Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 cc29bec6 编写于 作者: L lujun 提交者： GitHub
浏览文件
操作

Merge pull request #16 from PaddlePaddle/develop 

merge to local
上级 8ba17759 ed61d67c
展开全部 隐藏空白更改
内联 并排
Showing with 13382 addition and 2207 deletion
CMakeLists.txt
浏览文件 @ cc29bec6
......@@ -64,6 +64,7 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(WITH_PSLIB "Compile with pslib support" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
# TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
......@@ -190,7 +191,14 @@ include(configure) # add paddle env configuration
if(WITH_GPU)
include(cuda)
include(tensorrt)
include(anakin_subgraph)
endif()
if(WITH_GPU AND NOT WIN32)
message(STATUS "add dgc lib.")
include(external/dgc)
endif()
if(WITH_MKL OR WITH_MKLML)
include(external/anakin)
elseif()
......
CONTRIBUTING.md
浏览文件 @ cc29bec6
......@@ -156,7 +156,7 @@ python \
This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3. This suggests that we output overall messages in lower verbose levels, so they display with higher probability. When coding C++, please follow the verbose level convention as follows:
- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)
- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform)
- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math)
- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework)
- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)
- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform)
- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators/math/)
cmake/anakin_subgraph.cmake 0 → 100644
浏览文件 @ cc29bec6
if(NOT WITH_GPU)
return()
endif()
set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
find_path(ANAKIN_INCLUDE_DIR anakin_config.h
PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
$ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
NO_DEFAULT_PATH
)
find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
PATHS ${ANAKIN_ROOT}
$ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
NO_DEFAULT_PATH
DOC "Path to ANAKIN library.")
if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
if(WITH_DSO)
set(ANAKIN_FOUND ON)
endif(WITH_DSO)
else()
set(ANAKIN_FOUND OFF)
endif()
if(ANAKIN_FOUND)
message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
include_directories(${ANAKIN_ROOT}/include)
include_directories(${ANAKIN_ROOT}/include/saber)
link_directories(${ANAKIN_ROOT})
add_definitions(-DPADDLE_WITH_ANAKIN)
endif()
cmake/external/dgc.cmake 0 → 100644
浏览文件 @ cc29bec6
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
INCLUDE(ExternalProject)
SET(DGC_SOURCES_DIR "${THIRD_PARTY_PATH}/dgc")
SET(DGC_INSTALL_DIR "${THIRD_PARTY_PATH}/install/dgc")
SET(DGC_INCLUDE_DIR "${DGC_INSTALL_DIR}/include" CACHE PATH "dgc include directory." FORCE)
SET(DGC_LIBRARIES "${DGC_INSTALL_DIR}/lib/libdgc.a" CACHE FILEPATH "dgc library." FORCE)
INCLUDE_DIRECTORIES(${DGC_INCLUDE_DIR})
ExternalProject_Add(
extern_dgc
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/PaddlePaddle/Fleet"
GIT_TAG "2d04dc3800cdd0601f1b65d547dabcc60b0cf9dc"
SOURCE_DIR "${DGC_SOURCES_DIR}"
CONFIGURE_COMMAND ""
BUILD_COMMAND cd collective && make -j
INSTALL_COMMAND mkdir -p ${DGC_INSTALL_DIR}/lib/ ${DGC_INCLUDE_DIR}/dgc
&& cp ${DGC_SOURCES_DIR}/collective/build/lib/libdgc.a ${DGC_LIBRARIES}
&& cp ${DGC_SOURCES_DIR}/collective/build/include/dgc.h ${DGC_INCLUDE_DIR}/dgc/
BUILD_IN_SOURCE 1
)
ADD_LIBRARY(dgc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET dgc PROPERTY IMPORTED_LOCATION ${DGC_LIBRARIES})
ADD_DEPENDENCIES(dgc extern_dgc)
LIST(APPEND external_project_dependencies dgc)
cmake/external/ngraph.cmake
浏览文件 @ cc29bec6
......@@ -57,20 +57,25 @@ SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
ExternalProject_Add(
${NGRAPH_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_PROJECT} ${MKLML_PROJECT}
GIT_REPOSITORY ${NGRAPH_GIT_REPO}
GIT_TAG ${NGRAPH_GIT_TAG}
PREFIX ${NGRAPH_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
CMAKE_ARGS -DNGRAPH_UNIT_TEST_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_TOOLS_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_INTERPRETER_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
CMAKE_ARGS -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
DEPENDS ${MKLDNN_PROJECT} ${MKLML_PROJECT}
GIT_REPOSITORY ${NGRAPH_GIT_REPO}
GIT_TAG ${NGRAPH_GIT_TAG}
PREFIX ${NGRAPH_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_GENERATOR ${CMAKE_GENERATOR}
CMAKE_GENERATOR_PLATFORM ${CMAKE_GENERATOR_PLATFORM}
CMAKE_GENERATOR_TOOLSET ${CMAKE_GENERATOR_TOOLSET}
CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
CMAKE_ARGS -DNGRAPH_UNIT_TEST_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_TOOLS_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_INTERPRETER_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
CMAKE_ARGS -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
)
add_dependencies(ngraph ${NGRAPH_PROJECT})
......
cmake/inference_lib.cmake
浏览文件 @ cc29bec6
......@@ -131,6 +131,15 @@ elseif (NOT CBLAS_FOUND OR WIN32)
)
endif ()
if (WITH_GPU AND NOT WIN32)
set(dgc_dir "${FLUID_INSTALL_DIR}/third_party/install/dgc")
copy(dgc_lib
SRCS ${DGC_INSTALL_DIR}/lib ${DGC_INSTALL_DIR}/include
DSTS ${dgc_dir} ${dgc_dir}
DEPS dgc)
endif()
if (WITH_MKLDNN)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
copy(mkldnn_lib
......
cmake/operators.cmake
浏览文件 @ cc29bec6
......@@ -110,7 +110,7 @@ function(op_library TARGET)
# Define operators that don't need pybind here.
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op")
"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op" "dgc_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1)
endif()
......
cmake/tensorrt.cmake
浏览文件 @ cc29bec6
......@@ -33,5 +33,6 @@ if(TENSORRT_FOUND)
message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
"Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
include_directories(${TENSORRT_INCLUDE_DIR})
link_directories(${TENSORRT_LIBRARY})
add_definitions(-DPADDLE_WITH_TENSORRT)
endif()
paddle/fluid/API.spec
浏览文件 @ cc29bec6
......@@ -211,7 +211,7 @@ paddle.fluid.layers.mean (ArgSpec(args=['x', 'name'], varargs=None, keywords=Non
paddle.fluid.layers.mul (ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None)), ('document', 'ccd37fa6b53f074adbfb732d738c4c2d'))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits (ArgSpec(args=['x', 'label', 'ignore_index', 'name', 'normalize'], varargs=None, keywords=None, defaults=(-100, None, False)), ('document', '180c284317ea45ef89a460d8d79c0b72'))
paddle.fluid.layers.maxout (ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '891870d069a6aea746d34cc53b61690c'))
paddle.fluid.layers.space_to_depth (ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5f207ae10589ebe38a63575ef6ff8e1e'))
paddle.fluid.layers.space_to_depth (ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'a9221eaef53884a00654e028551b78e2'))
paddle.fluid.layers.affine_grid (ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '51def402b8910e163cbace9d0c0526ed'))
paddle.fluid.layers.sequence_reverse (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '77a6d80aa5551ca70324fc975c44507f'))
paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name', 'act'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None, None)), ('document', 'ab84fdc6dc60f3ad9aa397e6007e3bf9'))
......@@ -483,6 +483,11 @@ paddle.fluid.optimizer.LarsMomentumOptimizer.apply_gradients (ArgSpec(args=['sel
paddle.fluid.optimizer.LarsMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.LarsMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.LarsMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.DGCMomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'rampup_begin_step', 'rampup_step', 'sparsity', 'use_nesterov', 'local_grad_clip_norm', 'num_trainers', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1, [0.999], False, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.DGCMomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.DGCMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.DGCMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.DGCMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.backward.append_backward (ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '1a79bd7d10ae54ca763ec81bca36ba24'))
paddle.fluid.regularizer.L1DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.regularizer.L2DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
......@@ -520,6 +525,7 @@ paddle.fluid.unique_name.guard (ArgSpec(args=['new_generator'], varargs=None, ke
paddle.fluid.recordio_writer.convert_reader_to_recordio_file (ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', '65c7523e86f0c50bb729b01667f36310'))
paddle.fluid.recordio_writer.convert_reader_to_recordio_files (ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', 'bc643f0f5f1b9db57ff0d8a57d379bd7'))
paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope
paddle.fluid.install_check.run_check (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '66b7c84a17ed32fec2df9628367be2b9'))
paddle.reader.cache (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', '1676886070eb607cb608f7ba47be0d3c'))
paddle.reader.map_readers (ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None), ('document', '77cbadb09df588e21e5cc0819b69c87d'))
paddle.reader.buffered (ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None), ('document', '0d6186f109feceb99f60ec50a0a624cb'))
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ cc29bec6
......@@ -63,7 +63,7 @@ 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(garbage_collector SRCS garbage_collector.cc DEPS device_context memory gflags glog)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
cc_test(reader_test SRCS reader_test.cc DEPS reader)
......@@ -164,6 +164,8 @@ else()
set(NGRAPH_EXE_DEPS)
endif()
cc_library(executor_gc_helper SRCS executor_gc_helper.cc DEPS scope proto_desc operator garbage_collector)
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_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS})
......@@ -174,7 +176,7 @@ else()
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
target_link_libraries(executor garbage_collector while_op_helper)
target_link_libraries(executor while_op_helper executor_gc_helper)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS
threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor
......@@ -194,6 +196,7 @@ cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_con
cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
proto_desc)
cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper)
cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ cc29bec6
......@@ -5,6 +5,7 @@ cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_h
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(fetch_barrier_op_handle SRCS fetch_barrier_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
......@@ -22,7 +23,7 @@ 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)
dynload_cuda variable_visitor dgc)
nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
if(WITH_DISTRIBUTE)
......@@ -72,7 +73,7 @@ cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS grap
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
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 fused_broadcast_op_handle)
scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
......
paddle/fluid/framework/details/all_reduce_deps_pass.cc
浏览文件 @ cc29bec6
......@@ -13,6 +13,7 @@
// limitations under the License.
#include <algorithm>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
......@@ -41,8 +42,7 @@ VarHandle* GetValidInput(const OpHandleBase* a) {
return nullptr;
}
std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void AllReduceDepsPass::ApplyImpl(ir::Graph* graph) const {
auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
// get vars order
......@@ -52,13 +52,28 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
// Note that must assert topology sort is stable
auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
for (auto* op_desc : ops) {
auto outputs = op_desc->Outputs();
for (auto& o_it : outputs) {
for (auto& v : o_it.second) { // values
vars[v] = order;
try {
bool is_bk_op =
static_cast<bool>(boost::get<int>(op_desc->GetAttr(
OpProtoAndCheckerMaker::OpRoleAttrName())) &
static_cast<int>(OpRole::kBackward));
if (!is_bk_op) continue;
auto backward_vars =
boost::get<std::vector<std::string>>(op_desc->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
auto outputs = op_desc->Outputs();
for (auto& o_it : outputs) {
for (auto& v : o_it.second) { // values
vars[v] = order;
VLOG(1) << "in all_reduce_deps_pass:" << v;
}
}
order++;
} catch (boost::bad_get e) {
}
order++;
}
std::vector<OpHandleBase*> dist_ops;
......@@ -70,7 +85,8 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
}
}
VLOG(10) << "dist_ops size:" << dist_ops.size() << std::endl;
VLOG(10) << "dist_ops size:" << dist_ops.size()
<< ", outputs size:" << vars.size() << ", ops size:" << ops.size();
std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1,
OpHandleBase* op2) {
......@@ -83,6 +99,10 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
auto l_it = vars.find(i0->name());
auto r_it = vars.find(i1->name());
PADDLE_ENFORCE(l_it != vars.end() && r_it != vars.end(),
"can't find var's name %s and %s in opdesc", i0->name(),
i1->name());
if (l_it->second < r_it->second) return true;
if (l_it->second == r_it->second) {
......@@ -110,8 +130,6 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
VLOG(10) << "pre_op:" << pre_op->DebugString()
<< ", op:" << op->DebugString();
}
return graph;
}
} // namespace details
......
paddle/fluid/framework/details/all_reduce_deps_pass.h
浏览文件 @ cc29bec6
......@@ -24,8 +24,7 @@ namespace details {
// TODO(gongwb): overlap allreduce with backward computation.
class AllReduceDepsPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace details
......
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ cc29bec6
......@@ -16,6 +16,13 @@
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/framework/operator.h"
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "dgc/dgc.h"
#endif
#include "paddle/fluid/platform/gpu_info.h"
#include "paddle/fluid/platform/profiler.h"
// asynchronous nccl allreduce or synchronous issue:
......@@ -33,11 +40,14 @@ namespace details {
AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs)
const platform::NCCLContextMap *ctxs,
bool is_encoded, int nranks)
: OpHandleBase(node),
local_scopes_(local_scopes),
places_(places),
nccl_ctxs_(ctxs) {
nccl_ctxs_(ctxs),
is_encoded_(is_encoded),
nranks_(nranks) {
if (nccl_ctxs_) {
for (auto &p : places_) {
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
......@@ -51,7 +61,185 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
: OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
#endif
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void AllReduceOpHandle::RunImplEncoded() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(
in_var_handles.size(), places_.size(),
"The NoDummyInputSize should be equal to the number of places.");
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
std::vector<const LoDTensor *> ins;
std::vector<LoDTensor *> outs;
int k = -1;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &local_scope =
local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto original_name =
paddle::framework::GradOriginalVarName(in_var_handles[i]->name());
auto encode_var_name = original_name + g_dgc_encoded;
auto *in_var = local_scope->FindVar(encode_var_name);
PADDLE_ENFORCE_NOT_NULL(in_var);
auto &in = in_var->Get<LoDTensor>();
ins.emplace_back(&in);
auto *out = local_scope->FindVar(out_var_handles[i]->name())
->GetMutable<LoDTensor>();
outs.emplace_back(out);
if (k < 0) {
k = GetKValue(in_var_handles[i]->name());
}
}
PADDLE_ENFORCE(platform::is_gpu_place(ins[0]->place()));
PADDLE_ENFORCE(platform::is_gpu_place(outs[0]->place()));
PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
int dtype = -1;
size_t in_numel = 0;
size_t out_numel = 0;
PADDLE_ENFORCE(nranks_ > 1);
std::vector<std::function<void()>> all_reduce_calls;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &place = places_[i];
auto &in = *ins[i];
void *in_tensor_buf = const_cast<void *>(in.data<void>());
auto &out = *outs[i];
float *out_tensor_buf = out.data<float>();
dtype = (dtype == -1) ? platform::ToNCCLDataType(in.type()) : dtype;
in_numel = (in_numel == 0) ? static_cast<size_t>(in.numel()) : in_numel;
PADDLE_ENFORCE(in_numel % 2 == 0);
PADDLE_ENFORCE(in_numel / 2 == static_cast<size_t>(k));
out_numel = (out_numel == 0) ? static_cast<size_t>(out.numel()) : out_numel;
int dev_id = boost::get<platform::CUDAPlace>(place).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_;
auto &allocator =
platform::DeviceTemporaryAllocator::Instance().Get(place, stream);
int encode_size = 2 * k * sizeof(int);
// dgc use ncclAllGather to get all the encoded data
// so the buffer need nranks.
