提交 a1e7f2d5 编写于 作者: C chenweihang

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into unsqueeze_op

1 合并请求!11897Add Unsqueeze operator and unit testing
...@@ -4,7 +4,6 @@ ...@@ -4,7 +4,6 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) [![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) [![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
[![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases) [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......
...@@ -96,6 +96,20 @@ if(NOT APPLE AND NOT ANDROID) ...@@ -96,6 +96,20 @@ if(NOT APPLE AND NOT ANDROID)
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt") set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
endif(NOT APPLE AND NOT ANDROID) endif(NOT APPLE AND NOT ANDROID)
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
# for building inference libs
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
function(merge_static_libs TARGET_NAME) function(merge_static_libs TARGET_NAME)
set(libs ${ARGN}) set(libs ${ARGN})
list(REMOVE_DUPLICATES libs) list(REMOVE_DUPLICATES libs)
......
...@@ -12,19 +12,6 @@ ...@@ -12,19 +12,6 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
# make package for paddle fluid shared and static library # make package for paddle fluid shared and static library
function(copy TARGET) function(copy TARGET)
set(options "") set(options "")
......
...@@ -147,9 +147,9 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, ...@@ -147,9 +147,9 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
"Input tensor type is not supported: ", in.type().name()); "Input tensor type is not supported: ", in.type().name());
memory::data_type out_type = in_type; memory::data_type out_type = in_type;
auto in_format = MKLDNNFormatForSize(in_tz.size(), in.format()); auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
auto out_format = auto out_format =
MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout)); platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
void* in_data = GetDataFromTensor(in, in_type); void* in_data = GetDataFromTensor(in, in_type);
......
...@@ -62,12 +62,6 @@ inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) { ...@@ -62,12 +62,6 @@ inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
return MKLDNNDataType::data_undef; return MKLDNNDataType::data_undef;
} }
inline MKLDNNFormat MKLDNNFormatForSize(size_t dims_size,
MKLDNNFormat default_format) {
return (dims_size == 1
? mkldnn::memory::format::x
: dims_size == 2 ? mkldnn::memory::format::nc : default_format);
}
#endif #endif
void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
......
...@@ -18,6 +18,10 @@ limitations under the License. */ ...@@ -18,6 +18,10 @@ limitations under the License. */
#include "paddle/fluid/framework/data_layout_transform.h" #include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/data_type_transform.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle { namespace paddle {
namespace framework { namespace framework {
...@@ -48,8 +52,8 @@ void TransformData(const OpKernelType &expected_kernel_type, ...@@ -48,8 +52,8 @@ void TransformData(const OpKernelType &expected_kernel_type,
// Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel // Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
// Just set layout/format. No real transform occur // Just set layout/format. No real transform occur
auto out_format = auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
MKLDNNFormatForSize(in.dims().size(), ToMKLDNNFormat(lin)); ToMKLDNNFormat(lin));
out.ShareDataWith(input_tensor); out.ShareDataWith(input_tensor);
out.set_layout(DataLayout::kMKLDNN); out.set_layout(DataLayout::kMKLDNN);
......
...@@ -73,18 +73,12 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, ...@@ -73,18 +73,12 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
stream); stream);
} else { } else {
// NOTE(zcd): Because TensorCopy is an async operation, when the src_place
// and dst_place are two different GPU, to ensure that the operation can
// be carried out correctly, we should make ctx wait.
// If ctx_place and src_place are the same, we should add ctx.Wait()
// after memory::Copy; if ctx_place and dst_place are the same, we should
// add ctx.Wait() before memory::Copy.
if (platform::is_same_place(ctx_place, src_place)) { if (platform::is_same_place(ctx_place, src_place)) {
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
stream); stream);
ctx.Wait(); platform::DeviceContextPool::Instance().Get(src.place())->Wait();
} else if (platform::is_same_place(ctx_place, dst_place)) { } else if (platform::is_same_place(ctx_place, dst_place)) {
ctx.Wait(); platform::DeviceContextPool::Instance().Get(src.place())->Wait();
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
stream); stream);
} else { } else {
...@@ -97,13 +91,6 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, ...@@ -97,13 +91,6 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
void TensorCopy(const Tensor& src, const platform::Place& dst_place, void TensorCopy(const Tensor& src, const platform::Place& dst_place,
Tensor* dst) { Tensor* dst) {
// NOTE(zcd): If the src.place() and dst_place are two different GPU,
// the copy operation is carried out on the dst_place's stream. This is
// very important, because TensorCopy is an async operator, and in most
// case, once this copy operator returns, dst is to be used in dst_place's
// stream, if this copy operation is carried out on the src_place's stream,
// when dst is used in dst_place's stream the copy operation may be
// not completed.
