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

test=develop
panyx0718-patch-1
nhzlx 6 years ago
commit e5bf8616f0

@ -41,6 +41,7 @@ option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_F
option(WITH_AMD_GPU "Compile PaddlePaddle with AMD GPU" OFF)
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FOUND})
option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
@ -103,6 +104,8 @@ if(ANDROID OR IOS)
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when cross-compiling for Android and iOS" FORCE)
set(WITH_NGRAPH OFF CACHE STRING
"Disable nGraph when cross-compiling for Android and iOS" FORCE)
set(WITH_GOLANG OFF CACHE STRING
"Disable golang when cross-compiling for Android and iOS" FORCE)
@ -171,6 +174,7 @@ include(external/protobuf) # download, build, install protobuf
include(external/python) # download, build, install python
include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/ngraph) # download, build, install nGraph
include(external/swig) # download, build, install swig
include(external/boost) # download boost
include(external/any) # download libn::any

@ -37,7 +37,6 @@ SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h
IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
SET(MKLDNN_DEPENDS ${MKLML_PROJECT})

@ -0,0 +1,92 @@
# 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.
add_library(ngraph INTERFACE)
IF(WIN32 OR APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with nGraph in Paddle yet."
"Force WITH_NGRAPH=OFF")
SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph in Windows and MacOS" FORCE)
ENDIF()
IF(${WITH_NGRAPH} AND NOT ${WITH_MKLDNN})
MESSAGE(WARNING
"nGraph needs mkl-dnn to be enabled."
"Force WITH_NGRAPH=OFF")
SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph if mkl-dnn is disabled" FORCE)
ENDIF()
IF(NOT ${WITH_NGRAPH})
return()
ENDIF()
INCLUDE(ExternalProject)
SET(NGRAPH_PROJECT "extern_ngraph")
SET(NGRAPH_VERSION "0.9")
SET(NGRAPH_GIT_TAG "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0")
SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph)
SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph)
SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include)
SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION})
SET(NGRAPH_CPU_LIB_NAME libcpu_backend.so)
SET(NGRAPH_TBB_LIB_NAME libtbb.so.2)
SET(NGRAPH_GIT_REPO "https://github.com/NervanaSystems/ngraph.git")
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}/lib
)
if(UNIX AND NOT APPLE)
include(GNUInstallDirs)
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
else()
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib)
endif()
MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}")
SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
# Workaround for nGraph expecting mklml to be in mkldnn install directory.
ExternalProject_Add_Step(
${NGRAPH_PROJECT}
PrepareMKL
COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_LIB} ${MKLDNN_INSTALL_DIR}/lib/libmklml_intel.so
COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_IOMP_LIB} ${MKLDNN_INSTALL_DIR}/lib/libiomp5.so
DEPENDEES download
DEPENDERS configure
)
add_dependencies(ngraph ${NGRAPH_PROJECT})
target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
LIST(APPEND external_project_dependencies ngraph)

@ -30,9 +30,7 @@ UNSET_VAR(PROTOBUF_LITE_LIBRARY)
UNSET_VAR(PROTOBUF_LIBRARY)
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(Protobuf_PROTOC_EXECUTABLE)
if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined.
function(protobuf_generate_python SRCS)
function(protobuf_generate_python SRCS)
# shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
if(NOT ARGN)
message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
@ -52,7 +50,6 @@ if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate
else()
set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
endif()
if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
endif()
@ -77,19 +74,17 @@ if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate
set(FIL_WE "${FIL_DIR}/${FIL_WE}")
endif()
endif()
list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
add_custom_command(
OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
COMMAND ${Protobuf_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
DEPENDS ${ABS_FIL} ${Protobuf_PROTOC_EXECUTABLE}
COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
DEPENDS ${ABS_FIL} ${PROTOBUF_PROTOC_EXECUTABLE}
COMMENT "Running Python protocol buffer compiler on ${FIL}"
VERBATIM )
endforeach()
set(${SRCS} ${${SRCS}} PARENT_SCOPE)
endfunction()
endif()
endfunction()
# Print and set the protobuf library information,
# finish this cmake process and exit from this file.
@ -126,6 +121,7 @@ macro(PROMPT_PROTOBUF_LIB)
# FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`.
# make `protobuf_generate_cpp` happy.
SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE})
FOREACH(dep ${protobuf_DEPS})
ADD_DEPENDENCIES(protobuf ${dep})
ADD_DEPENDENCIES(protobuf_lite ${dep})

