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

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add_ut_for_trt

Browse Source
revert-13637-optimize-opyreader
nhzlx 7 years ago
parent baae7e4f63 f67483bf3b
commit 5c57e15044
71 changed files with 2497 additions and 1796 deletions
Show Stats Download Patch File Download Diff File

20
cmake/configure.cmake
Unescape View File

@ -62,8 +62,26 @@ if(NOT CMAKE_CROSSCOMPILING)
endif()
if(WIN32)
# windows stupid compile option for all targets.
# windows header option for all targets.
add_definitions(-D_XKEYCHECK_H)
# Use symbols instead of absolute path, reduce the cmake link command length.
SET(CMAKE_C_USE_RESPONSE_FILE_FOR_LIBRARIES 1)
SET(CMAKE_CXX_USE_RESPONSE_FILE_FOR_LIBRARIES 1)
SET(CMAKE_C_USE_RESPONSE_FILE_FOR_OBJECTS 1)
SET(CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS 1)
SET(CMAKE_C_USE_RESPONSE_FILE_FOR_INCLUDES 1)
SET(CMAKE_CXX_USE_RESPONSE_FILE_FOR_INCLUDES 1)
SET(CMAKE_C_RESPONSE_FILE_LINK_FLAG "@")
SET(CMAKE_CXX_RESPONSE_FILE_LINK_FLAG "@")
# Specify the program to use when building static libraries
SET(CMAKE_C_CREATE_STATIC_LIBRARY "<CMAKE_AR> lib <TARGET> <LINK_FLAGS> <OBJECTS>")
SET(CMAKE_CXX_CREATE_STATIC_LIBRARY "<CMAKE_AR> lib <TARGET> <LINK_FLAGS> <OBJECTS>")
# set defination for the dll export
if (NOT MSVC)
message(FATAL "Windows build only support msvc. Which was binded by the nvcc compiler of NVIDIA.")
endif(NOT MSVC)
endif(WIN32)
if(NOT WITH_GOLANG)

27
cmake/generic.cmake
Unescape View File

@ -110,6 +110,20 @@ function(find_fluid_modules TARGET_NAME)
endif()
endfunction(find_fluid_modules)
# find all third_party modules is used for paddle static library
# for reduce the dependency when building the inference libs.
set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY)
function(find_fluid_thirdparties TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "third_party" pos)
if(pos GREATER 1)
get_property(fluid_ GLOBAL PROPERTY FLUID_THIRD_PARTY)
set(fluid_third_partys ${fluid_third_partys} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY "${fluid_third_partys}")
endif()
endfunction(find_fluid_thirdparties)
function(merge_static_libs TARGET_NAME)
set(libs ${ARGN})
list(REMOVE_DUPLICATES libs)
@ -204,18 +218,13 @@ function(merge_static_libs TARGET_NAME)
foreach(lib ${libs})
# Get the file names of the libraries to be merged
#if(NOT $<TARGET_FILE:${lib}> MATCHES "lib.*\\.lib")
# message("library" ${lib})
# set(libfiles ${libfiles} lib$<TARGET_FILE:${lib}>)
#else()
set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
#endif()
endforeach()
# windows cmd return error in clean env.
# COMMAND del "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
# msvc will put libarary in directory of "/Release/xxxlib" by default
# COMMAND cmake -E remove "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.lib ${libfiles}
COMMAND cmake -E make_directory "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}"
COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/lib${TARGET_NAME}.lib ${libfiles}
)
endif(WIN32)
endfunction(merge_static_libs)

4
doc/README.md
Unescape View File

@ -2,6 +2,6 @@
Thanks for reading PaddlePaddle documentation.
Since **September 17th, 2018**, the **0.15.0 and develop** documentation source has been moved to [Fluiddoc Repo](https://github.com/PaddlePaddle/Paddle) and updated in Fluiddoc Repo.
Since **September 17th, 2018**, the **0.15.0 and develop** documentation source has been moved to [FluidDoc Repo](https://github.com/PaddlePaddle/FluidDoc) and updated there.
Please turn to Fluiddoc Repo for the latest documentation.
Please turn to FluidDoc Repo for the latest documentation.

