Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
1 年多 前同步成功

通知 699
Star 11112
Fork 2696
P
PaddleDetection
提交 547c1bd1 编写于 作者: T tangwei12
浏览文件
操作

Merge branch 'feature/add_prefech_hs' of https://github.com/JiabinYang/Paddle... 

Merge branch 'feature/add_prefech_hs' of https://github.com/JiabinYang/Paddle into add_prefetch_nce_hs
上级 e0c3c56b b5fa9164
展开全部 隐藏空白更改
内联 并排
Showing with 6331 addition and 1951 deletion
README.md
浏览文件 @ 547c1bd1
......@@ -19,6 +19,15 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
欢迎来到 PaddlePaddle GitHub
PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### Install Latest Stable Release:
```
......@@ -34,6 +43,23 @@ pip install paddlepaddle-gpu==1.2.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### 安装最新稳定版本:
```
# Linux CPU
pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.2.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.2.0.post85
# 其他平台上的安装指引请参考 http://paddlepaddle.org/
```
## Features
- **Flexibility**
......@@ -74,10 +100,38 @@ pip install paddlepaddle-gpu==1.2.0.post85
Baidu and it has achieved a significant impact. We hope you can also explore
the capability of PaddlePaddle to make an impact on your product.
## 特点
- **灵活性**
PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
- **高效性**
为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
- 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
- 通过MKL-DNN库优化CNN网络
- 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
- 针对高维稀疏数据模型，优化了局部和分布式训练。
- **稳定性**
有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
- **连接产品**
另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
## 安装
推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
......@@ -99,10 +153,37 @@ We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarte
We appreciate your contributions!
## 文档
我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
- [深度学习101](https://github.com/PaddlePaddle/book)
或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
可以在MPI集群上运行分布式训练任务
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
新的API支持代码更少更简洁的程序
- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
欢迎您的贡献!
## Ask Questions
You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues).
## 答疑
欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
## Copyright and License
PaddlePaddle is provided under the [Apache-2.0 license](LICENSE).
## 版权和许可证
PaddlePaddle由[Apache-2.0 license](LICENSE)提供
benchmark/fluid/fluid_benchmark.py
浏览文件 @ 547c1bd1
......@@ -81,9 +81,11 @@ def dist_transpile(trainer_id, args, train_prog, startup_prog):
# the role, should be either PSERVER or TRAINER
training_role = os.getenv("PADDLE_TRAINING_ROLE")
config = distribute_transpiler.DistributeTranspilerConfig()
config = fluid.DistributeTranspilerConfig()
config.slice_var_up = not args.no_split_var
config.min_block_size = 1048576
t = distribute_transpiler.DistributeTranspiler(config=config)
t.transpile(
trainer_id,
# NOTE: *MUST* use train_prog, for we are using with guard to
......
cmake/external/brpc.cmake
浏览文件 @ 547c1bd1
......@@ -14,14 +14,16 @@
INCLUDE(ExternalProject)
find_library(SSL_LIBRARY NAMES ssl)
find_package(OpenSSL REQUIRED)
message(STATUS "ssl:" ${OPENSSL_SSL_LIBRARY})
message(STATUS "crypto:" ${OPENSSL_CRYPTO_LIBRARY})
ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY})
SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY})
find_library(CRYPTO_LIBRARY NAMES crypto)
ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY})
SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY})
SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc)
SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc)
......@@ -31,14 +33,15 @@ SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc libr
INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR})
# Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib")
set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib|${THIRD_PARTY_PATH}/install/glog")
# If minimal .a is need, you can set WITH_DEBUG_SYMBOLS=OFF
ExternalProject_Add(
extern_brpc
${EXTERNAL_PROJECT_LOG_ARGS}
# TODO(gongwb): change to de newst repo when they changed.
GIT_REPOSITORY "https://github.com/gongweibao/brpc"
GIT_TAG "7dc04defad1fd4173aae170c3fcbde131b65155a"
GIT_TAG "e9b67ec1b7458f2af5fae76451afe1e27e01b4b4"
PREFIX ${BRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......@@ -50,7 +53,7 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-DCMAKE_PREFIX_PATH=${prefix_path}
-DBRPC_WITH_GLOG=ON
-DWITH_GLOG=ON
-DIOBUF_WITH_HUGE_BLOCK=ON
-DBRPC_WITH_RDMA=${WITH_BRPC_RDMA}
${EXTERNAL_OPTIONAL_ARGS}
......@@ -65,5 +68,6 @@ ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
ADD_DEPENDENCIES(brpc extern_brpc)
add_definitions(-DBRPC_WITH_GLOG)
LIST(APPEND external_project_dependencies brpc)
cmake/external/gtest.cmake
浏览文件 @ 547c1bd1
......@@ -12,8 +12,12 @@
# See the License for the specific language governing permissions and
# limitations under the License.
IF(WITH_TESTING)
ENABLE_TESTING()
#FIXME:(gongwb) Move brpc's gtest dependency.
IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
IF(WITH_TESTING)
ENABLE_TESTING()
ENDIF(WITH_TESTING)
INCLUDE(ExternalProject)
SET(GTEST_SOURCES_DIR ${THIRD_PARTY_PATH}/gtest)
......@@ -76,4 +80,4 @@ IF(WITH_TESTING)
ADD_DEPENDENCIES(gtest_main extern_gtest)
LIST(APPEND external_project_dependencies gtest gtest_main)
ENDIF(WITH_TESTING)
ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
cmake/external/leveldb.cmake
浏览文件 @ 547c1bd1
......@@ -24,8 +24,8 @@ ExternalProject_Add(
extern_leveldb
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${LEVELDB_SOURCES_DIR}
URL "https://github.com/google/leveldb/archive/v1.18.tar.gz"
URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0"
GIT_REPOSITORY "https://github.com/google/leveldb"
GIT_TAG v1.18
CONFIGURE_COMMAND ""
BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a
INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/
......
cmake/external/python.cmake
浏览文件 @ 547c1bd1
......@@ -18,8 +18,8 @@ ENDIF()
INCLUDE(python_module)
FIND_PACKAGE(PythonInterp ${PY_VERSION})
FIND_PACKAGE(PythonLibs ${PY_VERSION})
FIND_PACKAGE(PythonInterp ${PY_VERSION} REQUIRED)
FIND_PACKAGE(PythonLibs ${PY_VERSION} REQUIRED)
if(WIN32)
execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
......@@ -79,6 +79,5 @@ IF(PYTHONINTERP_FOUND)
"please use pip to upgrade protobuf. pip install -U protobuf")
ENDIF()
ENDIF(PYTHONINTERP_FOUND)
INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
paddle/fluid/API.spec
浏览文件 @ 547c1bd1
......@@ -77,6 +77,8 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name']
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.adaptive_pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
paddle.fluid.layers.adaptive_pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
......@@ -198,6 +200,7 @@ paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act
paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ 547c1bd1
......@@ -72,6 +72,8 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
cc_test(reader_test SRCS reader_test.cc DEPS reader)
......@@ -167,9 +169,12 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass variable_helper)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
if(WITH_NGRAPH)
if(NOT WIN32)
......@@ -183,6 +188,8 @@ else()
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
target_link_libraries(executor garbage_collector)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS
threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
graph build_strategy
......
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ 547c1bd1
......@@ -85,7 +85,7 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var,
out->mutable_data(expected_kernel_type.place_, in.type());
framework::VisitDataType(
framework::ToDataType(in.type()),
in.type(),
CastDataLayout(pool.Get(expected_kernel_type.place_), axis, in, out));
out->set_layout(expected_kernel_type.data_layout_);
......@@ -101,7 +101,7 @@ void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) {
case mkldnn::memory::data_type::f32:
return platform::to_void_cast(tensor.data<float>());
case mkldnn::memory::data_type::s8:
return platform::to_void_cast(tensor.data<char>());
return platform::to_void_cast(tensor.data<int8_t>());
case mkldnn::memory::data_type::u8:
return platform::to_void_cast(tensor.data<unsigned char>());
case mkldnn::memory::data_type::s16:
......@@ -144,7 +144,7 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
memory::data_type in_type = ToMKLDNNDataType(in.type());
PADDLE_ENFORCE(in_type != memory::data_type::data_undef,
"Input tensor type is not supported: ", in.type().name());
"Input tensor type is not supported: %s", in.type());
memory::data_type out_type = in_type;
auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
......
paddle/fluid/framework/data_layout_transform.h
浏览文件 @ 547c1bd1
......@@ -50,14 +50,14 @@ inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) {
}
}
inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
static const std::map<std::type_index, MKLDNNDataType> dict{
{std::type_index(typeid(float)), MKLDNNDataType::f32}, // NOLINT
{std::type_index(typeid(char)), MKLDNNDataType::s8}, // NOLINT
{std::type_index(typeid(unsigned char)), MKLDNNDataType::u8},
{std::type_index(typeid(int16_t)), MKLDNNDataType::s16},
{std::type_index(typeid(int32_t)), MKLDNNDataType::s32}};
auto iter = dict.find(type);
inline MKLDNNDataType ToMKLDNNDataType(proto::VarType::Type type) {
static std::unordered_map<int, MKLDNNDataType> dict{
{DataTypeTrait<float>::DataType, MKLDNNDataType::f32},
{DataTypeTrait<int8_t>::DataType, MKLDNNDataType::s8},
{DataTypeTrait<uint8_t>::DataType, MKLDNNDataType::u8},
{DataTypeTrait<int16_t>::DataType, MKLDNNDataType::s16},
{DataTypeTrait<int32_t>::DataType, MKLDNNDataType::s32}};
auto iter = dict.find(static_cast<int>(type));
if (iter != dict.end()) return iter->second;
return MKLDNNDataType::data_undef;
}
......
paddle/fluid/framework/data_type.cc
浏览文件 @ 547c1bd1
......@@ -26,7 +26,7 @@ struct DataTypeMap {
std::unordered_map<std::type_index, proto::VarType::Type> cpp_to_proto_;
std::unordered_map<int, std::type_index> proto_to_cpp_;
std::unordered_map<int, std::string> proto_to_str_;
std::unordered_map<std::type_index, size_t> cpp_to_size_;
std::unordered_map<int, size_t> proto_to_size_;
};
static DataTypeMap* InitDataTypeMap();
......@@ -45,7 +45,7 @@ static inline void RegisterType(DataTypeMap* map,
map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
map->cpp_to_proto_.emplace(typeid(T), proto_type);
map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
map->cpp_to_size_.emplace(typeid(T), sizeof(T));
map->proto_to_size_.emplace(static_cast<int>(proto_type), sizeof(T));
}
static DataTypeMap* InitDataTypeMap() {
......@@ -54,17 +54,7 @@ static DataTypeMap* InitDataTypeMap() {
#define RegType(cc_type, proto_type) \
RegisterType<cc_type>(retv, proto_type, #cc_type)
// NOTE: Add your customize type here.
RegType(float16, proto::VarType::FP16);
RegType(float, proto::VarType::FP32);
RegType(double, proto::VarType::FP64);
RegType(int, proto::VarType::INT32);
RegType(int64_t, proto::VarType::INT64);
RegType(bool, proto::VarType::BOOL);
RegType(size_t, proto::VarType::SIZE_T);
RegType(int16_t, proto::VarType::INT16);
RegType(uint8_t, proto::VarType::UINT8);
RegType(int8_t, proto::VarType::INT8);
_ForEachDataType_(RegType);
#undef RegType
return retv;
......@@ -96,12 +86,12 @@ std::string DataTypeToString(const proto::VarType::Type type) {
static_cast<int>(type));
}
size_t SizeOfType(std::type_index type) {
auto it = gDataTypeMap().cpp_to_size_.find(type);
if (it != gDataTypeMap().cpp_to_size_.end()) {
size_t SizeOfType(proto::VarType::Type type) {
auto it = gDataTypeMap().proto_to_size_.find(static_cast<int>(type));
if (it != gDataTypeMap().proto_to_size_.end()) {
return it->second;
}
PADDLE_THROW("Not support %s as tensor type", type.name());
PADDLE_THROW("Not support %s as tensor type", DataTypeToString(type));
}
} // namespace framework
......
paddle/fluid/framework/data_type.h
浏览文件 @ 547c1bd1
......@@ -22,46 +22,59 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <typename T>
struct DataTypeTrait {};
// Stub handle for void
template <>
struct DataTypeTrait<void> {
constexpr static auto DataType = proto::VarType::RAW;
};
#define _ForEachDataTypeHelper_(callback, cpp_type, proto_type) \
callback(cpp_type, ::paddle::framework::proto::VarType::proto_type);
#define _ForEachDataType_(callback) \
_ForEachDataTypeHelper_(callback, float, FP32); \
_ForEachDataTypeHelper_(callback, ::paddle::platform::float16, FP16); \
_ForEachDataTypeHelper_(callback, double, FP64); \
_ForEachDataTypeHelper_(callback, int, INT32); \
_ForEachDataTypeHelper_(callback, int64_t, INT64); \
_ForEachDataTypeHelper_(callback, bool, BOOL); \
_ForEachDataTypeHelper_(callback, uint8_t, UINT8); \
_ForEachDataTypeHelper_(callback, int16_t, INT16); \
_ForEachDataTypeHelper_(callback, int8_t, INT8)
#define DefineDataTypeTrait(cpp_type, proto_type) \
template <> \
struct DataTypeTrait<cpp_type> { \
constexpr static auto DataType = proto_type; \
}
_ForEachDataType_(DefineDataTypeTrait);
#undef DefineDataTypeTrait
extern proto::VarType::Type ToDataType(std::type_index type);
extern std::type_index ToTypeIndex(proto::VarType::Type type);
template <typename Visitor>
inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
switch (type) {
case proto::VarType::FP16:
visitor.template apply<platform::float16>();
break;
case proto::VarType::FP32:
visitor.template apply<float>();
break;
case proto::VarType::FP64:
visitor.template apply<double>();
break;
case proto::VarType::INT32:
visitor.template apply<int>();
break;
case proto::VarType::INT64:
visitor.template apply<int64_t>();
break;
case proto::VarType::BOOL:
visitor.template apply<bool>();
break;
case proto::VarType::UINT8:
visitor.template apply<uint8_t>();
break;
case proto::VarType::INT16:
visitor.template apply<int16_t>();
break;
case proto::VarType::INT8:
visitor.template apply<int8_t>();
break;
default:
PADDLE_THROW("Not supported %d", type);
}
#define VisitDataTypeCallback(cpp_type, proto_type) \
do { \
if (type == proto_type) { \
visitor.template apply<cpp_type>(); \
return; \
} \
} while (0)
_ForEachDataType_(VisitDataTypeCallback);
#undef VisitDataTypeCallback
PADDLE_THROW("Not supported %d", type);
}
extern std::string DataTypeToString(const proto::VarType::Type type);
extern size_t SizeOfType(std::type_index type);
extern size_t SizeOfType(proto::VarType::Type type);
inline std::ostream& operator<<(std::ostream& out,
const proto::VarType::Type& type) {
out << DataTypeToString(type);
......
paddle/fluid/framework/data_type_test.cc
浏览文件 @ 547c1bd1
......@@ -26,15 +26,15 @@ TEST(DataType, float16) {
Tensor tensor;
CPUPlace cpu;
tensor.mutable_data(cpu, f::ToTypeIndex(dtype));
tensor.mutable_data(cpu, dtype);
// test fp16 tensor
EXPECT_EQ(tensor.type(), std::type_index(typeid(float16)));
EXPECT_EQ(tensor.type(), f::ToDataType(typeid(float16)));
// test fp16 size
EXPECT_EQ(f::SizeOfType(f::ToTypeIndex(dtype)), 2u);
EXPECT_EQ(f::SizeOfType(dtype), 2u);
// test debug info
std::string type = "float16";
std::string type = "::paddle::platform::float16";
EXPECT_STREQ(f::DataTypeToString(dtype).c_str(), type.c_str());
}
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ 547c1bd1
......@@ -12,12 +12,19 @@ cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc
cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
if(WITH_DISTRIBUTE)
if(NOT WITH_GRPC)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(reduce_op_handle.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
endif()
endif()
if(WITH_GPU)
nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
if(WITH_DISTRIBUTE)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor sendrecvop_grpc)
ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
else()
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor)
......@@ -30,7 +37,7 @@ else()
variable_visitor)
if(WITH_DISTRIBUTE)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor sendrecvop_grpc)
ddim selected_rows_functor sendrecvop_rpc)
else()
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor)
......@@ -45,10 +52,10 @@ cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base s
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
if (WITH_GPU)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
endif()
cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper)
cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
......@@ -56,10 +63,7 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass)
if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
......