@ -103,7 +103,7 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode', 'return_softmax'], varargs=None, keywords=None, defaults=(False, -100, False, False))
paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
@ -179,10 +179,12 @@ paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], vara
paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
paddle.fluid.layers.similarity_focus ArgSpec(args=['input', 'axis', 'indexes', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
@ -201,6 +203,7 @@ paddle.fluid.layers.create_tensor ArgSpec(args=['dtype', 'name', 'persistable'],
paddle.fluid.layers.create_parameter ArgSpec(args=['shape', 'dtype', 'name', 'attr', 'is_bias', 'default_initializer'], varargs=None, keywords=None, defaults=(None, None, False, None))
paddle.fluid.layers.create_global_var ArgSpec(args=['shape', 'value', 'dtype', 'persistable', 'force_cpu', 'name'], varargs=None, keywords=None, defaults=(False, False, None))
paddle.fluid.layers.cast ArgSpec(args=['x', 'dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.tensor_array_to_tensor ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.concat ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.sums ArgSpec(args=['input', 'out'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.assign ArgSpec(args=['input', 'output'], varargs=None, keywords=None, defaults=(None,))
@ -271,6 +274,7 @@ paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, k
paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False))
paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, None))
paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False))
paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None))

@ -167,10 +167,12 @@ struct HitGroup {
bool Match(Node *node, PDNode *pat) {
if (nodes_.count(node)) {
if (!roles.count(pat)) return false;
return roles[pat] == node;
if (roles.count(pat) && roles[pat] == node) return true;
return false;
} else {
if (roles.count(pat) && roles[pat] != node) return false;
return true;
}
return !roles.count(pat) || roles.at(pat) == node;
}
void Register(Node *node, PDNode *pat) {
@ -198,7 +200,6 @@ GraphPatternDetector::DetectPatterns() {
std::vector<GraphPatternDetector::subgraph_t> result;
std::vector<HitGroup> init_groups;
std::array<std::vector<HitGroup>, 2> bi_records;
// PADDLE_ENFORCE(!pattern_.edges().empty(), "At least one edge is needed");
auto *first_pnode = pattern_.edges().empty() ? pattern().nodes().front().get()
: pattern_.edges().front().first;
if (!pdnodes2nodes_.count(first_pnode)) return result;
@ -228,11 +229,12 @@ GraphPatternDetector::DetectPatterns() {
VLOG(80) << "check " << source->id() << " -- " << target->id();
// TODO(Superjomn) add some prune strategies.
for (const auto &group : pre_groups) {
if (IsNodesLink(source, target)) {
HitGroup new_group = group;
if (IsNodesLink(source, target) &&
new_group.Match(source, edge.first)) {
bool flag = new_group.Match(source, edge.first) &&
new_group.Match(target, edge.second);
if (flag) {
new_group.Register(source, edge.first);
if (new_group.Match(target, edge.second)) {
new_group.Register(target, edge.second);
cur_groups.push_back(new_group);
// TODO(Superjomn) need to unique
@ -261,14 +263,16 @@ GraphPatternDetector::DetectPatterns() {
return result;
}
bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
struct GraphItemLessThan {
bool operator()(const std::pair<PDNode *, Node *> &a,
const std::pair<PDNode *, Node *> &b) {
if (a.first != b.first) {
return a.first < b.first;
} else {
return a.second < b.second;
}
}
}
};
// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550
@ -282,7 +286,7 @@ void GraphPatternDetector::UniquePatterns(
for (auto &g : *subgraphs) {
// Sort the items in the sub-graph, and transform to a string key.
std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemLessThan());
std::stringstream ss;
for (auto &item : sorted_keys) {
ss << item.first << ":" << item.second;

