m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'develop' of https://github.com/paddlepaddle/paddle into add_affine_channel_fuse

Browse Source 
test=develop
revert-15207-remove_op_handle_lock_and_fix_var
nhzlx 6 years ago
parent ce3782c193 bc16bcda49
commit 73b47df1f4
191 changed files with 8529 additions and 4152 deletions
Show Stats Download Patch File Download Diff File

8
CMakeLists.txt
Unescape View File

@ -126,16 +126,12 @@ if(ANDROID OR IOS)
add_definitions(-DPADDLE_MOBILE_INFERENCE)
endif()
if (APPLE OR WIN32)
if (APPLE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL for building on mac and windows" FORCE)
"Disable MKL for building on mac" FORCE)
endif()
if (WIN32)
set(WITH_DSO OFF CACHE STRING
"Disable DSO when compiling for Windows" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when compiling for Windows" FORCE)
set(WITH_DISTRIBUTE OFF CACHE STRING
"Disable DISTRIBUTE when compiling for Windows" FORCE)
set(WITH_C_API OFF CACHE STRING

4
cmake/cudnn.cmake
Unescape View File

@ -44,9 +44,9 @@ if(WIN32)
set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll")
endif(WIN32)
if(Apple)
if(APPLE)
set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
endif(Apple)
endif(APPLE)
find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a
PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist}

43
cmake/external/mkldnn.cmake vendored
Unescape View File

@ -23,15 +23,14 @@ SET(MKLDNN_SOURCES_DIR ${THIRD_PARTY_PATH}/mkldnn)
SET(MKLDNN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/mkldnn)
SET(MKLDNN_INC_DIR "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
IF(WIN32 OR APPLE)
IF(APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with MKLDNN in Paddle yet."
"Mac is not supported with MKLDNN in Paddle yet."
"Force WITH_MKLDNN=OFF")
SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE)
SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in MacOS" FORCE)
return()
ENDIF()
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
@ -44,10 +43,14 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
ELSE()
MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
ENDIF()
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
IF(NOT WIN32)
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
ENDIF(NOT WIN32)
ExternalProject_Add(
${MKLDNN_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
@ -58,8 +61,15 @@ ExternalProject_Add(
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
CMAKE_ARGS -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
CMAKE_ARGS -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
CMAKE_ARGS -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
CMAKE_ARGS -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON
CMAKE_ARGS -DMKLROOT=${MKLML_ROOT}
CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
@ -67,6 +77,11 @@ ExternalProject_Add(
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
-DMKLROOT:PATH=${MKLML_ROOT}
)
if(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
else(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
endif(WIN32)
ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
@ -85,10 +100,14 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
# copy the real so.0 lib to install dir
# it can be directly contained in wheel or capi
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
DEPENDS mkldnn)
if(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
else(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
DEPENDS mkldnn)
endif(WIN32)
ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
IF(WITH_C_API)

