Thanks to visit codestin.com
Credit goes to github.com

Skip to content

🐛 Fix CUDA for old GPUs without FP16 support #25880

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
merged 3 commits into from
Jul 10, 2024
Merged

🐛 Fix CUDA for old GPUs without FP16 support #25880

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5115dc6
🐛 Fix CUDA for old GPUs without FP16 support
Jamim Jul 8, 2024
cfb2bc3
Added CUDA FP16 availability check for target management.
asmorkalov Jul 9, 2024
cc91789
Update modules/dnn/src/registry.cpp
asmorkalov Jul 10, 2024
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
22 changes: 12 additions & 10 deletions modules/dnn/src/cuda4dnn/init.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@

namespace cv { namespace dnn { namespace cuda4dnn {

void checkVersions()
inline void checkVersions()
{
// https://docs.nvidia.com/deeplearning/cudnn/developer-guide/index.html#programming-model
// cuDNN API Compatibility
Expand Down Expand Up @@ -44,21 +44,23 @@ namespace cv { namespace dnn { namespace cuda4dnn {
}
}

int getDeviceCount()
inline int getDeviceCount()
{
return cuda::getCudaEnabledDeviceCount();
}

int getDevice()
inline int getDevice()
{
int device_id = -1;
CUDA4DNN_CHECK_CUDA(cudaGetDevice(&device_id));
return device_id;
}

bool isDeviceCompatible()
inline bool isDeviceCompatible(int device_id = -1)
{
int device_id = getDevice();
if (device_id < 0)
device_id = getDevice();

if (device_id < 0)
return false;

Expand All @@ -76,9 +78,11 @@ namespace cv { namespace dnn { namespace cuda4dnn {
return false;
}

bool doesDeviceSupportFP16()
inline bool doesDeviceSupportFP16(int device_id = -1)
{
int device_id = getDevice();
if (device_id < 0)
device_id = getDevice();

if (device_id < 0)
return false;

Expand All @@ -87,9 +91,7 @@ namespace cv { namespace dnn { namespace cuda4dnn {
CUDA4DNN_CHECK_CUDA(cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, device_id));

int version = major * 10 + minor;
if (version < 53)
return false;
return true;
return (version >= 53);
}

}}} /* namespace cv::dnn::cuda4dnn */
Expand Down
14 changes: 14 additions & 0 deletions modules/dnn/src/net_impl_backend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,10 @@
#include "backend.hpp"
#include "factory.hpp"

#ifdef HAVE_CUDA
#include "cuda4dnn/init.hpp"
#endif

namespace cv {
namespace dnn {
CV__DNN_INLINE_NS_BEGIN
Expand Down Expand Up @@ -242,6 +246,16 @@ void Net::Impl::setPreferableTarget(int targetId)
#endif
}

if (IS_DNN_CUDA_TARGET(targetId))
{
preferableTarget = DNN_TARGET_CPU;
#ifdef HAVE_CUDA
if (cuda4dnn::doesDeviceSupportFP16() && targetId == DNN_TARGET_CUDA_FP16)
preferableTarget = DNN_TARGET_CUDA_FP16;
else
preferableTarget = DNN_TARGET_CUDA;
#endif
}
#if !defined(__arm64__) || !__arm64__
if (targetId == DNN_TARGET_CPU_FP16)
{
Expand Down
26 changes: 24 additions & 2 deletions modules/dnn/src/registry.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,10 @@
#include "backend.hpp"
#include "factory.hpp"

#ifdef HAVE_CUDA
#include "cuda4dnn/init.hpp"
#endif

namespace cv {
namespace dnn {
CV__DNN_INLINE_NS_BEGIN
Expand Down Expand Up @@ -118,10 +122,28 @@ class BackendRegistry
#endif

#ifdef HAVE_CUDA
if (haveCUDA())
cuda4dnn::checkVersions();

bool hasCudaCompatible = false;
bool hasCudaFP16 = false;
for (int i = 0; i < cuda4dnn::getDeviceCount(); i++)
{
if (cuda4dnn::isDeviceCompatible(i))
{
hasCudaCompatible = true;
if (cuda4dnn::doesDeviceSupportFP16(i))
{
hasCudaFP16 = true;
break; // we already have all we need here
}
}
}

if (hasCudaCompatible)
{
backends.push_back(std::make_pair(DNN_BACKEND_CUDA, DNN_TARGET_CUDA));
backends.push_back(std::make_pair(DNN_BACKEND_CUDA, DNN_TARGET_CUDA_FP16));
if (hasCudaFP16)
backends.push_back(std::make_pair(DNN_BACKEND_CUDA, DNN_TARGET_CUDA_FP16));
}
#endif

Expand Down
2 changes: 1 addition & 1 deletion modules/dnn/test/test_common.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -211,7 +211,7 @@ class DNNTestLayer : public TestWithParam<tuple<Backend, Target> >
if ((!l->supportBackend(backend) || l->preferableTarget != target) && !fused)
{
hasFallbacks = true;
std::cout << "FALLBACK: Layer [" << l->type << "]:[" << l->name << "] is expected to has backend implementation" << endl;
std::cout << "FALLBACK: Layer [" << l->type << "]:[" << l->name << "] is expected to have backend implementation" << endl;
}
}
if (hasFallbacks && raiseError)
Expand Down
2 changes: 1 addition & 1 deletion modules/dnn/test/test_onnx_conformance.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1008,7 +1008,7 @@ class Test_ONNX_conformance : public TestWithParam<ONNXConfParams>
if ((!l->supportBackend(backend) || l->preferableTarget != target) && !fused)
{
hasFallbacks = true;
std::cout << "FALLBACK: Layer [" << l->type << "]:[" << l->name << "] is expected to has backend implementation" << endl;
std::cout << "FALLBACK: Layer [" << l->type << "]:[" << l->name << "] is expected to have backend implementation" << endl;
}
}
return hasFallbacks;
Expand Down