Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 d3e60959 编写于 作者: C Chen Weihang 提交者： GitHub
浏览文件
操作

[Cherry-pick] The Second part of new custom op extension in 2.0.1 (#31237) 

[Cherry-pick] The Second part of new custom op extension in 2.0.1
上级 34092ab3
显示空白变更内容
内联 并排
Showing with 1705 addition and 504 deletion
CMakeLists.txt
浏览文件 @ d3e60959
...@@ -293,6 +293,8 @@ set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build") ...@@ -293,6 +293,8 @@ set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG") set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
set(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG") set(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
add_definitions(-DPADDLE_DLL_EXPORT)
if(ON_INFER) if(ON_INFER)
# you can trun off the paddle fluid and inference lib by set ON_INFER=OFF # you can trun off the paddle fluid and inference lib by set ON_INFER=OFF
message(STATUS "On inference mode, will take place some specific optimization.") message(STATUS "On inference mode, will take place some specific optimization.")
......
cmake/generic.cmake
浏览文件 @ d3e60959
...@@ -793,16 +793,14 @@ function(py_test TARGET_NAME) ...@@ -793,16 +793,14 @@ function(py_test TARGET_NAME)
if(WITH_COVERAGE) if(WITH_COVERAGE)
add_test(NAME ${TARGET_NAME} add_test(NAME ${TARGET_NAME}
COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
FLAGS_cpu_deterministic=true FLAGS_cpu_deterministic=true ${py_test_ENVS}
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
COVERAGE_FILE=${PADDLE_BINARY_DIR}/python-coverage.data COVERAGE_FILE=${PADDLE_BINARY_DIR}/python-coverage.data
${PYTHON_EXECUTABLE} -m coverage run --branch -p ${py_test_SRCS} ${py_test_ARGS} ${PYTHON_EXECUTABLE} -m coverage run --branch -p ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
else() else()
add_test(NAME ${TARGET_NAME} add_test(NAME ${TARGET_NAME}
COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
FLAGS_cpu_deterministic=true FLAGS_cpu_deterministic=true ${py_test_ENVS}
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS} ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif() endif()
......
paddle/fluid/extension/include/all.h
浏览文件 @ d3e60959
...@@ -18,6 +18,12 @@ limitations under the License. */ ...@@ -18,6 +18,12 @@ limitations under the License. */
#error C++11 or later compatible compiler is required to use Paddle. #error C++11 or later compatible compiler is required to use Paddle.
#endif #endif
#ifdef _WIN32
#ifndef NOMINMAX
#define NOMINMAX // msvc max/min macro conflict with std::min/max
#endif
#endif
#include "paddle/fluid/extension/include/dispatch.h" #include "paddle/fluid/extension/include/dispatch.h"
#include "paddle/fluid/extension/include/dtype.h" #include "paddle/fluid/extension/include/dtype.h"
#include "paddle/fluid/extension/include/op_meta_info.h" #include "paddle/fluid/extension/include/op_meta_info.h"
......
paddle/fluid/extension/include/dll_decl.h 0 → 100644
浏览文件 @ d3e60959
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#if defined(_WIN32)
#ifndef PD_DLL_DECL
#ifdef PADDLE_DLL_EXPORT
#define PD_DLL_DECL __declspec(dllexport)
#else
#define PD_DLL_DECL __declspec(dllimport)
#endif // PADDLE_DLL_EXPORT
#endif // PD_DLL_DECL
#else
#define PD_DLL_DECL
#endif // _WIN32
paddle/fluid/extension/include/op_meta_info.h
浏览文件 @ d3e60959
...@@ -14,12 +14,14 @@ limitations under the License. */ ...@@ -14,12 +14,14 @@ limitations under the License. */
#pragma once #pragma once
#include <iostream>
#include <string> #include <string>
#include <unordered_map> #include <unordered_map>
#include <vector> #include <vector>
#include <boost/any.hpp> #include <boost/any.hpp>
#include "paddle/fluid/extension/include/dll_decl.h"
#include "paddle/fluid/extension/include/tensor.h" #include "paddle/fluid/extension/include/tensor.h"
/** /**
...@@ -31,7 +33,7 @@ limitations under the License. */ ...@@ -31,7 +33,7 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace framework { namespace framework {
class OpMetaInfoHelper; class PD_DLL_DECL OpMetaInfoHelper;
} // namespace framework } // namespace framework
using Tensor = paddle::Tensor; using Tensor = paddle::Tensor;
...@@ -43,6 +45,26 @@ using Tensor = paddle::Tensor; ...@@ -43,6 +45,26 @@ using Tensor = paddle::Tensor;
classname& operator=(const classname&) = delete; \ classname& operator=(const classname&) = delete; \
classname& operator=(classname&&) = delete classname& operator=(classname&&) = delete
#if defined _WIN32
#define HANDLE_THE_ERROR try {
#define END_HANDLE_THE_ERROR \
} \
catch (const std::exception& e) { \
std::cerr << e.what() << std::endl; \
throw e; \
}
#else
#define HANDLE_THE_ERROR
#define END_HANDLE_THE_ERROR
#endif
#define PD_THROW(err_msg) \
do { \
HANDLE_THE_ERROR \
throw std::runtime_error(err_msg); \
END_HANDLE_THE_ERROR \
} while (0)
///////////////// Util Define and Function //////////////// ///////////////// Util Define and Function ////////////////
inline std::string Grad(const std::string& var_name) { inline std::string Grad(const std::string& var_name) {
...@@ -59,6 +81,26 @@ inline std::string Grad(const std::string& var_name) { ...@@ -59,6 +81,26 @@ inline std::string Grad(const std::string& var_name) {
using KernelFunc = std::vector<Tensor> (*)(std::vector<Tensor> inputs, using KernelFunc = std::vector<Tensor> (*)(std::vector<Tensor> inputs,
std::vector<boost::any> attrs); std::vector<boost::any> attrs);
#define PD_SPECIALIZE_ComputeCallHelper(attr_type) \
template <typename... Tail> \
struct ComputeCallHelper<attr_type, Tail...> { \
template <int in_idx, int attr_idx, typename... PreviousArgs> \
static Return Compute(std::vector<Tensor> inputs, \
std::vector<boost::any> attrs, \
const PreviousArgs&... pargs) { \
try { \
attr_type arg = boost::any_cast<attr_type>(attrs[attr_idx]); \
return ComputeCallHelper<Tail...>::template Compute<in_idx, \
attr_idx + 1>( \
inputs, attrs, pargs..., arg); \
} catch (boost::bad_any_cast&) { \
PD_THROW( \
"Attribute cast error in custom operator. Expected " #attr_type \
" value."); \
} \
} \
}
template <typename T> template <typename T>
struct TypeTag {}; struct TypeTag {};
...@@ -92,26 +134,20 @@ struct KernelFuncImpl<Return (*)(Args...), impl_fn> { ...@@ -92,26 +134,20 @@ struct KernelFuncImpl<Return (*)(Args...), impl_fn> {
} }
}; };
// TODO(chenweihang): add support for attribute input PD_SPECIALIZE_ComputeCallHelper(bool);
// int attribute input (not used now) PD_SPECIALIZE_ComputeCallHelper(int);
template <typename... Tail> PD_SPECIALIZE_ComputeCallHelper(float);
struct ComputeCallHelper<int, Tail...> { PD_SPECIALIZE_ComputeCallHelper(int64_t);
template <int in_idx, int attr_idx, typename... PreviousArgs> PD_SPECIALIZE_ComputeCallHelper(std::string);
static Return Compute(std::vector<Tensor> inputs, PD_SPECIALIZE_ComputeCallHelper(std::vector<int>);
std::vector<boost::any> attrs, PD_SPECIALIZE_ComputeCallHelper(std::vector<float>);
const PreviousArgs&... pargs) { PD_SPECIALIZE_ComputeCallHelper(std::vector<int64_t>);
try { PD_SPECIALIZE_ComputeCallHelper(std::vector<std::string>);
int arg = boost::any_cast<int>(attrs[attr_idx]); // TODO(chenweihang): support other attribute type if needed.
return ComputeCallHelper<Tail...>::template Compute<in_idx, // Why not support other attribute type here?
attr_idx + 1>( // - boost::blank, std::vector<bool> and std::vector<double>
inputs, attrs, pargs..., arg); // are not used in op
} catch (boost::bad_any_cast&) { // - BlockDesc* and std::vector<BlockDesc*> are used in framework
throw std::runtime_error(
"Attribute cast error in custom operator. Expected int value.");
}
}
};
// end: base template // end: base template
template <typename T> template <typename T>
struct ComputeCallHelper<TypeTag<T>> { struct ComputeCallHelper<TypeTag<T>> {
...@@ -220,13 +256,26 @@ struct InferDtypeFuncImpl<Return (*)(Args...), impl_fn> { ...@@ -220,13 +256,26 @@ struct InferDtypeFuncImpl<Return (*)(Args...), impl_fn> {
////////////////////// Op Meta Info ////////////////////// ////////////////////// Op Meta Info //////////////////////
class OpMetaInfo { class PD_DLL_DECL OpMetaInfo {
public: public:
explicit OpMetaInfo(const std::string& op_name) : name_(op_name) {} explicit OpMetaInfo(const std::string& op_name) : name_(op_name) {}
// format: {"<name1>", "<name2>", ...}
OpMetaInfo& Inputs(std::vector<std::string>&& inputs); OpMetaInfo& Inputs(std::vector<std::string>&& inputs);
// format: {"<name1>", "<name2>", ...}
OpMetaInfo& Outputs(std::vector<std::string>&& outputs); OpMetaInfo& Outputs(std::vector<std::string>&& outputs);
// format: {"<name1>:<type1>", "<name1>:<type1>", ...}
OpMetaInfo& Attrs(std::vector<std::string>&& attrs);
// format: PD_KERNEL(...)
OpMetaInfo& SetKernelFn(KernelFunc&& func); OpMetaInfo& SetKernelFn(KernelFunc&& func);
// format: PD_INFER_SHAPE(...)
OpMetaInfo& SetInferShapeFn(InferShapeFunc&& func); OpMetaInfo& SetInferShapeFn(InferShapeFunc&& func);
// format: PD_INFER_DTYPE(...)
OpMetaInfo& SetInferDtypeFn(InferDtypeFunc&& func); OpMetaInfo& SetInferDtypeFn(InferDtypeFunc&& func);
private: private:
...@@ -246,7 +295,7 @@ class OpMetaInfo { ...@@ -246,7 +295,7 @@ class OpMetaInfo {
//////////////// Op Meta Info Map ///////////////// //////////////// Op Meta Info Map /////////////////
class OpMetaInfoMap { class PD_DLL_DECL OpMetaInfoMap {
public: public:
// this function's impl should keep in header file. // this function's impl should keep in header file.
// if move to cc file, meta info can not be added // if move to cc file, meta info can not be added
...@@ -270,14 +319,15 @@ class OpMetaInfoMap { ...@@ -270,14 +319,15 @@ class OpMetaInfoMap {
//////////////// Op Meta Info Builder ///////////////// //////////////// Op Meta Info Builder /////////////////
class OpMetaInfoBuilder { class PD_DLL_DECL OpMetaInfoBuilder {
public: public:
explicit OpMetaInfoBuilder(std::string&& name); explicit OpMetaInfoBuilder(std::string&& name);
OpMetaInfoBuilder& Inputs(std::vector<std::string>&& inputs); OpMetaInfoBuilder& Inputs(std::vector<std::string>&& inputs);
OpMetaInfoBuilder& Outputs(std::vector<std::string>&& outputs); OpMetaInfoBuilder& Outputs(std::vector<std::string>&& outputs);
OpMetaInfoBuilder& SetKernelFn(KernelFunc&& func); OpMetaInfoBuilder& Attrs(std::vector<std::string>&& attrs);
OpMetaInfoBuilder& SetInferShapeFn(InferShapeFunc&& func); OpMetaInfoBuilder& SetKernelFn(KernelFunc func);
OpMetaInfoBuilder& SetInferDtypeFn(InferDtypeFunc&& func); OpMetaInfoBuilder& SetInferShapeFn(InferShapeFunc func);
OpMetaInfoBuilder& SetInferDtypeFn(InferDtypeFunc func);
OpMetaInfoBuilder& SetBackwardOp(const std::string& bwd_op_name); OpMetaInfoBuilder& SetBackwardOp(const std::string& bwd_op_name);
private: private:
...@@ -317,8 +367,12 @@ void LoadCustomOperatorLib(const std::string& dso_name); ...@@ -317,8 +367,12 @@ void LoadCustomOperatorLib(const std::string& dso_name);
extern "C" { extern "C" {
#endif #endif
#if defined(_WIN32)
// C-API to get global OpMetaInfoMap. // C-API to get global OpMetaInfoMap.
paddle::OpMetaInfoMap& PD_GetOpMetaInfoMap(); __declspec(dllexport) inline paddle::OpMetaInfoMap& PD_GetOpMetaInfoMap() {
return paddle::OpMetaInfoMap::Instance();
}
#endif // _WIN32
#ifdef __cplusplus #ifdef __cplusplus
} }
......
paddle/fluid/extension/include/tensor.h
浏览文件 @ d3e60959
...@@ -16,6 +16,7 @@ limitations under the License. */ ...@@ -16,6 +16,7 @@ limitations under the License. */
#include <memory> #include <memory>
#include <vector> #include <vector>
#include "paddle/fluid/extension/include/dll_decl.h"
#include "paddle/fluid/extension/include/dtype.h" #include "paddle/fluid/extension/include/dtype.h"
#include "paddle/fluid/extension/include/place.h" #include "paddle/fluid/extension/include/place.h"
...@@ -23,7 +24,7 @@ namespace paddle { ...@@ -23,7 +24,7 @@ namespace paddle {
namespace framework { namespace framework {
class CustomTensorUtils; class CustomTensorUtils;
} // namespace framework } // namespace framework
class Tensor { class PD_DLL_DECL Tensor {
public: public:
/// \brief Construct a Tensor on target Place for CustomOp. /// \brief Construct a Tensor on target Place for CustomOp.
/// Generally it's only used for user to create Tensor. /// Generally it's only used for user to create Tensor.
......
paddle/fluid/extension/src/op_meta_info.cc
浏览文件 @ d3e60959
...@@ -32,6 +32,10 @@ OpMetaInfo& OpMetaInfo::Outputs(std::vector<std::string>&& outputs) { ...@@ -32,6 +32,10 @@ OpMetaInfo& OpMetaInfo::Outputs(std::vector<std::string>&& outputs) {
outputs_ = std::forward<std::vector<std::string>>(outputs); outputs_ = std::forward<std::vector<std::string>>(outputs);
return *this; return *this;
} }
OpMetaInfo& OpMetaInfo::Attrs(std::vector<std::string>&& attrs) {
attrs_ = std::forward<std::vector<std::string>>(attrs);
return *this;
}
OpMetaInfo& OpMetaInfo::SetKernelFn(KernelFunc&& func) { OpMetaInfo& OpMetaInfo::SetKernelFn(KernelFunc&& func) {
kernel_fn_ = std::forward<KernelFunc>(func); kernel_fn_ = std::forward<KernelFunc>(func);
return *this; return *this;
...@@ -78,17 +82,22 @@ OpMetaInfoBuilder& OpMetaInfoBuilder::Outputs( ...@@ -78,17 +82,22 @@ OpMetaInfoBuilder& OpMetaInfoBuilder::Outputs(
return *this; return *this;
} }
OpMetaInfoBuilder& OpMetaInfoBuilder::SetKernelFn(KernelFunc&& func) { OpMetaInfoBuilder& OpMetaInfoBuilder::Attrs(std::vector<std::string>&& attrs) {
info_ptr_->Attrs(std::forward<std::vector<std::string>>(attrs));
return *this;
}
OpMetaInfoBuilder& OpMetaInfoBuilder::SetKernelFn(KernelFunc func) {
info_ptr_->SetKernelFn(std::forward<KernelFunc>(func)); info_ptr_->SetKernelFn(std::forward<KernelFunc>(func));
return *this; return *this;
} }
OpMetaInfoBuilder& OpMetaInfoBuilder::SetInferShapeFn(InferShapeFunc&& func) { OpMetaInfoBuilder& OpMetaInfoBuilder::SetInferShapeFn(InferShapeFunc func) {
info_ptr_->SetInferShapeFn(std::forward<InferShapeFunc>(func)); info_ptr_->SetInferShapeFn(std::forward<InferShapeFunc>(func));
return *this; return *this;
} }
OpMetaInfoBuilder& OpMetaInfoBuilder::SetInferDtypeFn(InferDtypeFunc&& func) { OpMetaInfoBuilder& OpMetaInfoBuilder::SetInferDtypeFn(InferDtypeFunc func) {
info_ptr_->SetInferDtypeFn(std::forward<InferDtypeFunc>(func)); info_ptr_->SetInferDtypeFn(std::forward<InferDtypeFunc>(func));
return *this; return *this;
} }
...@@ -114,10 +123,17 @@ void LoadCustomOperatorLib(const std::string& dso_name) { ...@@ -114,10 +123,17 @@ void LoadCustomOperatorLib(const std::string& dso_name) {
} }
} // namespace paddle } // namespace paddle
#ifdef __cplusplus
extern "C" { extern "C" {
#endif
#ifndef _WIN32
// C-API to get global OpMetaInfoMap.