63
cmake/external/mklml.cmake vendored
Unescape View File

@ -16,56 +16,67 @@ IF(NOT ${WITH_MKLML})
return()
ENDIF(NOT ${WITH_MKLML})
IF(WIN32 OR APPLE)
IF(APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with MKLML in Paddle yet."
"Mac is not supported with MKLML in Paddle yet."
"Force WITH_MKLML=OFF")
SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE)
return()
ENDIF()
INCLUDE(ExternalProject)
SET(MKLML_PROJECT "extern_mklml")
IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
MESSAGE(STATUS "use pre defined download url")
SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
SET(MKLML_DST_DIR "mklml")
SET(MKLML_INSTALL_ROOT "${THIRD_PARTY_PATH}/install")
SET(MKLML_INSTALL_DIR ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
SET(MKLML_ROOT ${MKLML_INSTALL_DIR})
SET(MKLML_INC_DIR ${MKLML_ROOT}/include)
SET(MKLML_LIB_DIR ${MKLML_ROOT}/lib)
SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
if(WIN32)
SET(MKLML_LIB ${MKLML_LIB_DIR}/mklml.lib)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.lib)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/mklml.dll)
SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.dll)
else()
SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
endif()
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
MESSAGE(STATUS "use pre defined download url")
if(WIN32)
SET(MKLML_VER "mklml_win_2019.0.20180710" CACHE STRING "" FORCE)
SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
else()
SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
ENDIF()
endif()
FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
"PROJECT(MKLML)\n"
"cmake_minimum_required(VERSION 3.0)\n"
"install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n"
" DESTINATION ${MKLML_DST_DIR})\n")
SET(MKLML_PROJECT "extern_mklml")
MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
ExternalProject_Add(
${MKLML_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${MKLML_SOURCE_DIR}
PREFIX ${MKLML_SOURCE_DIR}
URL ${MKLML_URL}
DOWNLOAD_DIR ${MKLML_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz
&& tar zxf ${MKLML_VER}.tgz
DOWNLOAD_NO_PROGRESS 1
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLML_INSTALL_ROOT}
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
UPDATE_COMMAND ""
INSTALL_COMMAND
${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/include ${MKLML_INC_DIR} &&
${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/lib ${MKLML_LIB_DIR}
)
INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})

6
cmake/generic.cmake
Unescape View File

@ -267,7 +267,11 @@ function(cc_library TARGET_NAME)
list(APPEND cc_library_DEPS dynload_mklml)
endif()
add_dependencies(${TARGET_NAME} mklml)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
if(WIN32)
target_link_libraries(${TARGET_NAME} ${MKLML_IOMP_LIB})
else(WIN32)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif(WIN32)
endif()
# remove link to python, see notes at:
# https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually

16
cmake/inference_lib.cmake
Unescape View File

@ -115,20 +115,20 @@ if (NOT PROTOBUF_FOUND OR WIN32)
)
endif ()
if (NOT CBLAS_FOUND)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
DEPS extern_openblas
)
elseif (WITH_MKLML)
if (WITH_MKLML)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml")
copy(mklml_lib
SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
DEPS mklml
)
elseif (NOT CBLAS_FOUND OR WIN32)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
DEPS extern_openblas
)
endif ()
if (WITH_MKLDNN)

18
paddle/fluid/API.spec
Unescape View File

@ -208,6 +208,7 @@ paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act
paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
paddle.fluid.layers.py_func ArgSpec(args=['func', 'x', 'out', 'backward_func', 'skip_vars_in_backward_input'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
@ -350,6 +351,23 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b
paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.add_strategy ArgSpec(args=['self', 'strategy'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.CompressPass.apply ArgSpec(args=['self', 'graph'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.ImitationGraph.__init__ ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.ImitationGraph.all_parameters ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.__init__ ArgSpec(args=['self', 'pruner', 'start_epoch', 'end_epoch', 'delta_rate', 'acc_loss_threshold', 'sensitivities'], varargs=None, keywords=None, defaults=(None, 0, 10, 0.2, 0.2, None))
paddle.fluid.contrib.SensitivePruneStrategy.on_batch_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_batch_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_compress_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_compress_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.MagnitudePruner.__init__ ArgSpec(args=['self', 'threshold'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.MagnitudePruner.prune ArgSpec(args=['self', 'param', 'threshold'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.RatioPruner.__init__ ArgSpec(args=['self', 'ratios'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.RatioPruner.prune ArgSpec(args=['self', 'param', 'ratio'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.load_persistables_for_increment ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.load_persistables_for_inference ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.convert_dist_to_sparse_program ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None)