@ -259,6 +259,8 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
if (row_size >= 0) {
ss << "[row_size=" << row_size << "]";
}
std::string dtype = GetDtype(*scope, output.second[i]);
ss << ":" << dtype;
ss << "[" << GetDims(*scope, var_name, true) << "]";
ss << "(" << GetLoD(*scope, var_name) << ")";
}

@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/type_defs.h"
namespace paddle {
@ -24,5 +27,27 @@ class VarTypeInference {
virtual void operator()(const OpDesc& op_desc, BlockDesc* block) const = 0;
};
class PassInDtypeAndVarTypeToOutput : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc& op_desc,
framework::BlockDesc* block) const final {
auto in_out_var_names = this->GetInputOutputWithSameType();
for (auto& i_o_n : in_out_var_names) {
auto& x_name = op_desc.Input(i_o_n.first).at(0);
auto& out_name = op_desc.Output(i_o_n.second).at(0);
auto& x = block->FindRecursiveOrCreateVar(x_name);
auto& out = block->FindRecursiveOrCreateVar(out_name);
out.SetType(x.GetType());
out.SetDataType(x.GetDataType());
}
}
protected:
virtual std::unordered_map<std::string, std::string>
GetInputOutputWithSameType() const = 0;
};
} // namespace framework
} // namespace paddle

@ -113,7 +113,9 @@ void Analyzer::Run(Argument* argument) {
passes.push_back("infer_clean_graph_pass");
passes.push_back("graph_viz_pass"); // add graphviz for debug.
for (auto& pass : ir_passes_) {
if (!disabled_ir_passes_.count(pass)) {
// skip mkldnn pass when use_mkldnn_ = false;
bool skip_pass = (!use_mkldnn_) && pass.find("mkldnn") != std::string::npos;
if (!disabled_ir_passes_.count(pass) && !skip_pass) {
passes.push_back(pass);
passes.push_back("graph_viz_pass"); // add graphviz for debug.
}

@ -112,7 +112,7 @@ void DataFlowGraph::Build(const framework::proto::ProgramDesc &prog) {
out_alias->SetPbMsg(out->pb_msg());
var2id[out_alias->name()] =
out_alias->id(); // update variable's alias Node
LOG(INFO) << "loop found in graph, create SSA alias node ["
VLOG(40) << "loop found in graph, create SSA alias node ["
<< out_alias->repr() << "] for [" << out->repr() << "]";
out = out_alias;
}

@ -27,7 +27,7 @@ class ActivationOpConverter : public OpConverter {
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework::OpDesc op_desc(op, nullptr);
LOG(INFO)
VLOG(3)
<< "convert a fluid Activation op to tensorrt activation layer whose "
"type is "
<< op_type_;

@ -23,7 +23,7 @@ class BatchNormOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
LOG(INFO) << "convert a fluid batch norm op to tensorrt batch_norm";
VLOG(3) << "convert a fluid batch norm op to tensorrt batch_norm";
framework::OpDesc op_desc(op, nullptr);
PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);

@ -25,7 +25,7 @@ class ConcatOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40) << "convert a fluid mul op to tensorrt mul layer without bias";
VLOG(3) << "convert a fluid mul op to tensorrt mul layer without bias";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs

@ -37,8 +37,7 @@ class Conv2dOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
LOG(INFO)
<< "convert a fluid conv2d op to tensorrt conv layer without bias";
VLOG(3) << "convert a fluid conv2d op to tensorrt conv layer without bias";
framework::OpDesc op_desc(op, nullptr);
PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1);

@ -25,7 +25,7 @@ class DropoutOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40) << "convert a fluid dropout op to tensorrt dropout layer";
VLOG(3) << "convert a fluid dropout op to tensorrt dropout layer";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs
auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]);

@ -26,7 +26,7 @@ class ElementwiseWeightOpConverter : public OpConverter {
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework::OpDesc op_desc(op, nullptr);
LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer";
VLOG(3) << "convert a fluid elementwise op to tensorrt IScaleLayer";
PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight
@ -108,7 +108,7 @@ class ElementwiseTensorOpConverter : public OpConverter {
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework::OpDesc op_desc(op, nullptr);
LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer";
VLOG(3) << "convert a fluid elementwise op to tensorrt IScaleLayer";
PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight

@ -52,7 +52,7 @@ class FcOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40) << "convert a fluid fc op to tensorrt fc layer without bias";
VLOG(3) << "convert a fluid fc op to tensorrt fc layer without bias";
framework::OpDesc op_desc(op, nullptr);
PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);

@ -25,7 +25,7 @@ class MulOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40) << "convert a fluid mul op to tensorrt mul layer without bias";
VLOG(3) << "convert a fluid mul op to tensorrt mul layer without bias";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs

@ -25,7 +25,7 @@ class PadOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40) << "convert a fluid transpose op to tensorrt tranpose layer";
VLOG(3) << "convert a fluid transpose op to tensorrt tranpose layer";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs

@ -25,7 +25,7 @@ class Pool2dOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40)
VLOG(3)
<< "convert a fluid pool2d op to tensorrt pool2d layer without bias";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs

@ -25,7 +25,7 @@ class SoftMaxOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(40)
VLOG(3)
<< "convert a fluid softmax op to tensorrt softmax layer without bias";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs

@ -52,7 +52,7 @@ class NaiveLogger : public nvinfer1::ILogger {
void log(nvinfer1::ILogger::Severity severity, const char* msg) override {
switch (severity) {
case Severity::kINFO:
LOG(INFO) << msg;
VLOG(3) << msg;
break;
case Severity::kWARNING:
LOG(WARNING) << msg;

@ -110,5 +110,5 @@ if(WITH_GPU AND TENSORRT_FOUND)
endif()
cc_test(test_trt_models SRCS trt_models_tester.cc
ARGS --dirname=${TRT_MODEL_INSTALL_DIR}/trt_test_models
DEPS paddle_inference_tensorrt_subgraph_engine)
DEPS paddle_inference_tensorrt_subgraph_engine SERIAL)
endif()

@ -5,6 +5,8 @@ list(REMOVE_DUPLICATES GENERAL_OPS)
set(DEPS_OPS "")
set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operator/CMakeLists.txt. DO NOT EDIT!\n\n")
set(PART_CUDA_KERNEL_FILES)
function(op_library TARGET)
# op_library is a function to create op library. The interface is same as
# cc_library. But it handle split GPU/CPU code and link some common library
@ -37,6 +39,12 @@ function(op_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${TARGET}.cu)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu
${PART_CUDA_KERNEL_FILES} PARENT_SCOPE)
list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
endif()
@ -317,6 +325,7 @@ op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
op_library(save_combine_op DEPS lod_tensor)
op_library(load_combine_op DEPS lod_tensor)
op_library(tensor_array_to_tensor_op DEPS concat_op)
op_library(concat_op DEPS concat_and_split)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
@ -326,6 +335,8 @@ foreach(src ${GENERAL_OPS})
endforeach()
file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
if (NOT WIN32)
add_subdirectory(reader)
endif(NOT WIN32)
@ -352,3 +363,14 @@ if(NOT WIN32)
nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
endif()
nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)
if(WITH_GPU)
foreach(CUDA_KERNEL_FILE ${PART_CUDA_KERNEL_FILES})
file(READ ${CUDA_KERNEL_FILE} TARGET_CONTENT)
string(REGEX MATCH "REGISTER_OP_CUDA_KERNEL\\(\\n?([^,]+),.*" MATCHED ${TARGET_CONTENT})
if (MATCHED)
string(STRIP ${CMAKE_MATCH_1} MATCHED)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${MATCHED}, CUDA);\n")
endif()
endforeach()
endif()

@ -91,16 +91,12 @@ class ActivationOp : public framework::OperatorWithKernel {
}
};
class ActivationOpInferVarType : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc& op_desc,
framework::BlockDesc* block) const override {
auto x_name = op_desc.Input("X")[0];
auto out_name = op_desc.Output("Out")[0];
auto& x = block->FindRecursiveOrCreateVar(x_name);
auto& out = block->FindRecursiveOrCreateVar(out_name);
out.SetType(x.GetType());
out.SetDataType(x.GetDataType());
class ActivationOpInferVarType
: public framework::PassInDtypeAndVarTypeToOutput {
protected:
std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
const override {
return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
}
};

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

Loading…
Cancel
Save