void * embb_mtapi_alloc_allocate(unsigned int bytes) {
void * ptr = embb_alloc(bytes);
if (ptr != NULL) {
embb_internal__atomic_fetch_and_add_4(
&embb_mtapi_alloc_bytes_allocated, sizeof(unsigned int)+bytes);
}
return ptr;
}
int embb_mtapi_network_buffer_initialize(
embb_mtapi_network_buffer_t * that,
int capacity) {
int result = 1;
that->position = 0;
that->size = 0;
that->data = (char*)embb_alloc((size_t)capacity);
if (NULL != that->data) {
that->capacity = capacity;
} else {
that->capacity = 0;
result = 0;
}
return result;
}
int embb_thread_create(embb_thread_t* thread, const embb_core_set_t* core_set,
embb_thread_start_t func, void* arg) {
pthread_attr_t attr; /* Used to set thread affinities */
int status = pthread_attr_init(&attr);
if (status != 0) return EMBB_ERROR;
if (core_set != NULL) {
#if defined(EMBB_PLATFORM_HAS_GLIB_CPU) || \
defined(EMBB_PLATFORM_HAS_HEADER_CPUSET)
assert(embb_core_count_available() < CPU_SETSIZE &&
"Core sets are only supported up to CPU_SETSIZE processors!");
#ifdef EMBB_PLATFORM_HAS_GLIB_CPU
cpu_set_t cpuset;
#else
cpuset_t cpuset;
#endif
CPU_ZERO(&cpuset); /* Disable all processors */
for (unsigned int i = 0; i < embb_core_count_available(); i++) {
if (embb_core_set_contains(core_set, i)) {
CPU_SET(i, &cpuset);
}
}
status = pthread_attr_setaffinity_np(&attr, sizeof(cpuset), &cpuset);
if (status != 0) return EMBB_ERROR;
#else
embb_log_write("base_c", EMBB_LOG_LEVEL_WARNING, "Could not set thread "
"affinity, since no implementation available!\n");
#endif
}
/* Dynamic allocation of thread arguments. Freed on call of join. */
thread->embb_internal_arg = (embb_internal_thread_arg_t*)
embb_alloc(sizeof(embb_internal_thread_arg_t));
thread->embb_internal_arg->func = func;
thread->embb_internal_arg->arg = arg;
status = pthread_create(
&(thread->embb_internal_handle), /* pthread handle */
&attr, /* additional attributes,
e.g., affinities */
embb_internal_thread_start, /* thread start function */
(void*)(thread->embb_internal_arg)); /* arguments to thread start func. */
if (status != 0) return EMBB_ERROR;
status = pthread_attr_destroy(&attr);
if (status != 0) return EMBB_ERROR;
return EMBB_SUCCESS;
}
int embb_thread_create(embb_thread_t* thread, const embb_core_set_t* core_set,
embb_thread_start_t func, void *arg) {
assert(thread != NULL);
thread->embb_internal_arg = (embb_internal_thread_arg_t*)
embb_alloc(sizeof(embb_internal_thread_arg_t));
if (thread->embb_internal_arg == NULL) return EMBB_NOMEM;
thread->embb_internal_arg->func = func;
thread->embb_internal_arg->arg = arg;
thread->embb_internal_handle = CreateThread(
0, /* no security */
0, /* default stack size */
embb_internal_thread_start, /* entry function */
(LPVOID)thread->embb_internal_arg, /* parameters */
0, /* no creation arguments */
0); /* no system thread ID */
if (thread->embb_internal_handle == NULL) {
return EMBB_ERROR;
}
if (core_set != NULL) { /* Set thread affinity, if a core set is given */
DWORD_PTR core_mask = 0;
DWORD bit_mask = 1;
assert(embb_core_count_available() < 64);
