bool OpenCLPlatform::createContext() {
cl_int err;
cl_context_properties contextProperties[] = { CL_CONTEXT_PLATFORM,
(cl_context_properties) platform_(), 0 };
context_ = cl::Context(
CL_DEVICE_TYPE_CPU | CL_DEVICE_TYPE_GPU,
contextProperties,
NULL,
NULL,
&err);
if (err != CL_SUCCESS) {
LOG(ERROR)<< "failed to create context.";
return false;
}
LOG(INFO)<< "create OpenCL context for platform " << this->name();
std::vector<caffe::OpenCLDevice>::iterator it;
for (it = devices.begin(); it != devices.end(); it++) {
it->SetContext(
context_);
}
return true;
}
CL::CL() {
cl_int err;
std::vector<cl::Platform> platforms;
if(cl::Platform::get(&platforms) == CL_INVALID_VALUE) {
fprintf(stderr, "[OpenCL] No platforms available\n");
util_abort();
}
platform_ = platforms[0]; //Just select the first platform
std::string name, version, extensions;
platform_.getInfo(CL_PLATFORM_NAME, &name);
platform_.getInfo(CL_PLATFORM_VERSION, &version);
platform_.getInfo(CL_PLATFORM_EXTENSIONS, &extensions);
fprintf(verbose, "[OpenCL] Platform: %s %s\n"
" Extensions: %s\n", name.c_str(), version.c_str() ,extensions.c_str());
#if defined (__APPLE__) || defined(MACOSX)
CGLContextObj kCGLContext = CGLGetCurrentContext();
CGLShareGroupObj kCGLShareGroup = CGLGetShareGroup(kCGLContext);
cl_context_properties properties[] =
{
CL_CONTEXT_PROPERTY_USE_CGL_SHAREGROUP_APPLE, (cl_context_properties)kCGLShareGroup,
0
};
#elif defined WIN32
HGLRC current_context = wglGetCurrentContext();
HDC current_dc = wglGetCurrentDC();
if(current_dc == NULL || current_context == NULL) {
fprintf(stderr,"[OpenCL] No OpenGL context active\n");
util_abort();
}
cl_context_properties properties[] = {
CL_CONTEXT_PLATFORM, (cl_context_properties)platform_(),
CL_WGL_HDC_KHR, (intptr_t) current_dc,
CL_GL_CONTEXT_KHR, (intptr_t) current_context,
0
};
#else
if(glXGetCurrentContext() == NULL) {
fprintf(stderr, "[OpenCL] glXGetCurrentContex() return NULL. Make sure to create OpenGL context before create the CL-context\n");
util_abort();
}
cl_context_properties properties[] =
{
CL_GL_CONTEXT_KHR, (cl_context_properties)glXGetCurrentContext(),
CL_GLX_DISPLAY_KHR, (cl_context_properties)glXGetCurrentDisplay(),
CL_CONTEXT_PLATFORM, (cl_context_properties)(platform_)(),
0
};
#endif
static CL_API_ENTRY cl_int (CL_API_CALL
*clGetGLContextInfoKHR)(const cl_context_properties *properties,
cl_gl_context_info param_name,
size_t param_value_size,
void *param_value,
size_t *param_value_size_ret)=NULL;
clGetGLContextInfoKHR = (clGetGLContextInfoKHR_fn) clGetExtensionFunctionAddress("clGetGLContextInfoKHR");
cl_device_id devices[32];
size_t deviceSize = 0;
err = clGetGLContextInfoKHR(properties,
CL_DEVICES_FOR_GL_CONTEXT_KHR,
32 * sizeof(cl_device_id),
devices,
&deviceSize);
if(deviceSize == 0) {
fprintf(stderr, "[OpenCL] Interop not possible\n");
util_abort();
}
cl_bool image_support, available;
size_t max_width, max_height;
cl_uint num_cores, frequency;
cl_device_type _type;
std::string type;
fprintf(verbose, "[OpenCL] Available devices: \n");
for(int i=0; i< (deviceSize / sizeof(cl_device_id)); ++i) {
cl::Device device(devices[i]);
device.getInfo(CL_DEVICE_VENDOR, &name);
device.getInfo(CL_DEVICE_VERSION, &version);
device.getInfo(CL_DEVICE_EXTENSIONS, &extensions);
device.getInfo(CL_DEVICE_AVAILABLE, &available);
device.getInfo(CL_DEVICE_IMAGE_SUPPORT, &image_support);
device.getInfo(CL_DEVICE_IMAGE2D_MAX_WIDTH, &max_width);
device.getInfo(CL_DEVICE_IMAGE2D_MAX_HEIGHT, &max_height);
device.getInfo( CL_DEVICE_MAX_COMPUTE_UNITS , &num_cores);
device.getInfo(CL_DEVICE_MAX_CLOCK_FREQUENCY, &frequency);
device.getInfo(CL_DEVICE_TYPE, &_type);
switch(_type) {
case CL_DEVICE_TYPE_GPU:
type = "GPU";
break;
