void omp_set_num_threads (int p1)
{
#if defined(DEBUG)
fprintf (stderr, PACKAGE_NAME": omp_set_num_threads is at %p\n", omp_set_num_threads_real);
fprintf (stderr, PACKAGE_NAME": omp_set_num_threads params %d\n", p1);
#endif
if (omp_set_num_threads_real != NULL && EXTRAE_INITIALIZED())
{
Backend_ChangeNumberOfThreads (p1);
Backend_Enter_Instrumentation (2);
Probe_OpenMP_SetNumThreads_Entry (p1);
omp_set_num_threads_real (p1);
Probe_OpenMP_SetNumThreads_Exit ();
Backend_Leave_Instrumentation ();
}
else if (omp_set_num_threads_real != NULL)
{
omp_set_num_threads_real (p1);
}
else
{
fprintf (stderr, PACKAGE_NAME": omp_set_num_threads is not hooked! exiting!!\n");
exit (0);
}
}
JNIEXPORT void JNICALL Java_es_bsc_cepbatools_extrae_Wrapper_SetNumThreads(
JNIEnv *env, jclass jc, jint numthreads)
{
UNREFERENCED(env);
UNREFERENCED(jc);
NUMTHREADS = numthreads;
Backend_ChangeNumberOfThreads (numthreads);
Extrae_set_numthreads_function (get_num_threads);
}
/**************************************************************************
Lets change the number of active threads
*************************************************************************/
void Extrae_change_number_of_threads_Wrapper (unsigned nthreads)
{
Backend_ChangeNumberOfThreads (nthreads);
}
void Extrae_OpenCL_clCreateCommandQueue (cl_command_queue queue,
cl_device_id device, cl_command_queue_properties properties)
{
if (!Extrae_OpenCL_lookForOpenCLQueue (queue, NULL))
{
cl_int err;
char _threadname[THREAD_INFO_NAME_LEN];
char _hostname[HOST_NAME_MAX];
char *_device_type;
int prev_threadid, found, idx;
cl_device_type device_type;
cl_event event;
idx = nCommandQueues;
CommandQueues = (RegisteredCommandQueue_t*) realloc (
CommandQueues,
sizeof(RegisteredCommandQueue_t)*(nCommandQueues+1));
if (CommandQueues == NULL)
{
fprintf (stderr, PACKAGE_NAME": Fatal error! Failed to allocate memory for OpenCL Command Queues\n");
exit (-1);
}
CommandQueues[idx].queue = queue;
CommandQueues[idx].isOutOfOrder =
(properties & CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE) != 0;
err = clGetDeviceInfo (device, CL_DEVICE_TYPE, sizeof(device_type), &device_type, NULL);
if (err == CL_SUCCESS)
{
if (device_type == CL_DEVICE_TYPE_GPU)
_device_type = "GPU";
else if (device_type == CL_DEVICE_TYPE_CPU)
_device_type = "CPU";
else
_device_type = "Other";
}
else
_device_type = "Unknown";
/* Was the thread created before (i.e. did we executed a cudadevicereset?) */
if (gethostname(_hostname, HOST_NAME_MAX) == 0)
sprintf (_threadname, "OpenCL-%s-CQ%d-%s", _device_type, 1+idx,
_hostname);
else
sprintf (_threadname, "OpenCL-%s-CQ%d-%s", _device_type, 1+idx,
"unknown-host");
prev_threadid = Extrae_search_thread_name (_threadname, &found);
if (found)
{
/* If thread name existed, reuse its thread id */
CommandQueues[idx].threadid = prev_threadid;
}
else
{
/* For timing purposes we change num of threads here instead of doing Backend_getNumberOfThreads() + CUDAdevices*/
Backend_ChangeNumberOfThreads (Backend_getNumberOfThreads() + 1);
CommandQueues[idx].threadid = Backend_getNumberOfThreads()-1;
/* Set thread name */
Extrae_set_thread_name (CommandQueues[idx].threadid, _threadname);
}
CommandQueues[idx].nevents = 0;
#ifdef CL_VERSION_1_2
err = clEnqueueBarrierWithWaitList (queue, 0, NULL, &event);
#else
err = clEnqueueBarrier (queue);
if (err == CL_SUCCESS)
err = clEnqueueMarker (queue, &event);
#endif
CommandQueues[idx].host_reference_time = TIME;
if (err == CL_SUCCESS)
{
err = clFinish(queue);
if (err != CL_SUCCESS)
{
fprintf (stderr, PACKAGE_NAME": Error in clFinish (error = %d)! Dying...\n", err);
exit (-1);
}
err = clGetEventProfilingInfo (event, CL_PROFILING_COMMAND_SUBMIT,
sizeof(cl_ulong), &(CommandQueues[idx].device_reference_time),
NULL);
if (err != CL_SUCCESS)
{
fprintf (stderr, PACKAGE_NAME": Error in clGetEventProfilingInfo (error = %d)! Dying...\n", err);
exit (-1);
}
}
else
{
fprintf (stderr, PACKAGE_NAME": Error while looking for clock references in host & accelerator\n");
exit (-1);
}
nCommandQueues++;
}
}
