platform_ptr platform::create(cl_platform_id platform_id,
unsigned start_id) {
vector<unsigned> device_types = {CL_DEVICE_TYPE_GPU,
CL_DEVICE_TYPE_ACCELERATOR,
CL_DEVICE_TYPE_CPU};
vector<cl_device_id> ids;
for (cl_device_type device_type : device_types) {
auto known = ids.size();
cl_uint discoverd;
auto err = clGetDeviceIDs(platform_id, device_type, 0, nullptr, &discoverd);
if (err == CL_DEVICE_NOT_FOUND) {
continue; // no devices of the type found
} else if (err != CL_SUCCESS) {
throwcl("clGetDeviceIDs", err);
}
ids.resize(known + discoverd);
v2callcl(CAF_CLF(clGetDeviceIDs), platform_id, device_type,
discoverd, (ids.data() + known));
}
vector<detail::raw_device_ptr> devices;
devices.resize(ids.size());
auto lift = [](cl_device_id ptr) {
return detail::raw_device_ptr{ptr, false};
};
transform(ids.begin(), ids.end(), devices.begin(), lift);
detail::raw_context_ptr context;
context.reset(v2get(CAF_CLF(clCreateContext), nullptr,
static_cast<unsigned>(ids.size()),
ids.data(), pfn_notify, nullptr),
false);
vector<device_ptr> device_information;
for (auto& device_id : devices) {
device_information.push_back(device::create(context, device_id,
start_id++));
}
if (device_information.empty())
CAF_RAISE_ERROR("no devices for the platform found");
auto name = platform_info(platform_id, CL_PLATFORM_NAME);
auto vendor = platform_info(platform_id, CL_PLATFORM_VENDOR);
auto version = platform_info(platform_id, CL_PLATFORM_VERSION);
return make_counted<platform>(platform_id, move(context), move(name),
move(vendor), move(version),
move(device_information));
}
void opencl_metainfo::initialize() {
// get number of available platforms
auto num_platforms = v1get<cl_uint>(CAF_CLF(clGetPlatformIDs));
// get platform ids
std::vector<cl_platform_id> platforms(num_platforms);
v2callcl(CAF_CLF(clGetPlatformIDs), num_platforms, platforms.data());
if (platforms.empty()) {
throw std::runtime_error("no OpenCL platform found");
}
// support multiple platforms -> "for (auto platform : platforms)"?
auto platform = platforms.front();
// detect how many devices we got
cl_uint num_devs = 0;
cl_device_type dev_type = CL_DEVICE_TYPE_GPU;
// try get some GPU devices and try falling back to CPU devices on error
try {
num_devs = v1get<cl_uint>(CAF_CLF(clGetDeviceIDs), platform, dev_type);
}
catch (std::runtime_error&) {
dev_type = CL_DEVICE_TYPE_CPU;
num_devs = v1get<cl_uint>(CAF_CLF(clGetDeviceIDs), platform, dev_type);
}
// get available devices
std::vector<cl_device_id> ds(num_devs);
v2callcl(CAF_CLF(clGetDeviceIDs), platform, dev_type, num_devs, ds.data());
std::vector<device_ptr> devices(num_devs);
// lift raw pointer as returned by OpenCL to C++ smart pointers
auto lift = [](cl_device_id ptr) { return device_ptr{ptr, false}; };
std::transform(ds.begin(), ds.end(), devices.begin(), lift);
// create a context
context_.reset(v2get(CAF_CLF(clCreateContext), nullptr, num_devs, ds.data(),
pfn_notify, nullptr),
false);
for (auto& device : devices) {
CAF_LOG_DEBUG("creating command queue for device(s)");
command_queue_ptr cmd_queue;
try {
cmd_queue.reset(v2get(CAF_CLF(clCreateCommandQueue),
context_.get(), device.get(),
unsigned{CL_QUEUE_PROFILING_ENABLE}),
false);
}
catch (std::runtime_error&) {
CAF_LOG_DEBUG("unable to create command queue for device");
}
if (cmd_queue) {
auto max_wgs = v3get<size_t>(CAF_CLF(clGetDeviceInfo), device.get(),
unsigned{CL_DEVICE_MAX_WORK_GROUP_SIZE});
auto max_wid = v3get<cl_uint>(CAF_CLF(clGetDeviceInfo), device.get(),
unsigned{CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS});
dim_vec max_wi_per_dim(max_wid);
v2callcl(CAF_CLF(clGetDeviceInfo), device.get(),
unsigned{CL_DEVICE_MAX_WORK_ITEM_SIZES},
sizeof(size_t) * max_wid,
max_wi_per_dim.data());
devices_.push_back(device_info{std::move(device), std::move(cmd_queue),
max_wgs, max_wid, max_wi_per_dim});
}
}
if (devices_.empty()) {
std::string errstr = "could not create a command queue for any device";
CAF_LOG_ERROR(errstr);
throw std::runtime_error(std::move(errstr));
}
}