43
paddle/fluid/API.spec
Unescape View File

@ -22,9 +22,6 @@ paddle.fluid.Operator.rename_input ArgSpec(args=['self', 'old_name', 'new_name']
paddle.fluid.Operator.rename_output ArgSpec(args=['self', 'old_name', 'new_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Operator.set_attr ArgSpec(args=['self', 'name', 'val'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Operator.to_string ArgSpec(args=['self', 'throw_on_error'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Parameter.__init__ ArgSpec(args=['self', 'block', 'shape', 'dtype'], varargs=None, keywords='kwargs', defaults=None)
paddle.fluid.Parameter.astype ArgSpec(args=['self', 'dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Parameter.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
@ -35,29 +32,16 @@ paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None,
paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False))
paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
paddle.fluid.Trainer.__init__ ArgSpec(args=['self', 'train_func', 'optimizer_func', 'param_path', 'place', 'parallel', 'checkpoint_config'], varargs=None, keywords=None, defaults=(None, None, False, None))
paddle.fluid.Trainer.save_inference_model ArgSpec(args=['self', 'param_path', 'feeded_var_names', 'target_var_indexes'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.save_params ArgSpec(args=['self', 'param_path'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.test ArgSpec(args=['self', 'reader', 'feed_order'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.train ArgSpec(args=['self', 'num_epochs', 'event_handler', 'reader', 'feed_order'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.BeginEpochEvent.__init__ ArgSpec(args=['self', 'epoch_id'], varargs=None, keywords=None, defaults=None)
paddle.fluid.EndEpochEvent.__init__ ArgSpec(args=['self', 'epoch_id'], varargs=None, keywords=None, defaults=None)
paddle.fluid.BeginStepEvent.__init__ ArgSpec(args=['self', 'epoch_id', 'step_id'], varargs=None, keywords=None, defaults=None)
paddle.fluid.EndStepEvent.__init__ ArgSpec(args=['self', 'epoch_id', 'step_id', 'metrics'], varargs=None, keywords=None, defaults=None)
paddle.fluid.CheckpointConfig.__init__ ArgSpec(args=['self', 'checkpoint_dir', 'max_num_checkpoints', 'epoch_interval', 'step_interval'], varargs=None, keywords=None, defaults=(None, 3, 1, 10))
paddle.fluid.Inferencer.__init__ ArgSpec(args=['self', 'infer_func', 'param_path', 'place', 'parallel'], varargs=None, keywords=None, defaults=(None, False))
paddle.fluid.Inferencer.infer ArgSpec(args=['self', 'inputs', 'return_numpy'], varargs=None, keywords=None, defaults=(True,))
paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.DistributeTranspiler.get_trainer_program ArgSpec(args=['self', 'wait_port'], varargs=None, keywords=None, defaults=(True,))
paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None))
paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program', 'current_endpoint'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None, '127.0.0.1:6174'))
paddle.fluid.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0))
paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.DistributeTranspilerConfig.__init__
paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id', 'scope'], varargs=None, keywords='kwargs', defaults=(None, None, None, None, None, 1, 0, None))
paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id', 'scope'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 1, 0, None))
paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True))
paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None
paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None
@ -178,14 +162,14 @@ paddle.fluid.layers.unstack ArgSpec(args=['x', 'axis', 'num'], varargs=None, key
paddle.fluid.layers.sequence_enumerate ArgSpec(args=['input', 'win_size', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.expand ArgSpec(args=['x', 'expand_times', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_concat ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.scale ArgSpec(args=['x', 'scale', 'bias', 'bias_after_scale', 'act', 'name'], varargs=None, keywords=None, defaults=(1.0, 0.0, True, None, None))
paddle.fluid.layers.elementwise_add ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.elementwise_div ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.elementwise_sub ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.elementwise_mul ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.elementwise_max ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.elementwise_min ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.elementwise_pow ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
paddle.fluid.layers.scale ArgSpec(args=['x', 'scale', 'bias', 'bias_after_scale', 'out', 'act', 'name'], varargs=None, keywords=None, defaults=(1.0, 0.0, True, None, None, None))
