Skip to content
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

Enable tensor core for cudnn conv #9623

Merged
merged 2 commits into from
Apr 4, 2018
Merged

Enable tensor core for cudnn conv #9623

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
187ba08
enable tensor core for conv cudnn
kexinzhao Apr 4, 2018
9ba3660
fix cpplint error
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
22 changes: 22 additions & 0 deletions paddle/fluid/operators/conv_cudnn_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,10 +128,32 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
workspace_size_limit, &algo));

#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
// Tensor core is supported since the volta GPU and
// is only enabled when input and filter data are float16
if (dev_ctx.GetComputeCapability() >= 70 &&
std::type_index(typeid(T)) ==
std::type_index(typeid(platform::float16))) {
PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
cudnn_conv_desc, CUDNN_TENSOR_OP_MATH));
// Currently tensor core is only enabled using this algo
algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
} else {
PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
cudnn_conv_desc, CUDNN_DEFAULT_MATH));
}
#endif

// get workspace size able to allocate
PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
cudnn_output_desc, algo, &workspace_size_in_bytes));
// It is possible for float16 on Volta GPU to allocate more memory than
// the limit because the algo is overrided to use tensor core.
PADDLE_ENFORCE_LE(workspace_size_in_bytes, workspace_size_limit,
"workspace_size to be allocated exceeds the limit");

// Allocate on GPU memory
platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
Expand Down
4 changes: 3 additions & 1 deletion paddle/fluid/platform/cudnn_helper.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -257,9 +257,11 @@ class ScopedConvolutionDescriptor {
}
#endif

cudnnDataType_t compute_type =
(type == CUDNN_DATA_DOUBLE) ? CUDNN_DATA_DOUBLE : CUDNN_DATA_FLOAT;
PADDLE_ENFORCE(dynload::cudnnSetConvolutionNdDescriptor(
desc_, pads.size(), pads.data(), strides.data(), dilations.data(),
CUDNN_CROSS_CORRELATION, type));
CUDNN_CROSS_CORRELATION, compute_type));
return desc_;
}

Expand Down
5 changes: 3 additions & 2 deletions paddle/fluid/platform/dynload/cudnn.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ limitations under the License. */

#include <cudnn.h>
#include <dlfcn.h>
#include <mutex>
#include <mutex> // NOLINT
#include "paddle/fluid/platform/dynload/dynamic_loader.h"

namespace paddle {
Expand Down Expand Up @@ -140,7 +140,8 @@ CUDNN_DNN_ROUTINE_EACH_R5(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)

#if CUDNN_VERSION >= 7001
#define CUDNN_DNN_ROUTINE_EACH_R7(__macro) \
__macro(cudnnSetConvolutionGroupCount);
__macro(cudnnSetConvolutionGroupCount); \
__macro(cudnnSetConvolutionMathType);
CUDNN_DNN_ROUTINE_EACH_R7(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
#endif

Expand Down