paddle::OpMetaInfoMap& PD_GetOpMetaInfoMap() { paddle::OpMetaInfoMap& PD_GetOpMetaInfoMap() {
return paddle::OpMetaInfoMap::Instance(); return paddle::OpMetaInfoMap::Instance();
} }
#endif
#ifdef __cplusplus
} // end extern "C" } // end extern "C"
#endif
paddle/fluid/extension/src/tensor.cc
浏览文件 @ d3e60959
...@@ -207,73 +207,87 @@ Tensor Tensor::copy_to(const PlaceType &target_place) const { ...@@ -207,73 +207,87 @@ Tensor Tensor::copy_to(const PlaceType &target_place) const {
return target; return target;
} }
template Tensor Tensor::copy_to<paddle::platform::float16>( template PD_DLL_DECL Tensor
Tensor::copy_to<paddle::platform::float16>(const PlaceType &target_place) const;
template PD_DLL_DECL Tensor Tensor::copy_to<paddle::platform::bfloat16>(
const PlaceType &target_place) const; const PlaceType &target_place) const;
template Tensor Tensor::copy_to<paddle::platform::bfloat16>( template PD_DLL_DECL Tensor Tensor::copy_to<paddle::platform::complex64>(
const PlaceType &target_place) const; const PlaceType &target_place) const;
template Tensor Tensor::copy_to<paddle::platform::complex64>( template PD_DLL_DECL Tensor Tensor::copy_to<paddle::platform::complex128>(
const PlaceType &target_place) const; const PlaceType &target_place) const;
template Tensor Tensor::copy_to<paddle::platform::complex128>( template PD_DLL_DECL Tensor
const PlaceType &target_place) const; Tensor::copy_to<float>(const PlaceType &target_place) const;
template Tensor Tensor::copy_to<float>(const PlaceType &target_place) const; template PD_DLL_DECL Tensor
template Tensor Tensor::copy_to<double>(const PlaceType &target_place) const; Tensor::copy_to<double>(const PlaceType &target_place) const;
template Tensor Tensor::copy_to<int64_t>(const PlaceType &target_place) const; template PD_DLL_DECL Tensor
template Tensor Tensor::copy_to<int32_t>(const PlaceType &target_place) const; Tensor::copy_to<int64_t>(const PlaceType &target_place) const;
template Tensor Tensor::copy_to<uint8_t>(const PlaceType &target_place) const; template PD_DLL_DECL Tensor
template Tensor Tensor::copy_to<int8_t>(const PlaceType &target_place) const; Tensor::copy_to<int32_t>(const PlaceType &target_place) const;
template Tensor Tensor::copy_to<int16_t>(const PlaceType &target_place) const; template PD_DLL_DECL Tensor
template Tensor Tensor::copy_to<bool>(const PlaceType &target_place) const; Tensor::copy_to<uint8_t>(const PlaceType &target_place) const;
template PD_DLL_DECL Tensor
Tensor::copy_to<int8_t>(const PlaceType &target_place) const;
template PD_DLL_DECL Tensor
Tensor::copy_to<int16_t>(const PlaceType &target_place) const;
template PD_DLL_DECL Tensor
Tensor::copy_to<bool>(const PlaceType &target_place) const;
template float *Tensor::data<float>() const; template PD_DLL_DECL float *Tensor::data<float>() const;
template double *Tensor::data<double>() const; template PD_DLL_DECL double *Tensor::data<double>() const;
template int64_t *Tensor::data<int64_t>() const; template PD_DLL_DECL int64_t *Tensor::data<int64_t>() const;
template int32_t *Tensor::data<int32_t>() const; template PD_DLL_DECL int32_t *Tensor::data<int32_t>() const;
template uint8_t *Tensor::data<uint8_t>() const; template PD_DLL_DECL uint8_t *Tensor::data<uint8_t>() const;
template int8_t *Tensor::data<int8_t>() const; template PD_DLL_DECL int8_t *Tensor::data<int8_t>() const;
template paddle::platform::float16 *Tensor::data<paddle::platform::float16>() template PD_DLL_DECL paddle::platform::float16 *
const; Tensor::data<paddle::platform::float16>() const;
template paddle::platform::bfloat16 *Tensor::data<paddle::platform::bfloat16>() template PD_DLL_DECL paddle::platform::bfloat16 *
const; Tensor::data<paddle::platform::bfloat16>() const;
template paddle::platform::complex128 * template PD_DLL_DECL paddle::platform::complex128 *
Tensor::data<paddle::platform::complex128>() const; Tensor::data<paddle::platform::complex128>() const;
template paddle::platform::complex64 * template PD_DLL_DECL paddle::platform::complex64 *
Tensor::data<paddle::platform::complex64>() const; Tensor::data<paddle::platform::complex64>() const;
template int16_t *Tensor::data<int16_t>() const; template PD_DLL_DECL int16_t *Tensor::data<int16_t>() const;
template bool *Tensor::data<bool>() const; template PD_DLL_DECL bool *Tensor::data<bool>() const;
template float *Tensor::mutable_data<float>(); template PD_DLL_DECL float *Tensor::mutable_data<float>();
template double *Tensor::mutable_data<double>(); template PD_DLL_DECL double *Tensor::mutable_data<double>();
template int64_t *Tensor::mutable_data<int64_t>(); template PD_DLL_DECL int64_t *Tensor::mutable_data<int64_t>();
template int32_t *Tensor::mutable_data<int32_t>(); template PD_DLL_DECL int32_t *Tensor::mutable_data<int32_t>();
template uint8_t *Tensor::mutable_data<uint8_t>(); template PD_DLL_DECL uint8_t *Tensor::mutable_data<uint8_t>();
template int8_t *Tensor::mutable_data<int8_t>(); template PD_DLL_DECL int8_t *Tensor::mutable_data<int8_t>();
template paddle::platform::float16 * template PD_DLL_DECL paddle::platform::float16 *
Tensor::mutable_data<paddle::platform::float16>(); Tensor::mutable_data<paddle::platform::float16>();
template paddle::platform::bfloat16 * template PD_DLL_DECL paddle::platform::bfloat16 *
Tensor::mutable_data<paddle::platform::bfloat16>(); Tensor::mutable_data<paddle::platform::bfloat16>();
template paddle::platform::complex128 * template PD_DLL_DECL paddle::platform::complex128 *
Tensor::mutable_data<paddle::platform::complex128>(); Tensor::mutable_data<paddle::platform::complex128>();
template paddle::platform::complex64 * template PD_DLL_DECL paddle::platform::complex64 *
Tensor::mutable_data<paddle::platform::complex64>(); Tensor::mutable_data<paddle::platform::complex64>();
template int16_t *Tensor::mutable_data<int16_t>(); template PD_DLL_DECL int16_t *Tensor::mutable_data<int16_t>();
template bool *Tensor::mutable_data<bool>(); template PD_DLL_DECL bool *Tensor::mutable_data<bool>();
template float *Tensor::mutable_data<float>(const PlaceType &place); template PD_DLL_DECL float *Tensor::mutable_data<float>(const PlaceType &place);
template double *Tensor::mutable_data<double>(const PlaceType &place); template PD_DLL_DECL double *Tensor::mutable_data<double>(
template int64_t *Tensor::mutable_data<int64_t>(const PlaceType &place); const PlaceType &place);
template int32_t *Tensor::mutable_data<int32_t>(const PlaceType &place); template PD_DLL_DECL int64_t *Tensor::mutable_data<int64_t>(
template uint8_t *Tensor::mutable_data<uint8_t>(const PlaceType &place); const PlaceType &place);
template int8_t *Tensor::mutable_data<int8_t>(const PlaceType &place); template PD_DLL_DECL int32_t *Tensor::mutable_data<int32_t>(
template paddle::platform::float16 * const PlaceType &place);
template PD_DLL_DECL uint8_t *Tensor::mutable_data<uint8_t>(
const PlaceType &place);
template PD_DLL_DECL int8_t *Tensor::mutable_data<int8_t>(
const PlaceType &place);
template PD_DLL_DECL paddle::platform::float16 *
Tensor::mutable_data<paddle::platform::float16>(const PlaceType &place); Tensor::mutable_data<paddle::platform::float16>(const PlaceType &place);
template paddle::platform::bfloat16 * template PD_DLL_DECL paddle::platform::bfloat16 *
Tensor::mutable_data<paddle::platform::bfloat16>(const PlaceType &place); Tensor::mutable_data<paddle::platform::bfloat16>(const PlaceType &place);
template paddle::platform::complex128 * template PD_DLL_DECL paddle::platform::complex128 *
Tensor::mutable_data<paddle::platform::complex128>(const PlaceType &place); Tensor::mutable_data<paddle::platform::complex128>(const PlaceType &place);
template paddle::platform::complex64 * template PD_DLL_DECL paddle::platform::complex64 *
Tensor::mutable_data<paddle::platform::complex64>(const PlaceType &place); Tensor::mutable_data<paddle::platform::complex64>(const PlaceType &place);
template int16_t *Tensor::mutable_data<int16_t>(const PlaceType &place); template PD_DLL_DECL int16_t *Tensor::mutable_data<int16_t>(
template bool *Tensor::mutable_data<bool>(const PlaceType &place); const PlaceType &place);
template PD_DLL_DECL bool *Tensor::mutable_data<bool>(const PlaceType &place);
std::vector<int> Tensor::shape() const { std::vector<int> Tensor::shape() const {
GET_CASTED_TENSOR GET_CASTED_TENSOR
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ d3e60959
...@@ -321,9 +321,9 @@ message(STATUS "branch: ${PADDLE_BRANCH}") ...@@ -321,9 +321,9 @@ message(STATUS "branch: ${PADDLE_BRANCH}")
configure_file(commit.h.in commit.h) configure_file(commit.h.in commit.h)
cc_library(custom_tensor SRCS ../extension/src/tensor.cc DEPS lod_tensor) cc_library(custom_tensor SRCS ../extension/src/tensor.cc DEPS lod_tensor memory enforce)
cc_library(op_meta_info SRCS ../extension/src/op_meta_info.cc DEPS custom_tensor) cc_library(op_meta_info SRCS ../extension/src/op_meta_info.cc DEPS custom_tensor)
cc_library(custom_operator SRCS custom_operator.cc DEPS operator op_registry device_context dynamic_loader custom_tensor op_meta_info) cc_library(custom_operator SRCS custom_operator.cc DEPS tensor attribute framework_proto op_registry operator dynamic_loader string_helper custom_tensor op_meta_info)
cc_test(custom_tensor_test SRCS custom_tensor_test.cc DEPS custom_tensor glog) cc_test(custom_tensor_test SRCS custom_tensor_test.cc DEPS custom_tensor glog)
set(FLUID_FRAMEWORK_MODULES proto_desc memory lod_tensor executor data_feed_proto layer dynamic_loader custom_operator) set(FLUID_FRAMEWORK_MODULES proto_desc memory lod_tensor executor data_feed_proto layer dynamic_loader custom_operator)
...@@ -346,9 +346,12 @@ if (LINUX) ...@@ -346,9 +346,12 @@ if (LINUX)
endif() endif()
if (WIN32) if (WIN32)
set(FLUID_FRAMEWORK_SHARED_LIB set(FLUID_FRAMEWORK_IMPORT_LIB
${PADDLE_BINARY_DIR}/paddle/fluid/framework/libpaddle_framework.dll ${PADDLE_BINARY_DIR}/paddle/fluid/framework/${CMAKE_BUILD_TYPE}/paddle_framework.lib
CACHE INTERNAL "Fluid framework lib") CACHE INTERNAL "Fluid framework lib")
set(FLUID_FRAMEWORK_SHARED_LIB
${PADDLE_BINARY_DIR}/paddle/fluid/framework/${CMAKE_BUILD_TYPE}/paddle_framework.dll
CACHE INTERNAL "Fluid framework dll")
endif() endif()
if(APPLE) if(APPLE)
...@@ -359,3 +362,37 @@ endif() ...@@ -359,3 +362,37 @@ endif()
if(WITH_TESTING) if(WITH_TESTING)
set_tests_properties(selected_rows_test PROPERTIES TIMEOUT 120) set_tests_properties(selected_rows_test PROPERTIES TIMEOUT 120)
endif() endif()
# New custom op extension mechanism related
# if not deps `layer`, will cause: undefined symbol: _ZN6paddle10imperative7VarBase9name_set_
set(PADDLE_CUSTOM_OP_MODULES custom_tensor op_meta_info custom_operator layer)
cc_library(paddle_custom_op_shared
SHARED SRCS custom_operator.cc ../extension/src/tensor.cc ../extension/src/op_meta_info.cc
${CMAKE_SOURCE_DIR}/paddle/fluid/imperative/layer.cc
DEPS ${PADDLE_CUSTOM_OP_MODULES})
get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
set_target_properties(paddle_custom_op_shared PROPERTIES OUTPUT_NAME paddle_custom_op)
target_link_libraries(paddle_custom_op_shared ${os_dependency_modules})
if (LINUX)
set(PADDLE_CUSTOM_OP_SHARED_LIB
${PADDLE_BINARY_DIR}/paddle/fluid/framework/libpaddle_custom_op.so
CACHE INTERNAL "Paddle custom op lib")
endif()
if (WIN32)
set(PADDLE_CUSTOM_OP_SHARED_LIB
${PADDLE_BINARY_DIR}/paddle/fluid/framework/${CMAKE_BUILD_TYPE}/paddle_custom_op.lib
CACHE INTERNAL "Paddle custom op lib")
set(PADDLE_CUSTOM_OP_SHARED_LIB
${PADDLE_BINARY_DIR}/paddle/fluid/framework/${CMAKE_BUILD_TYPE}/paddle_custom_op.dll
CACHE INTERNAL "Paddle custom op dll")
endif()
if(APPLE)
set(PADDLE_CUSTOM_OP_SHARED_LIB
${PADDLE_BINARY_DIR}/paddle/fluid/framework/paddle_custom_op.dylib
CACHE INTERNAL "Paddle custom op lib")
endif()
paddle/fluid/framework/custom_operator.cc
浏览文件 @ d3e60959
...@@ -73,6 +73,24 @@ inline bool IsMemberOf(const std::vector<std::string>& vec, ...@@ -73,6 +73,24 @@ inline bool IsMemberOf(const std::vector<std::string>& vec,
return std::find(vec.cbegin(), vec.cend(), name) != vec.cend(); return std::find(vec.cbegin(), vec.cend(), name) != vec.cend();
} }
std::vector<std::string> ParseAttrStr(const std::string& attr) {
auto split_pos = attr.find_first_of(":");
PADDLE_ENFORCE_NE(split_pos, std::string::npos,
platform::errors::InvalidArgument(
"Invalid attribute string format. Attribute string "
"format is `<name>:<type>`."));
std::vector<std::string> rlt;
// 1. name
rlt.emplace_back(string::trim_spaces(attr.substr(0, split_pos)));
// 2. type
rlt.emplace_back(string::trim_spaces(attr.substr(split_pos + 1)));
VLOG(1) << "attr name: " << rlt[0] << ", attr type str: " << rlt[1];
return rlt;
}
} // namespace detail } // namespace detail
////////////////// Kernel Define //////////////////// ////////////////// Kernel Define ////////////////////
...@@ -81,7 +99,8 @@ inline bool IsMemberOf(const std::vector<std::string>& vec, ...@@ -81,7 +99,8 @@ inline bool IsMemberOf(const std::vector<std::string>& vec,
static void RunKernelFunc(const framework::ExecutionContext& ctx, static void RunKernelFunc(const framework::ExecutionContext& ctx,
const paddle::KernelFunc& func, const paddle::KernelFunc& func,
const std::vector<std::string>& inputs, const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) { const std::vector<std::string>& outputs,
const std::vector<std::string>& attrs) {
VLOG(1) << "Custom Operator: Start run KernelFunc."; VLOG(1) << "Custom Operator: Start run KernelFunc.";
std::vector<paddle::Tensor> custom_ins; std::vector<paddle::Tensor> custom_ins;
for (auto& in_name : inputs) { for (auto& in_name : inputs) {
...@@ -98,10 +117,43 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx, ...@@ -98,10 +117,43 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
custom_ins.emplace_back(custom_in); custom_ins.emplace_back(custom_in);
} }
std::vector<boost::any> attrs; std::vector<boost::any> custom_attrs;
for (auto& attr_str : attrs) {
auto attr_name_and_type = detail::ParseAttrStr(attr_str);
auto attr_name = attr_name_and_type[0];
auto attr_type_str = attr_name_and_type[1];
if (attr_type_str == "bool") {
custom_attrs.emplace_back(ctx.Attr<bool>(attr_name));
} else if (attr_type_str == "int") {
custom_attrs.emplace_back(ctx.Attr<int>(attr_name));
} else if (attr_type_str == "float") {
custom_attrs.emplace_back(ctx.Attr<float>(attr_name));
} else if (attr_type_str == "int64_t") {
custom_attrs.emplace_back(ctx.Attr<int64_t>(attr_name));
} else if (attr_type_str == "std::string") {
custom_attrs.emplace_back(ctx.Attr<std::string>(attr_name));
} else if (attr_type_str == "std::vector<int>") {
custom_attrs.emplace_back(ctx.Attr<std::vector<int>>(attr_name));
} else if (attr_type_str == "std::vector<float>") {
custom_attrs.emplace_back(ctx.Attr<std::vector<float>>(attr_name));
} else if (attr_type_str == "std::vector<int64_t>") {
custom_attrs.emplace_back(ctx.Attr<std::vector<int64_t>>(attr_name));
} else if (attr_type_str == "std::vector<std::string>") {
custom_attrs.emplace_back(ctx.Attr<std::vector<std::string>>(attr_name));
} else {
PADDLE_THROW(platform::errors::Unimplemented(
"Unsupported `%s` type value as custom attribute now. "
"Supported data types include `bool`, `int`, `float`, "
"`int64_t`, `std::string`, `std::vector<int>`, "
"`std::vector<float>`, `std::vector<int64_t>, "
"`std::vector<std::string>`, Please check whether "
"the attribute data type and data type string are matched.",
attr_type_str));
}
}
VLOG(1) << "Run ComputeFunc."; VLOG(1) << "Run ComputeFunc.";
auto outs = func(custom_ins, attrs); auto outs = func(custom_ins, custom_attrs);
VLOG(1) << "Custom Operator: Share outputs into ExecutionContext."; VLOG(1) << "Custom Operator: Share outputs into ExecutionContext.";
for (size_t i = 0; i < outputs.size(); ++i) { for (size_t i = 0; i < outputs.size(); ++i) {
...@@ -164,7 +216,51 @@ class CustomOpMaker : public OpProtoAndCheckerMaker { ...@@ -164,7 +216,51 @@ class CustomOpMaker : public OpProtoAndCheckerMaker {
for (auto& out_name : outputs_) { for (auto& out_name : outputs_) {
AddOutput(out_name, "The output " + out_name + "of Custom Operator."); AddOutput(out_name, "The output " + out_name + "of Custom Operator.");
} }
// TODO(chenweihang): support attrs in later PR for (auto& attr : attrs_) {
auto attr_name_and_type = detail::ParseAttrStr(attr);
auto attr_name = attr_name_and_type[0];
auto attr_type_str = attr_name_and_type[1];
if (attr_type_str == "bool") {
AddAttr<bool>(attr_name, "custom operator bool attribute.")
.SetDefault(false);
} else if (attr_type_str == "int") {
AddAttr<int>(attr_name, "custom operator int attribute.").SetDefault(1);
} else if (attr_type_str == "float") {
AddAttr<float>(attr_name, "custom operator float attribute.")
.SetDefault(1.0f);
} else if (attr_type_str == "int64_t") {
AddAttr<int64_t>(attr_name, "custom operator int64_t attribute.")
.SetDefault(1);
} else if (attr_type_str == "std::string") {
AddAttr<std::string>(attr_name, "custom operator int attribute.")
.SetDefault("");
} else if (attr_type_str == "std::vector<int>") {
AddAttr<std::vector<int>>(attr_name,
"custom operator std::vector<int> attribute.")
.SetDefault({});
} else if (attr_type_str == "std::vector<float>") {
AddAttr<std::vector<float>>(
attr_name, "custom operator std::vector<float> attribute.")
.SetDefault({});
} else if (attr_type_str == "std::vector<int64_t>") {
AddAttr<std::vector<int64_t>>(
attr_name, "custom operator std::vector<int64_t> attribute.")
.SetDefault({});
} else if (attr_type_str == "std::vector<std::string>") {
AddAttr<std::vector<std::string>>(
attr_name, "custom operator std::vector<std::string> attribute.")
.SetDefault({});
} else {
PADDLE_THROW(platform::errors::Unimplemented(
"Unsupported `%s` type value as custom attribute now. "
"Supported data types include `bool`, `int`, `float`, "
"`int64_t`, `std::string`, `std::vector<int>`, "
"`std::vector<float>`, `std::vector<int64_t>, "
"`std::vector<std::string>`, Please check whether "
"the attribute data type and data type string are matched.",
attr_type_str));
}
}
AddComment(R"DOC( AddComment(R"DOC(
Custom Operator. Custom Operator.
...@@ -227,7 +323,7 @@ class CustomGradOpMaker<OpDesc> : public SingleGradOpMaker<OpDesc> { ...@@ -227,7 +323,7 @@ class CustomGradOpMaker<OpDesc> : public SingleGradOpMaker<OpDesc> {
VLOG(1) << "Custom Operator: GradOpDescMaker - output: " << out_name; VLOG(1) << "Custom Operator: GradOpDescMaker - output: " << out_name;
grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(out_name))); grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(out_name)));
} }
// TODO(chenweihang): support attrs in later PR grad_op->SetAttrMap(this->Attrs());
} }
private: private:
...@@ -287,7 +383,7 @@ class CustomGradOpMaker<imperative::OpBase> ...@@ -287,7 +383,7 @@ class CustomGradOpMaker<imperative::OpBase>
VLOG(1) << "Custom Operator: GradOpBaseMaker - output: " << out_name; VLOG(1) << "Custom Operator: GradOpBaseMaker - output: " << out_name;
grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(out_name))); grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(out_name)));
} }
// TODO(chenweihang): support attrs in later PR grad_op->SetAttrMap(this->Attrs());
} }
private: private:
...@@ -303,21 +399,24 @@ void RegisterOperatorKernelWithPlace(const std::string& name, ...@@ -303,21 +399,24 @@ void RegisterOperatorKernelWithPlace(const std::string& name,
const proto::VarType::Type type, const proto::VarType::Type type,
const PlaceType& place, const PlaceType& place,
const std::vector<std::string>& inputs, const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) { const std::vector<std::string>& outputs,
const std::vector<std::string>& attrs) {
OpKernelType key(type, OpKernelType key(type,
CustomTensorUtils::ConvertEnumPlaceToInnerPlace(place)); CustomTensorUtils::ConvertEnumPlaceToInnerPlace(place));
VLOG(1) << "Custom Operator: op kernel key: " << key; VLOG(1) << "Custom Operator: op kernel key: " << key;
OperatorWithKernel::AllOpKernels()[name][key] = OperatorWithKernel::AllOpKernels()[name][key] =
[kernel_func, inputs, outputs](const framework::ExecutionContext& ctx) { [kernel_func, inputs, outputs,
attrs](const framework::ExecutionContext& ctx) {
VLOG(1) << "Custom Operator: run custom kernel func in lambda."; VLOG(1) << "Custom Operator: run custom kernel func in lambda.";
RunKernelFunc(ctx, kernel_func, inputs, outputs); RunKernelFunc(ctx, kernel_func, inputs, outputs, attrs);
}; };
} }
void RegisterOperatorKernel(const std::string& name, void RegisterOperatorKernel(const std::string& name,
const paddle::KernelFunc& kernel_func, const paddle::KernelFunc& kernel_func,
const std::vector<std::string>& inputs, const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) { const std::vector<std::string>& outputs,
const std::vector<std::string>& attrs) {
VLOG(1) << "Custom Operator: op name in kernel: " << name; VLOG(1) << "Custom Operator: op name in kernel: " << name;
// NOTE [ Dummy Op Kernel Key ] // NOTE [ Dummy Op Kernel Key ]
// TODO(chenweihang): Because execute engine need get device context based // TODO(chenweihang): Because execute engine need get device context based
...@@ -325,9 +424,11 @@ void RegisterOperatorKernel(const std::string& name, ...@@ -325,9 +424,11 @@ void RegisterOperatorKernel(const std::string& name,
// device. But this is not entirely correct, if user only give a cpu kernel, // device. But this is not entirely correct, if user only give a cpu kernel,
// but call api in gpu device, it will cause error. // but call api in gpu device, it will cause error.
