Skip to content

P
Paddle

PaddlePaddle / Paddle
9 个月 前同步成功

通知 2282
Star 20931
Fork 5422
P
Paddle

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

提交 b3a11fdf 编写于 作者: F fengjiayi
浏览文件
操作

Merge branch 'rm_reader_HasNext' into dev_double_buffer_for_cpp_reader

上级 b1f647fd 6e5736e2
展开全部 隐藏空白更改
内联 并排
Showing with 3105 addition and 795 deletion
.travis.yml
浏览文件 @ b3a11fdf
......@@ -56,7 +56,7 @@ script:
export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
export DOCS_DIR=`pwd`
cd ..
curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/v2
curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/
notifications:
email:
on_success: change
......
cmake/external/openblas.cmake
浏览文件 @ b3a11fdf
......@@ -77,7 +77,8 @@ IF(NOT ${CBLAS_FOUND})
INSTALL_DIR ${CBLAS_INSTALL_DIR}
BUILD_IN_SOURCE 1
BUILD_COMMAND ${CMAKE_MAKE_PROGRAM} ${COMMON_ARGS} ${OPTIONAL_ARGS}
INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 NO_LAPACK=1 PREFIX=<INSTALL_DIR>
INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 NO_LAPACK=1 PREFIX=<INSTALL_DIR>
&& rm -r ${CBLAS_INSTALL_DIR}/lib/cmake ${CBLAS_INSTALL_DIR}/lib/pkgconfig
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
)
......@@ -100,11 +101,6 @@ IF(NOT ${CBLAS_FOUND})
\"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
)"
)
INSTALL(CODE "execute_process(
COMMAND rm -r ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}/cmake
${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}/pkgconfig
)"
)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})
......
cmake/generic.cmake
浏览文件 @ b3a11fdf
......@@ -186,7 +186,9 @@ function(cc_library TARGET_NAME)
add_library(${TARGET_NAME} SHARED ${cc_library_SRCS})
else()
add_library(${TARGET_NAME} STATIC ${cc_library_SRCS})
find_fluid_modules(${TARGET_NAME})
endif()
if(cc_library_DEPS)
# Don't need link libwarpctc.so
if("${cc_library_DEPS};" MATCHES "warpctc;")
......@@ -263,7 +265,8 @@ function(nv_library TARGET_NAME)
if (nv_library_SHARED OR nv_library_shared) # build *.so
cuda_add_library(${TARGET_NAME} SHARED ${nv_library_SRCS})
else()
cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
find_fluid_modules(${TARGET_NAME})
endif()
if (nv_library_DEPS)
add_dependencies(${TARGET_NAME} ${nv_library_DEPS})
......
cmake/inference_lib.cmake
浏览文件 @ b3a11fdf
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
# make package for paddle fluid shared and static library
function(copy TARGET)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS DSTS DEPS)
cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE)
list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
......@@ -42,13 +55,21 @@ copy(glog_lib
DSTS ${dst_dir} ${dst_dir}/lib
)
IF(NOT PROTOBUF_FOUND)
if(NOT PROTOBUF_FOUND)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf")
copy(protobuf_lib
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LITE_LIBRARY}
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
DSTS ${dst_dir} ${dst_dir}/lib
)
ENDIF(NOT PROTOBUF_FOUND)
endif()
if(NOT CBLAS_FOUND)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
)
endif()
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
......@@ -66,8 +87,8 @@ copy(memory_lib
)
set(module "inference")
copy(inference_lib DEPENDS paddle_fluid_shared
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.so
copy(inference_lib DEPS paddle_fluid_shared paddle_fluid
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
DSTS ${dst_dir}/${module} ${dst_dir}/${module}
)
......@@ -83,6 +104,4 @@ copy(string_lib
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
)
add_custom_target(inference_lib_dist DEPENDS
inference_lib framework_lib memory_lib platform_lib string_lib
gflags_lib glog_lib protobuf_lib eigen3_lib)
add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep})
doc/design/cpp_data_feeding.md
浏览文件 @ b3a11fdf
......@@ -20,9 +20,8 @@ class ReaderBase {
PADDLE_ENFORCE(!shapes_.empty());
}
// Read the next batch of data. (A 'batch' can be only one instance)
// If the next batch doesn't exist, the '*out' will be an empty std::vector.
virtual void ReadNext(std::vector<LoDTensor>* out) = 0;
// Show whether the next bacth exists.
virtual bool HasNext() const = 0;
// Reinitialize the reader and read the file from the begin.
virtual void ReInit() = 0;
......
doc/fluid/howto/optimization/timeline.md 0 → 100644
浏览文件 @ b3a11fdf
## how to use timeline tool to do profile
1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
```python
with profiler.profiler('All', 'total', '/tmp/profile') as prof:
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[],
use_program_cache=True)
...
```
1. Run `python paddle/tools/timeline.py` to process `/tmp/profile`, it will generate another
file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at
[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details.
1. Open chrome and visit <chrome://tracing/>, use `load` button to load the generated `timeline` file.
![chrome tracing](./tracing.jpeg)
1. The resulting timeline should be like:
![chrome timeline](./timeline.jpeg)
doc/fluid/read_source.md
浏览文件 @ b3a11fdf
......@@ -2,17 +2,17 @@
Examples: https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid/tests/book
Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework
Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework
Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators
Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators
Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory
Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory
Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform
Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform
# Compile Time
The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto).
The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto).
```python
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
......@@ -29,10 +29,10 @@ sgd_optimizer.minimize(avg_cost)
- Variables: `x`, `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/framework.py#)
- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid/layers)
- Every Layer has one or more operators and variables/parameters
- All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files:
- Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h)
- Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h)
- Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h)
- All the operators are defined at [`paddle/fluid/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators). Other worth-looking files:
- Base class: [`paddle/fluid/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/operator.h)
- Operator Registration: [`paddle/fluid/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/op_registry.h)
- Operator Lookup: [`paddle/fluid/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/op_info.h)
- Optimizer: `fluid.optimizer.SGD`. It does the following
- Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/backward.py)]
- Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/optimizer.py)]
......@@ -55,13 +55,13 @@ exe.run(fluid.default_main_program(),
fetch_list=[avg_cost])
```
- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h)
- The device handle are at [paddle/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h)
- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)]
- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/place.h)
- The device handle are at [paddle/fluid/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/device_context.h)
- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/executor.cc)]
- Feeds the data: `feed=feeder.feed(data)`
- Evaluates all the operators
- Fetches the result: `fetch_list=[avg_cost]`
- Other worth looking files:
- Scope: [paddle/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h). Where all the variables live
- Variable: [paddle/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h). Where all the data (most likely tensors) live
- Tensor: [paddle/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h). Where we allocate memory through [`paddle/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory)
- Scope: [paddle/fluid/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/scope.h). Where all the variables live
- Variable: [paddle/fluid/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/variable.h). Where all the data (most likely tensors) live
- Tensor: [paddle/fluid/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/tensor.h). Where we allocate memory through [`paddle/fluid/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory)
doc/v2/build_and_install/pip_install_cn.rst
浏览文件 @ b3a11fdf
......@@ -39,7 +39,7 @@ PaddlePaddle可以使用常用的Python包管理工具
"cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddle.tgz>`_"
"cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
......
doc/v2/build_and_install/pip_install_en.rst
浏览文件 @ b3a11fdf
......@@ -42,7 +42,7 @@ If the links below shows up the login form, just click "Log in as guest" to star
"cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddle.tgz>`_"
"cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
......
doc/v2/howto/index_cn.rst
浏览文件 @ b3a11fdf
进阶使用
========
PaddlePaddle支持用户灵活地设置各种命令行参数，以实现对模型训练或预测流程的控制。使用方式请参考：
.. toctree::
:maxdepth: 1
cmd_parameter/index_cn.rst
PaddlePaddle支持在fabric集群、MPI集群、kubernetes集群上分布式训练任务，具体环境配置和使用说明请参考：
.. toctree::
:maxdepth: 1
cluster/index_cn.rst
PaddlePaddle提供了用于预测的C-API，关于C-API的使用，我们提供了如下指南:
.. toctree::
:maxdepth: 1
capi/index_cn.rst
PaddlePaddle支持多种灵活和高效的循环神经网络，具体配置使用方式请参考：
.. toctree::
:maxdepth: 1
rnn/index_cn.rst
关于如何使用内置的定时工具、nvprof 或 nvvp 来运行性能分析和调优，请参考：
.. toctree::
:maxdepth: 1
optimization/gpu_profiling_cn.rst
doc/v2/howto/rnn/index_cn.rst
浏览文件 @ b3a11fdf
RNN模型
===========
循环神经网络（RNN）是对序列数据建模的重要工具。PaddlePaddle提供了灵活的接口以支持复杂循环神经网络的构建。
这里将分为以下四个部分详细介绍如何使用PaddlePaddle搭建循环神经网络。
第一部分由浅入深的展示了使用PaddlePaddle搭建循环神经网络的全貌：首先以简单的循环神经网络（vanilla RNN）为例，
说明如何封装配置循环神经网络组件；然后更进一步的通过序列到序列（sequence to sequence）模型，逐步讲解如何构建完整而复杂的循环神经网络模型。
.. toctree::
:maxdepth: 1
rnn_config_cn.rst
Recurrent Group是PaddlePaddle中实现复杂循环神经网络的关键，第二部分阐述了PaddlePaddle中Recurrent Group的相关概念和原理，
对Recurrent Group接口进行了详细说明。另外，对双层RNN（对应的输入为双层序列）及Recurrent Group在其中的使用进行了介绍。
.. toctree::
:maxdepth: 1
recurrent_group_cn.md
第三部分对双层序列进行了解释说明，列出了PaddlePaddle中支持双层序列作为输入的Layer，并对其使用进行了逐一介绍。
.. toctree::
:maxdepth: 1
hierarchical_layer_cn.rst
第四部分以PaddlePaddle的双层RNN单元测试中的网络配置为示例，辅以效果相同的单层RNN网络配置作为对比，讲解了多种情况下双层RNN的使用。
.. toctree::
:maxdepth: 1
hrnn_rnn_api_compare_cn.rst
paddle/fluid/framework/data_type_transform.cc
浏览文件 @ b3a11fdf
......@@ -53,6 +53,7 @@ struct CastDataType {
auto* context = static_cast<const platform::CUDADeviceContext*>(ctx_);
trans(*context, in_begin, in_end, out_begin,
CastDataTypeFunctor<InType, OutType>());
context->Wait();
#endif
} else {
PADDLE_THROW("Unsupported place!");
......
paddle/fluid/framework/data_type_transform_test.cc
浏览文件 @ b3a11fdf
......@@ -50,13 +50,13 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp32, kernel_fp64, in, &out);
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(i / 3));
EXPECT_EQ(out_data_double[i], static_cast<double>(i / 3));
}
TransDataType(kernel_fp32, kernel_int32, in, &out);
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(i / 3));
EXPECT_EQ(out_data_int[i], static_cast<int>(i / 3));
}
}
......@@ -76,31 +76,31 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp16, kernel_fp32, in, &out);
float* out_data_float = out.data<float>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
EXPECT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_fp64, in, &out);
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int32, in, &out);
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int64, in, &out);
int64_t* out_data_int64 = out.data<int64_t>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
EXPECT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_bool, in, &out);
bool* out_data_bool = out.data<bool>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
EXPECT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
}
// transform float to float16
......@@ -112,7 +112,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp32, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
}
// transform double to float16
......@@ -124,7 +124,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp64, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
}
// transform int to float16
......@@ -136,7 +136,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_int32, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
}
// transform int64 to float16
......@@ -148,7 +148,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_int64, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
}
// transform bool to float16
......@@ -160,7 +160,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_bool, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
}
}
}
paddle/fluid/framework/data_type_transform_test.cu
浏览文件 @ b3a11fdf
......@@ -49,15 +49,16 @@ TEST(DataTypeTransform, GPUTransform) {
float arr[6] = {0, 1, 2, 3, 4, 5};
int data_number = sizeof(arr) / sizeof(arr[0]);
memcpy(in_ptr, arr, sizeof(arr));
TensorCopy(in, gpu_place, context, &in_gpu);
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp32, kernel_fp64, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(arr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(arr[i]));
}
TransDataType(kernel_fp32, kernel_int32, in_gpu, &out_gpu);
......@@ -66,7 +67,7 @@ TEST(DataTypeTransform, GPUTransform) {
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(arr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(arr[i]));
}
}
......@@ -83,6 +84,7 @@ TEST(DataTypeTransform, GPUTransform) {
int data_number = sizeof(arr) / sizeof(arr[0]);
memcpy(ptr, arr, sizeof(arr));
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
// transform from float16 to other data types
TransDataType(kernel_fp16, kernel_fp32, in_gpu, &out_gpu);
......@@ -91,7 +93,7 @@ TEST(DataTypeTransform, GPUTransform) {
float* out_data_float = out.data<float>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
EXPECT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_fp64, in_gpu, &out_gpu);
......@@ -100,7 +102,7 @@ TEST(DataTypeTransform, GPUTransform) {
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int32, in_gpu, &out_gpu);
......@@ -109,7 +111,7 @@ TEST(DataTypeTransform, GPUTransform) {
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int64, in_gpu, &out_gpu);
......@@ -118,7 +120,7 @@ TEST(DataTypeTransform, GPUTransform) {
int64_t* out_data_int64 = out.data<int64_t>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
EXPECT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_bool, in_gpu, &out_gpu);
......@@ -127,7 +129,7 @@ TEST(DataTypeTransform, GPUTransform) {
bool* out_data_bool = out.data<bool>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
EXPECT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
}
// transform float to float16
......@@ -137,13 +139,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp32, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
}
// transform double to float16
......@@ -154,13 +157,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp64, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
}
// transform int to float16
......@@ -170,13 +174,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_int32, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
}
// transform int64 to float16
......@@ -187,13 +192,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_int64, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
}
// transform bool to float16
......@@ -203,13 +209,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_bool, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
}
}
}
paddle/fluid/framework/executor.cc
浏览文件 @ b3a11fdf
......@@ -125,8 +125,9 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
for (auto& op_desc : block.AllOps()) {
auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
VLOG(4) << place_ << " " << op->DebugStringEx(local_scope);
op->Run(*local_scope, place_);
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
if (FLAGS_benchmark) {
VLOG(2) << "Memory used after operator " + op->Type() + " running: "
......
paddle/fluid/framework/reader.h
浏览文件 @ b3a11fdf
......@@ -26,7 +26,6 @@ class ReaderBase {
PADDLE_ENFORCE(!shapes_.empty());
}
virtual void ReadNext(std::vector<LoDTensor>* out) = 0;
virtual bool HasNext() const = 0;
virtual void ReInit() = 0;
......@@ -52,8 +51,6 @@ class DecoratedReader : public ReaderBase {
PADDLE_ENFORCE_NOT_NULL(reader_);
}
bool HasNext() const override { return reader_->HasNext(); }
void ReInit() override { reader_->ReInit(); }
protected:
......@@ -69,7 +66,6 @@ class ReaderHolder {
ReaderBase* Get() const { return reader_.get(); }
void ReadNext(std::vector<LoDTensor>* out) { reader_->ReadNext(out); }
bool HasNext() const { return reader_->HasNext(); }
void ReInit() { reader_->ReInit(); }
DDim shape(size_t idx) const { return reader_->shape(idx); }
......
paddle/fluid/framework/tensor_util.cc
浏览文件 @ b3a11fdf
......@@ -187,7 +187,6 @@ bool TensorContainsInf(const framework::Tensor& tensor) {
void TensorToStream(std::ostream& os, const Tensor& tensor,
const platform::DeviceContext& dev_ctx) {
// TODO(typhoonzero): serialize to ostream
{ // the 1st field, uint32_t version
constexpr uint32_t version = 0;
os.write(reinterpret_cast<const char*>(&version), sizeof(version));
......
paddle/fluid/framework/threadpool.h
浏览文件 @ b3a11fdf
......@@ -67,10 +67,10 @@ class ThreadPool {
} catch (platform::EnforceNotMet ex) {
return std::unique_ptr<platform::EnforceNotMet>(
new platform::EnforceNotMet(ex));
} catch (...) {
LOG(FATAL)
<< "Unexpected exception is catched in thread pool. All "
"throwable exception in Fluid should be an EnforceNotMet.";
} catch (const std::exception& e) {
LOG(FATAL) << "Unexpected exception is catched in thread pool. All "
"throwable exception in Fluid should be an EnforceNotMet."
<< e.what();
}
return nullptr;
});
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ b3a11fdf
......@@ -5,7 +5,8 @@ cc_library(paddle_fluid_api
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
# Create static library
cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
cc_library(paddle_fluid DEPS ${fluid_modules})
# Create shared library
cc_library(paddle_fluid_shared SHARED
......
paddle/fluid/inference/io.cc
浏览文件 @ b3a11fdf
......@@ -22,14 +22,14 @@ namespace paddle {
namespace inference {
void ReadBinaryFile(const std::string& filename, std::string& contents) {
VLOG(3) << "loading model from " << filename;
std::ifstream inputfs(filename, std::ios::in | std::ios::binary);
inputfs.seekg(0, std::ios::end);
std::ifstream fin(filename, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", filename);
fin.seekg(0, std::ios::end);
contents.clear();
contents.resize(inputfs.tellg());
inputfs.seekg(0, std::ios::beg);
inputfs.read(&contents[0], contents.size());
inputfs.close();
contents.resize(fin.tellg());
fin.seekg(0, std::ios::beg);
fin.read(&contents[0], contents.size());
fin.close();
}
bool IsPersistable(const framework::VarDesc* var) {
......@@ -97,6 +97,7 @@ std::unique_ptr<framework::ProgramDesc> Load(framework::Executor& executor,
const std::string& dirname) {
std::string model_filename = dirname + "/__model__";
std::string program_desc_str;
VLOG(3) << "loading model from " << model_filename;
ReadBinaryFile(model_filename, program_desc_str);
std::unique_ptr<framework::ProgramDesc> main_program(
......
paddle/fluid/inference/tests/book/test_inference_image_classification.cc
浏览文件 @ b3a11fdf
......@@ -17,10 +17,13 @@ limitations under the License. */
#include "paddle/fluid/inference/tests/test_helper.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_int32(batch_size, 1, "Batch size of input data");
DEFINE_int32(repeat, 1, "Running the inference program repeat times");
TEST(inference, image_classification) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model "
"--batch_size=1 --repeat=1";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
......@@ -29,13 +32,11 @@ TEST(inference, image_classification) {
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
int64_t batch_size = 1;
paddle::framework::LoDTensor input;
// Use normilized image pixels as input data,
// which should be in the range [0.0, 1.0].