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ 547c1bd1
......@@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() {
// Reduce All Tensor to trg in CPU
ReduceLoDTensor func(lod_tensors, &trg);
VisitDataType(ToDataType(lod_tensors[0]->type()), func);
VisitDataType(lod_tensors[0]->type(), func);
for (size_t i = 1; i < local_scopes_.size(); ++i) {
auto &scope =
......
paddle/fluid/framework/details/computation_op_handle.cc
浏览文件 @ 547c1bd1
......@@ -20,11 +20,13 @@ namespace paddle {
namespace framework {
namespace details {
ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
platform::Place place)
platform::Place place,
size_t scope_idx)
: OpHandleBase(node),
op_(framework::OpRegistry::CreateOp(*node->Op())),
scope_(scope),
place_(place) {}
place_(place),
scope_idx_(scope_idx) {}
void ComputationOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ 547c1bd1
......@@ -28,7 +28,8 @@ namespace framework {
namespace details {
struct ComputationOpHandle : public OpHandleBase {
public:
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place);
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
size_t scope_idx);
std::string Name() const override;
......@@ -38,6 +39,8 @@ struct ComputationOpHandle : public OpHandleBase {
void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; }
size_t GetScopeIdx() const { return scope_idx_; }
protected:
void RunImpl() override;
......@@ -47,6 +50,7 @@ struct ComputationOpHandle : public OpHandleBase {
std::unique_ptr<OperatorBase> op_;
Scope *scope_;
platform::Place place_;
size_t scope_idx_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
......
paddle/fluid/framework/details/eager_deletion_op_handle.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
namespace paddle {
namespace framework {
namespace details {
EagerDeletionOpHandle::EagerDeletionOpHandle(
ir::Node *node, const Scope *scope, const platform::Place &place,
const std::unordered_set<std::string> &var_names, GarbageCollector *gc,
AtomicReferenceCountMap *ref_cnts)
: OpHandleBase(node),
scope_(scope),
var_names_(var_names),
gc_(gc),
ref_cnts_(ref_cnts) {
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place)) {
dev_ctx_ = reinterpret_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (dynamic_cast<StreamGarbageCollector *>(gc_)) {
platform::CUDADeviceGuard guard(
boost::get<platform::CUDAPlace>(place).device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
PADDLE_ENFORCE_NOT_NULL(event_);
}
}
#endif
}
EagerDeletionOpHandle::~EagerDeletionOpHandle() {
#ifdef PADDLE_WITH_CUDA
if (event_) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
platform::CUDADeviceGuard guard(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
#endif
}
std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; }
void EagerDeletionOpHandle::RunImpl() {
auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
std::deque<std::shared_ptr<memory::Allocation>> garbages;
for (auto &name : var_names_) {
auto it = ref_cnts_->find(name);
// Var not found, not reference count has not decreased to 0
if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) {
continue;
}
auto *var = exec_scope->FindVar(name);
if (var == nullptr) {
continue;
}
VLOG(2) << "Erase variable " << name;
if (var->IsType<LoDTensor>()) {
garbages.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
garbages.emplace_back(
var->GetMutable<SelectedRows>()->mutable_value()->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
auto *tensor_arr = var->GetMutable<LoDTensorArray>();
for (auto &t : *tensor_arr) {
garbages.emplace_back(t.MoveMemoryHolder());
}
} else {
PADDLE_THROW("Type %s of %s is not supported eager deletion",
var->Type().name(), name);
}
}
if (!garbages.empty()) {
ClearGarbages(&garbages);
}
}
void EagerDeletionOpHandle::ClearGarbages(
std::deque<std::shared_ptr<memory::Allocation>> *garbages) {
#ifdef PADDLE_WITH_CUDA
if (event_) {
auto compute_stream = dev_ctx_->stream();
auto callback_stream =
reinterpret_cast<StreamGarbageCollector *>(gc_)->stream();
auto callback_func = [=]() {
PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
};
gc_->Add(std::move(*garbages), callback_func);
} else {
#endif
gc_->Add(std::move(*garbages));
#ifdef PADDLE_WITH_CUDA
}
#endif
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/eager_deletion_op_handle.h 0 → 100644
浏览文件 @ 547c1bd1
// 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 <deque>
#include <string>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
namespace paddle {
namespace framework {
class Scope;
namespace details {
class EagerDeletionOpHandle : public OpHandleBase {
public:
EagerDeletionOpHandle(ir::Node *node, const Scope *scope,
const platform::Place &place,
const std::unordered_set<std::string> &var_names,
GarbageCollector *gc,
AtomicReferenceCountMap *ref_cnts);
~EagerDeletionOpHandle();
std::string Name() const override;
protected:
void RunImpl() override;
private:
void ClearGarbages(std::deque<std::shared_ptr<memory::Allocation>> *garbages);
const Scope *scope_;
std::unordered_set<std::string> var_names_;
GarbageCollector *gc_; // not own
AtomicReferenceCountMap *ref_cnts_; // not own
#ifdef PADDLE_WITH_CUDA
platform::CUDADeviceContext *dev_ctx_{nullptr};
cudaEvent_t event_{nullptr};
#endif
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/eager_deletion_pass.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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 <queue>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_pass.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts =
Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
PADDLE_ENFORCE(ref_cnts.empty(),
"kRuntimeReferenceCount should be initialized here!");
const auto &vars = graph->Get<GraphVars>(kGraphVars);
ref_cnts.resize(vars.size());
const auto &last_live_ops =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
const auto &gcs = Get<GarbageCollectorMap>(kGarbageCollector);
const auto &places = Get<std::vector<platform::Place>>(kAllPlaces);
// a reverse map of last_live_ops
// i.e., last op --> variable names which can be deleted.
std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>
op_vars_map;
for (auto &var_ops_map : last_live_ops) {
for (auto &var_ops_pair : var_ops_map) {
const std::string &var_name = var_ops_pair.first;
for (auto *op : var_ops_pair.second) {
op_vars_map[op].insert(var_name);
}
}
}
for (auto &pair : op_vars_map) {
auto *op = pair.first;
auto &var_names = pair.second;
auto *eager_deletion_node =
graph->CreateEmptyNode("eager_deletion", ir::Node::Type::kOperation);
auto *eager_deletion_op = new EagerDeletionOpHandle(
eager_deletion_node, op->GetScope(), op->GetPlace(), var_names,
gcs.at(places[op->GetScopeIdx()]).get(),
&(ref_cnts[op->GetScopeIdx()]));
auto it = std::find_if(
op->Outputs().begin(), op->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != op->Outputs().end()) {
eager_deletion_op->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
op->AddOutput(dep_var);
eager_deletion_op->AddInput(dep_var);
}
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
eager_deletion_op->AddOutput(dummy_leaf);
}
VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(eager_deletion_pass,
paddle::framework::details::EagerDeletionPass)
.RequirePassAttr(paddle::framework::details::kRuntimeReferenceCount)
.RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars)
.RequirePassAttr(paddle::framework::details::kAllPlaces)
.RequirePassAttr(paddle::framework::details::kGarbageCollector);
paddle/fluid/framework/details/eager_deletion_pass.h 0 → 100644
浏览文件 @ 547c1bd1
// 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/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class EagerDeletionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fuse_vars_op_handle.h
浏览文件 @ 547c1bd1
......@@ -33,7 +33,7 @@ struct FuseVarsOpHandle : public OpHandleBase {
FuseVarsOpHandle(ir::Node *node, Scope *local_scope,
const platform::Place &place,
const std::unordered_map<std::string, int64_t> &inputs_numel,
const std::type_index &var_type)
const proto::VarType::Type var_type)
: OpHandleBase(node),
local_scope_(local_scope),
place_(place),
......@@ -57,7 +57,7 @@ struct FuseVarsOpHandle : public OpHandleBase {
Scope *local_scope_;
const platform::Place place_;
const std::unordered_map<std::string, int64_t> inputs_numel_;
const std::type_index type_;
const proto::VarType::Type type_;
int64_t total_numel_;
};
} // namespace details
......
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ 547c1bd1
......@@ -565,7 +565,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
int dev_id) const {
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()),
local_scopes_[dev_id], places_[dev_id]));
local_scopes_[dev_id], places_[dev_id], dev_id));
CreateOpHandleIOs(result, node, dev_id);
}
......@@ -688,8 +688,8 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) {
auto p = places_[scope_idx];
auto s = local_scopes_[scope_idx];
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p));
result->Get<GraphOps>(kGraphOps).emplace_back(new ComputationOpHandle(
result->CreateOpNode(node->Op()), s, p, scope_idx));
CreateOpHandleIOs(result, node, scope_idx);
}
}
......
paddle/fluid/framework/details/op_graph_view.cc
浏览文件 @ 547c1bd1
......@@ -23,6 +23,8 @@ namespace details {
OpGraphView::OpGraphView(const std::vector<OpHandleBase *> &ops) { Build(ops); }
void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
preceding_ops_.clear();
pending_ops_.clear();
for (auto &op : ops) {
preceding_ops_[op];
pending_ops_[op];
......@@ -40,6 +42,7 @@ void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
std::unordered_set<OpHandleBase *> OpGraphView::AllOps() const {
std::unordered_set<OpHandleBase *> ret;
ret.reserve(preceding_ops_.size());
for (auto &pair : preceding_ops_) {
ret.insert(pair.first);
}
......
paddle/fluid/framework/details/op_graph_view.h
浏览文件 @ 547c1bd1
......@@ -14,7 +14,7 @@
#pragma once
#include <memory>
#include <queue>
#include <unordered_map>
#include <unordered_set>
#include <vector>
......@@ -34,6 +34,11 @@ class OpGraphView {
bool HasOp(OpHandleBase *op) const;
// Use a visitor to visit all pending ops of op
// Stop when callback returns false
template <typename Callback>
bool VisitAllPendingOps(OpHandleBase *op, Callback &&callback) const;
private:
void Build(const std::vector<OpHandleBase *> &ops);
void EnforceHasOp(OpHandleBase *op) const;
......@@ -44,6 +49,28 @@ class OpGraphView {
pending_ops_;
};
template <typename Callback>
bool OpGraphView::VisitAllPendingOps(OpHandleBase *op,
Callback &&callback) const {
EnforceHasOp(op);
std::unordered_set<OpHandleBase *> visited;
std::queue<OpHandleBase *> q;
q.push(op);
do {
op = q.front();
q.pop();
for (auto &pending_op : pending_ops_.at(op)) {
if (visited.count(pending_op) == 0) {
visited.insert(pending_op);
if (!callback(pending_op)) {
return false;
}
}
}
} while (!q.empty());
return true;
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @ 547c1bd1
......@@ -218,18 +218,18 @@ void ReduceOpHandle::RunImpl() {
}
#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
if (framework::IsType<const float>(in_selected_rows[0]->value().type())) {
if (in_selected_rows[0]->value().type() ==
framework::proto::VarType::FP32) {
GatherSelectedRows<platform::CUDADeviceContext, float>(
in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
} else if (framework::IsType<const double>(
in_selected_rows[0]->value().type())) {
} else if (in_selected_rows[0]->value().type() ==
framework::proto::VarType::FP64) {
GatherSelectedRows<platform::CUDADeviceContext, double>(
in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
} else {
PADDLE_ENFORCE(false,
"only support double or float when gahter SelectedRows");
PADDLE_THROW("only support double or float when gather SelectedRows");
}
#endif
});
......@@ -246,7 +246,7 @@ void ReduceOpHandle::RunImpl() {
if (!FLAGS_cpu_deterministic) {
ReduceLoDTensor func(lod_tensors,
out_var->GetMutable<framework::LoDTensor>());
VisitDataType(ToDataType(lod_tensors[0]->type()), func);
VisitDataType(lod_tensors[0]->type(), func);
} else {
// We sum lod_tensors to reduce_sum_trg which is in local_scopes_0
// here, but it doesn't mean reduce_sum_trg must be in local_scopes_0.
......@@ -256,7 +256,7 @@ void ReduceOpHandle::RunImpl() {
->FindVar(out_var_handle->name_)
->GetMutable<framework::LoDTensor>();
ReduceLoDTensor func(lod_tensors, &reduce_sum_trg);
VisitDataType(ToDataType(lod_tensors[0]->type()), func);
VisitDataType(lod_tensors[0]->type(), func);
auto trg = out_var->GetMutable<framework::LoDTensor>();
if (reduce_sum_trg.data<void>() != trg->data<void>()) {
......
paddle/fluid/framework/details/reference_count_op_handle.h 已删除 100644 → 0
浏览文件 @ e0c3c56b
// 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 <atomic>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/tensor.h"
namespace paddle {
namespace framework {
namespace details {
using ReferenceCountMap = std::unordered_map<std::string, int>;
using AtomicReferenceCountMap =
std::unordered_map<std::string, std::atomic<int>>;
using DeviceReferenceCountMap =
std::unordered_map<int, std::unique_ptr<ReferenceCountMap>>;
using AtomicDeviceReferenceCountMap =
std::unordered_map<int, std::unique_ptr<AtomicReferenceCountMap>>;
using DeviceGarbageCollectorMap =
std::unordered_map<int,
std::unique_ptr<GarbageCollector<framework::Tensor>>>;
class ReferenceCountOpHandle : public OpHandleBase {
public:
ReferenceCountOpHandle(ir::Node *node, const Scope *scope,
const platform::CUDAPlace &place,
const std::vector<std::string> &var_names,
GarbageCollector<Tensor> *gc,
AtomicReferenceCountMap *ref_cnts)
: OpHandleBase(node), scope_(scope), gc_(gc), ref_cnts_(ref_cnts) {
dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (IsStreamGarabageCollector()) {
platform::SetDeviceId(place.device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
for (auto &name : var_names) AddVar(name);
}
~ReferenceCountOpHandle() {
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
platform::SetDeviceId(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
}
std::string Name() const override { return "reference_count"; }
void AddVar(const std::string &name) {
auto it = var_names_.find(name);
if (it != var_names_.end())
++(it->second);
else
var_names_[name] = 1;
}
protected:
void RunImpl() override {
auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
std::vector<Tensor *> tensors;
for (auto &pair : var_names_) {
auto &name = pair.first;
auto it = ref_cnts_->find(name);
if (it == ref_cnts_->end()) continue;
auto *var = exec_scope->FindVar(name);
if (var == nullptr) continue;
if (var->IsType<LoDTensor>()) {
if (it->second.fetch_sub(pair.second) <= pair.second) {
tensors.emplace_back(var->GetMutable<LoDTensor>());
}
} else if (var->IsType<SelectedRows>()) {
if (it->second.fetch_sub(pair.second) <= pair.second) {
tensors.emplace_back(
var->GetMutable<SelectedRows>()->mutable_value());
}
}
}
if (!tensors.empty()) {
ClearTensors(tensors);
}
}
private:
void ClearTensors(const std::vector<Tensor *> &tensors) {
auto *gc = dynamic_cast<StreamGarbageCollector<Tensor> *>(gc_);
if (gc != nullptr) {
auto compute_stream = dev_ctx_->stream();
auto callback_stream = gc->stream();
auto callback_func = [=]() {
PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
};
gc_->Add(tensors, callback_func);
} else {
gc_->Add(tensors);
}
}
bool IsStreamGarabageCollector() const {
return dynamic_cast<const StreamGarbageCollector<Tensor> *>(gc_) != nullptr;
}
const Scope *scope_;
platform::CUDADeviceContext *dev_ctx_;
std::unordered_map<std::string, int> var_names_;
GarbageCollector<Tensor> *gc_; // not own
AtomicReferenceCountMap *ref_cnts_; // not own
cudaEvent_t event_;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reference_count_pass.cc
浏览文件 @ 547c1bd1
......@@ -14,187 +14,240 @@
#include <queue>
#include <string>
#include <type_traits>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/op_graph_view.h"
#include "paddle/fluid/framework/details/reference_count_pass.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) {
std::queue<VarHandleBase *> queue;
queue.push(var_in);
do {
auto *var = queue.front();
queue.pop();
for (auto *op : var->PendingOps()) {
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op != nullptr && compute_op->GetPlace() == var_in->place_) {
return compute_op;
// A functor to shrink/remove operators who depend on other operators in a set
class ShrinkDepsOpFunctor {
private:
enum RelationShip { kSame = 0, kNoDeps = 1, kBefore = 2, kAfter = 3 };
public:
explicit ShrinkDepsOpFunctor(const std::vector<OpHandleBase *> &all_ops)
: graph_(all_ops) {}
template <typename OpSet>
OpSet operator()(const OpSet &op_set) const {
using KeyType = typename OpSet::key_type;
static_assert(
std::is_base_of<OpHandleBase,
typename std::remove_pointer<KeyType>::type>::value,
"Key type of OpSet must be OpHandleBase, or derived of OpHandleBase");
if (op_set.size() <= 1) return op_set;
std::vector<OpHandleBase *> ops(op_set.begin(), op_set.end());
OpSet ret;
auto rels = GetRelations(ops);
auto not_before = [](RelationShip r) { return r != kBefore; };
for (size_t i = 0; i < rels.size(); ++i) {
if (std::all_of(rels[i].begin(), rels[i].end(), not_before)) {
ret.emplace(static_cast<KeyType>(ops[i]));
}
for (auto *out_var : op->Outputs()) {
queue.push(out_var);
}
return ret;
}
private:
std::vector<std::vector<RelationShip>> GetRelations(
const std::vector<OpHandleBase *> &ops) const {
std::unordered_map<OpHandleBase *, size_t> op_to_idx;
for (size_t i = 0; i < ops.size(); ++i) {
PADDLE_ENFORCE(graph_.HasOp(ops[i]), "Op does not exist in graph");
op_to_idx[ops[i]] = i;
}
PADDLE_ENFORCE(op_to_idx.size() == ops.size(), "Duplicate ops");
std::vector<std::vector<RelationShip>> ret(ops.size());
for (auto &e : ret) {
e.assign(ops.size(), kSame);
}
size_t found_num = ops.size();
size_t total_num = ops.size() * ops.size();
auto visitor = [&](OpHandleBase *op, size_t i) {
auto it = op_to_idx.find(op);
if (it != op_to_idx.end()) {
size_t j = it->second;
if (i != j && ret[i][j] == kSame) {
ret[i][j] = kBefore;
ret[j][i] = kAfter;
found_num += 2;
if (found_num == total_num) {
return false;
}
}
}
return true;
};
for (size_t i = 0; i < ops.size(); ++i) {
auto sub_visitor = [&, i](OpHandleBase *op) { return visitor(op, i); };
if (!graph_.VisitAllPendingOps(ops[i], sub_visitor)) {
break;
}
}
for (size_t i = 0; i < ops.size(); ++i) {
for (size_t j = i + 1; j < ops.size(); ++j) {
if (ret[i][j] != kSame) continue;
ret[i][j] = kNoDeps;
ret[j][i] = kNoDeps;
}
}
return ret;
}
const OpGraphView graph_;
};
/**
* Find the nearest downstream computation op handle. If the op is a
* computation op, just return itself.