RegisterOperatorKernelWithPlace(name, kernel_func, proto::VarType::RAW, RegisterOperatorKernelWithPlace(name, kernel_func, proto::VarType::RAW,
PlaceType::kCPU, inputs, outputs); PlaceType::kCPU, inputs, outputs, attrs);
#ifdef PADDLE_WITH_CUDA
RegisterOperatorKernelWithPlace(name, kernel_func, proto::VarType::RAW, RegisterOperatorKernelWithPlace(name, kernel_func, proto::VarType::RAW,
PlaceType::kGPU, inputs, outputs); PlaceType::kGPU, inputs, outputs, attrs);
#endif
} }
void RegisterOperatorWithMetaInfo( void RegisterOperatorWithMetaInfo(
...@@ -350,6 +451,8 @@ void RegisterOperatorWithMetaInfo( ...@@ -350,6 +451,8 @@ void RegisterOperatorWithMetaInfo(
<< string::join_strings(op_inputs, ','); << string::join_strings(op_inputs, ',');
VLOG(1) << "Custom Operator: forward, op outputs: " VLOG(1) << "Custom Operator: forward, op outputs: "
<< string::join_strings(op_outputs, ','); << string::join_strings(op_outputs, ',');
VLOG(1) << "Custom Operator: forward, op attrs: "
<< string::join_strings(op_attrs, ',');
// Op // Op
info.creator_ = [](const std::string& op_name, const VariableNameMap& inputs, info.creator_ = [](const std::string& op_name, const VariableNameMap& inputs,
...@@ -426,7 +529,7 @@ void RegisterOperatorWithMetaInfo( ...@@ -426,7 +529,7 @@ void RegisterOperatorWithMetaInfo(
}; };
// Kernel func // Kernel func
RegisterOperatorKernel(op_name, kernel_fn, op_inputs, op_outputs); RegisterOperatorKernel(op_name, kernel_fn, op_inputs, op_outputs, op_attrs);
// If grad op or double grad op exists // If grad op or double grad op exists
std::string cur_op_name = op_name; std::string cur_op_name = op_name;
...@@ -436,6 +539,7 @@ void RegisterOperatorWithMetaInfo( ...@@ -436,6 +539,7 @@ void RegisterOperatorWithMetaInfo(
auto& grad_op_name = OpMetaInfoHelper::GetOpName(cur_grad_op); auto& grad_op_name = OpMetaInfoHelper::GetOpName(cur_grad_op);
auto& grad_op_inputs = OpMetaInfoHelper::GetInputs(cur_grad_op); auto& grad_op_inputs = OpMetaInfoHelper::GetInputs(cur_grad_op);
auto& grad_op_outputs = OpMetaInfoHelper::GetOutputs(cur_grad_op); auto& grad_op_outputs = OpMetaInfoHelper::GetOutputs(cur_grad_op);
auto& grad_op_attrs = OpMetaInfoHelper::GetAttrs(cur_grad_op);
auto& grad_kernel_fn = OpMetaInfoHelper::GetKernelFn(cur_grad_op); auto& grad_kernel_fn = OpMetaInfoHelper::GetKernelFn(cur_grad_op);
VLOG(1) << "Custom Operator: backward, op name: " << grad_op_name; VLOG(1) << "Custom Operator: backward, op name: " << grad_op_name;
...@@ -489,7 +593,7 @@ void RegisterOperatorWithMetaInfo( ...@@ -489,7 +593,7 @@ void RegisterOperatorWithMetaInfo(
// Kernel func // Kernel func
RegisterOperatorKernel(grad_op_name, grad_kernel_fn, grad_op_inputs, RegisterOperatorKernel(grad_op_name, grad_kernel_fn, grad_op_inputs,
grad_op_outputs); grad_op_outputs, grad_op_attrs);
// update current info // update current info
OpInfoMap::Instance().Insert(cur_op_name, info); OpInfoMap::Instance().Insert(cur_op_name, info);
......
paddle/fluid/platform/dynload/dynamic_loader.cc
浏览文件 @ d3e60959
...@@ -378,9 +378,6 @@ void* GetOpDsoHandle(const std::string& dso_name) { ...@@ -378,9 +378,6 @@ void* GetOpDsoHandle(const std::string& dso_name) {
#if defined(__APPLE__) || defined(__OSX__) #if defined(__APPLE__) || defined(__OSX__)
PADDLE_THROW(platform::errors::Unimplemented( PADDLE_THROW(platform::errors::Unimplemented(
"Create custom cpp op outside framework do not support Apple.")); "Create custom cpp op outside framework do not support Apple."));
#elif defined(_WIN32) && defined(PADDLE_WITH_CUDA)
PADDLE_THROW(platform::errors::Unimplemented(
"Create custom cpp op outside framework do not support Windows."));
#else #else
return GetDsoHandleFromSearchPath(FLAGS_op_dir, dso_name); return GetDsoHandleFromSearchPath(FLAGS_op_dir, dso_name);
#endif #endif
......
paddle/scripts/paddle_build.bat
浏览文件 @ d3e60959
...@@ -114,23 +114,25 @@ rem ------pre install python requirement---------- ...@@ -114,23 +114,25 @@ rem ------pre install python requirement----------
where python where python
where pip where pip
pip install wheel --user pip install wheel --user
pip install -r %work_dir%\python\requirements.txt --user
pip install -r %work_dir%\python\unittest_py\requirements.txt --user pip install -r %work_dir%\python\unittest_py\requirements.txt --user
pip install -r %work_dir%\python\requirements.txt --user
if %ERRORLEVEL% NEQ 0 ( if %ERRORLEVEL% NEQ 0 (
echo pip install requirements.txt failed! echo pip install requirements.txt failed!
exit /b 7 exit /b 7
) )
rem ------pre install clcache and init config---------- rem ------pre install clcache and init config----------
pip install clcache --user rem pip install clcache --user
pip uninstall -y clcache
:: set USE_CLCACHE to enable clcache :: set USE_CLCACHE to enable clcache
set USE_CLCACHE=1 rem set USE_CLCACHE=1
:: In some scenarios, CLCACHE_HARDLINK can save one file copy. :: In some scenarios, CLCACHE_HARDLINK can save one file copy.
set CLCACHE_HARDLINK=1 rem set CLCACHE_HARDLINK=1
:: If it takes more than 1000s to obtain the right to use the cache, an error will be reported :: If it takes more than 1000s to obtain the right to use the cache, an error will be reported
set CLCACHE_OBJECT_CACHE_TIMEOUT_MS=1000000 rem set CLCACHE_OBJECT_CACHE_TIMEOUT_MS=1000000
:: set maximum cache size to 20G :: set maximum cache size to 20G
clcache.exe -M 21474836480 rem clcache.exe -M 21474836480
rem ------show summary of current environment---------- rem ------show summary of current environment----------
python %work_dir%\tools\summary_env.py python %work_dir%\tools\summary_env.py
...@@ -194,11 +196,28 @@ set start=%start:~4,10% ...@@ -194,11 +196,28 @@ set start=%start:~4,10%
@ECHO ON @ECHO ON
if not defined CUDA_TOOLKIT_ROOT_DIR set CUDA_TOOLKIT_ROOT_DIR=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0 if not defined CUDA_TOOLKIT_ROOT_DIR set CUDA_TOOLKIT_ROOT_DIR=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0
set PATH=%CUDA_TOOLKIT_ROOT_DIR%\bin;%CUDA_TOOLKIT_ROOT_DIR%\libnvvp;%PATH% set PATH=%TENSORRT_ROOT:/=\%\lib;%CUDA_TOOLKIT_ROOT_DIR%\bin;%CUDA_TOOLKIT_ROOT_DIR%\libnvvp;%PATH%
set CUDA_PATH=%CUDA_TOOLKIT_ROOT_DIR%
rem ------set third_party cache dir------ rem ------set third_party cache dir------
: clear third party cache every once in a while
for /F %%# in ('wmic os get localdatetime^|findstr 20') do set datetime=%%#
set day_now=%datetime:~6,2%
set day_before=-1
set /p day_before=< %cache_dir%\day.txt
if %day_now% NEQ %day_before% (
echo %day_now% > %cache_dir%\day.txt
type %cache_dir%\day.txt
if %day_now% EQU 25 (
rmdir %cache_dir%\third_party_GPU/ /s/q
rmdir %cache_dir%\third_party/ /s/q
)
if %day_now% EQU 10 (
rmdir %cache_dir%\third_party_GPU/ /s/q
rmdir %cache_dir%\third_party/ /s/q
)
)
if "%WITH_TPCACHE%"=="OFF" ( if "%WITH_TPCACHE%"=="OFF" (
set THIRD_PARTY_PATH=%work_dir:\=/%/build/third_party set THIRD_PARTY_PATH=%work_dir:\=/%/build/third_party
goto :cmake_impl goto :cmake_impl
...@@ -263,6 +282,9 @@ echo Build third_party successfully! ...@@ -263,6 +282,9 @@ echo Build third_party successfully!
set build_times=1 set build_times=1
:build_paddle :build_paddle
:: reset clcache zero stats for collect PR's actual hit rate
rem clcache.exe -z
echo Build Paddle the %build_times% time: echo Build Paddle the %build_times% time:
if "%WITH_CLCACHE%"=="OFF" ( if "%WITH_CLCACHE%"=="OFF" (
msbuild /m:%PARALLEL_PROJECT_COUNT% /p:Configuration=Release /verbosity:minimal paddle.sln msbuild /m:%PARALLEL_PROJECT_COUNT% /p:Configuration=Release /verbosity:minimal paddle.sln
...@@ -281,6 +303,11 @@ if %ERRORLEVEL% NEQ 0 ( ...@@ -281,6 +303,11 @@ if %ERRORLEVEL% NEQ 0 (
) )
echo Build Paddle successfully! echo Build Paddle successfully!
echo 0 > %cache_dir%\error_code.txt
type %cache_dir%\error_code.txt
:: ci will collect clcache hit rate
rem goto :collect_clcache_hits
goto:eof goto:eof
...@@ -319,13 +346,14 @@ set /p PADDLE_WHL_FILE_WIN=< whl_file.txt ...@@ -319,13 +346,14 @@ set /p PADDLE_WHL_FILE_WIN=< whl_file.txt
@ECHO ON @ECHO ON
pip uninstall -y paddlepaddle pip uninstall -y paddlepaddle
pip uninstall -y paddlepaddle-gpu pip uninstall -y paddlepaddle-gpu
pip install -U %PADDLE_WHL_FILE_WIN% --user pip install %PADDLE_WHL_FILE_WIN% --user
if %ERRORLEVEL% NEQ 0 ( if %ERRORLEVEL% NEQ 0 (
call paddle_winci\Scripts\deactivate.bat 2>NUL call paddle_winci\Scripts\deactivate.bat 2>NUL
echo pip install whl package failed! echo pip install whl package failed!
exit /b 1 exit /b 1
) )
set CUDA_VISIBLE_DEVICES=0 set CUDA_VISIBLE_DEVICES=0
python %work_dir%\paddle\scripts\installation_validate.py python %work_dir%\paddle\scripts\installation_validate.py
goto:eof goto:eof
...@@ -383,7 +411,7 @@ if "%WITH_GPU%"=="ON" ( ...@@ -383,7 +411,7 @@ if "%WITH_GPU%"=="ON" (
:parallel_test_base_gpu :parallel_test_base_gpu
echo ======================================== echo ========================================
echo Running GPU unit tests... echo Running GPU unit tests in parallel way ...
echo ======================================== echo ========================================
setlocal enabledelayedexpansion setlocal enabledelayedexpansion
...@@ -451,6 +479,7 @@ goto:eof ...@@ -451,6 +479,7 @@ goto:eof
echo ======================================== echo ========================================
echo Running CPU unit tests in parallel way ... echo Running CPU unit tests in parallel way ...
echo ======================================== echo ========================================
ctest.exe -E "(%disable_ut_quickly%)" -LE %nightly_label% --output-on-failure -C Release -j 8 --repeat until-pass:4 after-timeout:4 ctest.exe -E "(%disable_ut_quickly%)" -LE %nightly_label% --output-on-failure -C Release -j 8 --repeat until-pass:4 after-timeout:4
goto:eof goto:eof
...@@ -622,6 +651,7 @@ taskkill /f /im vctip.exe 2>NUL ...@@ -622,6 +651,7 @@ taskkill /f /im vctip.exe 2>NUL
taskkill /f /im cvtres.exe 2>NUL taskkill /f /im cvtres.exe 2>NUL
taskkill /f /im rc.exe 2>NUL taskkill /f /im rc.exe 2>NUL
wmic process where name="op_function_generator.exe" call terminate 2>NUL wmic process where name="op_function_generator.exe" call terminate 2>NUL
wmic process where name="python.exe" call terminate 2>NUL
taskkill /f /im python.exe 2>NUL taskkill /f /im python.exe 2>NUL
echo 0 > %cache_dir%\error_code.txt echo 0 > %cache_dir%\error_code.txt
type %cache_dir%\error_code.txt type %cache_dir%\error_code.txt
......
python/paddle/fluid/tests/CMakeLists.txt
浏览文件 @ d3e60959
...@@ -9,7 +9,9 @@ endforeach() ...@@ -9,7 +9,9 @@ endforeach()
add_subdirectory(unittests) add_subdirectory(unittests)
add_subdirectory(book) add_subdirectory(book)
if(NOT APPLE AND NOT WIN32) # TODO: support New Custom OP on Mac
if(NOT APPLE)
add_subdirectory(custom_op) add_subdirectory(custom_op)
endif() endif()
set_tests_properties(test_beam_search_decoder PROPERTIES TIMEOUT 120) set_tests_properties(test_beam_search_decoder PROPERTIES TIMEOUT 120)
python/paddle/fluid/tests/custom_op/CMakeLists.txt
浏览文件 @ d3e60959
if (WITH_GPU) # New custom OP can support Windows/Linux now
if(WITH_GPU)
# 'test_custom_relu_op_setup/jit' compile .cc and .cu file
py_test(test_custom_relu_op_setup SRCS test_custom_relu_op_setup.py)
py_test(test_custom_relu_op_jit SRCS test_custom_relu_op_jit.py)
# Compiling shared library will cost some time, but running process is very fast.
set_tests_properties(test_custom_relu_op_setup PROPERTIES TIMEOUT 250)
set_tests_properties(test_custom_relu_op_jit PROPERTIES TIMEOUT 180)
endif()
py_test(test_sysconfig SRCS test_sysconfig.py)
# 'test_dispatch' compile .cc file
py_test(test_dispatch_jit SRCS test_dispatch_jit.py)
set_tests_properties(test_dispatch_jit PROPERTIES TIMEOUT 120)
py_test(test_multi_out_jit SRCS test_multi_out_jit.py)
set_tests_properties(test_multi_out_jit PROPERTIES TIMEOUT 120)
py_test(test_custom_attrs_jit SRCS test_custom_attrs_jit.py)
set_tests_properties(test_custom_attrs_jit PROPERTIES TIMEOUT 120)
if(NOT LINUX)
return()
endif()
# TODO(zhouwei): support test_check_abi and abi check on Windows
py_test(test_check_abi SRCS test_check_abi.py)
# Old custom OP only support Linux, only run on Linux
py_test(test_custom_op SRCS test_custom_op.py)
py_test(test_jit_load SRCS test_jit_load.py)
py_test(test_setup_install SRCS test_setup_install.py)
py_test(test_setup_build SRCS test_setup_build.py)
set_tests_properties(test_jit_load PROPERTIES TIMEOUT 180)
set_tests_properties(test_setup_install PROPERTIES TIMEOUT 180)
set_tests_properties(test_setup_build PROPERTIES TIMEOUT 180)
if(WITH_ROCM)
hip_library(relu_op_shared SHARED SRCS relu_op.cc relu_op.cu DEPS paddle_framework_shared)
elseif(WITH_GPU)
nv_library(relu_op_shared SHARED SRCS relu_op.cc relu_op.cu DEPS paddle_framework_shared) nv_library(relu_op_shared SHARED SRCS relu_op.cc relu_op.cu DEPS paddle_framework_shared)
else() else()
cc_library(relu_op_shared SHARED SRCS relu_op.cc DEPS paddle_framework_shared) cc_library(relu_op_shared SHARED SRCS relu_op.cc DEPS paddle_framework_shared)
...@@ -16,19 +59,3 @@ get_target_property(TARGET_LIBRARIES relu_op_shared LINK_LIBRARIES) ...@@ -16,19 +59,3 @@ get_target_property(TARGET_LIBRARIES relu_op_shared LINK_LIBRARIES)
LIST(REMOVE_ITEM TARGET_LIBRARIES glog) LIST(REMOVE_ITEM TARGET_LIBRARIES glog)
LIST(REMOVE_ITEM TARGET_LIBRARIES gflags) LIST(REMOVE_ITEM TARGET_LIBRARIES gflags)
set_property(TARGET relu_op_shared PROPERTY LINK_LIBRARIES ${TARGET_LIBRARIES} ) set_property(TARGET relu_op_shared PROPERTY LINK_LIBRARIES ${TARGET_LIBRARIES} )
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
endforeach()