int buf_size = nranks_ * encode_size;
auto tmp_ious_data = allocator.Allocate(buf_size);
void *gather_buff = reinterpret_cast<void *>(tmp_ious_data->ptr());
VLOG(10) << "in_numel:" << in_numel << ", out_numel:" << out_numel
<< ", nranks:" << nranks_ << ", gather_buf size:" << buf_size
<< ", k:" << k << ", place:" << place << ", dtype:" << dtype;
all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(paddle::communication::dgc::sparseAllGReduce(
in_tensor_buf, gather_buff, k, out_tensor_buf, out_numel, comm,
stream));
});
}
this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device
all_reduce_calls[0]();
} else {
platform::NCCLGroupGuard guard;
for (auto &call : all_reduce_calls) {
call();
}
}
});
if (FLAGS_sync_nccl_allreduce) {
for (auto &p : places_) {
int dev_id = boost::get<platform::CUDAPlace>(p).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
cudaError_t e_sync = cudaStreamSynchronize(stream);
if (e_sync != 0) {
LOG(FATAL) << "cudaStreamSynchronize " << cudaGetErrorString(e_sync);
}
cudaError_t e_get = cudaGetLastError();
if (e_get != 0) {
LOG(FATAL) << "cudaGetLastError " << cudaGetErrorString(e_get)
<< " errno:" << e_get;
}
}
}
}
int AllReduceOpHandle::GetKValue(const std::string &grad_name) {
auto original_name = paddle::framework::GradOriginalVarName(grad_name);
auto var_name = original_name + g_dgc_k;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto var = local_scope->FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(var);
auto tensor = var->Get<LoDTensor>().data<float>();
return *tensor;
}
#endif
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
bool AllReduceOpHandle::IsEncoded() {
if (!is_encoded_) {
return false;
}
auto counter_name = g_dgc_counter_name;
auto step_name = g_dgc_rampup_begin_step;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto count_var = local_scope->FindVar(counter_name);
auto step_var = local_scope->FindVar(step_name);
if (count_var == nullptr || step_var == nullptr) {
PADDLE_THROW("not find count_var:%s or step_var:%s", counter_name,
step_var);
}
float count = *count_var->Get<LoDTensor>().data<float>();
float step = *step_var->Get<LoDTensor>().data<float>();
if (static_cast<int>(count) < static_cast<int>(step)) {
VLOG(10) << "in all_reduce currentstep:" << count
<< " < rampup_begin_step:" << step
<< " so not use sparse all reduce";
return false;
}
return true;
}
#else
bool AllReduceOpHandle::IsEncoded() { return false; }
#endif
void AllReduceOpHandle::RunImpl() {
if (!IsEncoded()) {
RunImplNormal();
return;
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
RunImplEncoded();
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
}
void AllReduceOpHandle::RunImplNormal() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
......@@ -72,6 +260,8 @@ void AllReduceOpHandle::RunImpl() {
auto &lod_tensor =
local_scope.FindVar(in_var_handles[i]->name())->Get<LoDTensor>();
lod_tensors.emplace_back(&lod_tensor);
VLOG(10) << "place:" << i << ", input_name:" << in_var_handles[i]->name()
<< ", out_name:" << out_var_handles[i]->name();
PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(),
"The name of input and output should be equal.");
}
......@@ -99,13 +289,17 @@ void AllReduceOpHandle::RunImpl() {
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_;
VLOG(10) << "before all reduce buffer:" << buffer << ", numel:" << numel
<< ", dev_id:" << dev_id << ", dtype:" << dtype
<< ", place:" << p;
all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
buffer, buffer, numel, static_cast<ncclDataType_t>(dtype), ncclSum,
comm, stream));
});
}
this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device
......
paddle/fluid/framework/details/all_reduce_op_handle.h
浏览文件 @ cc29bec6
......@@ -28,11 +28,19 @@ namespace paddle {
namespace framework {
namespace details {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
constexpr char g_dgc_counter_name[] = "__g_dgc_counter__";
constexpr char g_dgc_rampup_begin_step[] = "__g_rampup_begin_step__";
constexpr char g_dgc_encoded[] = "__dgc_encoded__";
constexpr char g_dgc_k[] = "__dgc_k__";
#endif
struct AllReduceOpHandle : public OpHandleBase {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs);
const platform::NCCLContextMap *ctxs,
bool is_encoded = false, int nranks = -1);
#else
AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places);
......@@ -50,8 +58,14 @@ struct AllReduceOpHandle : public OpHandleBase {
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void RunImplEncoded();
const platform::NCCLContextMap *nccl_ctxs_;
bool is_encoded_{false};
int nranks_{-1};
int GetKValue(const std::string &grad_name);
#endif
void RunImplNormal();
bool IsEncoded();
};
} // namespace details
......
paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.cc
浏览文件 @ cc29bec6
......@@ -46,8 +46,7 @@ static framework::proto::VarType::Type kDefaultDtype =
class AllocContinuousSpaceForGradPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
void ApplyImpl(ir::Graph *graph) const override {
ir::Graph &result = *graph;
auto &places = Get<const std::vector<platform::Place>>(kPlaces);
......@@ -65,7 +64,7 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
if (params_grads.size() == 0) {
VLOG(10) << "Doesn't find gradients";
return std::move(graph);
return;
}
std::unordered_map<std::string, ir::Node *> vars;
......@@ -124,8 +123,6 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars,
fused_var_name, params_grads);
return std::move(graph);
}
template <typename AttrType>
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ cc29bec6
......@@ -204,15 +204,16 @@ bool BuildStrategy::IsMultiDevPass(const std::string &pass_name) const {
return framework::details::MultiDevSSAGraphBuilder().count(pass_name) > 0;
}
std::unique_ptr<ir::Graph> BuildStrategy::Apply(
std::unique_ptr<ir::Graph> graph,
const std::vector<platform::Place> &places,
const std::string &loss_var_name, const std::vector<Scope *> &local_scopes,
const size_t &nranks,
ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::vector<Scope *> &local_scopes,
const size_t &nranks,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const {
const bool use_cuda,
platform::NCCLContextMap *nccl_ctxs) const {
#else
const bool use_cuda) const {
const bool use_cuda) const {
#endif
// Create a default one if not finalized by user.
CreatePassesFromStrategy(false);
......@@ -265,7 +266,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
}
}
VLOG(3) << "Start Apply Pass " << pass->Type();
graph = pass->Apply(std::move(graph));
graph = pass->Apply(graph);
VLOG(3) << "Finish Apply Pass " << pass->Type();
}
return graph;
......
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ cc29bec6
......@@ -120,16 +120,15 @@ struct BuildStrategy {
// Apply the passes built by the pass_builder_. The passes will be
// applied to the Program and output an ir::Graph.
std::unique_ptr<ir::Graph> Apply(std::unique_ptr<ir::Graph> graph,
const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::vector<Scope *> &local_scopes,
const size_t &nranks,
ir::Graph *Apply(ir::Graph *graph, const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::vector<Scope *> &local_scopes,
const size_t &nranks,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const bool use_cuda,
platform::NCCLContextMap *nccl_ctxs) const;
const bool use_cuda,
platform::NCCLContextMap *nccl_ctxs) const;
#else
const bool use_cuda) const;
const bool use_cuda) const;
#endif
// If set true, ParallelExecutor would build the main_program into multiple
......
paddle/fluid/framework/details/eager_deletion_pass.cc
浏览文件 @ cc29bec6
......@@ -22,14 +22,9 @@
#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/garbage_collector.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
"Fraction of eager deletion. If less than 1.0, all variables in "
"the program would be sorted according to its memory size, and "
"only the FLAGS_memory_fraction_of_eager_deletion of the largest "
"variables would be deleted.");
namespace paddle {
namespace framework {
namespace details {
......@@ -175,12 +170,10 @@ static OpToVarNameSetMap ShrinkGCVars(
class EagerDeletionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph *graph) const override;
};
std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void EagerDeletionPass::ApplyImpl(ir::Graph *graph) const {
auto &ref_cnts =
Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
PADDLE_ENFORCE(ref_cnts.empty(),
......@@ -206,8 +199,9 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
}
}
op_vars_map = ShrinkGCVars(op_vars_map, vars, places,
FLAGS_memory_fraction_of_eager_deletion);
double memory_fraction = framework::GetEagerDeletionMemoryFraction();
op_vars_map = ShrinkGCVars(op_vars_map, vars, places, memory_fraction);
for (auto &pair : op_vars_map) {
auto *op = pair.first;
......@@ -239,13 +233,12 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
eager_deletion_op->AddOutput(dummy_leaf);
}
VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = "
<< FLAGS_memory_fraction_of_eager_deletion;
VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " << memory_fraction;
VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
auto while_op_eager_deletion_pass =
ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass");
return while_op_eager_deletion_pass->Apply(std::move(graph));
while_op_eager_deletion_pass->Apply(graph);
}
} // namespace details
......
paddle/fluid/framework/details/early_delete_op_handle.h 已删除 100644 → 0
浏览文件 @ 8ba17759
// 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 <string>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/var_handle.h"
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/lod_tensor_array.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 {
class EarlyDeleteOpHandle : public OpHandleBase {
public:
EarlyDeleteOpHandle(ir::Node* node, const Scope* scope,
const platform::Place& place,
const std::vector<std::string>& names,
GarbageCollector* gc)
: OpHandleBase(node),
scope_(scope),
place_(place),
names_(names),
gc_(gc) {
#ifdef PADDLE_WITH_CUDA
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(place);
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
#endif
}
~EarlyDeleteOpHandle() {
#ifdef PADDLE_WITH_CUDA
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
#endif
}
std::string Name() const override { return "early_delete"; }
protected:
void RunImpl() override {
std::vector<std::shared_ptr<memory::Allocation>> tensors;
auto* local_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope*>();
for (auto& var_name : names_) {
auto* var = local_scope->FindVar(var_name);
PADDLE_ENFORCE(var != nullptr,
string::Sprintf("Local Scope not has var %s", var_name));
if (var->IsType<LoDTensor>()) {
tensors.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
tensors.emplace_back(var->GetMutable<SelectedRows>()
->mutable_value()
->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
LoDTensorArray* tensor_array = var->GetMutable<LoDTensorArray>();
for (auto& tensor : *tensor_array) {
tensors.emplace_back(tensor.MoveMemoryHolder());
}
}
}
if (!tensors.empty()) {
ClearTensors(tensors);
}
}
private:
void ClearTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
if (platform::is_cpu_place(place_)) {
ClearCPUTensors(tensors);
} else {
ClearGPUTensors(tensors);
}
}
void ClearCPUTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
auto* gc = dynamic_cast<CPUGarbageCollector*>(gc_);
if (gc != nullptr) {
gc->Add(tensors);
}
}
void ClearGPUTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
#ifdef PADDLE_WITH_CUDA
auto* gc = dynamic_cast<StreamGarbageCollector*>(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*>(gc_) != nullptr;
#endif
}
const Scope* scope_;
const platform::Place place_;
std::vector<std::string> names_;
GarbageCollector* gc_;
#ifdef PADDLE_WITH_CUDA
platform::CUDADeviceContext* dev_ctx_;
cudaEvent_t event_;
#endif
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fetch_barrier_op_handle.cc 0 → 100644
浏览文件 @ cc29bec6
// 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/fetch_barrier_op_handle.h"
#include <string>
namespace paddle {
namespace framework {
namespace details {
FetchBarrierOpHandle::FetchBarrierOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places)
// fetch_barrier op always run on place0, but output on all places.
: OpHandleBase(node),
op_(framework::OpRegistry::CreateOp(*node->Op())),
local_scopes_(local_scopes),
places_(places),
run_scope_(local_scopes[0]),
place_(places[0]) {
for (auto &p : places) {
this->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p));
}
}
bool FetchBarrierOpHandle::IsMultiDeviceTransfer() {
// override IsMultiDeviceTransfer to return true
return true;
}
void FetchBarrierOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
auto run_func = [this]() {
op_->Run(*run_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
};
if (is_lock_and_record_event_free_) {
run_func();
} else {
this->RunAndRecordEvent(run_func);
}
}
bool FetchBarrierOpHandle::NeedWait(VarHandleBase *in_var) {
bool need_wait =
in_var && in_var->GeneratedOp() &&
in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
return need_wait;
}
std::string FetchBarrierOpHandle::Name() const { return op_->Type(); }
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fetch_barrier_op_handle.h 0 → 100644
浏览文件 @ cc29bec6
// 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 <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace framework {
namespace details {
// **NOTE**: fetch_barrier op is special it outputs all recved variables on
// all places if there are multiple places, must init with
// multiple dev_ctxes_ !!!!
struct FetchBarrierOpHandle : public OpHandleBase {
public:
FetchBarrierOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places);
bool IsMultiDeviceTransfer() override;
std::string Name() const override;
protected:
void RunImpl() override;
bool NeedWait(VarHandleBase *in_var) override;
private:
std::unique_ptr<OperatorBase> op_;
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
Scope *run_scope_;
platform::Place place_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fuse_all_reduce_op_pass.cc
浏览文件 @ cc29bec6
......@@ -28,8 +28,7 @@ namespace details {
class FuseAllReduceOpPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
void ApplyImpl(ir::Graph *graph) const override {
ir::Graph &result = *graph;
auto &places = Get<const std::vector<platform::Place>>(kPlaces);
......@@ -71,7 +70,7 @@ class FuseAllReduceOpPass : public ir::Pass {
VLOG(10) << "Find all_reduce_ops: " << all_reduce_ops.size();
if (all_reduce_ops.size() == 0) {
return std::move(graph);
return;
}
PADDLE_ENFORCE_EQ(all_reduce_ops.size(), grads.size(),
......@@ -99,7 +98,6 @@ class FuseAllReduceOpPass : public ir::Pass {
group_all_reduce_ops, &result);
#endif
}
return std::move(graph);
}
void InsertFusedAllReduce(const std::vector<platform::Place> &places,
......
paddle/fluid/framework/details/inplace_op_pass.cc
浏览文件 @ cc29bec6
......@@ -17,6 +17,8 @@
#include <deque>
#include <iterator>
#include <memory>
#include <queue>
#include <sstream>
#include <stack>
#include <string>
#include <unordered_map>
......@@ -142,20 +144,19 @@ void InplacePass::InitSSAGraphNodes() const {
}
}
std::unique_ptr<ir::Graph> InplacePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void InplacePass::ApplyImpl(ir::Graph* graph) const {
var_nodes_.clear();
view_.Build(graph.get());
view_.Build(graph);
InitSSAGraphNodes();
auto cnt = 0;
for (auto* op : view_.AllOps()) {
VLOG(4) << "Handle op " << cnt++ << ": " << op->Name();
if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
continue;
TryInplaceOpInputOutput(op, graph.get());
TryInplaceOpInputOutput(op, graph);
}
graph->ResolveHazard(var_nodes_);
return graph;
// graph->ResolveHazard(var_nodes_);
}
void InplacePass::InplaceModifyDesc(const std::string& var,
......@@ -264,13 +265,10 @@ void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
ir::Graph* graph) const {
VLOG(4) << "Try to inplace op " << op->Name();
// FIXME(liuwei1031): Graph is not aware of the existence of BlockDescs and
// ProgramDescs.
// The operations related to BlockDesc or ProgramDesc should perform on Graph
// or Node directly!
PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
"op_desc is nullptr");
// PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
// "op_desc is nullptr");
// some pre-requirments need to meet if the op want to inplaced.
PADDLE_ENFORCE(op->Op() != nullptr, "op_desc is nullptr");
auto* op_desc = op->Op();
auto& infer_inplace =
......@@ -281,21 +279,58 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
PADDLE_ENFORCE(static_cast<bool>(infer_inplace),
"%s's infer_inplace has not been registered", op_desc->Type());
auto* block = op_desc->Block();
auto in_to_outs = infer_inplace(*op_desc, block);
auto in_to_outs = infer_inplace(*op_desc);
auto& all_ops = view_.AllOps();
auto cursor = std::find(all_ops.begin(), all_ops.end(), op);
size_t idx = std::distance(all_ops.begin(), cursor);
for (auto& pair : in_to_outs) {
auto& in_var_name = pair.first;
auto& out_var_name = pair.second;
auto& in_para_name = pair.first;
auto& out_para_name = pair.second;
auto input_vars = op->Op()->Input(in_para_name);
if (!input_vars.size()) {
VLOG(4) << "Parameter " << in_para_name << " is empty skip "
<< in_para_name << " => " << out_para_name << " pair";
continue;
}
auto output_vars = op->Op()->Output(out_para_name);
if (!output_vars.size()) {
VLOG(4) << "Parameter " << out_para_name << " is empty skip "
<< in_para_name << " => " << out_para_name << " pair";
continue;
}
auto in_var_name = input_vars.at(0);
auto out_var_name = output_vars.at(0);
auto* in_node = view_.GetNodeByName(in_var_name, op->inputs);
auto* out_node = view_.GetNodeByName(out_var_name, op->outputs);
VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
bool can_replace = true;
if (in_var_name == out_var_name) {
can_replace = false;
VLOG(4) << "SKIP: Input variable " << in_var_name << " & Output variable "
<< out_var_name << " are the same";
} else if (!NodeCanReused(in_node)) {
can_replace = false;
VLOG(4) << "SKIP: Input varialbe " << in_var_name << "cannot be reused";
} else if (!NodeCanReused(out_node)) {
can_replace = false;
VLOG(4) << "SKIP: Output variable " << out_var_name
<< " cannot be reused";
} else if (details::NodeSize(*in_node->Var()) !=
details::NodeSize(*out_node->Var())) {
can_replace = false;
VLOG(4) << "SKIP: Input and Output varialbe size not match";
}
if (!can_replace) continue;
// 2. there is no external pending op on the input node
if (view_.PendingOpsOnVar(in_node).size() > 1) {
// if (view_.PendingOpsOnVar(in_node).size() > 1) {
if (in_node->outputs.size() > 1 && !view_.CheckDeps(in_node, op)) {
VLOG(4) << string::Sprintf(
"Skiped pair %s => %s. %s input has external dependency."