*/
static ComputationOpHandle *FindNextComputationOpHandleOrReturnItself(
OpHandleBase *op, size_t scope_idx) {
std::queue<OpHandleBase *> q;
std::unordered_set<OpHandleBase *> visited;
q.push(op);
do {
auto *op = q.front();
q.pop();
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op != nullptr && compute_op->GetScopeIdx() == scope_idx) {
return compute_op;
}
for (auto *out_var : op->Outputs()) {
for (auto *pending_op : out_var->PendingOps()) {
if (visited.count(pending_op)) continue;
visited.insert(pending_op);
}
}
} while (!queue.empty());
} while (!q.empty());
return nullptr;
}
static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out,
ir::Graph *graph) {
auto it = std::find_if(
in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != in->Outputs().end()) {
out->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
in->AddOutput(dep_var);
out->AddInput(dep_var);
static std::unordered_set<ComputationOpHandle *>
ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx,
const ShrinkDepsOpFunctor &shrink_func,
bool *ok) {
// stage one. Get last op for variable.
std::unordered_set<OpHandleBase *> candidates;
{
if (var->PendingOps().empty() && var->GeneratedOp()) {
// No operator depends on this variable. So the last operator is the op
// who generates this variable.
candidates.emplace(var->GeneratedOp());
} else {
candidates = var->PendingOps();
}
// No pending ops or generated op is nullptr
if (candidates.empty()) {
*ok = false;
return {};
}
}
// stage two. Try to cast them to computation op.
// return (*ok=false) when failed.
//
// The reason why we cannot make any types of op handle to be the last lived
// op is:
// some op handle may operate on many DeviceContext, however, our garbage
// collector can only wait one DeviceContext for now. So currently, we wait
// the nearest compute op.
std::unordered_set<ComputationOpHandle *> computation_op;
{
for (auto *op : candidates) {
auto *compute_op =
FindNextComputationOpHandleOrReturnItself(op, scope_idx);
if (compute_op == nullptr) {
*ok = false;
return {};
}
computation_op.emplace(compute_op);
}
}
// stage three. Try to shrink computation op if they depend on each other.
// Get the smallest set of the most ops.
*ok = true;
return shrink_func(computation_op);
}
static VarDesc *TryGetLatestVarDesc(const std::vector<VarHandle *> &vars) {
VarDesc *var_desc = nullptr;
std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool {
var_desc = var_handle->Node()->Var();
return var_desc != nullptr;
});
return var_desc;
}
std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
auto &cur_ref_cnts = Get<AtomicDeviceReferenceCountMap>(kCurReferenceCount);
auto &gcs = Get<DeviceGarbageCollectorMap>(kGarbageCollector);
// It is not easy to find the right reference counts of varaibles in graph
// Step 1: Find all variables in computation ops
// Step 2: Find all variables in non-computation ops which refers to variables
// in computation ops
std::unordered_set<std::string> names;
std::unordered_map<OpHandleBase *, ReferenceCountOpHandle *>
compute_ref_cnt_map;
auto get_ref_cnts_from_compute_op = [&](
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
std::vector<std::string> var_names_in_op;
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op == nullptr ||
!platform::is_gpu_place(compute_op->GetPlace()))
return var_names_in_op;
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
if (!platform::is_gpu_place(var_handle->place_) ||
boost::get<platform::CUDAPlace>(var_handle->place_) != place)
continue;
auto &ref_cnts = Get<std::vector<ReferenceCountMap>>(kGlobalReferenceCount);
auto &last_live_ops_of_vars =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
PADDLE_ENFORCE(last_live_ops_of_vars.empty() && ref_cnts.empty(),
"Last Live Ops and Reference Counts of vars should be "
"initialized at here.");
VarDesc *var_desc = var_handle->Node()->Var();
auto var_name = var_handle->Node()->Name();
const auto &vars = graph->Get<GraphVars>(kGraphVars);
// This is weird but there is really some variables without var_desc
// in computation_op
if (var_desc == nullptr) {
var_desc = compute_op->Node()->Op()->Block()->FindVar(var_name);
if (var_desc == nullptr) continue;
last_live_ops_of_vars.resize(vars.size());
ref_cnts.resize(vars.size());
ShrinkDepsOpFunctor shrink_func(
ir::FilterByNodeWrapper<OpHandleBase>(*graph));
for (size_t i = 0; i < vars.size(); ++i) {
for (auto &name_var_pair : vars[i]) {
// Whether this variable can be reused or deleted? If not, we do not
// compute reference counts and dependencies.
VarDesc *var_desc = TryGetLatestVarDesc(name_var_pair.second);
if (var_desc == nullptr || var_desc->Persistable()) {
continue;
}
if (var_desc->Persistable()) continue;
auto var_type = var_desc->Proto()->type().type();
if (var_type != proto::VarType::LOD_TENSOR &&
var_type != proto::VarType::SELECTED_ROWS) {
var_type != proto::VarType::SELECTED_ROWS &&
var_type != proto::VarType::LOD_TENSOR_ARRAY) {
// Var type cannot be deleted
continue;
}
// compute op only runs in one device
if (ref_cnts[place.device]->count(var_name))
++(*ref_cnts[place.device])[var_name];
else
(*ref_cnts[place.device])[var_name] = 1;
bool ok;
auto result = ExtractComputationOpFromLastLivedVar(
name_var_pair.second.back(), i, shrink_func, &ok);
names.insert(var_name);
var_names_in_op.push_back(var_name);
}
return var_names_in_op;
};
auto update_ref_cnts_from_non_compute_op = [&](
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
if (dynamic_cast<ComputationOpHandle *>(op) != nullptr) return;
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
auto var_name = var_handle->Node()->Name();
auto var_place = var_handle->place_;
if (!platform::is_gpu_place(var_place)) continue;
auto place = boost::get<platform::CUDAPlace>(var_place);
if (names.count(var_name) == 0) continue;
if (ref_cnts.count(place.device) &&
ref_cnts[place.device]->count(var_name)) {
++(*ref_cnts[place.device])[var_name];
auto *next_compute_op = FindNextComputationOpHandle(var_handle);
if (next_compute_op != nullptr) {
if (compute_ref_cnt_map.count(next_compute_op)) {
compute_ref_cnt_map[next_compute_op]->AddVar(var_name);
VLOG(5) << "Add reference count of " << var_name << " to Operator "
<< next_compute_op->Name();
} else {
// Create new reference_count_op_handle
ir::Node *ref_cnt_node = graph->CreateEmptyNode(
"reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[next_compute_op] = ref_cnt_handle;
}
}
if (ok) {
auto &var_name = name_var_pair.first;
PADDLE_ENFORCE(!result.empty(), "Last living ops of %s cannot be empty",
var_name);
ref_cnts[i].emplace(var_name, result.size());
last_live_ops_of_vars[i].emplace(var_name, std::move(result));
}
}
};
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto &op : all_ops) {
auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs());
auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs());
if (in_var_names.empty() && out_var_names.empty()) continue;
in_var_names.insert(in_var_names.end(), out_var_names.begin(),
out_var_names.end());
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
ir::Node *ref_cnt_node =
graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, compute_op->GetScope(), place, in_var_names,
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[compute_op] = ref_cnt_handle;
}
for (auto &op : all_ops) {
update_ref_cnts_from_non_compute_op(op, op->Inputs());
update_ref_cnts_from_non_compute_op(op, op->Outputs());
}
std::vector<OpHandleBase *> new_all_ops;
new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size());
for (auto &op : all_ops) {
new_all_ops.emplace_back(std::move(op));
auto it = compute_ref_cnt_map.find(new_all_ops.back());
if (it != compute_ref_cnt_map.end()) {
// Add LeafNode to ReferenceCountOpHandle
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
it->second->AddOutput(dummy_leaf);
new_all_ops.emplace_back(std::move(it->second));
}
}
all_ops.swap(new_all_ops);
return graph;
}
......@@ -205,5 +258,4 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
REGISTER_PASS(reference_count_pass,
paddle::framework::details::ReferenceCountPass)
.RequirePassAttr(paddle::framework::details::kGlobalReferenceCount)
.RequirePassAttr(paddle::framework::details::kCurReferenceCount)
.RequirePassAttr(paddle::framework::details::kGarbageCollector);
.RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars);
paddle/fluid/framework/details/reference_count_pass.h
浏览文件 @ 547c1bd1
......@@ -14,7 +14,6 @@
#pragma once
#include "paddle/fluid/framework/details/reference_count_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
......@@ -22,10 +21,6 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kGlobalReferenceCount[] = "reference_count";
constexpr char kCurReferenceCount[] = "current_reference_count";
constexpr char kGarbageCollector[] = "garbage_collector";
class ReferenceCountPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
......
paddle/fluid/framework/details/reference_count_pass_helper.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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/details/reference_count_pass_helper.h"
namespace paddle {
namespace framework {
namespace details {} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reference_count_pass_helper.h 0 → 100644
浏览文件 @ 547c1bd1
// 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 <atomic>
#include <map>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
namespace details {
class ComputationOpHandle;
using ReferenceCountMap = std::unordered_map<std::string, size_t>;
using AtomicReferenceCountMap =
std::unordered_map<std::string, std::atomic<size_t>>;
using GarbageCollectorMap =
std::map<platform::Place, std::unique_ptr<GarbageCollector>>;
const char kGlobalReferenceCount[] = "global_reference_count";
const char kRuntimeReferenceCount[] = "runtime_reference_count";
const char kGarbageCollector[] = "garbage_collector";
const char kAllPlaces[] = "all_places";
using LastLiveOpsOfVars =
std::unordered_map<std::string, std::unordered_set<ComputationOpHandle*>>;
const char kLastLiveOpsOfVars[] = "last_live_ops_of_var";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
浏览文件 @ 547c1bd1
......@@ -18,9 +18,6 @@
#include <vector>
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/platform/profiler.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/details/reference_count_op_handle.h"
#endif
namespace paddle {
namespace framework {
......@@ -69,27 +66,12 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
drop_scope_counter_ += 1;
#ifdef PADDLE_WITH_CUDA
const std::string gc_name = "garbage_collector";
DeviceGarbageCollectorMap *gc =
Graph().Has(gc_name) ? &(Graph().Get<DeviceGarbageCollectorMap>(gc_name))
: nullptr;
#endif
if (!fetch_tensors.empty() ||
drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
drop_scope_counter_ = 0;
// Wait All computational streams
for (auto p : places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
#ifdef PADDLE_WITH_CUDA
if (gc != nullptr && platform::is_gpu_place(p)) {
auto gpu_place = boost::get<platform::CUDAPlace>(p);
auto &gc_at_place = gc->at(gpu_place.device);
gc_at_place->Wait();
gc_at_place->Reset();
}
#endif
}
for (auto &scope : local_scopes_) {
auto &local_scope =
......
paddle/fluid/framework/dlpack_tensor.cc
浏览文件 @ 547c1bd1
......@@ -13,7 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/dlpack_tensor.h"
#include "paddle/fluid/framework/data_type.h"
namespace paddle {
namespace framework {
......@@ -36,26 +36,23 @@ static ::DLDataType GetDLDataTypeCode() {
return dtype;
}
static DLDataType GetDLDataTypeFromTypeIndex(const std::type_index &type) {
#define REG_DL_DATA_TYPE(type) \
{ std::type_index(typeid(type)), GetDLDataTypeCode<type>() }
static const std::unordered_map<std::type_index, ::DLDataType>
type_to_dtype_map({
REG_DL_DATA_TYPE(platform::float16), // NOLINT
REG_DL_DATA_TYPE(float), // NOLINT
REG_DL_DATA_TYPE(double), // NOLINT
REG_DL_DATA_TYPE(int), // NOLINT
REG_DL_DATA_TYPE(int64_t), // NOLINT
REG_DL_DATA_TYPE(bool), // NOLINT
REG_DL_DATA_TYPE(size_t), // NOLINT
REG_DL_DATA_TYPE(int16_t), // NOLINT
REG_DL_DATA_TYPE(uint8_t), // NOLINT
REG_DL_DATA_TYPE(int8_t) // NOLINT
});
static std::unordered_map<int, ::DLDataType> CreateDLDataTypeMap() {
static std::unordered_map<int, ::DLDataType> result;
#define REG_DL_DATA_TYPE(cpp_type, proto_type) \
result[static_cast<int>(proto_type)] = GetDLDataTypeCode<cpp_type>()
_ForEachDataType_(REG_DL_DATA_TYPE);
#undef REG_DL_DATA_TYPE
return result;
}
static DLDataType GetDLDataTypeFromTypeIndex(proto::VarType::Type type) {
static auto type_to_dtype_map = CreateDLDataTypeMap();
static auto type_to_dtype_map_end_it = type_to_dtype_map.end();
auto it = type_to_dtype_map.find(type);
PADDLE_ENFORCE(it != type_to_dtype_map_end_it, "Unsupported data type %s",
type.name());
auto it = type_to_dtype_map.find(static_cast<int>(type));
PADDLE_ENFORCE(it != type_to_dtype_map_end_it, "Unsupported data type %d",
type);
return it->second;
#undef REG_DL_DATA_TYPE
}
......