paddle.fluid.layers.elementwise_add ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
paddle.fluid.layers.elementwise_div ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
paddle.fluid.layers.elementwise_sub ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
paddle.fluid.layers.elementwise_mul ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
paddle.fluid.layers.elementwise_max ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
paddle.fluid.layers.elementwise_min ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
paddle.fluid.layers.elementwise_pow ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, 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)
@ -265,12 +249,12 @@ paddle.fluid.layers.sum ArgSpec(args=[], varargs='args', keywords='kwargs', defa
paddle.fluid.layers.slice ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.layers.shape ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.layers.maxout ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.layers.softshrink ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.layers.sigmoid ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.logsigmoid ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.exp ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.tanh ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.tanh_shrink ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.softshrink ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sqrt ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.abs ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.ceil ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
@ -336,7 +320,7 @@ paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['
paddle.fluid.transpiler.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.transpiler.DistributeTranspiler.get_trainer_program ArgSpec(args=['self', 'wait_port'], varargs=None, keywords=None, defaults=(True,))
paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None))
paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program', 'current_endpoint'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None, '127.0.0.1:6174'))
paddle.fluid.transpiler.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0))
paddle.fluid.transpiler.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.transpiler.HashName.__init__ ArgSpec(args=['self', 'pserver_endpoints'], varargs=None, keywords=None, defaults=None)
@ -350,6 +334,7 @@ paddle.fluid.nets.simple_img_conv_pool ArgSpec(args=['input', 'num_filters', 'fi
paddle.fluid.nets.sequence_conv_pool ArgSpec(args=['input', 'num_filters', 'filter_size', 'param_attr', 'act', 'pool_type'], varargs=None, keywords=None, defaults=(None, 'sigmoid', 'max'))
paddle.fluid.nets.glu ArgSpec(args=['input', 'dim'], varargs=None, keywords=None, defaults=(-1,))
paddle.fluid.nets.scaled_dot_product_attention ArgSpec(args=['queries', 'keys', 'values', 'num_heads', 'dropout_rate'], varargs=None, keywords=None, defaults=(1, 0.0))
paddle.fluid.nets.img_conv_group ArgSpec(args=['input', 'conv_num_filter', 'pool_size', 'conv_padding', 'conv_filter_size', 'conv_act', 'param_attr', 'conv_with_batchnorm', 'conv_batchnorm_drop_rate', 'pool_stride', 'pool_type', 'use_cudnn', 'use_mkldnn'], varargs=None, keywords=None, defaults=(1, 3, None, None, False, 0.0, 1, 'max', True, False))
paddle.fluid.optimizer.SGDOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'regularization', 'name'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.optimizer.SGDOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'use_nesterov', 'regularization', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
@ -389,7 +374,7 @@ paddle.fluid.CPUPlace.__init__ __init__(self: paddle.fluid.core.CPUPlace) -> Non
paddle.fluid.CUDAPlace.__init__ __init__(self: paddle.fluid.core.CUDAPlace, arg0: int) -> None
paddle.fluid.CUDAPinnedPlace.__init__ __init__(self: paddle.fluid.core.CUDAPinnedPlace) -> None
paddle.fluid.ParamAttr.__init__ ArgSpec(args=['self', 'name', 'initializer', 'learning_rate', 'regularizer', 'trainable', 'gradient_clip', 'do_model_average'], varargs=None, keywords=None, defaults=(None, None, 1.0, None, True, None, False))
paddle.fluid.WeightNormParamAttr.__init__ ArgSpec(args=['self', 'dim'], varargs=None, keywords='kwargs', defaults=(None,))
paddle.fluid.WeightNormParamAttr.__init__ ArgSpec(args=['self', 'dim', 'name', 'initializer', 'learning_rate', 'regularizer', 'trainable', 'gradient_clip', 'do_model_average'], varargs=None, keywords=None, defaults=(None, None, None, 1.0, None, True, None, False))
paddle.fluid.DataFeeder.__init__ ArgSpec(args=['self', 'feed_list', 'place', 'program'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.DataFeeder.decorate_reader ArgSpec(args=['self', 'reader', 'multi_devices', 'num_places', 'drop_last'], varargs=None, keywords=None, defaults=(None, True))
paddle.fluid.DataFeeder.feed ArgSpec(args=['self', 'iterable'], varargs=None, keywords=None, defaults=None)