"inplace such pair will overwrite the memory.",
......@@ -342,6 +377,97 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
}
}
void GraphView::TopoSort(ir::Graph* graph) {
//
ops_.clear();
auto deps_num = [](ir::Node* op) {
auto cnt = 0;
for (auto& var : op->inputs)
if (var->inputs.size() > 0) ++cnt;
return cnt;
};
std::queue<std::pair<ir::Node*, uint32_t>> ready_ops;
int level = 0;
auto nodes = graph->Nodes();
std::unordered_map<ir::Node*, uint32_t> deps_map;
for (auto& node : nodes) {
if (node->IsOp() && node->Op() != nullptr) {
deps_map[node] = deps_num(node);
if (0 == deps_map[node]) {
ready_ops.push({node, level});
}
}
}
while (!ready_ops.empty()) {
auto item = ready_ops.front();
ready_ops.pop();
ops_.emplace_back(item.first);
// record level when pop from queue
op_level_[item.first] = item.second;
for (auto node : item.first->outputs) {
for (auto op : node->outputs) {
--deps_map[op];
if (deps_map[op] == 0) ready_ops.push({op, item.second + 1});
}
}
}
bool all_ops_checked = true;
for (auto& node : nodes) {
if (node->IsOp() && node->Op() != nullptr && deps_map[node] > 0) {
all_ops_checked = false;
break;
}
}
PADDLE_ENFORCE(all_ops_checked, "All ops deps should be 0 after analysis");
}
// return true if current op node depeneds on all other op that use the same
// variable node
bool GraphView::CheckDeps(ir::Node* var, ir::Node* current_op) const {
// get op list that rely on the same variable
auto op_list = var->outputs;
for (auto& op : op_list) {
if (op == current_op) continue;
VLOG(4) << " GraphView::CheckDeps : " << op->Name() << " & "
<< current_op->Name();
if (!CheckOpDeps(op, current_op)) return false;
VLOG(4) << "";
}
return true;
}
// check if op2 depends on op1's output
bool GraphView::CheckOpDeps(ir::Node* op1, ir::Node* op2) const {
auto print_op = [&](ir::Node* op, const char* name) {
std::ostringstream os;
os << " " << name << " : " << op->Name() << " ";
os << "Input args : ";
for (auto& arg : op->inputs) os << arg->Name() << " ";
os << "Output args : ";
for (auto& arg : op->outputs) os << arg->Name() << " ";
os << "Level : " << op_level_.at(op);
VLOG(4) << os.str();
};
print_op(op1, "OP1");
print_op(op2, "OP2");
if (op1 == op2) return true;
if (op_level_.at(op1) >= op_level_.at(op2)) return false;
for (auto& var : op2->inputs)
if (var->inputs.size() > 0 && CheckOpDeps(op1, var->inputs[0])) return true;
return false;
}
ir::Node* GraphView::GetNodeByName(const std::string& name,
const std::vector<ir::Node*>& nodes) const {
// nodes should be op->inputs/outputs
......@@ -387,22 +513,7 @@ void GraphView::Build(ir::Graph* g) {
// Because we insert some new created node. Which may have data race between
// nodes.
// resolve data harzards depends on the var nodes in right order.
ops_ = SortOpLikeDescOrder(*g);
// 1. track the nodes which reused previous node in Python memory optimize.
// these node can not be inplaced, otherwise may generate a circle in graph.
std::unordered_set<std::string> all_vars;
for (auto& node : g->Nodes()) {
if (node->IsVar()) continue;
for (auto& out : node->outputs) {
if (out->IsCtrlVar() || out->Var() == nullptr) continue;
if (all_vars.count(out->Name())) {
dup_nodes_.emplace(out->Name());
} else {
all_vars.emplace(out->Name());
}
}
}
TopoSort(g);
// 2. track the nodes which used by parameter server.
// these node can not be inplaced, otherwise trainer
......
paddle/fluid/framework/details/inplace_op_pass.h
浏览文件 @ cc29bec6
......@@ -14,6 +14,7 @@
#pragma once
#include <map>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
......@@ -50,10 +51,15 @@ class GraphView {
// map the parameter and gradient, must be skipped.
bool InSkipSet(const std::string& var) const;
bool CheckDeps(ir::Node* var, ir::Node* current_op) const;
bool CheckOpDeps(ir::Node* op1, ir::Node* op2) const;
void TopoSort(ir::Graph* g);
private:
std::vector<ir::Node*> ops_;
std::unordered_set<std::string> dup_nodes_; // mem opt affect nodes
std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
std::unordered_map<ir::Node*, uint32_t> op_level_;
};
// swap pairs in sequence
......@@ -63,8 +69,7 @@ class InplacePass : public ir::Pass {
InplacePass();
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
void InitSSAGraphNodes() const;
......
paddle/fluid/framework/details/memory_optimize_helper.cc
浏览文件 @ cc29bec6
......@@ -190,7 +190,7 @@ struct NodeComparator {
auto rhs_shape = rhs_desc->GetShape();
if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) ||
(lhs_shape[0] != -1 && rhs_shape[0] != -1)) {
return NodeSize(lhs) <= NodeSize(rhs);
return NodeSize(lhs) == NodeSize(rhs);
} else {
return false;
}
......@@ -449,6 +449,7 @@ void ControlFlowGraph::LiveVariableAnalysis() {
live_in_[op].insert(var);
}
for (auto& var : defs_[op]) {
if (uses_[op].count(var)) continue;
live_in_[op].erase(var);
}
......
paddle/fluid/framework/details/memory_optimize_helper_test.cc
浏览文件 @ cc29bec6
......@@ -142,15 +142,16 @@ TEST(OrderedSet, FindBestFitNode) {
for (auto& node : nodes) {
pool.Insert(node.get());
}
// FIXME(liuwei1031) this API has changed,
// disable these tests temporarily
// FindNextBestFitNode
auto* n = nodes[0].get();
auto* cache = pool.FindBestFitNode(n);
PADDLE_ENFORCE(cache->Name() == "a");
cache = pool.FindNextBestFitNode(n, cache);
PADDLE_ENFORCE(cache->Name() == "c");
cache = pool.FindNextBestFitNode(n, cache);
PADDLE_ENFORCE(cache->Name() == "b");
// auto* n = nodes[0].get();
// auto* cache = pool.FindBestFitNode(n);
// PADDLE_ENFORCE(cache->Name() == "a");
// cache = pool.FindNextBestFitNode(n, cache);
// PADDLE_ENFORCE(cache->Name() == "c");
// cache = pool.FindNextBestFitNode(n, cache);
// PADDLE_ENFORCE(cache->Name() == "b");
}
} // namespace details
......
paddle/fluid/framework/details/memory_optimize_pass.cc
浏览文件 @ cc29bec6
......@@ -44,8 +44,7 @@ namespace paddle {
namespace framework {
namespace details {
std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void MemoryOptimizePass::ApplyImpl(ir::Graph* graph) const {
auto nodes = graph->Nodes();
CollectSkipVarsSet(nodes);
......@@ -113,7 +112,7 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
cfg_->RenameVarInCFGGraph(var_name, cache_name, idx);
RenameVarInGraphDesc(var_name, cache_name, idx);
RenameVarInGraphNode(var_name, cache_name, idx, graph.get());
RenameVarInGraphNode(var_name, cache_name, idx, graph);
pool_.Erase(cache_name);
}
}
......@@ -128,8 +127,6 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
}
}
graph->ResolveHazard(var_nodes_);
return graph;
}
void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
......
paddle/fluid/framework/details/memory_optimize_pass.h
浏览文件 @ cc29bec6
......@@ -21,6 +21,7 @@
#include <set>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
......@@ -35,8 +36,7 @@ namespace details {
class MemoryOptimizePass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
// fill the variable map(var_nodes) by version.
void InitSSAGraphNodes() const;
......
paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc
浏览文件 @ cc29bec6
......@@ -34,8 +34,7 @@ static bool IsLockAndRecordEventFreeComputationOpHandle(
return true;
}
std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
std::unique_ptr<ir::Graph> ir_graph) const {
void ModifyOpLockAndRecordEventPass::ApplyImpl(ir::Graph *ir_graph) const {
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*ir_graph);
OpGraphView graph_view(all_ops);
for (auto &op : all_ops) {
......@@ -49,7 +48,6 @@ std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
<< compute_op->DebugString();
}
}
return ir_graph;
}
} // namespace details
......
paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h
浏览文件 @ cc29bec6
......@@ -23,8 +23,7 @@ namespace details {
class ModifyOpLockAndRecordEventPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace details
......
paddle/fluid/framework/details/multi_devices_graph_check_pass.cc
浏览文件 @ cc29bec6
......@@ -23,10 +23,8 @@ namespace details {
class SSAGraghBuilderWithChecker : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
PADDLE_ENFORCE(IsValidGraph(graph.get()));
return graph;
void ApplyImpl(ir::Graph *graph) const override {
PADDLE_ENFORCE(IsValidGraph(graph));
}
bool IsValidGraph(const ir::Graph *graph) const {
......
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ cc29bec6
......@@ -23,6 +23,7 @@
#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/fetch_barrier_op_handle.h"
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
#include "paddle/fluid/framework/details/reduce_op_handle.h"
#include "paddle/fluid/framework/details/rpc_op_handle.h"
......@@ -31,6 +32,7 @@
#include "paddle/fluid/framework/ir/node.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace framework {
......@@ -151,8 +153,7 @@ void MultiDevSSAGraphBuilderBase::Init() const {
PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
}
std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void MultiDevSSAGraphBuilderBase::ApplyImpl(ir::Graph *graph) const {
Init();
CheckGraph(*graph);
std::vector<ir::Node *> sorted_ops = SortOperations(*graph);
......@@ -208,7 +209,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
for (size_t i = 0; i < backward_vars.size(); i += 2) {
auto &p_name = backward_vars[i];
auto &g_name = backward_vars[i + 1];
VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
VLOG(10) << "Bcast " << g_name << " for parameter " << p_name
<< " op_type " << node->Op()->Type();
if (NeedCollectiveForGrad(g_name, sorted_ops)) {
InsertCollectiveOp(&result, p_name, g_name);
}
......@@ -233,7 +235,6 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
AddOutputToLeafOps(&result);
result.Erase(kGraphOps);
return graph;
}
void MultiDevSSAGraphBuilderBase::InsertScaleLossGradOp(
......@@ -413,8 +414,9 @@ void MultiDevSSAGraphBuilderBase::CreateComputationalOp(ir::Graph *result,
CreateOpHandleIOs(result, node, dev_id);
}
void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
ir::Graph *result, const std::string &og) const {
void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(ir::Graph *result,
const std::string &og,
bool is_encoded) const {
OpHandleBase *op_handle = nullptr;
auto append_allreduce_op = [&](
......@@ -423,7 +425,9 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places, nccl_ctxs_));
scopes, places, nccl_ctxs_, is_encoded,
static_cast<int>(strategy_.trainers_endpoints_.size()) *
places_.size()));
#else
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
......@@ -445,12 +449,15 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
op_handle->AddInput(prev_grad);
VLOG(10) << "all_reduce_op_handle add input " << prev_grad->DebugString();
auto var =
new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
vars.size(), i, og, places_[i]);
vars.emplace_back(var);
op_handle->AddOutput(var);
VLOG(10) << "all_reduce_op_handle add output " << og
<< ", handle:" << var->DebugString();
}
}
......@@ -851,9 +858,17 @@ int DistSSAGraphBuilder::CreateRPCOp(ir::Graph *result, ir::Node *node) const {
PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
node->Op()->Type());
result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
node->Op()->Type(), places_[op_dev_id]));
// Create fetch_barrier op handle to enable output on all devices.
// **NOTE** fetch_barrier should output variables list same as recv op does.
if (node->Op()->Type() == "fetch_barrier") {
result->Get<GraphOps>(kGraphOps).emplace_back(new FetchBarrierOpHandle(
result->CreateOpNode(node->Op()), local_scopes_, places_));
} else {
result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
node->Op()->Type(), places_[op_dev_id]));
}
if (node->Op()->Type() == "send") {
CreateOpHandleIOs(result, node, op_dev_id);
......@@ -932,6 +947,17 @@ int DistSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
return op_dev_id;
}
bool DistSSAGraphBuilder::IsEncoded(const std::string &p_name) const {
auto u_name = p_name + "__dgc_u__";
auto it = all_vars_.find(u_name);
if (it == all_vars_.end()) {
VLOG(10) << "can't find u_name, so it's not encoded:" << u_name;
return false;
}
return true;
}
void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
const std::string &p_name,
const std::string &g_name) const {
......@@ -947,7 +973,11 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
CreateReduceOp(result, g_name, 0);
CreateBroadcastOp(result, g_name, 0);
} else {
CreateAllReduceOp(result, g_name);
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
CreateAllReduceOp(result, g_name, IsEncoded(p_name));
#else
PADDLE_ENFORCE(false, "Compiled withoud cuda!");
#endif
}
break;
default:
......
paddle/fluid/framework/details/multi_devices_graph_pass.h
浏览文件 @ cc29bec6
......@@ -36,8 +36,7 @@ namespace details {
class MultiDevSSAGraphBuilderBase : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph *graph) const override;
virtual void Init() const;
......@@ -75,7 +74,8 @@ class MultiDevSSAGraphBuilderBase : public ir::Pass {
bool IsSparseGradient(const std::string &og) const;
void CreateAllReduceOp(ir::Graph *result, const std::string &og) const;
void CreateAllReduceOp(ir::Graph *result, const std::string &og,
bool is_encoded = false) const;
void CreateBroadcastOp(ir::Graph *result, const std::string &p_name,
size_t src_dev_id) const;
......@@ -171,6 +171,8 @@ class DistSSAGraphBuilder : public BalanceVarSSAGraphBuilder {
mutable std::vector<std::unordered_set<std::string>> bcast_var_name_set_;
mutable bool need_broadcast_var_{false};
bool IsEncoded(const std::string &p_name) const;
};
std::unordered_set<std::string> &MultiDevSSAGraphBuilder();
......
paddle/fluid/framework/details/multi_devices_graph_print_pass.cc
浏览文件 @ cc29bec6
......@@ -13,7 +13,9 @@
// limitations under the License.
