Skip to content

CL_DEVICE_HALF_FP_CONFIG returns CL_INVALID_VALUE #3068

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
umar456 merged 1 commit into from
Feb 18, 2021
Merged

CL_DEVICE_HALF_FP_CONFIG returns CL_INVALID_VALUE #3068

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 15 additions & 23 deletions src/backend/opencl/platform.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,6 +184,7 @@ string getDeviceInfo() noexcept {
nDevices++;
}
} catch (const AfError& err) {
UNUSED(err);
info << "No platforms found.\n";
// Don't throw an exception here. Info should pass even if the system
// doesn't have the correct drivers installed.
Expand Down Expand Up @@ -215,8 +216,9 @@ int getDeviceCount() noexcept try {
DeviceManager& devMngr = DeviceManager::getInstance();

common::lock_guard_t lock(devMngr.deviceMutex);
return devMngr.mQueues.size();
return static_cast<int>(devMngr.mQueues.size());
} catch (const AfError& err) {
UNUSED(err);
// If device manager threw an error then return 0 because no platforms
// were found
return 0;
Expand All @@ -233,7 +235,7 @@ int getDeviceIdFromNativeId(cl_device_id id) {

common::lock_guard_t lock(devMngr.deviceMutex);

int nDevices = devMngr.mDevices.size();
int nDevices = static_cast<int>(devMngr.mDevices.size());
int devId = 0;
for (devId = 0; devId < nDevices; ++devId) {
if (id == devMngr.mDevices[devId]->operator()()) { break; }
Expand Down Expand Up @@ -359,8 +361,9 @@ bool isDoubleSupported(unsigned device) {
common::lock_guard_t lock(devMngr.deviceMutex);
dev = *devMngr.mDevices[device];
}

return (dev.getInfo<CL_DEVICE_DOUBLE_FP_CONFIG>() > 0);
// 64bit fp is an optional extension
return (dev.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_fp64") !=
string::npos);

@9prady9 9prady9 Dec 18, 2020

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

👍🏾

Although I wonder if the vendor implementation caches the device type support in internal cache. If cached in such manner, then fetching that compared to doing a string search is more efficient.

@willyborn willyborn Dec 18, 2020

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I will get through the debugger step by step and report what I find.

@willyborn willyborn Dec 18, 2020

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have the impression that it is straight copy into a vector.
For NVIDIA

  • Total length of vector: 453 chars
  • cl_khr_fp16 NA
  • cl_khr_fp64 at pos 138
    For AMD
  • Total length of vector: 661 chars
  • cl_khr_fp16 NA
  • cl_khr_fp64 at pos 0

The complete string is cached in the driver. Copy happens in cl2.hpp at line 1427.
Although I can not see inside the function, I come to this conclusion because the exact length is provided upfront during the construction of the vector.

}

bool isHalfSupported(unsigned device) {
Expand All @@ -371,21 +374,9 @@ bool isHalfSupported(unsigned device) {
common::lock_guard_t lock(devMngr.deviceMutex);
dev = *devMngr.mDevices[device];
}
cl_device_fp_config config = 0;
size_t ret_size = 0;
// NVIDIA OpenCL seems to return error codes for CL_DEVICE_HALF_FP_CONFIG.
// It seems to be a bug in their implementation. Assuming if this function
// fails that the implemenation does not support f16 type. Using the C API
// to avoid exceptions
cl_int err =
clGetDeviceInfo(dev(), CL_DEVICE_HALF_FP_CONFIG,
sizeof(cl_device_fp_config), &config, &ret_size);

if (err) {
return false;
} else {
return config > 0;
}
// 16bit fp is an option extension
return (dev.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_fp16") !=
string::npos);
}

void devprop(char* d_name, char* d_platform, char* d_toolkit, char* d_compute) {
Expand Down Expand Up @@ -481,12 +472,13 @@ void addDeviceContext(cl_device_id dev, cl_context ctx, cl_command_queue que) {
devMngr.mPlatforms.push_back(getPlatformEnum(*tDevice));
// FIXME: add OpenGL Interop for user provided contexts later
devMngr.mIsGLSharingOn.push_back(false);
devMngr.mDeviceTypes.push_back(tDevice->getInfo<CL_DEVICE_TYPE>());
devMngr.mDeviceTypes.push_back(
static_cast<int>(tDevice->getInfo<CL_DEVICE_TYPE>()));

devMngr.mDevices.push_back(move(tDevice));
devMngr.mContexts.push_back(move(tContext));
devMngr.mQueues.push_back(move(tQueue));
nDevices = devMngr.mDevices.size() - 1;
nDevices = static_cast<int>(devMngr.mDevices.size()) - 1;

// cache the boost program_cache object, clean up done on program exit
// not during removeDeviceContext
Expand All @@ -507,7 +499,7 @@ void setDeviceContext(cl_device_id dev, cl_context ctx) {

common::lock_guard_t lock(devMngr.deviceMutex);

const int dCount = devMngr.mDevices.size();
const int dCount = static_cast<int>(devMngr.mDevices.size());
for (int i = 0; i < dCount; ++i) {
Comment thread
9prady9 marked this conversation as resolved.
if (devMngr.mDevices[i]->operator()() == dev &&
devMngr.mContexts[i]->operator()() == ctx) {
Expand All @@ -529,7 +521,7 @@ void removeDeviceContext(cl_device_id dev, cl_context ctx) {
{
common::lock_guard_t lock(devMngr.deviceMutex);

const int dCount = devMngr.mDevices.size();
const int dCount = static_cast<int>(devMngr.mDevices.size());
for (int i = 0; i < dCount; ++i) {
Comment thread
9prady9 marked this conversation as resolved.
if (devMngr.mDevices[i]->operator()() == dev &&
devMngr.mContexts[i]->operator()() == ctx) {
Expand Down