# Compiling .so will cost some time, but running process is very fast.
set_tests_properties(test_jit_load PROPERTIES TIMEOUT 180)
set_tests_properties(test_setup_install PROPERTIES TIMEOUT 180)
set_tests_properties(test_setup_build PROPERTIES TIMEOUT 180)
set_tests_properties(test_dispatch PROPERTIES TIMEOUT 180)
set_tests_properties(test_simple_custom_op_setup PROPERTIES TIMEOUT 250)
set_tests_properties(test_simple_custom_op_jit PROPERTIES TIMEOUT 180)
python/paddle/fluid/tests/custom_op/attr_test_op.cc 0 → 100644
浏览文件 @ d3e60959
// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <cstdlib>
#include <iostream>
#include <vector>
#include "paddle/extension.h"
template <typename data_t>
void assign_cpu_kernel(const data_t* x_data,
data_t* out_data,
int64_t x_numel) {
for (int i = 0; i < x_numel; ++i) {
out_data[i] = x_data[i];
}
}
std::vector<paddle::Tensor> AttrTestForward(
const paddle::Tensor& x,
bool bool_attr,
int int_attr,
float float_attr,
int64_t int64_attr,
std::string str_attr,
std::vector<int> int_vec_attr,
std::vector<float> float_vec_attr,
std::vector<int64_t> int64_vec_attr,
std::vector<std::string> str_vec_attr) {
auto out = paddle::Tensor(paddle::PlaceType::kCPU);
out.reshape(x.shape());
PD_DISPATCH_FLOATING_TYPES(
x.type(), "assign_cpu_kernel", ([&] {
assign_cpu_kernel<data_t>(
x.data<data_t>(), out.mutable_data<data_t>(), x.size());
}));
// Check attrs value
if (bool_attr != true) {
throw std::runtime_error("bool_attr value error.");
}
if (int_attr != 10) {
throw std::runtime_error("int_attr value error.");
}
if (std::abs(float_attr - 3.14) > 1e-6) {
throw std::runtime_error("float_attr value error.");
}
if (int64_attr != 10000000000) {
throw std::runtime_error("int64_attr value error.");
}
if (str_attr != "StrAttr") {
throw std::runtime_error("str_attr value error.");
}
if (int_vec_attr.size() != 3) {
throw std::runtime_error("int_vec_attr size error.");
} else {
for (auto& value : int_vec_attr) {
if (value != 10) {
throw std::runtime_error("int_vec_attr value error.");
}
}
}
if (float_vec_attr.size() != 3) {
throw std::runtime_error("float_vec_attr size error.");
} else {
for (auto& value : float_vec_attr) {
if (std::abs(value - 3.14) > 1e-6) {
throw std::runtime_error("float_vec_attr value error.");
}
}
}
if (int64_vec_attr.size() != 3) {
throw std::runtime_error("int64_vec_attr size error.");
} else {
for (auto& value : int64_vec_attr) {
if (value != 10000000000) {
throw std::runtime_error("int64_vec_attr value error.");
}
}
}
if (str_vec_attr.size() != 3) {
throw std::runtime_error("str_vec_attr size error.");
} else {
for (auto& value : str_vec_attr) {
if (value != "StrAttr") {
throw std::runtime_error("str_vec_attr value error.");
}
}
}
return {out};
}
// The attrs of backward op must be the subset of attrs of forward op
std::vector<paddle::Tensor> AttrTestBackward(
const paddle::Tensor& grad_out,
int int_attr,
std::vector<float> float_vec_attr,
std::vector<std::string> str_vec_attr) {
auto grad_x = paddle::Tensor(paddle::PlaceType::kCPU);
grad_x.reshape(grad_out.shape());
PD_DISPATCH_FLOATING_TYPES(grad_out.type(), "assign_cpu_kernel", ([&] {
assign_cpu_kernel<data_t>(
grad_out.data<data_t>(),
grad_x.mutable_data<data_t>(),
grad_out.size());
}));
if (int_attr != 10) {
throw std::runtime_error("int_attr value error.");
}
if (float_vec_attr.size() != 3) {
throw std::runtime_error("float_vec_attr size error.");
} else {
for (auto& value : float_vec_attr) {
if (std::abs(value - 3.14) > 1e-6) {
throw std::runtime_error("float_vec_attr value error.");
}
}
}
if (str_vec_attr.size() != 3) {
throw std::runtime_error("str_vec_attr size error.");
} else {
for (auto& value : str_vec_attr) {
if (value != "StrAttr") {
throw std::runtime_error("str_vec_attr value error.");
}
}
}
return {grad_x};
}
std::vector<std::vector<int64_t>> InferShape(std::vector<int64_t> x_shape) {
return {x_shape};
}
std::vector<paddle::DataType> InferDType(paddle::DataType x_dtype) {
return {x_dtype};
}
PD_BUILD_OP("attr_test")
.Inputs({"X"})
.Outputs({"Out"})
.Attrs({"bool_attr: bool",
"int_attr: int",
"float_attr: float",
"int64_attr: int64_t",
"str_attr: std::string",
"int_vec_attr: std::vector<int>",
"float_vec_attr: std::vector<float>",
"int64_vec_attr: std::vector<int64_t>",
"str_vec_attr: std::vector<std::string>"})
.SetKernelFn(PD_KERNEL(AttrTestForward))
.SetInferShapeFn(PD_INFER_SHAPE(InferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(InferDType))
.SetBackwardOp("attr_test_grad")
.Inputs({paddle::Grad("Out")})
.Outputs({paddle::Grad("X")})
.Attrs({"int_attr: int",
"float_vec_attr: std::vector<float>",
"str_vec_attr: std::vector<std::string>"})
.SetKernelFn(PD_KERNEL(AttrTestBackward));
python/paddle/fluid/tests/custom_op/relu_op_simple.cc → python/paddle/fluid/tests/custom_op/custom_relu_op.cc
浏览文件 @ d3e60959
...@@ -17,13 +17,6 @@ ...@@ -17,13 +17,6 @@
#include "paddle/extension.h" #include "paddle/extension.h"
template <typename data_t>
void fill_constant_cpu_kernel(data_t* out_data, int64_t x_numel, data_t value) {
for (int i = 0; i < x_numel; ++i) {
out_data[i] = value;
}
}
template <typename data_t> template <typename data_t>
void relu_cpu_forward_kernel(const data_t* x_data, void relu_cpu_forward_kernel(const data_t* x_data,
data_t* out_data, data_t* out_data,
...@@ -53,21 +46,8 @@ std::vector<paddle::Tensor> relu_cpu_forward(const paddle::Tensor& x) { ...@@ -53,21 +46,8 @@ std::vector<paddle::Tensor> relu_cpu_forward(const paddle::Tensor& x) {
relu_cpu_forward_kernel<data_t>( relu_cpu_forward_kernel<data_t>(
x.data<data_t>(), out.mutable_data<data_t>(x.place()), x.size()); x.data<data_t>(), out.mutable_data<data_t>(x.place()), x.size());
})); }));
// fake multi output: Fake_float64 with float64 dtype
auto fake_float64 = paddle::Tensor(paddle::PlaceType::kCPU);
fake_float64.reshape(x.shape());
fill_constant_cpu_kernel<double>(
fake_float64.mutable_data<double>(x.place()), x.size(), 0.);
// fake multi output: ZFake_int32 with int32 dtype
auto zfake_int32 = paddle::Tensor(paddle::PlaceType::kCPU);
zfake_int32.reshape(x.shape());
fill_constant_cpu_kernel<int32_t>(
zfake_int32.mutable_data<int32_t>(x.place()), x.size(), 1);
return {out, fake_float64, zfake_int32}; return {out};
} }
std::vector<paddle::Tensor> relu_cpu_backward(const paddle::Tensor& x, std::vector<paddle::Tensor> relu_cpu_backward(const paddle::Tensor& x,
...@@ -117,16 +97,16 @@ std::vector<paddle::Tensor> ReluBackward(const paddle::Tensor& x, ...@@ -117,16 +97,16 @@ std::vector<paddle::Tensor> ReluBackward(const paddle::Tensor& x,
} }
std::vector<std::vector<int64_t>> ReluInferShape(std::vector<int64_t> x_shape) { std::vector<std::vector<int64_t>> ReluInferShape(std::vector<int64_t> x_shape) {
return {x_shape, x_shape, x_shape}; return {x_shape};
} }
std::vector<paddle::DataType> ReluInferDType(paddle::DataType x_dtype) { std::vector<paddle::DataType> ReluInferDType(paddle::DataType x_dtype) {
return {x_dtype, paddle::DataType::FLOAT64, paddle::DataType::INT32}; return {x_dtype};
} }
PD_BUILD_OP("relu2") PD_BUILD_OP("custom_relu")
.Inputs({"X"}) .Inputs({"X"})
.Outputs({"Out", "Fake_float64", "ZFake_int32"}) .Outputs({"Out"})
.SetKernelFn(PD_KERNEL(ReluForward)) .SetKernelFn(PD_KERNEL(ReluForward))
.SetInferShapeFn(PD_INFER_SHAPE(ReluInferShape)) .SetInferShapeFn(PD_INFER_SHAPE(ReluInferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(ReluInferDType)) .SetInferDtypeFn(PD_INFER_DTYPE(ReluInferDType))
......
python/paddle/fluid/tests/custom_op/relu_op_simple.cu → python/paddle/fluid/tests/custom_op/custom_relu_op.cu
浏览文件 @ d3e60959
...@@ -14,16 +14,6 @@ ...@@ -14,16 +14,6 @@
#include "paddle/extension.h" #include "paddle/extension.h"
template <typename data_t>
__global__ void fill_constant_cuda_kernel(data_t* y,
const int num,
data_t value) {
int gid = blockIdx.x * blockDim.x + threadIdx.x;
for (int i = gid; i < num; i += blockDim.x * gridDim.x) {
y[i] = value;
}
}
template <typename data_t> template <typename data_t>
__global__ void relu_cuda_forward_kernel(const data_t* x, __global__ void relu_cuda_forward_kernel(const data_t* x,
data_t* y, data_t* y,
...@@ -57,18 +47,8 @@ std::vector<paddle::Tensor> relu_cuda_forward(const paddle::Tensor& x) { ...@@ -57,18 +47,8 @@ std::vector<paddle::Tensor> relu_cuda_forward(const paddle::Tensor& x) {
relu_cuda_forward_kernel<data_t><<<grid, block>>>( relu_cuda_forward_kernel<data_t><<<grid, block>>>(
x.data<data_t>(), out.mutable_data<data_t>(x.place()), numel); x.data<data_t>(), out.mutable_data<data_t>(x.place()), numel);
})); }));
// fake multi output: Fake_1
auto fake_float64 = paddle::Tensor(paddle::PlaceType::kGPU);
fake_float64.reshape(x.shape());
fill_constant_cuda_kernel<double><<<grid, block>>>(
fake_float64.mutable_data<double>(x.place()), numel, 0.);
// fake multi output: ZFake_1
auto zfake_int32 = paddle::Tensor(paddle::PlaceType::kGPU);
zfake_int32.reshape(x.shape());
fill_constant_cuda_kernel<int32_t><<<grid, block>>>(
zfake_int32.mutable_data<int32_t>(x.place()), numel, 1);
return {out, fake_float64, zfake_int32}; return {out};
} }
std::vector<paddle::Tensor> relu_cuda_backward(const paddle::Tensor& x, std::vector<paddle::Tensor> relu_cuda_backward(const paddle::Tensor& x,
......
python/paddle/fluid/tests/custom_op/relu_op3_simple.cc → python/paddle/fluid/tests/custom_op/custom_relu_op_dup.cc
浏览文件 @ d3e60959
...@@ -29,11 +29,11 @@ std::vector<std::vector<int64_t>> ReluInferShape(std::vector<int64_t> x_shape); ...@@ -29,11 +29,11 @@ std::vector<std::vector<int64_t>> ReluInferShape(std::vector<int64_t> x_shape);
std::vector<paddle::DataType> ReluInferDType(paddle::DataType x_dtype); std::vector<paddle::DataType> ReluInferDType(paddle::DataType x_dtype);
// Reuse codes in `relu_op_simple.cc/cu` to register another custom operator // Reuse codes in `custom_relu_op.cc/cu` to register another custom operator
// to test jointly compile multi operators at same time. // to test jointly compile multi operators at same time.
PD_BUILD_OP("relu3") PD_BUILD_OP("custom_relu_dup")
.Inputs({"X"}) .Inputs({"X"})
.Outputs({"Out", "Fake_float64", "ZFake_int32"}) .Outputs({"Out"})
.SetKernelFn(PD_KERNEL(ReluForward)) .SetKernelFn(PD_KERNEL(ReluForward))
.SetInferShapeFn(PD_INFER_SHAPE(ReluInferShape)) .SetInferShapeFn(PD_INFER_SHAPE(ReluInferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(ReluInferDType)) .SetInferDtypeFn(PD_INFER_DTYPE(ReluInferDType))
......
python/paddle/fluid/tests/custom_op/setup_install_simple.py → python/paddle/fluid/tests/custom_op/custom_relu_setup.py
浏览文件 @ d3e60959
...@@ -17,11 +17,14 @@ import os ...@@ -17,11 +17,14 @@ import os
from utils import paddle_includes, extra_compile_args from utils import paddle_includes, extra_compile_args
from paddle.utils.cpp_extension import CUDAExtension, setup from paddle.utils.cpp_extension import CUDAExtension, setup
# custom_relu_op_dup.cc is only used for multi ops test,
# not a new op, if you want to test only one op, remove this
# source file
setup( setup(
name='simple_setup_relu2', name='custom_relu_module_setup',
ext_modules=CUDAExtension( # test for not specific name here. ext_modules=CUDAExtension( # test for not specific name here.
sources=[ sources=[
'relu_op_simple.cc', 'relu_op_simple.cu', 'relu_op3_simple.cc' 'custom_relu_op.cc', 'custom_relu_op.cu', 'custom_relu_op_dup.cc'
], # test for multi ops ], # test for multi ops
include_dirs=paddle_includes, include_dirs=paddle_includes,
extra_compile_args=extra_compile_args)) extra_compile_args=extra_compile_args))
python/paddle/fluid/tests/custom_op/multi_out_test_op.cc 0 → 100644
浏览文件 @ d3e60959
// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <iostream>
#include <vector>
#include "paddle/extension.h"
template <typename data_t>
void assign_cpu_kernel(const data_t* x_data,
data_t* out_data,
int64_t x_numel) {
for (int i = 0; i < x_numel; ++i) {
out_data[i] = x_data[i];
}
}
template <typename data_t>
void fill_constant_cpu_kernel(data_t* out_data, int64_t x_numel, data_t value) {
for (int i = 0; i < x_numel; ++i) {
out_data[i] = value;
}
}
std::vector<paddle::Tensor> MultiOutCPU(const paddle::Tensor& x) {
auto out = paddle::Tensor(paddle::PlaceType::kCPU);
out.reshape(x.shape());
PD_DISPATCH_FLOATING_TYPES(
x.type(), "assign_cpu_kernel", ([&] {
assign_cpu_kernel<data_t>(
x.data<data_t>(), out.mutable_data<data_t>(x.place()), x.size());
}));
// fake multi output: Fake_float64 with float64 dtype
auto fake_float64 = paddle::Tensor(paddle::PlaceType::kCPU);
fake_float64.reshape(x.shape());
fill_constant_cpu_kernel<double>(
fake_float64.mutable_data<double>(x.place()), x.size(), 0.);
// fake multi output: ZFake_int32 with int32 dtype
auto zfake_int32 = paddle::Tensor(paddle::PlaceType::kCPU);
zfake_int32.reshape(x.shape());
fill_constant_cpu_kernel<int32_t>(
zfake_int32.mutable_data<int32_t>(x.place()), x.size(), 1);
return {out, fake_float64, zfake_int32};
}
std::vector<std::vector<int64_t>> InferShape(std::vector<int64_t> x_shape) {
return {x_shape, x_shape, x_shape};
}
std::vector<paddle::DataType> InferDtype(paddle::DataType x_dtype) {
return {x_dtype, paddle::DataType::FLOAT64, paddle::DataType::INT32};
}
PD_BUILD_OP("multi_out")
.Inputs({"X"})
.Outputs({"Out", "Fake_float64", "ZFake_int32"})
.SetKernelFn(PD_KERNEL(MultiOutCPU))
.SetInferShapeFn(PD_INFER_SHAPE(InferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(InferDtype));
python/paddle/fluid/tests/custom_op/test_custom_attrs_jit.py 0 → 100644
浏览文件 @ d3e60959
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import unittest
import numpy as np
import paddle
from paddle.utils.cpp_extension import load, get_build_directory
from utils import paddle_includes, extra_compile_args
from paddle.utils.cpp_extension.extension_utils import run_cmd
# Because Windows don't use docker, the shared lib already exists in the
# cache dir, it will not be compiled again unless the shared lib is removed.
file = '{}\\custom_attrs_jit\\custom_attrs_jit.pyd'.format(get_build_directory(
))
if os.name == 'nt' and os.path.isfile(file):
cmd = 'del {}'.format(file)
run_cmd(cmd, True)
# Compile and load custom op Just-In-Time.
custom_attrs = load(
name='custom_attrs_jit',
sources=['attr_test_op.cc'],
extra_include_paths=paddle_includes, # add for Coverage CI
extra_cxx_cflags=extra_compile_args, # add for Coverage CI
verbose=True)
class TestJitCustomAttrs(unittest.TestCase):
def test_attr_value(self):
paddle.set_device('cpu')
# prepare test value
bool_attr = True
int_attr = 10
float_attr = 3.14
int64_attr = 10000000000
str_attr = "StrAttr"
int_vec_attr = [10, 10, 10]
float_vec_attr = [3.14, 3.14, 3.14]
int64_vec_attr = [10000000000, 10000000000, 10000000000]
str_vec_attr = ["StrAttr", "StrAttr", "StrAttr"]
x = paddle.ones([2, 2], dtype='float32')
x.stop_gradient = False
out = custom_attrs.attr_test(
x, bool_attr, int_attr, float_attr, int64_attr, str_attr,
int_vec_attr, float_vec_attr, int64_vec_attr, str_vec_attr)
out.stop_gradient = False
out.backward()
self.assertTrue(np.array_equal(x.numpy(), out.numpy()))
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/custom_op/test_custom_relu_op_jit.py 0 → 100644
浏览文件 @ d3e60959
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import subprocess
import unittest
import paddle
import numpy as np
from paddle.utils.cpp_extension import load, get_build_directory
from paddle.utils.cpp_extension.extension_utils import run_cmd
from utils import paddle_includes, extra_compile_args
from test_custom_relu_op_setup import custom_relu_dynamic, custom_relu_static
# Because Windows don't use docker, the shared lib already exists in the
# cache dir, it will not be compiled again unless the shared lib is removed.
file = '{}\\custom_relu_module_jit\\custom_relu_module_jit.pyd'.format(
get_build_directory())
if os.name == 'nt' and os.path.isfile(file):
cmd = 'del {}'.format(file)
run_cmd(cmd, True)