21
paddle/fluid/framework/eigen.h
Unescape View File

@ -13,6 +13,11 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
// logging.h and windows.h conflict
#define GLOG_NO_ABBREVIATED_SEVERITIES
// solve static linking error in windows
// https://github.com/google/glog/issues/301
#define GOOGLE_GLOG_DLL_DECL
#include "paddle/fluid/framework/tensor.h"
#include "unsupported/Eigen/CXX11/Tensor"
@ -46,11 +51,13 @@ struct EigenTensor {
using ConstType =
Eigen::TensorMap<Eigen::Tensor<const T, D, MajorType, IndexType>>;
static Type From(Tensor& tensor, DDim dims) {
static Type From(Tensor& tensor, DDim dims) { // NOLINT
return Type(tensor.data<T>(), EigenDim<D>::From(dims));
}
static Type From(Tensor& tensor) { return From(tensor, tensor.dims_); }
static Type From(Tensor& tensor) { // NOLINT
return From(tensor, tensor.dims_);
} // NOLINT
static ConstType From(const Tensor& tensor, DDim dims) {
return ConstType(tensor.data<T>(), EigenDim<D>::From(dims));
@ -64,7 +71,8 @@ struct EigenTensor {
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {
static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) {
static typename EigenMatrix::Type Reshape(Tensor& tensor, // NOLINT
int num_col_dims) {
int rank = tensor.dims_.size();
PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
"`num_col_dims` must be between (0, rank_of_tensor).");
@ -86,11 +94,12 @@ template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
// Flatten reshapes a Tensor into an EigenVector.
static typename EigenVector::Type Flatten(Tensor& tensor) {
static typename EigenVector::Type Flatten(Tensor& tensor) { // NOLINT
return EigenVector::From(tensor, {product(tensor.dims_)});
}
static typename EigenVector::ConstType Flatten(const Tensor& tensor) {
static typename EigenVector::ConstType Flatten(
const Tensor& tensor) { // NOLINT
return EigenVector::From(tensor, {product(tensor.dims_)});
}
};
@ -104,7 +113,7 @@ struct EigenScalar {
using ConstType = Eigen::TensorMap<
Eigen::TensorFixedSize<const T, Eigen::Sizes<>, MajorType, IndexType>>;
static Type From(Tensor& tensor) { return Type(tensor.data<T>()); }
static Type From(Tensor& tensor) { return Type(tensor.data<T>()); } // NOLINT
static ConstType From(const Tensor& tensor) {
return ConstType(tensor.data<T>());

34
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc
Unescape View File

@ -26,8 +26,6 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("conv_relu_mkldnn_fuse", graph.get());
std::unordered_set<Node*> nodes2delete;
GraphPatternDetector gpd;
auto* conv_input = gpd.mutable_pattern()
->NewNode("conv_relu_mkldnn_fuse/conv_input")
@ -42,36 +40,20 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
Graph* g) {
VLOG(4) << "handle ConvReLU fuse";
GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
conv_relu_pattern); // Filter
GET_IR_NODE_FROM_SUBGRAPH(conv_bias, conv_bias, conv_relu_pattern); // Bias
GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern); // tmp
conv_relu_pattern); // Filter
GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern); // tmp
GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_relu_pattern); // CONV op
GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern); // Out
GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern); // ReLU op
// Create an ConvReLU Node.
OpDesc desc;
std::string conv_relu_i_in = subgraph.at(conv_input)->Name();
std::string conv_relu_w_in = conv_weight->Name();
std::string conv_relu_b_in = conv_bias->Name();
std::string conv_relu_out = relu_out->Name();
desc.SetInput("Input", std::vector<std::string>({conv_relu_i_in}));
desc.SetInput("Filter", std::vector<std::string>({conv_relu_w_in}));
desc.SetInput("Bias", std::vector<std::string>({conv_relu_b_in}));
desc.SetOutput("Output", std::vector<std::string>({conv_relu_out}));
desc.SetType("conv2d");
for (auto& attr : conv->Op()->GetAttrMap()) {
desc.SetAttr(attr.first, attr.second);
}
desc.SetAttr("fuse_relu", true);
auto conv_relu_node = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {conv, relu, conv_out});
// Transform Conv node into ConvReLU node.
OpDesc* desc = conv->Op();
desc->SetOutput("Output", std::vector<std::string>({relu_out->Name()}));
desc->SetAttr("fuse_relu", true);
GraphSafeRemoveNodes(graph.get(), {relu, conv_out});
PADDLE_ENFORCE(subgraph.count(conv_input));
IR_NODE_LINK_TO(subgraph.at(conv_input), conv_relu_node);
IR_NODE_LINK_TO(conv_weight, conv_relu_node);
IR_NODE_LINK_TO(conv_bias, conv_relu_node);
IR_NODE_LINK_TO(conv_relu_node, relu_out);
IR_NODE_LINK_TO(conv, relu_out);
found_conv_relu_count++;
};