#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
#include <memory>
#include <string>
#include <unordered_map>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
......
paddle/fluid/framework/details/multi_devices_graph_print_pass.h
浏览文件 @ cc29bec6
......@@ -17,6 +17,7 @@
#include <glog/logging.h>
#include <fstream>
#include <iosfwd>
#include <memory>
#include <ostream>
#include <string>
#include "paddle/fluid/framework/details/multi_devices_helper.h"
......@@ -40,13 +41,11 @@ class GraphvizSSAGraphPrinter : public SSAGraphPrinter {
class SSAGraghBuilderWithPrinter : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
void ApplyImpl(ir::Graph* graph) const override {
std::unique_ptr<std::ostream> fout(
new std::ofstream(Get<std::string>(kGraphvizPath)));
PADDLE_ENFORCE(fout->good());
Get<GraphvizSSAGraphPrinter>("graph_printer").Print(*graph, *fout);
return graph;
}
};
......
paddle/fluid/framework/details/op_handle_base.cc
浏览文件 @ cc29bec6
......@@ -55,7 +55,7 @@ void OpHandleBase::Run(bool use_cuda) {
if (out_var_handle) {
int dev_id =
boost::get<platform::CUDAPlace>(out_var_handle->place()).device;
out_var_handle->SetGenerateEvent(events_[dev_id]);
out_var_handle->SetGenerateEvent(events_.at(dev_id));
}
}
} else {
......@@ -71,7 +71,7 @@ void OpHandleBase::Run(bool use_cuda) {
"The place of input(%s) is not consistent with the "
"place of current op(%s).",
out_var_handle->Name(), Name());
out_var_handle->SetGenerateEvent(events_[dev_id]);
out_var_handle->SetGenerateEvent(events_.at(dev_id));
}
}
}
......
paddle/fluid/framework/details/op_registry.h
浏览文件 @ cc29bec6
......@@ -21,6 +21,7 @@ limitations under the License. */
#include <vector>
#include "paddle/fluid/framework/grad_op_desc_maker.h"
#include "paddle/fluid/framework/inplace_op_inference.h"
#include "paddle/fluid/framework/no_need_buffer_vars_inference.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/framework/operator.h"
......@@ -36,27 +37,86 @@ enum OpInfoFillType {
kGradOpDescMaker = 2,
kVarTypeInference = 3,
kShapeInference = 4,
kInplaceOpInference = 5
kInplaceOpInference = 5,
kNoNeedBufferVarsInference = 6,
kUnknown = -1
};
namespace internal {
template <typename T, OpInfoFillType kType>
struct TypePair {
using Type = T;
static constexpr OpInfoFillType kFillType = kType;
};
using OpRegistryClasses = std::tuple< // NOLINT
TypePair<OperatorBase, kOperator>, // NOLINT
TypePair<OpProtoAndCheckerMaker, kOpProtoAndCheckerMaker>, // NOLINT
TypePair<GradOpDescMakerBase, kGradOpDescMaker>, // NOLINT
TypePair<VarTypeInference, kVarTypeInference>, // NOLINT
TypePair<InferShapeBase, kShapeInference>, // NOLINT
TypePair<InplaceOpInference, kInplaceOpInference>, // NOLINT
TypePair<NoNeedBufferVarsInference, kNoNeedBufferVarsInference> // NOLINT
>;
static constexpr int kOpRegistryClassNumber =
std::tuple_size<OpRegistryClasses>::value;
template <typename T, int kPos, bool kIsBounded /* = true*/>
struct IsMatchedBaseTypeImpl {
using PairType = typename std::tuple_element<kPos, OpRegistryClasses>::type;
static constexpr bool kValue =
std::is_base_of<typename PairType::Type, T>::value;
};
template <typename T, int kPos>
struct IsMatchedBaseTypeImpl<T, kPos, false> {
static constexpr bool kValue = false;
};
template <typename T, int kPos>
static inline constexpr bool IsMatchedBaseType() {
return IsMatchedBaseTypeImpl<
T, kPos, (kPos >= 0 && kPos < kOpRegistryClassNumber)>::kValue;
}
template <typename T, int kStart, int kEnd, bool kIsEnd, bool kIsMatched>
struct OpInfoFillTypeGetterImpl {};
// This case should not happen
template <typename T, int kStart, int kEnd>
struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, true, true> {};
template <typename T, int kStart, int kEnd>
struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, true, false> {
static constexpr OpInfoFillType kType = kUnknown;
};
template <typename T, int kStart, int kEnd>
struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, false, false> {
static constexpr OpInfoFillType kType =
OpInfoFillTypeGetterImpl<T, kStart + 1, kEnd, kStart + 1 == kEnd,
IsMatchedBaseType<T, kStart + 1>()>::kType;
};
template <typename T, int kStart, int kEnd>
struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, false, true> {
using PairType = typename std::tuple_element<kStart, OpRegistryClasses>::type;
static constexpr OpInfoFillType kType = PairType::kFillType;
};
template <typename T>
using OpInfoFillTypeGetter =
OpInfoFillTypeGetterImpl<T, 0, kOpRegistryClassNumber,
kOpRegistryClassNumber == 0,
IsMatchedBaseType<T, 0>()>;
} // namespace internal
template <typename T>
struct OpInfoFillTypeID {
static constexpr OpInfoFillType ID() {
return std::is_base_of<OperatorBase, T>::value
? kOperator
: (std::is_base_of<OpProtoAndCheckerMaker, T>::value
? kOpProtoAndCheckerMaker
: (std::is_base_of<GradOpDescMakerBase, T>::value
? kGradOpDescMaker
: (std::is_base_of<VarTypeInference, T>::value
? kVarTypeInference
: (std::is_base_of<InferShapeBase, T>::value
? kShapeInference
: (std::is_base_of<
InplaceOpInference, T>::value
? kInplaceOpInference
: static_cast<OpInfoFillType>(
-1))))));
return internal::OpInfoFillTypeGetter<T>::kType;
}
};
......@@ -149,9 +209,21 @@ struct OpInfoFiller<T, kShapeInference> {
template <typename T>
struct OpInfoFiller<T, kInplaceOpInference> {
void operator()(const char* op_type, OpInfo* info) const {
info->infer_inplace_ = [](const OpDesc& op_desc, BlockDesc* block) {
info->infer_inplace_ = [](const OpDesc& op_desc) {
T infer;
return infer(op_desc, block);
return infer(op_desc);
};
}
};
template <typename T>
struct OpInfoFiller<T, kNoNeedBufferVarsInference> {
void operator()(const char* op_type, OpInfo* info) const {
info->infer_no_need_buffer_vars_ = [](const VariableNameMap& inputs,
const VariableNameMap& outputs,
const AttributeMap& attrs) {
T infer(inputs, outputs, attrs);
return infer();
};
}
};
......
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
浏览文件 @ cc29bec6
......@@ -96,7 +96,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
auto seq_allreduce_pass =
ir::PassRegistry::Instance().Get("all_reduce_deps_pass");
for (size_t i = 0; i < graphs_.size(); ++i) {
graphs_[i] = seq_allreduce_pass->Apply(std::move(graphs_[i]));
graphs_[i].reset(seq_allreduce_pass->Apply(graphs_[i].release()));
}
// set the correct size of thread pool to each device.
......
paddle/fluid/framework/details/reference_count_pass.cc
浏览文件 @ cc29bec6
......@@ -193,8 +193,80 @@ ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx,
return shrink_func(computation_op);
}
std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
/**
* Shrink op dependencies according to no need buffer vars.
*
* If some ops do not need Tensor buffer of any input,
* just remove the dependency of this op, i.e, decrease reference count.
*
* For example, input Y of elementwise_add_grad op is only used to infer shape
* and lod of Y@GRAD, we do not need the buffer of input Y. Data buffer of
* input Y can be collected before elementwise_add_grad op runs.
*
* This method returns whether the dependency count decreases to 0, and
* shrinks op dependency if possible.
*/
static bool ShrinkNoNeedBufferVarOpDependency(
const std::string &var_name,
std::unordered_set<ComputationOpHandle *> *op_handles) {
std::vector<ComputationOpHandle *> skip_ops;
for (auto *op_handle : *op_handles) {
auto *op_base = op_handle->GetOp();
auto &inferer = op_base->Info().NoNeedBufferVarsInferer();
if (!inferer) {
continue;
}
std::unordered_set<std::string> no_need_buffer_vars =
inferer(op_base->Inputs(), op_base->Outputs(), op_base->Attrs());
// Check whether var_name occurs in other inputs or outputs of the op
// If it occurs, we cannot decrease the dependency number.
bool occurred_in_other_vars = false;
for (auto &in_pair : op_base->Inputs()) {
if (no_need_buffer_vars.count(in_pair.first) > 0) {
continue;
}
auto &args = in_pair.second;
auto iter = std::find(args.begin(), args.end(), var_name);
if (iter != args.end()) {
occurred_in_other_vars = true;
break;
}
}
if (occurred_in_other_vars) {
continue;
}
for (auto &out_pair : op_base->Outputs()) {
auto &args = out_pair.second;
auto iter = std::find(args.begin(), args.end(), var_name);
if (iter != args.end()) {
occurred_in_other_vars = true;
break;
}
}
if (!occurred_in_other_vars) {
VLOG(2) << "Shrink var " << var_name << " in op " << op_handle->Name();
skip_ops.emplace_back(op_handle);
}
}
if (skip_ops.size() == op_handles->size()) {
op_handles->clear();
return true;
} else {
for (auto *skip_op : skip_ops) {
op_handles->erase(skip_op);
}
return false;
}
}
void ReferenceCountPass::ApplyImpl(ir::Graph *graph) const {
auto &ref_cnts = Get<std::vector<ReferenceCountMap>>(kGlobalReferenceCount);
auto &last_live_ops_of_vars =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
......@@ -229,21 +301,46 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
continue;
}
bool ok;
auto result = ExtractComputationOpFromLastLivedVar(
name_var_pair.second.back(), i, shrink_func, &ok);
auto &var_name = name_var_pair.first;
auto &var_handles = name_var_pair.second;
for (auto iter = var_handles.rbegin(); iter != var_handles.rend();
++iter) {
bool ok;
auto result =
ExtractComputationOpFromLastLivedVar(*iter, i, shrink_func, &ok);
// Seldomly, some vars may have no pending or preceding computation ops
// Just break;
if (!ok) break;
VLOG(10) << "Extract " << result.size() << " ops of var " << var_name;
size_t original_op_deps = result.size();
// If all ops do not need buffer of var_name, calculate reference count
// of the previous version of var_name.
if (ShrinkNoNeedBufferVarOpDependency(var_name, &result)) {
VLOG(10) << "Try to precede reference count computing at var "
<< var_name;
continue;
}
size_t final_op_deps = result.size();
if (final_op_deps < original_op_deps) {
VLOG(5) << "Shrink op deps from " << original_op_deps << " to "
<< final_op_deps;
}
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));
break;
}
// Seldomly, all preceding trying failed.
// Just skip this corner case
}
}
return graph;
}
} // namespace details
......
paddle/fluid/framework/details/reference_count_pass.h
浏览文件 @ cc29bec6
......@@ -23,8 +23,7 @@ namespace details {
class ReferenceCountPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace details
......
paddle/fluid/framework/details/sequential_execution_pass.cc
浏览文件 @ cc29bec6
......@@ -29,8 +29,7 @@ static bool IsSameOpDesc(OpDesc *op1, OpDesc *op2) {
op1->Outputs() == op2->Outputs();
}
std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void SequentialExecutionPass::ApplyImpl(ir::Graph *graph) const {
// FIXME(zjl): Insert dependencies between some distributed ops may cause
// the multi_devices_graph_pass fails. So we skip these ops here.
// Indeed, maybe we should not insert dependencies between these ops
......@@ -98,7 +97,6 @@ std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
VLOG(10) << "Add dependencies between " << op_node_list[i - 1]->Name()
<< " and " << op_node_list[i]->Name();
}
return graph;
}
} // namespace details
......
paddle/fluid/framework/details/sequential_execution_pass.h
浏览文件 @ cc29bec6
......@@ -23,8 +23,7 @@ namespace details {
class SequentialExecutionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace details
......
paddle/fluid/framework/details/var_handle.cc
浏览文件 @ cc29bec6
......@@ -24,7 +24,8 @@ VarHandle::~VarHandle() { VLOG(4) << "deleting var handle " << DebugString(); }
std::string VarHandle::DebugString() const {
std::stringstream ss;
ss << name_ << ":" << place_;
ss << "name:" << name_ << ", place:" << place_ << ", version:" << version_
<< ", scope_idx:" << scope_idx_;
return ss.str();
}
......
paddle/fluid/framework/details/while_op_eager_deletion_pass.cc
浏览文件 @ cc29bec6
......@@ -23,8 +23,7 @@ namespace details {
class WhileOpEagerDeletionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
void ApplyImpl(ir::Graph *graph) const override {
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
// Find all while_op and while_grad_op
......@@ -50,7 +49,6 @@ class WhileOpEagerDeletionPass : public ir::Pass {
operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(
while_ops, while_grad_ops);
}
return graph;
}
};
......
paddle/fluid/framework/executor.cc
浏览文件 @ cc29bec6
......@@ -19,6 +19,7 @@ limitations under the License. */
#include <unordered_set>
#include <utility>
#include "paddle/fluid/framework/executor_gc_helper.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
......@@ -48,97 +49,23 @@ 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 std::vector<std::string>& keep_vars, bool force_disable_gc)
: prog_(prog), block_id_(block_id), force_disable_gc_(force_disable_gc) {
if (GetEagerDeletionThreshold() >= 0 && !force_disable_gc_) {
global_ref_cnts_ =
GetNonPersistableReferenceCounts(prog.Block(block_id), keep_vars);
const framework::ProgramDesc& prog, size_t block_id)
: prog_(prog), block_id_(block_id) {}
void ExecutorPrepareContext::PrepareUnusedVars(
const std::vector<std::string>& keep_vars, bool force_disable_gc) {
force_disable_gc_ = force_disable_gc;
if (GetEagerDeletionThreshold() < 0 || force_disable_gc_) {
return;
}
unused_vars_ = GetUnusedVars(prog_.Block(block_id_), ops_, keep_vars);
}
ExecutorPrepareContext::~ExecutorPrepareContext() {
VLOG(5) << "destroy ExecutorPrepareContext";
}
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) != 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",
framework::ToTypeName(var->Type()), name);
}
}
}
};
handler(op->Inputs());
handler(op->Outputs());
if (!garbages.empty()) {
gc->Add(std::move(garbages));
}
}
Executor::Executor(const platform::Place& place) : place_(place) {}
void Executor::Close() {
......@@ -362,8 +289,8 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars, bool force_disable_gc) {
std::unique_ptr<ExecutorPrepareContext> ctx(new ExecutorPrepareContext(
program, block_id, skip_ref_cnt_vars, force_disable_gc));
std::unique_ptr<ExecutorPrepareContext> ctx(
new ExecutorPrepareContext(program, block_id));
PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
auto& block = program.Block(block_id);
for (auto& op_desc : block.AllOps()) {
......@@ -375,6 +302,7 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
ctx->prog_.Block(ctx->block_id_), &ctx->ops_);
}
#endif
ctx->PrepareUnusedVars(skip_ref_cnt_vars, force_disable_gc);
return ctx;
}
......@@ -389,19 +317,17 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
std::vector<std::shared_ptr<ExecutorPrepareContext>> result;
size_t idx = 0;
for (auto& bid : block_ids) {
ExecutorPrepareContext* ctx;
if (skip_ref_cnt_vars.empty()) {
ctx = new ExecutorPrepareContext(program, bid, std::vector<std::string>(),
force_disable_gc);
} else {
ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx],
force_disable_gc);
}
PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
auto* ctx = new ExecutorPrepareContext(program, bid);
auto& block = program.Block(bid);
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
if (skip_ref_cnt_vars.empty()) {
ctx->PrepareUnusedVars(std::vector<std::string>(), force_disable_gc);
} else {
ctx->PrepareUnusedVars(skip_ref_cnt_vars[idx], force_disable_gc);
}
result.push_back(std::shared_ptr<ExecutorPrepareContext>(ctx));
++idx;
}
......@@ -425,7 +351,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
// FIXME(zjl): recurrent_op is rather complex, we would
// disable gc forcely in recurrent_op
if (!ctx->force_disable_gc_ && max_memory_size >= 0) {
ctx->ResetReferenceCount();
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place_)) {
if (IsFastEagerDeletionModeEnabled()) {
......@@ -453,8 +378,7 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
op->Run(*local_scope, place_);
if (gc) {
DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
&(ctx->runtime_ref_cnts_));
DeleteUnusedTensors(*local_scope, op.get(), ctx->unused_vars_, gc.get());
}
}
......