8
paddle/fluid/framework/details/multi_devices_graph_pass.cc
Unescape View File

@ -355,9 +355,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
BuildStrategy::GradientScaleStrategy::kCustomized) {
// TODO(paddle-dev): Why is there no input for this op_handle?
auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
auto out_dtype = all_vars_.at(loss_grad_name)->GetDataType();
CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0],
out_dtype);
CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
}
// This assumes the backward generating code will ensure IsScaleLossOp
// is true only for the op that scale the final scalar loss.
@ -660,13 +658,13 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(
void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
ir::Graph *result, const std::string &loss_grad_name,
ir::Node *out_var_node, proto::VarType::Type dtype) const {
ir::Node *out_var_node) const {
for (size_t i = 0; i < places_.size(); ++i) {
// Insert ScaleCost OpHandle
auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
auto *op_handle = new ScaleLossGradOpHandle(
result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation),
local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx, dtype);
local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx);
result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
// FIXME: Currently ScaleLossGradOp only use device_count as scale

3
paddle/fluid/framework/details/multi_devices_graph_pass.h
Unescape View File

@ -68,8 +68,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
void CreateScaleLossGradOp(ir::Graph *result,
const std::string &loss_grad_name,
ir::Node *out_var_node,
proto::VarType::Type dtype) const;
ir::Node *out_var_node) const;
VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
int dst_dev_id) const;

61
paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
Unescape View File

@ -22,66 +22,39 @@ namespace details {
ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev,
Scope *scope,
platform::Place place,
platform::DeviceContext *dev_ctx,
proto::VarType::Type dtype)
platform::DeviceContext *dev_ctx)
: OpHandleBase(node),
coeff_(static_cast<float>(1.0 / num_dev)),
scope_(scope),
place_(place),
out_dtype_(dtype) {
place_(place) {
this->SetDeviceContext(place_, dev_ctx);
}
ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
struct ScaleLossGradFunctor {
float coeff_;
Tensor *out_;
platform::Place place_;
OpHandleBase *op_handle_;
proto::VarType::Type out_dtype_;
platform::DeviceContext *ctx_;
ScaleLossGradFunctor(float coeff, Tensor *out, platform::Place place,
OpHandleBase *op_handle, proto::VarType::Type dtype,
platform::DeviceContext *ctx)
: coeff_(coeff), out_(out), place_(place), out_dtype_(dtype), ctx_(ctx) {}
template <typename OutT>
void apply() const {
auto *out_data = out_->mutable_data<OutT>(place_);
if (platform::is_cpu_place(place_)) {
*out_data = static_cast<OutT>(coeff_);
} else {
#ifdef PADDLE_WITH_CUDA
OutT cast_coeff = static_cast<OutT>(coeff_);
auto stream = static_cast<platform::CUDADeviceContext *>(ctx_)->stream();
memory::Copy(boost::get<platform::CUDAPlace>(place_), out_data,
platform::CPUPlace(), &cast_coeff, SizeOfType(out_dtype_),
stream);
VLOG(10) << place_ << "RUN Scale loss grad op";
#endif
}
}
};
void ScaleLossGradOpHandle::RunImpl() {
// Doesn't wait any event
std::string var_name = static_cast<VarHandle *>(this->outputs_[0])->name_;
auto &local_scope = *scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto *tensor = local_scope.FindVar(var_name)->GetMutable<LoDTensor>();
tensor->Resize(make_ddim({1}));
float *tmp = local_scope.FindVar(var_name)
->GetMutable<LoDTensor>()
->mutable_data<float>(make_ddim({1}), place_);
if (platform::is_cpu_place(place_)) {
*tmp = coeff_;
} else {
#ifdef PADDLE_WITH_CUDA
ScaleLossGradFunctor func(coeff_, tensor, place_, this, out_dtype_,
this->dev_ctxes_.at(place_));
this->RunAndRecordEvent([&] { framework::VisitDataType(out_dtype_, func); });
#else
ScaleLossGradFunctor func(coeff_, tensor, place_, this, out_dtype_, nullptr);
framework::VisitDataType(out_dtype_, func);
this->RunAndRecordEvent([&] {
auto stream = static_cast<platform::CUDADeviceContext *>(
this->dev_ctxes_.at(place_))
->stream();
memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
platform::CPUPlace(), &coeff_, sizeof(float), stream);
VLOG(10) << place_ << "RUN Scale loss grad op";
});
#endif
}
}
std::string ScaleLossGradOpHandle::Name() const { return "Scale LossGrad"; }