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
const platform::DeviceContext* dev_ctx; const platform::DeviceContext* dev_ctx;
if (platform::is_gpu_place(dst_place)) { if (platform::is_gpu_place(dst_place)) {
......
...@@ -23,10 +23,25 @@ limitations under the License. */ ...@@ -23,10 +23,25 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace framework { namespace framework {
// NOTE(zcd): Because TensorCopy is an async operation, when the src_place
// and dst_place are two different GPU, to ensure that the operation can
// be carried out correctly, there is a src_ctx wait operation in TensorCopy.
// If ctx_place and src_place are the same, src_ctx.Wait() is added
// after memory::Copy; if ctx_place and dst_place are the same,
// src_ctx.Wait() is added before memory::Copy.
void TensorCopy(const Tensor& src, const platform::Place& dst_place, void TensorCopy(const Tensor& src, const platform::Place& dst_place,
const platform::DeviceContext& ctx, Tensor* dst); const platform::DeviceContext& ctx, Tensor* dst);
// NOTE(zcd): If the src.place() and dst_place are two different GPU,
// the copy operation is carried out on the dst_place's stream. This is
// very important, because TensorCopy is an async operator, and in most
// case, once this copy operator returns, dst is to be used in dst_place's
// stream, if this copy operation is carried out on the src_place's stream,
// when dst is used in dst_place's stream the copy operation may be
// not completed.
void TensorCopy(const Tensor& src, const platform::Place& dst_place, void TensorCopy(const Tensor& src, const platform::Place& dst_place,
Tensor* dst); Tensor* dst);
void TensorCopySync(const Tensor& src, const platform::Place& dst_place, void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
Tensor* dst); Tensor* dst);
......
...@@ -115,9 +115,12 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -115,9 +115,12 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
if (fuse_with_relu) flags |= mkldnn::fuse_bn_relu; if (fuse_with_relu) flags |= mkldnn::fuse_bn_relu;
// create mkldnn memory from input x tensor // create mkldnn memory from input x tensor
auto src_memory = mkldnn::memory::format input_format =
memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine}, platform::MKLDNNFormatForSize(src_tz.size(), x->format());
to_void_cast(x_data));
auto src_memory = memory(
{{{src_tz}, memory::data_type::f32, input_format}, mkldnn_engine},
to_void_cast(x_data));
// create primitive descriptor for batch norm forward // create primitive descriptor for batch norm forward
using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>; using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
...@@ -251,15 +254,21 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> { ...@@ -251,15 +254,21 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>; using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>;
// create mkldnn memory from input diff_y tensor // create mkldnn memory from input diff_y tensor
auto user_diff_dst_memory =
memory({{{diff_dst_tz}, memory::data_type::f32, diff_y->format()}, mkldnn::memory::format dst_format =
mkldnn_engine}, platform::MKLDNNFormatForSize(src_tz.size(), diff_y->format());
to_void_cast(diff_y_data));
auto user_diff_dst_memory = memory(
{{{diff_dst_tz}, memory::data_type::f32, dst_format}, mkldnn_engine},
to_void_cast(diff_y_data));
// create mkldnn memory from input x tensor // create mkldnn memory from input x tensor
auto src_memory = mkldnn::memory::format input_format =
memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine}, platform::MKLDNNFormatForSize(src_tz.size(), x->format());
to_void_cast(x_data));
auto src_memory = memory(
{{{src_tz}, memory::data_type::f32, input_format}, mkldnn_engine},
to_void_cast(x_data));
// for diff_dst, try to use same format as dst in forward pass // for diff_dst, try to use same format as dst in forward pass
auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc(); auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc();
......
...@@ -228,7 +228,7 @@ class MKLDNNHandler { ...@@ -228,7 +228,7 @@ class MKLDNNHandler {
return dstr; return dstr;
}; };
return dims2str(operand_dims) + suffix; return dims2str(operand_dims) + suffix;
}; }
protected: protected:
const MKLDNNDeviceContext& dev_ctx_; const MKLDNNDeviceContext& dev_ctx_;
...@@ -237,5 +237,15 @@ class MKLDNNHandler { ...@@ -237,5 +237,15 @@ class MKLDNNHandler {
bool is_reusing_; bool is_reusing_;
}; };
inline mkldnn::memory::format MKLDNNFormatForSize(
size_t dims_size, mkldnn::memory::format data_format) {
if (dims_size == 1) {
return mkldnn::memory::format::x;
} else if (dims_size == 2) {
return mkldnn::memory::format::nc;
}
return data_format;
}
} // namespace platform } // namespace platform
} // namespace paddle } // namespace paddle
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册