case CL_DEVICE_TYPE_CPU:
type = "CPU";
break;
case CL_DEVICE_TYPE_ACCELERATOR:
type = "Accelerator";
break;
}
fprintf(verbose, "[OpenCL] Device (%p): %s %s (%s)\n"
" Cores: %u, Frequency: %u MHz, Available: %s,"
" Image support: %s, max size: %lux%lu\n"
" Extensions: %s\n --- \n", (device)(), name.c_str(), version.c_str(), type.c_str(), num_cores, frequency,available?"YES":"NO",image_support?"YES":"NO", max_width, max_height, extensions.c_str());
devices_.push_back(device);
}
fprintf(verbose, "\n-------------------\n");
cl_device_id device_id;
context_ = cl::Context(devices_, properties, &CL::cl_error_callback, nullptr, &err);
if(err != CL_SUCCESS) {
fprintf(stderr, "[OpenCL] Failed to create context: %s\n", errorString(err));
util_abort();
}
err = clGetGLContextInfoKHR(properties, CL_CURRENT_DEVICE_FOR_GL_CONTEXT_KHR, sizeof(device_id), &device_id, NULL);
if(err != CL_SUCCESS) {
fprintf(stderr, "[OpenCL] Failed to get current device for context: %s\n", errorString(err));
util_abort();
}
context_device_ = cl::Device(device_id);
context_device_.getInfo(CL_DEVICE_VENDOR, &name);
context_device_.getInfo(CL_DEVICE_VERSION, &version);
fprintf(verbose, "[OpenCL] Context Device (%p): %s %s\n",(context_device_)(), name.c_str(), version.c_str());
queue_ = cl::CommandQueue(context_, context_device_, 0, &err);
if(err != CL_SUCCESS) {
fprintf(stderr, "[OpenCL] Failed to create a command queue: %s\n", errorString(err));
util_abort();
}
}
bool OpenCLPlatform::Query() {
Check();
// get the name
cl_int err = 0;
name_ = platform_.getInfo<CL_PLATFORM_NAME>(
&err);
if (!CL_CHECK(
err)) {
name_ = "Failed to get platform name.";
}
// get the vendor
vendor_ = platform_.getInfo<CL_PLATFORM_VENDOR>(
&err);
if (!CL_CHECK(
err)) {
vendor_ = "Failed to get platform vendor.";
}
version_ = platform_.getInfo<CL_PLATFORM_VERSION>(
&err);
if (!CL_CHECK(
err)) {
version_ = "Failed to get platform version.";
}
extensions_ = platform_.getInfo<CL_PLATFORM_EXTENSIONS>(
&err);
if (!CL_CHECK(
err)) {
extensions_ = "failed to get platform extensions.";
}
profile_ = platform_.getInfo<CL_PLATFORM_PROFILE>(
&err);
if (!CL_CHECK(
err)) {
profile_ = "Failed to get platform profile.";
}
// CPU & GPU devices
if (!CL_CHECK(
clGetDeviceIDs(platform_(), CL_DEVICE_TYPE_CPU | CL_DEVICE_TYPE_GPU,
0, NULL, &(numDevices)))) {
return false;
}
devicePtr = reinterpret_cast<cl_device_id*>(malloc(
numDevices * sizeof(cl_device_id)));
if (!CL_CHECK(
clGetDeviceIDs(platform_(), CL_DEVICE_TYPE_CPU | CL_DEVICE_TYPE_GPU,
numDevices, devicePtr, NULL))) {
return false;
}
// CPU devices
if (clGetDeviceIDs(
platform_(),
CL_DEVICE_TYPE_CPU,
0,
NULL,
&(numCPUDevices)) == CL_SUCCESS) {
cpuDevicePtr = reinterpret_cast<cl_device_id*>(malloc(
numCPUDevices * sizeof(cl_device_id)));
if (!CL_CHECK(clGetDeviceIDs(platform_(), CL_DEVICE_TYPE_CPU,
numCPUDevices, cpuDevicePtr, NULL))) {
return false;
}
for (int i = 0; i < numCPUDevices; i++) {
OpenCLDevice d(
platform_(),
cpuDevicePtr[i]);
d.query();
devices.push_back(
d);
}
}
// GPU DEVICES
if (!CL_CHECK(clGetDeviceIDs(platform_(), CL_DEVICE_TYPE_GPU,
0, NULL, &(numGPUDevices)))) {
return false;
}
gpuDevicePtr = reinterpret_cast<cl_device_id*>(malloc(
numGPUDevices * sizeof(cl_device_id)));
if (!CL_CHECK(clGetDeviceIDs(platform_(), CL_DEVICE_TYPE_GPU,
numGPUDevices, gpuDevicePtr, NULL))) {
return false;
}
for (int i = 0; i < numGPUDevices; i++) {
OpenCLDevice d(
platform_(),
gpuDevicePtr[i]);
d.query();
devices.push_back(
d);
}
return true;
}
// Check that we have an assigned cl::Platform object.
void OpenCLPlatform::Check() {
if (!platform_()) {
LOG(FATAL)<< "Unassigned platform.";
}
}
cl_platform_id OpenCLPlatform::id() {
return platform_();
}