paddle/fluid/framework/dlpack_tensor_test.cc
浏览文件 @ 547c1bd1
......@@ -91,23 +91,11 @@ void TestMainLoop() {
}
}
}
TEST(dlpack, test_all) {
#define TestCallback(cpp_type, proto_type) TestMainLoop<cpp_type>()
#define PADDLE_DLPACK_TEST(type) \
TEST(dlpack, test_##type) { TestMainLoop<type>(); }
using float16 = platform::float16;
PADDLE_DLPACK_TEST(float16);
PADDLE_DLPACK_TEST(float);
PADDLE_DLPACK_TEST(double);
PADDLE_DLPACK_TEST(int);
PADDLE_DLPACK_TEST(int64_t);
PADDLE_DLPACK_TEST(bool);
PADDLE_DLPACK_TEST(size_t);
PADDLE_DLPACK_TEST(int16_t);
PADDLE_DLPACK_TEST(uint8_t);
PADDLE_DLPACK_TEST(int8_t);
#undef PADDLE_DLPACK_TEST
_ForEachDataType_(TestCallback);
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/executor.cc
浏览文件 @ 547c1bd1
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/executor.h"
#include <deque>
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
......@@ -41,11 +42,43 @@ namespace {
int kProgramId = -1;
} // namespace
static std::unordered_map<std::string, size_t> GetNonPersistableReferenceCounts(
const BlockDesc& block, const std::vector<std::string>& skip_var_list) {
std::unordered_map<std::string, size_t> ref_cnts;
std::unordered_set<std::string> skip_vars(skip_var_list.begin(),
skip_var_list.end());
auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
if (skip_vars.count(name)) continue;
auto* var_desc = block.FindVar(name);
if (var_desc == nullptr || var_desc->Persistable()) continue;
auto type = var_desc->Proto()->type().type();
if (type != proto::VarType::LOD_TENSOR &&
type != proto::VarType::SELECTED_ROWS &&
type != proto::VarType::LOD_TENSOR_ARRAY) {
continue;
}
++ref_cnts[name];
}
}
};
for (auto op_desc : block.AllOps()) {
update_ref_cnts(op_desc, op_desc->Inputs());
update_ref_cnts(op_desc, op_desc->Outputs());
}
return ref_cnts;
}
ExecutorPrepareContext::ExecutorPrepareContext(
const framework::ProgramDesc& prog, size_t block_id)
const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars)
: prog_(prog), block_id_(block_id) {
if (GetEagerDeletionThreshold() >= 0) {
ref_cnts_ = GetNonPersistableReferenceCount<int>(prog_, block_id_);
global_ref_cnts_ = GetNonPersistableReferenceCounts(prog.Block(block_id),
skip_ref_cnt_vars);
}
}
......@@ -53,28 +86,40 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
VLOG(5) << "destroy ExecutorPrepareContext";
}
template <typename RefCntMap>
static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
GarbageCollector<Tensor>* gc,
RefCntMap* ref_cnts) {
std::unordered_set<Tensor*> erase_tensors;
static void DeleteUnusedTensors(
const Scope& scope, const OperatorBase* op, GarbageCollector* gc,
std::unordered_map<std::string, size_t>* ref_cnts) {
std::deque<std::shared_ptr<memory::Allocation>> garbages;
auto handler = [&](const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
auto it = ref_cnts->find(name);
if (it == ref_cnts->end()) continue;
if ((it->second)-- == 1) {
auto* var = scope.FindVar(name);
if (var != nullptr) {
VLOG(10) << "Erase tensor \'" << name << "\'";
if (var->IsType<LoDTensor>()) {
erase_tensors.insert(var->GetMutable<LoDTensor>());
} else if (var->IsType<SelectedRows>()) {
erase_tensors.insert(
var->GetMutable<SelectedRows>()->mutable_value());
}
if (--(it->second) != 0) {
continue;
}
auto* var = scope.FindVar(name);
if (var != nullptr) {
continue;
}
VLOG(2) << "Erase variable " << name;
if (var->IsType<LoDTensor>()) {
garbages.emplace_back(
var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
garbages.emplace_back(var->GetMutable<SelectedRows>()
->mutable_value()
->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
auto* lod_tensor_arr = var->GetMutable<LoDTensorArray>();
for (auto& t : *lod_tensor_arr) {
garbages.emplace_back(t.MoveMemoryHolder());
}
} else {
PADDLE_THROW("Type %s of %s is not supported eager deletion",
var->Type().name(), name);
}
}
}
......@@ -83,8 +128,8 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
handler(op->Inputs());
handler(op->Outputs());
if (!erase_tensors.empty()) {
gc->Add(erase_tensors);
if (!garbages.empty()) {
gc->Add(std::move(garbages));
}
}
......@@ -112,9 +157,9 @@ void Executor::Close() {
#ifdef PADDLE_WITH_DISTRIBUTE
// TODO(typhoonzero): complete message will need to use real trainer_id,
// except 0.
::paddle::operators::distributed::RPCClient::GetInstance<
::paddle::operators::distributed::GRPCClient>(0)
->SendComplete();
auto client =
paddle::operators::distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
client->SendComplete();
#endif
}
......@@ -325,9 +370,10 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
}
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
const ProgramDesc& program, int block_id) {
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars) {
std::unique_ptr<ExecutorPrepareContext> ctx(
new ExecutorPrepareContext(program, block_id));
new ExecutorPrepareContext(program, block_id, skip_ref_cnt_vars));
PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
auto& block = program.Block(block_id);
for (auto& op_desc : block.AllOps()) {
......@@ -338,16 +384,28 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
}
std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids) {
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars) {
PADDLE_ENFORCE(
skip_ref_cnt_vars.empty() || skip_ref_cnt_vars.size() == block_ids.size(),
"skip_ref_cnt_vars should be either empty or equals to block number %d",
block_ids.size());
std::vector<std::shared_ptr<ExecutorPrepareContext>> result;
size_t idx = 0;
for (auto& bid : block_ids) {
auto* ctx = new ExecutorPrepareContext(program, bid);
ExecutorPrepareContext* ctx;
if (skip_ref_cnt_vars.empty()) {
ctx = new ExecutorPrepareContext(program, bid);
} else {
ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx]);
}
PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
auto& block = program.Block(bid);
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
result.push_back(std::shared_ptr<ExecutorPrepareContext>(ctx));
++idx;
}
return result;
}
......@@ -365,22 +423,23 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
int64_t max_memory_size = GetEagerDeletionThreshold();
std::unique_ptr<GarbageCollector<Tensor>> gc;
// WhileOp would set keep_kids to true,
// because WhileGradOp needs the scopes created in WhileOp.
// Perhaps, we should not perform eager deletion in WhileOp
// The scopes and variables created by WhileOp would be deleted
// in WhileGradOp.
std::unique_ptr<GarbageCollector> gc;
// skip while_op and while_grad_op temporarily
if (max_memory_size >= 0 && !keep_kids) {
ctx->ResetReferenceCount();
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place_)) {
gc.reset(new DefaultStreamGarbageCollector<Tensor>(
boost::get<platform::CUDAPlace>(place_), max_memory_size));
} else {
if (IsFastEagerDeletionModeEnabled()) {
gc.reset(new UnsafeFastGPUGarbageCollector(
boost::get<platform::CUDAPlace>(place_), max_memory_size));
} else {
gc.reset(new DefaultStreamGarbageCollector(
boost::get<platform::CUDAPlace>(place_), max_memory_size));
}
} else if (platform::is_cpu_place(place_)) {
#endif
gc.reset(new CPUGarbageCollector<Tensor>(
boost::get<platform::CPUPlace>(place_), max_memory_size));
gc.reset(new CPUGarbageCollector(boost::get<platform::CPUPlace>(place_),
max_memory_size));
#ifdef PADDLE_WITH_CUDA
}
#endif
......@@ -389,17 +448,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
for (auto& op : ctx->ops_) {
op->Run(*local_scope, place_);
if (gc != nullptr) {
if (gc) {
DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
&(ctx->cur_ref_cnts_));
&(ctx->runtime_ref_cnts_));
}
}
if (gc != nullptr) {
gc->Wait();
} else {
platform::DeviceContextPool::Instance().Get(place_)->Wait();
}
platform::DeviceContextPool::Instance().Get(place_)->Wait();
if (local_scope != scope) {
scope->DeleteScope(local_scope);
......
paddle/fluid/framework/executor.h
浏览文件 @ 547c1bd1
......@@ -27,52 +27,21 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <typename T>
std::unordered_map<std::string, T> GetNonPersistableReferenceCount(
const ProgramDesc& prog, size_t block_id) {
auto& block = prog.Block(block_id);
std::unordered_map<std::string, T> ref_cnts;
auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
auto* var_desc = block.FindVar(name);
if (var_desc == nullptr || var_desc->Persistable()) continue;
auto type = var_desc->Proto()->type().type();
if (type != proto::VarType::LOD_TENSOR &&
type != proto::VarType::SELECTED_ROWS) {
continue;
}
auto it = ref_cnts.find(name);
if (it != ref_cnts.end()) {
++it->second;
} else {
ref_cnts[name] = 1;
}
}
}
};
for (auto op_desc : block.AllOps()) {
update_ref_cnts(op_desc, op_desc->Inputs());
update_ref_cnts(op_desc, op_desc->Outputs());
}
return ref_cnts;
}
struct ExecutorPrepareContext {
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
~ExecutorPrepareContext();
void ResetReferenceCount() { cur_ref_cnts_ = ref_cnts_; }
void ResetReferenceCount() { runtime_ref_cnts_ = global_ref_cnts_; }
const framework::ProgramDesc& prog_;
size_t block_id_;
std::vector<std::unique_ptr<OperatorBase>> ops_;
std::unordered_map<std::string, int> ref_cnts_;
std::unordered_map<std::string, int> cur_ref_cnts_;
std::unordered_map<std::string, size_t> global_ref_cnts_;
std::unordered_map<std::string, size_t> runtime_ref_cnts_;
};
class Executor {
......@@ -108,10 +77,14 @@ class Executor {
const std::string& fetch_holder_name = "fetch");
static std::unique_ptr<ExecutorPrepareContext> Prepare(
const ProgramDesc& program, int block_id);
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
static std::vector<std::shared_ptr<ExecutorPrepareContext>> Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids);
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars =
std::vector<std::vector<std::string>>());
void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id);
......
paddle/fluid/framework/executor_thread_worker.cc
浏览文件 @ 547c1bd1
......@@ -26,6 +26,7 @@ limitations under the License. */
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
namespace paddle {
......@@ -138,42 +139,24 @@ void print_lod_tensor(std::string var_name, const LoDTensor& lod_tensor) {
std::cout << sstream.str() << std::endl;
}
void print_fetch_var(Scope* scope, std::string var_name) {
const LoDTensor& tensor = scope->FindVar(var_name)->Get<LoDTensor>();
if (std::type_index(tensor.type()) ==
std::type_index(typeid(platform::float16))) {
print_lod_tensor<platform::float16>(var_name, tensor);
} else if (std::type_index(tensor.type()) == std::type_index(typeid(float))) {
print_lod_tensor<float>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(double))) {
print_lod_tensor<double>(var_name, tensor);
} else if (std::type_index(tensor.type()) == std::type_index(typeid(int))) {
print_lod_tensor<int>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(int64_t))) {
print_lod_tensor<int64_t>(var_name, tensor);
} else if (std::type_index(tensor.type()) == std::type_index(typeid(bool))) {
print_lod_tensor<bool>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(uint8_t))) {
print_lod_tensor<uint8_t>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(int16_t))) {
print_lod_tensor<int16_t>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(int8_t))) {
print_lod_tensor<int8_t>(var_name, tensor);
} else {
VLOG(1) << "print_fetch_var: unrecognized data type:"
<< tensor.type().name();
}
static void print_fetch_var(Scope* scope, const std::string& var_name) {
auto& tensor = scope->FindVar(var_name)->Get<LoDTensor>();
return;
#define PrintLoDTensorCallback(cpp_type, proto_type) \
do { \
if (tensor.type() == proto_type) { \
print_lod_tensor<cpp_type>(var_name, tensor); \
return; \
} \
} while (0)
_ForEachDataType_(PrintLoDTensorCallback);
VLOG(1) << "print_fetch_var: unrecognized data type:" << tensor.type();
}
void ExecutorThreadWorker::TrainFiles() {
platform::SetNumThreads(1);
// todo: configurable
SetDevice();
......
paddle/fluid/framework/garbage_collector.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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 <algorithm>
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
GarbageCollector::GarbageCollector(const platform::Place &place,
size_t max_memory_size)
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
garbages_.reset(new GarbageQueue());
dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}
CPUGarbageCollector::CPUGarbageCollector(const platform::CPUPlace &place,
size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void CPUGarbageCollector::ClearCallback(const std::function<void()> &callback) {
callback();
}
#ifdef PADDLE_WITH_CUDA
UnsafeFastGPUGarbageCollector::UnsafeFastGPUGarbageCollector(
const platform::CUDAPlace &place, size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void UnsafeFastGPUGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
callback();
}
DefaultStreamGarbageCollector::DefaultStreamGarbageCollector(
const platform::CUDAPlace &place, size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void DefaultStreamGarbageCollector::Wait() const {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->WaitStreamCallback();
}
void DefaultStreamGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->AddStreamCallback(callback);
}
StreamGarbageCollector::StreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {
platform::CUDADeviceGuard guard(place.device);
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
callback_manager_.reset(new platform::StreamCallbackManager(stream_));
}
StreamGarbageCollector::~StreamGarbageCollector() {
auto place = boost::get<platform::CUDAPlace>(this->dev_ctx_->GetPlace());
platform::CUDADeviceGuard guard(place.device);
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}
cudaStream_t StreamGarbageCollector::stream() const { return stream_; }
void StreamGarbageCollector::Wait() const { callback_manager_->Wait(); }
void StreamGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
callback_manager_->AddCallback(callback);
}
#endif
} // namespace framework
} // namespace paddle
paddle/fluid/framework/garbage_collector.h
浏览文件 @ 547c1bd1
......@@ -14,7 +14,6 @@
#pragma once
#include <algorithm>
#include <deque>
#include <functional>
#include <memory>
......@@ -24,134 +23,74 @@
namespace paddle {
namespace framework {
// T should have memory_size() and clear() method
template <typename T>
class GarbageCollector {
public:
GarbageCollector(const platform::Place &place, size_t max_memory_size)
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
garbages_.reset(new std::deque<T *>());
dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}
using GarbageQueue = std::deque<std::shared_ptr<memory::Allocation>>;
virtual ~GarbageCollector() {}
GarbageCollector(const platform::Place &place, size_t max_memory_size);
void Reset() {
std::lock_guard<std::mutex> guard(mutex_);
garbages_.reset(new std::deque<T *>());
cur_memory_size_ = 0;
}
virtual ~GarbageCollector() = default;
virtual void Wait() const {}
template <typename Container>
void Add(const Container &objs) {
Add(objs, []() {});
}
void Add(Container &&objs);
template <typename Container, typename Callback>
void Add(const Container &objs, Callback &&callback) {
std::shared_ptr<std::deque<T *>> clear_deque;
{
std::lock_guard<std::mutex> guard(mutex_);
for (auto *obj : objs) {
garbages_->push_back(obj);
cur_memory_size_ += obj->memory_size();
}
if (cur_memory_size_ >= max_memory_size_) {
cur_memory_size_ = 0;
clear_deque = garbages_;
garbages_.reset(new std::deque<T *>());
}
}
if (clear_deque != nullptr) {
callback();
ClearCallback([=]() {
for (auto *obj : *clear_deque) obj->clear();
});
}
}
virtual void Wait() const {}
void Add(Container &&objs, Callback &&callback);
protected:
virtual void ClearCallback(const std::function<void()> &callback) = 0;
platform::DeviceContext *dev_ctx_;
std::shared_ptr<std::deque<T *>> garbages_;
std::unique_ptr<GarbageQueue> garbages_;
mutable std::mutex mutex_;
const size_t max_memory_size_;
size_t cur_memory_size_ = 0;
size_t cur_memory_size_{0};
};
template <typename T>
class CPUGarbageCollector : public GarbageCollector<T> {
class CPUGarbageCollector : public GarbageCollector {
public:
CPUGarbageCollector(const platform::CPUPlace &place, size_t max_memory_size)
: GarbageCollector<T>(place, max_memory_size) {}
CPUGarbageCollector(const platform::CPUPlace &place, size_t max_memory_size);
protected:
void ClearCallback(const std::function<void()> &callback) override {
callback();
}
void ClearCallback(const std::function<void()> &callback) override;
};
#ifdef PADDLE_WITH_CUDA
template <typename T>
class DefaultStreamGarbageCollector : public GarbageCollector<T> {
class UnsafeFastGPUGarbageCollector : public GarbageCollector {
public:
DefaultStreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector<T>(place, max_memory_size) {}
UnsafeFastGPUGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size);
cudaStream_t stream() const {
return static_cast<const platform::CUDADeviceContext *>(this->dev_ctx_)
->stream();
}
protected:
void ClearCallback(const std::function<void()> &callback) override;
};
void Wait() const override {
this->dev_ctx_->Wait();
static_cast<const platform::CUDADeviceContext *>(this->dev_ctx_)
->WaitStreamCallback();
}
class DefaultStreamGarbageCollector : public GarbageCollector {
public:
DefaultStreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size);
void Wait() const override;
protected:
void ClearCallback(const std::function<void()> &callback) override {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->AddStreamCallback(callback);
}
void ClearCallback(const std::function<void()> &callback) override;
};
template <typename T>
class StreamGarbageCollector : public GarbageCollector<T> {
class StreamGarbageCollector : public GarbageCollector {
public:
StreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector<T>(place, max_memory_size) {
PADDLE_ENFORCE(cudaSetDevice(place.device));
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
callback_manager_.reset(new platform::StreamCallbackManager(stream_));
}
size_t max_memory_size);
~StreamGarbageCollector() {
auto place = boost::get<platform::CUDAPlace>(this->dev_ctx_->GetPlace());
PADDLE_ENFORCE(cudaSetDevice(place.device));
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}
~StreamGarbageCollector();
void Wait() const override {
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
std::lock_guard<std::mutex> guard(this->mutex_);
callback_manager_->Wait();
}
void Wait() const override;
cudaStream_t stream() const { return stream_; }
cudaStream_t stream() const;
protected:
void ClearCallback(const std::function<void()> &callback) override {
std::lock_guard<std::mutex> guard(this->mutex_);
callback_manager_->AddCallback(callback);
}
void ClearCallback(const std::function<void()> &callback) override;
private:
cudaStream_t stream_;
......@@ -159,5 +98,33 @@ class StreamGarbageCollector : public GarbageCollector<T> {
};
#endif
template <typename Container>
void GarbageCollector::Add(Container &&objs) {
Add(std::forward<Container>(objs), []() {});
}
template <typename Container, typename Callback>
void GarbageCollector::Add(Container &&objs, Callback &&callback) {
GarbageQueue *garbage_queue = nullptr;
{
std::lock_guard<std::mutex> guard(mutex_);
for (auto &obj : objs) {
if (!obj) continue;
cur_memory_size_ += obj->size();
garbages_->push_back(std::move(obj));
}
if (cur_memory_size_ >= max_memory_size_) {
cur_memory_size_ = 0;
garbage_queue = garbages_.release();
garbages_.reset(new GarbageQueue());
}
}
if (garbage_queue) {
callback();
ClearCallback([garbage_queue]() { delete garbage_queue; });
}
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 547c1bd1
......@@ -42,6 +42,8 @@ pass_library(multi_batch_merge_pass base)
pass_library(conv_bn_fuse_pass inference)
pass_library(seqconv_eltadd_relu_fuse_pass inference)
pass_library(is_test_pass base)
pass_library(conv_elementwise_add_act_fuse_pass inference)
pass_library(conv_elementwise_add2_act_fuse_pass inference)
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base)
pass_library(depthwise_conv_mkldnn_pass base)
......
