提交 c86e7e2a 编写于 作者: T Tao Luo 提交者: GitHub

Merge pull request #4101 from luotao1/use_op

auto generate USE_OP() in pybind.cc
......@@ -22,6 +22,7 @@ cmake-build-*
# generated while compiling
python/paddle/v2/framework/core.so
paddle/pybind/pybind.h
CMakeFiles
cmake_install.cmake
paddle/.timestamp
......
......@@ -26,9 +26,9 @@ set(IGNORE_PATTERN
.*ImportanceSampler.*
.*cblas\\.h.*
.*\\.pb\\.txt
.*LtrDataProvider.*
.*MultiDataProvider.*
.*pb.*)
.*pb.*
.*pybind.h)
# add_style_check_target
#
......
......@@ -34,7 +34,7 @@ Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中,否则,CPU
注册Op | Op注册实现在`.cc`文件;Kernel注册CPU实现在`.cc`文件中,GPU实现在`.cu`文件中
实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下,文件命名以`*_op.h`(如有) 、 `*_op.cc``*_op.cu`(如有)结尾。
实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下,文件命名以`*_op.h`(如有) 、 `*_op.cc``*_op.cu`(如有)结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
下面以矩阵乘操作,即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
......@@ -224,45 +224,15 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
### 5. 编译
- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。
- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如,`mul_op` 依赖 `math_function`,需要在`CMakeLists.txt`中添加如下内容:
运行下面命令可以进行编译:
```
op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) +
```
- 运行下面命令可以进行编译:
```
make mul_op
```
```
make mul_op
```
## 绑定Python
- 绑定Python
在 [`paddle/pybind/pybind.cc
`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op,具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
```
USE_OP(mul);
```
如果只实现了CPU版本,则使用`USE_CPU_ONLY_OP`:
```
USE_CPU_ONLY_OP(gather);
```
如果OP不带Kernel,则使用`USE_NO_KENREL_OP`:
```
USE_NO_KENREL_OP(recurrent);
```
- 生成库
`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
系统会对新增的op自动绑定Python,并链接到生成的lib库中。
## 实现单元测试
......@@ -367,3 +337,10 @@ make test ARGS="-R test_mul_op -V"
```bash
ctest -R test_mul_op
```
## 注意事项
- 为每个Op创建单独的`*_op.h`(如有)、`*_op.cc``*_op.cu`(如有)。不允许一个文件中包含多个Op,这将会导致编译出错。
- 注册Op时的类型名,需要和该Op的名字一样。即不允许在`A_op.cc`里面,注册`REGISTER_OP(B, ...)`等,这将会导致单元测试出错。
- 如果Op没有实现GPU Kernel,请不要创建空的`*_op.cu`,这将会导致单元测试出错。
- 如果多个Op依赖一些共用的函数,可以创建非`*_op.*`格式的文件来存放,如`gather.h`文件。
file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt. DO NOT EDIT!\n\n")
function(op_library TARGET)
# op_library is a function to create op library. The interface is same as
# cc_library. But it handle split GPU/CPU code and link some common library
......@@ -11,6 +13,7 @@ function(op_library TARGET)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
set(pybind_flag 0)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
......@@ -46,15 +49,40 @@ function(op_library TARGET)
cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
endif()
# net_op doesn't need pybind
if ("${TARGET}" STREQUAL "net_op")
set(pybind_flag 1)
endif()
# pybind USE_NO_KERNEL_OP
file(READ ${TARGET}.cc TARGET_CONTENT)
string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")
string(REPLACE "_op" "" TARGET "${TARGET}")
if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
# pybind USE_CPU_ONLY_OP
list(LENGTH cu_srcs cu_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0)
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
# pybind USE_OP
if (${pybind_flag} EQUAL 0)
file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
endif()
endfunction()
add_subdirectory(math)
set(DEPS_OPS
recurrent_op)
op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS framework_proto tensor operator net_op)
DEPS framework_proto tensor net_op)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
foreach(src ${GENERAL_OPS})
......
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#define EIGEN_USE_GPU
#include "paddle/operators/concat_op.h"
namespace ops = paddle::operators;
// TODO(Yancey1989) Add GPU kernel
......@@ -12,7 +12,7 @@ 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 "paddle/operators/cross_entropy_op.h"
#include "paddle/operators/onehot_cross_entropy_op.h"
namespace paddle {
namespace operators {
......
......@@ -23,6 +23,7 @@ limitations under the License. */
#include "paddle/operators/recurrent_op.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "paddle/pybind/pybind.h"
#include "paddle/pybind/tensor_py.h"
#include "paddle/string/to_string.h"
#include "pybind11/numpy.h"
......@@ -31,34 +32,6 @@ limitations under the License. */
namespace py = pybind11;
USE_OP(add);
USE_OP(onehot_cross_entropy);
USE_OP(sgd);
USE_OP(mul);
USE_OP(elementwise_mul);
USE_OP(mean);
USE_OP(sigmoid);
USE_OP(softmax);
USE_OP(rowwise_add);
USE_OP(fill_zeros_like);
USE_NO_KERNEL_OP(recurrent);
USE_OP(gaussian_random);
USE_OP(uniform_random);
USE_OP(lookup_table);
USE_OP(scale);
USE_NO_KERNEL_OP(identity);
USE_OP(minus);
USE_OP(cos_sim);
USE_CPU_ONLY_OP(gather);
USE_OP(pad);
USE_CPU_ONLY_OP(scatter);
USE_OP(accuracy);
USE_CPU_ONLY_OP(concat);
USE_OP(top_k);
USE_OP(squared_l2_distance);
USE_OP(sum);
USE_OP(reshape);
namespace paddle {
namespace framework {
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册