SetupTensor<float>(input,
{batch_size, 3, 32, 32},
{FLAGS_batch_size, 3, 32, 32},
static_cast<float>(0),
static_cast<float>(1));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
......@@ -46,7 +47,9 @@ TEST(inference, image_classification) {
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << "--- CPU Runs: ---";
TestInference<paddle::platform::CPUPlace>(
dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
......@@ -55,7 +58,9 @@ TEST(inference, image_classification) {
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << "--- GPU Runs: ---";
TestInference<paddle::platform::CUDAPlace>(
dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);
......
paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc
浏览文件 @ b3a11fdf
......@@ -17,10 +17,13 @@ limitations under the License. */
#include "paddle/fluid/inference/tests/test_helper.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_int32(batch_size, 1, "Batch size of input data");
DEFINE_int32(repeat, 1, "Running the inference program repeat times");
TEST(inference, recognize_digits) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model "
"--batch_size=1 --repeat=1";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
......@@ -29,77 +32,39 @@ TEST(inference, recognize_digits) {
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
int64_t batch_size = 1;
paddle::framework::LoDTensor input;
// Use normilized image pixels as input data,
// which should be in the range [-1.0, 1.0].
SetupTensor<float>(input,
{batch_size, 1, 28, 28},
{FLAGS_batch_size, 1, 28, 28},
static_cast<float>(-1),
static_cast<float>(1));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
cpu_feeds.push_back(&input);
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
for (auto is_combined : {false, true}) {
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << output1.dims();
// Run inference on CPU
LOG(INFO) << "--- CPU Runs: is_combined=" << is_combined << " ---";
TestInference<paddle::platform::CPUPlace>(
dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << output2.dims();
// Run inference on CUDA GPU
LOG(INFO) << "--- GPU Runs: is_combined=" << is_combined << " ---";
TestInference<paddle::platform::CUDAPlace>(
dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat, is_combined);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);
CheckError<float>(output1, output2);
#endif
}
TEST(inference, recognize_digits_combine) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
std::string dirname = FLAGS_dirname;
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
paddle::framework::LoDTensor input;
// Use normilized image pixels as input data,
// which should be in the range [-1.0, 1.0].
SetupTensor<float>(
input, {1, 1, 28, 28}, static_cast<float>(-1), static_cast<float>(1));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
cpu_feeds.push_back(&input);
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace, true>(
dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace, true>(
dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);
#endif
}
paddle/fluid/inference/tests/test_helper.h
浏览文件 @ b3a11fdf
......@@ -15,6 +15,7 @@ limitations under the License. */
#include <time.h>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/profiler.h"
template <typename T>
void SetupTensor(paddle::framework::LoDTensor& input,
......@@ -87,31 +88,60 @@ void CheckError(paddle::framework::LoDTensor& output1,
EXPECT_EQ(count, 0U) << "There are " << count << " different elements.";
}
template <typename Place, bool IsCombined = false>
template <typename Place>
void TestInference(const std::string& dirname,
const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
std::vector<paddle::framework::LoDTensor*>& cpu_fetchs,
const int repeat = 1,
const bool is_combined = false) {
// 1. Define place, executor, scope
auto place = Place();
auto executor = paddle::framework::Executor(place);
auto* scope = new paddle::framework::Scope();
// Profile the performance
paddle::platform::ProfilerState state;
if (paddle::platform::is_cpu_place(place)) {
state = paddle::platform::ProfilerState::kCPU;
} else {
#ifdef PADDLE_WITH_CUDA
state = paddle::platform::ProfilerState::kCUDA;
// The default device_id of paddle::platform::CUDAPlace is 0.
// Users can get the device_id using:
// int device_id = place.GetDeviceId();
paddle::platform::SetDeviceId(0);
#else
PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
#endif
}
// Enable the profiler
paddle::platform::EnableProfiler(state);
// 2. Initialize the inference_program and load parameters
std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
if (IsCombined) {
// All parameters are saved in a single file.
// Hard-coding the file names of program and parameters in unittest.
// The file names should be consistent with that used in Python API
// `fluid.io.save_inference_model`.
std::string prog_filename = "__model_combined__";
std::string param_filename = "__params_combined__";
inference_program = paddle::inference::Load(executor,
*scope,
dirname + "/" + prog_filename,
dirname + "/" + param_filename);
} else {
// Parameters are saved in separate files sited in the specified `dirname`.
inference_program = paddle::inference::Load(executor, *scope, dirname);
{
paddle::platform::RecordEvent record_event(
"init_program",
paddle::platform::DeviceContextPool::Instance().Get(place));
if (is_combined) {
// All parameters are saved in a single file.
// Hard-coding the file names of program and parameters in unittest.
// The file names should be consistent with that used in Python API
// `fluid.io.save_inference_model`.
std::string prog_filename = "__model_combined__";
std::string param_filename = "__params_combined__";
inference_program =
paddle::inference::Load(executor,
*scope,
dirname + "/" + prog_filename,
dirname + "/" + param_filename);
} else {
// Parameters are saved in separate files sited in the specified
// `dirname`.
inference_program = paddle::inference::Load(executor, *scope, dirname);
}
}
// 3. Get the feed_target_names and fetch_target_names
......@@ -134,7 +164,21 @@ void TestInference(const std::string& dirname,
}
// 6. Run the inference program
executor.Run(*inference_program, scope, feed_targets, fetch_targets);
{
// Run repeat times to profile the performance
for (int i = 0; i < repeat; ++i) {
paddle::platform::RecordEvent record_event(
"run_inference",
paddle::platform::DeviceContextPool::Instance().Get(place));
executor.Run(*inference_program, scope, feed_targets, fetch_targets);
}
}
// Disable the profiler and print the timing information
paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
"profiler.txt");
paddle::platform::ResetProfiler();
delete scope;
}
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ b3a11fdf
file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE "_mkldnn" "" GENERAL_OPS "${GENERAL_OPS}")
string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
list(REMOVE_DUPLICATES GENERAL_OPS)
set(DEPS_OPS "")
set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/fluid/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt. DO NOT EDIT!\n\n")
......@@ -13,6 +15,8 @@ function(op_library TARGET)
set(cu_cc_srcs)
set(cudnn_cu_cc_srcs)
set(CUDNN_FILE)
set(mkldnn_cc_srcs)
set(MKLDNN_FILE)
set(op_common_deps operator op_registry math_function)
set(options "")
set(oneValueArgs "")
......@@ -36,12 +40,20 @@ function(op_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
endif()
if(WITH_MKLDNN)
string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
endif()
endif()
else()
foreach(src ${op_library_SRCS})
if (${src} MATCHES ".*\\.cu$")
list(APPEND cu_srcs ${src})
elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
list(APPEND cudnn_cu_cc_srcs ${src})
elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
list(APPEND mkldnn_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cu.cc$")
list(APPEND cu_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cc$")
......@@ -62,11 +74,11 @@ function(op_library TARGET)
set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
endif()
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
else()
cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
endif()
# Define operators that don't need pybind here.
......@@ -101,7 +113,8 @@ function(op_library TARGET)
# pybind USE_CPU_ONLY_OP
list(LENGTH cu_srcs cu_srcs_len)
list(LENGTH cu_cc_srcs cu_cc_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
......@@ -112,6 +125,11 @@ function(op_library TARGET)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
endif()
# pybind USE_OP_DEVICE_KERNEL for MKLDNN
if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
endif()
# pybind USE_OP
if (${pybind_flag} EQUAL 0)
file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
......@@ -172,17 +190,18 @@ op_library(cos_sim_op DEPS cos_sim_functor)
op_library(parallel_do_op DEPS executor)
if (WITH_GPU)
op_library(conv_op DEPS vol2col depthwise_conv)
op_library(conv_op DEPS vol2col depthwise_conv im2col)
else()
op_library(conv_op DEPS vol2col)
op_library(conv_op DEPS vol2col im2col)
endif()
op_library(conv_transpose_op DEPS vol2col)
op_library(conv_transpose_op DEPS vol2col im2col)
# FIXME(typhoonzero): save/load depends lodtensor serialization functions
op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
op_library(save_combine_op DEPS lod_tensor)
op_library(load_combine_op DEPS lod_tensor)
op_library(concat_op DEPS concat)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
foreach(src ${GENERAL_OPS})
......
paddle/fluid/operators/concat_op.cc
浏览文件 @ b3a11fdf
......@@ -100,7 +100,8 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
namespace ops = paddle::operators;
REGISTER_OP_EX(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad,
ops::ConcatOpGrad, false)
REGISTER_OP_CPU_KERNEL(concat,
ops::ConcatKernel<paddle::platform::CPUPlace, float>)
REGISTER_OP_CPU_KERNEL(concat_grad,
ops::ConcatGradKernel<paddle::platform::CPUPlace, float>)
REGISTER_OP_CPU_KERNEL(
concat, ops::ConcatKernel<paddle::platform::CPUDeviceContext, float>)
REGISTER_OP_CPU_KERNEL(
concat_grad,
ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, float>)
paddle/fluid/operators/concat_op.h
浏览文件 @ b3a11fdf
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <utility>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/concat.h"
#include "paddle/fluid/operators/strided_memcpy.h"
namespace paddle {
......@@ -27,54 +28,30 @@ class ConcatKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto ins = ctx.MultiInput<framework::Tensor>("X");
auto* out = ctx.Output<framework::Tensor>("Out");
framework::Tensor* out = ctx.Output<framework::Tensor>("Out");
int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
auto place = ctx.GetPlace();
out->mutable_data<T>(place);
auto out_stride = framework::stride_numel(out->dims());
size_t output_offset = 0;
// If axis >=1, copy to out immediately need to call many times
// of cuda memcpy. Copy the input to cpu and do the stride copy,
// then copy to gpu output.
if (platform::is_gpu_place(place) && axis >= 1) {
platform::CPUPlace copy_place;
auto& cpu_ctx = *platform::DeviceContextPool::Instance().Get(copy_place);
framework::Tensor cpu_out;
cpu_out.Resize(out->dims());
cpu_out.mutable_data<T>(copy_place);
auto& dev_ctx = ctx.device_context();
std::vector<std::unique_ptr<framework::Tensor>> cpu_ins;
for (auto* in : ins) {
std::unique_ptr<framework::Tensor> cpu_in(new framework::Tensor);
framework::TensorCopy(*in, copy_place, dev_ctx, cpu_in.get());
cpu_ins.emplace_back(std::move(cpu_in));
}
// TODO(dzhwinter): overlap copy and compute stream
// https://devblogs.nvidia.com/how-overlap-data-transfers-cuda-cc/
dev_ctx.Wait();
for (auto& in : cpu_ins) {
auto& cpu_in = *in.get();
auto in_stride = framework::stride_numel(cpu_in.dims());
StridedNumelCopyWithAxis<T>(
cpu_ctx, axis, cpu_out.data<T>() + output_offset, out_stride,
cpu_in.data<T>(), in_stride, in_stride[axis]);
output_offset += in_stride[axis];
}
framework::TensorCopy(cpu_out, place, dev_ctx, out);
} else {
// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && ins.size() < 10) {
size_t output_offset = 0;
for (auto* in : ins) {
auto in_stride = framework::stride_numel(in->dims());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis,
out->data<T>() + output_offset, out_stride,
in->data<T>(), in_stride, in_stride[axis]);
output_offset += in_stride[axis];
}
} else {
std::vector<framework::Tensor> inputs(ins.size());
for (size_t j = 0; j < ins.size(); ++j) {
inputs[j] = *ins[j];
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatFunctor<DeviceContext, T> concat_functor;
concat_functor(dev_ctx, inputs, static_cast<int>(axis), out);
}
}
};
......@@ -86,16 +63,31 @@ class ConcatGradKernel : public framework::OpKernel<T> {
auto* in = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
size_t input_offset = 0;
auto in_stride = framework::stride_numel(in->dims());
for (auto& out : outs) {
out->mutable_data<T>(ctx.GetPlace());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
out_stride, in->data<T>() + input_offset,
in_stride, out_stride[axis]);
input_offset += out_stride[axis];
// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && outs.size() < 10) {
size_t input_offset = 0;
auto in_stride = framework::stride_numel(in->dims());
for (auto& out : outs) {
out->mutable_data<T>(ctx.GetPlace());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
out_stride, in->data<T>() + input_offset,
in_stride, out_stride[axis]);
input_offset += out_stride[axis];
}
} else {
std::vector<framework::Tensor> outputs(outs.size());
for (size_t j = 0; j < outs.size(); ++j) {
outs[j]->mutable_data<T>(ctx.GetPlace());
outputs[j] = *outs[j];
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
concat_grad_functor;
concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), outputs);
}
}
};
......
paddle/fluid/operators/conv_mkldnn_op.cc 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2018 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 "mkldnn.hpp"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/conv_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
namespace paddle {
namespace operators {
using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;
using paddle::platform::MKLDNNMemDesc;
using mkldnn::memory; // Note: paddle has also "memory" namespace
using mkldnn::primitive;
using mkldnn::convolution_forward;
using mkldnn::convolution_backward_weights;
using mkldnn::convolution_backward_data;
using mkldnn::convolution_direct;
using mkldnn::prop_kind;
using mkldnn::padding_kind;
using mkldnn::stream;
namespace {
std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
const memory::desc& dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const mkldnn::engine& engine);
convolution_backward_weights::primitive_desc ConvBwdWeightsPrimitiveDesc(
const memory::desc& src, const memory::desc& diff_weights,
const memory::desc& diff_dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const convolution_forward::primitive_desc& conv_pd,
const mkldnn::engine& engine);
convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc(
const memory::desc& diff_src, const memory::desc& weights,
const memory::desc& diff_dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const convolution_forward::primitive_desc& conv_pd,
const mkldnn::engine& engine);
} // anonymous namespace
template <typename T>
class ConvOpMkldnnKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine();
auto* input = ctx.Input<Tensor>("Input");
auto* filter = ctx.Input<Tensor>("Filter");
auto* output = ctx.Output<Tensor>("Output");
// Get an unique name from "argument" name of "Output" variable
// This name will be used as key when saving info into device context
const std::string key = ctx.op().Output("Output");
const std::string key_conv_pd = key + "@conv_pd";
std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
int groups = ctx.Attr<int>("groups");
// TODO(pzelazko-intel) add support for group convolution and dilation
PADDLE_ENFORCE(groups == 1, "group convolution is not implemented yet");
PADDLE_ENFORCE(
dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
"dilation in convolution is not implemented yet");
const T* input_data = input->data<T>();
const T* filter_data = filter->data<T>();
// allocate memory for output
T* output_data = output->mutable_data<T>(ctx.GetPlace());
PADDLE_ENFORCE(input->dims().size() == 4,
"Input must be with 4 dimensions, i.e. NCHW");
PADDLE_ENFORCE(filter->dims().size() == 4,
"Filter must be with 4 dimensions, i.e. OIHW");
std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
std::vector<int> weights_tz =
paddle::framework::vectorize2int(filter->dims());
std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
// TODO(pzelazko-intel): support more formats
// memory descriptors for convolution src/weight/dst
auto conv_src_md =
MKLDNNMemDesc(src_tz, memory::data_type::f32, memory::format::nchw);
auto conv_weights_md =
MKLDNNMemDesc(weights_tz, memory::data_type::f32, memory::format::oihw);
auto conv_dst_md =
MKLDNNMemDesc(dst_tz, memory::data_type::f32, memory::format::nchw);
// create memory primitives
auto conv_src_memory =
memory({conv_src_md, mkldnn_engine}, (void*)input_data);
auto conv_weights_memory =
memory({conv_weights_md, mkldnn_engine}, (void*)filter_data);
auto conv_dst_memory = memory({conv_dst_md, mkldnn_engine}, output_data);
std::unique_ptr<convolution_forward::primitive_desc> conv_pd =
ConvFwdPrimitiveDesc(conv_src_md, conv_weights_md, conv_dst_md, strides,
paddings, mkldnn_engine);
// save p_conv_pd into dev_ctx to be referred in backward path
auto p_conv_pd = conv_pd.get();
std::shared_ptr<void> conv_pd_value = std::move(conv_pd);
dev_ctx.SetBlob(key_conv_pd, conv_pd_value);
// create convolution op primitive
auto conv_prim = convolution_forward(*p_conv_pd, conv_src_memory,
conv_weights_memory, conv_dst_memory);
// push op to stream and wait MKLDNN until it's executed
std::vector<primitive> pipeline{conv_prim};
stream(stream::kind::eager).submit(pipeline).wait();
}
};
template <typename T>
class ConvGradOpMkldnnKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine();
const Tensor* input = ctx.Input<Tensor>("Input");
const Tensor* filter = ctx.Input<Tensor>("Filter");
const Tensor* output = ctx.Input<Tensor>("Output");
const Tensor* output_grad =
ctx.Input<Tensor>(framework::GradVarName("Output"));
Tensor* input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
Tensor* filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));
if (!input_grad && !filter_grad) return;
// Get an unique name from "argument" name of "Output" variable
// This name will be used as key when saving info into device context
const std::string key = ctx.op().Input("Output");
const std::string key_conv_pd = key + "@conv_pd";
std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
const T* input_data = input->data<T>();
const T* filter_data = filter->data<T>();
const T* output_grad_data = output_grad->data<T>();
T* input_grad_data = nullptr;
T* filter_grad_data = nullptr;
// allocate memory for gradient of input/filter
if (input_grad) {
input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
}
if (filter_grad) {
filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
}
std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
std::vector<int> weights_tz =
paddle::framework::vectorize2int(filter->dims());
std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
// TODO(pzelazko-intel): support more formats
auto conv_src_md =
MKLDNNMemDesc(src_tz, memory::data_type::f32, memory::format::nchw);
auto conv_diff_src_md =
MKLDNNMemDesc(src_tz, memory::data_type::f32, memory::format::nchw);
auto conv_weights_md =
MKLDNNMemDesc(weights_tz, memory::data_type::f32, memory::format::oihw);
auto conv_diff_weights_md =
MKLDNNMemDesc(weights_tz, memory::data_type::f32, memory::format::oihw);
auto conv_diff_dst_md =
MKLDNNMemDesc(dst_tz, memory::data_type::f32, memory::format::nchw);
// create memory
auto conv_diff_dst_memory =
memory({conv_diff_weights_md, mkldnn_engine}, (void*)output_grad_data);
// Retrieve conv_pd from device context
std::shared_ptr<void> conv_pd;
convolution_forward::primitive_desc* p_conv_pd;
conv_pd = dev_ctx.GetBlob(key_conv_pd);
PADDLE_ENFORCE(conv_pd != nullptr,
"Fail to find conv_pd in device context");
p_conv_pd =
static_cast<convolution_forward::primitive_desc*>(conv_pd.get());
// create backward conv primitive for weights
if (filter_grad) {
// create primitive descriptor
convolution_backward_weights::primitive_desc conv_bwd_weights_pd =
ConvBwdWeightsPrimitiveDesc(conv_src_md, conv_diff_weights_md,
conv_diff_dst_md, strides, paddings,
*p_conv_pd, mkldnn_engine);
// create memory
auto conv_diff_weights_memory = memory(
{conv_diff_weights_md, mkldnn_engine}, (void*)filter_grad_data);
auto conv_src_memory =
memory({conv_src_md, mkldnn_engine}, (void*)input_data);
// create backward conv primitive for weights
auto conv_bwd_weights_prim = convolution_backward_weights(
conv_bwd_weights_pd, conv_src_memory, conv_diff_dst_memory,
conv_diff_weights_memory);
// push primitive and execute it
std::vector<primitive> pipeline{conv_bwd_weights_prim};
stream(stream::kind::eager).submit(pipeline).wait();
}
if (input_grad) {
// create primitive descriptor
convolution_backward_data::primitive_desc conv_bwd_data_pd =
ConvBwdDataPrimitiveDesc(conv_diff_src_md, conv_weights_md,
conv_diff_dst_md, strides, paddings,
*p_conv_pd, mkldnn_engine);
// create memory
auto conv_diff_src_memory =
memory({conv_diff_src_md, mkldnn_engine}, (void*)input_grad_data);
auto conv_weights_memory =
memory({conv_weights_md, mkldnn_engine}, (void*)filter_data);
// create backward conv primitive for data
auto conv_bwd_data_prim =
convolution_backward_data(conv_bwd_data_pd, conv_diff_dst_memory,
conv_weights_memory, conv_diff_src_memory);
// push primitive and execute it
std::vector<primitive> pipeline{conv_bwd_data_prim};
stream(stream::kind::eager).submit(pipeline).wait();
}
} // Compute()
};
namespace {
std::unique_ptr<convolution_forward::primitive_desc> ConvFwdPrimitiveDesc(
const memory::desc& src, const memory::desc& weights,
const memory::desc& dst, const std::vector<int>& strides,
const std::vector<int>& paddings, const mkldnn::engine& engine) {
mkldnn::memory::dims stride_dims = {strides[0], strides[1]};
mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]};
auto conv_desc = mkldnn::convolution_forward::desc(
mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights, dst,
stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
auto p_conv_pd = new convolution_forward::primitive_desc(conv_desc, engine);
return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
p_conv_pd);
}
convolution_backward_weights::primitive_desc ConvBwdWeightsPrimitiveDesc(
const memory::desc& src, const memory::desc& diff_weights,
const memory::desc& diff_dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const convolution_forward::primitive_desc& conv_pd,
const mkldnn::engine& engine) {
auto conv_bwd_weights_desc = convolution_backward_weights::desc(
convolution_direct, src, diff_weights, diff_dst, strides, paddings,
paddings, padding_kind::zero);
return convolution_backward_weights::primitive_desc(conv_bwd_weights_desc,
engine, conv_pd);
}
convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc(
const memory::desc& diff_src, const memory::desc& weights,
const memory::desc& diff_dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const convolution_forward::primitive_desc& conv_pd,
const mkldnn::engine& engine) {
auto conv_bwd_data_desc = convolution_backward_data::desc(
convolution_direct, diff_src, weights, diff_dst, strides, paddings,
paddings, padding_kind::zero);
return convolution_backward_data::primitive_desc(conv_bwd_data_desc, engine,
conv_pd);
}
} // anonymous namespace
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_KERNEL(conv2d, MKLDNN, ::paddle::platform::CPUPlace,
ops::ConvOpMkldnnKernel<float>);
REGISTER_OP_KERNEL(conv2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
ops::ConvGradOpMkldnnKernel<float>);
paddle/fluid/operators/conv_op.cc
浏览文件 @ b3a11fdf
......@@ -13,6 +13,12 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/conv_op.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cudnn_helper.h"
#endif
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace operators {
......@@ -64,22 +70,21 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
framework::LibraryType library_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
} else {
library_ = framework::LibraryType::kPlain;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
......@@ -131,6 +136,9 @@ Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
"use_cudnn",
"(bool, default false) Only used in cudnn kernel, need install cudnn")
.SetDefault(false);
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddAttr<std::string>(
"data_format",
"(string, default NCHW) Only used in "
......@@ -224,6 +232,9 @@ Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
"use_cudnn",
"(bool, default false) Only used in cudnn kernel, need install cudnn")
.SetDefault(false);
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddAttr<std::string>(
"data_format",
"(string, default NCHW) Only used in "
......@@ -284,23 +295,21 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
framework::LibraryType library_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
} else {
library_ = framework::LibraryType::kPlain;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
......