# Compile and load custom op Just-In-Time.
# custom_relu_op_dup.cc is only used for multi ops test,
# not a new op, if you want to test only one op, remove this
# source file
custom_module = load(
name='custom_relu_module_jit',
sources=[
'custom_relu_op.cc', 'custom_relu_op.cu', 'custom_relu_op_dup.cc'
],
extra_include_paths=paddle_includes, # add for Coverage CI
extra_cxx_cflags=extra_compile_args, # add for Coverage CI
extra_cuda_cflags=extra_compile_args, # add for Coverage CI
verbose=True)
class TestJITLoad(unittest.TestCase):
def setUp(self):
self.custom_ops = [
custom_module.custom_relu, custom_module.custom_relu_dup
]
self.dtypes = ['float32', 'float64']
self.devices = ['cpu', 'gpu']
def test_static(self):
for device in self.devices:
for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops:
out = custom_relu_static(custom_op, device, dtype, x)
pd_out = custom_relu_static(custom_op, device, dtype, x,
False)
self.assertTrue(
np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format(
out, pd_out))
def test_dynamic(self):
for device in self.devices:
for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops:
out, x_grad = custom_relu_dynamic(custom_op, device, dtype,
x)
pd_out, pd_x_grad = custom_relu_dynamic(custom_op, device,
dtype, x, False)
self.assertTrue(
np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format(
out, pd_out))
self.assertTrue(
np.array_equal(x_grad, pd_x_grad),
"custom op x grad: {},\n paddle api x grad: {}".format(
x_grad, pd_x_grad))
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/custom_op/test_simple_custom_op_setup.py → python/paddle/fluid/tests/custom_op/test_custom_relu_op_setup.py
浏览文件 @ d3e60959
...@@ -23,13 +23,13 @@ import numpy as np ...@@ -23,13 +23,13 @@ import numpy as np
from paddle.utils.cpp_extension.extension_utils import run_cmd from paddle.utils.cpp_extension.extension_utils import run_cmd
def relu2_dynamic(func, device, dtype, np_x, use_func=True): def custom_relu_dynamic(func, device, dtype, np_x, use_func=True):
paddle.set_device(device) paddle.set_device(device)
t = paddle.to_tensor(np_x) t = paddle.to_tensor(np_x)
t.stop_gradient = False t.stop_gradient = False
out = func(t)[0] if use_func else paddle.nn.functional.relu(t) out = func(t) if use_func else paddle.nn.functional.relu(t)
out.stop_gradient = False out.stop_gradient = False
out.backward() out.backward()
...@@ -37,7 +37,12 @@ def relu2_dynamic(func, device, dtype, np_x, use_func=True): ...@@ -37,7 +37,12 @@ def relu2_dynamic(func, device, dtype, np_x, use_func=True):
return out.numpy(), t.grad return out.numpy(), t.grad
def relu2_static(func, device, dtype, np_x, use_func=True): def custom_relu_static(func,
device,
dtype,
np_x,
use_func=True,
test_infer=False):
paddle.enable_static() paddle.enable_static()
paddle.set_device(device) paddle.set_device(device)
...@@ -45,8 +50,7 @@ def relu2_static(func, device, dtype, np_x, use_func=True): ...@@ -45,8 +50,7 @@ def relu2_static(func, device, dtype, np_x, use_func=True):
with static.program_guard(static.Program()): with static.program_guard(static.Program()):
x = static.data(name='X', shape=[None, 8], dtype=dtype) x = static.data(name='X', shape=[None, 8], dtype=dtype)
x.stop_gradient = False x.stop_gradient = False
# out, fake_float64, fake_int32 out = func(x) if use_func else paddle.nn.functional.relu(x)
out = func(x)[0] if use_func else paddle.nn.functional.relu(x)
static.append_backward(out) static.append_backward(out)
exe = static.Executor() exe = static.Executor()
...@@ -60,7 +64,7 @@ def relu2_static(func, device, dtype, np_x, use_func=True): ...@@ -60,7 +64,7 @@ def relu2_static(func, device, dtype, np_x, use_func=True):
return out_v return out_v
def relu2_static_pe(func, device, dtype, np_x, use_func=True): def custom_relu_static_pe(func, device, dtype, np_x, use_func=True):
paddle.enable_static() paddle.enable_static()
paddle.set_device(device) paddle.set_device(device)
...@@ -69,7 +73,7 @@ def relu2_static_pe(func, device, dtype, np_x, use_func=True): ...@@ -69,7 +73,7 @@ def relu2_static_pe(func, device, dtype, np_x, use_func=True):
with static.program_guard(static.Program()): with static.program_guard(static.Program()):
x = static.data(name='X', shape=[None, 8], dtype=dtype) x = static.data(name='X', shape=[None, 8], dtype=dtype)
x.stop_gradient = False x.stop_gradient = False
out = func(x)[0] if use_func else paddle.nn.functional.relu(x) out = func(x) if use_func else paddle.nn.functional.relu(x)
static.append_backward(out) static.append_backward(out)
exe = static.Executor() exe = static.Executor()
...@@ -87,11 +91,58 @@ def relu2_static_pe(func, device, dtype, np_x, use_func=True): ...@@ -87,11 +91,58 @@ def relu2_static_pe(func, device, dtype, np_x, use_func=True):
return out_v return out_v
def custom_relu_static_inference(func, device, np_data, np_label, path_prefix):
paddle.set_device(device)
with static.scope_guard(static.Scope()):
with static.program_guard(static.Program()):
# simple module
data = static.data(
name='data', shape=[None, 1, 28, 28], dtype='float32')
label = static.data(name='label', shape=[None, 1], dtype='int64')
hidden = static.nn.fc(data, size=128)
hidden = func(hidden)
hidden = static.nn.fc(hidden, size=128)
predict = static.nn.fc(hidden, size=10, activation='softmax')
loss = paddle.nn.functional.cross_entropy(input=hidden, label=label)
avg_loss = paddle.mean(loss)
opt = paddle.optimizer.SGD(learning_rate=0.1)
opt.minimize(avg_loss)
# run start up model
exe = static.Executor()
exe.run(static.default_startup_program())
# train
for i in range(4):
avg_loss_v = exe.run(static.default_main_program(),
feed={'data': np_data,
'label': np_label},
fetch_list=[avg_loss])
# save inference model
static.save_inference_model(path_prefix, [data], [predict], exe)
# get train predict value
predict_v = exe.run(static.default_main_program(),
feed={'data': np_data,
'label': np_label},
fetch_list=[predict])
return predict_v
class TestNewCustomOpSetUpInstall(unittest.TestCase): class TestNewCustomOpSetUpInstall(unittest.TestCase):
def setUp(self): def setUp(self):
cur_dir = os.path.dirname(os.path.abspath(__file__)) cur_dir = os.path.dirname(os.path.abspath(__file__))
# compile, install the custom op egg into site-packages under background # compile, install the custom op egg into site-packages under background
cmd = 'cd {} && python setup_install_simple.py install'.format(cur_dir) if os.name == 'nt':
cmd = 'cd /d {} && python custom_relu_setup.py install'.format(
cur_dir)
else:
cmd = 'cd {} && python custom_relu_setup.py install'.format(cur_dir)
run_cmd(cmd) run_cmd(cmd)
# NOTE(Aurelius84): Normally, it's no need to add following codes for users. # NOTE(Aurelius84): Normally, it's no need to add following codes for users.
...@@ -99,28 +150,42 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase): ...@@ -99,28 +150,42 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase):
# sys.path has been updated. So we update it manually. # sys.path has been updated. So we update it manually.
# See: https://stackoverflow.com/questions/56974185/import-runtime-installed-module-using-pip-in-python-3 # See: https://stackoverflow.com/questions/56974185/import-runtime-installed-module-using-pip-in-python-3
if os.name == 'nt':
# NOTE(zhouwei25): getsitepackages on windows will return a list: [python install dir, site packages dir]
site_dir = site.getsitepackages()[1]
else:
site_dir = site.getsitepackages()[0] site_dir = site.getsitepackages()[0]
custom_egg_path = [ custom_egg_path = [
x for x in os.listdir(site_dir) if 'simple_setup_relu2' in x x for x in os.listdir(site_dir) if 'custom_relu_module_setup' in x
] ]
assert len(custom_egg_path) == 1, "Matched egg number is %d." % len( assert len(custom_egg_path) == 1, "Matched egg number is %d." % len(
custom_egg_path) custom_egg_path)
sys.path.append(os.path.join(site_dir, custom_egg_path[0])) sys.path.append(os.path.join(site_dir, custom_egg_path[0]))
# usage: import the package directly # usage: import the package directly
import simple_setup_relu2 import custom_relu_module_setup
self.custom_ops = [simple_setup_relu2.relu2, simple_setup_relu2.relu3] # `custom_relu_dup` is same as `custom_relu_dup`
self.custom_ops = [
custom_relu_module_setup.custom_relu,
custom_relu_module_setup.custom_relu_dup
]
self.dtypes = ['float32', 'float64'] self.dtypes = ['float32', 'float64']
self.devices = ['cpu', 'gpu'] self.devices = ['cpu', 'gpu']
# config seed
SEED = 2021
paddle.seed(SEED)
paddle.framework.random._manual_program_seed(SEED)
def test_static(self): def test_static(self):
for device in self.devices: for device in self.devices:
for dtype in self.dtypes: for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype) x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops: for custom_op in self.custom_ops:
out = relu2_static(custom_op, device, dtype, x) out = custom_relu_static(custom_op, device, dtype, x)
pd_out = relu2_static(custom_op, device, dtype, x, False) pd_out = custom_relu_static(custom_op, device, dtype, x,
False)
self.assertTrue( self.assertTrue(
np.array_equal(out, pd_out), np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format( "custom op out: {},\n paddle api out: {}".format(
...@@ -131,8 +196,9 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase): ...@@ -131,8 +196,9 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase):
for dtype in self.dtypes: for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype) x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops: for custom_op in self.custom_ops:
out = relu2_static_pe(custom_op, device, dtype, x) out = custom_relu_static_pe(custom_op, device, dtype, x)
pd_out = relu2_static_pe(custom_op, device, dtype, x, False) pd_out = custom_relu_static_pe(custom_op, device, dtype, x,
False)
self.assertTrue( self.assertTrue(
np.array_equal(out, pd_out), np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format( "custom op out: {},\n paddle api out: {}".format(
...@@ -143,9 +209,10 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase): ...@@ -143,9 +209,10 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase):
for dtype in self.dtypes: for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype) x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops: for custom_op in self.custom_ops:
out, x_grad = relu2_dynamic(custom_op, device, dtype, x) out, x_grad = custom_relu_dynamic(custom_op, device, dtype,
pd_out, pd_x_grad = relu2_dynamic(custom_op, device, dtype, x)
x, False) pd_out, pd_x_grad = custom_relu_dynamic(custom_op, device,
dtype, x, False)
self.assertTrue( self.assertTrue(
np.array_equal(out, pd_out), np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format( "custom op out: {},\n paddle api out: {}".format(
...@@ -155,6 +222,28 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase): ...@@ -155,6 +222,28 @@ class TestNewCustomOpSetUpInstall(unittest.TestCase):
"custom op x grad: {},\n paddle api x grad: {}".format( "custom op x grad: {},\n paddle api x grad: {}".format(
x_grad, pd_x_grad)) x_grad, pd_x_grad))
def test_static_save_and_load_inference_model(self):
paddle.enable_static()
np_data = np.random.random((1, 1, 28, 28)).astype("float32")
np_label = np.random.random((1, 1)).astype("int64")
path_prefix = "custom_op_inference/custom_relu"
for device in self.devices:
predict = custom_relu_static_inference(
self.custom_ops[0], device, np_data, np_label, path_prefix)
# load inference model
with static.scope_guard(static.Scope()):
exe = static.Executor()
[inference_program, feed_target_names,
fetch_targets] = static.load_inference_model(path_prefix, exe)
predict_infer = exe.run(inference_program,
feed={feed_target_names[0]: np_data},
fetch_list=fetch_targets)
self.assertTrue(
np.array_equal(predict, predict_infer),
"custom op predict: {},\n custom op infer predict: {}".
format(predict, predict_infer))
paddle.disable_static()
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()
python/paddle/fluid/tests/custom_op/test_dispatch.py → python/paddle/fluid/tests/custom_op/test_dispatch_jit.py
浏览文件 @ d3e60959
...@@ -16,14 +16,23 @@ import os ...@@ -16,14 +16,23 @@ import os
import unittest import unittest
import paddle import paddle
import numpy as np import numpy as np
from paddle.utils.cpp_extension import load from paddle.utils.cpp_extension import load, get_build_directory
from utils import paddle_includes, extra_compile_args from utils import paddle_includes, extra_compile_args
from paddle.utils.cpp_extension.extension_utils import run_cmd
# Because Windows don't use docker, the shared lib already exists in the
# cache dir, it will not be compiled again unless the shared lib is removed.
file = '{}\\dispatch_op\\dispatch_op.pyd'.format(get_build_directory())
if os.name == 'nt' and os.path.isfile(file):
cmd = 'del {}'.format(file)
run_cmd(cmd, True)
dispatch_op = load( dispatch_op = load(
name='dispatch_op', name='dispatch_op',
sources=['dispatch_test_op.cc'], sources=['dispatch_test_op.cc'],
extra_include_paths=paddle_includes, # add for Coverage CI extra_include_paths=paddle_includes, # add for Coverage CI
extra_cflags=extra_compile_args) # add for Coverage CI extra_cxx_cflags=extra_compile_args,
verbose=True)
class TestJitDispatch(unittest.TestCase): class TestJitDispatch(unittest.TestCase):
......
python/paddle/fluid/tests/custom_op/test_jit_load.py
浏览文件 @ d3e60959
...@@ -29,7 +29,8 @@ custom_module = load( ...@@ -29,7 +29,8 @@ custom_module = load(
sources=['relu_op.cc', 'relu_op.cu', 'relu_op3.cc', 'relu_op3.cu'], sources=['relu_op.cc', 'relu_op.cu', 'relu_op3.cc', 'relu_op3.cu'],
interpreter='python', # add for unittest interpreter='python', # add for unittest
extra_include_paths=paddle_includes, # add for Coverage CI extra_include_paths=paddle_includes, # add for Coverage CI
extra_cflags=extra_compile_args, # add for Coverage CI extra_cxx_cflags=extra_compile_args, # add for Coverage CI,
extra_cuda_cflags=extra_compile_args, # add for split cpp/cuda flags
verbose=True # add for unittest verbose=True # add for unittest
) )
......
python/paddle/fluid/tests/custom_op/test_simple_custom_op_jit.py → python/paddle/fluid/tests/custom_op/test_multi_out_jit.py
浏览文件 @ d3e60959
...@@ -13,81 +13,54 @@ ...@@ -13,81 +13,54 @@
# limitations under the License. # limitations under the License.
import os import os
import subprocess
import unittest import unittest
import paddle
import numpy as np import numpy as np
import paddle
from paddle.utils.cpp_extension import load from paddle.utils.cpp_extension import load
from paddle.utils.cpp_extension import load, get_build_directory
from paddle.utils.cpp_extension.extension_utils import run_cmd
from utils import paddle_includes, extra_compile_args from utils import paddle_includes, extra_compile_args
from test_simple_custom_op_setup import relu2_dynamic, relu2_static
# Because Windows don't use docker, the shared lib already exists in the
# cache dir, it will not be compiled again unless the shared lib is removed.
file = '{}\\multi_out_jit\\multi_out_jit.pyd'.format(get_build_directory())
if os.name == 'nt' and os.path.isfile(file):
cmd = 'del {}'.format(file)
run_cmd(cmd, True)
# Compile and load custom op Just-In-Time. # Compile and load custom op Just-In-Time.
custom_module = load( multi_out_module = load(
name='simple_jit_relu2', name='multi_out_jit',
sources=['relu_op_simple.cc', 'relu_op_simple.cu', 'relu_op3_simple.cc'], sources=['multi_out_test_op.cc'],
extra_include_paths=paddle_includes, # add for Coverage CI extra_include_paths=paddle_includes, # add for Coverage CI
extra_cflags=extra_compile_args) # add for Coverage CI extra_cxx_cflags=extra_compile_args, # add for Coverage CI
verbose=True)
class TestJITLoad(unittest.TestCase):
def setUp(self):
self.custom_ops = [custom_module.relu2, custom_module.relu3]
self.dtypes = ['float32', 'float64']
self.devices = ['cpu', 'gpu']
def test_static(self):
for device in self.devices:
for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops:
out = relu2_static(custom_op, device, dtype, x)
pd_out = relu2_static(custom_op, device, dtype, x, False)
self.assertTrue(
np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format(
out, pd_out))
def test_dynamic(self):
for device in self.devices:
for dtype in self.dtypes:
x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
for custom_op in self.custom_ops:
out, x_grad = relu2_dynamic(custom_op, device, dtype, x)
pd_out, pd_x_grad = relu2_dynamic(custom_op, device, dtype,
x, False)
self.assertTrue(
np.array_equal(out, pd_out),
"custom op out: {},\n paddle api out: {}".format(
out, pd_out))
self.assertTrue(
np.array_equal(x_grad, pd_x_grad),
"custom op x grad: {},\n paddle api x grad: {}".format(
x_grad, pd_x_grad))
class TestMultiOutputDtypes(unittest.TestCase): class TestMultiOutputDtypes(unittest.TestCase):
def setUp(self): def setUp(self):
self.custom_op = custom_module.relu2 self.custom_op = multi_out_module.multi_out
self.dtypes = ['float32', 'float64'] self.dtypes = ['float32', 'float64']
self.devices = ['cpu', 'gpu'] self.devices = ['cpu']
def test_static(self):
paddle.enable_static()
for device in self.devices:
for dtype in self.dtypes:
res = self.run_static(device, dtype)
self.check_multi_outputs(res)
paddle.disable_static()
def test_dynamic(self): def run_static(self, device, dtype):
for device in self.devices:
for dtype in self.dtypes:
paddle.set_device(device) paddle.set_device(device)
x_data = np.random.uniform(-1, 1, [4, 8]).astype(dtype) x_data = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
x = paddle.to_tensor(x_data)
with paddle.static.scope_guard(paddle.static.Scope()):
with paddle.static.program_guard(paddle.static.Program()):
x = paddle.static.data(name='X', shape=[None, 8], dtype=dtype)
outs = self.custom_op(x) outs = self.custom_op(x)
self.assertTrue(len(outs) == 3) exe = paddle.static.Executor()
self.check_multi_outputs(outs, True) exe.run(paddle.static.default_startup_program())
res = exe.run(paddle.static.default_main_program(),
feed={'X': x_data},
fetch_list=outs)
return res
def check_multi_outputs(self, outs, is_dynamic=False): def check_multi_outputs(self, outs, is_dynamic=False):
out, zero_float64, one_int32 = outs out, zero_float64, one_int32 = outs
...@@ -103,22 +76,24 @@ class TestMultiOutputDtypes(unittest.TestCase): ...@@ -103,22 +76,24 @@ class TestMultiOutputDtypes(unittest.TestCase):
self.assertTrue( self.assertTrue(
np.array_equal(one_int32, np.ones([4, 8]).astype('int32'))) np.array_equal(one_int32, np.ones([4, 8]).astype('int32')))
def run_static(self, device, dtype): def test_static(self):
paddle.enable_static()
for device in self.devices:
for dtype in self.dtypes:
res = self.run_static(device, dtype)
self.check_multi_outputs(res)
paddle.disable_static()
def test_dynamic(self):
for device in self.devices:
for dtype in self.dtypes:
paddle.set_device(device) paddle.set_device(device)
x_data = np.random.uniform(-1, 1, [4, 8]).astype(dtype) x_data = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
x = paddle.to_tensor(x_data)
with paddle.static.scope_guard(paddle.static.Scope()):
with paddle.static.program_guard(paddle.static.Program()):
x = paddle.static.data(name='X', shape=[None, 8], dtype=dtype)
outs = self.custom_op(x) outs = self.custom_op(x)
exe = paddle.static.Executor() self.assertTrue(len(outs) == 3)
exe.run(paddle.static.default_startup_program()) self.check_multi_outputs(outs, True)
res = exe.run(paddle.static.default_main_program(),
feed={'X': x_data},
fetch_list=outs)
return res
if __name__ == '__main__': if __name__ == '__main__':
......
python/paddle/fluid/tests/custom_op/utils.py
浏览文件 @ d3e60959
...@@ -23,8 +23,8 @@ site_packages_path = get_python_lib() ...@@ -23,8 +23,8 @@ site_packages_path = get_python_lib()
# paddle include directory. Because the following path is generated after insalling # paddle include directory. Because the following path is generated after insalling
# PaddlePaddle whl. So here we specific `include_dirs` to avoid errors in CI. # PaddlePaddle whl. So here we specific `include_dirs` to avoid errors in CI.
paddle_includes = [ paddle_includes = [
os.path.join(site_packages_path, 'paddle/include'), os.path.join(site_packages_path, 'paddle', 'include'),
os.path.join(site_packages_path, 'paddle/include/third_party') os.path.join(site_packages_path, 'paddle', 'include', 'third_party')
] ]
# TODO(Aurelius84): Memory layout is different if build paddle with PADDLE_WITH_MKLDNN=ON, # TODO(Aurelius84): Memory layout is different if build paddle with PADDLE_WITH_MKLDNN=ON,
......
python/paddle/utils/cpp_extension/__init__.py
浏览文件 @ d3e60959
...@@ -25,6 +25,5 @@ from . import cpp_extension ...@@ -25,6 +25,5 @@ from . import cpp_extension
from . import extension_utils from . import extension_utils
__all__ = [ __all__ = [
'CppExtension', 'CUDAExtension', 'BuildExtension', 'load', 'setup', 'CppExtension', 'CUDAExtension', 'load', 'setup', 'get_build_directory'
'get_build_directory'
] ]
python/paddle/utils/cpp_extension/cpp_extension.py
浏览文件 @ d3e60959
...@@ -14,47 +14,124 @@ ...@@ -14,47 +14,124 @@
import os import os
import six import six
import sys
import textwrap
import copy import copy
import re
import setuptools import setuptools
from setuptools.command.easy_install import easy_install from setuptools.command.easy_install import easy_install
from setuptools.command.build_ext import build_ext from setuptools.command.build_ext import build_ext
from distutils.command.build import build
from .extension_utils import find_cuda_home, normalize_extension_kwargs, add_compile_flag, bootstrap_context from .extension_utils import find_cuda_home, normalize_extension_kwargs, add_compile_flag, bootstrap_context
from .extension_utils import is_cuda_file, prepare_unix_cflags, add_std_without_repeat, get_build_directory from .extension_utils import is_cuda_file, prepare_unix_cudaflags, prepare_win_cudaflags, add_std_without_repeat, get_build_directory
from .extension_utils import _import_module_from_library, CustomOpInfo, _write_setup_file, _jit_compile, parse_op_name_from from .extension_utils import _import_module_from_library, CustomOpInfo, _write_setup_file, _jit_compile, parse_op_name_from
from .extension_utils import check_abi_compatibility, log_v, IS_WINDOWS from .extension_utils import check_abi_compatibility, log_v, IS_WINDOWS, OS_NAME
from .extension_utils import use_new_custom_op_load_method from .extension_utils import use_new_custom_op_load_method, MSVC_COMPILE_FLAGS
# Note(zhouwei): On windows, it will export function 'PyInit_[name]' by default,
# The solution is: 1.User add function PyInit_[name] 2. set not to export
# refer to https://stackoverflow.com/questions/34689210/error-exporting-symbol-when-building-python-c-extension-in-windows
if IS_WINDOWS and six.PY3:
from distutils.command.build_ext import build_ext as _du_build_ext
from unittest.mock import Mock
_du_build_ext.get_export_symbols = Mock(return_value=None)
CUDA_HOME = find_cuda_home() CUDA_HOME = find_cuda_home()
def setup(**attr): def setup(**attr):
""" """
Wrapper setuptools.setup function to valid `build_ext` command and The interface is used to config the process of compiling customized operators,
implement paddle api code injection by switching `write_stub` mainly includes how to complile shared library, automatically generate python API
function in bdist_egg with `custom_write_stub`. and install it into site-package. It supports using customized operators directly with
``import`` statement.
