Skip to content

Add float16 support to non-cudnn softmax op on GPU #9686

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
Apr 6, 2018
Merged

Add float16 support to non-cudnn softmax op on GPU #9686

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
aa82bf8
initial commit
kexinzhao Apr 4, 2018
7491c3e
fix error
kexinzhao Apr 4, 2018
4c73ad3
fix typo and order
kexinzhao Apr 4, 2018
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
3 changes: 3 additions & 0 deletions paddle/fluid/operators/math/softmax.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,8 @@ limitations under the License. */

#define EIGEN_USE_GPU

#include <vector>

#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/softmax.h"
#include "paddle/fluid/operators/math/softmax_impl.h"
Expand Down Expand Up @@ -95,6 +97,7 @@ template class SoftmaxCUDNNFunctor<double>;
template class SoftmaxGradCUDNNFunctor<float>;
template class SoftmaxGradCUDNNFunctor<double>;

template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16>;
template class SoftmaxFunctor<platform::CUDADeviceContext, float>;
template class SoftmaxFunctor<platform::CUDADeviceContext, double>;
template class SoftmaxGradFunctor<platform::CUDADeviceContext, float>;
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/operators/math/softmax_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,7 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename T>
struct ValueClip {
HOSTDEVICE T operator()(const T& x) const {
const T kThreshold = -64.;
const T kThreshold = static_cast<T>(-64.);
Copy link
Collaborator

Choose a reason for hiding this comment

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

Do we need a comment here about the criteria for choosing the value of the threshold?

return x < kThreshold ? kThreshold : x;
}
};
Expand Down
9 changes: 7 additions & 2 deletions paddle/fluid/operators/softmax_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,13 +13,17 @@ See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/softmax_op.h"

#include <string>

#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cudnn_helper.h"
#endif

#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

namespace paddle {
namespace operators {

Expand Down Expand Up @@ -60,8 +64,8 @@ class SoftmaxOp : public framework::OperatorWithKernel {
auto input_data_type =
framework::ToDataType(ctx.Input<Tensor>("X")->type());
if (input_data_type == framework::proto::VarType::FP16) {
PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN,
"float16 can only be used when CUDNN is used");
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"float16 can only be used on GPU place");
}

std::string data_format = ctx.Attr<std::string>("data_format");
Expand All @@ -70,6 +74,7 @@ class SoftmaxOp : public framework::OperatorWithKernel {
library_);
}
};

class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
Expand Down
11 changes: 6 additions & 5 deletions paddle/fluid/operators/softmax_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,11 +13,12 @@ See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/softmax_op.h"
#include "paddle/fluid/platform/float16.h"

namespace ops = paddle::operators;

REGISTER_OP_CUDA_KERNEL(
softmax, ops::SoftmaxKernel<paddle::platform::CUDADeviceContext, float>);
namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(
softmax_grad,
ops::SoftmaxGradKernel<paddle::platform::CUDADeviceContext, float>);
softmax, ops::SoftmaxKernel<plat::CUDADeviceContext, float>,
ops::SoftmaxKernel<plat::CUDADeviceContext, plat::float16>);
REGISTER_OP_CUDA_KERNEL(softmax_grad,
ops::SoftmaxGradKernel<plat::CUDADeviceContext, float>);
Loading