paddle/fluid/operators/detail/CMakeLists.txt
浏览文件 @ b3a11fdf
if(WITH_DISTRIBUTE)
grpc_library(sendrecvop_grpc SRCS sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(test_serde.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
cc_test(serde_test SRCS test_serde.cc DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc)
endif()
paddle/fluid/operators/detail/bytebuffer_stream.cc 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2016 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. */
// NOTE: This file was originally created by tensorflow
// (https://github.com/tensorflow/tensorflow/) we borrow this
// file and did some modifications so that we can send gRPC
// requests without too much copying of the tensor data.
#include "bytebuffer_stream.h"
namespace paddle {
namespace operators {
namespace detail {
GrpcByteBufferSource::GrpcByteBufferSource() {}
bool GrpcByteBufferSource::Init(const grpc::ByteBuffer& src) {
cur_ = -1;
left_ = 0;
ptr_ = nullptr;
byte_count_ = 0;
bool ok = src.Dump(&slices_).ok();
if (!ok) {
slices_.clear();
}
return ok;
}
bool GrpcByteBufferSource::Next(const void** data, int* size) {
// Use loop instead of if in case buffer contained empty slices.
while (left_ == 0) {
// Advance to next slice.
cur_++;
if (cur_ >= slices_.size()) {
return false;
}
const ::grpc::Slice& s = slices_[cur_];
left_ = s.size();
ptr_ = reinterpret_cast<const char*>(s.begin());
}
*data = ptr_;
*size = left_;
byte_count_ += left_;
ptr_ += left_;
left_ = 0;
return true;
}
void GrpcByteBufferSource::BackUp(int count) {
ptr_ -= count;
left_ += count;
byte_count_ -= count;
}
bool GrpcByteBufferSource::Skip(int count) {
const void* data;
int size;
while (Next(&data, &size)) {
if (size >= count) {
BackUp(size - count);
return true;
}
// size < count;
count -= size;
}
// error or we have too large count;
return false;
}
google::protobuf::int64 GrpcByteBufferSource::ByteCount() const {
return byte_count_;
}
} // namespace detail
} // namespace operators
} // namespace paddle
\ No newline at end of file
paddle/fluid/operators/detail/bytebuffer_stream.h 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2016 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. */
// NOTE: This file was originally created by tensorflow
// (https://github.com/tensorflow/tensorflow/) we borrow this
// file and did some modifications so that we can send gRPC
// requests without too much copying of the tensor data.
#pragma once
#include <grpc++/grpc++.h>
#include "google/protobuf/io/coded_stream.h"
#include "google/protobuf/io/zero_copy_stream.h"
namespace paddle {
namespace operators {
namespace detail {
// A ZeroCopyInputStream that reads from a grpc::ByteBuffer.
class GrpcByteBufferSource
: public ::google::protobuf::io::ZeroCopyInputStream {
public:
GrpcByteBufferSource();
bool Init(const ::grpc::ByteBuffer& src); // Can be called multiple times.
bool Next(const void** data, int* size) override;
void BackUp(int count) override;
bool Skip(int count) override;
::google::protobuf::int64 ByteCount() const override;
private:
std::vector<::grpc::Slice> slices_;
size_t cur_; // Current slice index.
int left_; // Number of bytes in slices_[cur_] left to yield.
const char* ptr_; // Address of next byte in slices_[cur_] to yield.
::google::protobuf::int64 byte_count_;
};
} // namespace detail
} // namespace operators
} // namespace paddle
paddle/fluid/operators/detail/proto_encoder_helper.h 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2016 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. */
// NOTE: This file was originally created by tensorflow
// (https://github.com/tensorflow/tensorflow/) we borrow this
// file and did some modifications so that we can send gRPC
// requests without too much copying of the tensor data.
#pragma once
#include <grpc++/grpc++.h>
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace operators {
namespace detail {
char* EncodeVarint32(char* dst, uint32_t v) {
// Operate on characters as unsigneds
unsigned char* ptr = reinterpret_cast<unsigned char*>(dst);
static const int B = 128;
if (v < (1 << 7)) {
*(ptr++) = v;
} else if (v < (1 << 14)) {
*(ptr++) = v | B;
*(ptr++) = v >> 7;
} else if (v < (1 << 21)) {
*(ptr++) = v | B;
*(ptr++) = (v >> 7) | B;
*(ptr++) = v >> 14;
} else if (v < (1 << 28)) {
*(ptr++) = v | B;
*(ptr++) = (v >> 7) | B;
*(ptr++) = (v >> 14) | B;
*(ptr++) = v >> 21;
} else {
*(ptr++) = v | B;
*(ptr++) = (v >> 7) | B;
*(ptr++) = (v >> 14) | B;
*(ptr++) = (v >> 21) | B;
*(ptr++) = v >> 28;
}
return reinterpret_cast<char*>(ptr);
}
char* EncodeVarint64(char* dst, uint64_t v) {
static const int B = 128;
unsigned char* ptr = reinterpret_cast<unsigned char*>(dst);
while (v >= B) {
*(ptr++) = (v & (B - 1)) | B;
v >>= 7;
}
*(ptr++) = static_cast<unsigned char>(v);
return reinterpret_cast<char*>(ptr);
}
int VarintLength(uint64_t v) {
int len = 1;
while (v >= 128) {
v >>= 7;
len++;
}
return len;
}
class ProtoEncodeHelper {
public:
ProtoEncodeHelper(char* buf, int max_size)
: base_(buf), p_(buf), limit_(base_ + max_size) {}
~ProtoEncodeHelper() {
// Make sure callers didn't do operations that went over max_size promised
PADDLE_ENFORCE_LE(p_, limit_);
}
const char* data() const { return base_; }
size_t size() const { return p_ - base_; }
void WriteUint64(int tag, uint64_t v) {
Encode32(combine(tag, WIRETYPE_VARINT));
Encode64(v);
}
void WriteBool(int tag, bool v) {
Encode32(combine(tag, WIRETYPE_VARINT));
EncodeBool(v);
}
void WriteString(int tag, const std::string& v) {
Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED));
Encode32(v.size());
EncodeBytes(v.data(), v.size());
}
void WriteVarlengthBeginning(int tag, uint32_t len) {
Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED));
Encode32(len);
}
void WriteRawBytes(const std::string& v) { EncodeBytes(v.data(), v.size()); }
private:
// Note: this module's behavior must match the protocol buffer wire encoding
// format.
enum {
WIRETYPE_VARINT = 0,
WIRETYPE_LENGTH_DELIMITED = 2,
};
static uint32_t combine(uint32_t tag, uint32_t type) {
return ((tag << 3) | type);
}
inline void Encode32(uint32_t v) {
if (v < 128) {
// Fast path for single-byte values. Many of the calls will use a
// constant value for v, so the comparison will get optimized away
// when Encode32 is inlined into the caller.
*p_ = v;
p_++;
} else {
p_ = EncodeVarint32(p_, v);
}
}
void Encode64(uint64_t v) { p_ = EncodeVarint64(p_, v); }
void EncodeBool(bool v) {
*p_ = (v ? 1 : 0); // Equal to varint32 encoding of 0 or 1
p_++;
}
void EncodeBytes(const char* bytes, int N) {
memcpy(p_, bytes, N);
p_ += N;
}
char* base_;
char* p_;
char* limit_; // Just for CHECKs
};
} // detail
} // operators
} // paddle
paddle/fluid/operators/detail/send_recv.proto
浏览文件 @ b3a11fdf
......@@ -33,10 +33,34 @@ enum VarType {
}
message VariableMessage {
enum Type {
// Pod Types
BOOL = 0;
INT16 = 1;
INT32 = 2;
INT64 = 3;
FP16 = 4;
FP32 = 5;
FP64 = 6;
}
message LodData { repeated int64 lod_data = 1; }
string varname = 1;
// TODO(Yancey1989): reference framework::proto::VarDesc::VarType
VarType type = 2;
bytes serialized = 3;
// bool persistable is not needed for sending.
// tensor info:
Type data_type = 3;
repeated int64 dims = 4;
// lod details:
int64 lod_level = 5;
repeated LodData lod = 6;
// tensor data
bytes serialized = 7;
// selected_rows data
bytes rows = 8;
}
message VoidMessage {}
paddle/fluid/operators/detail/sendrecvop_utils.cc
浏览文件 @ b3a11fdf
......@@ -13,6 +13,11 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
#include "google/protobuf/io/coded_stream.h"
#include "google/protobuf/io/zero_copy_stream.h"
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/operators/detail/bytebuffer_stream.h"
#include "paddle/fluid/operators/detail/proto_encoder_helper.h"
namespace paddle {
namespace operators {
......@@ -63,6 +68,242 @@ void DeserializeFromMessage(const sendrecv::VariableMessage& msg,
}
}
void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
const platform::DeviceContext& ctx,
::grpc::ByteBuffer* msg) {
using VarMsg = sendrecv::VariableMessage;
sendrecv::VariableMessage request;
std::string header;
request.AppendToString(&header);
// When using GPU, need to free the copied CPU buffer
// when the ByteBuffer destroies
// TODO(typhoonzero): add unref here, if we have dependent
// parallelism execution, need to know when to free the tensor.
DestroyCallback destroy_callback = [](void* backing) {};
void* buf = malloc(1024);
void* payload;
size_t payload_size;
ProtoEncodeHelper e((char*)buf, 1024);
e.WriteString(VarMsg::kVarnameFieldNumber, name);
if (var->IsType<framework::LoDTensor>()) {
e.WriteUint64(VarMsg::kTypeFieldNumber, 0);
} else if (var->IsType<framework::SelectedRows>()) {
e.WriteUint64(VarMsg::kTypeFieldNumber, 1);
}
switch (framework::ToVarType(var->Type())) {
case framework::proto::VarType_Type_LOD_TENSOR: {
auto tensor = var->Get<framework::LoDTensor>();
e.WriteUint64(VarMsg::kDataTypeFieldNumber,
framework::ToDataType(tensor.type()));
for (auto& dim : framework::vectorize(tensor.dims())) {
e.WriteUint64(VarMsg::kDimsFieldNumber, dim);
}
auto lod = tensor.lod(); // std::vector<Vector<size_t>>
if (lod.size() > 0) {
e.WriteUint64(VarMsg::kLodLevelFieldNumber, lod.size());
for (auto& each : lod) {
e.WriteVarlengthBeginning(VarMsg::kLodFieldNumber,
2 + // tag + varintlength of submessage
1 + // kLodDataFieldNumber
each.size());
// auto copied from GPU
for (auto& d : each) {
e.WriteUint64(VarMsg::LodData::kLodDataFieldNumber, d);
}
}
}
if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
PADDLE_ENFORCE(platform::is_gpu_place(tensor.place()));
platform::CPUPlace cpu;
auto& gpu_dev_ctx =
static_cast<const platform::CUDADeviceContext&>(ctx);
auto copy_size = tensor.memory_size();
payload = memory::Alloc(cpu, copy_size);
memory::Copy(cpu, payload,
boost::get<platform::CUDAPlace>(tensor.place()),
reinterpret_cast<const void*>(tensor.data<void>()),
copy_size, gpu_dev_ctx.stream());
destroy_callback = [](void* backing) {
std::cout << "destroy payload" << std::endl;
platform::CPUPlace cpu;
memory::Free(cpu, backing);
};
#endif
} else {
payload = tensor.data<void>();
}
payload_size = tensor.memory_size();
std::string tmp(reinterpret_cast<char*>(payload), payload_size);
for (int i = 0; i < tmp.size(); ++i) {
printf("%02X ", tmp.data()[i]);
}
printf("\n");
e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
} break;
case framework::proto::VarType_Type_SELECTED_ROWS: {
// TODO(typhoonzero): selectedrows implement should not use unique_ptr
auto* slr = var->GetMutable<framework::SelectedRows>();
e.WriteUint64(VarMsg::kDataTypeFieldNumber,
framework::ToDataType(slr->value().type()));
for (auto& dim : framework::vectorize(slr->value().dims())) {
e.WriteUint64(VarMsg::kDimsFieldNumber, dim);
}
e.WriteUint64(VarMsg::kLodLevelFieldNumber, 0);
auto* tensor = slr->mutable_value();
if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
platform::CPUPlace cpu;
auto& gpu_dev_ctx =
static_cast<const platform::CUDADeviceContext&>(ctx);
auto copy_size = tensor->memory_size();
payload = memory::Alloc(cpu, copy_size);
memory::Copy(cpu, payload,
boost::get<platform::CUDAPlace>(tensor->place()),
reinterpret_cast<const void*>(tensor->data<void>()),
copy_size, gpu_dev_ctx.stream());
ctx.Wait();
float* ttt = reinterpret_cast<float*>(payload);
for (int i = 0; i < copy_size / 4; i++) {
std::cout << "copied to cpu: " << ttt[i] << std::endl;
}
destroy_callback = [](void* backing) {
std::cout << "destroy..." << std::endl;
// platform::CPUPlace cpu;
// memory::Free(cpu, backing);
};
#endif
} else {
payload = slr->mutable_value()->data<void>();
}
payload_size = tensor->memory_size();
e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
} break;
default:
PADDLE_THROW("Serialize does not support type: %s",
typeid(var->Type()).name());
break;
}
// steal reference of tensor data
::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows
int num_slices = 2; // only SelectedRows have rows buffer
slices[0] = ::grpc::Slice(e.size());
memcpy(const_cast<uint8_t*>(slices[0].begin()), e.data(), e.size());
slices[1] = ::grpc::Slice(
grpc_slice_new_with_user_data(payload, payload_size, destroy_callback,
static_cast<char*>(payload)),
::grpc::Slice::STEAL_REF);
if (framework::ToVarType(var->Type()) ==
framework::proto::VarType_Type_SELECTED_ROWS) {
auto* slr = var->GetMutable<framework::SelectedRows>();
ProtoEncodeHelper e2((char*)buf, 128);
// NOTE: rows is of type int64_t
size_t rows_memory_size =
slr->rows().capacity() * framework::SizeOfType(typeid(int64_t));
e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size);
slices[2] = ::grpc::Slice(e2.size());
memcpy(const_cast<uint8_t*>(slices[2].begin()), e2.data(), e2.size());
slices[3] = ::grpc::Slice(
grpc_slice_new_with_user_data(
const_cast<void*>(
reinterpret_cast<const void*>(slr->rows().data())),
rows_memory_size,
[](void* backing) {
// TODO(typhoonzero): add unref here, same as above.