It encapsulates the python built-in ``setuptools.setup`` function and keeps arguments
and usage same as the native interface. Meanwhile, it hiddens Paddle inner framework
concepts, such as necessary compiling flags, included paths of head files, and linking
flags. It also will automatically search and valid local enviromment and versions of ``cc`` and
``nvcc`` , then compiles customized operators supporting CPU or GPU device according to
the specified Extension type.
Moreover, `ABI compatibility <https://gcc.gnu.org/onlinedocs/libstdc++/manual/abi.html>`_
will be checked to ensure that compiler version from ``cc``
on local machine is compatible with pre-installed Paddle whl in python site-packages.
For example if Paddle with CUDA 10.1 is built with GCC 8.2, then the version of user's
local machine should satisfy GCC >= 8.2. Otherwise, a fatal error will occur because of
ABI compatibility.
.. note::
1. Compiler ABI compatibility is forward compatible. On Linux platform,
we recommend to use GCC 8.2 as soft linking condidate of ``/usr/bin/cc`` .
2. Using ``which cc`` to ensure location of ``cc`` and using ``cc --version``
to ensure linking GCC version on Linux.
3. Currently we support Linux and Windows platfrom. MacOS is supporting...
Compared with Just-In-Time ``load`` interface, it only compiles once by executing
``python setup.py install`` . Then customized operators API will be available everywhere
after importing it.
Its usage is almost same as `setuptools.setup` except for `ext_modules` A simple example of ``setup.py`` as followed:
arguments. For compiling multi custom operators, all necessary source files
can be include into just one Extension (CppExtension/CUDAExtension).
Moreover, only one `name` argument is required in `setup` and no need to spcific
`name` in Extension.
Example: .. code-block:: text
>> from paddle.utils.cpp_extension import CUDAExtension, setup # setup.py
>> setup(name='custom_module',
# Case 1: Compiling customized operators supporting CPU and GPU devices
from paddle.utils.cpp_extension import CUDAExtension, setup
setup(
name='custom_op', # name of package used by "import"
ext_modules=CUDAExtension( ext_modules=CUDAExtension(
sources=['relu_op.cc', 'relu_op.cu'], sources=['relu_op.cc', 'relu_op.cu', 'tanh_op.cc', 'tanh_op.cu'] # Support for compilation of multiple OPs
include_dirs=[], # specific user-defined include dirs )
extra_compile_args=[]) # specific user-defined compil arguments. )
# Case 2: Compiling customized operators supporting only CPU device
from paddle.utils.cpp_extension import CppExtension, setup
setup(
name='custom_op', # name of package used by "import"
ext_modules=CppExtension(
sources=['relu_op.cc', 'tanh_op.cc'] # Support for compilation of multiple OPs
)
)
Applying compilation and installation by executing ``python setup.py install`` under source files directory.
Then we can use the layer api as followed:
.. code-block:: text
import paddle
from custom_op import relu, tanh
x = paddle.randn([4, 10], dtype='float32')
relu_out = relu(x)
tanh_out = tanh(x)
Args:
name(str): Specify the name of shared library file and installed python package.
ext_modules(Extension): Specify the Extension instance including customized operator source files, compiling flags et.al.
If only compile operator supporting CPU device, please use ``CppExtension`` ; If compile operator
supporting CPU and GPU devices, please use ``CUDAExtension`` .
include_dirs(list[str], optional): Specify the extra include directoies to search head files. The interface will automatically add
``site-package/paddle/include`` . Please add the corresponding directory path if including third-party
head files. Default is None.
extra_compile_args(list[str] | dict, optional): Specify the extra compiling flags such as ``-O3`` . If set ``list[str]`` , all these flags
will be applied for ``cc`` and ``nvcc`` compiler. It support specify flags only applied ``cc`` or ``nvcc``
compiler using dict type with ``{'cxx': [...], 'nvcc': [...]}`` . Default is None.
**attr(dict, optional): Specify other arguments same as ``setuptools.setup`` .
Returns: None
""" """
cmdclass = attr.get('cmdclass', {}) cmdclass = attr.get('cmdclass', {})
assert isinstance(cmdclass, dict) assert isinstance(cmdclass, dict)
# if not specific cmdclass in setup, add it automaticaly. # if not specific cmdclass in setup, add it automatically.
if 'build_ext' not in cmdclass: if 'build_ext' not in cmdclass:
cmdclass['build_ext'] = BuildExtension.with_options( cmdclass['build_ext'] = BuildExtension.with_options(
no_python_abi_suffix=True) no_python_abi_suffix=True)
...@@ -71,18 +148,22 @@ def setup(**attr): ...@@ -71,18 +148,22 @@ def setup(**attr):
sources=['relu_op.cc', 'relu_op.cu']) sources=['relu_op.cc', 'relu_op.cu'])
# After running `python setup.py install` # After running `python setup.py install`
from custom_module import relue from custom_module import relu
""" """
# name argument is required # name argument is required
if 'name' not in attr: if 'name' not in attr:
raise ValueError(error_msg) raise ValueError(error_msg)
assert not attr['name'].endswith('module'), \
"Please don't use 'module' as suffix in `name` argument, "
"it will be stripped in setuptools.bdist_egg and cause import error."
ext_modules = attr.get('ext_modules', []) ext_modules = attr.get('ext_modules', [])
if not isinstance(ext_modules, list): if not isinstance(ext_modules, list):
ext_modules = [ext_modules] ext_modules = [ext_modules]
assert len( assert len(
ext_modules ext_modules
) == 1, "Required only one Extension, but received {}. If you want to compile multi operators, you can include all necessary source files in one Extenion.".format( ) == 1, "Required only one Extension, but received {}. If you want to compile multi operators, you can include all necessary source files in one Extension.".format(
len(ext_modules)) len(ext_modules))
# replace Extension.name with attr['name] to keep consistant with Package name. # replace Extension.name with attr['name] to keep consistant with Package name.
for ext_module in ext_modules: for ext_module in ext_modules:
...@@ -94,6 +175,13 @@ def setup(**attr): ...@@ -94,6 +175,13 @@ def setup(**attr):
assert 'easy_install' not in cmdclass assert 'easy_install' not in cmdclass
cmdclass['easy_install'] = EasyInstallCommand cmdclass['easy_install'] = EasyInstallCommand
# Note(Aurelius84): Add rename build_base directory hook in build command.
# To avoid using same build directory that will lead to remove the directory
# by mistake while parallelling execute setup.py, for example on CI.
assert 'build' not in cmdclass
build_base = os.path.join('build', attr['name'])
cmdclass['build'] = BuildCommand.with_options(build_base=build_base)
# Always set zip_safe=False to make compatible in PY2 and PY3 # Always set zip_safe=False to make compatible in PY2 and PY3
# See http://peak.telecommunity.com/DevCenter/setuptools#setting-the-zip-safe-flag # See http://peak.telecommunity.com/DevCenter/setuptools#setting-the-zip-safe-flag
attr['zip_safe'] = False attr['zip_safe'] = False
...@@ -105,16 +193,41 @@ def setup(**attr): ...@@ -105,16 +193,41 @@ def setup(**attr):
def CppExtension(sources, *args, **kwargs): def CppExtension(sources, *args, **kwargs):
""" """
Returns setuptools.CppExtension instance for setup.py to make it easy The interface is used to config source files of customized operators and complies
to specify compile flags while building C++ custommed op kernel. Op Kernel only supporting CPU device. Please use ``CUDAExtension`` if you want to
compile Op Kernel that supports both CPU and GPU devices.
It furtherly encapsulates python built-in ``setuptools.Extension`` .The arguments and
usage are same as the native interface, except for no need to explicitly specify
``name`` .
**A simple example:**
.. code-block:: text
# setup.py
# Compiling customized operators supporting only CPU device
from paddle.utils.cpp_extension import CppExtension, setup
setup(
name='custom_op',
ext_modules=CppExtension(sources=['relu_op.cc'])
)
.. note::
It is mainly used in ``setup`` and the nama of built shared library keeps same
as ``name`` argument specified in ``setup`` interface.
Args: Args:
sources(list[str]): The C++/CUDA source file names sources(list[str]): Specify the C++/CUDA source files of customized operators.
args(list[options]): list of config options used to compile shared library *args(list[options], optional): Specify other arguments same as ``setuptools.Extension`` .
kwargs(dict[option]): dict of config options used to compile shared library **kwargs(dict[option], optional): Specify other arguments same as ``setuptools.Extension`` .
Returns: Returns:
Extension: An instance of setuptools.Extension setuptools.Extension: An instance of ``setuptools.Extension``
""" """
kwargs = normalize_extension_kwargs(kwargs, use_cuda=False) kwargs = normalize_extension_kwargs(kwargs, use_cuda=False)
# Note(Aurelius84): While using `setup` and `jit`, the Extension `name` will # Note(Aurelius84): While using `setup` and `jit`, the Extension `name` will
...@@ -130,16 +243,43 @@ def CppExtension(sources, *args, **kwargs): ...@@ -130,16 +243,43 @@ def CppExtension(sources, *args, **kwargs):
def CUDAExtension(sources, *args, **kwargs): def CUDAExtension(sources, *args, **kwargs):
""" """
Returns setuptools.CppExtension instance for setup.py to make it easy The interface is used to config source files of customized operators and complies
to specify compile flags while build CUDA custommed op kernel. Op Kernel supporting both CPU and GPU devices. Please use ``CppExtension`` if you want to
compile Op Kernel that supports only CPU device.
It furtherly encapsulates python built-in ``setuptools.Extension`` .The arguments and
usage are same as the native interface, except for no need to explicitly specify
``name`` .
**A simple example:**
.. code-block:: text
# setup.py
# Compiling customized operators supporting CPU and GPU devices
from paddle.utils.cpp_extension import CUDAExtension, setup
setup(
name='custom_op',
ext_modules=CUDAExtension(
sources=['relu_op.cc', 'relu_op.cu']
)
)
.. note::
It is mainly used in ``setup`` and the nama of built shared library keeps same
as ``name`` argument specified in ``setup`` interface.
Args: Args:
sources(list[str]): The C++/CUDA source file names sources(list[str]): Specify the C++/CUDA source files of customized operators.
args(list[options]): list of config options used to compile shared library *args(list[options], optional): Specify other arguments same as ``setuptools.Extension`` .
kwargs(dict[option]): dict of config options used to compile shared library **kwargs(dict[option], optional): Specify other arguments same as ``setuptools.Extension`` .
Returns: Returns:
Extension: An instance of setuptools.Extension setuptools.Extension: An instance of setuptools.Extension
""" """
kwargs = normalize_extension_kwargs(kwargs, use_cuda=True) kwargs = normalize_extension_kwargs(kwargs, use_cuda=True)
# Note(Aurelius84): While using `setup` and `jit`, the Extension `name` will # Note(Aurelius84): While using `setup` and `jit`, the Extension `name` will
...@@ -202,9 +342,6 @@ class BuildExtension(build_ext, object): ...@@ -202,9 +342,6 @@ class BuildExtension(build_ext, object):
super(BuildExtension, self).__init__(*args, **kwargs) super(BuildExtension, self).__init__(*args, **kwargs)
self.no_python_abi_suffix = kwargs.get("no_python_abi_suffix", True) self.no_python_abi_suffix = kwargs.get("no_python_abi_suffix", True)
self.output_dir = kwargs.get("output_dir", None) self.output_dir = kwargs.get("output_dir", None)
# for compatible two custom op define method
use_new_custom_op_load_method(
kwargs.get("use_new_method", use_new_custom_op_load_method()))
def initialize_options(self): def initialize_options(self):
super(BuildExtension, self).initialize_options() super(BuildExtension, self).initialize_options()
...@@ -219,20 +356,14 @@ class BuildExtension(build_ext, object): ...@@ -219,20 +356,14 @@ class BuildExtension(build_ext, object):
def build_extensions(self): def build_extensions(self):
self._check_abi() self._check_abi()
for extension in self.extensions:
# check settings of compiler
if isinstance(extension.extra_compile_args, dict):
for compiler in ['cxx', 'nvcc']:
if compiler not in extension.extra_compile_args:
extension.extra_compile_args[compiler] = []
# add determine compile flags
add_compile_flag(extension, '-std=c++11')
# Consider .cu, .cu.cc as valid source extensions. # Consider .cu, .cu.cc as valid source extensions.
self.compiler.src_extensions += ['.cu', '.cu.cc'] self.compiler.src_extensions += ['.cu', '.cu.cc']
# Save the original _compile method for later. # Save the original _compile method for later.
if self.compiler.compiler_type == 'msvc' or IS_WINDOWS: if self.compiler.compiler_type == 'msvc':
raise NotImplementedError("Not support on MSVC currently.") self.compiler._cpp_extensions += ['.cu', '.cuh']
original_compile = self.compiler.compile
original_spawn = self.compiler.spawn
else: else:
original_compile = self.compiler._compile original_compile = self.compiler._compile
...@@ -255,8 +386,8 @@ class BuildExtension(build_ext, object): ...@@ -255,8 +386,8 @@ class BuildExtension(build_ext, object):
# {'nvcc': {}, 'cxx: {}} # {'nvcc': {}, 'cxx: {}}
if isinstance(cflags, dict): if isinstance(cflags, dict):
cflags = cflags['nvcc'] cflags = cflags['nvcc']
else:
cflags = prepare_unix_cflags(cflags) cflags = prepare_unix_cudaflags(cflags)
# cxx compile Cpp source # cxx compile Cpp source
elif isinstance(cflags, dict): elif isinstance(cflags, dict):
cflags = cflags['cxx'] cflags = cflags['cxx']
...@@ -268,6 +399,81 @@ class BuildExtension(build_ext, object): ...@@ -268,6 +399,81 @@ class BuildExtension(build_ext, object):
# restore original_compiler # restore original_compiler
self.compiler.compiler_so = original_compiler self.compiler.compiler_so = original_compiler
def win_custom_single_compiler(sources,
output_dir=None,
macros=None,
include_dirs=None,
debug=0,
extra_preargs=None,
extra_postargs=None,
depends=None):
self.cflags = copy.deepcopy(extra_postargs)
extra_postargs = None
def win_custom_spawn(cmd):
# Using regex to modify compile options
compile_options = self.compiler.compile_options
for i in range(len(cmd)):
if re.search('/MD', cmd[i]) is not None:
cmd[i] = '/MT'
if re.search('/W[1-4]', cmd[i]) is not None:
cmd[i] = '/W0'
# Using regex to match src, obj and include files
src_regex = re.compile('/T(p|c)(.*)')
src_list = [
m.group(2) for m in (src_regex.match(elem) for elem in cmd)
if m
]
obj_regex = re.compile('/Fo(.*)')
obj_list = [
m.group(1) for m in (obj_regex.match(elem) for elem in cmd)
if m
]
include_regex = re.compile(r'((\-|\/)I.*)')
include_list = [
m.group(1)
for m in (include_regex.match(elem) for elem in cmd) if m
]
assert len(src_list) == 1 and len(obj_list) == 1
src = src_list[0]
obj = obj_list[0]
if is_cuda_file(src):
assert CUDA_HOME is not None
nvcc_cmd = os.path.join(CUDA_HOME, 'bin', 'nvcc')
if isinstance(self.cflags, dict):
cflags = self.cflags['nvcc']
elif isinstance(self.cflags, list):
cflags = self.cflags
else:
cflags = []
cflags = prepare_win_cudaflags(cflags) + ['--use-local-env']
for flag in MSVC_COMPILE_FLAGS:
cflags = ['-Xcompiler', flag] + cflags
cmd = [nvcc_cmd, '-c', src, '-o', obj
] + include_list + cflags
elif isinstance(self.cflags, dict):
cflags = MSVC_COMPILE_FLAGS + self.cflags['cxx']
cmd += cflags
elif isinstance(self.cflags, list):
cflags = MSVC_COMPILE_FLAGS + self.cflags
cmd += cflags
return original_spawn(cmd)
try:
self.compiler.spawn = win_custom_spawn
return original_compile(sources, output_dir, macros,
include_dirs, debug, extra_preargs,
extra_postargs, depends)
finally:
self.compiler.spawn = original_spawn
def object_filenames_with_cuda(origina_func, build_directory): def object_filenames_with_cuda(origina_func, build_directory):
""" """
Decorated the function to add customized naming machanism. Decorated the function to add customized naming machanism.
...@@ -280,9 +486,12 @@ class BuildExtension(build_ext, object): ...@@ -280,9 +486,12 @@ class BuildExtension(build_ext, object):
objects = origina_func(source_filenames, strip_dir, objects = origina_func(source_filenames, strip_dir,
output_dir) output_dir)
for i, source in enumerate(source_filenames): for i, source in enumerate(source_filenames):
# modify xx.o -> xx.cu.o # modify xx.o -> xx.cu.o/xx.cu.obj
if is_cuda_file(source): if is_cuda_file(source):
old_obj = objects[i] old_obj = objects[i]
if self.compiler.compiler_type == 'msvc':
objects[i] = old_obj[:-3] + 'cu.obj'
else:
objects[i] = old_obj[:-1] + 'cu.o' objects[i] = old_obj[:-1] + 'cu.o'
# if user set build_directory, output objects there. # if user set build_directory, output objects there.
if build_directory is not None: if build_directory is not None:
...@@ -300,10 +509,13 @@ class BuildExtension(build_ext, object): ...@@ -300,10 +509,13 @@ class BuildExtension(build_ext, object):
return wrapper return wrapper
# customized compile process # customized compile process
if self.compiler.compiler_type == 'msvc':
self.compiler.compile = win_custom_single_compiler
else:
self.compiler._compile = unix_custom_single_compiler self.compiler._compile = unix_custom_single_compiler
self.compiler.object_filenames = object_filenames_with_cuda( self.compiler.object_filenames = object_filenames_with_cuda(
self.compiler.object_filenames, self.build_lib) self.compiler.object_filenames, self.build_lib)
self._record_op_info() self._record_op_info()
print("Compiling user custom op, it will cost a few seconds.....") print("Compiling user custom op, it will cost a few seconds.....")
...@@ -333,15 +545,21 @@ class BuildExtension(build_ext, object): ...@@ -333,15 +545,21 @@ class BuildExtension(build_ext, object):
compiler = self.compiler.compiler_cxx[0] compiler = self.compiler.compiler_cxx[0]
elif IS_WINDOWS: elif IS_WINDOWS:
compiler = os.environ.get('CXX', 'cl') compiler = os.environ.get('CXX', 'cl')
raise NotImplementedError("We don't support Windows Currently.")
else: else:
compiler = os.environ.get('CXX', 'c++') compiler = os.environ.get('CXX', 'c++')
check_abi_compatibility(compiler) check_abi_compatibility(compiler)
# Warn user if VC env is activated but `DISTUTILS_USE_SDK` is not set.
if IS_WINDOWS and 'VSCMD_ARG_TGT_ARCH' in os.environ and 'DISTUTILS_USE_SDK' not in os.environ:
msg = (
'It seems that the VC environment is activated but DISTUTILS_USE_SDK is not set.'
