This might not the the most narrowed down problem but..
The program implements a wrapper around all of the OpenCL stuff. The wrapper detects all of the OpenCL devices then wraps these into yet another wrappers. The device wrappers contain all the objects related to it such as allocated cl_mem buffers the associated context etc.
I have chacked multiple times if there are no mistake, if no pointers are reused like if due to some error device wrappers from different platforms would share the same platform pointer. but no.
The problem: When I divide work between all the computational devices on my laptop (CPU+Intel GPU + Nvidia GPU), kernel execution issued to the NVIDIA GPU crahses with CL_INVALID_COMMAND_QUEUE.
I've checked everything.
I've tried the following scenarios:
- Intel GPU and CPU simultaneously => everything works fine
- two CPUs simultaneously (server) => everything works fine
- if I mix devices from two platforms on the laptop => it crashes with CL_INVALID_COMMAND_QUEUE. it crashes ONLY on the Nvidia GPU.
most of the initialization code is below.
std::cout << "Initializing the OpenCL engine..\n";
cl_int ret;
unsigned int nrOfActiveContexts = 0;
ret = clGetPlatformIDs(0, NULL, &mRetNumPlatforms);
if (mRetNumPlatforms > 0)
{
this->mPlatforms.resize(mRetNumPlatforms);
}
else
{
fprintf(stderr, "No OpenCL platform available.\n");
exit(1);
}
ret = clGetPlatformIDs(mRetNumPlatforms, mPlatforms.data(), NULL);
std::vector<cl_device_id> devices;
cl_context context;
cl_uint numberOfDevices;
//query for available compute platforms
for (int i = 0; i < mPlatforms.size() ; i++)
{
bool error = false;
numberOfDevices = 0;
devices.clear();
context = NULL;
cl_device_type deviceTypes = CL_DEVICE_TYPE_ALL;
if (useCPU &&useGPU)
deviceTypes = CL_DEVICE_TYPE_ALL;
else if (useCPU)
deviceTypes = CL_DEVICE_TYPE_CPU;
else if (useGPU)
deviceTypes = CL_DEVICE_TYPE_GPU;
ret = clGetDeviceIDs(mPlatforms[i], deviceTypes, 0, NULL, &numberOfDevices);
if (numberOfDevices > 0)
{
devices.resize( numberOfDevices);
ret = clGetDeviceIDs(mPlatforms[i], deviceTypes,
numberOfDevices, devices.data(), NULL);
}
else continue;
context = clCreateContext(NULL, numberOfDevices, devices.data(), NULL, NULL, &ret);
if (ret != CL_SUCCESS)
throw(std::abort);
mContexts.push_back(context);
if (ret != CL_SUCCESS)
{
error = true;
}
//query device properties create Workers
size_t ret_size;
cl_uint compute_units;
cl_ulong max_alloc;
size_t max_work_size;
std::string name;
std::vector<char> c_name;
for (int y = 0; y < devices.size(); y++)
{
ret_size = compute_units = max_alloc = max_work_size = 0;
c_name.clear();
ret = clGetDeviceInfo(devices[y], CL_DEVICE_NAME, NULL, NULL, &ret_size);
if (ret != CL_SUCCESS)
{
error = true; goto errored;
}
c_name.resize(ret_size);
ret = clGetDeviceInfo(devices[y], CL_DEVICE_NAME, c_name.size(), c_name.data(), &ret_size);
if (ret != CL_SUCCESS)
{
error = true; goto errored;
}
name = std::string(c_name.begin(), c_name.end());
name = std::regex_replace(name, std::regex("[' ']{2,}"), " ");
cl_device_type devType;
ret = clGetDeviceInfo(devices[y], CL_DEVICE_TYPE, sizeof(cl_device_type), (void *)&devType, NULL);
if (ret != CL_SUCCESS)
{
error = true; goto errored;
}
ret = clGetDeviceInfo(devices[y], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(cl_uint), (void *)&compute_units, NULL);
if (ret != CL_SUCCESS)
{
error = true; goto errored;
}
ret = clGetDeviceInfo(devices[y], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), (void *)&max_work_size, NULL);
if (ret != CL_SUCCESS)
{
error = true; goto errored;
}
CWorker::eWorkerType type;
if (devType & CL_DEVICE_TYPE_GPU)
type = CWorker::eWorkerType::GPU;
else
if (devType & CL_DEVICE_TYPE_CPU)
type = CWorker::eWorkerType::CPU;
if (type == CWorker::eWorkerType::CPU)
{
if (compute_units > 8)
max_work_size = compute_units / 4;
else if (compute_units == 8)
max_work_size = 2;
else
max_work_size = 1;
}
ret = clGetDeviceInfo(devices[y], CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(cl_ulong), (void *)&max_alloc, NULL);
if (ret != CL_SUCCESS)
{
error = true;
goto errored;
}
errored:
if (error != true)
{
CWorker * w = new CWorker();
w->setDevice(devices[y]);
w->setMaxComputeUnits(compute_units);
w->setMaxMemAlloc(max_alloc);
w->setMaxWorkGroupSize(max_work_size);
w->setName(name);
std::cmatch cm;
if (std::regex_search(name.data(), cm, std::regex("\\w\+")))
w->setShortName(std::string(cm[0]) +"-"+ std::to_string(mWorkers.size()+1));
w->setContext(context);
w->setType(type);
mWorkers.push_back(w);
}
}
nrOfActiveContexts++;
}
if (mWorkers.size() > 0)
mInitialised = true;
if (mWorkers.size() > 0)
return true;
else return false;