for (unsigned int i = 0; i < embb_core_count_available(); i++) {
if (embb_core_set_contains(core_set, i)) {
core_mask |= bit_mask;
}
bit_mask <<= 1;
}
if (SetThreadAffinityMask(thread->embb_internal_handle, core_mask)
== (DWORD_PTR)NULL) {
return EMBB_ERROR;
}
}
return EMBB_SUCCESS;
}
mtapi_action_hndl_t mtapi_opencl_action_create(
MTAPI_IN mtapi_job_id_t job_id,
MTAPI_IN char* kernel_source,
MTAPI_IN char* kernel_name,
MTAPI_IN mtapi_size_t local_work_size,
MTAPI_IN mtapi_size_t element_size,
MTAPI_IN void* node_local_data,
MTAPI_IN mtapi_size_t node_local_data_size,
MTAPI_OUT mtapi_status_t* status) {
mtapi_status_t local_status = MTAPI_ERR_UNKNOWN;
cl_int err;
embb_mtapi_opencl_plugin_t * plugin = &embb_mtapi_opencl_plugin;
embb_mtapi_opencl_action_t * action =
(embb_mtapi_opencl_action_t*)embb_alloc(
sizeof(embb_mtapi_opencl_action_t));
mtapi_action_hndl_t action_hndl = { 0, 0 }; // invalid handle
size_t kernel_length = strlen(kernel_source);
mtapi_boolean_t free_program_on_error = MTAPI_FALSE;
mtapi_boolean_t free_kernel_on_error = MTAPI_FALSE;
mtapi_boolean_t free_node_local_data_on_error = MTAPI_FALSE;
action->local_work_size = local_work_size;
action->element_size = element_size;
/* initialization */
action->program = clCreateProgramWithSource(plugin->context,
1, &kernel_source, &kernel_length, &err);
if (CL_SUCCESS == err) {
free_program_on_error = MTAPI_TRUE;
err = clBuildProgram(action->program, 1, &plugin->device_id,
NULL, NULL, NULL);
} else {
err = clGetProgramBuildInfo(action->program, plugin->device_id,
CL_PROGRAM_BUILD_LOG, 1024, buffer, NULL);
}
if (CL_SUCCESS == err) {
action->kernel = clCreateKernel(action->program, kernel_name, &err);
if (CL_SUCCESS == err) {
free_kernel_on_error = MTAPI_TRUE;
}
}
if (0 < node_local_data_size) {
action->node_local_data = clCreateBuffer(plugin->context, CL_MEM_READ_ONLY,
node_local_data_size, NULL, &err);
if (CL_SUCCESS == err) {
free_node_local_data_on_error = MTAPI_TRUE;
}
action->node_local_data_size = (int)node_local_data_size;
if (CL_SUCCESS == err) {
err = clEnqueueWriteBuffer(plugin->command_queue,
action->node_local_data, CL_TRUE, 0,
(size_t)action->node_local_data_size, node_local_data, 0, NULL, NULL);
}
} else {
action->node_local_data = NULL;
action->node_local_data_size = 0;
}
if (CL_SUCCESS == err) {
err = clSetKernelArg(action->kernel, 4, sizeof(cl_mem),
(const void*)&action->node_local_data);
}
if (CL_SUCCESS == err) {
err = clSetKernelArg(action->kernel, 5, sizeof(cl_int),
(const void*)&action->node_local_data_size);
}
if (CL_SUCCESS == err) {
action_hndl = mtapi_ext_plugin_action_create(
job_id,
opencl_task_start,
opencl_task_cancel,
opencl_action_finalize,
action,
node_local_data,
node_local_data_size,
MTAPI_NULL,
&local_status);
} else {
if (free_node_local_data_on_error) {
clReleaseMemObject(action->node_local_data);
}
if (free_kernel_on_error) {
clReleaseKernel(action->kernel);
}
if (free_program_on_error) {
clReleaseProgram(action->program);
}
embb_free(action);
}
mtapi_status_set(status, local_status);
return action_hndl;
}