'This may lead to multiple activations of the VC env.'
'Please set `DISTUTILS_USE_SDK=1` and try again.')
raise UserWarning(msg)
def _record_op_info(self): def _record_op_info(self):
""" """
Record custum op inforomation. Record custom op information.
""" """
# parse shared library abs path # parse shared library abs path
outputs = self.get_outputs() outputs = self.get_outputs()
...@@ -380,16 +598,59 @@ class EasyInstallCommand(easy_install, object): ...@@ -380,16 +598,59 @@ class EasyInstallCommand(easy_install, object):
# .so shared library to another name. # .so shared library to another name.
for egg_file in self.outputs: for egg_file in self.outputs:
filename, ext = os.path.splitext(egg_file) filename, ext = os.path.splitext(egg_file)
if ext == '.so': will_rename = False
if OS_NAME.startswith('linux') and ext == '.so':
will_rename = True
elif IS_WINDOWS and ext == '.pyd':
will_rename = True
if will_rename:
new_so_path = filename + "_pd_" + ext new_so_path = filename + "_pd_" + ext
if not os.path.exists(new_so_path): if not os.path.exists(new_so_path):
os.rename(r'%s' % egg_file, r'%s' % new_so_path) os.rename(r'%s' % egg_file, r'%s' % new_so_path)
assert os.path.exists(new_so_path) assert os.path.exists(new_so_path)
class BuildCommand(build, object):
"""
Extend build Command to control the behavior of specifying `build_base` root directory.
NOTE(Aurelius84): This is a hook subclass inherited Command used to specify customized
build_base directory.
"""
@classmethod
def with_options(cls, **options):
"""
Returns a BuildCommand subclass containing use-defined options.
"""
class cls_with_options(cls):
def __init__(self, *args, **kwargs):
kwargs.update(options)
cls.__init__(self, *args, **kwargs)
return cls_with_options
def __init__(self, *args, **kwargs):
# Note: shall put before super()
self._specified_build_base = kwargs.get('build_base', None)
super(BuildCommand, self).__init__(*args, **kwargs)
def initialize_options(self):
"""
build_base is root directory for all sub-command, such as
build_lib, build_temp. See `distutils.command.build` for details.
"""
super(BuildCommand, self).initialize_options()
if self._specified_build_base is not None:
self.build_base = self._specified_build_base
def load(name, def load(name,
sources, sources,
extra_cflags=None, extra_cxx_cflags=None,
extra_cuda_cflags=None, extra_cuda_cflags=None,
extra_ldflags=None, extra_ldflags=None,
extra_include_paths=None, extra_include_paths=None,
...@@ -399,48 +660,86 @@ def load(name, ...@@ -399,48 +660,86 @@ def load(name,
""" """
An Interface to automatically compile C++/CUDA source files Just-In-Time An Interface to automatically compile C++/CUDA source files Just-In-Time
and return callable python function as other Paddle layers API. It will and return callable python function as other Paddle layers API. It will
append user defined custom op in background. append user defined custom operators in background while building models.
It will perform compiling, linking, Python API generation and module loading
processes under a individual subprocess. It does not require CMake or Ninja environment
and only ``g++/nvcc`` on Linux and clang++ on MacOS. For example it requires
GCC compiler with version is greater than 5.4 and linked into ``/usr/bin/cc`` .
If compiling Operators supporting GPU device, please make sure ``nvcc`` compiler
is installed in local environment.
Moreover, `ABI compatibility <https://gcc.gnu.org/onlinedocs/libstdc++/manual/abi.html>`_
will be checked to ensure that compiler version from ``cc``
on local machine is compatible with pre-installed Paddle whl in python site-packages.
For example if Paddle with CUDA 10.1 is built with GCC 8.2, then the version of user's
local machine should satisfy GCC >= 8.2. Otherwise, a fatal error will occur because of
ABI compatibility.
Compared with ``setup`` interface, it doesn't need extra ``setup.py`` and excute
``python setup.py install`` command. The interface contains all compiling and installing
process underground.
.. note::
1. Compiler ABI compatibility is forward compatible. On Linux platform,
we recommend to use GCC 8.2 as soft linking condidate of ``/usr/bin/cc`` .
2. Using ``which cc`` to ensure location of ``cc`` and using ``cc --version``
to ensure linking GCC version on Linux.
3. Currenly we support Linux and Windows platfrom. MacOS is supporting...
**A simple example:**
.. code-block:: text
import paddle
from paddle.utils.cpp_extension import load
custom_op_module = load(
name="op_shared_libary_name", # name of shared library
sources=['relu_op.cc', 'relu_op.cu'], # source files of cusomized op
extra_cxx_cflags=['-DPADDLE_WITH_MKLDNN'], # need to specify the flag if pre-installed Paddle supports MKLDNN
extra_cuda_cflags=['-DPADDLE_WITH_MKLDNN'], # need to specify the flag if pre-installed Paddle supports MKLDNN
interpreter='python3.7', # optional, specify another python interpreter
verbose=True # output log information
)
x = paddle.randn([4, 10], dtype='float32')
out = custom_op_module.relu(x)
This module will perform compiling, linking, api generation and module loading
processes for users. It does not require CMake or Ninja environment and only
g++/nvcc on Linux and clang++ on MacOS. Moreover, ABI compatibility will be
checked to ensure that compiler version on local machine is compatible with
pre-installed Paddle whl in python site-packages. For example if Paddle is built
with GCC5.4, the version of user's local machine should satisfy GCC >= 5.4.
Otherwise, a fatal error will occur because ABI compatibility.
Args: Args:
name(str): generated shared library file name. name(str): Specify the name of generated shared library file name, not including ``.so`` and ``.dll`` suffix.
sources(list[str]): custom op source files name with .cc/.cu suffix. sources(list[str]): Specify source files name of customized operators. Supporting ``.cc`` , ``.cpp`` for CPP file
extra_cflag(list[str]): additional flags used to compile CPP files. By default and ``.cu`` for CUDA file.
extra_cxx_cflags(list[str], optional): Specify additional flags used to compile CPP files. By default
all basic and framework related flags have been included. all basic and framework related flags have been included.
If your pre-insall Paddle supported MKLDNN, please add If your pre-insall Paddle supported MKLDNN, please add
'-DPADDLE_WITH_MKLDNN'. Default None. ``-DPADDLE_WITH_MKLDNN`` . Default is None.
extra_cuda_cflags(list[str]): additonal flags used to compile CUDA files. See extra_cuda_cflags(list[str], optional): Specify additional flags used to compile CUDA files. By default
https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html all basic and framework related flags have been included. If your pre-insall Paddle supported MKLDNN,
for details. Default None. please add ``-DPADDLE_WITH_MKLDNN`` . Default None. See `Cuda Compiler Driver NVCC <https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html>`_
extra_ldflags(list[str]): additonal flags used to link shared library. See for details. Default is None.
https://gcc.gnu.org/onlinedocs/gcc/Link-Options.html for details. extra_ldflags(list[str], optional): Specify additional flags used to link shared library. See
Default None. `GCC Link Options <https://gcc.gnu.org/onlinedocs/gcc/Link-Options.html>`_ for details.
extra_include_paths(list[str]): additional include path used to search header files. Default is None.
Default None. extra_include_paths(list[str], optional): Specify additional include path used to search header files. By default
build_directory(str): specific directory path to put shared library file. If set None, all basic headers are included implicitly from ``site-package/paddle/include`` .
it will use `PADDLE_EXTENSION_DIR` from os.environ. Use Default is None.
`paddle.utils.cpp_extension.get_build_directory()` to see the location. build_directory(str, optional): Specify root directory path to put shared library file. If set None,
interpreter(str): alias or full interpreter path to specific which one to use if have installed multiple. it will use ``PADDLE_EXTENSION_DIR`` from os.environ. Use
If set None, will use `python` as default interpreter. ``paddle.utils.cpp_extension.get_build_directory()`` to see the location. Default is None.
verbose(bool): whether to verbose compiled log information interpreter(str, optional): Specify nterpreter path, supporting alias and full path.
If set None, it will use `python` as default interpreter. If local environment contains
more than one python interpreters and want to use new interpreter to apply compilation,
please specify this parameter, such as ``python3.7`` . Default is None.
verbose(bool, optional): whether to verbose compiled log information. Default is False
Returns: Returns:
custom api: A callable python function with same signature as CustomOp Kernel defination. Moudle: A callable python module contains all CustomOp Layer APIs.
Example:
>> from paddle.utils.cpp_extension import load
>> relu2 = load(name='relu2',
sources=['relu_op.cc', 'relu_op.cu'])
>> x = paddle.rand([4, 10]], dtype='float32')
>> out = relu2(x)
""" """
if build_directory is None: if build_directory is None:
...@@ -448,24 +747,37 @@ def load(name, ...@@ -448,24 +747,37 @@ def load(name,
# ensure to use abs path # ensure to use abs path
build_directory = os.path.abspath(build_directory) build_directory = os.path.abspath(build_directory)
# Will load shared library from 'path' on windows
if IS_WINDOWS:
os.environ['path'] = build_directory + ';' + os.environ['path']
log_v("build_directory: {}".format(build_directory), verbose) log_v("build_directory: {}".format(build_directory), verbose)
file_path = os.path.join(build_directory, "setup.py") file_path = os.path.join(build_directory, "{}_setup.py".format(name))
sources = [os.path.abspath(source) for source in sources] sources = [os.path.abspath(source) for source in sources]
# TODO(Aurelius84): split cflags and cuda_flags if extra_cxx_cflags is None: extra_cxx_cflags = []
if extra_cflags is None: extra_cflags = []
if extra_cuda_cflags is None: extra_cuda_cflags = [] if extra_cuda_cflags is None: extra_cuda_cflags = []
compile_flags = extra_cflags + extra_cuda_cflags assert isinstance(
log_v("additonal compile_flags: [{}]".format(' '.join(compile_flags)), extra_cxx_cflags, list
verbose) ), "Required type(extra_cxx_cflags) == list[str], but received {}".format(
extra_cxx_cflags)
assert isinstance(
extra_cuda_cflags, list
), "Required type(extra_cuda_cflags) == list[str], but received {}".format(
extra_cuda_cflags)
log_v("additional extra_cxx_cflags: [{}], extra_cuda_cflags: [{}]".format(
' '.join(extra_cxx_cflags), ' '.join(extra_cuda_cflags)), verbose)
# write setup.py file and compile it # write setup.py file and compile it
_write_setup_file(name, sources, file_path, extra_include_paths, build_base_dir = os.path.join(build_directory, name)
compile_flags, extra_ldflags, verbose) _write_setup_file(name, sources, file_path, build_base_dir,
extra_include_paths, extra_cxx_cflags, extra_cuda_cflags,
extra_ldflags, verbose)
_jit_compile(file_path, interpreter, verbose) _jit_compile(file_path, interpreter, verbose)
# import as callable python api # import as callable python api
custom_op_api = _import_module_from_library(name, build_directory, verbose) custom_op_api = _import_module_from_library(name, build_base_dir, verbose)
return custom_op_api return custom_op_api
python/paddle/utils/cpp_extension/extension_utils.py
浏览文件 @ d3e60959
...@@ -16,7 +16,6 @@ import os ...@@ -16,7 +16,6 @@ import os
import re import re
import six import six
import sys import sys
import copy
import glob import glob
import logging import logging
import collections import collections
...@@ -38,11 +37,17 @@ logger = logging.getLogger("utils.cpp_extension") ...@@ -38,11 +37,17 @@ logger = logging.getLogger("utils.cpp_extension")
OS_NAME = sys.platform OS_NAME = sys.platform
IS_WINDOWS = OS_NAME.startswith('win') IS_WINDOWS = OS_NAME.startswith('win')
NVCC_COMPILE_FLAGS = [
'-ccbin', 'cc', '-DPADDLE_WITH_CUDA', '-DEIGEN_USE_GPU', '-DPADDLE_USE_DSO', MSVC_COMPILE_FLAGS = [
'-Xcompiler', '-fPIC', '-w', '--expt-relaxed-constexpr', '-O3', '-DNVCC' '/MT', '/wd4819', '/wd4251', '/wd4244', '/wd4267', '/wd4275', '/wd4018',
'/wd4190', '/EHsc', '/w', '/DGOOGLE_GLOG_DLL_DECL',
'/DBOOST_HAS_STATIC_ASSERT', '/DNDEBUG', '/DPADDLE_USE_DSO'
] ]
MSVC_LINK_FLAGS = ['/MACHINE:X64', 'paddle_framework.lib']
COMMON_NVCC_FLAGS = ['-DPADDLE_WITH_CUDA', '-DEIGEN_USE_GPU', '-O3']
GCC_MINI_VERSION = (5, 4, 0) GCC_MINI_VERSION = (5, 4, 0)
# Give warning if using wrong compiler # Give warning if using wrong compiler
WRONG_COMPILER_WARNING = ''' WRONG_COMPILER_WARNING = '''
...@@ -80,6 +85,14 @@ information ...@@ -80,6 +85,14 @@ information
''' '''
USING_NEW_CUSTOM_OP_LOAD_METHOD = True USING_NEW_CUSTOM_OP_LOAD_METHOD = True
DEFAULT_OP_ATTR_NAMES = [
core.op_proto_and_checker_maker.kOpRoleAttrName(),
core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
core.op_proto_and_checker_maker.kOpCreationCallstackAttrName(),
core.op_proto_and_checker_maker.kOpDeviceAttrName()
]
# NOTE(chenweihang): In order to be compatible with # NOTE(chenweihang): In order to be compatible with
# the two custom op define method, after removing # the two custom op define method, after removing
...@@ -206,11 +219,23 @@ class CustomOpInfo: ...@@ -206,11 +219,23 @@ class CustomOpInfo:
return next(reversed(self.op_info_map.items())) return next(reversed(self.op_info_map.items()))
def prepare_unix_cflags(cflags): def prepare_unix_cudaflags(cflags):
""" """
Prepare all necessary compiled flags for nvcc compiling CUDA files. Prepare all necessary compiled flags for nvcc compiling CUDA files.
""" """
cflags = NVCC_COMPILE_FLAGS + cflags + get_cuda_arch_flags(cflags) cflags = COMMON_NVCC_FLAGS + [
'-ccbin', 'cc', '-Xcompiler', '-fPIC', '-w', '--expt-relaxed-constexpr',
'-DNVCC'
] + cflags + get_cuda_arch_flags(cflags)
return cflags
def prepare_win_cudaflags(cflags):
"""
Prepare all necessary compiled flags for nvcc compiling CUDA files.
"""
cflags = COMMON_NVCC_FLAGS + ['-w'] + cflags + get_cuda_arch_flags(cflags)
return cflags return cflags
...@@ -245,6 +270,7 @@ def normalize_extension_kwargs(kwargs, use_cuda=False): ...@@ -245,6 +270,7 @@ def normalize_extension_kwargs(kwargs, use_cuda=False):
# append necessary include dir path of paddle # append necessary include dir path of paddle
include_dirs = kwargs.get('include_dirs', []) include_dirs = kwargs.get('include_dirs', [])
include_dirs.extend(find_paddle_includes(use_cuda)) include_dirs.extend(find_paddle_includes(use_cuda))
kwargs['include_dirs'] = include_dirs kwargs['include_dirs'] = include_dirs
# append necessary lib path of paddle # append necessary lib path of paddle
...@@ -252,50 +278,46 @@ def normalize_extension_kwargs(kwargs, use_cuda=False): ...@@ -252,50 +278,46 @@ def normalize_extension_kwargs(kwargs, use_cuda=False):
library_dirs.extend(find_paddle_libraries(use_cuda)) library_dirs.extend(find_paddle_libraries(use_cuda))
kwargs['library_dirs'] = library_dirs kwargs['library_dirs'] = library_dirs
# add runtime library dirs # append compile flags and check settings of compiler
runtime_library_dirs = kwargs.get('runtime_library_dirs', []) extra_compile_args = kwargs.get('extra_compile_args', [])
runtime_library_dirs.extend(find_paddle_libraries(use_cuda)) if isinstance(extra_compile_args, dict):
kwargs['runtime_library_dirs'] = runtime_library_dirs for compiler in ['cxx', 'nvcc']:
if compiler not in extra_compile_args:
extra_compile_args[compiler] = []
if IS_WINDOWS:
# TODO(zhouwei): may append compile flags in future
pass
# append link flags
extra_link_args = kwargs.get('extra_link_args', [])
extra_link_args.extend(MSVC_LINK_FLAGS)
if use_cuda:
extra_link_args.extend(['cudadevrt.lib', 'cudart_static.lib'])
kwargs['extra_link_args'] = extra_link_args
else:
# append compile flags # append compile flags
extra_compile_args = kwargs.get('extra_compile_args', []) add_compile_flag(extra_compile_args, ['-g', '-w']) # disable warnings
extra_compile_args.extend(['-g', '-w']) # diable warnings
kwargs['extra_compile_args'] = extra_compile_args
# append link flags # append link flags
extra_link_args = kwargs.get('extra_link_args', []) extra_link_args = kwargs.get('extra_link_args', [])
if use_new_custom_op_load_method():
extra_link_args.append('-lpaddle_custom_op')
else:
extra_link_args.append('-lpaddle_framework') extra_link_args.append('-lpaddle_framework')
if use_cuda: if use_cuda:
extra_link_args.append('-lcudart') extra_link_args.append('-lcudart')
kwargs['extra_link_args'] = extra_link_args kwargs['extra_link_args'] = extra_link_args
kwargs['language'] = 'c++' # add runtime library dirs
return kwargs runtime_library_dirs = kwargs.get('runtime_library_dirs', [])
runtime_library_dirs.extend(find_paddle_libraries(use_cuda))
kwargs['runtime_library_dirs'] = runtime_library_dirs
def find_paddle_includes(use_cuda=False):
"""
Return Paddle necessary include dir path.
"""
# pythonXX/site-packages/paddle/include
paddle_include_dir = get_include()
third_party_dir = os.path.join(paddle_include_dir, 'third_party')
include_dirs = [paddle_include_dir, third_party_dir]
return include_dirs
def find_cuda_includes():
cuda_home = find_cuda_home() kwargs['extra_compile_args'] = extra_compile_args
if cuda_home is None:
raise ValueError(
"Not found CUDA runtime, please use `export CUDA_HOME=XXX` to specific it."
)
return [os.path.join(cuda_home, 'lib64')] kwargs['language'] = 'c++'
return kwargs
def find_cuda_home(): def find_cuda_home():
...@@ -315,19 +337,22 @@ def find_cuda_home(): ...@@ -315,19 +337,22 @@ def find_cuda_home():
if six.PY3: if six.PY3:
nvcc_path = nvcc_path.decode() nvcc_path = nvcc_path.decode()
nvcc_path = nvcc_path.rstrip('\r\n') nvcc_path = nvcc_path.rstrip('\r\n')
# for example: /usr/local/cuda/bin/nvcc # for example: /usr/local/cuda/bin/nvcc
cuda_home = os.path.dirname(os.path.dirname(nvcc_path)) cuda_home = os.path.dirname(os.path.dirname(nvcc_path))
except: except:
if IS_WINDOWS: if IS_WINDOWS:
# search from default NVIDIA GPU path # search from default NVIDIA GPU path
candidate_paths = glob.glob( candidate_paths = glob.glob(
'C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v*.*') 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*.*'
)
if len(candidate_paths) > 0: if len(candidate_paths) > 0:
cuda_home = candidate_paths[0] cuda_home = candidate_paths[0]
else: else:
cuda_home = "/usr/local/cuda" cuda_home = "/usr/local/cuda"
# step 3. check whether path is valid # step 3. check whether path is valid
if not os.path.exists(cuda_home) and core.is_compiled_with_cuda(): if cuda_home and not os.path.exists(
cuda_home) and core.is_compiled_with_cuda():
cuda_home = None cuda_home = None
warnings.warn( warnings.warn(
"Not found CUDA runtime, please use `export CUDA_HOME= XXX` to specific it." "Not found CUDA runtime, please use `export CUDA_HOME= XXX` to specific it."