5
paddle/fluid/framework/details/scale_loss_grad_op_handle.h
Unescape View File

@ -26,8 +26,8 @@ namespace details {
struct ScaleLossGradOpHandle : public OpHandleBase {
ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, Scope *scope,
platform::Place place, platform::DeviceContext *context,
proto::VarType::Type dtype);
platform::Place place,
platform::DeviceContext *context);
~ScaleLossGradOpHandle() final;
@ -40,7 +40,6 @@ struct ScaleLossGradOpHandle : public OpHandleBase {
float coeff_;
Scope *scope_;
platform::Place place_;
proto::VarType::Type out_dtype_;
};
} // namespace details

9
paddle/fluid/framework/lod_tensor.cc
Unescape View File

@ -157,13 +157,8 @@ bool CheckLoD(const LoD &in, int tensor_height) {
if (level.size() < 2) return false;
// check: the first offset(the begin offset) of each level should be 0.
if (level.front() != 0) return false;
// check: all the offsets in a level should be ascending(no same items
// allows).
if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) {
if (a < b) return true;
return false;
})) {
LOG(INFO) << "ascending error";
// check: all the offsets in a level should be ascending(allow same items)
if (!std::is_sorted(level.begin(), level.end())) {
return false;
}
}

5
paddle/fluid/framework/lod_tensor_test.cc
Unescape View File

@ -217,6 +217,11 @@ TEST(LoD, CheckLoD) {
// check with underlying tensor storage.
ASSERT_TRUE(CheckLoD(relative_lod, 5));
ASSERT_FALSE(CheckLoD(relative_lod, 9));
// check whether lod is ascending-sorted (allow same items)
ASSERT_TRUE(CheckLoD({{0, 1, 2, 3, 4, 5}}, 5));
ASSERT_TRUE(CheckLoD({{0, 1, 3, 3, 4, 5}}, 5));
ASSERT_FALSE(CheckLoD({{0, 1, 3, 2, 5}}, 5));
}
TEST(LoD, CheckAbsLoD) {

4
paddle/fluid/framework/ngraph_bridge.cc
Unescape View File

@ -31,10 +31,12 @@ std::map<std::string,
std::shared_ptr<std::unordered_map<
std::string, std::shared_ptr<ngraph::Node>>>)>>
NgraphBridge::NG_NODE_MAP = {
{"fill_constant", paddle::operators::ngraphs::BuildFillConstantNode},
{"mul", paddle::operators::ngraphs::BuildMulNode},
{"mul_grad", paddle::operators::ngraphs::BuildMulGradNode},
{"relu", paddle::operators::ngraphs::BuildUnaryNode<ngraph::op::Relu>},
{"tanh", paddle::operators::ngraphs::BuildUnaryNode<ngraph::op::Tanh>}};
{"tanh", paddle::operators::ngraphs::BuildUnaryNode<ngraph::op::Tanh>},
{"top_k", paddle::operators::ngraphs::BuildTopKNode}};
void NgraphBridge::BuildNgNode(const std::shared_ptr<OperatorBase>& op) {
auto& op_type = op->Type();

2
paddle/fluid/framework/op_desc.h
Unescape View File

@ -123,6 +123,8 @@ class OpDesc {
BlockDesc *Block() { return this->block_; }
const BlockDesc *Block() const { return this->block_; }
private:
template <typename MapType>
static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

22
paddle/fluid/framework/operator.cc
Unescape View File

@ -476,6 +476,28 @@ const Tensor* ExecutionContext::LegacyInput<Tensor>(
template <>
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
const std::string& name) const {
auto it = ctx_.inputs.find(name);
if (it == ctx_.inputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<const Tensor*> res;
res.reserve(vars.size());
std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](Variable* var) -> const Tensor* {
if (var == nullptr) return nullptr;
PADDLE_ENFORCE(
var->IsType<LoDTensor>(),
"should be LoDTensor, but the received type is %s",
var->Type().name());
return &(var->Get<LoDTensor>());
});
return res;
}
template <>
const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
const std::string& name) const {
auto names = op().Inputs(name);
std::vector<const Tensor*> res;
res.reserve(names.size());