paddle/fluid/framework/executor.h
浏览文件 @ cc29bec6
......@@ -30,22 +30,20 @@ namespace paddle {
namespace framework {
struct ExecutorPrepareContext {
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>(),
bool force_disable_gc = false);
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
~ExecutorPrepareContext();
void ResetReferenceCount() { runtime_ref_cnts_ = global_ref_cnts_; }
void PrepareUnusedVars(const std::vector<std::string>& keep_vars,
bool force_disable_gc = false);
const framework::ProgramDesc& prog_;
size_t block_id_;
bool force_disable_gc_;
const size_t block_id_;
std::vector<std::unique_ptr<OperatorBase>> ops_;
std::unordered_map<std::string, size_t> global_ref_cnts_;
std::unordered_map<std::string, size_t> runtime_ref_cnts_;
std::unordered_map<OperatorBase*, std::vector<std::string>> unused_vars_;
bool force_disable_gc_{false};
};
class Executor {
......
paddle/fluid/framework/executor_gc_helper.cc 0 → 100644
浏览文件 @ cc29bec6
// Copyright (c) 2019 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/executor_gc_helper.h"
#include <deque>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace framework {
struct OpInOutInfo {
public:
void Build(const OperatorBase *op) {
is_built_ = true;
auto &inferer = op->Info().NoNeedBufferVarsInferer();
if (inferer) {
no_need_buffer_ins_ = inferer(op->Inputs(), op->Outputs(), op->Attrs());
if (no_need_buffer_ins_.empty()) return;
for (auto &in_name_pair : op->Inputs()) {
if (no_need_buffer_ins_.count(in_name_pair.first) != 0) {
continue;
}
for (auto &in_arg_name : in_name_pair.second) {
other_args_set_.insert(in_arg_name);
}
}
for (auto &out_name_pair : op->Outputs()) {
for (auto &out_arg_name : out_name_pair.second) {
other_args_set_.insert(out_arg_name);
}
}
}
}
bool IsBuilt() const { return is_built_; }
bool IsInArgBufferNeeded(const std::string &in_arg_name) const {
return no_need_buffer_ins_.empty() ||
other_args_set_.count(in_arg_name) != 0;
}
private:
// A set to record unused buffer input vars of op
std::unordered_set<std::string> no_need_buffer_ins_;
// A set to record other args of op (including in, out)
std::unordered_set<std::string> other_args_set_;
bool is_built_{false};
};
static bool VarCanBeDeleted(const std::string &name, const BlockDesc &block,
const std::unordered_set<std::string> &skip_vars) {
if (skip_vars.count(name) != 0) {
return false;
}
auto *var_desc = block.FindVar(name);
if (var_desc == nullptr || var_desc->Persistable()) {
return false;
}
auto type = var_desc->Proto()->type().type();
return type == proto::VarType::LOD_TENSOR ||
type == proto::VarType::SELECTED_ROWS ||
type == proto::VarType::LOD_TENSOR_ARRAY;
}
std::unordered_map<OperatorBase *, std::vector<std::string>> GetUnusedVars(
const BlockDesc &block,
const std::vector<std::unique_ptr<OperatorBase>> &ops,
const std::vector<std::string> &skip_var_list) {
std::unordered_set<std::string> skip_vars(skip_var_list.begin(),
skip_var_list.end());
std::unordered_map<std::string, size_t> var_op_idx_map;
for (size_t i = 0; i < ops.size(); ++i) {
auto *op = ops[i].get();
OpInOutInfo info;
for (auto &name_pair : op->Inputs()) {
for (auto &name : name_pair.second) {
if (!VarCanBeDeleted(name, block, skip_vars)) {
continue;
}
// var can be gc-ed
if (!info.IsBuilt()) {
info.Build(op);
}
if (info.IsInArgBufferNeeded(name)) {
// Update the last living op of variable to current op
var_op_idx_map[name] = i;
} else {
VLOG(10) << "Skip reference count computing of variable "
<< name_pair.first << "(" << name << ") in Operator "
<< op->Type();
}
}
}
for (auto &name_pair : op->Outputs()) {
for (auto &name : name_pair.second) {
if (VarCanBeDeleted(name, block, skip_vars)) {
// Update the last living op of variable to current op
var_op_idx_map[name] = i;
}
}
}
}
std::unordered_map<OperatorBase *, std::vector<std::string>> result;
for (auto &name_op_idx_pair : var_op_idx_map) {
auto &name = name_op_idx_pair.first;
size_t op_idx = name_op_idx_pair.second;
result[ops[op_idx].get()].emplace_back(name);
}
return result;
}
void DeleteUnusedTensors(
const Scope &scope, OperatorBase *op,
const std::unordered_map<OperatorBase *, std::vector<std::string>>
&delete_vars_map,
GarbageCollector *gc) {
auto iter = delete_vars_map.find(op);
if (iter == delete_vars_map.end()) {
return;
}
auto &delete_vars = iter->second;
std::deque<std::shared_ptr<memory::Allocation>> garbages;
for (auto &var_name : delete_vars) {
auto *var = scope.FindVar(var_name);
if (var == nullptr) {
continue;
}
VLOG(2) << "Erase variable " << var_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",
framework::ToTypeName(var->Type()), var_name);
}
}
if (!garbages.empty()) {
gc->Add(std::move(garbages));
}
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/executor_gc_helper.h 0 → 100644
浏览文件 @ cc29bec6
// Copyright (c) 2019 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 <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/scope.h"
namespace paddle {
namespace framework {
// Result map: op -> variable names that can be deleted after op runs
std::unordered_map<OperatorBase *, std::vector<std::string>> GetUnusedVars(
const BlockDesc &block,
const std::vector<std::unique_ptr<OperatorBase>> &ops,
const std::vector<std::string> &skip_vars);
// Collect unused tensors after op runs
void DeleteUnusedTensors(
const Scope &scope, OperatorBase *op,
const std::unordered_map<OperatorBase *, std::vector<std::string>>
&delete_vars_map,
GarbageCollector *gc);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/garbage_collector.cc
浏览文件 @ cc29bec6
......@@ -13,14 +13,36 @@
// limitations under the License.
#include <algorithm>
#include <deque>
#include <functional>
#include <memory>
#include <mutex> // NOLINT
#include <utility>
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
#include "gflags/gflags.h"
#include "glog/logging.h"
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
DEFINE_double(
eager_delete_tensor_gb, -1.0,
"Memory size threshold (GB) when the garbage collector clear tensors."
"Disabled when this value is less than 0");
DEFINE_bool(fast_eager_deletion_mode, true,
"Fast eager deletion mode. If enabled, memory would release "
"immediately without waiting GPU kernel ends.");
DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
"Fraction of eager deletion. If less than 1.0, all variables in "
"the program would be sorted according to its memory size, and "
"only the FLAGS_memory_fraction_of_eager_deletion of the largest "
"variables would be deleted.");
GarbageCollector::GarbageCollector(const platform::Place &place,
size_t max_memory_size)
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
......@@ -85,5 +107,25 @@ void StreamGarbageCollector::ClearCallback(
callback_manager_->AddCallback(callback);
}
#endif
int64_t GetEagerDeletionThreshold() {
return FLAGS_eager_delete_tensor_gb < 0
? -1
: static_cast<int64_t>(FLAGS_eager_delete_tensor_gb *
(static_cast<int64_t>(1) << 30));
}
bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; }
void SetEagerDeletionMode(double threshold, double fraction, bool fast_mode) {
FLAGS_eager_delete_tensor_gb = threshold;
FLAGS_memory_fraction_of_eager_deletion = fraction;
FLAGS_fast_eager_deletion_mode = fast_mode;
}
double GetEagerDeletionMemoryFraction() {
return FLAGS_memory_fraction_of_eager_deletion;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/garbage_collector.h
浏览文件 @ cc29bec6
......@@ -18,6 +18,8 @@
#include <functional>
#include <memory>
#include <mutex> // NOLINT
#include <utility>
#include "gflags/gflags.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
......@@ -126,5 +128,12 @@ void GarbageCollector::Add(Container &&objs, Callback &&callback) {
}
}
int64_t GetEagerDeletionThreshold();
bool IsFastEagerDeletionModeEnabled();
void SetEagerDeletionMode(double threshold, double fraction, bool fast_mode);
double GetEagerDeletionMemoryFraction();
} // namespace framework
} // namespace paddle
paddle/fluid/framework/inplace_op_inference.h
浏览文件 @ cc29bec6
......@@ -17,8 +17,8 @@
#include <numeric>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include "glog/logging.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/type_defs.h"
......@@ -32,55 +32,22 @@ namespace framework {
then Out will inplaced use X's memory. The base class will do
legality validation for both variables.
*/
class InplaceOpInference {
public:
virtual ~InplaceOpInference() {}
virtual std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc, BlockDesc* block) const = 0;
};
class InplaceInToOut : public InplaceOpInference {
public:
std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc, BlockDesc* block) const {
std::unordered_map<std::string, std::string> ret;
auto in_out_var_names_pair = this->Apply(op_desc, block);
for (auto& pair : in_out_var_names_pair) {
PADDLE_ENFORCE(!op_desc.Input(pair.first).empty(),
string::Sprintf("op %s do not have input of %s!",
op_desc.Type(), pair.first));
PADDLE_ENFORCE(!op_desc.Output(pair.second).empty(),
string::Sprintf("op %s do not have output of %s!",
op_desc.Type(), pair.second));
auto& in_name = op_desc.Input(pair.first).at(0);
auto& out_name = op_desc.Output(pair.second).at(0);
auto in = block->FindRecursiveOrCreateVar(in_name);
auto out = block->FindRecursiveOrCreateVar(out_name);
if (TryInplaceInputOutput(in, out)) ret.insert({in_name, out_name});
}
return ret;
}
protected:
virtual std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const = 0;
bool TryInplaceInputOutput(const VarDesc& in, const VarDesc& out) const {
return in.Name() != out.Name() && details::NodeCanReused(in) &&
details::NodeCanReused(out) &&
details::NodeSize(out) <= details::NodeSize(in);
}
const OpDesc& op_desc) const = 0;
};
/*
Inplace In and Out for operator only have an Input and an Output.
For example, activation op.
*/
class SingleOpInplaceInToOut : public InplaceInToOut {
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
class SingleOpInplaceInToOut : public InplaceOpInference {
public:
std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc) const override {
PADDLE_ENFORCE(!op_desc.InputNames().empty(),
"Op inputs must not be empty");
PADDLE_ENFORCE(!op_desc.OutputNames().empty(),
......@@ -95,10 +62,10 @@ class SingleOpInplaceInToOut : public InplaceInToOut {
Gradient op. Inplace output use it's Input.
For example, Input@Grad->Input reuse strategy.
*/
class GradOpInplaceInToOut : public InplaceInToOut {
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
class GradOpInplaceInToOut : public InplaceOpInference {
public:
std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc) const override {
std::unordered_map<std::string, std::string> ret;
std::unordered_set<std::string> output_names(op_desc.OutputNames().begin(),
op_desc.OutputNames().end());
......
paddle/fluid/framework/inplace_op_inference_test.cc
浏览文件 @ cc29bec6
......@@ -127,26 +127,20 @@ class MultiOutGradShapeInference : public framework::InferShapeBase {
}
};
class MultiOutInplaceInToOut : public framework::InplaceInToOut {
class MultiOutInplaceInToOut : public framework::InplaceOpInference {
public:
using framework::InplaceInToOut::InplaceInToOut;
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc) const override {
return std::unordered_map<std::string, std::string>{
{"X", "Out"}, {"Y", "YOut"}, {"Z", "ZOut"},
};
}
};
class MultiOutGradInplaceInToOut : public framework::InplaceInToOut {
class MultiOutGradInplaceInToOut : public framework::InplaceOpInference {
public:
using framework::InplaceInToOut::InplaceInToOut;
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc) const override {
return std::unordered_map<std::string, std::string>{
{framework::GradVarName("YOut"), framework::GradVarName("Y")},
{framework::GradVarName("Out"), framework::GradVarName("X")},
......@@ -171,118 +165,118 @@ REGISTER_OPERATOR(multi_out_grad, f::NOP, f::MultiOutGradInplaceInToOut,
namespace paddle {
namespace framework {
TEST(InferInplace, SingleOpInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("single_op");
op->SetInput("X", {"test2_a", "test2_b", "test2_c"});
op->SetOutput("Out", {"test2_out"});
prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128});
prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 1ul);
auto it = in_to_outs.begin();
EXPECT_EQ(it->first, "test2_a");
EXPECT_EQ(it->second, "test2_out");
}
TEST(InferInplace, SingleGradOpInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("single_op_grad");
op->SetInput(GradVarName("Out"), {"test2_out"});
op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"});
prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 1ul);
auto it = in_to_outs.begin();
EXPECT_EQ(it->first, "test2_out");
EXPECT_EQ(it->second, "test2_a");
}
TEST(InferInplace, MultiOutInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("multi_out_op");
op->SetInput("X", {"a0", "a1"});
op->SetInput("Y", {"b0"});
op->SetInput("Z", {"c0", "c1"});
op->SetOutput("Out", {"o0"});
op->SetOutput("YOut", {"y0"});
op->SetOutput("ZOut", {"z0"});
prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("o0");
prog.MutableBlock(0)->Var("y0");
prog.MutableBlock(0)->Var("z0");
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 3ul);
std::unordered_map<std::string, std::string> expects = {
{"a0", "o0"}, {"b0", "y0"}, {"c0", "z0"},
};
EXPECT_TRUE(expects == in_to_outs);
}
TEST(InferInplace, MultiGradInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("multi_out_grad");
op->SetInput(GradVarName("Out"), {"o0"});
op->SetInput(GradVarName("YOut"), {"y0"});
op->SetInput(GradVarName("ZOut"), {"z0"});
op->SetOutput(GradVarName("X"), {"a0", "a1"});
op->SetOutput(GradVarName("Y"), {"b0"});
op->SetOutput(GradVarName("Z"), {"c0", "c1"});
prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("o0");
prog.MutableBlock(0)->Var("y0");
prog.MutableBlock(0)->Var("z0");
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 3ul);
std::unordered_map<std::string, std::string> expects = {
{"o0", "a0"}, {"y0", "b0"}, {"z0", "c0"},
};
EXPECT_TRUE(expects == in_to_outs);
}
// TEST(InferInplace, SingleOpInplaceInToOut) {
// ProgramDesc prog;
// auto* op = prog.MutableBlock(0)->AppendOp();
// op->SetType("single_op");
// op->SetInput("X", {"test2_a", "test2_b", "test2_c"});
// op->SetOutput("Out", {"test2_out"});
//
// prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128});
// prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("test2_out");
// prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128});
//
// auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
// auto in_to_outs = infer_inplace(*op);
// EXPECT_EQ(in_to_outs.size(), 1ul);
// auto it = in_to_outs.begin();
// EXPECT_EQ(it->first, "test2_a");
// EXPECT_EQ(it->second, "test2_out");
// }
//
// TEST(InferInplace, SingleGradOpInplaceInToOut) {
// ProgramDesc prog;
// auto* op = prog.MutableBlock(0)->AppendOp();
// op->SetType("single_op_grad");
// op->SetInput(GradVarName("Out"), {"test2_out"});
// op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"});
//
// prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("test2_out");
// prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024});
//
// auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
// auto in_to_outs = infer_inplace(*op);
// EXPECT_EQ(in_to_outs.size(), 1ul);
// auto it = in_to_outs.begin();
// EXPECT_EQ(it->first, "test2_out");
// EXPECT_EQ(it->second, "test2_a");
// }
//
// TEST(InferInplace, MultiOutInplaceInToOut) {
// ProgramDesc prog;
// auto* op = prog.MutableBlock(0)->AppendOp();
// op->SetType("multi_out_op");
// op->SetInput("X", {"a0", "a1"});
// op->SetInput("Y", {"b0"});
// op->SetInput("Z", {"c0", "c1"});
// op->SetOutput("Out", {"o0"});
// op->SetOutput("YOut", {"y0"});
// op->SetOutput("ZOut", {"z0"});
//
// prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("o0");
// prog.MutableBlock(0)->Var("y0");
// prog.MutableBlock(0)->Var("z0");
// prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
//
// auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
// auto in_to_outs = infer_inplace(*op);
// EXPECT_EQ(in_to_outs.size(), 3ul);
// std::unordered_map<std::string, std::string> expects = {
// {"a0", "o0"}, {"b0", "y0"}, {"c0", "z0"},
// };
// EXPECT_TRUE(expects == in_to_outs);
// }
//
// TEST(InferInplace, MultiGradInplaceInToOut) {
// ProgramDesc prog;
// auto* op = prog.MutableBlock(0)->AppendOp();
// op->SetType("multi_out_grad");
// op->SetInput(GradVarName("Out"), {"o0"});
// op->SetInput(GradVarName("YOut"), {"y0"});
// op->SetInput(GradVarName("ZOut"), {"z0"});
// op->SetOutput(GradVarName("X"), {"a0", "a1"});
// op->SetOutput(GradVarName("Y"), {"b0"});
// op->SetOutput(GradVarName("Z"), {"c0", "c1"});
//
// prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
// prog.MutableBlock(0)->Var("o0");
// prog.MutableBlock(0)->Var("y0");
// prog.MutableBlock(0)->Var("z0");
// prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
// prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
//
// auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
// auto in_to_outs = infer_inplace(*op);
//
// EXPECT_EQ(in_to_outs.size(), 3ul);
// std::unordered_map<std::string, std::string> expects = {
// {"o0", "a0"}, {"y0", "b0"}, {"z0", "c0"},
// };
// EXPECT_TRUE(expects == in_to_outs);
// }
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ cc29bec6
......@@ -46,9 +46,6 @@ cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
pass_library(graph_to_program_pass base)
pass_library(graph_viz_pass base)
pass_library(lock_free_optimize_pass base)
pass_library(cpu_quantize_placement_pass base)
pass_library(cpu_quantize_pass inference)
pass_library(cpu_quantize_squash_pass inference)
pass_library(fc_fuse_pass inference)
pass_library(attention_lstm_fuse_pass inference)
pass_library(infer_clean_graph_pass inference)
......@@ -71,22 +68,31 @@ pass_library(transpose_flatten_concat_fuse_pass inference)
pass_library(identity_scale_op_clean_pass base)
pass_library(sync_batch_norm_pass base)
pass_library(runtime_context_cache_pass base)
pass_library(simplify_anakin_detection_pattern_pass inference)
pass_library(anakin_fillconstant_elementwisemul_fuse inference)