...@@ -336,27 +361,73 @@ def find_cuda_home(): ...@@ -336,27 +361,73 @@ def find_cuda_home():
return cuda_home return cuda_home
def find_cuda_includes():
"""
Use heuristic method to find cuda include path
"""
cuda_home = find_cuda_home()
if cuda_home is None:
raise ValueError(
"Not found CUDA runtime, please use `export CUDA_HOME=XXX` to specific it."
)
return [os.path.join(cuda_home, 'include')]
def find_paddle_includes(use_cuda=False):
"""
Return Paddle necessary include dir path.
"""
# pythonXX/site-packages/paddle/include
paddle_include_dir = get_include()
third_party_dir = os.path.join(paddle_include_dir, 'third_party')
include_dirs = [paddle_include_dir, third_party_dir]
if use_cuda:
cuda_include_dir = find_cuda_includes()
include_dirs.extend(cuda_include_dir)
return include_dirs
def find_cuda_libraries():
"""
Use heuristic method to find cuda static lib path
"""
cuda_home = find_cuda_home()
if cuda_home is None:
raise ValueError(
"Not found CUDA runtime, please use `export CUDA_HOME=XXX` to specific it."
)
if IS_WINDOWS:
cuda_lib_dir = [os.path.join(cuda_home, 'lib', 'x64')]
else:
cuda_lib_dir = [os.path.join(cuda_home, 'lib64')]
return cuda_lib_dir
def find_paddle_libraries(use_cuda=False): def find_paddle_libraries(use_cuda=False):
""" """
Return Paddle necessary library dir path. Return Paddle necessary library dir path.
""" """
# pythonXX/site-packages/paddle/libs # pythonXX/site-packages/paddle/libs
paddle_lib_dirs = [get_lib()] paddle_lib_dirs = [get_lib()]
if use_cuda: if use_cuda:
cuda_dirs = find_cuda_includes() cuda_lib_dir = find_cuda_libraries()
paddle_lib_dirs.extend(cuda_dirs) paddle_lib_dirs.extend(cuda_lib_dir)
return paddle_lib_dirs return paddle_lib_dirs
def add_compile_flag(extension, flag): def add_compile_flag(extra_compile_args, flags):
extra_compile_args = copy.deepcopy(extension.extra_compile_args) assert isinstance(flags, list)
if isinstance(extra_compile_args, dict): if isinstance(extra_compile_args, dict):
for args in extra_compile_args.values(): for args in extra_compile_args.values():
args.append(flag) args.extend(flags)
else: else:
extra_compile_args.append(flag) extra_compile_args.extend(flags)
extension.extra_compile_args = extra_compile_args
def is_cuda_file(path): def is_cuda_file(path):
...@@ -369,17 +440,34 @@ def is_cuda_file(path): ...@@ -369,17 +440,34 @@ def is_cuda_file(path):
def get_build_directory(verbose=False): def get_build_directory(verbose=False):
""" """
Return paddle extension root directory, default specific by `PADDLE_EXTENSION_DIR` Return paddle extension root directory to put shared library. It could be specified by
``export PADDLE_EXTENSION_DIR=XXX`` . If not set, ``~/.cache/paddle_extension`` will be used
by default.
Returns:
The root directory of compiling customized operators.
Examples:
.. code-block:: python
from paddle.utils.cpp_extension import get_build_directory
build_dir = get_build_directory()
print(build_dir)
""" """
root_extensions_directory = os.environ.get('PADDLE_EXTENSION_DIR') root_extensions_directory = os.environ.get('PADDLE_EXTENSION_DIR')
if root_extensions_directory is None: if root_extensions_directory is None:
dir_name = "paddle_extensions" dir_name = "paddle_extensions"
if OS_NAME.startswith('linux'):
root_extensions_directory = os.path.join( root_extensions_directory = os.path.join(
os.path.expanduser('~/.cache'), dir_name) os.path.expanduser('~/.cache'), dir_name)
else: if IS_WINDOWS:
# TODO(Aurelius84): consider wind32/macOs root_extensions_directory = os.path.normpath(
raise NotImplementedError("Only support Linux now.") root_extensions_directory)
elif OS_NAME.startswith('darwin'):
# TODO(Aurelius84): consider macOs
raise NotImplementedError("Not support Mac now.")
log_v("$PADDLE_EXTENSION_DIR is not set, using path: {} by default.". log_v("$PADDLE_EXTENSION_DIR is not set, using path: {} by default.".
format(root_extensions_directory), verbose) format(root_extensions_directory), verbose)
...@@ -404,16 +492,22 @@ def parse_op_info(op_name): ...@@ -404,16 +492,22 @@ def parse_op_info(op_name):
in_names = [x.name for x in op_proto.inputs] in_names = [x.name for x in op_proto.inputs]
out_names = [x.name for x in op_proto.outputs] out_names = [x.name for x in op_proto.outputs]
attr_names = [
x.name for x in op_proto.attrs if x.name not in DEFAULT_OP_ATTR_NAMES
]
return in_names, out_names return in_names, out_names, attr_names
def _import_module_from_library(module_name, build_directory, verbose=False): def _import_module_from_library(module_name, build_directory, verbose=False):
""" """
Load .so shared library and import it as callable python module. Load shared library and import it as callable python module.
""" """
# TODO(Aurelius84): Consider file suffix is .dll on Windows Platform. if IS_WINDOWS:
ext_path = os.path.join(build_directory, module_name + '.so') dynamic_suffix = '.pyd'
else:
dynamic_suffix = '.so'
ext_path = os.path.join(build_directory, module_name + dynamic_suffix)
if not os.path.exists(ext_path): if not os.path.exists(ext_path):
raise FileNotFoundError("Extension path: {} does not exist.".format( raise FileNotFoundError("Extension path: {} does not exist.".format(
ext_path)) ext_path))
...@@ -448,7 +542,7 @@ def _generate_python_module(module_name, ...@@ -448,7 +542,7 @@ def _generate_python_module(module_name,
def _custom_api_content(op_name): def _custom_api_content(op_name):
params_str, ins_str, outs_str = _get_api_inputs_str(op_name) params_str, ins_str, attrs_str, outs_str = _get_api_inputs_str(op_name)
API_TEMPLATE = textwrap.dedent(""" API_TEMPLATE = textwrap.dedent("""
from paddle.fluid.layer_helper import LayerHelper from paddle.fluid.layer_helper import LayerHelper
...@@ -456,8 +550,9 @@ def _custom_api_content(op_name): ...@@ -456,8 +550,9 @@ def _custom_api_content(op_name):
def {op_name}({inputs}): def {op_name}({inputs}):
helper = LayerHelper("{op_name}", **locals()) helper = LayerHelper("{op_name}", **locals())
# prepare inputs and output # prepare inputs and outputs
ins = {ins} ins = {ins}
attrs = {attrs}
outs = {{}} outs = {{}}
out_names = {out_names} out_names = {out_names}
for out_name in out_names: for out_name in out_names:
...@@ -465,7 +560,7 @@ def _custom_api_content(op_name): ...@@ -465,7 +560,7 @@ def _custom_api_content(op_name):
# in runtime. # in runtime.
outs[out_name] = helper.create_variable(dtype='float32') outs[out_name] = helper.create_variable(dtype='float32')
helper.append_op(type="{op_name}", inputs=ins, outputs=outs) helper.append_op(type="{op_name}", inputs=ins, outputs=outs, attrs=attrs)
res = [outs[out_name] for out_name in out_names] res = [outs[out_name] for out_name in out_names]
...@@ -474,7 +569,11 @@ def _custom_api_content(op_name): ...@@ -474,7 +569,11 @@ def _custom_api_content(op_name):
# generate python api file # generate python api file
api_content = API_TEMPLATE.format( api_content = API_TEMPLATE.format(
op_name=op_name, inputs=params_str, ins=ins_str, out_names=outs_str) op_name=op_name,
inputs=params_str,
ins=ins_str,
attrs=attrs_str,
out_names=outs_str)
return api_content return api_content
...@@ -505,22 +604,30 @@ def _get_api_inputs_str(op_name): ...@@ -505,22 +604,30 @@ def _get_api_inputs_str(op_name):
""" """
Returns string of api parameters and inputs dict. Returns string of api parameters and inputs dict.
""" """
in_names, out_names = parse_op_info(op_name) in_names, out_names, attr_names = parse_op_info(op_name)
# e.g: x, y, z # e.g: x, y, z
params_str = ','.join([p.lower() for p in in_names]) param_names = in_names + attr_names
params_str = ','.join([p.lower() for p in param_names])
# e.g: {'X': x, 'Y': y, 'Z': z} # e.g: {'X': x, 'Y': y, 'Z': z}
ins_str = "{%s}" % ','.join( ins_str = "{%s}" % ','.join(
["'{}' : {}".format(in_name, in_name.lower()) for in_name in in_names]) ["'{}' : {}".format(in_name, in_name.lower()) for in_name in in_names])
# e.g: {'num': n}
attrs_str = "{%s}" % ",".join([
"'{}' : {}".format(attr_name, attr_name.lower())
for attr_name in attr_names
])
# e.g: ['Out', 'Index'] # e.g: ['Out', 'Index']
outs_str = "[%s]" % ','.join(["'{}'".format(name) for name in out_names]) outs_str = "[%s]" % ','.join(["'{}'".format(name) for name in out_names])
return params_str, ins_str, outs_str return params_str, ins_str, attrs_str, outs_str
def _write_setup_file(name, def _write_setup_file(name,
sources, sources,
file_path, file_path,
build_dir,
include_dirs, include_dirs,
compile_flags, extra_cxx_cflags,
extra_cuda_cflags,
link_args, link_args,
verbose=False): verbose=False):
""" """
...@@ -530,18 +637,21 @@ def _write_setup_file(name, ...@@ -530,18 +637,21 @@ def _write_setup_file(name,
import os import os
from paddle.utils.cpp_extension import CppExtension, CUDAExtension, BuildExtension, setup from paddle.utils.cpp_extension import CppExtension, CUDAExtension, BuildExtension, setup
from paddle.utils.cpp_extension import get_build_directory from paddle.utils.cpp_extension import get_build_directory
from paddle.utils.cpp_extension.extension_utils import use_new_custom_op_load_method
use_new_custom_op_load_method({use_new_method})
setup( setup(
name='{name}', name='{name}',
ext_modules=[ ext_modules=[
{prefix}Extension( {prefix}Extension(
sources={sources}, sources={sources},
include_dirs={include_dirs}, include_dirs={include_dirs},
extra_compile_args={extra_compile_args}, extra_compile_args={{'cxx':{extra_cxx_cflags}, 'nvcc':{extra_cuda_cflags}}},
extra_link_args={extra_link_args})], extra_link_args={extra_link_args})],
cmdclass={{"build_ext" : BuildExtension.with_options( cmdclass={{"build_ext" : BuildExtension.with_options(
output_dir=get_build_directory(), output_dir=r'{build_dir}',
no_python_abi_suffix=True, no_python_abi_suffix=True)
use_new_method={use_new_method})
}})""").lstrip() }})""").lstrip()
with_cuda = False with_cuda = False
...@@ -554,8 +664,10 @@ def _write_setup_file(name, ...@@ -554,8 +664,10 @@ def _write_setup_file(name,
prefix='CUDA' if with_cuda else 'Cpp', prefix='CUDA' if with_cuda else 'Cpp',
sources=list2str(sources), sources=list2str(sources),
include_dirs=list2str(include_dirs), include_dirs=list2str(include_dirs),
extra_compile_args=list2str(compile_flags), extra_cxx_cflags=list2str(extra_cxx_cflags),
extra_cuda_cflags=list2str(extra_cuda_cflags),
extra_link_args=list2str(link_args), extra_link_args=list2str(link_args),
build_dir=build_dir,
use_new_method=use_new_custom_op_load_method()) use_new_method=use_new_custom_op_load_method())
log_v('write setup.py into {}'.format(file_path), verbose) log_v('write setup.py into {}'.format(file_path), verbose)
...@@ -565,12 +677,12 @@ def _write_setup_file(name, ...@@ -565,12 +677,12 @@ def _write_setup_file(name,
def list2str(args): def list2str(args):
""" """
Convert list[str] into string. For example: [x, y] -> "['x', 'y']" Convert list[str] into string. For example: ['x', 'y'] -> "['x', 'y']"
""" """
if args is None: return '[]' if args is None: return '[]'
assert isinstance(args, (list, tuple)) assert isinstance(args, (list, tuple))
args = ["'{}'".format(arg) for arg in args] args = ["{}".format(arg) for arg in args]
return '[' + ','.join(args) + ']' return repr(args)
def _jit_compile(file_path, interpreter=None, verbose=False): def _jit_compile(file_path, interpreter=None, verbose=False):
...@@ -583,7 +695,8 @@ def _jit_compile(file_path, interpreter=None, verbose=False): ...@@ -583,7 +695,8 @@ def _jit_compile(file_path, interpreter=None, verbose=False):
if interpreter is None: if interpreter is None:
interpreter = 'python' interpreter = 'python'
try: try:
py_path = subprocess.check_output(['which', interpreter]) which = 'where' if IS_WINDOWS else 'which'
py_path = subprocess.check_output([which, interpreter])
py_version = subprocess.check_output([interpreter, '-V']) py_version = subprocess.check_output([interpreter, '-V'])
if six.PY3: if six.PY3:
py_path = py_path.decode() py_path = py_path.decode()
...@@ -596,8 +709,13 @@ def _jit_compile(file_path, interpreter=None, verbose=False): ...@@ -596,8 +709,13 @@ def _jit_compile(file_path, interpreter=None, verbose=False):
'Failed to check Python interpreter with `{}`, errors: {}'.format( 'Failed to check Python interpreter with `{}`, errors: {}'.format(
interpreter, error)) interpreter, error))
if IS_WINDOWS:
compile_cmd = 'cd /d {} && {} {} build'.format(ext_dir, interpreter,
setup_file)
else:
compile_cmd = 'cd {} && {} {} build'.format(ext_dir, interpreter, compile_cmd = 'cd {} && {} {} build'.format(ext_dir, interpreter,
setup_file) setup_file)
print("Compiling user custom op, it will cost a few seconds.....") print("Compiling user custom op, it will cost a few seconds.....")
run_cmd(compile_cmd, verbose) run_cmd(compile_cmd, verbose)
...@@ -682,7 +800,7 @@ def check_abi_compatibility(compiler, verbose=False): ...@@ -682,7 +800,7 @@ def check_abi_compatibility(compiler, verbose=False):
try: try:
if OS_NAME.startswith('linux'): if OS_NAME.startswith('linux'):
version_info = subprocess.check_output( version_info = subprocess.check_output(
[compiler, '-dumpfullversion']) [compiler, '-dumpfullversion', '-dumpversion'])
if six.PY3: if six.PY3:
version_info = version_info.decode() version_info = version_info.decode()
version = version_info.strip().split('.') version = version_info.strip().split('.')
...@@ -694,8 +812,8 @@ def check_abi_compatibility(compiler, verbose=False): ...@@ -694,8 +812,8 @@ def check_abi_compatibility(compiler, verbose=False):
warnings.warn( warnings.warn(
ABI_INCOMPATIBILITY_WARNING.format( ABI_INCOMPATIBILITY_WARNING.format(
user_compiler=compiler, version=version_info.strip())) user_compiler=compiler, version=version_info.strip()))
# TODO(Aurelius84): check version compatibility on windows
elif IS_WINDOWS: elif IS_WINDOWS:
# TODO(zhouwei): support check abi compatibility on windows
warnings.warn("We don't support Windows now.") warnings.warn("We don't support Windows now.")
except Exception: except Exception:
_, error, _ = sys.exc_info() _, error, _ = sys.exc_info()
...@@ -714,7 +832,7 @@ def _expected_compiler_current_platform(): ...@@ -714,7 +832,7 @@ def _expected_compiler_current_platform():
return expect_compilers return expect_compilers
def log_v(info, verbose): def log_v(info, verbose=True):
""" """
Print log information on stdout. Print log information on stdout.
""" """
......
python/requirements.txt
浏览文件 @ d3e60959
...@@ -3,7 +3,8 @@ numpy>=1.13, <=1.16.4 ; python_version<"3.5" ...@@ -3,7 +3,8 @@ numpy>=1.13, <=1.16.4 ; python_version<"3.5"
numpy>=1.13 ; python_version>="3.5" and platform_system != "Windows" numpy>=1.13 ; python_version>="3.5" and platform_system != "Windows"
numpy>=1.13, <=1.19.3 ; python_version>="3.5" and platform_system == "Windows" numpy>=1.13, <=1.19.3 ; python_version>="3.5" and platform_system == "Windows"
protobuf>=3.1.0 protobuf>=3.1.0
gast==0.3.3 gast>=0.3.3 ; platform_system != "Windows"
gast==0.3.3 ; platform_system == "Windows"
Pillow Pillow
six six
decorator decorator
......
python/setup.py.in
浏览文件 @ d3e60959
...@@ -334,11 +334,21 @@ if '${WITH_XPU_BKCL}' == 'ON': ...@@ -334,11 +334,21 @@ if '${WITH_XPU_BKCL}' == 'ON':
shutil.copy('${XPU_BKCL_LIB}', libs_path) shutil.copy('${XPU_BKCL_LIB}', libs_path)
package_data['paddle.libs']+=['${XPU_BKCL_LIB_NAME}'] package_data['paddle.libs']+=['${XPU_BKCL_LIB_NAME}']
# copy libfuild_framework.so to libs # copy libpaddle_framework.so to libs on linux
if os.name != 'nt' and sys.platform != 'darwin': if sys.platform.startswith('linux'):
paddle_framework_lib='${FLUID_FRAMEWORK_SHARED_LIB}' shutil.copy('${FLUID_FRAMEWORK_SHARED_LIB}', libs_path)
shutil.copy(paddle_framework_lib, libs_path) package_data['paddle.libs'] += ['libpaddle_framework.so']
package_data['paddle.libs'] += [('libpaddle_framework' if os.name != 'nt' else 'paddle_framework') + ext_name]
# copy libpaddle_custom_op.so to libs on linux
if sys.platform.startswith('linux'):
shutil.copy('${PADDLE_CUSTOM_OP_SHARED_LIB}', libs_path)
package_data['paddle.libs'] += ['libpaddle_custom_op.so']
# copy paddle_framework.lib/paddle_framework.dll to libs on windows
if os.name == 'nt':
shutil.copy('${FLUID_FRAMEWORK_IMPORT_LIB}', libs_path)
shutil.copy('${FLUID_FRAMEWORK_SHARED_LIB}', libs_path)
package_data['paddle.libs'] += ['paddle_framework.lib', 'paddle_framework.dll']
# remove unused paddle/libs/__init__.py # remove unused paddle/libs/__init__.py
if os.path.isfile(libs_path+'/__init__.py'): if os.path.isfile(libs_path+'/__init__.py'):
...@@ -409,9 +419,9 @@ if '${WITH_GPU}' == 'ON': ...@@ -409,9 +419,9 @@ if '${WITH_GPU}' == 'ON':
class InstallCommand(InstallCommandBase): class InstallCommand(InstallCommandBase):
def finalize_options(self): def finalize_options(self):
ret = InstallCommandBase.finalize_options(self) ret = InstallCommandBase.finalize_options(self)
self.install_headers = os.path.join(self.install_purelib, 'paddle',
'include')
self.install_lib = self.install_platlib self.install_lib = self.install_platlib
self.install_headers = os.path.join(self.install_platlib, 'paddle',
'include')
return ret return ret
...@@ -462,11 +472,6 @@ class InstallHeaders(Command): ...@@ -462,11 +472,6 @@ class InstallHeaders(Command):
return self.copy_file(header, install_dir) return self.copy_file(header, install_dir)
def run(self): def run(self):
# only copy third_party/cudaErrorMessage.pb for cudaErrorMessage on mac or windows
if os.name == 'nt' or sys.platform == 'darwin':
if '${WITH_GPU}' == 'ON':
self.mkdir_and_copy_file('${cudaerror_INCLUDE_DIR}/cudaErrorMessage.pb')
return
hdrs = self.distribution.headers hdrs = self.distribution.headers
if not hdrs: if not hdrs:
return return
......
Markdown is supported
0% 或 .
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 或