},
const_cast<char*>(
reinterpret_cast<const char*>(slr->rows().data()))),
::grpc::Slice::STEAL_REF);
num_slices = 4;
}
::grpc::ByteBuffer tmp(&slices[0], num_slices);
msg->Swap(&tmp);
}
void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
const platform::DeviceContext& ctx,
framework::Variable* var) {
sendrecv::VariableMessage meta;
GrpcByteBufferSource source;
source.Init(msg);
::google::protobuf::io::CodedInputStream input(&source);
// do zerocopy parsing
PADDLE_ENFORCE(meta.ParseFromCodedStream(&input));
PADDLE_ENFORCE(input.ConsumedEntireMessage());
// dims is needed by both tensor and selectedrows
std::vector<int> vecdims;
for (auto& d : meta.dims()) {
vecdims.push_back(d);
}
framework::DDim dims = framework::make_ddim(vecdims);
if (meta.type() == sendrecv::LOD_TENSOR) {
auto* tensor = var->GetMutable<framework::LoDTensor>();
tensor->Resize(dims);
void* tensor_data = tensor->mutable_data(
ctx.GetPlace(),
paddle::operators::detail::ToTypeIndex(meta.data_type()));
framework::LoD lod;
for (int i = 0; i < meta.lod_level(); ++i) {
framework::Vector<size_t> v;
for (int j = 0; j < meta.lod(i).lod_data_size(); ++j) {
v.push_back(meta.lod(i).lod_data(j));
}
lod.push_back(v);
}
tensor->set_lod(lod);
// How to avoid copying and use the message buffer directly?
// Maybe need to find a way to release all memory except tensor content.
if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
platform::CPUPlace cpu;
auto& gpu_dev_ctx = static_cast<const platform::CUDADeviceContext&>(ctx);
memory::Copy(boost::get<platform::CUDAPlace>(tensor->place()),
tensor_data, cpu,
reinterpret_cast<const void*>(meta.serialized().data()),
meta.serialized().size(), gpu_dev_ctx.stream());
#endif
} else {
memcpy(tensor_data,
reinterpret_cast<const void*>(meta.serialized().data()),
meta.serialized().size());
}
} else if (meta.type() == sendrecv::SELECTED_ROWS) {
auto* slr = var->GetMutable<framework::SelectedRows>();
auto* tensor = slr->mutable_value();
int64_t* rows_data = slr->mutable_rows()->data();
tensor->Resize(dims);
void* tensor_data = tensor->mutable_data(
ctx.GetPlace(),
paddle::operators::detail::ToTypeIndex(meta.data_type()));
if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
platform::CPUPlace cpu;
auto& gpu_dev_ctx = static_cast<const platform::CUDADeviceContext&>(ctx);
memory::Copy(boost::get<platform::CUDAPlace>(tensor->place()),
tensor_data, cpu,
reinterpret_cast<const void*>(meta.serialized().data()),
meta.serialized().size(), gpu_dev_ctx.stream());
#endif
} else {
memcpy(tensor_data,
reinterpret_cast<const void*>(meta.serialized().data()),
meta.serialized().size());
}
// copy rows CPU data, GPU data will be copied lazly
memcpy(rows_data, reinterpret_cast<const void*>(meta.rows().data()),
meta.rows().size());
}
}
} // namespace detail
} // namespace operators
} // namespace paddle
} // namespace paddle
\ No newline at end of file
paddle/fluid/operators/detail/sendrecvop_utils.h
浏览文件 @ b3a11fdf
......@@ -33,6 +33,14 @@ namespace detail {
#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
typedef void (*DestroyCallback)(void*);
inline int64_t GetTimestamp() {
return std::chrono::duration_cast<std::chrono::milliseconds>(
std::chrono::system_clock::now().time_since_epoch())
.count();
}
void SerializeToMessage(const std::string& name, const framework::Variable* var,
const platform::DeviceContext& ctx,
sendrecv::VariableMessage* msg);
......@@ -40,6 +48,32 @@ void SerializeToMessage(const std::string& name, const framework::Variable* var,
void DeserializeFromMessage(const sendrecv::VariableMessage& msg,
const platform::DeviceContext& ctx,
framework::Variable* var);
void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
const platform::DeviceContext& ctx,
::grpc::ByteBuffer* msg);
void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
const platform::DeviceContext& ctx,
framework::Variable* var);
inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) {
switch (type) {
case sendrecv::VariableMessage::FP32:
return typeid(float); // NOLINT
case sendrecv::VariableMessage::FP64:
return typeid(double); // NOLINT
case sendrecv::VariableMessage::INT32:
return typeid(int); // NOLINT
case sendrecv::VariableMessage::INT64:
return typeid(int64_t); // NOLINT
case sendrecv::VariableMessage::BOOL:
return typeid(bool); // NOLINT
default:
PADDLE_THROW("Not support type %d", type);
}
}
} // namespace detail
} // namespace operators
} // namespace paddle
paddle/fluid/operators/detail/test_serde.cc 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2016 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 <unistd.h>
#include <string>
#include <thread>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/string/printf.h"
namespace framework = paddle::framework;
namespace platform = paddle::platform;
namespace operators = paddle::operators;
namespace math = paddle::operators::math;
namespace memory = paddle::memory;
void RunSerdeTestTensor(platform::Place place) {
// serialize var to ByteBuffer
framework::Variable var;
auto* tensor = var.GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim({4, 8, 4, 2}));
framework::LoD lod;
lod.push_back(framework::Vector<size_t>({1, 3, 8}));
tensor->set_lod(lod);
int tensor_numel = 4 * 8 * 4 * 2;
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto& ctx = *pool.Get(place);
float* orig_tensor_data = tensor->mutable_data<float>(place);
math::set_constant(ctx, tensor, 31.9);
::grpc::ByteBuffer msg;
operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg);
EXPECT_GT(msg.Length(), 0);
// deserialize
std::vector<::grpc::Slice> slices;
(void)msg.Dump(&slices);
std::string tmp;
for (const auto& s : slices) {
tmp.append(reinterpret_cast<const char*>(s.begin()), s.size());
}
sendrecv::VariableMessage varmsg;
EXPECT_TRUE(varmsg.ParseFromString(tmp));
EXPECT_EQ(varmsg.varname(), "myvar");
EXPECT_EQ(varmsg.type(), 0);
EXPECT_EQ(varmsg.dims()[0], 4);
EXPECT_EQ(varmsg.dims()[1], 8);
EXPECT_EQ(varmsg.dims()[2], 4);
EXPECT_EQ(varmsg.dims()[3], 2);
EXPECT_EQ(varmsg.lod_level(), 1);
EXPECT_EQ(varmsg.lod(0).lod_data(0), 1);
EXPECT_EQ(varmsg.lod(0).lod_data(1), 3);
EXPECT_EQ(varmsg.lod(0).lod_data(2), 8);
const float* tensor_data =
reinterpret_cast<const float*>(varmsg.serialized().data());
for (int i = 0; i < varmsg.serialized().size(); ++i) {
printf("%02X ", varmsg.serialized().data()[i]);
}
printf("\n");
for (int i = 0; i < tensor_numel; ++i) {
std::cout << "#####tensor data: " << tensor_data[i] << std::endl;
EXPECT_EQ(tensor_data[i], orig_tensor_data[i]);
std::cout << "test end 1 " << std::endl;
}
std::cout << "tensor data end " << std::endl;
// deserialize zero-copy
framework::Variable var2;
operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2);
auto tensor2 = var2.Get<framework::LoDTensor>();
float* tensor_data2 = nullptr;
framework::Tensor tmp_tensor;
if (platform::is_gpu_place(ctx.GetPlace())) {
platform::CPUPlace cpu;
framework::TensorCopy(tensor2, cpu, &tmp_tensor);
tensor_data2 = tmp_tensor.data<float>();
} else {
tensor_data2 = const_cast<float*>(tensor2.data<float>());
}
EXPECT_EQ(varmsg.lod_level(), 1);
EXPECT_EQ(varmsg.lod(0).lod_data(0), 1);
EXPECT_EQ(varmsg.lod(0).lod_data(1), 3);
EXPECT_EQ(varmsg.lod(0).lod_data(2), 8);
for (int i = 0; i < tensor_numel; ++i)
EXPECT_EQ(tensor_data2[i], orig_tensor_data[i]);
}
void RunSerdeTestSelectedRows(platform::Place place) {
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto& ctx = *pool.Get(place);
// serialize var to ByteBuffer
framework::Variable var;
auto* slr = var.GetMutable<framework::SelectedRows>();
auto* tensor = slr->mutable_value();
auto* rows = slr->mutable_rows();
tensor->Resize(framework::make_ddim({2, 10}));
int tensor_numel = 2 * 10;
float* orig_tensor_data = tensor->mutable_data<float>(place);
math::set_constant(ctx, tensor, 32.7);
rows->push_back(3);
rows->push_back(10);
::grpc::ByteBuffer msg;
operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg);
EXPECT_GT(msg.Length(), 0);
// deserialize
std::vector<::grpc::Slice> slices;
(void)msg.Dump(&slices);
std::string tmp;
for (const auto& s : slices) {
tmp.append(reinterpret_cast<const char*>(s.begin()), s.size());
}
sendrecv::VariableMessage varmsg;
EXPECT_TRUE(varmsg.ParseFromString(tmp));
EXPECT_EQ(varmsg.varname(), "myvar");
EXPECT_EQ(varmsg.type(), 1);
const float* tensor_data =
reinterpret_cast<const float*>(varmsg.serialized().data());
const int64_t* rows_data =
reinterpret_cast<const int64_t*>(varmsg.rows().data());
for (int i = 0; i < tensor_numel; ++i) {
EXPECT_EQ(tensor_data[i], orig_tensor_data[i]);
}
EXPECT_EQ(rows_data[0], 3);
EXPECT_EQ(rows_data[1], 10);
// deserialize zero-copy
framework::Variable var2;
operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2);
auto* slr2 = var2.GetMutable<framework::SelectedRows>();
auto* tensor2 = slr2->mutable_value();
auto* rows2 = slr2->mutable_rows();
float* tensor_data2 = nullptr;
framework::Tensor tmp_tensor;
if (platform::is_gpu_place(ctx.GetPlace())) {
platform::CPUPlace cpu;
framework::TensorCopy(*tensor2, cpu, &tmp_tensor);
tensor_data2 = tmp_tensor.data<float>();
} else {
tensor_data2 = const_cast<float*>(tensor2->data<float>());
}
const int64_t* rows_data2 = rows2->data();
for (int i = 0; i < tensor_numel; ++i) {
EXPECT_EQ(tensor_data2[i], orig_tensor_data[i]);
}
EXPECT_EQ(rows_data2[0], 3);
EXPECT_EQ(rows_data2[1], 10);
}
// TEST(SelectedRows, CPU) {
// platform::CPUPlace place;
// RunSerdeTestSelectedRows(place);
// }
// TEST(SelectedRows, GPU) {
// platform::CUDAPlace place;
// RunSerdeTestSelectedRows(place);
// }
TEST(Tensor, CPU) {
platform::CPUPlace place;
RunSerdeTestTensor(place);
}
TEST(Tensor, GPU) {
platform::CUDAPlace place;
RunSerdeTestTensor(place);
}
\ No newline at end of file
paddle/fluid/operators/detection_map_op.cc
浏览文件 @ b3a11fdf
......@@ -71,7 +71,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
return framework::OpKernelType(
framework::ToDataType(
ctx.Input<framework::Tensor>("DetectRes")->type()),
ctx.device_context());
platform::CPUPlace());
}
};
......
paddle/fluid/operators/detection_map_op.h
浏览文件 @ b3a11fdf
......@@ -144,6 +144,15 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
}
}
inline void ClipBBox(const Box& bbox, Box* clipped_bbox) const {
T one = static_cast<T>(1.0);
T zero = static_cast<T>(0.0);
clipped_bbox->xmin = std::max(std::min(bbox.xmin, one), zero);
clipped_bbox->ymin = std::max(std::min(bbox.ymin, one), zero);
clipped_bbox->xmax = std::max(std::min(bbox.xmax, one), zero);
clipped_bbox->ymax = std::max(std::min(bbox.ymax, one), zero);
}
void GetBoxes(const framework::LoDTensor& input_label,
const framework::LoDTensor& input_detect,
std::vector<std::map<int, std::vector<Box>>>& gt_boxes,
......@@ -360,7 +369,9 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
size_t max_idx = 0;
auto score = pred_boxes[i].first;
for (size_t j = 0; j < matched_bboxes.size(); ++j) {
T overlap = JaccardOverlap(pred_boxes[i].second, matched_bboxes[j]);
Box& pred_box = pred_boxes[i].second;
ClipBBox(pred_box, &pred_box);
T overlap = JaccardOverlap(pred_box, matched_bboxes[j]);
if (overlap > max_overlap) {
max_overlap = overlap;
max_idx = j;
......
paddle/fluid/operators/elementwise_div_op.h
浏览文件 @ b3a11fdf
......@@ -41,77 +41,14 @@ class ElementwiseDivKernel : public framework::OpKernel<T> {
};
template <typename T>
struct ElementwiseDivGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto z_e = framework::EigenVector<T>::Flatten(*z);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e / y_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = -1.0 * dz_e * z_e / y_e;
}
}
};
template <typename T>
struct ElementwiseDivBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e / y_e_bcast;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
struct DivGradDX {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout / y; }
};
template <typename T>
struct ElementwiseDivBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e / y_e_bcast;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
struct DivGradDY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return -dout * x / (y * y);
}
};
......@@ -128,10 +65,8 @@ class ElementwiseDivGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElementwiseGradCompute<DeviceContext, T, ElementwiseDivGradFunctor<T>,
ElementwiseDivBroadCastGradFunctor<T>,
ElementwiseDivBroadCast2GradFunctor<T>>(
ctx, x, y, out, dout, axis, dx, dy);
ElemwiseGradCompute<DeviceContext, T, DivGradDX<T>, DivGradDY<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, DivGradDX<T>(), DivGradDY<T>());
}
};
......
paddle/fluid/operators/elementwise_max_op.h
浏览文件 @ b3a11fdf
......@@ -41,76 +41,16 @@ class ElementwiseMaxKernel : public framework::OpKernel<T> {
};
template <typename T>
struct ElementwiseMaxGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e > y_e).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (x_e <= y_e).template cast<T>() * dz_e;
}
struct MaxGradDx {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return dout * (x > y);
}
};
template <typename T>
struct ElementwiseMaxBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e > y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e <= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
};
template <typename T>
struct ElementwiseMaxBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e > y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e <= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
struct MaxGradDy {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return dout * (x <= y);
}
};
......@@ -127,12 +67,9 @@ class ElementwiseMaxGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElementwiseGradCompute<DeviceContext, T, ElementwiseMaxGradFunctor<T>,
ElementwiseMaxBroadCastGradFunctor<T>,
ElementwiseMaxBroadCast2GradFunctor<T>>(
ctx, x, y, out, dout, axis, dx, dy);
ElemwiseGradCompute<DeviceContext, T, MaxGradDx<T>, MaxGradDy<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, MaxGradDx<T>(), MaxGradDy<T>());
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/elementwise_min_op.h
浏览文件 @ b3a11fdf
......@@ -41,76 +41,16 @@ class ElementwiseMinKernel : public framework::OpKernel<T> {
};
template <typename T>
struct ElementwiseMinGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e < y_e).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (x_e >= y_e).template cast<T>() * dz_e;
}
struct MinGradDx {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return dout * (x < y);
}
};
template <typename T>
struct ElementwiseMinBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e < y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e >= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
};
template <typename T>
struct ElementwiseMinBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e < y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e >= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
struct MinGradDy {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return dout * (x >= y);
}
};
......@@ -127,12 +67,9 @@ class ElementwiseMinGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElementwiseGradCompute<DeviceContext, T, ElementwiseMinGradFunctor<T>,
ElementwiseMinBroadCastGradFunctor<T>,
ElementwiseMinBroadCast2GradFunctor<T>>(
ctx, x, y, out, dout, axis, dx, dy);
ElemwiseGradCompute<DeviceContext, T, MinGradDx<T>, MinGradDy<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, MinGradDx<T>(), MinGradDy<T>());
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/elementwise_mul_op.h
浏览文件 @ b3a11fdf
......@@ -40,14 +40,15 @@ class ElementwiseMulKernel : public framework::OpKernel<T> {
};
template <typename T>
struct IdentityGrad_DX {
struct MulGradDX {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout * y; }
};
template <typename T>
struct IdentityGrad_DY {
struct MulGradDY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout * x; }
};
template <typename DeviceContext, typename T>
class ElementwiseMulGradKernel : public framework::OpKernel<T> {
public:
......@@ -61,10 +62,8 @@ class ElementwiseMulGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElemwiseGradCompute<DeviceContext, T, IdentityGrad_DX<T>,
IdentityGrad_DY<T>>(ctx, *x, *y, *out, *dout, axis, dx,
dy, IdentityGrad_DX<T>(),
IdentityGrad_DY<T>());
ElemwiseGradCompute<DeviceContext, T, MulGradDX<T>, MulGradDY<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, MulGradDX<T>(), MulGradDY<T>());
}
};
} // namespace operators
......