65
paddle/fluid/framework/operator.h
Unescape View File

@ -197,8 +197,31 @@ class ExecutionContext {
const std::vector<const Variable*> MultiInputVar(
const std::string& name) const {
auto names = op_.Inputs(name);
auto it = ctx_.inputs.find(name);
if (it == ctx_.inputs.end()) {
return {};
}
std::vector<const Variable*> res;
res.reserve(it->second.size());
std::transform(it->second.begin(), it->second.end(),
std::back_inserter(res),
[this](Variable* var) { return var; });
return res;
}
std::vector<Variable*> MultiOutputVar(const std::string& name) const {
auto names = op_.Outputs(name);
auto it = ctx_.outputs.find(name);
if (it == ctx_.outputs.end()) {
return {};
}
return it->second;
}
const std::vector<Variable*> LegacyMultiInputVar(
const std::string& name) const {
auto names = op_.Inputs(name);
std::vector<Variable*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[this](const std::string& name) {
@ -208,7 +231,7 @@ class ExecutionContext {
return res;
}
std::vector<Variable*> MultiOutputVar(const std::string& name) const {
std::vector<Variable*> LegacyMultiOutputVar(const std::string& name) const {
auto names = op_.Outputs(name);
std::vector<Variable*> res;
res.reserve(names.size());
@ -250,6 +273,38 @@ class ExecutionContext {
template <typename T>
const std::vector<const T*> MultiInput(const std::string& name) const {
auto it = ctx_.inputs.find(name);
if (it == ctx_.inputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<const T*> res;
res.reserve(vars.size());
std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](Variable* var) -> const T* {
return var == nullptr ? nullptr : &var->Get<T>();
});
return res;
}
template <typename T>
std::vector<T*> MultiOutput(const std::string& name) const {
auto it = ctx_.outputs.find(name);
if (it == ctx_.outputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<T*> res;
res.reserve(vars.size());
std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](Variable* var) -> T* {
return var == nullptr ? nullptr : var->GetMutable<T>();
});
return res;
}
template <typename T>
const std::vector<const T*> LegacyMultiInput(const std::string& name) const {
auto names = op_.Inputs(name);
std::vector<const T*> res;
res.reserve(names.size());
@ -262,7 +317,7 @@ class ExecutionContext {
}
template <typename T>
std::vector<T*> MultiOutput(const std::string& name) const {
std::vector<T*> LegacyMultiOutput(const std::string& name) const {
auto names = op_.Outputs(name);
std::vector<T*> res;
res.reserve(names.size());
@ -321,6 +376,10 @@ template <>
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
const std::string& name) const;
template <>
const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
const std::string& name) const;
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;

2
paddle/fluid/framework/shape_inference.h
Unescape View File

@ -25,6 +25,8 @@ limitations under the License. */
namespace paddle {
namespace framework {
class OperatorBase;
using InferShapeVarPtr = boost::variant<VarDesc *, Variable *>;
class InferShapeContext {

11
paddle/fluid/framework/tensor.cc
Unescape View File

@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/var_type.h"
namespace paddle {
namespace framework {
@ -27,6 +28,9 @@ void Tensor::check_memory_size() const {
"or maybe the required data-type mismatches the data already stored.");
}
Tensor::Tensor(std::type_index type)
: type_(framework::ToDataType(type)), offset_(0) {}
size_t Tensor::memory_size() const {
return holder_ == nullptr ? 0UL : holder_->size() - offset_;
}
@ -101,5 +105,12 @@ const DDim& Tensor::dims() const { return dims_; }
int64_t Tensor::numel() const { return product(dims_); }
void Tensor::ResetHolder(std::shared_ptr<memory::Allocation> holder) {
if (holder_) {
PADDLE_ENFORCE_EQ(numel() * SizeOfType(type()), holder->size());
}
holder_ = holder;
}
} // namespace framework
} // namespace paddle