# There may be many transpose-flatten structures in a model, and the output of
# these structures will be used as inputs to the concat Op. This pattern will
# be detected by our pass. The index here represents the number of structures in the
# pattern. We use index 3 ~ 6, because these quantities of structures are
# common in the models.
foreach (index RANGE 3 6)
foreach (index RANGE 2 6)
file(APPEND ${pass_file} "USE_PASS(transpose_flatten${index}_concat_fuse_pass);\n")
endforeach()
foreach (index RANGE 2 6)
file(APPEND ${pass_file} "USE_PASS(simplify_anakin_detection_pattern_pass${index});\n")
endforeach()
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base mkldnn)
pass_library(depthwise_conv_mkldnn_pass base mkldnn)
pass_library(conv_bias_mkldnn_fuse_pass inference mkldnn)
pass_library(conv_relu_mkldnn_fuse_pass inference mkldnn)
pass_library(conv_elementwise_add_mkldnn_fuse_pass inference mkldnn)
pass_library(cpu_quantize_placement_pass base mkldnn)
pass_library(cpu_quantize_pass inference mkldnn)
pass_library(cpu_quantize_squash_pass inference mkldnn)
endif()
cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
......@@ -105,9 +111,6 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
cc_test(test_cpu_quantize_placement_pass SRCS cpu_quantize_placement_pass_tester.cc DEPS cpu_quantize_placement_pass)
cc_test(test_cpu_quantize_pass SRCS cpu_quantize_pass_tester.cc DEPS cpu_quantize_pass naive_executor)
cc_test(test_cpu_quantize_squash_pass SRCS cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
if(NOT WIN32)
cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
endif()
......@@ -117,4 +120,7 @@ if (WITH_MKLDNN)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
cc_test(test_mkldnn_placement_pass SRCS mkldnn/mkldnn_placement_pass_tester.cc DEPS mkldnn_placement_pass)
cc_test(test_cpu_quantize_placement_pass SRCS mkldnn/cpu_quantize_placement_pass_tester.cc DEPS cpu_quantize_placement_pass)
cc_test(test_cpu_quantize_pass SRCS mkldnn/cpu_quantize_pass_tester.cc DEPS cpu_quantize_pass naive_executor)
cc_test(test_cpu_quantize_squash_pass SRCS mkldnn/cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
endif ()
paddle/fluid/framework/ir/anakin_fillconstant_elementwisemul_fuse.cc 0 → 100644
浏览文件 @ cc29bec6
// 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 <memory>
#include <string>
#include "paddle/fluid/framework/ir/anakin_fillconstant_elementwisemul_fuse.h"
#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(fill_constant); \
GET_IR_NODE(fill_constant_out); \
GET_IR_NODE(elementwise_mul); \
GET_IR_NODE(elementwise_mul_out);
void AnakinFillconstantElementwisemulFuse::ApplyImpl(ir::Graph* graph) const {
const std::string pattern_name = "anakin_fillconstant_elementwisemul_fuse";
FusePassBase::Init(pattern_name, graph);
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()
->NewNode("x")
->assert_is_op_input("elementwise_mul", "X")
->AsInput();
patterns::AnakinFillConstantElementWiseMulFuse pattern(gpd.mutable_pattern(),
pattern_name);
pattern(x);
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_NODES;
PADDLE_ENFORCE(subgraph.count(x));
auto* elementwise_in = subgraph.at(x);
float constant_value =
boost::get<float>(fill_constant->Op()->GetAttr("value"));
framework::OpDesc new_op_desc;
new_op_desc.SetType("scale");
new_op_desc.SetInput("X", {elementwise_in->Name()});
new_op_desc.SetAttr("scale", constant_value);
new_op_desc.SetAttr("bias", static_cast<float>(0.0));
new_op_desc.SetAttr("bias_after_scale", true);
new_op_desc.SetOutput("Out", {elementwise_mul_out->Name()});
new_op_desc.Flush();
// Create a new node for the fused op.
auto* scale_op = graph->CreateOpNode(&new_op_desc);
IR_NODE_LINK_TO(elementwise_in, scale_op); // Input
IR_NODE_LINK_TO(scale_op, elementwise_mul_out); // Output
// Delete the unneeded nodes.
GraphSafeRemoveNodes(graph,
{fill_constant, fill_constant_out, elementwise_mul});
};
gpd(graph, handler);
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(anakin_fillconstant_elementwisemul_fuse,
paddle::framework::ir::AnakinFillconstantElementwisemulFuse);
paddle/fluid/framework/ir/anakin_fillconstant_elementwisemul_fuse.h 0 → 100644
浏览文件 @ cc29bec6
// 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 <memory>
#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 AnakinFillconstantElementwisemulFuse : public FusePassBase {
public:
virtual ~AnakinFillconstantElementwisemulFuse() {}
protected:
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -14,6 +14,7 @@
#include "paddle/fluid/framework/ir/attention_lstm_fuse_pass.h"
#include <string>
#include <unordered_set>
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
#include "paddle/fluid/framework/lod_tensor.h"
......@@ -253,8 +254,7 @@ void PrepareLSTMBias(const LoDTensor& B_forget, const LoDTensor& B_input,
// Parameters
std::unique_ptr<ir::Graph> AttentionLSTMFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void AttentionLSTMFusePass::ApplyImpl(ir::Graph* graph) const {
PDPattern external_pattern, subblock_pattern;
// Use the following variables to tell whether this model is RNN1.
......@@ -269,12 +269,11 @@ std::unique_ptr<ir::Graph> AttentionLSTMFusePass::ApplyImpl(
}
}
if (count < specified_vars.size()) {
return graph;
return;
}
// Continue to fuse.
FindWhileOp(graph.get());
return graph;
FindWhileOp(graph);
}
} // namespace ir
......
paddle/fluid/framework/ir/attention_lstm_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -22,8 +22,7 @@ namespace ir {
class AttentionLSTMFusePass : public FusePassBase {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/conv_affine_channel_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -77,10 +77,9 @@ void recompute_bias_and_weights(const Scope* scope, ir::Node* conv_weight,
weights_array_2d.colwise() *= scale_array;
}
std::unique_ptr<ir::Graph> ConvAffineChannelFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
void ConvAffineChannelFusePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init(name_scope_, graph);
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
......@@ -139,7 +138,7 @@ std::unique_ptr<ir::Graph> ConvAffineChannelFusePass::ApplyImpl(
desc.SetAttr("axis", 1);
auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {ac_scale, ac_bias, affine_channel});
GraphSafeRemoveNodes(graph, {ac_scale, ac_bias, affine_channel});
IR_NODE_LINK_TO(conv_out, eltwise_op);
IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
......@@ -147,16 +146,14 @@ std::unique_ptr<ir::Graph> ConvAffineChannelFusePass::ApplyImpl(
found_conv_ac_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_conv_ac_count);
return graph;
}
std::unique_ptr<ir::Graph> ConvEltwiseAddAffineChannelFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
void ConvEltwiseAddAffineChannelFusePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init(name_scope_, graph);
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
......@@ -199,7 +196,7 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddAffineChannelFusePass::ApplyImpl(
eltwise->Op()->SetAttr("axis", 1);
eltwise->Op()->SetOutput("Out", std::vector<std::string>({ac_out->Name()}));
GraphSafeRemoveNodes(graph.get(),
GraphSafeRemoveNodes(graph,
{ac_scale, ac_bias, affine_channel, eltwise_out});
IR_NODE_LINK_TO(eltwise, ac_out);
......@@ -207,9 +204,8 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddAffineChannelFusePass::ApplyImpl(
found_conv_ac_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_conv_ac_count);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_affine_channel_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -31,8 +31,7 @@ class ConvAffineChannelFusePass : public FusePassBase {
virtual ~ConvAffineChannelFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph*) const override;
const std::string name_scope_{"conv_affine_channel_fuse"};
};
......@@ -41,8 +40,7 @@ class ConvEltwiseAddAffineChannelFusePass : public FusePassBase {
virtual ~ConvEltwiseAddAffineChannelFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph*) const override;
const std::string name_scope_{"conv_eltwiseadd_affine_channel_fuse"};
};
......
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -101,10 +101,9 @@ void recompute_bias_and_weights(const Scope* scope,
weights_array_2d.colwise() *= variance_array;
}
std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
void ConvBNFusePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init(name_scope_, graph);
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
......@@ -187,7 +186,7 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
std::vector<std::string>({bn_out->Name()}));
GraphSafeRemoveNodes(
graph.get(),
graph,
{conv_out, bn_scale, bn_bias, bn_mean, bn_variance, batch_norm,
bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance});
......@@ -203,10 +202,9 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
desc.SetAttr("axis", 1);
auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(
graph.get(),
{bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
bn_variance_out, bn_saved_mean, bn_saved_variance});
GraphSafeRemoveNodes(graph, {bn_scale, bn_bias, bn_mean, bn_variance,
batch_norm, bn_mean_out, bn_variance_out,
bn_saved_mean, bn_saved_variance});
IR_NODE_LINK_TO(conv_out, eltwise_op);
IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
......@@ -215,16 +213,14 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
}
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_conv_bn_count);
return graph;
}
std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
void ConvEltwiseAddBNFusePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init(name_scope_, graph);
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
......@@ -274,7 +270,7 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
eltwise->Op()->SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
GraphSafeRemoveNodes(
graph.get(),
graph,
{bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
bn_variance_out, bn_saved_mean, bn_saved_variance, eltwise_out});
......@@ -283,10 +279,9 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
found_conv_bn_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_conv_bn_count);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_bn_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -31,8 +31,7 @@ class ConvBNFusePass : public FusePassBase {
virtual ~ConvBNFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"conv_bn_fuse"};
};
......@@ -41,8 +40,7 @@ class ConvEltwiseAddBNFusePass : public FusePassBase {
virtual ~ConvEltwiseAddBNFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"conv_eltwiseadd_bn_fuse"};
};
......
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse.cc
浏览文件 @ cc29bec6
......@@ -50,10 +50,9 @@ framework::proto::OpDesc PrepareOpDesc(
return *desc.Proto();
}
std::unique_ptr<ir::Graph> ConvElementwiseAddActFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void ConvElementwiseAddActFusePass::ApplyImpl(ir::Graph* graph) const {
const std::string pattern_name = "conv_elementwise_add_act_fuse";
FusePassBase::Init(pattern_name, graph.get());
FusePassBase::Init(pattern_name, graph);
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()->NewNode("x")->AsInput()->assert_is_op_input(
......@@ -95,7 +94,6 @@ std::unique_ptr<ir::Graph> ConvElementwiseAddActFusePass::ApplyImpl(
elementwise_add_out});
};
gpd(graph.get(), handler);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -51,10 +51,9 @@ framework::proto::OpDesc PrepareOpDesc(
return *desc.Proto();
}
std::unique_ptr<ir::Graph> ConvElementwiseAdd2ActFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void ConvElementwiseAdd2ActFusePass::ApplyImpl(ir::Graph* graph) const {
const std::string pattern_name = "conv_elementwise_add2_act_fuse";
FusePassBase::Init(pattern_name, graph.get());
FusePassBase::Init(pattern_name, graph);
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()->NewNode("x")->AsInput()->assert_is_op_input(
......@@ -92,12 +91,10 @@ std::unique_ptr<ir::Graph> ConvElementwiseAdd2ActFusePass::ApplyImpl(
// Delete the unneeded nodes.
GraphSafeRemoveNodes(
graph.get(),
{conv_op, conv_out, elementwise_add_op, elementwise_add_op_1,
elementwise_add_out, elementwise_add_out_1, act_op});
graph, {conv_op, conv_out, elementwise_add_op, elementwise_add_op_1,
elementwise_add_out, elementwise_add_out_1, act_op});
};
gpd(graph.get(), handler);
return graph;
gpd(graph, handler);
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -25,8 +25,7 @@ class ConvElementwiseAdd2ActFusePass : public FusePassBase {
virtual ~ConvElementwiseAdd2ActFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -48,10 +48,9 @@ framework::proto::OpDesc PrepareOpDesc(
return *desc.Proto();
}
std::unique_ptr<ir::Graph> ConvElementwiseAddActFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void ConvElementwiseAddActFusePass::ApplyImpl(ir::Graph* graph) const {
const std::string pattern_name = "conv_elementwise_add_act_fuse";
FusePassBase::Init(pattern_name, graph.get());
FusePassBase::Init(pattern_name, graph);
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()
......@@ -88,12 +87,11 @@ std::unique_ptr<ir::Graph> ConvElementwiseAddActFusePass::ApplyImpl(
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});
GraphSafeRemoveNodes(graph, {conv_op, conv_out, elementwise_add_op,
elementwise_add_out, act_op});
};
gpd(graph.get(), handler);
return graph;
gpd(graph, handler);
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -25,8 +25,7 @@ class ConvElementwiseAddActFusePass : public FusePassBase {
virtual ~ConvElementwiseAddActFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -30,10 +30,9 @@ namespace ir {
GET_IR_NODE(elementwise_add_in_y); \
GET_IR_NODE(elementwise_add_out);
std::unique_ptr<ir::Graph> ConvElementwiseAddFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void ConvElementwiseAddFusePass::ApplyImpl(ir::Graph* graph) const {
const std::string pattern_name = "conv_elementwise_add_fuse";
FusePassBase::Init(pattern_name, graph.get());
FusePassBase::Init(pattern_name, graph);
GraphPatternDetector gpd;
auto* x = gpd.mutable_pattern()
......@@ -76,11 +75,10 @@ std::unique_ptr<ir::Graph> ConvElementwiseAddFusePass::ApplyImpl(
IR_NODE_LINK_TO(new_conv_op, elementwise_add_out); // Output
// Delete the unneeded nodes.
GraphSafeRemoveNodes(graph.get(), {conv_op, conv_out, elementwise_add_op});
GraphSafeRemoveNodes(graph, {conv_op, conv_out, elementwise_add_op});
};
gpd(graph.get(), handler);
return graph;
gpd(graph, handler);
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -25,8 +25,7 @@ class ConvElementwiseAddFusePass : public FusePassBase {
virtual ~ConvElementwiseAddFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -15,6 +15,8 @@
#include "paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h"
#include <algorithm>
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/operators/math/blas.h"
......@@ -201,7 +203,7 @@ static int BuildFusion(Graph* graph, const std::string& name_scope,
// Remove unneeded nodes.