paddle/fluid/operators/elementwise_sub_op.h
浏览文件 @ b3a11fdf
......@@ -40,61 +40,13 @@ class ElementwiseSubKernel : public framework::OpKernel<T> {
};
template <typename T>
struct ElementwiseSubGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0) * dz_e;
}
}
struct SubGradDX {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout; }
};
template <typename T>
struct ElementwiseSubBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0) *
dz_e.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
};
template <typename T>
struct ElementwiseSubBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0) *
dz_e.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
}
struct SubGradDY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return -dout; }
};
template <typename DeviceContext, typename T>
......@@ -110,12 +62,9 @@ class ElementwiseSubGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElementwiseGradCompute<DeviceContext, T, ElementwiseSubGradFunctor<T>,
ElementwiseSubBroadCastGradFunctor<T>,
ElementwiseSubBroadCast2GradFunctor<T>>(
ctx, x, y, out, dout, axis, dx, dy);
ElemwiseGradCompute<DeviceContext, T, SubGradDX<T>, SubGradDY<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, SubGradDX<T>(), SubGradDY<T>());
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/math/CMakeLists.txt
浏览文件 @ b3a11fdf
add_subdirectory(detail)
if(WITH_GPU)
nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context framework_proto)
nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function tensor)
nv_library(selected_rows_functor SRCS selected_rows_functor.cc selected_rows_functor.cu DEPS selected_rows math_function)
nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor)
nv_library(softmax SRCS softmax.cc softmax.cu DEPS device_context)
nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS device_context)
nv_library(pooling SRCS pooling.cc pooling.cu DEPS device_context)
nv_library(depthwise_conv SRCS depthwise_conv.cu DEPS device_context)
nv_library(sequence_pooling SRCS sequence_pooling.cc sequence_pooling.cu DEPS device_context math_function)
nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context tensor)
nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function)
nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context tensor math_function)
nv_library(sequence_padding SRCS sequence_padding.cc sequence_padding.cu DEPS lod_tensor device_context)
nv_library(sequence_scale SRCS sequence_scale.cc sequence_scale.cu DEPS lod_tensor device_context)
nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
nv_library(maxouting SRCS maxouting.cc maxouting.cu DEPS device_context)
nv_library(unpooling SRCS unpooling.cc unpooling.cu DEPS device_context)
nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
nv_library(cos_sim_functor SRCS cos_sim_functor.cc cos_sim_functor.cu DEPS device_context)
else()
cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto)
cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function)
cc_library(softmax SRCS softmax.cc DEPS device_context)
cc_library(cross_entropy SRCS cross_entropy.cc DEPS device_context)
cc_library(pooling SRCS pooling.cc DEPS device_context)
cc_library(sequence_pooling SRCS sequence_pooling.cc DEPS device_context math_function)
cc_library(vol2col SRCS vol2col.cc DEPS device_context tensor)
cc_library(context_project SRCS context_project.cc DEPS device_context math_function)
cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context tensor math_function)
cc_library(sequence_padding SRCS sequence_padding.cc DEPS lod_tensor device_context)
cc_library(sequence_scale SRCS sequence_scale.cc DEPS lod_tensor device_context)
cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions)
cc_library(maxouting SRCS maxouting.cc DEPS device_context)
cc_library(unpooling SRCS unpooling.cc DEPS device_context)
cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function)
cc_library(cos_sim_functor SRCS cos_sim_functor.cc DEPS device_context)
endif()
function(math_library TARGET)
# math_library is a function to create math library.
# The interface is the same as cc_library.
# But it handle split GPU/CPU code and link some common library.
set(cc_srcs)
set(cu_srcs)
set(math_common_deps device_context framework_proto)
set(multiValueArgs DEPS)
cmake_parse_arguments(math_library "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
list(APPEND cc_srcs ${TARGET}.cc)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${TARGET}.cu)
endif()
list(LENGTH cc_srcs cc_srcs_len)
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${math_library_DEPS} ${math_common_deps})
elseif(${cc_srcs_len} GREATER 0)
cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${math_library_DEPS} ${math_common_deps})
endif()
endfunction()
cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)
# please add new math_library in alphabetical order
math_library(concat)
math_library(context_project DEPS im2col math_function)
math_library(cross_entropy)
math_library(cos_sim_functor)
math_library(depthwise_conv)
math_library(gru_compute DEPS activation_functions math_function)
math_library(im2col)
math_library(lstm_compute DEPS activation_functions)
math_library(math_function DEPS cblas)
math_library(maxouting)
math_library(pooling)
math_library(selected_rows_functor DEPS selected_rows)
math_library(sequence2batch)
math_library(sequence_padding)
math_library(sequence_pooling DEPS math_function)
math_library(sequence_scale)
math_library(softmax)
math_library(unpooling)
math_library(vol2col)
cc_test(math_function_test SRCS math_function_test.cc)
cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selected_rows_functor)
cc_test(im2col_test SRCS im2col_test.cc DEPS math_function tensor)
cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col tensor)
cc_test(im2col_test SRCS im2col_test.cc DEPS im2col)
cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col)
cc_test(sequence_padding_test SRCS sequence_padding_test.cc DEPS sequence_padding)
if(WITH_GPU)
nv_test(math_function_gpu_test SRCS math_function_test.cu)
nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor)
endif()
cc_test(concat_test SRCS concat_test.cc DEPS concat)
paddle/fluid/operators/math/concat.cc 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2018 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 "paddle/fluid/operators/math/concat.h"
namespace paddle {
namespace operators {
namespace math {
/*
* All tensors' dimension should be the same and the values of
* each dimension are the same, except the axis dimension.
*/
template <typename T>
class ConcatFunctor<platform::CPUDeviceContext, T> {
public:
void operator()(const platform::CPUDeviceContext& context,
const std::vector<framework::Tensor>& input, const int axis,
framework::Tensor* output) {
// TODO(zcd): Add input data validity checking
int num = input.size();
int rows = 1;
auto dim_0 = input[0].dims();
for (int i = 0; i < axis; ++i) {
rows *= dim_0[i];
}
int out_rows = rows, out_cols = 0;
std::vector<int64_t> input_cols(input.size());
for (int i = 0; i < num; ++i) {
int t_cols = input[i].numel() / rows;
out_cols += t_cols;
input_cols[i] = t_cols;
}
auto& cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
// computation
for (int k = 0; k < out_rows; ++k) {
T* dst_ptr = output->data<T>() + k * out_cols;
int col_idx = 0;
for (int j = 0; j < num; ++j) {
int col_len = input_cols[j];
const T* src_prt = input[j].data<T>() + k * col_len;
memory::Copy(cpu_place, dst_ptr + col_idx, cpu_place, src_prt,
sizeof(T) * col_len);
col_idx += col_len;
}
}
}
};
/*
* All tensors' dimension should be the same and the values of
* each dimension are the same, except the axis dimension.
*/
template <typename T>
class ConcatGradFunctor<platform::CPUDeviceContext, T> {
public:
void operator()(const platform::CPUDeviceContext& context,
const framework::Tensor& input, const int axis,
std::vector<framework::Tensor>& outputs) {
// TODO(zcd): Add input data validity checking
int num = outputs.size();
int input_rows = 1;
auto dim_0 = outputs[0].dims();
for (int i = 0; i < axis; ++i) {
input_rows *= dim_0[i];
}
int input_cols = 0;
std::vector<int64_t> output_cols(outputs.size());
for (int i = 0; i < num; ++i) {
int t_cols = outputs[i].numel() / input_rows;
input_cols += t_cols;
output_cols[i] = t_cols;
}
auto& cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
// computation
for (int k = 0; k < input_rows; ++k) {
const T* src_ptr = input.data<T>() + k * input_cols;
int col_idx = 0;
for (int j = 0; j < num; ++j) {
int col_len = output_cols[j];
T* dst_ptr = outputs[j].data<T>() + k * col_len;
memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx,
sizeof(T) * col_len);
col_idx += col_len;
}
}
}
};
template class ConcatFunctor<platform::CPUDeviceContext, int>;
template class ConcatFunctor<platform::CPUDeviceContext, int64_t>;
template class ConcatFunctor<platform::CPUDeviceContext, float>;
template class ConcatFunctor<platform::CPUDeviceContext, double>;
template class ConcatGradFunctor<platform::CPUDeviceContext, int>;
template class ConcatGradFunctor<platform::CPUDeviceContext, int64_t>;
template class ConcatGradFunctor<platform::CPUDeviceContext, float>;
template class ConcatGradFunctor<platform::CPUDeviceContext, double>;
} // namespace math
} // namespace operators
} // namespace paddle
paddle/fluid/operators/math/concat.cu 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2018 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 "paddle/fluid/framework/mixed_vector.h"
#include "paddle/fluid/operators/math/concat.h"
#include "paddle/fluid/platform/cuda_helper.h"
namespace paddle {
namespace operators {
namespace math {
template <typename T>
__device__ T upper_bound(const T* first, T count, T val) {
const T* orig = first;
const T* it = nullptr;
T step = 0;
while (count > 0) {
it = first;
step = count / 2;
it += step;
if (!(val < *it)) {
first = ++it;
count -= step + 1;
} else {
count = step;
}
}
return first - orig;
}
template <typename T>
__global__ void KernelConcat(T** inputs, const int* input_cols, int col_size,
const int output_rows, const int output_cols,
T* output) {
int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
int segment = upper_bound<int>(input_cols, col_size, tid_x) - 1;
int curr_offset = input_cols[segment];
int curr_segment = segment;
for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) {
T curr_col_offset;
while ((curr_col_offset = input_cols[curr_segment + 1]) <= tid_x) {
curr_offset = curr_col_offset;
++curr_segment;
}
int local_col = tid_x - curr_offset;
int segment_width = curr_col_offset - curr_offset;
T* input_ptr = inputs[curr_segment];
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y)
output[tid_y * output_cols + tid_x] =
input_ptr[tid_y * segment_width + local_col];
}
}
template <typename T>
__global__ void KernelConcat(T** inputs, const int input_col,
const int output_rows, const int output_cols,
T* output) {
int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
double inv_input_col = 1.0 / input_col;
for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) {
int split = tid_x * inv_input_col;
int in_offset = tid_x - split * input_col;
T* input_ptr = inputs[split];
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y) {
output[tid_y * output_cols + tid_x] =
input_ptr[tid_y * input_col + in_offset];
}
}
}
template <typename T>
__global__ void KernelConcatGrad(const T* input, const int input_row,
const int input_col, const int* output_cols,
int col_size, T** outputs) {
int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
int segment = upper_bound<int>(output_cols, col_size, tid_x) - 1;
int curr_offset = output_cols[segment];
int curr_segment = segment;
for (; tid_x < input_col; tid_x += blockDim.x * gridDim.x) {
T curr_col_offset;
while ((curr_col_offset = output_cols[curr_segment + 1]) <= tid_x) {
curr_offset = curr_col_offset;
++curr_segment;
}
int local_col = tid_x - curr_offset;
int segment_width = curr_col_offset - curr_offset;
T* output_ptr = outputs[curr_segment];
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < input_row; tid_y += blockDim.y * gridDim.y)
output_ptr[tid_y * segment_width + local_col] =
input[tid_y * input_col + tid_x];
}
}
template <typename T>
__global__ void KernelConcatGrad(const T* input, const int input_row,
const int input_col, const int output_cols,
T** outputs) {
int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
double inv_input_col = 1.0 / input_col;
for (; tid_x < input_col; tid_x += blockDim.x * gridDim.x) {
int split = tid_x * inv_input_col;
int in_offset = tid_x - split * input_col;
T* output_ptr = outputs[split];
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < input_row; tid_y += blockDim.y * gridDim.y)
output_ptr[tid_y * output_cols + in_offset] =
input[tid_y * input_col + tid_x];
}
}
/*
* All tensors' dimension should be the same and the values of
* each dimension are the same, except the axis dimension.
*/
template <typename T>
class ConcatFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const std::vector<framework::Tensor>& input, const int axis,
framework::Tensor* output) {
// TODO(zcd): Add input data validity checking
int num = input.size();
int rows = 1;
auto dim_0 = input[0].dims();
for (int i = 0; i < axis; ++i) {
rows *= dim_0[i];
}
int cols = input[0].numel() / rows;
int out_rows = rows, out_cols = 0;
framework::Vector<int16_t> inputs_data(num * sizeof(T*) / 2);
framework::Vector<int> inputs_cols(num + 1);
inputs_cols[0] = 0;
T** inputs_ptr = reinterpret_cast<T**>(inputs_data.data());
bool sameShape = true;
for (int i = 0; i < num; ++i) {
int t_cols = input[i].numel() / rows;
if (sameShape) {
if (t_cols != cols) sameShape = false;
}
out_cols += t_cols;
inputs_cols[i + 1] = out_cols;
inputs_ptr[i] = const_cast<T*>(input[i].data<T>());
}
T** ins_gpu =
reinterpret_cast<T**>(inputs_data.CUDAMutableData(context.GetPlace()));
const int* ins_col_gpu = inputs_cols.CUDAData(context.GetPlace());
// computation
// set the thread block and grid according to CurrentDeviceId
const int kThreadsPerBlock = 1024;
int block_cols = kThreadsPerBlock;
if (out_cols < kThreadsPerBlock) { // block_cols is aligned by 32.
block_cols = ((out_cols + 31) >> 5) << 5;
}
int block_rows = kThreadsPerBlock / block_cols;
dim3 block_size = dim3(block_cols, block_rows, 1);
int max_threads = context.GetMaxPhysicalThreadCount();
int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);
int grid_cols =
std::min((out_cols + block_cols - 1) / block_cols, max_blocks);
int grid_rows =
std::min(max_blocks / grid_cols, std::max(out_rows / block_rows, 1));
dim3 grid_size = dim3(grid_cols, grid_rows, 1);
if (sameShape) {
KernelConcat<<<grid_size, block_size, 0, context.stream()>>>(
ins_gpu, cols, out_rows, out_cols, output->data<T>());
} else {
KernelConcat<<<grid_size, block_size, 0, context.stream()>>>(
ins_gpu, ins_col_gpu, static_cast<int>(inputs_cols.size()), out_rows,
out_cols, output->data<T>());
}
}
};
/*
* All tensors' dimension should be the same and the values of
* each dimension are the same, except the axis dimension.
*/
template <typename T>
class ConcatGradFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const framework::Tensor& input, const int axis,
std::vector<framework::Tensor>& outputs) {
// TODO(zcd): Add input data validity checking
int num = outputs.size();
int input_row = 1;
auto dim_0 = outputs[0].dims();
for (int i = 0; i < axis; ++i) {
input_row *= dim_0[i];
}
int output_col_0 = outputs[0].numel() / input_row;
int input_col = 0;
bool sameShape = true;
framework::Vector<int16_t> outputs_data(num * sizeof(T*) / 2);
framework::Vector<int> outputs_cols(num + 1);
outputs_cols[0] = 0;
T** outputs_ptr = reinterpret_cast<T**>(outputs_data.data());
for (int i = 0; i < num; ++i) {
int t_col = outputs[i].numel() / input_row;
if (sameShape) {
if (t_col != output_col_0) sameShape = false;
}
input_col += t_col;
outputs_cols[i + 1] = input_col;
outputs_ptr[i] = outputs[i].data<T>();
}
T** outs_gpu =
reinterpret_cast<T**>(outputs_data.CUDAMutableData(context.GetPlace()));
const int* outs_col_gpu = outputs_cols.CUDAData(context.GetPlace());
// computation
const int kThreadsPerBlock = 1024;
int block_cols = kThreadsPerBlock;
if (input_col < kThreadsPerBlock) { // block_cols is aligned by 32.
block_cols = ((input_col + 31) >> 5) << 5;
}
int block_rows = kThreadsPerBlock / block_cols;
dim3 block_size = dim3(block_cols, block_rows, 1);
int max_threads = context.GetMaxPhysicalThreadCount();
int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);
int grid_cols =
std::min((input_col + block_cols - 1) / block_cols, max_blocks);
int grid_rows =
std::min(max_blocks / grid_cols, std::max(input_row / block_rows, 1));
dim3 grid_size = dim3(grid_cols, grid_rows, 1);
if (sameShape) {
KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
input.data<T>(), input_row, input_col, output_col_0, outs_gpu);
} else {
KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
input.data<T>(), input_row, input_col, outs_col_gpu,
static_cast<int>(outputs_cols.size()), outs_gpu);
}
}
};
template class ConcatFunctor<platform::CUDADeviceContext, int>;
template class ConcatFunctor<platform::CUDADeviceContext, int64_t>;
template class ConcatFunctor<platform::CUDADeviceContext, float>;
template class ConcatFunctor<platform::CUDADeviceContext, double>;
template class ConcatGradFunctor<platform::CUDADeviceContext, int>;
template class ConcatGradFunctor<platform::CUDADeviceContext, int64_t>;
template class ConcatGradFunctor<platform::CUDADeviceContext, float>;
template class ConcatGradFunctor<platform::CUDADeviceContext, double>;
} // namespace math
} // namespace operators
} // namespace paddle
paddle/fluid/operators/math/concat.h 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2018 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
#include "paddle/fluid/framework/tensor.h"
namespace paddle {
namespace operators {
namespace math {
/*
* \brief Concatenate the input tensors along the dimension axis.
* TODO(zcd): maybe it needs to be more detailed.
* Examples:
* Input[0] = [[1,2],[3,4]]
* Input[1] = [[5,6]]
* axis = 0
*
* Output = [[1,2],
* [3,4],
* [5,6]]
*/
template <typename DeviceContext, typename T>
class ConcatFunctor {
public:
void operator()(const DeviceContext& context,
const std::vector<framework::Tensor>& input, const int axis,
framework::Tensor* output);
};
/*
* \brief Split the input tensors along the dimension axis into outputs.
* TODO(zcd): maybe it needs to be more detailed.