17
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc
Unescape View File

@ -85,16 +85,13 @@ TEST(ConvReLUFusePass, basic) {
for (auto* node : graph->Nodes()) {
if (node->IsOp() && node->Op()->Type() == "conv2d") {
if (node->Op()->HasAttr("use_mkldnn")) {
bool use_mkldnn = boost::get<bool>(node->Op()->GetAttr("use_mkldnn"));
if (use_mkldnn) {
if (node->Op()->HasAttr("fuse_relu")) {
bool fuse_relu = boost::get<bool>(node->Op()->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
}
}
}
auto* op = node->Op();
ASSERT_TRUE(op->HasAttr("use_mkldnn"));
EXPECT_TRUE(boost::get<bool>(op->GetAttr("use_mkldnn")));
ASSERT_TRUE(op->HasAttr("fuse_relu"));
bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
}
}
}

8
paddle/fluid/framework/ir/graph_pattern_detector.cc
Unescape View File

@ -638,11 +638,6 @@ PDNode *patterns::ConvReLU::operator()(
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Filter");
// Bias
auto *conv_bias_var = pattern->NewNode(conv_bias_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Bias");
// intermediate variable, will be removed in the IR after fuse.
auto *conv_out_var = pattern->NewNode(conv_out_repr())
->AsIntermediate()
@ -653,8 +648,7 @@ PDNode *patterns::ConvReLU::operator()(
->AsOutput()
->assert_is_op_output("relu");
conv_op->LinksFrom({conv_input, conv_weight_var, conv_bias_var})
.LinksTo({conv_out_var});
conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
relu_op->LinksFrom({conv_out_var}).LinksTo({relu_out_var});
return relu_out_var;
}

3
paddle/fluid/framework/ir/graph_pattern_detector.h
Unescape View File

@ -379,7 +379,7 @@ struct PatternBase {
// op: conv + relu
// named nodes:
// conv_input, conv_weight,
// conv_bias, conv_out, conv,
// conv_out, conv,
// relu_out, relu
struct ConvReLU : public PatternBase {
ConvReLU(PDPattern* pattern, const std::string& name_scope)
@ -392,7 +392,6 @@ struct ConvReLU : public PatternBase {
PATTERN_DECL_NODE(relu);
// declare variable node's name
PATTERN_DECL_NODE(conv_weight);
PATTERN_DECL_NODE(conv_bias);
PATTERN_DECL_NODE(conv_out);
PATTERN_DECL_NODE(relu_out);
};

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

@ -132,7 +132,9 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
AddAttr<std::string>(OpNamescopeAttrName(), "Operator name with namesope.")
.SetDefault("");
AddAttr<std::vector<std::string>>(OpCreationCallstackAttrName(),
"Callstack for Op Creatation.")
.SetDefault({});
Validate();
}

1
paddle/fluid/framework/op_proto_maker.h
Unescape View File

@ -46,6 +46,7 @@ class OpProtoAndCheckerMaker {
static const char *OpRoleAttrName() { return "op_role"; }
static const char *OpRoleVarAttrName() { return "op_role_var"; }
static const char *OpNamescopeAttrName() { return "op_namescope"; }
static const char *OpCreationCallstackAttrName() { return "op_callstack"; }
void operator()(proto::OpProto *proto, OpAttrChecker *attr_checker);