// TODO(jczaja): Proper removing of lookup table
std::unordered_set<const Node*> marked_nodes(
//{lookup_table, mul, lstm, elementwise_add, fc_bias, W});
// {lookup_table, mul, lstm, elementwise_add, fc_bias, W});
{mul, lstm, elementwise_add, fc_bias});
GraphSafeRemoveNodes(graph, marked_nodes);
} else {
......@@ -224,15 +226,13 @@ static int BuildFusion(Graph* graph, const std::string& name_scope,
return fusion_count;
}
std::unique_ptr<ir::Graph> EmbeddingFCLSTMFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
void EmbeddingFCLSTMFusePass::ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init(name_scope_, graph);
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope(),
true /*with_fc_bias*/);
int fusion_count =
BuildFusion(graph, name_scope_, param_scope(), true /*with_fc_bias*/);
AddStatis(fusion_count);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -32,8 +32,7 @@ class EmbeddingFCLSTMFusePass : public FusePassBase {
virtual ~EmbeddingFCLSTMFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"embedding_fc_lstm_fuse"};
};
......
paddle/fluid/framework/ir/fc_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -14,6 +14,7 @@
#include "paddle/fluid/framework/ir/fc_fuse_pass.h"
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/platform/enforce.h"
......@@ -22,10 +23,9 @@ namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("fc_fuse", graph.get());
void FCFusePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init("fc_fuse", graph);
std::unordered_set<Node*> nodes2delete;
......@@ -61,7 +61,7 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
desc.SetAttr("in_num_col_dims", mul->Op()->GetAttr("x_num_col_dims"));
desc.SetType("fc");
auto fc_node = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {mul, elementwise_add, mul_out});
GraphSafeRemoveNodes(graph, {mul, elementwise_add, mul_out});
PADDLE_ENFORCE(subgraph.count(x));
IR_NODE_LINK_TO(subgraph.at(x), fc_node);
......@@ -72,10 +72,9 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
found_fc_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_fc_count);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/fc_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -31,8 +31,7 @@ class FCFusePass : public FusePassBase {
virtual ~FCFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/fc_fuse_pass_tester.cc
浏览文件 @ cc29bec6
......@@ -73,7 +73,7 @@ TEST(FCFusePass, basic) {
int pre_nodes = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
graph.reset(pass->Apply(graph.release()));
int after_nodes = graph->Nodes().size();
......
paddle/fluid/framework/ir/fc_gru_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -14,6 +14,7 @@
#include "paddle/fluid/framework/ir/fc_gru_fuse_pass.h"
#include <string>
#include <unordered_set>
#include "paddle/fluid/framework/lod_tensor.h"
namespace paddle {
......@@ -39,7 +40,6 @@ static int BuildFusion(Graph* graph, const std::string& name_scope,
// Create New OpDesc
auto gru_creater = [&](Node* gru, Node* x, Node* weight_x, Node* weight_h,
Node* bias, Node* hidden, Node* fc_bias) {
OpDesc op_desc;
op_desc.SetType("fusion_gru");
......@@ -155,26 +155,22 @@ static int BuildFusion(Graph* graph, const std::string& name_scope,
return fusion_count;
}
std::unique_ptr<ir::Graph> MulGRUFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
void MulGRUFusePass::ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init(name_scope_, graph);
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope(),
false /*with_fc_bias*/);
int fusion_count =
BuildFusion(graph, name_scope_, param_scope(), false /*with_fc_bias*/);
AddStatis(fusion_count);
return graph;
}
std::unique_ptr<ir::Graph> FCGRUFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
void FCGRUFusePass::ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init(name_scope_, graph);
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope(),
true /*with_fc_bias*/);
int fusion_count =
BuildFusion(graph, name_scope_, param_scope(), true /*with_fc_bias*/);
AddStatis(fusion_count);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/fc_gru_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -30,8 +30,7 @@ class FCGRUFusePass : public FusePassBase {
virtual ~FCGRUFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"fc_gru_fuse"};
};
......@@ -42,8 +41,7 @@ class MulGRUFusePass : public FusePassBase {
virtual ~MulGRUFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"fc_nobias_gru_fuse"};
};
......
paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -14,6 +14,7 @@
#include "paddle/fluid/framework/ir/fc_lstm_fuse_pass.h"
#include <string>
#include <unordered_set>
#include "paddle/fluid/framework/lod_tensor.h"
namespace paddle {
......@@ -157,26 +158,22 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
return fusion_count;
}
std::unique_ptr<ir::Graph> MulLstmFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
void MulLstmFusePass::ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init(name_scope_, graph);
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope(),
false /*with_fc_bias*/);
int fusion_count =
BuildFusion(graph, name_scope_, param_scope(), false /*with_fc_bias*/);
AddStatis(fusion_count);
return graph;
}
std::unique_ptr<ir::Graph> FCLstmFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
void FCLstmFusePass::ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init(name_scope_, graph);
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope(),
true /*with_fc_bias*/);
int fusion_count =
BuildFusion(graph, name_scope_, param_scope(), true /*with_fc_bias*/);
AddStatis(fusion_count);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/fc_lstm_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -32,8 +32,7 @@ class FCLstmFusePass : public FusePassBase {
virtual ~FCLstmFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"fc_lstm_fuse"};
};
......@@ -43,8 +42,7 @@ class MulLstmFusePass : public FusePassBase {
virtual ~MulLstmFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"fc_nobias_lstm_fuse"};
};
......
paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc
浏览文件 @ cc29bec6
......@@ -15,6 +15,8 @@
#include "paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h"
#include <algorithm>
#include <string>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/enforce.h"
......@@ -23,29 +25,25 @@ namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> FuseElewiseAddActPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void FuseElewiseAddActPass::ApplyImpl(ir::Graph *graph) const {
std::unordered_set<std::string> act_types = {"relu", "scale"};
graph = FuseActElewiseAdd(std::move(graph), act_types);
graph = FuseElewiseAddAct(std::move(graph), act_types);
graph = FuseActElewiseAdd(graph, act_types);
graph = FuseElewiseAddAct(graph, act_types);
// backward
{
std::unordered_set<std::string> in_place_act_types = {"relu_grad"};
graph = FuseElewiseAddActInplaceGrad(std::move(graph), in_place_act_types);
graph = FuseElewiseAddActInplaceGrad(graph, in_place_act_types);
}
// Remove the removable intermediate_out.
RemoveIntermediateOut(graph.get());
return graph;
RemoveIntermediateOut(graph);
}
// ele_add(x, act(y))
std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddAct(
std::unique_ptr<ir::Graph> graph,
const std::unordered_set<std::string> &act_types) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("elewise_add_act", graph.get());
ir::Graph *FuseElewiseAddActPass::FuseElewiseAddAct(
ir::Graph *graph, const std::unordered_set<std::string> &act_types) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init("elewise_add_act", graph);
GraphPatternDetector gpd;
auto *x = gpd.mutable_pattern()
......@@ -86,18 +84,17 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddAct(
found_elewise_add_act_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_elewise_add_act_count);
return graph;
}
// act(ele_add(x,y))
std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
std::unique_ptr<ir::Graph> graph,
const std::unordered_set<std::string> &act_types) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("act_elewise_add", graph.get());
ir::Graph *FuseElewiseAddActPass::FuseActElewiseAdd(
ir::Graph *graph, const std::unordered_set<std::string> &act_types) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init("act_elewise_add", graph);
GraphPatternDetector gpd;
auto *x = gpd.mutable_pattern()
......@@ -137,7 +134,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
found_elewise_add_act_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_elewise_add_act_count);
return graph;
......@@ -146,11 +143,10 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
// the backward of act(ele_add(x,y))
// act_grad: in["Out", "Out@GRAD"], out["X@GRAD"]
// ele_add_grad: in["Y", "Out@GRAD"], out["X@GRAD", "Y@GRAD"]
std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
std::unique_ptr<ir::Graph> graph,
const std::unordered_set<std::string> &act_types) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("elewise_add_act_grad", graph.get());
ir::Graph *FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
ir::Graph *graph, const std::unordered_set<std::string> &act_types) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init("elewise_add_act_grad", graph);
GraphPatternDetector gpd;
auto *d_act_out = gpd.mutable_pattern()
......@@ -217,7 +213,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
found_elewise_add_act_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_elewise_add_act_count);
return graph;
......
paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h
浏览文件 @ cc29bec6
......@@ -14,6 +14,8 @@
#pragma once
#include <string>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
......@@ -32,20 +34,16 @@ class FuseElewiseAddActPass : public FusePassBase {
virtual ~FuseElewiseAddActPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph *graph) const override;
std::unique_ptr<ir::Graph> FuseElewiseAddAct(
std::unique_ptr<ir::Graph> graph,
const std::unordered_set<std::string> &act_types) const;
ir::Graph *FuseElewiseAddAct(
ir::Graph *graph, const std::unordered_set<std::string> &act_types) const;
std::unique_ptr<ir::Graph> FuseActElewiseAdd(
std::unique_ptr<ir::Graph> graph,
const std::unordered_set<std::string> &act_types) const;
ir::Graph *FuseActElewiseAdd(
ir::Graph *graph, const std::unordered_set<std::string> &act_types) const;
std::unique_ptr<ir::Graph> FuseElewiseAddActInplaceGrad(
std::unique_ptr<ir::Graph> graph,
const std::unordered_set<std::string> &act_types) const;
ir::Graph *FuseElewiseAddActInplaceGrad(
ir::Graph *graph, const std::unordered_set<std::string> &act_types) const;
/**
* Remove the removable intermediate_out.
......
paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc
浏览文件 @ cc29bec6
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h"
#include <algorithm>
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/enforce.h"
......@@ -23,20 +24,18 @@ namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> FuseReluDepthwiseConvPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
graph = FuseReluDepthwiseConv(std::move(graph), true);
graph = FuseReluDepthwiseConv(std::move(graph), false);
return graph;
void FuseReluDepthwiseConvPass::ApplyImpl(ir::Graph *graph) const {
graph = FuseReluDepthwiseConv(graph, true);
graph = FuseReluDepthwiseConv(graph, false);
}
std::unique_ptr<ir::Graph> FuseReluDepthwiseConvPass::FuseReluDepthwiseConv(
std::unique_ptr<ir::Graph> graph, bool only_forward) const {
PADDLE_ENFORCE(graph.get());
ir::Graph *FuseReluDepthwiseConvPass::FuseReluDepthwiseConv(
ir::Graph *graph, bool only_forward) const {
PADDLE_ENFORCE(graph);
if (only_forward)
FusePassBase::Init("relu_depthwise_conv_only_forward", graph.get());
FusePassBase::Init("relu_depthwise_conv_only_forward", graph);
else
FusePassBase::Init("relu_depthwise_conv", graph.get());
FusePassBase::Init("relu_depthwise_conv", graph);
/*
x ---act--> y ---layer-> z
+----------+
......@@ -144,10 +143,9 @@ std::unique_ptr<ir::Graph> FuseReluDepthwiseConvPass::FuseReluDepthwiseConv(
}
count++;
};
gpd(graph.get(), handler);
GraphSafeRemoveNodes(graph.get(), need_removed_nodes);
gpd(graph, handler);
GraphSafeRemoveNodes(graph, need_removed_nodes);
AddStatis(count);
return graph;
}
......
paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h
浏览文件 @ cc29bec6
......@@ -32,10 +32,8 @@ class FuseReluDepthwiseConvPass : public FusePassBase {
virtual ~FuseReluDepthwiseConvPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
std::unique_ptr<ir::Graph> FuseReluDepthwiseConv(
std::unique_ptr<ir::Graph> graph, bool only_forward) const;
void ApplyImpl(ir::Graph* graph) const override;
ir::Graph* FuseReluDepthwiseConv(ir::Graph* graph, bool only_forward) const;
};
} // namespace ir
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ cc29bec6
......@@ -1470,6 +1470,171 @@ PDNode *patterns::TransposeFlattenConcat::operator()(
return concat_out;
}
PDNode *patterns::AnakinDetectionPattern::operator()(
std::vector<PDNode *> conv_in, int times) {
// The times represents the repeat times of the
// {prior_box, prior_box_loc_out, flatten, prior_box_var_out, reshape}
const int kNumFields = 7;
const int kPriorBoxLocOffset = 1;
const int kReshape1Offset = 2;
const int kReshape1OutOffset = 3;
const int kPriorBoxVarOffset = 4;
const int kReshape2Offset = 5;
const int kReshape2OutOffset = 6;
const int kBoxCoderThirdInputOffset = times;
const int kMultiClassSecondInputNmsOffset = times + 1;
std::vector<PDNode *> nodes;
for (int i = 0; i < times; i++) {
nodes.push_back(
pattern->NewNode(GetNodeName("prior_box" + std::to_string(i)))
->assert_is_op("density_prior_box"));
nodes.push_back(pattern->NewNode(GetNodeName("box_out" + std::to_string(i)))
->assert_is_op_output("density_prior_box", "Boxes")
->assert_is_op_input("reshape2", "X")
->AsIntermediate());
nodes.push_back(
pattern->NewNode(GetNodeName("reshape1" + std::to_string(i)))
->assert_is_op("reshape2"));
nodes.push_back(
pattern->NewNode(GetNodeName("reshape1_out" + std::to_string(i)))
->assert_is_op_output("reshape2")
->assert_is_op_nth_input("concat", "X", i)
->AsIntermediate());
nodes.push_back(
pattern->NewNode(GetNodeName("box_var_out" + std::to_string(i)))
->assert_is_op_output("density_prior_box", "Variances")
->assert_is_op_input("reshape2", "X")
->AsIntermediate());
nodes.push_back(
pattern->NewNode(GetNodeName("reshape2" + std::to_string(i)))
->assert_is_op("reshape2"));
nodes.push_back(
pattern->NewNode(GetNodeName("reshape2_out" + std::to_string(i)))
->assert_is_op_output("reshape2")
->assert_is_op_nth_input("concat", "X", i)
->AsIntermediate());
}
auto concat_op1 = pattern->NewNode(GetNodeName("concat1"))
->assert_is_op("concat")
->assert_op_has_n_inputs("concat", times);
auto concat_out1 = pattern->NewNode(GetNodeName("concat1_out"))
->assert_is_op_output("concat")
->AsIntermediate();
auto concat_op2 = pattern->NewNode(GetNodeName("concat2"))
->assert_is_op("concat")
->assert_op_has_n_inputs("concat", times);
auto concat_out2 = pattern->NewNode(GetNodeName("concat2_out"))
->assert_is_op_output("concat")
->AsIntermediate();
auto box_coder_op = pattern->NewNode(GetNodeName("box_coder"))
->assert_is_op("box_coder")
->assert_op_has_n_inputs("box_coder", 3);
auto box_coder_out = pattern->NewNode(GetNodeName("box_coder_out"))
->assert_is_op_output("box_coder")
->AsIntermediate();
auto transpose_before_nms =
pattern->NewNode(GetNodeName("transpose_before_nms"))
->assert_is_op("transpose2");
auto transpose_before_nms_out =
pattern->NewNode(GetNodeName("transpose_before_nms_out"))
->assert_is_op_output("transpose2")
->assert_is_op_input("multiclass_nms", "Scores")
->AsIntermediate();
auto multiclass_nms_op = pattern->NewNode(GetNodeName("multiclass_nms"))
->assert_is_op("multiclass_nms")
->assert_op_has_n_inputs("multiclass_nms", 2);
auto multiclass_nms_out = pattern->NewNode(GetNodeName("multiclass_nms_out"))
->assert_is_op_output("multiclass_nms")
->AsOutput();
std::vector<PDNode *> reshape1_outs;
std::vector<PDNode *> reshape2_outs;
for (int i = 0; i < times; i++) {
conv_in[i]->AsInput();
// prior_box
nodes[i * kNumFields]->LinksFrom({conv_in[i]});
// prior_box box out
nodes[i * kNumFields + kPriorBoxLocOffset]->LinksFrom(
{nodes[i * kNumFields]});
// reshape
nodes[i * kNumFields + kReshape1Offset]->LinksFrom(
{nodes[i * kNumFields + kPriorBoxLocOffset]});
// reshape_out
nodes[i * kNumFields + kReshape1OutOffset]->LinksFrom(
{nodes[i * kNumFields + kReshape1Offset]});
nodes[i * kNumFields + kPriorBoxVarOffset]->LinksFrom(
{nodes[i * kNumFields]});
// reshape
nodes[i * kNumFields + kReshape2Offset]->LinksFrom(
{nodes[i * kNumFields + kPriorBoxVarOffset]});
// reshape_out
nodes[i * kNumFields + kReshape2OutOffset]->LinksFrom(
{nodes[i * kNumFields + kReshape2Offset]});
reshape1_outs.push_back(nodes[i * kNumFields + kReshape1OutOffset]);
reshape2_outs.push_back(nodes[i * kNumFields + kReshape2OutOffset]);
}
concat_op1->LinksFrom(reshape1_outs);
concat_op2->LinksFrom(reshape2_outs);
concat_out1->LinksFrom({concat_op1});
concat_out2->LinksFrom({concat_op2});
conv_in[kBoxCoderThirdInputOffset]->AsInput();
conv_in[kMultiClassSecondInputNmsOffset]->AsInput();
box_coder_op->LinksFrom(
{concat_out1, concat_out2, conv_in[kBoxCoderThirdInputOffset]});
box_coder_out->LinksFrom({box_coder_op});
transpose_before_nms->LinksFrom({conv_in[kMultiClassSecondInputNmsOffset]});
transpose_before_nms_out->LinksFrom({transpose_before_nms});
multiclass_nms_op->LinksFrom({box_coder_out, transpose_before_nms_out})
.LinksTo({multiclass_nms_out});
return multiclass_nms_out;
}
PDNode *patterns::AnakinFillConstantElementWiseMulFuse::operator()(
PDNode *elementwise_op_input) {
auto fill_constant =
pattern->NewNode(fill_constant_repr())->assert_is_op("fill_constant");
auto fill_constant_out = pattern->NewNode(fill_constant_out_repr())
->assert_is_op_output("fill_constant")
->assert_is_op_input("elementwise_mul", "Y")
->AsIntermediate();
auto elementwise_mul_op =
pattern->NewNode(elementwise_mul_repr())->assert_is_op("elementwise_mul");
auto elementwise_mul_out = pattern->NewNode(elementwise_mul_out_repr())
->assert_is_op_output("elementwise_mul")
->AsOutput();
fill_constant_out->LinksFrom({fill_constant});
elementwise_mul_op->LinksFrom({elementwise_op_input, fill_constant_out});
elementwise_mul_out->LinksFrom({elementwise_mul_op});
return elementwise_mul_out;
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ cc29bec6
......@@ -844,6 +844,36 @@ struct TransposeFlattenConcat : public PatternBase {
}
};
struct AnakinDetectionPattern : public PatternBase {
AnakinDetectionPattern(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "anakin_detect_pattern") {}
PDNode* operator()(std::vector<PDNode*> conv_inputs, int times);
std::string GetNodeName(const std::string& op_type) {
return PDNodeName(name_scope_, repr_, id_, op_type);
}
PDNode* GetPDNode(const std::string& op_type) {
return pattern->RetrieveNode(GetNodeName(op_type));
}
};
struct AnakinFillConstantElementWiseMulFuse : public PatternBase {
AnakinFillConstantElementWiseMulFuse(PDPattern* pattern,
const std::string& name_scope)
: PatternBase(pattern, name_scope,
"anakin_fillconstant_elementwisemul_fuse") {}
PDNode* operator()(PDNode* elementwise_op_input);
// declare operator node's name
PATTERN_DECL_NODE(fill_constant);
PATTERN_DECL_NODE(fill_constant_out);
PATTERN_DECL_NODE(elementwise_mul);
PATTERN_DECL_NODE(elementwise_mul_out);
};
} // namespace patterns
// Link two ir::Nodes from each other.