* Examples:
* Input = [[1,2],
* [3,4],
* [5,6]]
* axis = 0
*
* Output[0] = [[1,2],[3,4]]
* Output[1] = [[5,6]]
*/
template <typename DeviceContext, typename T>
class ConcatGradFunctor {
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
const int axis, std::vector<framework::Tensor>& outputs);
};
} // namespace math
} // namespace operators
} // namespace paddle
paddle/fluid/operators/math/concat_test.cc 0 → 100644
浏览文件 @ b3a11fdf
/* Copyright (c) 2018 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 "paddle/fluid/operators/math/concat.h"
#include <gtest/gtest.h>
#include <vector>
#include "paddle/fluid/framework/tensor_util.h"
using namespace paddle::framework;
using namespace paddle::platform;
template <typename DeviceContext, typename Place>
void testConcat() {
Tensor input_a_cpu;
Tensor input_b_cpu;
Tensor out_cpu;
Tensor input_a;
Tensor input_b;
Tensor out;
DeviceContext* context = new DeviceContext(Place());
// DeviceContext context(Place());
/**
* cast1:
* inputs:
* t_a.shape: [2, 3, 4]
* t_b.shape: [3, 3, 4]
* output:
* out.shape: [5, 3, 4]
*/
auto dim_a = make_ddim({2, 3, 4});
auto dim_b = make_ddim({3, 3, 4});
auto dim_out = make_ddim({5, 3, 4});
input_a.mutable_data<int>(dim_a, Place());
input_b.mutable_data<int>(dim_b, Place());
out.mutable_data<int>(dim_out, Place());
if (paddle::platform::is_gpu_place(Place())) {
input_a_cpu.mutable_data<int>(dim_a, CPUPlace());
input_b_cpu.mutable_data<int>(dim_b, CPUPlace());
out_cpu.mutable_data<int>(dim_out, CPUPlace());
}
int* a_ptr;
int* b_ptr;
if (paddle::platform::is_gpu_place(Place())) {
a_ptr = input_a_cpu.data<int>();
b_ptr = input_b_cpu.data<int>();
} else {
a_ptr = input_a.data<int>();
b_ptr = input_b.data<int>();
}
for (int i = 0; i < 2 * 3 * 4; ++i) {
a_ptr[i] = i;
}
for (int i = 0; i < 3 * 3 * 4; ++i) {
b_ptr[i] = i;
}
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(input_a_cpu, Place(), *context, &input_a);
TensorCopy(input_b_cpu, Place(), *context, &input_b);
}
std::vector<Tensor> input;
input.push_back(input_a);
input.push_back(input_b);
paddle::operators::math::ConcatFunctor<DeviceContext, int> concat_functor;
concat_functor(*context, input, 0, &out);
// check the dim of input_a, input_b
PADDLE_ENFORCE_EQ(input_a.dims(), dim_a);
PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
int* out_ptr;
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(out, CPUPlace(), *context, &out_cpu);
out_ptr = out_cpu.data<int>();
} else {
out_ptr = out.data<int>();
}
int cols = 2 * 3 * 4;
int idx_a = 0, idx_b = 0;
for (int j = 0; j < 5 * 3 * 4; ++j) {
if (j >= cols) {
PADDLE_ENFORCE_EQ(out_ptr[j], b_ptr[idx_b]);
++idx_b;
} else {
PADDLE_ENFORCE_EQ(out_ptr[j], a_ptr[idx_a]);
++idx_a;
}
}
//
/**
* cast2:
* inputs:
* t_a.shape: [2, 3, 4]
* t_b.shape: [2, 4, 4]
* output:
* out.shape: [2, 7, 4]
*/
dim_a = make_ddim({2, 3, 4});
dim_b = make_ddim({2, 4, 4});
dim_out = make_ddim({2, 7, 4});
input_a.Resize(dim_a);
input_b.Resize(dim_b);
out.Resize(dim_out);
if (paddle::platform::is_gpu_place(Place())) {
input_a_cpu.Resize(dim_a);
input_b_cpu.Resize(dim_b);
out_cpu.Resize(dim_out);
}
if (paddle::platform::is_gpu_place(Place())) {
a_ptr = input_a_cpu.data<int>();
b_ptr = input_b_cpu.data<int>();
} else {
a_ptr = input_a.data<int>();
b_ptr = input_b.data<int>();
}
for (int i = 0; i < 2 * 3 * 4; ++i) {
a_ptr[i] = i;
}
for (int i = 0; i < 2 * 4 * 4; ++i) {
b_ptr[i] = i;
}
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(input_a_cpu, Place(), *context, &input_a);
TensorCopy(input_b_cpu, Place(), *context, &input_b);
}
input.clear();
input.push_back(input_a);
input.push_back(input_b);
concat_functor(*context, input, 1, &out);
// check the dim of input_a, input_b
PADDLE_ENFORCE_EQ(input_a.dims(), dim_a);
PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(out, CPUPlace(), *context, &out_cpu);
out_ptr = out_cpu.data<int>();
} else {
out_ptr = out.data<int>();
}
cols = 3 * 4;
idx_a = 0, idx_b = 0;
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 28; ++j) {
if (j >= cols) {
PADDLE_ENFORCE_EQ(out_ptr[i * 28 + j], b_ptr[idx_b]);
++idx_b;
} else {
PADDLE_ENFORCE_EQ(out_ptr[i * 28 + j], a_ptr[idx_a]);
++idx_a;
}
}
}
/**
* cast3:
* inputs:
* t_a.shape: [2, 3, 5]
* t_b.shape: [2, 3, 4]
* output:
* out.shape: [2, 3, 9]
*/
dim_a = make_ddim({2, 3, 4});
dim_b = make_ddim({2, 3, 5});
dim_out = make_ddim({2, 3, 9});
input_a.Resize(dim_a);
input_b.Resize(dim_b);
out.Resize(dim_out);
if (paddle::platform::is_gpu_place(Place())) {
input_a_cpu.Resize(dim_a);
input_b_cpu.Resize(dim_b);
out_cpu.Resize(dim_out);
}
if (paddle::platform::is_gpu_place(Place())) {
a_ptr = input_a_cpu.data<int>();
b_ptr = input_b_cpu.data<int>();
} else {
a_ptr = input_a.data<int>();
b_ptr = input_b.data<int>();
}
for (int i = 0; i < 2 * 3 * 4; ++i) {
a_ptr[i] = i;
}
for (int i = 0; i < 2 * 3 * 5; ++i) {
b_ptr[i] = i;
}
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(input_a_cpu, Place(), *context, &input_a);
TensorCopy(input_b_cpu, Place(), *context, &input_b);
}
input.clear();
input.push_back(input_a);
input.push_back(input_b);
concat_functor(*context, input, 2, &out);
// check the dim of input_a, input_b
PADDLE_ENFORCE_EQ(input_a.dims(), dim_a);
PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(out, CPUPlace(), *context, &out_cpu);
out_ptr = out_cpu.data<int>();
} else {
out_ptr = out.data<int>();
}
// check the data
cols = 4;
idx_a = 0, idx_b = 0;
for (int i = 0; i < 6; ++i) {
for (int j = 0; j < 9; ++j) {
if (j >= cols) {
PADDLE_ENFORCE_EQ(out_ptr[i * 9 + j], b_ptr[idx_b]);
++idx_b;
} else {
PADDLE_ENFORCE_EQ(out_ptr[i * 9 + j], a_ptr[idx_a]);
++idx_a;
}
}
}
/**
* cast4:
* inputs:
* axis = 1
* t_a.shape: [2, 3, 4]
* t_b.shape: [2, 3, 4]
* output:
* out.shape: [2, 6, 4]
*/
dim_a = make_ddim({2, 3, 4});
dim_b = make_ddim({2, 3, 4});
dim_out = make_ddim({2, 6, 4});
input_a.Resize(dim_a);
input_b.Resize(dim_b);
out.Resize(dim_out);
if (paddle::platform::is_gpu_place(Place())) {
input_a_cpu.Resize(dim_a);
input_b_cpu.Resize(dim_b);
out_cpu.Resize(dim_out);
}
if (paddle::platform::is_gpu_place(Place())) {
a_ptr = input_a_cpu.data<int>();
b_ptr = input_b_cpu.data<int>();
} else {
a_ptr = input_a.data<int>();
b_ptr = input_b.data<int>();
}
for (int i = 0; i < 2 * 3 * 4; ++i) {
a_ptr[i] = i;
}
for (int i = 0; i < 2 * 3 * 4; ++i) {
b_ptr[i] = i;
}
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(input_a_cpu, Place(), *context, &input_a);
TensorCopy(input_b_cpu, Place(), *context, &input_b);
}
input.clear();
input.push_back(input_a);
input.push_back(input_b);
concat_functor(*context, input, 1, &out);
// check the dim of input_a, input_b
PADDLE_ENFORCE_EQ(input_a.dims(), dim_a);
PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
if (paddle::platform::is_gpu_place(Place())) {
TensorCopy(out, CPUPlace(), *context, &out_cpu);
out_ptr = out_cpu.data<int>();
} else {
out_ptr = out.data<int>();
}
// check the data
cols = 12;
idx_a = 0, idx_b = 0;
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 24; ++j) {
if (j >= cols) {
PADDLE_ENFORCE_EQ(out_ptr[i * 24 + j], b_ptr[idx_b]);
++idx_b;
} else {
PADDLE_ENFORCE_EQ(out_ptr[i * 24 + j], a_ptr[idx_a]);
++idx_a;
}
}
}
}
TEST(math, concat) {
testConcat<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
#ifdef PADDLE_WITH_CUDA
testConcat<paddle::platform::CUDADeviceContext,
paddle::platform::CUDAPlace>();
#endif
}
paddle/fluid/operators/math/math_function.cc
浏览文件 @ b3a11fdf
......@@ -15,11 +15,23 @@ limitations under the License. */
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/operators/math/math_function_impl.h"
#include "paddle/fluid/platform/float16.h"
namespace paddle {
namespace operators {
namespace math {
using float16 = paddle::platform::float16;
template <>
void gemm<platform::CPUDeviceContext, float16>(
const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
const float16 alpha, const float16* A, const float16* B, const float16 beta,
float16* C) {
PADDLE_THROW("float16 GEMM not supported on CPU");
}
template <>
void gemm<platform::CPUDeviceContext, float>(
const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
......@@ -46,6 +58,15 @@ void gemm<platform::CPUDeviceContext, double>(
beta, C, ldc);
}
template <>
void gemm<platform::CPUDeviceContext, float16>(
const platform::CPUDeviceContext& context, const bool transA,
const bool transB, const int M, const int N, const int K,
const float16 alpha, const float16* A, const int lda, const float16* B,
const int ldb, const float16 beta, float16* C, const int ldc) {
PADDLE_THROW("float16 GEMM not supported on CPU");
}
template <>
void gemm<platform::CPUDeviceContext, float>(
const platform::CPUDeviceContext& context, const bool transA,
......@@ -68,6 +89,15 @@ void gemm<platform::CPUDeviceContext, double>(
lda, B, ldb, beta, C, ldc);
}
template <>
void matmul<platform::CPUDeviceContext, float16>(
const platform::CPUDeviceContext& context,
const framework::Tensor& matrix_a, bool trans_a,
const framework::Tensor& matrix_b, bool trans_b, float16 alpha,
framework::Tensor* matrix_out, float16 beta) {
PADDLE_THROW("float16 matmul not supported on CPU");
}
template <>
void matmul<platform::CPUDeviceContext, float>(
const platform::CPUDeviceContext& context,
......@@ -126,6 +156,15 @@ void matmul<platform::CPUDeviceContext, double>(
matrix_b.data<double>(), beta, matrix_out->data<double>());
}
template <>
void batched_gemm<platform::CPUDeviceContext, float16>(
const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
const float16 alpha, const float16* A, const float16* B, const float16 beta,
float16* C, const int batchCount, const int strideA, const int strideB) {
PADDLE_THROW("float16 batched_gemm not supported on CPU");
}
#ifdef PADDLE_WITH_MKLML
// Use cblas_{s,d}gemm_batched if available: Run with 1 group of size batchSize.
template <>
......
paddle/fluid/operators/math/math_function.cu
浏览文件 @ b3a11fdf
......@@ -16,11 +16,40 @@ limitations under the License. */
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/math_function_impl.h"
#include "paddle/fluid/platform/float16.h"
namespace paddle {
namespace operators {
namespace math {
using float16 = paddle::platform::float16;
template <>
void gemm<platform::CUDADeviceContext, float16>(
const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
const float16 alpha, const float16* A, const float16* B, const float16 beta,
float16* C) {
// Note that cublas follows fortran order, so the order is different from
// the cblas convention.
int lda = (transA == CblasNoTrans) ? K : M;
int ldb = (transB == CblasNoTrans) ? N : K;
cublasOperation_t cuTransA =
(transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
cublasOperation_t cuTransB =
(transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
const half h_alpha = static_cast<const half>(alpha);
const half h_beta = static_cast<const half>(beta);
const half* h_A = reinterpret_cast<const half*>(A);
const half* h_B = reinterpret_cast<const half*>(B);
half* h_C = reinterpret_cast<half*>(C);
PADDLE_ENFORCE(platform::dynload::cublasHgemm(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
h_A, lda, &h_beta, h_C, N));
}
template <>
void gemm<platform::CUDADeviceContext, float>(
const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
......@@ -60,6 +89,28 @@ void gemm<platform::CUDADeviceContext, double>(
lda, &beta, C, N));
}
template <>
void gemm<platform::CUDADeviceContext, float16>(
const platform::CUDADeviceContext& context, const bool transA,
const bool transB, const int M, const int N, const int K,
const float16 alpha, const float16* A, const int lda, const float16* B,
const int ldb, const float16 beta, float16* C, const int ldc) {
// Note that cublas follows fortran order, so the order is different from
// the cblas convention.
cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T;
cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T;
const half h_alpha = static_cast<const half>(alpha);
const half h_beta = static_cast<const half>(beta);
const half* h_A = reinterpret_cast<const half*>(A);
const half* h_B = reinterpret_cast<const half*>(B);
half* h_C = reinterpret_cast<half*>(C);
PADDLE_ENFORCE(platform::dynload::cublasHgemm(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
h_A, lda, &h_beta, h_C, ldc));
}
template <>
void gemm<platform::CUDADeviceContext, float>(
const platform::CUDADeviceContext& context, const bool transA,
......@@ -90,6 +141,35 @@ void gemm<platform::CUDADeviceContext, double>(
lda, &beta, C, ldc));
}
template <>
void matmul<platform::CUDADeviceContext, float16>(
const platform::CUDADeviceContext& context,
const framework::Tensor& matrix_a, bool trans_a,
const framework::Tensor& matrix_b, bool trans_b, float16 alpha,
framework::Tensor* matrix_out, float16 beta) {
auto dim_a = matrix_a.dims();
auto dim_b = matrix_b.dims();
auto dim_out = matrix_out->dims();
PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
"The input and output of matmul be matrix");
PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
platform::is_gpu_place(matrix_b.place()) &&
platform::is_gpu_place(matrix_out->place()),
"Matrix must all be in CUDAPlace");
int M = dim_out[0];
int N = dim_out[1];
int K = (trans_a == false) ? dim_a[1] : dim_a[0];
CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
gemm<platform::CUDADeviceContext, float16>(
context, transA, transB, M, N, K, alpha, matrix_a.data<float16>(),
matrix_b.data<float16>(), beta, matrix_out->data<float16>());
}
template <>
void matmul<platform::CUDADeviceContext, float>(
const platform::CUDADeviceContext& context,
......@@ -148,6 +228,34 @@ void matmul<platform::CUDADeviceContext, double>(
matrix_b.data<double>(), beta, matrix_out->data<double>());
}
template <>
void batched_gemm<platform::CUDADeviceContext, float16>(
const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
const float16 alpha, const float16* A, const float16* B, const float16 beta,
float16* C, const int batchCount, const int strideA, const int strideB) {
// Note that cublas follows fortran order, so the order is different from
// the cblas convention.
int lda = (transA == CblasNoTrans) ? K : M;
int ldb = (transB == CblasNoTrans) ? N : K;
int ldc = N;
cublasOperation_t cuTransA =
(transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
cublasOperation_t cuTransB =
(transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
const int strideC = M * N;
const half h_alpha = static_cast<const half>(alpha);
const half h_beta = static_cast<const half>(beta);
const half* h_A = reinterpret_cast<const half*>(A);
const half* h_B = reinterpret_cast<const half*>(B);
half* h_C = reinterpret_cast<half*>(C);
PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
strideB, h_A, lda, strideA, &h_beta, h_C, ldc, strideC, batchCount));
}
template <>
void batched_gemm<platform::CUDADeviceContext, float>(
const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
......