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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 <string>
#include "paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h"
namespace paddle {
namespace framework {
namespace ir {
#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern);
#define GET_NODES \
GET_IR_NODE(conv_op); \
GET_IR_NODE(conv_out); \
GET_IR_NODE(conv_filter); \
GET_IR_NODE(elementwise_add_op); \
GET_IR_NODE(elementwise_add_in_y); \
GET_IR_NODE(elementwise_add_out); \
GET_IR_NODE(elementwise_add_op_1); \
GET_IR_NODE(elementwise_add_in_y_1); \
GET_IR_NODE(elementwise_add_out_1); \
GET_IR_NODE(act_op); \
GET_IR_NODE(act_out);
// Inherient the basic infomation from `base_desc`, and modify some fields.
framework::proto::OpDesc PrepareOpDesc(
const framework::proto::OpDesc& base_desc, const std::string& bias,
const std::string& bias1, const std::string& activation,
const std::string& output) {
auto proto = base_desc;
framework::OpDesc desc(proto, nullptr);
desc.SetInput("Bias", {bias});
desc.SetInput("ResidualData", {bias1});
desc.SetAttr("activation", activation);
desc.SetOutput("Output", {output});
desc.SetAttr("is_test", true);
desc.SetAttr("use_cudnn", false);
return *desc.Proto();
}
std::unique_ptr<ir::Graph> ConvElementwiseAddActFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
const std::string pattern_name = "conv_elementwise_add_act_fuse";
FusePassBase::Init(pattern_name, graph.get());
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()->NewNode("x")->AsInput()->assert_is_op_input(
"conv2d", "Input");
patterns::ConvElementwiseaddAct pattern(gpd.mutable_pattern(), pattern_name);
pattern(x);
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_NODES;
auto base_op_desc = *conv_op->Op()->Proto();
std::string bias_name = elementwise_add_in_y->Name();
std::string bias1_name = elementwise_add_in_y_1->Name();
std::string act_op_type = act_op->Op()->Type();
std::string act_op_out = act_out->Name();
auto new_op_proto = PrepareOpDesc(base_op_desc, bias_name, bias1_name,
act_op_type, act_op_out);
framework::OpDesc new_op_desc(new_op_proto, nullptr);
// Create a new node for the fused op.
auto new_conv_op = graph->CreateOpNode(&new_op_desc);
// Link inputs and outputs.
PADDLE_ENFORCE(subgraph.count(x));
auto* conv_in_node = subgraph.at(x);
IR_NODE_LINK_TO(conv_in_node, new_conv_op); // Input
IR_NODE_LINK_TO(conv_filter, new_conv_op); // Filter
IR_NODE_LINK_TO(elementwise_add_in_y, new_conv_op); // Bias
IR_NODE_LINK_TO(elementwise_add_in_y_1, new_conv_op); // ResidualData
IR_NODE_LINK_TO(new_conv_op, act_out); // Output
// Delete the unneeded nodes.
GraphSafeRemoveNodes(graph.get(),
{conv_op, elementwise_add_op, elementwise_add_op_1,
elementwise_add_out});
};
gpd(graph.get(), handler);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_elementwise_add2_act_fuse_pass,
paddle::framework::ir::ConvElementwiseAdd2ActFusePass);
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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/ir/conv_elementwise_add2_act_fuse_pass.h"
#include <string>
namespace paddle {
namespace framework {
namespace ir {
#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern);
#define GET_NODES \
GET_IR_NODE(conv_op); \
GET_IR_NODE(conv_out); \
GET_IR_NODE(conv_filter); \
GET_IR_NODE(elementwise_add_op); \
GET_IR_NODE(elementwise_add_in_y); \
GET_IR_NODE(elementwise_add_out); \
GET_IR_NODE(elementwise_add_op_1); \
GET_IR_NODE(elementwise_add_in_y_1); \
GET_IR_NODE(elementwise_add_out_1); \
GET_IR_NODE(act_op); \
GET_IR_NODE(act_out);
// Inherient the basic infomation from `base_desc`, and modify some fields.
framework::proto::OpDesc PrepareOpDesc(
const framework::proto::OpDesc& base_desc, const std::string& bias,
const std::string& bias1, const std::string& activation,
const std::string& output) {
auto proto = base_desc;
framework::OpDesc desc(proto, nullptr);
desc.SetInput("Bias", {bias});
desc.SetInput("ResidualData", {bias1});
desc.SetAttr("activation", activation);
desc.SetOutput("Output", {output});
desc.SetAttr("is_test", true);
return *desc.Proto();
}
std::unique_ptr<ir::Graph> ConvElementwiseAdd2ActFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
const std::string pattern_name = "conv_elementwise_add_act_fuse";
FusePassBase::Init(pattern_name, graph.get());
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()->NewNode("x")->AsInput()->assert_is_op_input(
"conv2d", "Input");
patterns::ConvElementwiseadd2Act pattern(gpd.mutable_pattern(), pattern_name);
pattern(x);
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_NODES;
auto base_op_desc = *conv_op->Op()->Proto();
std::string bias_name = elementwise_add_in_y->Name();
std::string bias1_name = elementwise_add_in_y_1->Name();
std::string act_op_type = act_op->Op()->Type();
std::string act_op_out = act_out->Name();
auto new_op_proto = PrepareOpDesc(base_op_desc, bias_name, bias1_name,
act_op_type, act_op_out);
framework::OpDesc new_op_desc(new_op_proto, nullptr);
// Create a new node for the fused op.
graph->CreateOpNode(&new_op_desc);
// Link inputs and outputs.
PADDLE_ENFORCE(subgraph.count(x));
auto* conv_in_node = subgraph.at(x);
IR_NODE_LINK_TO(conv_in_node, conv_op); // Input
IR_NODE_LINK_TO(conv_filter, conv_op); // Filter
IR_NODE_LINK_TO(conv_op, conv_out); // Output
IR_NODE_LINK_TO(elementwise_add_in_y, conv_op); // Bias
IR_NODE_LINK_TO(elementwise_add_in_y_1, conv_op); // Bias
// Delete the unneeded nodes.
GraphSafeRemoveNodes(graph.get(),
{conv_op, elementwise_add_op, elementwise_add_op_1,
elementwise_add_out});
};
gpd(graph.get(), handler);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_elementwise_add2_act_fuse_pass,
paddle::framework::ir::ConvElementwiseAdd2ActFusePass);
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h 0 → 100644
浏览文件 @ 547c1bd1
// 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/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class ConvElementwiseAdd2ActFusePass : public FusePassBase {
public:
virtual ~ConvElementwiseAdd2ActFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc 0 → 100644
浏览文件 @ 547c1bd1
// 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/ir/conv_elementwise_add_act_fuse_pass.h"
#include <string>
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
namespace paddle {
namespace framework {
namespace ir {
#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern);
#define GET_NODES \
GET_IR_NODE(conv_op); \
GET_IR_NODE(conv_out); \
GET_IR_NODE(conv_filter); \
GET_IR_NODE(elementwise_add_op); \
GET_IR_NODE(elementwise_add_in_y); \
GET_IR_NODE(elementwise_add_out); \
GET_IR_NODE(act_op); \
GET_IR_NODE(act_out);
// Inherient the basic infomation from `base_desc`, and modify some fields.
framework::proto::OpDesc PrepareOpDesc(
const framework::proto::OpDesc& base_desc, const std::string& bias,
const std::string& activation, const std::string& output) {
auto proto = base_desc;
framework::OpDesc desc(proto, nullptr);
desc.SetType("conv2d_fusion");
desc.SetInput("Bias", {bias});
desc.SetInput("ResidualData", {});
desc.SetAttr("activation", activation);
desc.SetOutput("Output", {output});
desc.SetAttr("is_test", true);
desc.SetAttr("use_cudnn", false);
desc.Flush();
return *desc.Proto();
}
std::unique_ptr<ir::Graph> ConvElementwiseAddActFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
const std::string pattern_name = "conv_elementwise_add_act_fuse";
FusePassBase::Init(pattern_name, graph.get());
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()
->NewNode("x")
->assert_is_op_input("conv2d", "Input")
->AsInput();
patterns::ConvElementwiseaddAct pattern(gpd.mutable_pattern(), pattern_name);
pattern(x);
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_NODES;
auto base_op_desc = *conv_op->Op()->Proto();
std::string bias_name = elementwise_add_in_y->Name();
std::string act_op_type = act_op->Op()->Type();
std::string act_op_out = act_out->Name();
auto new_op_proto =
PrepareOpDesc(base_op_desc, bias_name, act_op_type, act_op_out);
framework::OpDesc new_op_desc(new_op_proto, nullptr);
// Create a new node for the fused op.
auto* new_conv_op = graph->CreateOpNode(&new_op_desc);
// Link inputs and outputs.
PADDLE_ENFORCE(subgraph.count(x));
auto* conv_in_node = subgraph.at(x);
IR_NODE_LINK_TO(conv_in_node, new_conv_op); // Input
IR_NODE_LINK_TO(conv_filter, new_conv_op); // Filter
IR_NODE_LINK_TO(elementwise_add_in_y, new_conv_op); // Bias
IR_NODE_LINK_TO(new_conv_op, act_out); // Output
// Delete the unneeded nodes.
GraphSafeRemoveNodes(graph.get(), {conv_op, conv_out, elementwise_add_op,
elementwise_add_out, act_op});
};
gpd(graph.get(), handler);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_elementwise_add_act_fuse_pass,
paddle::framework::ir::ConvElementwiseAddActFusePass);
paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h 0 → 100644
浏览文件 @ 547c1bd1
// 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/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class ConvElementwiseAddActFusePass : public FusePassBase {
public:
virtual ~ConvElementwiseAddActFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph.h
浏览文件 @ 547c1bd1
......@@ -73,14 +73,21 @@ class Graph {
}
bool Has(const std::string &attr_name) const {
return attrs_.find(attr_name) != attrs_.end();
return attrs_.count(attr_name) > 0;
}
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const {
PADDLE_ENFORCE(Has(attr_name), "%s attr not registered for graph.",
attr_name);
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
try {
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
} catch (boost::bad_any_cast &) {
PADDLE_THROW(
"Invalid attribute type of %s error, expected: %s, actual: %s",
attr_name, typeid(AttrType *).name(),
attrs_.at(attr_name).type().name());
}
}
template <typename AttrType>
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ 547c1bd1
......@@ -17,6 +17,7 @@
#include <string>
#include <vector>
#include "graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/graph_traits.h"
......@@ -25,6 +26,7 @@
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/string/pretty_log.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {
namespace framework {
namespace ir {
......@@ -104,7 +106,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
for (auto &node : GraphTraits::DFS(graph)) {
for (const auto &pdnode : pattern_.nodes()) {
if (pdnode->Tell(&node)) {
VLOG(4) << "pdnode " << pdnode->name() << " marked";
VLOG(4) << "Node " << node.Name() << " marked as " << pdnode->name();
pdnodes2nodes_[pdnode.get()].insert(&node);
}
}
......@@ -1099,6 +1101,115 @@ PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) {
return out_var;
}
std::unordered_set<std::string> conv_act_set({"identity", "sigmoid", "relu",
"relu6", "relux", "tanh",
"band_pass"});
PDNode *patterns::ConvElementwiseaddAct::operator()(PDNode *conv_in) {
conv_in->AsInput();
auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
auto conv_out = pattern->NewNode(conv_out_repr())
->assert_is_op_output("conv2d")
->assert_is_op_input("elementwise_add", "X")
->AsIntermediate();
auto conv_filter = pattern->NewNode(conv_filter_repr())
->assert_is_op_input("conv2d", "Filter")
->AsInput();
auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
->assert_is_op("elementwise_add");
auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr())
->assert_is_op_input("elementwise_add", "Y")
->AsInput();
auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr())
->assert_is_op_output("elementwise_add")
->AsIntermediate();
auto act_op = pattern->NewNode(act_op_repr())
->assert_is_op()
->assert_more([&](Node *node) {
auto op_type = node->Name();
return conv_act_set.count(op_type);
});
auto act_out = pattern->NewNode(act_out_repr())
->assert_is_var()
// is activation op's output.