4
paddle/fluid/framework/tensor.h
Unescape View File

@ -69,6 +69,8 @@ class Tensor {
public:
Tensor() : type_(proto::VarType::FP32), offset_(0) {}
explicit Tensor(std::type_index type);
/*! Return a pointer to mutable memory block. */
template <typename T>
T* data();
@ -162,6 +164,8 @@ class Tensor {
return std::move(holder_);
}
void ResetHolder(std::shared_ptr<memory::Allocation> holder);
private:
/*! holds the memory block if allocated. */
std::shared_ptr<memory::Allocation> holder_;

7
paddle/fluid/inference/api/analysis_predictor.cc
Unescape View File

@ -231,11 +231,14 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
inputs[i].data.length());
} else {
#ifdef PADDLE_WITH_CUDA
platform::DeviceContextPool &pool =
platform::DeviceContextPool::Instance();
auto *dev_ctx =
static_cast<const platform::CUDADeviceContext *>(pool.Get(place_));
auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
platform::CPUPlace(), inputs[i].data.data(),
inputs[i].data.length(),
0); // stream 0 for sync copy
inputs[i].data.length(), dev_ctx->stream());
#else
PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif

7
paddle/fluid/inference/api/api_impl.cc
Unescape View File

@ -208,11 +208,14 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
inputs[i].data.length());
} else {
#ifdef PADDLE_WITH_CUDA
platform::DeviceContextPool &pool =
platform::DeviceContextPool::Instance();
auto *dev_ctx =
static_cast<const platform::CUDADeviceContext *>(pool.Get(place_));
auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
platform::CPUPlace(), inputs[i].data.data(),
inputs[i].data.length(),
0); // stream 0 for sync copy
inputs[i].data.length(), dev_ctx->stream());
#else
PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif

15
paddle/fluid/inference/api/demo_ci/CMakeLists.txt
Unescape View File

@ -89,12 +89,21 @@ endif()
if(WITH_MKL)
include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
if(NOT WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
else(WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5md${CMAKE_SHARED_LIBRARY_SUFFIX})
endif(WIN32)
set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
if(EXISTS ${MKLDNN_PATH})
include_directories("${MKLDNN_PATH}/include")
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
if(WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
else(WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
endif(WIN32)
endif()
else()
set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})

5
paddle/fluid/inference/tests/api/CMakeLists.txt
Unescape View File

@ -75,6 +75,11 @@ set(LAC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/lac")
download_model_and_data(${LAC_INSTALL_DIR} "lac_model.tar.gz" "lac_data.txt.tar.gz")
inference_analysis_api_test(test_analyzer_lac ${LAC_INSTALL_DIR} analyzer_lac_tester.cc)
# MM DNN
set(MM_DNN_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/mm_dnn")
download_model_and_data(${MM_DNN_INSTALL_DIR} "MM_DNN_model.tar.gz" "MM_DNN_data.txt.tar.gz")
inference_analysis_api_test(test_analyzer_mm_dnn ${MM_DNN_INSTALL_DIR} analyzer_mm_dnn_tester.cc)
# text_classification
set(TEXT_CLASSIFICATION_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/text_classification")
download_model_and_data(${TEXT_CLASSIFICATION_INSTALL_DIR} "text-classification-Senta.tar.gz" "text_classification_data.txt.tar.gz")

11
paddle/fluid/inference/tests/api/analyzer_dam_tester.cc
Unescape View File

@ -254,5 +254,16 @@ TEST(Analyzer_dam, compare) { compare(); }
TEST(Analyzer_dam, compare_mkldnn) { compare(true /* use_mkldnn */); }
#endif
// Compare Deterministic result
TEST(Analyzer_dam, compare_determine) {
AnalysisConfig cfg;
SetConfig(&cfg);
std::vector<std::vector<PaddleTensor>> input_slots_all;
SetInput(&input_slots_all);
CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
input_slots_all);
}
} // namespace inference
} // namespace paddle

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save