL_FALSE, 0,
buf_n * sizeof(cl_uint), a_host, NULL, &err);
HANDLE_ERROR(err);
evt_write2 = ccl_buffer_enqueue_write(b_dev, queue, CL_FALSE, 0,
buf_n * sizeof(cl_uint), b_host, NULL, &err);
HANDLE_ERROR(err);
/* Initialize event wait list and add the two transfer events. */
ccl_event_wait_list_add(&ewl, evt_write1, evt_write2, NULL);
/* Execute program kernel, waiting for the two transfer events
* to terminate (this will empty the event wait list). */
evt_exec = ccl_program_enqueue_kernel(prg, KERNEL_NAME, queue, 1,
NULL, &gws, &lws, &ewl, &err,
/* Kernel arguments. */
a_dev, b_dev, c_dev,
ccl_arg_priv(d_host, cl_uint), ccl_arg_priv(buf_n, cl_uint),
NULL);
HANDLE_ERROR(err);
/* Add the kernel termination event to the wait list. */
ccl_event_wait_list_add(&ewl, evt_exec, NULL);
/* Sync. queue for events in wait list (just the execute event in
* this case) to terminate before going forward... */
ccl_enqueue_barrier(queue, &ewl, &err);
HANDLE_ERROR(err);
/* Read back results from host waiting for transfer to terminate
* before continuing host program. */
ccl_buffer_enqueue_read(c_dev, queue, CL_TRUE, 0,
buf_n * sizeof(cl_uint), c_host, NULL, &err);
HANDLE_ERROR(err);
/* Check results are as expected (not available with OpenCL stub). */
check_result = CL_TRUE;
for (cl_uint i = 0; i < buf_n; ++i) {
if(c_host[i] != a_host[i] + b_host[i] + d_host) {
check_result = CL_FALSE;
break;
}
}
if (check_result) {
fprintf(stdout, " * Kernel execution produced the expected results.\n");
ret_val = EXIT_SUCCESS;
} else {
fprintf(stderr,
" * Kernel execution failed to produce the expected results.\n");
ret_val = EXIT_FAILURE;
}
/* Perform profiling. */
prof = ccl_prof_new();
ccl_prof_add_queue(prof, "queue1", queue);
ccl_prof_calc(prof, &err);
HANDLE_ERROR(err);
/* Show profiling info. */
ccl_prof_print_summary(prof);
/* Export profiling info. */
ccl_prof_export_info_file(prof, "out.tsv", &err);
HANDLE_ERROR(err);
/* Destroy profiler object. */
ccl_prof_destroy(prof);
/* Destroy host buffers. */
free(a_host);
free(b_host);
free(c_host);
/* Destroy wrappers. */
ccl_buffer_destroy(a_dev);
ccl_buffer_destroy(b_dev);
ccl_buffer_destroy(c_dev);
ccl_queue_destroy(queue);
ccl_program_destroy(prg);
ccl_context_destroy(ctx);
/* Confirm that memory allocated by wrappers has been properly freed. */
assert(ccl_wrapper_memcheck());
/* Bye. */
return ret_val;
}