->assert_more([&](Node *node) {
for (auto *in_op : node->inputs) {
if (conv_act_set.count(in_op->Name())) {
return true;
}
}
return false;
})
->AsOutput();
conv_op->LinksFrom({conv_in, conv_filter});
conv_out->LinksFrom({conv_op});
elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y})
.LinksTo({elementwise_add_out});
act_op->LinksFrom({elementwise_add_out}).LinksTo({act_out});
return act_out;
}
PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) {
auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
auto conv_filter = pattern->NewNode(conv_filter_repr())
->assert_is_op_input("conv2d", "Filter")
->AsInput();
auto conv_out = pattern->NewNode(conv_out_repr())
->assert_is_op_output("conv2d")
->assert_is_op_input("elementwise_add", "X")
->AsIntermediate();
auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
->assert_is_op("elementwise_add");
auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr())
->assert_is_op_input("elementwise_add", "Y")
->AsInput();
auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr())
->assert_is_op_output("elementwise_add")
->assert_is_op_input("elementwise_add", "X")
->AsIntermediate();
auto elementwise_add_op_1 = pattern->NewNode(elementwise_add_op_1_repr())
->assert_is_op("elementwise_add");
auto elementwise_add_in_y_1 = pattern->NewNode(elementwise_add_in_y_1_repr())
->assert_is_op_input("elementwise_add", "Y")
->AsInput();
auto elementwise_add_out_1 = pattern->NewNode(elementwise_add_out_1_repr())
->assert_is_op_output("elementwise_add")
->AsIntermediate();
auto act_op = pattern->NewNode(act_op_repr())
->assert_is_op()
->assert_more([&](Node *node) {
auto op_type = node->Name();
return conv_act_set.count(op_type);
});
auto act_out = pattern->NewNode(act_out_repr())
->assert_is_var()
// is activation op's output.
->assert_more([&](Node *node) {
for (auto *in_op : node->inputs) {
if (conv_act_set.count(in_op->Name())) {
return true;
}
}
return false;
})
->AsOutput();
conv_op->LinksFrom({conv_in, conv_filter}).LinksTo({conv_out});
elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y})
.LinksTo({elementwise_add_out});
elementwise_add_op_1->LinksFrom(
{elementwise_add_out, elementwise_add_in_y_1});
act_op->LinksFrom({elementwise_add_out_1}).LinksTo({act_out});
return act_out;
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ 547c1bd1
......@@ -671,6 +671,51 @@ struct ElementwiseAdd : public PatternBase {
PATTERN_DECL_NODE(elementwise_add_y);
PATTERN_DECL_NODE(elementwise_add_out);
};
// Conv + ElementwiseAdd + an activation
// This pattern can futher fuse the conv related ops after the conv+bn fusion.
struct ConvElementwiseaddAct : public PatternBase {
ConvElementwiseaddAct(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "conv_elementwiseadd_act") {}
PDNode* operator()(PDNode* conv_in);
PATTERN_DECL_NODE(conv_op);
PATTERN_DECL_NODE(conv_out);
PATTERN_DECL_NODE(conv_filter);
PATTERN_DECL_NODE(elementwise_add_op);
PATTERN_DECL_NODE(elementwise_add_in_y); // input
PATTERN_DECL_NODE(elementwise_add_out);
PATTERN_DECL_NODE(act_op);
PATTERN_DECL_NODE(act_out);
};
// Conv + ElementwiseAdd + ElementwiseAdd + Activation
struct ConvElementwiseadd2Act : public PatternBase {
ConvElementwiseadd2Act(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope,
"conv_elementwiseadd2_elementwiseadd_act") {}
PDNode* operator()(PDNode* conv_in);
PATTERN_DECL_NODE(conv_op);
PATTERN_DECL_NODE(conv_filter);
PATTERN_DECL_NODE(conv_out);
PATTERN_DECL_NODE(elementwise_add_op);
PATTERN_DECL_NODE(elementwise_add_in_y); // input
PATTERN_DECL_NODE(elementwise_add_out);
PATTERN_DECL_NODE(elementwise_add_op_1);
PATTERN_DECL_NODE(elementwise_add_in_y_1); // input
PATTERN_DECL_NODE(elementwise_add_out_1);
PATTERN_DECL_NODE(act_op);
PATTERN_DECL_NODE(act_out);
};
} // namespace patterns
// Link two ir::Nodes from each other.
......
paddle/fluid/framework/ir/pass.h
浏览文件 @ 547c1bd1
......@@ -51,11 +51,18 @@ class Pass {
AttrType &Get(const std::string &attr_name) const {
PADDLE_ENFORCE(attrs_.find(attr_name) != attrs_.end(),
"%s attr not registered for pass.", attr_name);
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
try {
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
} catch (boost::bad_any_cast &) {
PADDLE_THROW(
"Invalid attribute type of %s error, expected: %s, actual: %s",
attr_name, typeid(AttrType *).name(),
attrs_.at(attr_name).type().name());
}
}
bool Has(const std::string &attr_name) const {
return attrs_.find(attr_name) != attrs_.end();
return attrs_.count(attr_name) > 0;
}
void Erase(const std::string &attr_name) {
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ 547c1bd1
......@@ -70,9 +70,9 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
// only print first ten elements
int64_t size = t.numel() < 10 ? t.numel() : 10;
for (int64_t i = 0; i < size; ++i) {
if (IsType<float>(t.type())) {
if (t.type() == proto::VarType::FP32) {
os << t.data<float>()[i] << " ";
} else if (IsType<int64_t>(t.type())) {
} else if (t.type() == proto::VarType::INT64) {
os << t.data<int64_t>()[i] << " ";
} else {
PADDLE_THROW("LoDTensor data type not in [float, int64_t]");
......@@ -387,7 +387,7 @@ void LoDTensor::MergeLoDTensor(
PADDLE_ENFORCE(!lod_tensors.empty());
framework::DDim new_dim = lod_tensors[0]->dims();
std::type_index new_type = lod_tensors[0]->type();
auto new_type = lod_tensors[0]->type();
framework::DataLayout new_layout = lod_tensors[0]->layout();
LoD new_lod = lod_tensors[0]->lod();
for (size_t i = 1; i < lod_tensors.size(); ++i) {
......
paddle/fluid/framework/ngraph_operator.cc
浏览文件 @ 547c1bd1
......@@ -471,27 +471,23 @@ void NgraphEngine::Run(const Scope& scope, const platform::Place& place) const {
auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()),
"Ensure ngraph tensor layout align with paddle tensor");
if (tensor_pd->type().hash_code() ==
typeid(float).hash_code()) { // NOLINT
if (tensor_pd->type() == proto::VarType::FP32) {
const float* arr = tensor_pd->data<float>();
ti = backend_->create_tensor(ngraph::element::f32, sp,
const_cast<float*>(arr));
} else if (tensor_pd->type().hash_code() ==
typeid(int).hash_code()) { // NOLINT
} else if (tensor_pd->type() == proto::VarType::INT32) {
const int* arr = tensor_pd->data<int>();
ti = backend_->create_tensor(ngraph::element::i32, sp,
const_cast<int*>(arr));
} else if (tensor_pd->type().hash_code() == typeid(int64_t).hash_code()) {
} else if (tensor_pd->type() == proto::VarType::INT64) {
const int64_t* arr = tensor_pd->data<int64_t>();
ti = backend_->create_tensor(ngraph::element::i64, sp,
const_cast<int64_t*>(arr));
} else if (tensor_pd->type().hash_code() ==
typeid(double).hash_code()) { // NOLINT
} else if (tensor_pd->type() == proto::VarType::FP64) {
const double* arr = tensor_pd->data<double>();
ti = backend_->create_tensor(ngraph::element::f64, sp,
const_cast<double*>(arr));
} else if (tensor_pd->type().hash_code() ==
typeid(bool).hash_code()) { // NOLINT
} else if (tensor_pd->type() == proto::VarType::BOOL) {
const bool* arr = tensor_pd->data<bool>();
ti = backend_->create_tensor(ngraph::element::boolean, sp,
const_cast<bool*>(arr));
......
paddle/fluid/framework/op_kernel_type_test.cc
浏览文件 @ 547c1bd1
......@@ -34,7 +34,8 @@ TEST(OpKernelType, ToString) {
OpKernelType op_kernel_type2(DataType::FP16, CUDAPlace(0), DataLayout::kNCHW,
LibraryType::kCUDNN);
ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type2),
"data_type[float16]:data_layout[NCHW]:place[CUDAPlace(0)]:library_"
"data_type[::paddle::platform::float16]:data_layout[NCHW]:place["
"CUDAPlace(0)]:library_"
"type[CUDNN]");
}
......
paddle/fluid/framework/operator.cc
浏览文件 @ 547c1bd1
......@@ -43,10 +43,9 @@ std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority = {
proto::VarType::Type GetDataTypeOfVar(const Variable* var) {
if (var->IsType<framework::LoDTensor>()) {
return framework::ToDataType(var->Get<framework::LoDTensor>().type());
return var->Get<framework::LoDTensor>().type();
} else if (var->IsType<framework::SelectedRows>()) {
return framework::ToDataType(
var->Get<framework::SelectedRows>().value().type());
return var->Get<framework::SelectedRows>().value().type();
} else {
PADDLE_THROW("Var should be LoDTensor or SelectedRows");
}
......@@ -93,13 +92,13 @@ static std::string GetDtype(const Scope& scope, const std::string& name) {
if (UNLIKELY(!tensor.IsInitialized())) {
return "";
}
return DataTypeToString(ToDataType(tensor.type()));
return DataTypeToString(tensor.type());
} else if (var->IsType<SelectedRows>()) {
auto tensor = var->Get<SelectedRows>().value();
if (UNLIKELY(!tensor.IsInitialized())) {
return "uninited";
} else {
return DataTypeToString(ToDataType(tensor.type()));
return DataTypeToString(tensor.type());
}
} else {
return "";
......@@ -686,7 +685,8 @@ static void CheckTensorNANOrInf(const std::string& name,
if (tensor.memory_size() == 0) {
return;
}
if (!IsType<float>(tensor.type()) && !IsType<double>(tensor.type())) {
if (tensor.type() != proto::VarType::FP32 &&
tensor.type() != proto::VarType::FP64) {
return;
}
PADDLE_ENFORCE(!framework::TensorContainsInf(tensor),
......@@ -879,7 +879,9 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
t = &(var->Get<SelectedRows>().value());
}
if (t != nullptr) {
int tmp = static_cast<int>(ToDataType(t->type()));
PADDLE_ENFORCE(t->IsInitialized(), "Input %s is not initialized: %s",
ipt_name, DebugString());
int tmp = static_cast<int>(t->type());
PADDLE_ENFORCE(
tmp == data_type || data_type == -1,
"DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)",
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ 547c1bd1
......@@ -26,6 +26,7 @@ limitations under the License. */
#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/platform/profiler.h"
......@@ -72,6 +73,26 @@ class ParallelExecutorPrivate {
}
}
}
std::unique_ptr<ir::Graph> PrepareGCAndRefCnts(
std::unique_ptr<ir::Graph> graph, size_t max_memory_size);
inline bool HasGarbageCollectors() const { return !gcs_.empty(); }
void ResetRuntimeReferenceCount(const std::vector<std::string> &fetch_tensors,
const std::string &fetched_var_name) {
for (size_t i = 0; i < runtime_ref_cnts_.size(); ++i) {
for (auto &pair : global_ref_cnts_[i]) {
runtime_ref_cnts_[i][pair.first] = pair.second;
}
for (auto &fetch_name : fetch_tensors) {
runtime_ref_cnts_[i].erase(fetch_name);
}
runtime_ref_cnts_[i].erase(fetched_var_name);
}
}
std::vector<platform::Place> places_;
std::vector<Scope *> local_scopes_;
Scope *global_scope_; // not owned
......@@ -83,8 +104,76 @@ class ParallelExecutorPrivate {
bool own_local_scope_;
bool use_cuda_;
bool use_all_reduce_;
// global_ref_cnts_ is only initialized when ParallelExecutor constructs, and
// then keeps unchanged
// Before each iteration, runtime_ref_cnts_ is reset to global_ref_cnts_
std::vector<details::ReferenceCountMap> global_ref_cnts_;
std::vector<details::AtomicReferenceCountMap> runtime_ref_cnts_;
details::GarbageCollectorMap gcs_;
};
std::unique_ptr<ir::Graph> ParallelExecutorPrivate::PrepareGCAndRefCnts(
std::unique_ptr<ir::Graph> graph, size_t max_memory_size) {
for (size_t i = 0; i < places_.size(); ++i) {
auto &place = places_[i];
if (gcs_.count(place) > 0) {
continue;
}
std::unique_ptr<GarbageCollector> gc;
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place)) {
if (IsFastEagerDeletionModeEnabled()) {
gc.reset(new UnsafeFastGPUGarbageCollector(
boost::get<platform::CUDAPlace>(place), max_memory_size));
} else {
gc.reset(new StreamGarbageCollector(
boost::get<platform::CUDAPlace>(place), max_memory_size));
}
VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
} else {
#endif
if (platform::is_cpu_place(place)) {
gc.reset(new CPUGarbageCollector(boost::get<platform::CPUPlace>(place),
max_memory_size));
VLOG(10) << "Created GarbageCollector at " << place;
} else {
PADDLE_THROW("Unsupported place for garbage collection");
}
#ifdef PADDLE_WITH_CUDA
}
#endif
gcs_.emplace(place, std::move(gc));
}
if (!gcs_.empty()) {
std::vector<details::LastLiveOpsOfVars> last_live_ops_of_vars;
auto ref_cnt_pass =
ir::PassRegistry::Instance().Get("reference_count_pass");
ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount,
&global_ref_cnts_);
ref_cnt_pass->SetNotOwned(details::kLastLiveOpsOfVars,
&last_live_ops_of_vars);
graph = ref_cnt_pass->Apply(std::move(graph));
VLOG(10) << "ReferenceCountPass Applied";
auto eager_deletion_pass =
ir::PassRegistry::Instance().Get("eager_deletion_pass");
eager_deletion_pass->SetNotOwned(details::kRuntimeReferenceCount,
&runtime_ref_cnts_);
eager_deletion_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
eager_deletion_pass->SetNotOwned(details::kLastLiveOpsOfVars,
&last_live_ops_of_vars);
eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_);
graph = eager_deletion_pass->Apply(std::move(graph));
VLOG(10) << "EagerDeletionPass Applied";
}
return graph;
}
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
return member_->local_scopes_;
}
......@@ -151,36 +240,18 @@ ParallelExecutor::ParallelExecutor(
std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
main_program, member_->places_, loss_var_name, params,
member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
auto max_memory_size = GetEagerDeletionThreshold();
if (max_memory_size >= 0) {
for (auto &place : member_->places_) {
if (!platform::is_gpu_place(place)) continue;
auto gpu_place = boost::get<platform::CUDAPlace>(place);
if (gcs_[gpu_place.device] == nullptr) {
ref_cnts_[gpu_place.device].reset(new details::ReferenceCountMap());
cur_ref_cnts_[gpu_place.device].reset(
new details::AtomicReferenceCountMap());
gcs_[gpu_place.device].reset(
new StreamGarbageCollector<Tensor>(gpu_place, max_memory_size));
}
}
if (!gcs_.empty()) {
auto ref_cnt_pass =
ir::PassRegistry::Instance().Get("reference_count_pass");
ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount, &ref_cnts_);
ref_cnt_pass->SetNotOwned(details::kCurReferenceCount, &cur_ref_cnts_);
ref_cnt_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
graph = ref_cnt_pass->Apply(std::move(graph));
graph->SetNotOwned("garbage_collector", &gcs_);
}
}
#else
std::unique_ptr<ir::Graph> graph =
build_strategy.Apply(main_program, member_->places_, loss_var_name,
params, member_->local_scopes_, member_->use_cuda_);
#endif
auto max_memory_size = GetEagerDeletionThreshold();
if (max_memory_size >= 0) {
graph = member_->PrepareGCAndRefCnts(std::move(graph),
static_cast<size_t>(max_memory_size));
}
// Step 3. Create vars in each scope. Passes may also create new vars.