static void opencl_task_start(
MTAPI_IN mtapi_task_hndl_t task,
MTAPI_OUT mtapi_status_t* status) {
mtapi_status_t local_status = MTAPI_ERR_UNKNOWN;
cl_int err;
if (embb_mtapi_node_is_initialized()) {
embb_mtapi_node_t * node = embb_mtapi_node_get_instance();
if (embb_mtapi_task_pool_is_handle_valid(node->task_pool, task)) {
embb_mtapi_task_t * local_task =
embb_mtapi_task_pool_get_storage_for_handle(node->task_pool, task);
if (embb_mtapi_action_pool_is_handle_valid(
node->action_pool, local_task->action)) {
embb_mtapi_action_t * local_action =
embb_mtapi_action_pool_get_storage_for_handle(
node->action_pool, local_task->action);
embb_mtapi_opencl_plugin_t * plugin = &embb_mtapi_opencl_plugin;
embb_mtapi_opencl_action_t * opencl_action =
(embb_mtapi_opencl_action_t*)local_action->plugin_data;
embb_mtapi_opencl_task_t * opencl_task =
(embb_mtapi_opencl_task_t*)embb_alloc(
sizeof(embb_mtapi_opencl_task_t));
size_t elements = local_task->result_size /
opencl_action->element_size;
size_t global_work_size;
if (0 == elements)
elements = 1;
global_work_size =
round_up(opencl_action->local_work_size, elements);
opencl_task->task = task;
opencl_task->arguments_size = (int)local_task->arguments_size;
if (0 < local_task->arguments_size) {
opencl_task->arguments = clCreateBuffer(plugin->context,
CL_MEM_READ_ONLY, local_task->arguments_size, NULL, &err);
} else {
opencl_task->arguments = NULL;
}
opencl_task->result_buffer_size = (int)local_task->result_size;
if (0 < local_task->result_size) {
opencl_task->result_buffer = clCreateBuffer(plugin->context,
CL_MEM_WRITE_ONLY, local_task->result_size, NULL, &err);
} else {
opencl_task->result_buffer = NULL;
}
err = clSetKernelArg(opencl_action->kernel, 0, sizeof(cl_mem),
(const void*)&opencl_task->arguments);
err |= clSetKernelArg(opencl_action->kernel, 1, sizeof(cl_int),
(const void*)&opencl_task->arguments_size);
err |= clSetKernelArg(opencl_action->kernel, 2, sizeof(cl_mem),
(const void*)&opencl_task->result_buffer);
err |= clSetKernelArg(opencl_action->kernel, 3, sizeof(cl_int),
(const void*)&opencl_task->result_buffer_size);
err |= clEnqueueWriteBuffer(plugin->command_queue,
opencl_task->arguments, CL_FALSE, 0,
(size_t)opencl_task->arguments_size, local_task->arguments,
0, NULL, NULL);
if (CL_SUCCESS == err) {
embb_mtapi_task_set_state(local_task, MTAPI_TASK_RUNNING);
err |= clEnqueueNDRangeKernel(plugin->command_queue,
opencl_action->kernel, 1, NULL,
&global_work_size, &opencl_action->local_work_size, 0, NULL, NULL);
err |= clEnqueueReadBuffer(plugin->command_queue,
opencl_task->result_buffer, CL_FALSE, 0,
(size_t)opencl_task->result_buffer_size, local_task->result_buffer,
0, NULL, &opencl_task->kernel_finish_event);
err |= clSetEventCallback(opencl_task->kernel_finish_event,
CL_COMPLETE, opencl_task_complete, opencl_task);
}
err |= clFlush(plugin->command_queue);
if (CL_SUCCESS != err) {
embb_mtapi_task_set_state(local_task, MTAPI_TASK_ERROR);
local_status = MTAPI_ERR_ACTION_FAILED;
} else {
local_status = MTAPI_SUCCESS;
}
}
}
}
mtapi_status_set(status, local_status);
}