23
paddle/fluid/framework/op_registry.h
Unescape View File

@ -23,6 +23,11 @@ limitations under the License. */
#include <unordered_map>
#include <unordered_set>
#if defined(_WIN32)
#define GLOG_NO_ABBREVIATED_SEVERITIES // msvc conflict logging with windows.h
#define GOOGLE_GLOG_DLL_DECL
#endif
#include "glog/logging.h" // For VLOG()
#include "paddle/fluid/framework/attribute.h"
#include "paddle/fluid/framework/details/op_registry.h"
@ -241,22 +246,20 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
* we will use and tell the compiler to
* link them into target.
*/
#define USE_OP_ITSELF(op_type) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__use_op_itself_##op_type, \
"USE_OP_ITSELF must be called in global namespace"); \
extern int TouchOpRegistrar_##op_type(); \
static int use_op_itself_##op_type##_ __attribute__((unused)) = \
TouchOpRegistrar_##op_type()
#define USE_OP_ITSELF(op_type) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__use_op_itself_##op_type, \
"USE_OP_ITSELF must be called in global namespace"); \
extern int TouchOpRegistrar_##op_type(); \
UNUSED static int use_op_itself_##op_type##_ = TouchOpRegistrar_##op_type()
#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__use_op_kernel_##op_type##_##LIBRARY_TYPE##__, \
"USE_OP_DEVICE_KERNEL must be in global namespace"); \
extern int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE(); \
static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_ \
__attribute__((unused)) = \
TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE()
UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_ = \
TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE()
// TODO(fengjiayi): The following macros
// seems ugly, do we have better method?

60
paddle/fluid/framework/operator.cc
Unescape View File

@ -11,15 +11,20 @@ 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. */
#define GLOG_NO_ABBREVIATED_SEVERITIES
#define GOOGLE_GLOG_DLL_DECL
#include "paddle/fluid/framework/operator.h"
#include <gflags/gflags.h>
#include <glog/logging.h>
#include <algorithm>
#include <sstream>
#include <string>
#include <vector>
#include "paddle/fluid/framework/data_transform.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/framework/shape_inference.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/platform/profiler.h"
@ -137,19 +142,48 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
}
void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
VLOG(4) << place << " " << DebugStringEx(&scope);
if (platform::is_gpu_place(place)) {
try {
if (VLOG_IS_ON(4)) {
VLOG(4) << place << " " << DebugStringEx(&scope);
}
if (platform::is_gpu_place(place)) {
#ifndef PADDLE_WITH_CUDA
PADDLE_THROW("Cannot run operator on place %s", place);
PADDLE_THROW("Cannot run operator on place %s", place);
#else
auto dev_id = boost::get<platform::CUDAPlace>(place).device;
platform::SetDeviceId(dev_id);
auto dev_id = boost::get<platform::CUDAPlace>(place).device;
platform::SetDeviceId(dev_id);
#endif
}
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
platform::RecordEvent record_event(Type(), pool.Get(place));
RunImpl(scope, place);
if (VLOG_IS_ON(3)) {
VLOG(3) << place << " " << DebugStringEx(&scope);
}
} catch (platform::EnforceNotMet exception) {
if (Attrs().count("sub_block") != 0) {
throw exception;
}
auto& callstack = Attr<std::vector<std::string>>(
OpProtoAndCheckerMaker::OpCreationCallstackAttrName());
if (callstack.empty()) {
throw exception;
}
std::ostringstream sout;
sout << "Invoke operator " << Type() << " error.\n";
sout << "Python Callstacks: \n";
for (auto& line : callstack) {
sout << line;
}
sout << "C++ Callstacks: \n";
sout << exception.err_str_;
exception.err_str_ = sout.str();
throw exception;
} catch (...) {
std::rethrow_exception(std::current_exception());
}
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
platform::RecordEvent record_event(Type(), pool.Get(place));
RunImpl(scope, place);
VLOG(3) << place << " " << DebugStringEx(&scope);
}
bool OperatorBase::HasInputs(const std::string& name) const {
@ -177,7 +211,7 @@ const std::vector<std::string>& OperatorBase::Inputs(
}
bool OperatorBase::HasOutputs(const std::string& name) const {
if (outputs_.find(name) != outputs_.end()) {
if (outputs_.end() != outputs_.find(name)) {
return true;
} else {
return false;