// skip control vars and empty vars
std::vector<details::VariableInfo> var_infos;
......@@ -300,18 +371,9 @@ void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
#endif
platform::RecordBlock b(0);
#ifdef PADDLE_WITH_CUDA
if (!gcs_.empty()) {
ResetReferenceCount();
for (auto &pair : cur_ref_cnts_) {
auto &name_map = *(pair.second);
for (auto &fetch_name : fetch_tensors) {
name_map.erase(fetch_name);
}
name_map.erase(fetched_var_name);
}
if (member_->HasGarbageCollectors()) {
member_->ResetRuntimeReferenceCount(fetch_tensors, fetched_var_name);
}
#endif
auto fetch_data = member_->executor_->Run(fetch_tensors);
*member_->global_scope_->Var(fetched_var_name)->GetMutable<FeedFetchList>() =
fetch_data;
......@@ -355,13 +417,11 @@ ParallelExecutor::~ParallelExecutor() {
for (auto &p : member_->places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
}
// member_ must be destructed before gcs_ since the destructor of
// ReferenceCountOpHandle use raw pointers of gcs_ inside.
member_.reset();
delete member_;
}
} // namespace framework
} // namespace paddle
#ifdef PADDLE_WITH_CUDA
USE_PASS(reference_count_pass);
#endif
USE_PASS(eager_deletion_pass);
paddle/fluid/framework/parallel_executor.h
浏览文件 @ 547c1bd1
......@@ -14,7 +14,6 @@ limitations under the License. */
#pragma once
#include <atomic>
#include <string>
#include <unordered_map>
#include <unordered_set>
......@@ -29,10 +28,6 @@ limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/details/reference_count_pass.h"
#endif
namespace paddle {
namespace framework {
......@@ -75,24 +70,7 @@ class ParallelExecutor {
private:
void BCastParamsToDevices(const std::unordered_set<std::string> &vars) const;
std::unique_ptr<ParallelExecutorPrivate> member_;
#ifdef PADDLE_WITH_CUDA
// ref_cnts_ is only initialized when ParallelExecutor constructs, and then
// keeps unchanged
// Before each iteration, cur_ref_cnts_ is reset to ref_cnts_
details::DeviceReferenceCountMap ref_cnts_;
details::AtomicDeviceReferenceCountMap cur_ref_cnts_;
details::DeviceGarbageCollectorMap gcs_;
void ResetReferenceCount() {
for (auto &pair1 : ref_cnts_) {
for (auto &pair2 : *(pair1.second)) {
(*(cur_ref_cnts_[pair1.first]))[pair2.first] = pair2.second;
}
}
}
#endif
ParallelExecutorPrivate *member_;
};
} // namespace framework
......
paddle/fluid/framework/scope.cc
浏览文件 @ 547c1bd1
......@@ -38,6 +38,10 @@ DEFINE_double(
"Memory size threshold (GB) when the garbage collector clear tensors."
"Disabled when this value is less than 0");
DEFINE_bool(fast_eager_deletion_mode, false,
"Fast eager deletion mode. If enabled, memory would release "
"immediately without waiting GPU kernel ends.");
// When in inference scenario, the scopes will not be written by two threads in
// a mean time, but a scope may be read by multiple threads concurrently, and
// the mutex will cause serious performance issue.
......@@ -58,6 +62,8 @@ int64_t GetEagerDeletionThreshold() {
(static_cast<int64_t>(1) << 30));
}
bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; }
Scope::~Scope() { DropKids(); }
Scope& Scope::NewScope() const {
......
paddle/fluid/framework/scope.h
浏览文件 @ 547c1bd1
......@@ -27,6 +27,7 @@ namespace paddle {
namespace framework {
int64_t GetEagerDeletionThreshold();
bool IsFastEagerDeletionModeEnabled();
class Scope;
......
paddle/fluid/framework/selected_rows.cc
浏览文件 @ 547c1bd1
......@@ -218,11 +218,11 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
if (index < 0) {
VLOG(5) << "id " << id << " not in the table, return 0";
framework::VisitDataType(
framework::ToDataType(value_->type()),
value_->type(),
TensorFillVisitor(value, i * value_width, value_width, 0.0));
} else {
framework::VisitDataType(
framework::ToDataType(value_->type()),
value_->type(),
TensorCopyVisitor(value, i * value_width, *value_.get(),
index * value_width, value_width));
}
......
paddle/fluid/framework/tensor.cc
浏览文件 @ 547c1bd1
......@@ -16,7 +16,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
extern size_t SizeOfType(std::type_index type);
extern size_t SizeOfType(proto::VarType::Type type);
void Tensor::check_memory_size() const {
PADDLE_ENFORCE_NOT_NULL(
holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
......@@ -31,7 +31,7 @@ size_t Tensor::memory_size() const {
return holder_ == nullptr ? 0UL : holder_->size() - offset_;
}
void* Tensor::mutable_data(platform::Place place, std::type_index type,
void* Tensor::mutable_data(platform::Place place, proto::VarType::Type type,
memory::Allocator::Attr attr,
size_t requested_size) {
type_ = type;
......
paddle/fluid/framework/tensor.h
浏览文件 @ 547c1bd1
......@@ -19,9 +19,9 @@ limitations under the License. */
#include <memory>
#include <typeindex>
#include <vector>
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/memory/memory.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
......@@ -67,7 +67,7 @@ class Tensor {
friend struct EigenVector;
public:
Tensor() : type_(typeid(float)), offset_(0) {}
Tensor() : type_(proto::VarType::FP32), offset_(0) {}
/*! Return a pointer to mutable memory block. */
template <typename T>
......@@ -88,7 +88,7 @@ class Tensor {
memory::Allocator::Attr attr = memory::Allocator::kDefault,
size_t requested_size = 0);
void* mutable_data(platform::Place place, std::type_index type,
void* mutable_data(platform::Place place, proto::VarType::Type type,
memory::Allocator::Attr attr = memory::Allocator::kDefault,
size_t requested_size = 0);
......@@ -138,7 +138,7 @@ class Tensor {
return holder_->place();
}
std::type_index type() const {
proto::VarType::Type type() const {
PADDLE_ENFORCE_NOT_NULL(
holder_, "Tensor not initialized yet when Tensor::type() is called.");
return type_;
......@@ -158,10 +158,14 @@ class Tensor {
const std::shared_ptr<memory::Allocation>& Holder() const { return holder_; }
size_t offset() const { return offset_; }
std::shared_ptr<memory::Allocation> MoveMemoryHolder() {
return std::move(holder_);
}
private:
/*! holds the memory block if allocated. */
std::shared_ptr<memory::Allocation> holder_;
std::type_index type_;
proto::VarType::Type type_;
/**
* @brief points to elements dimensions.
*
......
paddle/fluid/framework/tensor_impl.h
浏览文件 @ 547c1bd1
......@@ -24,9 +24,8 @@ template <typename T>
inline const T* Tensor::data() const {
check_memory_size();
bool valid =
std::is_same<T, void>::value || type_ == std::type_index(typeid(T));
PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s",
type_.name());
std::is_same<T, void>::value || type_ == DataTypeTrait<T>::DataType;
PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %d", type_);
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
......@@ -38,9 +37,8 @@ template <typename T>
inline T* Tensor::data() {
check_memory_size();
bool valid =
std::is_same<T, void>::value || type_ == std::type_index(typeid(T));
PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s",
type_.name());
std::is_same<T, void>::value || type_ == DataTypeTrait<T>::DataType;
PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s", type_);
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
......@@ -60,7 +58,7 @@ inline T* Tensor::mutable_data(platform::Place place,
size_t requested_size) {
static_assert(std::is_pod<T>::value, "T must be POD");
return reinterpret_cast<T*>(
mutable_data(place, typeid(T), attr, requested_size));
mutable_data(place, DataTypeTrait<T>::DataType, attr, requested_size));
}
inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
......
paddle/fluid/framework/tensor_util.cc
浏览文件 @ 547c1bd1
......@@ -186,8 +186,8 @@ struct AnyDTypeVisitor {
template <typename Predicate, typename DevCtx>
inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor,
const DevCtx& ctx, framework::Tensor* out) {
VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor<Predicate, DevCtx>(
predicate, tensor, ctx, out));
VisitDataType(tensor.type(), AnyDTypeVisitor<Predicate, DevCtx>(
predicate, tensor, ctx, out));
}
template <typename Predicate>
......@@ -379,7 +379,7 @@ void TensorToStream(std::ostream& os, const Tensor& tensor,
// int32_t size
// void* protobuf message
proto::VarType::TensorDesc desc;
desc.set_data_type(framework::ToDataType(tensor.type()));
desc.set_data_type(tensor.type());
auto dims = framework::vectorize(tensor.dims());
auto* pb_dims = desc.mutable_dims();
pb_dims->Resize(static_cast<int>(dims.size()), 0);
......@@ -461,9 +461,7 @@ void TensorFromStream(std::istream& is, Tensor* tensor,
tensor->Resize(framework::make_ddim(dims));
void* buf;
auto ctx = platform::CPUDeviceContext();
size_t size =
tensor->numel() *
framework::SizeOfType(framework::ToTypeIndex(desc.data_type()));
size_t size = tensor->numel() * framework::SizeOfType(desc.data_type());
if (platform::is_gpu_place(dev_ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
Tensor cpu_tensor;
......
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ 547c1bd1
......@@ -289,10 +289,10 @@ bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
auto type = fetch.type();
auto output = &(outputs->at(i));
output->name = fetchs_[idx]->Input("X")[0];
if (type == typeid(float)) {
if (type == framework::proto::VarType::FP32) {
GetFetchOne<float>(fetch, output);
output->dtype = PaddleDType::FLOAT32;
} else if (type == typeid(int64_t)) {
} else if (type == framework::proto::VarType::INT64) {
GetFetchOne<int64_t>(fetch, output);
output->dtype = PaddleDType::INT64;
} else {
......
paddle/fluid/inference/api/analysis_predictor_tester.cc
浏览文件 @ 547c1bd1
......@@ -55,7 +55,12 @@ TEST(AnalysisPredictor, analysis_off) {
}
TEST(AnalysisPredictor, analysis_on) {
AnalysisConfig config(false);
#ifdef PADDLE_WITH_CUDA
AnalysisConfig config(true);
config.fraction_of_gpu_memory = 0.15;
#else
AnalysisConfig config;
#endif
config.model_dir = FLAGS_dirname;
config.enable_ir_optim = true;
......
paddle/fluid/inference/api/api_impl.cc
浏览文件 @ 547c1bd1
......@@ -266,10 +266,10 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
auto type = fetch.type();
auto output = &(outputs->at(i));
output->name = fetchs_[idx]->Input("X")[0];
if (type == typeid(float)) {
if (type == framework::DataTypeTrait<float>::DataType) {
GetFetchOne<float>(fetch, output);
output->dtype = PaddleDType::FLOAT32;
} else if (type == typeid(int64_t)) {
} else if (type == framework::DataTypeTrait<int64_t>::DataType) {
GetFetchOne<int64_t>(fetch, output);
output->dtype = PaddleDType::INT64;
} else {
......
paddle/fluid/inference/api/api_impl_tester.cc
浏览文件 @ 547c1bd1
......@@ -36,10 +36,10 @@ namespace paddle {
PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
PaddleTensor pt;
if (t->type() == typeid(int64_t)) {
if (t->type() == framework::proto::VarType::INT64) {
pt.data.Reset(t->data<void>(), t->numel() * sizeof(int64_t));
pt.dtype = PaddleDType::INT64;
} else if (t->type() == typeid(float)) {
} else if (t->type() == framework::proto::VarType::FP32) {
pt.data.Reset(t->data<void>(), t->numel() * sizeof(float));
pt.dtype = PaddleDType::FLOAT32;
} else {
......
paddle/fluid/inference/api/paddle_pass_builder.h
浏览文件 @ 547c1bd1
......@@ -118,7 +118,10 @@ class GpuPassStrategy : public PassStrategy {
public:
GpuPassStrategy() : PassStrategy({}) {
passes_.assign({
"infer_clean_graph_pass", "conv_bn_fuse_pass",
"infer_clean_graph_pass", //
"conv_bn_fuse_pass", //
"conv_elementwise_add_act_fuse_pass", //
"conv_elementwise_add2_act_fuse_pass", //
});
}
......
paddle/fluid/inference/io.cc
浏览文件 @ 547c1bd1
......@@ -79,7 +79,7 @@ void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
for (auto* var : global_block.AllVars()) {
if (IsPersistable(var)) {
VLOG(3) << "persistable variable's name: " << var->Name();
VLOG(4) << "persistable variable's name: " << var->Name();
framework::VarDesc* new_var = load_block->Var(var->Name());
new_var->SetShape(var->GetShape());
......
paddle/fluid/inference/tests/api/tester_helper.h
浏览文件 @ 547c1bd1
......@@ -373,7 +373,7 @@ static bool CompareTensorData(const framework::LoDTensor &a,
}
for (size_t i = 0; i < a_size; i++) {
if (a.type() == typeid(float)) {
if (a.type() == framework::proto::VarType::FP32) {
const auto *a_data = a.data<float>();
const auto *b_data = b.data<float>();
if (std::abs(a_data[i] - b_data[i]) > 1e-3) {
......@@ -382,7 +382,7 @@ static bool CompareTensorData(const framework::LoDTensor &a,
b_data[i]);
return false;
}
} else if (a.type() == typeid(int64_t)) {
} else if (a.type() == framework::proto::VarType::INT64) {
const auto *a_data = a.data<int64_t>();
const auto *b_data = b.data<int64_t>();
if (std::abs(a_data[i] - b_data[i]) > 1e-3) {
......
paddle/fluid/inference/tests/api/trt_models_tester.cc
浏览文件 @ 547c1bd1
......@@ -78,6 +78,7 @@ void profile(std::string model_dir, bool use_analysis, bool use_tensorrt) {
std::vector<PaddleTensor> outputs;
if (use_analysis || use_tensorrt) {
contrib::AnalysisConfig config(true);
config.pass_builder()->TurnOnDebug();
SetConfig<contrib::AnalysisConfig>(&config, model_dir, true, use_tensorrt,
FLAGS_batch_size);
TestPrediction(reinterpret_cast<PaddlePredictor::Config*>(&config),
......@@ -141,9 +142,31 @@ TEST(TensorRT_resnext50, profile) {
profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
}
TEST(resnext50, compare_analysis_native) {
std::string model_dir = FLAGS_infer_model + "/resnext50";
compare(model_dir, false /*use tensorrt*/);
}
TEST(TensorRT_mobilenet, analysis) {
std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
compare(model_dir, /* use_tensorrt */ false);
compare(model_dir, false /* use_tensorrt */);
}
TEST(AnalysisPredictor, use_gpu) {
std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
AnalysisConfig config(true);
config.model_dir = model_dir;
config.fraction_of_gpu_memory = 0.15;
config.pass_builder()->TurnOnDebug();
std::vector<std::vector<PaddleTensor>> inputs_all;
auto predictor = CreatePaddlePredictor(config);
SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
std::vector<PaddleTensor> outputs;
for (auto& input : inputs_all) {
ASSERT_TRUE(predictor->Run(input, &outputs));
}
}
} // namespace inference
......
paddle/fluid/operators/affine_grid_op.cc
浏览文件 @ 547c1bd1
......@@ -78,7 +78,7 @@ class AffineGridOp : public framework::OperatorWithKernel {
library = framework::LibraryType::kCUDNN;
}
#endif
auto data_type = framework::ToDataType(ctx.Input<Tensor>("Theta")->type());
auto data_type = ctx.Input<Tensor>("Theta")->type();
return framework::OpKernelType(data_type, ctx.GetPlace(),
framework::DataLayout::kAnyLayout, library);
}
......@@ -188,9 +188,9 @@ class AffineGridOpGrad : public framework::OperatorWithKernel {
library_ = framework::LibraryType::kCUDNN;
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Theta")->type()),
ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_);
return framework::OpKernelType(ctx.Input<Tensor>("Theta")->type(),
ctx.GetPlace(),
framework::DataLayout::kAnyLayout, library_);
}
};
......
paddle/fluid/operators/arg_max_op.cc
浏览文件 @ 547c1bd1
......@@ -28,6 +28,5 @@ REGISTER_OP_CPU_KERNEL(
int32_t>,
paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
int16_t>,
paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, size_t>,
paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
uint8_t>);
paddle/fluid/operators/arg_max_op.cu
浏览文件 @ 547c1bd1
......@@ -25,7 +25,5 @@ REGISTER_OP_CUDA_KERNEL(
int32_t>,
paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
int16_t>,
paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
size_t>,
paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
uint8_t>);
paddle/fluid/operators/arg_min_op.cc
浏览文件 @ 547c1bd1
......@@ -28,6 +28,5 @@ REGISTER_OP_CPU_KERNEL(
int32_t>,
paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
int16_t>,
paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, size_t>,
paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
uint8_t>);
paddle/fluid/operators/arg_min_op.cu
浏览文件 @ 547c1bd1
......@@ -25,7 +25,5 @@ REGISTER_OP_CUDA_KERNEL(
int32_t>,
paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
int16_t>,
paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
size_t>,
paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
uint8_t>);
paddle/fluid/operators/array_to_lod_tensor_op.cc
浏览文件 @ 547c1bd1
......@@ -58,7 +58,7 @@ struct ArrayToLoDFunctor : public boost::static_visitor<void> {
ArrayToLoDFunctorImpl<DeviceContext> functor;
functor.dev_ctx_ = dev_ctx;
functor.prev_functor_ = this;
framework::VisitDataType(framework::ToDataType(out->type()), functor);
framework::VisitDataType(out->type(), functor);
}
};
......@@ -91,7 +91,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase {
PADDLE_ENFORCE(!x.empty(), "There's no element in the input array.");
int rank = x[0].dims().size();
platform::Place place = x[0].place();
std::type_index data_type = x[0].type();
auto data_type = x[0].type();
int64_t batch_size = x[0].dims()[0];
framework::DDim ins_dims = rank > 1
? framework::slice_ddim(x[0].dims(), 1, rank)
......
paddle/fluid/operators/attention_lstm_op.cc
浏览文件 @ 547c1bd1
......@@ -121,9 +121,8 @@ void AttentionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType AttentionLSTMOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<framework::LoDTensor>("X")->type(),
ctx.device_context());
}
void AttentionLSTMOpMaker::Make() {
......
paddle/fluid/operators/average_accumulates_op.cc
浏览文件 @ 547c1bd1
......@@ -103,9 +103,8 @@ class AverageAccumulatesOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("param")->type()),
ctx.GetPlace());
return framework::OpKernelType(ctx.Input<Tensor>("param")->type(),
ctx.GetPlace());
}
};
......
paddle/fluid/operators/batch_norm_op.cc
浏览文件 @ 547c1bd1
......@@ -72,8 +72,7 @@ class BatchNormOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
auto input_data_type =
framework::ToDataType(ctx.Input<Tensor>("X")->type());
auto input_data_type = ctx.Input<Tensor>("X")->type();
// By default, the type of the scale, bias, mean,
// and var tensors should both be float. (For float or float16 input tensor)
// or double (For double input tensor).