......
paddle/fluid/framework/ir/graph_to_program_pass.cc
浏览文件 @ cc29bec6
......@@ -15,7 +15,9 @@ limitations under the License. */
#include "paddle/fluid/framework/ir/graph_to_program_pass.h"
#include <map>
#include <memory>
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
......@@ -26,8 +28,7 @@ namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<Graph> GraphToProgramPass::ApplyImpl(
std::unique_ptr<Graph> graph) const {
void GraphToProgramPass::ApplyImpl(ir::Graph* graph) const {
// Remove the unneeded variables after memory optimization.
std::unordered_set<std::string> vars2remove;
if (graph->Has(kGraphToProgramVarsToRemove)) {
......@@ -73,7 +74,6 @@ std::unique_ptr<Graph> GraphToProgramPass::ApplyImpl(
}
program.CopyFrom(*program_pb);
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/graph_to_program_pass.h
浏览文件 @ cc29bec6
......@@ -26,7 +26,7 @@ const char kGraphToProgramSortKind[] = "__graph_to_program_sort_kind__";
class GraphToProgramPass : public Pass {
protected:
std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/graph_to_program_pass_test.cc
浏览文件 @ cc29bec6
......@@ -14,7 +14,9 @@ limitations under the License. */
#include "paddle/fluid/framework/ir/graph_to_program_pass.h"
#include <memory>
#include <string>
#include <unordered_set>
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/program_desc.h"
......@@ -84,7 +86,7 @@ TEST(GraphToProgramPass, Basic) {
ProgramDesc compiled_prog;
pass->SetNotOwned<paddle::framework::ProgramDesc>("program", &compiled_prog);
pass->Apply(std::move(g));
pass->Apply(g.get());
std::vector<OpDesc*> ops = compiled_prog.Block(0).AllOps();
EXPECT_EQ(ops[0]->Type(), "op1");
EXPECT_EQ(ops[1]->Type(), "op2");
......
paddle/fluid/framework/ir/graph_viz_pass.cc
浏览文件 @ cc29bec6
......@@ -12,10 +12,10 @@ 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/graph_viz_pass.h"
#include <algorithm>
#include <unordered_map>
#include <unordered_set>
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
#include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/inference/analysis/dot.h"
#include "paddle/fluid/string/printf.h"
......@@ -38,8 +38,7 @@ std::string FormatName(const Node* node) {
}
} // namespace
std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void GraphVizPass::ApplyImpl(ir::Graph* graph) const {
const std::string graph_viz_path = Get<std::string>(kGraphVizPath);
VLOG(3) << "draw IR graph viz to " << graph_viz_path;
std::unique_ptr<std::ostream> fout(new std::ofstream(graph_viz_path));
......@@ -82,7 +81,7 @@ std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
{Dot::Attr("style", "filled,rounded"), Dot::Attr("shape", "box"),
Dot::Attr("fillcolor", "yellow")});
auto marked_nodes = ConsumeMarkedNodes(graph.get());
auto marked_nodes = ConsumeMarkedNodes(graph);
// Create nodes
for (const Node* n : graph->Nodes()) {
std::string node_id = FormatName(n) + "(" + std::to_string(n->id()) + ")";
......@@ -115,8 +114,6 @@ std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
}
sout << dot.Build();
return graph;
}
GraphVizPass::marked_nodes_t GraphVizPass::ConsumeMarkedNodes(
......@@ -135,4 +132,4 @@ GraphVizPass::marked_nodes_t GraphVizPass::ConsumeMarkedNodes(
} // namespace paddle
REGISTER_PASS(graph_viz_pass, paddle::framework::ir::GraphVizPass)
.RequirePassAttr(paddle::framework::ir::kGraphVizPath);
\ No newline at end of file
.RequirePassAttr(paddle::framework::ir::kGraphVizPath);
paddle/fluid/framework/ir/graph_viz_pass.h
浏览文件 @ cc29bec6
......@@ -18,6 +18,7 @@ limitations under the License. */
#include <map>
#include <memory>
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
......@@ -34,8 +35,7 @@ class GraphVizPass : public Pass {
using marked_nodes_t = std::unordered_set<const Node*>;
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
// Tell whether there are any marked nodes in the graph. Consume the
// corresponding attribute.
......
paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc
浏览文件 @ cc29bec6
......@@ -20,9 +20,8 @@ namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init("identity_scale_op_clean", graph.get());
void IdentityScaleOpCleanPass::ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init("identity_scale_op_clean", graph);
// pre_op -> scale_in -> scale_op -> scale_out
// ->
......@@ -72,8 +71,7 @@ std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
IR_NODE_LINK_TO(pre_op_var, scale_out_var);
};
detector(graph.get(), handler);
return graph;
detector(graph, handler);
}
} // namespace ir
......
paddle/fluid/framework/ir/identity_scale_op_clean_pass.h
浏览文件 @ cc29bec6
......@@ -22,8 +22,7 @@ namespace ir {
class IdentityScaleOpCleanPass : public FusePassBase {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
private:
virtual ~IdentityScaleOpCleanPass() = default;
......
paddle/fluid/framework/ir/infer_clean_graph_pass.cc
浏览文件 @ cc29bec6
......@@ -26,9 +26,9 @@ class InferCleanGraphPass : public FusePassBase {
virtual ~InferCleanGraphPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init("original_graph", graph.get());
PADDLE_ENFORCE(graph.get());
void ApplyImpl(ir::Graph* graph) const {
FusePassBase::Init("original_graph", graph);
PADDLE_ENFORCE(graph);
auto is_valid_node = [](Node* x) {
return x && IsControlDepVar(*x) && x->IsVar() && !x->Var();
......@@ -46,11 +46,9 @@ class InferCleanGraphPass : public FusePassBase {
}
}
GraphSafeRemoveNodes(graph.get(), invalid_nodes);
GraphSafeRemoveNodes(graph, invalid_nodes);
AddStatis(valid_op);
return graph;
}
void CleanEdges(std::vector<Node*>* nodes,
......
paddle/fluid/framework/ir/is_test_pass.cc
浏览文件 @ cc29bec6
......@@ -20,8 +20,7 @@ namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> IsTestPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
void IsTestPass::ApplyImpl(ir::Graph* graph) const {
VLOG(3) << "Sets is_test attrbiute to true and if it is missing, inserts it "
"for activations and pooling.";
auto op_list = {"pool2d", "sigmoid", "logsigmoid",
......@@ -47,7 +46,6 @@ std::unique_ptr<ir::Graph> IsTestPass::ApplyImpl(
}
}
}
return graph;
}
} // namespace ir
......
paddle/fluid/framework/ir/is_test_pass.h
浏览文件 @ cc29bec6
......@@ -22,8 +22,7 @@ namespace ir {
class IsTestPass : public Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/is_test_pass_tester.cc
浏览文件 @ cc29bec6
......@@ -97,7 +97,7 @@ TEST(IsTestPass, basic) {
auto pass = PassRegistry::Instance().Get("is_test_pass");
graph = pass->Apply(std::move(graph));
graph.reset(pass->Apply(graph.release()));
for (auto* node : graph->Nodes()) {
if (node->IsOp()) {
......
paddle/fluid/framework/ir/lock_free_optimize_pass.cc
浏览文件 @ cc29bec6
......@@ -32,9 +32,8 @@ const char kSumGradOpName[] = "sum";
// other optimizers later.
const char kOptimizerType[] = "sgd";
std::unique_ptr<ir::Graph> LockFreeOptimizePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
void LockFreeOptimizePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
// We could collect all weights' name from SGD, where
// W1 <- SGD(W0, Grad0)
......@@ -92,14 +91,14 @@ std::unique_ptr<ir::Graph> LockFreeOptimizePass::ApplyImpl(
// find the forward op related to the backward op
ir::Node* forward_op =
FindForwardOpViaBackwardOp(graph.get(), backward_op);
FindForwardOpViaBackwardOp(graph, backward_op);
VLOG(3) << "Found forward_op " << forward_op->Name();
PADDLE_ENFORCE(forward_op);
Node* new_optimizer_node = CreateNewSGDNode(
graph.get(), forward_op, backward_op, node, opt_node);
graph, forward_op, backward_op, node, opt_node);
PADDLE_ENFORCE(new_optimizer_node);
}
......@@ -140,8 +139,6 @@ std::unique_ptr<ir::Graph> LockFreeOptimizePass::ApplyImpl(
}
}
}
return graph;
}
ir::Node* LockFreeOptimizePass::CreateNewSGDNode(
......
paddle/fluid/framework/ir/lock_free_optimize_pass.h
浏览文件 @ cc29bec6
......@@ -60,8 +60,7 @@ class LockFreeOptimizePass : public Pass {
virtual ~LockFreeOptimizePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
private:
// Create a new sgd node via current optimizer node
......
paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc
浏览文件 @ cc29bec6
......@@ -38,10 +38,9 @@ LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b,
return vec_y;
}
std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
void ConvBiasFusePass::ApplyImpl(ir::Graph* graph) const {
PADDLE_ENFORCE(graph);
FusePassBase::Init(name_scope_, graph);
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
......@@ -99,7 +98,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
conv->Op()->SetOutput("Output",
std::vector<std::string>({eltwise_out->Name()}));
GraphSafeRemoveNodes(graph.get(), {eltwise, conv_out});
GraphSafeRemoveNodes(graph, {eltwise, conv_out});
IR_NODE_LINK_TO(conv, eltwise_out);
} else {
......@@ -123,14 +122,13 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
IR_NODE_LINK_TO(eltwise_bias, conv_bias_node);
IR_NODE_LINK_TO(conv_bias_node, eltwise_out);
GraphSafeRemoveNodes(graph.get(), {conv, eltwise, conv_out});
GraphSafeRemoveNodes(graph, {conv, eltwise, conv_out});
}
found_conv_bias_count++;
};
gpd(graph.get(), handler);
gpd(graph, handler);
AddStatis(found_conv_bias_count);
return graph;
}
} // namespace ir
} // namespace framework
......
paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h
浏览文件 @ cc29bec6
......@@ -29,8 +29,7 @@ class ConvBiasFusePass : public FusePassBase {
virtual bool is_conv3d() const { return false; }
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void ApplyImpl(ir::Graph* graph) const override;
const std::string name_scope_{"conv_bias_mkldnn_fuse"};
};
/*
......
paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc
浏览文件 @ cc29bec6
......@@ -13,10 +13,10 @@
// limitations under the License.
#include "paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/platform/place.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
......@@ -103,7 +103,7 @@ void MainTest(bool convWithExistingBias) {
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
graph.reset(pass->Apply(graph.release()));
int current_nodes_num = graph->Nodes().size();
......
paddle/fluid/framework/ir/pass.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/ir/pass.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/no_need_buffer_vars_inference.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/op_desc.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/op_desc.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/op_info.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/scope.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/scope.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/framework/type_defs.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/imperative/layer.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/imperative/tracer.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/activation.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/batch_norm.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/batch_norm.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/concat.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/concat.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/conv2d.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/conv2d_fusion.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/conv2d_fusion.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/detection_out.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/detection_out.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/dropout.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/dropout.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/elementwise.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/elementwise.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/flatten.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/flatten.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/im2sequence.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/im2sequence.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/pool2d.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/pool2d.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/relu.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/relu.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/reshape.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/reshape.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/scale.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/scale.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/softmax.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/softmax.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/split.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/split.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/sum.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/sum.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_concat_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_conv2d_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_pool2d_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_relu_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_split_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_sum_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/transpose.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/convert/transpose.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/op_teller.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/anakin/op_teller.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/api/mkldnn_quantizer.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/api/mkldnn_quantizer.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/inference/api/mkldnn_quantizer_config.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/anakin/CMakeLists.txt 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/anakin/anakin_engine_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/anakin/anakin_engine_op.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/clip_op.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/concat_op.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/crop_op.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/dgc_clip_by_norm_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/dgc_clip_by_norm_op.cu 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/dgc_clip_by_norm_op.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/dgc_op.cc 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/dgc_op.cu 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/dgc_op.h 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/gather.cu.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/platform/assert.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/platform/init.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ cc29bec6
此差异已折叠。
paddle/fluid/string/printf.h
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/dataset/wmt14.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/install_check.py 0 → 100644
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ cc29bec6
此差异已折叠。
python/paddle/fluid/tests/unittests/test_dgc_op.py 0 → 100644
浏览文件 @ cc29bec6
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