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

@ -20,6 +20,8 @@ limitations under the License. */
#include <tuple>
#include <unordered_map>
#include <vector>
#define GLOG_NO_ABBREVIATED_SEVERITIES
#define GOOGLE_GLOG_DLL_DECL
#include "glog/logging.h" // For VLOG
#include "paddle/fluid/framework/attribute.h"

21
paddle/fluid/inference/api/analysis_predictor.cc
Unescape View File

@ -71,15 +71,15 @@ bool AnalysisPredictor::Init(
inference_program_ = paddle::inference::Load(
executor_.get(), scope_.get(), config_.prog_file, config_.param_file);
} else {
LOG(ERROR) << "fail to load inference model.";
LOG(ERROR) << "fail to load inference model from " << config_.model_dir;
return false;
}
OptimizeInferenceProgram();
ctx_ = executor_->Prepare(*inference_program_, 0);
if (config_._use_mkldnn) {
executor_->EnableMKLDNN(*inference_program_);
}
ctx_ = executor_->Prepare(*inference_program_, 0);
VLOG(5) << "to create variables";
PADDLE_ENFORCE(scope_.get());
@ -109,8 +109,9 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
}
argument_.origin_program_desc.reset(
new ProgramDesc(*inference_program_->Proto()));
PADDLE_ENFORCE(config_.ir_mode == AnalysisConfig::IrPassMode::kExclude,
"Only kExclude is supported yet.");
PADDLE_ENFORCE(
config_.ir_mode == contrib::AnalysisConfig::IrPassMode::kExclude,
"Only kExclude is supported yet.");
Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_);
CHECK(argument_.transformed_program_desc);
@ -126,8 +127,9 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
}
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig& config) {
std::unique_ptr<PaddlePredictor>
CreatePaddlePredictor<contrib::AnalysisConfig, PaddleEngineKind::kAnalysis>(
const contrib::AnalysisConfig& config) {
VLOG(3) << "create AnalysisConfig";
if (config.use_gpu) {
// 1. GPU memeroy
@ -154,4 +156,11 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
return predictor;
}
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>(
const contrib::AnalysisConfig& config) {
return CreatePaddlePredictor<contrib::AnalysisConfig,
PaddleEngineKind::kAnalysis>(config);
}
} // namespace paddle

4
paddle/fluid/inference/api/analysis_predictor.h
Unescape View File

@ -30,7 +30,7 @@ using framework::proto::ProgramDesc;
*/
class AnalysisPredictor : public NativePaddlePredictor {
public:
explicit AnalysisPredictor(const AnalysisConfig& config)
explicit AnalysisPredictor(const contrib::AnalysisConfig& config)
: NativePaddlePredictor(config), config_(config) {}
bool Init(const std::shared_ptr<framework::Scope>& parent_scope);
@ -46,7 +46,7 @@ class AnalysisPredictor : public NativePaddlePredictor {
Argument& analysis_argument() { return argument_; }
private:
AnalysisConfig config_;
contrib::AnalysisConfig config_;
Argument argument_;
};