......@@ -81,17 +80,13 @@ class BatchNormOp : public framework::OperatorWithKernel {
if (input_data_type == framework::proto::VarType::FP64) {
bn_param_type = framework::proto::VarType::FP64;
}
PADDLE_ENFORCE_EQ(bn_param_type,
framework::ToDataType(ctx.Input<Tensor>("Scale")->type()),
PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input<Tensor>("Scale")->type(),
"Scale input should be of float type");
PADDLE_ENFORCE_EQ(bn_param_type,
framework::ToDataType(ctx.Input<Tensor>("Bias")->type()),
PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input<Tensor>("Bias")->type(),
"Bias input should be of float type");
PADDLE_ENFORCE_EQ(bn_param_type,
framework::ToDataType(ctx.Input<Tensor>("Mean")->type()),
PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input<Tensor>("Mean")->type(),
"Mean input should be of float type");
PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType(
ctx.Input<Tensor>("Variance")->type()),
PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input<Tensor>("Variance")->type(),
"Variance input should be of float type");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
......@@ -413,9 +408,8 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout, library);
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
ctx.GetPlace(), layout, library);
}
};
......
paddle/fluid/operators/beam_search_decode_op.cc
浏览文件 @ 547c1bd1
......@@ -145,7 +145,7 @@ class BeamSearchDecodeOp : public framework::OperatorBase {
LoDTensor* sentenceScores = ctx.Output<LoDTensor>("SentenceScores");
framework::VisitDataType(
framework::ToDataType(scores->at(0).type()),
scores->at(0).type(),
BeamSearchDecodeFunctor(*ids, *scores, sentenceIds, sentenceScores,
beam_size, end_id));
}
......
paddle/fluid/operators/beam_search_op.cc
浏览文件 @ 547c1bd1
......@@ -282,8 +282,7 @@ class BeamSearchOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
framework::OpKernelType kt = framework::OpKernelType(
framework::ToDataType(
ctx.Input<framework::LoDTensor>("pre_ids")->type()),
ctx.Input<framework::LoDTensor>("pre_ids")->type(),
platform::CPUPlace());
return kt;
}
......
paddle/fluid/operators/bpr_loss_op.cc
浏览文件 @ 547c1bd1
......@@ -47,9 +47,8 @@ class BprLossOp : public framework::OperatorWithKernel {
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()),
platform::CPUPlace());
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
platform::CPUPlace());
}
};
......@@ -94,9 +93,8 @@ class BprLossGradientOp : public framework::OperatorWithKernel {
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()),
platform::CPUPlace());
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
platform::CPUPlace());
}
};
......
paddle/fluid/operators/controlflow/CMakeLists.txt
浏览文件 @ 547c1bd1
include(operators)
register_operators()
register_operators(DEPS naive_executor)
file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
paddle/fluid/operators/controlflow/conditional_block_op.cc
浏览文件 @ 547c1bd1
......@@ -48,13 +48,12 @@ class ConditionalOp : public framework::OperatorBase {
if (!(ips.size() == 1UL && ips[0]->IsInitialized())) {
PADDLE_THROW("should have one initialized input as condition");
}
if (!(framework::IsType<bool>(ips[0]->type()) && // NOLINT
ips[0]->numel() == 1)) {
PADDLE_THROW(
"condition input's data type should be bool, "
"numel should be 1, actual numel is %d",
ips[0]->numel());
}
PADDLE_ENFORCE(ips[0]->type() == framework::proto::VarType::BOOL &&
ips[0]->numel() == 1,
"condition input's data type should be bool, "
"numel should be 1, actual numel is %d",
ips[0]->numel());
bool res = false;
if (platform::is_gpu_place(ips[0]->place())) {
#ifdef PADDLE_WITH_CUDA
......
paddle/fluid/operators/controlflow/while_op.cc
浏览文件 @ 547c1bd1
......@@ -32,6 +32,20 @@ static constexpr char kStepScopes[] = "StepScopes";
static constexpr char kX[] = "X";
static constexpr char kXGRAD[] = "X@GRAD";
static constexpr char kOutputs[] = "Out";
static constexpr char kSkipEagerDeletionVars[] = "skip_eager_deletion_vars";
namespace { // NOLINT
static std::string GetSkipEagerDeletionVarsDebugString(
const std::vector<std::string> &vars) {
std::string str = "Skip " + std::to_string(vars.size()) +
" var(s) in eager deletion mode: ";
for (auto &var : vars) {
str.append(var);
str.push_back(' ');
}
return str;
}
} // NOLINT
class WhileOp : public framework::OperatorBase {
public:
......@@ -59,7 +73,10 @@ class WhileOp : public framework::OperatorBase {
"Condition of while op must in CPU memory.");
bool is_test = Attr<bool>("is_test");
auto ctx = executor.Prepare(*program, block->ID());
auto &skip_vars = Attr<std::vector<std::string>>(kSkipEagerDeletionVars);
VLOG(2) << GetSkipEagerDeletionVarsDebugString(skip_vars);
auto ctx = executor.Prepare(*program, block->ID(), skip_vars);
while (cond.data<bool>()[0]) {
auto &current_scope = scope.NewScope();
step_scopes->push_back(&current_scope);
......@@ -96,6 +113,10 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker {
"(bool, default false) Set to true for inference only, false "
"for training. Some layers may run faster when this is true.")
.SetDefault(false);
AddAttr<std::vector<std::string>>(kSkipEagerDeletionVars,
"Vars that would skip eager deletion."
"Users should not set this manually.")
.SetDefault(std::vector<std::string>());
AddComment(R"DOC(
)DOC");
}
......@@ -119,7 +140,10 @@ class WhileGradOp : public framework::OperatorBase {
framework::Executor executor(dev_place);
auto *block = Attr<framework::BlockDesc *>(kStepBlock);
auto *program = block->Program();
auto ctx = executor.Prepare(*program, block->ID());
auto &skip_vars = Attr<std::vector<std::string>>(kSkipEagerDeletionVars);
VLOG(2) << GetSkipEagerDeletionVarsDebugString(skip_vars);
auto ctx = executor.Prepare(*program, block->ID(), skip_vars);
auto *step_scopes =
scope.FindVar(Input(kStepScopes))->GetMutable<StepScopeVar>();
......@@ -237,7 +261,7 @@ class WhileGradOp : public framework::OperatorBase {
if (var->IsType<LoDTensor>()) {
auto &inside_tensor = var->Get<framework::LoDTensor>();
framework::AttributeMap attrs;
attrs["dtype"] = framework::ToDataType(inside_tensor.type());
attrs["dtype"] = inside_tensor.type();
attrs["shape"] = framework::vectorize2int(inside_tensor.dims());
attrs["value"] = 0.0f;
......@@ -341,6 +365,8 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker {
// while operator could be renamed.
while_grad->SetAttr("original_output_grad", output_grads_list);
while_grad->SetAttr(kSkipEagerDeletionVars, std::vector<std::string>());
return std::unique_ptr<framework::OpDesc>(while_grad);
}
};
......
paddle/fluid/operators/conv_op.cc
浏览文件 @ 547c1bd1
......@@ -44,7 +44,9 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
std::vector<int> dilations = ctx->Attrs().Get<std::vector<int>>("dilations");
PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5,
"Conv intput should be 4-D or 5-D tensor.");
"Conv intput should be 4-D or 5-D tensor, get %u",
in_dims.size());
PADDLE_ENFORCE_EQ(
in_dims.size(), filter_dims.size(),
"Conv input dimension and filter dimension should be the same.");
......@@ -95,10 +97,8 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
}
#endif
auto input_data_type =
framework::ToDataType(ctx.Input<Tensor>("Input")->type());
auto filter_data_type =
framework::ToDataType(ctx.Input<Tensor>("Filter")->type());
auto input_data_type = ctx.Input<Tensor>("Input")->type();
auto filter_data_type = ctx.Input<Tensor>("Filter")->type();
PADDLE_ENFORCE_EQ(input_data_type, filter_data_type,
"input and filter data type should be consistent");
......@@ -382,9 +382,9 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
layout_, library_, customized_type_value);
return framework::OpKernelType(ctx.Input<Tensor>("Input")->type(),
ctx.GetPlace(), layout_, library_,
customized_type_value);
}
} // namespace operators
......
paddle/fluid/operators/conv_transpose_op.cc
浏览文件 @ 547c1bd1
......@@ -104,9 +104,8 @@ framework::OpKernelType ConvTransposeOp::GetExpectedKernelType(
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
layout_, library_);
return framework::OpKernelType(ctx.Input<Tensor>("Input")->type(),
ctx.GetPlace(), layout_, library_);
}
void Conv2DTransposeOpMaker::Make() {
......@@ -335,9 +334,8 @@ framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType(
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
layout_, library_);
return framework::OpKernelType(ctx.Input<Tensor>("Input")->type(),
ctx.GetPlace(), layout_, library_);
}
} // namespace operators
......
paddle/fluid/operators/crf_decoding_op.cc
浏览文件 @ 547c1bd1
......@@ -118,9 +118,8 @@ class CRFDecodingOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
platform::CPUPlace());
return framework::OpKernelType(ctx.Input<LoDTensor>("Emission")->type(),
platform::CPUPlace());
}
};
} // namespace operators
......
paddle/fluid/operators/crop_op.cc
浏览文件 @ 547c1bd1
......@@ -51,9 +51,8 @@ class CropOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<framework::LoDTensor>("X")->type(),
ctx.device_context());
}
};
......@@ -174,9 +173,7 @@ class CropOpGrad : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))
->type()),
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->type(),
ctx.device_context());
}
};
......
paddle/fluid/operators/cross_entropy_op.cc
浏览文件 @ 547c1bd1
......@@ -57,9 +57,8 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
ctx.device_context());
}
};
......@@ -111,9 +110,8 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
ctx.device_context());
}
};
......
paddle/fluid/operators/ctc_align_op.cc
浏览文件 @ 547c1bd1
......@@ -36,9 +36,8 @@ class CTCAlignOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<Tensor>("Input")->type(),
ctx.device_context());
}
};
......
paddle/fluid/operators/cudnn_lstm_op.cu.cc
浏览文件 @ 547c1bd1
......@@ -300,9 +300,11 @@ class CudnnLSTMGPUKernel : public framework::OpKernel<T> {
}
CudnnRNNCache *cudnn_rnn_cache = nullptr;
if (cache_var->IsInitialized()) {
// const_cast is usually bad.
cudnn_rnn_cache = const_cast<framework::Variable *>(cache_var)
->GetMutable<CudnnRNNCache>();
} else {
// const_cast is usually bad.
cudnn_rnn_cache = const_cast<framework::Variable *>(cache_var)
->GetMutable<CudnnRNNCache>();
std::random_device rnd;
......
paddle/fluid/operators/detection/anchor_generator_op.cc
浏览文件 @ 547c1bd1
......@@ -53,8 +53,7 @@ class AnchorGeneratorOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("Input")->type()),
ctx.device_context());
ctx.Input<framework::Tensor>("Input")->type(), ctx.device_context());
}
};
......
paddle/fluid/operators/detection/bipartite_match_op.cc
浏览文件 @ 547c1bd1
......@@ -45,9 +45,8 @@ class BipartiteMatchOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<LoDTensor>("DistMat")->type()),
platform::CPUPlace());
return framework::OpKernelType(ctx.Input<LoDTensor>("DistMat")->type(),
platform::CPUPlace());
}
};
......
paddle/fluid/operators/detection/density_prior_box_op.cc
浏览文件 @ 547c1bd1
......@@ -66,8 +66,7 @@ class DensityPriorBoxOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("Input")->type()),
ctx.GetPlace());
ctx.Input<framework::Tensor>("Input")->type(), ctx.GetPlace());
}
};
......
paddle/fluid/operators/detection/generate_proposals_op.cc
浏览文件 @ 547c1bd1
......@@ -66,9 +66,8 @@ class GenerateProposalsOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Anchors")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<Tensor>("Anchors")->type(),
ctx.device_context());
}
};
......
paddle/fluid/operators/detection/mine_hard_examples_op.cc
浏览文件 @ 547c1bd1
......@@ -249,8 +249,7 @@ class MineHardExamplesOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("ClsLoss")->type()),
platform::CPUPlace());
ctx.Input<framework::Tensor>("ClsLoss")->type(), platform::CPUPlace());
}
};
......
paddle/fluid/operators/detection/multiclass_nms_op.cc
浏览文件 @ 547c1bd1
......@@ -65,8 +65,7 @@ class MultiClassNMSOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(
ctx.Input<framework::LoDTensor>("Scores")->type()),
ctx.Input<framework::LoDTensor>("Scores")->type(),
platform::CPUPlace());
}
};
......
paddle/fluid/operators/detection/prior_box_op.cc
浏览文件 @ 547c1bd1
......@@ -72,8 +72,7 @@ class PriorBoxOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("Input")->type()),
ctx.device_context());
ctx.Input<framework::Tensor>("Input")->type(), ctx.device_context());
}
};
......
paddle/fluid/operators/detection/roi_perspective_transform_op.cc
浏览文件 @ 547c1bd1
......@@ -498,9 +498,8 @@ class ROIPerspectiveTransformOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<framework::Tensor>("X")->type(),
ctx.device_context());
}
};
......@@ -519,9 +518,8 @@ class ROIPerspectiveTransformGradOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
ctx.device_context());
return framework::OpKernelType(ctx.Input<framework::Tensor>("X")->type(),
ctx.device_context());
}
};
......
paddle/fluid/operators/distributed/brpc_rdma_pool.cc 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/distributed/brpc_rdma_pool.h 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/distributed/brpc_serde_test.cc 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/fill_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/lrn_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/lstm_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/lstmp_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/mean_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/nce_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/pad2d_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/pool_op.h
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/prelu_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/print_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.cc 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.cu 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.h 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/slice_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/spp_op.h
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/sum_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/operators/unpool_op.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/platform/stream_callback_manager.cc 0 → 100644
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ 547c1bd1
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 547c1bd1
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 547c1bd1
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 547c1bd1
此差异已折叠。
python/paddle/fluid/tests/unittests/ngraph/__init__.py 0 → 100644
浏览文件 @ 547c1bd1
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