paddle/fluid/operators/math/math_function_test.cu
浏览文件 @ b3a11fdf
......@@ -14,30 +14,41 @@
#include "gtest/gtest.h"
#include "paddle/fluid/operators/math/math_function.h"
TEST(math_function, notrans_mul_trans) {
paddle::framework::Tensor input1;
paddle::framework::Tensor input1_gpu;
paddle::framework::Tensor input2_gpu;
paddle::framework::Tensor out_gpu;
paddle::framework::Tensor out;
auto* cpu_place = new paddle::platform::CPUPlace();
float* input1_ptr = input1.mutable_data<float>({2, 3}, *cpu_place);
void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size,
const std::vector<float>& data) {
PADDLE_ENFORCE_EQ(size, data.size());
for (size_t i = 0; i < data.size(); ++i) {
in_ptr[i] = paddle::platform::float16(data[i]);
}
}
TEST(math_function, notrans_mul_trans_fp32) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor out_gpu;
Tensor out;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
float* input1_ptr = input1.mutable_data<float>({2, 3}, cpu_place);
float arr[6] = {0, 1, 2, 3, 4, 5};
memcpy(input1_ptr, arr, 6 * sizeof(float));
auto* gpu_place = new paddle::platform::CUDAPlace(0);
paddle::platform::CUDADeviceContext context(*gpu_place);
paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu);
paddle::framework::TensorCopy(input1, *gpu_place, context, &input2_gpu);
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input1, gpu_place, context, &input2_gpu);
out_gpu.mutable_data<float>({2, 2}, *gpu_place);
out_gpu.mutable_data<float>({2, 2}, gpu_place);
paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float>(
paddle::operators::math::matmul<CUDADeviceContext, float>(
context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0);
paddle::framework::TensorCopy(out_gpu, *cpu_place, context, &out);
TensorCopy(out_gpu, cpu_place, context, &out);
float* out_ptr = out.data<float>();
context.Wait();
......@@ -45,33 +56,71 @@ TEST(math_function, notrans_mul_trans) {
EXPECT_EQ(out_ptr[1], 14);
EXPECT_EQ(out_ptr[2], 14);
EXPECT_EQ(out_ptr[3], 50);
delete gpu_place;
}
TEST(math_function, trans_mul_notrans) {
paddle::framework::Tensor input1;
paddle::framework::Tensor input1_gpu;
paddle::framework::Tensor input2_gpu;
paddle::framework::Tensor out_gpu;
paddle::framework::Tensor out;
TEST(math_function, notrans_mul_trans_fp16) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor out_gpu;
Tensor out;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input1, gpu_place, context, &input2_gpu);
out_gpu.mutable_data<float16>({2, 2}, gpu_place);
paddle::operators::math::matmul<CUDADeviceContext, float16>(
context, input1_gpu, false, input2_gpu, true, float16(1), &out_gpu,
float16(0));
TensorCopy(out_gpu, cpu_place, context, &out);
float16* out_ptr = out.data<float16>();
context.Wait();
EXPECT_EQ(static_cast<float>(out_ptr[0]), 5);
EXPECT_EQ(static_cast<float>(out_ptr[1]), 14);
EXPECT_EQ(static_cast<float>(out_ptr[2]), 14);
EXPECT_EQ(static_cast<float>(out_ptr[3]), 50);
}
TEST(math_function, trans_mul_notrans_fp32) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor out_gpu;
Tensor out;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
auto* cpu_place = new paddle::platform::CPUPlace();
float* input1_ptr = input1.mutable_data<float>({2, 3}, *cpu_place);
float* input1_ptr = input1.mutable_data<float>({2, 3}, cpu_place);
float arr[6] = {0, 1, 2, 3, 4, 5};
memcpy(input1_ptr, arr, 6 * sizeof(float));
auto* gpu_place = new paddle::platform::CUDAPlace(0);
paddle::platform::CUDADeviceContext context(*gpu_place);
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input1, gpu_place, context, &input2_gpu);
paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu);
paddle::framework::TensorCopy(input1, *gpu_place, context, &input2_gpu);
out_gpu.mutable_data<float>({3, 3}, *gpu_place);
out_gpu.mutable_data<float>({3, 3}, gpu_place);
paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float>(
context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0);
paddle::framework::TensorCopy(out_gpu, *cpu_place, context, &out);
TensorCopy(out_gpu, cpu_place, context, &out);
float* out_ptr = out.data<float>();
context.Wait();
......@@ -84,45 +133,88 @@ TEST(math_function, trans_mul_notrans) {
EXPECT_EQ(out_ptr[6], 15);
EXPECT_EQ(out_ptr[7], 22);
EXPECT_EQ(out_ptr[8], 29);
delete gpu_place;
}
TEST(math_function, gemm_notrans_cublas) {
paddle::framework::Tensor input1;
paddle::framework::Tensor input2;
paddle::framework::Tensor input3;
paddle::framework::Tensor input1_gpu;
paddle::framework::Tensor input2_gpu;
paddle::framework::Tensor input3_gpu;
TEST(math_function, trans_mul_notrans_fp16) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor out_gpu;
Tensor out;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input1, gpu_place, context, &input2_gpu);
out_gpu.mutable_data<float16>({3, 3}, gpu_place);
paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float16>(
context, input1_gpu, true, input2_gpu, false, float16(1), &out_gpu,
float16(0));
TensorCopy(out_gpu, cpu_place, context, &out);
float16* out_ptr = out.data<float16>();
context.Wait();
EXPECT_EQ(static_cast<float>(out_ptr[0]), 9);
EXPECT_EQ(static_cast<float>(out_ptr[1]), 12);
EXPECT_EQ(static_cast<float>(out_ptr[2]), 15);
EXPECT_EQ(static_cast<float>(out_ptr[3]), 12);
EXPECT_EQ(static_cast<float>(out_ptr[4]), 17);
EXPECT_EQ(static_cast<float>(out_ptr[5]), 22);
EXPECT_EQ(static_cast<float>(out_ptr[6]), 15);
EXPECT_EQ(static_cast<float>(out_ptr[7]), 22);
EXPECT_EQ(static_cast<float>(out_ptr[8]), 29);
}
TEST(math_function, gemm_notrans_cublas_fp32) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input2;
Tensor input3;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor input3_gpu;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
int m = 2;
int n = 3;
int k = 3;
auto* cpu_place = new paddle::platform::CPUPlace();
float* input1_ptr = input1.mutable_data<float>({2, 3}, *cpu_place);
float* input1_ptr = input1.mutable_data<float>({2, 3}, cpu_place);
float arr1[6] = {0, 1, 2, 3, 4, 5};
memcpy(input1_ptr, arr1, 6 * sizeof(float));
float* input2_ptr = input2.mutable_data<float>({3, 4}, *cpu_place);
float* input2_ptr = input2.mutable_data<float>({3, 4}, cpu_place);
float arr2[12] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
memcpy(input2_ptr, arr2, 12 * sizeof(float));
float* input3_ptr = input3.mutable_data<float>({2, 4}, *cpu_place);
float* input3_ptr = input3.mutable_data<float>({2, 4}, cpu_place);
float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
memcpy(input3_ptr, arr3, 8 * sizeof(float));
auto* gpu_place = new paddle::platform::CUDAPlace(0);
paddle::platform::CUDADeviceContext context(*gpu_place);
paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu);
paddle::framework::TensorCopy(input2, *gpu_place, context, &input2_gpu);
paddle::framework::TensorCopy(input3, *gpu_place, context, &input3_gpu);
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input2, gpu_place, context, &input2_gpu);
TensorCopy(input3, gpu_place, context, &input3_gpu);
float* a = input1_gpu.data<float>();
float* b = input2_gpu.data<float>();
float* c = input3_gpu.mutable_data<float>(*gpu_place);
float* c = input3_gpu.mutable_data<float>(gpu_place);
paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float>(
context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4);
paddle::framework::TensorCopy(input3_gpu, *cpu_place, context, &input3);
TensorCopy(input3_gpu, cpu_place, context, &input3);
// numpy code:
// a = np.arange(6).reshape(2, 3)
......@@ -139,47 +231,105 @@ TEST(math_function, gemm_notrans_cublas) {
EXPECT_EQ(input3_ptr[5], 73);
EXPECT_EQ(input3_ptr[6], 86);
EXPECT_EQ(input3_ptr[7], 99);
delete gpu_place;
}
TEST(math_function, gemm_trans_cublas) {
paddle::framework::Tensor input1;
paddle::framework::Tensor input2;
paddle::framework::Tensor input3;
paddle::framework::Tensor input1_gpu;
paddle::framework::Tensor input2_gpu;
paddle::framework::Tensor input3_gpu;
TEST(math_function, gemm_notrans_cublas_fp16) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input2;
Tensor input3;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor input3_gpu;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
int m = 2;
int n = 3;
int k = 3;
float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
float16* input2_ptr = input2.mutable_data<float16>({3, 4}, cpu_place);
fill_fp16_data(input2_ptr, input2.numel(),
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11});
float16* input3_ptr = input3.mutable_data<float16>({2, 4}, cpu_place);
fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7});
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input2, gpu_place, context, &input2_gpu);
TensorCopy(input3, gpu_place, context, &input3_gpu);
float16* a = input1_gpu.data<float16>();
float16* b = input2_gpu.data<float16>();
float16* c = input3_gpu.mutable_data<float16>(gpu_place);
paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float16>(
context, false, false, m, n, k, float16(1), a, 3, b + 1, 4, float16(1),
c + 1, 4);
TensorCopy(input3_gpu, cpu_place, context, &input3);
// numpy code:
// a = np.arange(6).reshape(2, 3)
// b = np.arange(12).reshape(3, 4)[:, 1:]
// c = np.arange(8).reshape(2, 4)[:, 1:]
// out = np.arange(8).reshape(2, 4)
// out[:, 1:] = np.dot(a, b) + c
context.Wait();
EXPECT_EQ(static_cast<float>(input3_ptr[0]), 0);
EXPECT_EQ(static_cast<float>(input3_ptr[1]), 24);
EXPECT_EQ(static_cast<float>(input3_ptr[2]), 28);
EXPECT_EQ(static_cast<float>(input3_ptr[3]), 32);
EXPECT_EQ(static_cast<float>(input3_ptr[4]), 4);
EXPECT_EQ(static_cast<float>(input3_ptr[5]), 73);
EXPECT_EQ(static_cast<float>(input3_ptr[6]), 86);
EXPECT_EQ(static_cast<float>(input3_ptr[7]), 99);
}
TEST(math_function, gemm_trans_cublas_fp32) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input2;
Tensor input3;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor input3_gpu;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
int m = 2;
int n = 3;
int k = 3;
auto* cpu_place = new paddle::platform::CPUPlace();
float* input1_ptr = input1.mutable_data<float>({2, 3}, *cpu_place);
float* input1_ptr = input1.mutable_data<float>({2, 3}, cpu_place);
float arr1[6] = {0, 1, 2, 3, 4, 5};
memcpy(input1_ptr, arr1, 6 * sizeof(float));
float* input2_ptr = input2.mutable_data<float>({4, 3}, *cpu_place);
float* input2_ptr = input2.mutable_data<float>({4, 3}, cpu_place);
float arr2[12] = {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11};
memcpy(input2_ptr, arr2, 12 * sizeof(float));
float* input3_ptr = input3.mutable_data<float>({2, 4}, *cpu_place);
float* input3_ptr = input3.mutable_data<float>({2, 4}, cpu_place);
float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
memcpy(input3_ptr, arr3, 8 * sizeof(float));
auto* gpu_place = new paddle::platform::CUDAPlace(0);
paddle::platform::CUDADeviceContext context(*gpu_place);
paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu);
paddle::framework::TensorCopy(input2, *gpu_place, context, &input2_gpu);
paddle::framework::TensorCopy(input3, *gpu_place, context, &input3_gpu);
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input2, gpu_place, context, &input2_gpu);
TensorCopy(input3, gpu_place, context, &input3_gpu);
float* a = input1_gpu.data<float>();
float* b = input2_gpu.data<float>();
float* c = input3_gpu.mutable_data<float>(*gpu_place);
float* c = input3_gpu.mutable_data<float>(gpu_place);
paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float>(
context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4);
paddle::framework::TensorCopy(input3_gpu, *cpu_place, context, &input3);
context.Wait();
TensorCopy(input3_gpu, cpu_place, context, &input3);
context.Wait();
EXPECT_EQ(input3_ptr[0], 0);
EXPECT_EQ(input3_ptr[1], 24);
EXPECT_EQ(input3_ptr[2], 28);
......@@ -188,27 +338,81 @@ TEST(math_function, gemm_trans_cublas) {
EXPECT_EQ(input3_ptr[5], 73);
EXPECT_EQ(input3_ptr[6], 86);
EXPECT_EQ(input3_ptr[7], 99);
delete gpu_place;
}
TEST(math_function, gemm_trans_cublas_fp16) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor input1;
Tensor input2;
Tensor input3;
Tensor input1_gpu;
Tensor input2_gpu;
Tensor input3_gpu;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
int m = 2;
int n = 3;
int k = 3;
float16* input1_ptr = input1.mutable_data<float16>({2, 3}, cpu_place);
fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5});
float16* input2_ptr = input2.mutable_data<float16>({4, 3}, cpu_place);
fill_fp16_data(input2_ptr, input2.numel(),
{0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11});
float16* input3_ptr = input3.mutable_data<float16>({2, 4}, cpu_place);
fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7});
TensorCopy(input1, gpu_place, context, &input1_gpu);
TensorCopy(input2, gpu_place, context, &input2_gpu);
TensorCopy(input3, gpu_place, context, &input3_gpu);
float16* a = input1_gpu.data<float16>();
float16* b = input2_gpu.data<float16>();
float16* c = input3_gpu.mutable_data<float16>(gpu_place);
paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float16>(
context, false, true, m, n, k, float16(1), a, 3, b + 3, 3, float16(1),
c + 1, 4);
TensorCopy(input3_gpu, cpu_place, context, &input3);
context.Wait();
EXPECT_EQ(static_cast<float>(input3_ptr[0]), 0);
EXPECT_EQ(static_cast<float>(input3_ptr[1]), 24);
EXPECT_EQ(static_cast<float>(input3_ptr[2]), 28);
EXPECT_EQ(static_cast<float>(input3_ptr[3]), 32);
EXPECT_EQ(static_cast<float>(input3_ptr[4]), 4);
EXPECT_EQ(static_cast<float>(input3_ptr[5]), 73);
EXPECT_EQ(static_cast<float>(input3_ptr[6]), 86);
EXPECT_EQ(static_cast<float>(input3_ptr[7]), 99);
}
template <typename T>
void GemvTest(int m, int n, bool trans) {
paddle::framework::Tensor mat_a;
paddle::framework::Tensor vec_b;
paddle::framework::Tensor vec_c;
auto* cpu_place = new paddle::platform::CPUPlace();
T* data_a = mat_a.mutable_data<T>({m, n}, *cpu_place);
T* data_b = vec_b.mutable_data<T>({trans ? m : n}, *cpu_place);
T* data_c = vec_c.mutable_data<T>({trans ? n : m}, *cpu_place);
auto* gpu_place = new paddle::platform::CUDAPlace(0);
paddle::framework::Tensor g_mat_a;
paddle::framework::Tensor g_vec_b;
paddle::framework::Tensor g_vec_c;
T* g_data_a = g_mat_a.mutable_data<T>(mat_a.dims(), *gpu_place);
T* g_data_b = g_vec_b.mutable_data<T>(vec_b.dims(), *gpu_place);
T* g_data_c = g_vec_c.mutable_data<T>(vec_c.dims(), *gpu_place);
using namespace paddle::framework;
using namespace paddle::platform;
Tensor mat_a;
Tensor vec_b;
Tensor vec_c;
CPUPlace cpu_place;
CUDAPlace gpu_place(0);
CUDADeviceContext context(gpu_place);
T* data_a = mat_a.mutable_data<T>({m, n}, cpu_place);
T* data_b = vec_b.mutable_data<T>({trans ? m : n}, cpu_place);
T* data_c = vec_c.mutable_data<T>({trans ? n : m}, cpu_place);
Tensor g_mat_a;
Tensor g_vec_b;
Tensor g_vec_c;
T* g_data_a = g_mat_a.mutable_data<T>(mat_a.dims(), gpu_place);
T* g_data_b = g_vec_b.mutable_data<T>(vec_b.dims(), gpu_place);
T* g_data_c = g_vec_c.mutable_data<T>(vec_c.dims(), gpu_place);
for (int i = 0; i < mat_a.numel(); ++i) {
data_a[i] = static_cast<T>(i);
......@@ -217,16 +421,14 @@ void GemvTest(int m, int n, bool trans) {
data_b[i] = static_cast<T>(i);
}
paddle::platform::CUDADeviceContext context(*gpu_place);
paddle::framework::TensorCopy(mat_a, *gpu_place, context, &g_mat_a);
paddle::framework::TensorCopy(vec_b, *gpu_place, context, &g_vec_b);
TensorCopy(mat_a, gpu_place, context, &g_mat_a);
TensorCopy(vec_b, gpu_place, context, &g_vec_b);
paddle::operators::math::gemv<paddle::platform::CUDADeviceContext, T>(
paddle::operators::math::gemv<CUDADeviceContext, T>(
context, trans, static_cast<int>(m), static_cast<int>(n), 1., g_data_a,
g_data_b, 0., g_data_c);
paddle::framework::TensorCopy(g_vec_c, paddle::platform::CPUPlace(), context,
&vec_c);
TensorCopy(g_vec_c, cpu_place, context, &vec_c);
if (!trans) {
for (int i = 0; i < m; ++i) {
......
paddle/fluid/operators/math/sequence2batch.cc
浏览文件 @ b3a11fdf
......@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/math/sequence2batch.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
......
paddle/fluid/operators/read_op.cc
浏览文件 @ b3a11fdf
......@@ -60,15 +60,16 @@ class ReadOp : public framework::OperatorBase {
const platform::Place& dev_place) const override {
framework::ReaderHolder* reader =
scope.FindVar(Input("Reader"))->GetMutable<framework::ReaderHolder>();
if (!reader->HasNext()) {
std::vector<std::string> out_arg_names = Outputs("Out");
std::vector<framework::LoDTensor> ins;
reader->ReadNext(&ins);
if (ins.empty()) {
reader->ReInit();
reader->ReadNext(&ins);
PADDLE_ENFORCE(
reader->HasNext(),
!ins.empty(),
"Reader can not read the next data even it has been re-initialized.");
}
std::vector<std::string> out_arg_names = Outputs("Out");
std::vector<framework::LoDTensor> ins;
reader->ReadNext(&ins);
PADDLE_ENFORCE_EQ(ins.size(), out_arg_names.size());
for (size_t i = 0; i < ins.size(); ++i) {
auto* out =
......
paddle/fluid/operators/reader/create_batch_reader_op.cc
浏览文件 @ b3a11fdf
......@@ -68,10 +68,10 @@ void BatchReader::ReadNext(std::vector<framework::LoDTensor>* out) {
buffer_.clear();
buffer_.reserve(batch_size_);
for (int i = 0; i < batch_size_; ++i) {
if (reader_->HasNext()) {
buffer_.push_back(std::vector<framework::LoDTensor>());
reader_->ReadNext(&buffer_.back());
} else {
buffer_.push_back(std::vector<framework::LoDTensor>());
reader_->ReadNext(&buffer_.back());
if (buffer_.back().empty()) {
buffer_.pop_back();
break;
}
}
......
paddle/fluid/operators/reader/create_random_data_generator_op.cc
浏览文件 @ b3a11fdf
......@@ -50,8 +50,6 @@ class RandomDataGenerator : public framework::FileReader {
}
}
bool HasNext() const override { return true; }
void ReInit() override { return; }
private:
......
paddle/fluid/operators/reader/create_shuffle_reader_op.cc
浏览文件 @ b3a11fdf
......@@ -39,10 +39,10 @@ void ShuffleReader::ReadNext(std::vector<framework::LoDTensor>* out) {
buffer_.clear();
buffer_.reserve(buffer_size_);
for (int i = 0; i < buffer_size_; ++i) {
if (reader_->HasNext()) {
buffer_.push_back(std::vector<framework::LoDTensor>());
reader_->ReadNext(&buffer_.back());
} else {
buffer_.push_back(std::vector<framework::LoDTensor>());
reader_->ReadNext(&buffer_.back());
if (buffer_.back().empty()) {
buffer_.pop_back();
break;
}
}
......
paddle/fluid/operators/reshape_op.cc
浏览文件 @ b3a11fdf
......@@ -84,6 +84,9 @@ class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
AddAttr<std::vector<int>>("shape",
"(vector<int>) "
"Target shape of reshape operator.");
AddAttr<bool>("inplace",
"Change the source tensor's shape without copy memory.")
.SetDefault(true);
AddComment(R"DOC(
Reshape Operator.
......
paddle/fluid/operators/reshape_op.h
浏览文件 @ b3a11fdf
......@@ -26,10 +26,16 @@ class ReshapeKernel : public framework::OpKernel<T> {
void Compute(const framework::ExecutionContext& ctx) const {
auto* out = ctx.Output<framework::Tensor>("Out");
auto* in = ctx.Input<framework::Tensor>("X");
bool inplace = ctx.Attr<bool>("inplace");
auto out_dims = out->dims();
out->mutable_data<T>(ctx.GetPlace());
framework::TensorCopy(*in, ctx.GetPlace(), ctx.device_context(), out);
out->Resize(out_dims);
if (!inplace) {
out->mutable_data<T>(ctx.GetPlace());
framework::TensorCopy(*in, ctx.GetPlace(), ctx.device_context(), out);
out->Resize(out_dims);
} else {
out->ShareDataWith(*in);
out->Resize(out_dims);
}
}
};
......@@ -40,10 +46,16 @@ class ReshapeGradKernel : public framework::OpKernel<T> {
auto* d_out = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto* d_x = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
d_x->mutable_data<T>(ctx.GetPlace());
bool inplace = ctx.Attr<bool>("inplace");
auto in_dims = d_x->dims();
framework::TensorCopy(*d_out, ctx.GetPlace(), ctx.device_context(), d_x);
d_x->Resize(in_dims);
if (!inplace) {
framework::TensorCopy(*d_out, ctx.GetPlace(), ctx.device_context(), d_x);
d_x->Resize(in_dims);
} else {
d_x->ShareDataWith(*d_out);
d_x->Resize(in_dims);
}
}
};
} // namespace operators
......
paddle/fluid/platform/cudnn_helper.h
浏览文件 @ b3a11fdf
......@@ -15,6 +15,8 @@ limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/dynload/cudnn.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/macros.h"
......@@ -282,5 +284,17 @@ class ScopedPoolingDescriptor {
DISABLE_COPY_AND_ASSIGN(ScopedPoolingDescriptor);
};
inline bool CanCUDNNBeUsed(const framework::ExecutionContext& ctx) {
bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= paddle::platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (use_cudnn) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
return use_cudnn;
}
} // namespace platform
} // namespace paddle
paddle/fluid/platform/device_context.cc
浏览文件 @ b3a11fdf
......@@ -33,9 +33,15 @@ DeviceContextPool::DeviceContextPool(
PADDLE_ENFORCE_GT(places.size(), 0);
for (size_t i = 0; i < places.size(); i++) {
if (platform::is_cpu_place(places[i])) {
#ifdef PADDLE_WITH_MKLDNN
device_contexts_.emplace(places[i],
new platform::MKLDNNDeviceContext(
boost::get<platform::CPUPlace>(places[i])));
#else
device_contexts_.emplace(places[i],
new platform::CPUDeviceContext(
boost::get<platform::CPUPlace>(places[i])));
#endif
} else if (platform::is_gpu_place(places[i])) {
#ifdef PADDLE_WITH_CUDA
device_contexts_.emplace(places[i],
......@@ -121,6 +127,8 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface {
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
SetDeviceId(place_.device);
multi_process = GetCUDAMultiProcessors(place_.device);
max_threads_per_mp = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
eigen_stream_.reset(new EigenCudaStreamDevice());
eigen_stream_->Reinitialize(&stream_, place);
......@@ -154,6 +162,10 @@ void CUDADeviceContext::Wait() const {
PADDLE_ENFORCE(cudaGetLastError());
}
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
return multi_process * max_threads_per_mp;
}
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
return eigen_device_.get();
}
......@@ -170,64 +182,38 @@ cudaStream_t CUDADeviceContext::stream() const { return stream_; }
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
: CPUDeviceContext(place), ready_(false) {
stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0));
: CPUDeviceContext(place), engine_(mkldnn::engine::cpu, 0), p_blobs_() {
p_blobs_.reset(new std::unordered_map<std::string, std::shared_ptr<void>>());
}
template <typename T>
void MKLDNNDeviceContext::AddElement(const std::string& op_key,
const T& value) {
if (GetElement<T>(op_key)) {
return;
}
GetElementPool<T>().emplace(op_key, std::move(value));
}
void MKLDNNDeviceContext::SetBlob(const std::string& name,
std::shared_ptr<void> data) const {
std::unordered_map<std::string, std::shared_ptr<void>>* p;
p = p_blobs_.get();
template <typename T>
const T& MKLDNNDeviceContext::GetElement(const std::string& op_key) const {
auto it = GetElementPool<T>().find(op_key);
return it == GetElementPool<T>().end() ? nullptr : it->second;
}
auto it = p->find(name);
template <>
const std::unordered_map<const std::string, const MKLDNNMemoryPtr,
std::hash<std::string>>&
MKLDNNDeviceContext::GetElementPool<MKLDNNMemoryPtr>() const {
return memory_pool_;
}
if (it == p->end()) {
(*p)[name] = data; // create new blob
} else {
it->second = data; // set data to existing blob
}
template <>
const std::unordered_map<const std::string, const MKLDNNPrimitivePtr,
std::hash<std::string>>&
MKLDNNDeviceContext::GetElementPool<MKLDNNPrimitivePtr>() const {
return primitive_pool_;
return;
}
template <>
const std::unordered_map<const std::string, const MKLDNNPrimitiveDescPtr,
std::hash<std::string>>&
MKLDNNDeviceContext::GetElementPool<MKLDNNPrimitiveDescPtr>() const {
return primitive_desc_pool_;
}
std::shared_ptr<void> MKLDNNDeviceContext::GetBlob(
const std::string& name) const {
std::unordered_map<std::string, std::shared_ptr<void>>* p;
p = p_blobs_.get();
void MKLDNNDeviceContext::Execute(bool block) {
if (pipeline_.empty()) {
return;
}
ResetStream();
stream_->submit(pipeline_).wait(block);
ready_ = false;
pipeline_.clear();
}
auto it = p->find(name);
void MKLDNNDeviceContext::ResetStream() {
if (ready_) {
return;
if (it != p->end()) {
return it->second;
}
// TODO(TJ): change me when mkldnn have specific method to reset this state
stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
ready_ = true;
return nullptr;
}
#endif
......
paddle/fluid/platform/device_context.h
浏览文件 @ b3a11fdf
......@@ -22,7 +22,7 @@ limitations under the License. */
#endif
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#include <mkldnn.hpp>
#endif
#include "paddle/fluid/platform/enforce.h"
......@@ -79,6 +79,9 @@ class CUDADeviceContext : public DeviceContext {
/*! \brief Return place in the device context. */
Place GetPlace() const override;
/*! \brief Return the max physical thread count in the device context */
int GetMaxPhysicalThreadCount() const;
/*! \brief Return eigen device in the device context. */
Eigen::GpuDevice* eigen_device() const;
......@@ -100,6 +103,9 @@ class CUDADeviceContext : public DeviceContext {
cudaStream_t stream_;
cudnnHandle_t cudnn_handle_;
cublasHandle_t cublas_handle_;
int multi_process;
int max_threads_per_mp;
};
template <>
......@@ -114,46 +120,19 @@ class MKLDNNDeviceContext : public CPUDeviceContext {
public:
explicit MKLDNNDeviceContext(CPUPlace place);
/* \brief Add new element: memory, primitive or primitive desc */
template <typename T>
void AddElement(const std::string& op_key, const T& value);
/* \brief Get existed element: memory, primitive or primitive desc */
template <typename T>
const T& GetElement(const std::string& op_key) const;
/* \brief Get element pool: memory, primitive or primitive desc pool */
template <typename T>
const std::unordered_map<const std::string, const T, std::hash<std::string>>&
GetElementPool() const;
/* \brief Get the active engine */
const MKLDNNEngine& engine() const { return *engine_; }
/* \brief Submit primitive to pipeline */
void Submit(const MKLDNNPrimitivePtr& p) { pipeline_.push_back(*p); }
const mkldnn::engine& GetEngine() const { return engine_; }
/*! \brief Execute all submitted primitives in pipeline */
void Execute(bool block = true);
// Set data to blob (i.e. name/data pair). Create blob if not existing
void SetBlob(const std::string& name, std::shared_ptr<void> data) const;
protected:
/*! \brief Reset the stream to prepare next exectue */
void ResetStream();
// Find a saved blob. Return nullptr if not found
std::shared_ptr<void> GetBlob(const std::string& name) const;
private:
std::unordered_map<const std::string, const MKLDNNMemoryPtr,
std::hash<std::string>>
memory_pool_;
std::unordered_map<const std::string, const MKLDNNPrimitivePtr,
std::hash<std::string>>
primitive_pool_;
std::unordered_map<const std::string, const MKLDNNPrimitiveDescPtr,
std::hash<std::string>>
primitive_desc_pool_;
std::vector<MKLDNNPrimitive> pipeline_;
MKLDNNStreamPtr stream_;
MKLDNNEnginePtr engine_;
bool ready_;
mkldnn::engine engine_;
std::shared_ptr<std::unordered_map<std::string, std::shared_ptr<void>>>
p_blobs_;
};
#endif
......
paddle/fluid/platform/device_tracer.cc
浏览文件 @ b3a11fdf
......@@ -199,20 +199,29 @@ class DeviceTracerImpl : public DeviceTracer {
return;
}
std::lock_guard<std::mutex> l(trace_mu_);
cpu_records_.push_back(
CPURecord{anno, start_ns, end_ns,
std::hash<std::thread::id>{}(std::this_thread::get_id())});
cpu_records_.push_back(CPURecord{anno, start_ns, end_ns, 0});
}
void AddMemRecords(const std::string &name, uint64_t start_ns,
uint64_t end_ns, uint32_t device_id, uint32_t stream_id,
uint32_t correlation_id, uint64_t bytes) {
// 0 means timestamp information could not be collected for the kernel.
if (start_ns == 0 || end_ns == 0) {
VLOG(3) << name << " cannot be traced";
return;
}
std::lock_guard<std::mutex> l(trace_mu_);
mem_records_.push_back(MemRecord{name, start_ns, end_ns, device_id,
stream_id, correlation_id, bytes});
}
void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id,
uint32_t stream_id, uint32_t correlation_id) {
// 0 means timestamp information could not be collected for the kernel.
if (start == 0 || end == 0) {
VLOG(3) << correlation_id << " cannot be traced";
return;
}
std::lock_guard<std::mutex> l(trace_mu_);
kernel_records_.push_back(
KernelRecord{start, end, device_id, stream_id, correlation_id});
......@@ -285,10 +294,10 @@ class DeviceTracerImpl : public DeviceTracer {
event->set_device_id(r.device_id);
event->mutable_memcopy()->set_bytes(r.bytes);
}
std::string profile_str;
google::protobuf::TextFormat::PrintToString(profile_pb, &profile_str);
std::ofstream profile_f;
profile_f.open(profile_path, std::ios::out | std::ios::trunc);
std::string profile_str;
profile_pb.SerializeToString(&profile_str);
profile_f << profile_str;
profile_f.close();
return profile_pb;
......
paddle/fluid/platform/dynload/cublas.h
浏览文件 @ b3a11fdf
......@@ -68,6 +68,8 @@ extern void *cublas_dso_handle;
__macro(cublasDgemv_v2); \
__macro(cublasSgemm_v2); \
__macro(cublasDgemm_v2); \
__macro(cublasHgemm); \
__macro(cublasSgemmEx); \
__macro(cublasSgeam_v2); \
__macro(cublasDgeam_v2); \
__macro(cublasCreate_v2); \
......@@ -83,6 +85,7 @@ extern void *cublas_dso_handle;
__macro(cublasDgemmStridedBatched); \
__macro(cublasCgemmStridedBatched); \
__macro(cublasZgemmStridedBatched); \
__macro(cublasHgemmStridedBatched); \
__macro(cublasSgetrfBatched); \
__macro(cublasSgetriBatched); \
__macro(cublasDgetrfBatched); \
......
paddle/fluid/platform/gpu_info.cc
浏览文件 @ b3a11fdf
......@@ -33,6 +33,26 @@ int GetCUDADeviceCount() {
return count;
}
int GetCUDAMultiProcessors(int id) {
PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
int count;
PADDLE_ENFORCE(
cudaDeviceGetAttribute(&count, cudaDevAttrMultiProcessorCount, id),
"cudaDeviceGetAttribute failed in "
"paddle::platform::GetCUDAMultiProcessors");
return count;
}
int GetCUDAMaxThreadsPerMultiProcessor(int id) {
PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
int count;
PADDLE_ENFORCE(cudaDeviceGetAttribute(
&count, cudaDevAttrMaxThreadsPerMultiProcessor, id),
"cudaDeviceGetAttribute failed in "
"paddle::platform::GetCUDAMaxThreadsPerMultiProcessor");
return count;
}
int GetCurrentDeviceId() {
int device_id;
PADDLE_ENFORCE(
......
paddle/fluid/platform/gpu_info.h
浏览文件 @ b3a11fdf
......@@ -30,6 +30,12 @@ const std::string kEnvFractionGpuMemoryToUse =
//! Get the total number of GPU devices in system.
int GetCUDADeviceCount();
//! Get the MultiProcessors of the ith GPU.
int GetCUDAMultiProcessors(int i);
//! Get the MaxThreads of each MultiProcessor of the ith GPU.
int GetCUDAMaxThreadsPerMultiProcessor(int i);
//! Get the current GPU device id in system.
int GetCurrentDeviceId();
......
paddle/fluid/platform/mkldnn_helper.h
浏览文件 @ b3a11fdf
......@@ -16,12 +16,15 @@ limitations under the License. */
#include <mkldnn.hpp>
#include "paddle/fluid/framework/operator.h"
namespace paddle {
namespace platform {
using MKLDNNStream = mkldnn::stream;
using MKLDNNEngine = mkldnn::engine;
using MKLDNNMemory = mkldnn::memory;
using MKLDNNMemoryDescriptor = mkldnn::memory::desc;
using MKLDNNPrimitive = mkldnn::primitive;
using MKLDNNPrimitiveDesc = mkldnn::handle<mkldnn_primitive_desc_t>;
......@@ -31,5 +34,17 @@ typedef std::unique_ptr<MKLDNNMemory> MKLDNNMemoryPtr;
typedef std::unique_ptr<MKLDNNPrimitive> MKLDNNPrimitivePtr;
typedef std::unique_ptr<MKLDNNPrimitiveDesc> MKLDNNPrimitiveDescPtr;
inline mkldnn::memory::desc MKLDNNMemDesc(const std::vector<int>& dims,
mkldnn::memory::data_type data_type,
mkldnn::memory::format format) {
mkldnn::memory::dims tz = dims;
return mkldnn::memory::desc({tz}, data_type, format);
}
inline bool CanMKLDNNBeUsed(const framework::ExecutionContext& ctx) {
bool use_mkldnn = ctx.Attr<bool>("use_mkldnn");
return use_mkldnn && platform::is_cpu_place(ctx.GetPlace());
}
} // namespace platform
} // namespace paddle
paddle/fluid/platform/profiler.cc
浏览文件 @ b3a11fdf
......@@ -178,7 +178,7 @@ void EnableProfiler(ProfilerState state) {
}
#ifdef PADDLE_WITH_CUDA
if (g_state == ProfilerState::kCUDA) {
// Generate some dummy evenets first to reduce the startup overhead.
// Generate some dummy events first to reduce the startup overhead.
for (int i = 0; i < 5; i++) {
ForEachDevice([](int d) {
DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(d));
......
paddle/fluid/platform/profiler.proto
浏览文件 @ b3a11fdf
......@@ -15,7 +15,7 @@ limitations under the License. */
syntax = "proto2";
package paddle.platform.proto;
message MemCopy { optional uint64 bytes = 3; }
message MemCopy { optional uint64 bytes = 1; }
message Event {
optional string name = 1;
......
paddle/fluid/platform/profiler_test.cc
浏览文件 @ b3a11fdf
......@@ -75,6 +75,7 @@ TEST(RecordEvent, RecordEvent) {
* ...
* PopEvent(evt_name, dev_ctx);
*/
LOG(INFO) << "Usage 1: PushEvent & PopEvent";
for (int loop = 0; loop < 3; ++loop) {
for (int i = 1; i < 5; ++i) {
std::string name = "op_" + std::to_string(i);
......@@ -93,6 +94,7 @@ TEST(RecordEvent, RecordEvent) {
* ...
* }
*/
LOG(INFO) << "Usage 2: RecordEvent";
for (int i = 1; i < 5; ++i) {
std::string name = "evs_op_" + std::to_string(i);
RecordEvent record_event(name, dev_ctx);
......@@ -100,6 +102,34 @@ TEST(RecordEvent, RecordEvent) {
while (counter != i * 1000) counter++;
}
/* Usage 3
* {
* RecordEvent record_event(name1, dev_ctx);
* ...
* code to be analyzed
* ...
* {
* RecordEvent nested_record_event(name2, dev_ctx);
* ...
* code to be analyzed
* ...
* }
* }
*/
LOG(INFO) << "Usage 3: nested RecordEvent";
for (int i = 1; i < 5; ++i) {
std::string name = "ano_evs_op_" + std::to_string(i);
RecordEvent record_event(name, dev_ctx);
int counter = 1;
while (counter != i * 100) counter++;
{
std::string nested_name = "nested_ano_evs_op_" + std::to_string(i);
RecordEvent nested_record_event(nested_name, dev_ctx);
int nested_counter = 1;
while (nested_counter != i * 100) nested_counter++;
}
}
// Bad Usage:
PushEvent("event_without_pop", dev_ctx);
PopEvent("event_without_push", dev_ctx);
......
paddle/scripts/docker/build.sh
浏览文件 @ b3a11fdf
......@@ -213,7 +213,7 @@ function gen_fluid_inference_lib() {
if [ ${WITH_C_API:-OFF} == "OFF" ] ; then
cat <<EOF
========================================
Building fluid inference library ...
Deploying fluid inference library ...
========================================
EOF
make inference_lib_dist
......
python/paddle/fluid/layers/nn.py
浏览文件 @ b3a11fdf
......@@ -1111,6 +1111,7 @@ def conv2d(input,
param_attr=None,
bias_attr=None,
use_cudnn=True,
use_mkldnn=False,
act=None):
"""
**Convlution2D Layer**
......@@ -1252,7 +1253,8 @@ def conv2d(input,
'strides': stride,
'paddings': padding,
'groups': groups,
'use_cudnn': use_cudnn
'use_cudnn': use_cudnn,
'use_mkldnn': use_mkldnn
})
pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
......
python/paddle/fluid/memory_optimization_transpiler.py
浏览文件 @ b3a11fdf
......@@ -31,6 +31,8 @@ dtype_to_size = {
sub_block_ops = ["while", "while_grad", "parallel_do", "parallel_do_grad"]
PRINT_LOG = False
class ControlFlowGraph(object):
def __init__(self, Program, ops, forward_num, skip_opt):
......@@ -171,12 +173,14 @@ class ControlFlowGraph(object):
# TODO(qijun): actually, we should compare dtype_to_size[x_dtype]
# and dtype_to_size[cache_dtype]
if x_dtype == cache_dtype:
print(("Hit Cache !!!! cache pool index "
"is %d, var name is %s, "
"cached var name is %s, "
"var shape is %s ") %
(index, x, cache_var,
str(cache_shape)))
if PRINT_LOG:
print(
("Hit Cache !!!! cache pool index "
"is %d, var name is %s, "
"cached var name is %s, "
"var shape is %s ") %
(index, x, cache_var,
str(cache_shape)))
self.pool.pop(index)
if x == cache_var:
break
......@@ -277,7 +281,9 @@ def _get_cfgs(input_program):
return cfgs
def memory_optimize(input_program):
def memory_optimize(input_program, print_log=False):
global PRINT_LOG
PRINT_LOG = print_log
cfgs = _get_cfgs(input_program)
for cfg in cfgs:
cfg.memory_optimize()
python/paddle/fluid/nets.py
浏览文件 @ b3a11fdf
......@@ -29,14 +29,16 @@ def simple_img_conv_pool(input,
act,
param_attr=None,
pool_type='max',
use_cudnn=True):
use_cudnn=True,
use_mkldnn=False):
conv_out = layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
param_attr=param_attr,
act=act,
use_cudnn=use_cudnn)
use_cudnn=use_cudnn,
use_mkldnn=use_mkldnn)
pool_out = layers.pool2d(
input=conv_out,
......@@ -58,7 +60,8 @@ def img_conv_group(input,
conv_batchnorm_drop_rate=0.0,
pool_stride=1,
pool_type=None,
use_cudnn=True):
use_cudnn=True,
use_mkldnn=False):
"""
Image Convolution Group, Used for vgg net.
"""
......@@ -90,7 +93,8 @@ def img_conv_group(input,
padding=conv_padding[i],
param_attr=param_attr[i],
act=local_conv_act,
use_cudnn=use_cudnn)
use_cudnn=use_cudnn,
use_mkldnn=use_mkldnn)
if conv_with_batchnorm[i]:
tmp = layers.batch_norm(input=tmp, act=conv_act)
......
python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
浏览文件 @ b3a11fdf
......@@ -49,7 +49,7 @@ avg_cost = fluid.layers.mean(x=cost)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.01)
sgd_optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
fluid.memory_optimize(fluid.default_main_program(), print_log=True)
BATCH_SIZE = 200
......
tools/manylinux1/Dockerfile.x64
浏览文件 @ b3a11fdf
此差异已折叠。
tools/manylinux1/build_scripts/install_nccl2.sh 0 → 100644
浏览文件 @ b3a11fdf
此差异已折叠。
tools/timeline.py
浏览文件 @ b3a11fdf
此差异已折叠。
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