18
paddle/fluid/inference/api/api_anakin_engine.cc
Unescape View File

@ -31,21 +31,24 @@
namespace paddle {
using paddle::contrib::AnakinConfig;
template <typename Target>
PaddleInferenceAnakinPredictor<Target>::PaddleInferenceAnakinPredictor(
const AnakinConfig &config) {
const contrib::AnakinConfig &config) {
CHECK(Init(config));
}
template <>
PaddleInferenceAnakinPredictor<anakin::X86>::PaddleInferenceAnakinPredictor(
const AnakinConfig &config) {
const contrib::AnakinConfig &config) {
omp_set_dynamic(0);
omp_set_num_threads(1);
mkl_set_num_threads(1);
CHECK(Init(config));
}
template <typename Target>
bool PaddleInferenceAnakinPredictor<Target>::Init(const AnakinConfig &config) {
bool PaddleInferenceAnakinPredictor<Target>::Init(
const contrib::AnakinConfig &config) {
if (!(graph_.load(config.model_file))) {
VLOG(3) << "fail to load graph from " << config.model_file;
return false;
@ -200,10 +203,11 @@ template class PaddleInferenceAnakinPredictor<anakin::X86>;
// A factory to help create difference predictor.
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
AnakinConfig, PaddleEngineKind::kAnakin>(const AnakinConfig &config) {
std::unique_ptr<PaddlePredictor>
CreatePaddlePredictor<contrib::AnakinConfig, PaddleEngineKind::kAnakin>(
const contrib::AnakinConfig &config) {
VLOG(3) << "Anakin Predictor create.";
if (config.target_type == AnakinConfig::NVGPU) {
if (config.target_type == contrib::AnakinConfig::NVGPU) {
#ifdef PADDLE_WITH_CUDA
VLOG(3) << "Anakin Predictor create on [ NVIDIA GPU ].";
std::unique_ptr<PaddlePredictor> x(
@ -213,7 +217,7 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
LOG(ERROR) << "AnakinConfig::NVGPU could not used in ONLY-CPU environment";
return nullptr;
#endif
} else if (config.target_type == AnakinConfig::X86) {
} else if (config.target_type == contrib::AnakinConfig::X86) {
VLOG(3) << "Anakin Predictor create on [ Intel X86 ].";
std::unique_ptr<PaddlePredictor> x(
new PaddleInferenceAnakinPredictor<anakin::X86>(config));

2
paddle/fluid/inference/api/api_anakin_engine.h
Unescape View File

@ -29,6 +29,8 @@ limitations under the License. */
namespace paddle {
using contrib::AnakinConfig;
template <typename Target>
class PaddleInferenceAnakinPredictor : public PaddlePredictor {
public:

77
paddle/fluid/inference/api/api_impl.cc
Unescape View File

@ -22,6 +22,7 @@ limitations under the License. */
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/inference/api/api_impl.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/timer.h"
#include "paddle/fluid/platform/profiler.h"
@ -101,14 +102,11 @@ bool NativePaddlePredictor::Init(
inference_program_ = paddle::inference::Load(
executor_.get(), scope_.get(), config_.prog_file, config_.param_file);
} else {
LOG(ERROR) << "fail to load inference model.";
LOG(ERROR) << "fail to load inference model from " << config_.model_dir;
return false;
}
ctx_ = executor_->Prepare(*inference_program_, 0);
if (config_._use_mkldnn) {
executor_->EnableMKLDNN(*inference_program_);
}
executor_->CreateVariables(*inference_program_,
sub_scope_ ? sub_scope_ : scope_.get(), 0);
@ -218,57 +216,20 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
template <typename T>
void NativePaddlePredictor::GetFetchOne(const framework::LoDTensor &fetch,
PaddleTensor *output) {
std::vector<int> shape;
auto dims_i = fetch.dims();
auto lod = fetch.lod();
const T *output_ptr = fetch.data<T>();
auto num = fetch.numel();
std::vector<T> data;
if (0 == lod.size()) {
std::copy(output_ptr, output_ptr + num, std::back_inserter(data));
for (int j = 0; j < dims_i.size(); ++j) {
shape.push_back(dims_i[j]);
}
} else {
// for batch detection
// image[0] -> output[0] shape {145, 6}
// image[1] -> output[1] shape {176, 6}
// then,
// the batch output shape {321, 6}
// the lod {{0, 145, 321}}
// so we should append output[0] to {176, 6}
size_t max_dim = 0;
for (size_t j = 1; j < lod[0].size(); j++) {
max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]);
}
size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back();
if (max_dim > 0) {
data.resize((lod[0].size() - 1) * max_dim * common_dim, 0);
}
for (size_t j = 1; j < lod[0].size(); j++) {
size_t start = lod[0][j - 1] * common_dim;
size_t end = lod[0][j] * common_dim;
if (end > start) {
std::copy(output_ptr + start, output_ptr + end,
data.begin() + (j - 1) * max_dim * common_dim);
}
}
shape.push_back(lod[0].size() - 1);
shape.push_back(max_dim);
for (int j = 1; j < dims_i.size(); ++j) {
shape.push_back(dims_i[j]);
}
}
output->shape = shape;
auto &buffer = output->data;
if (buffer.empty() || buffer.length() < sizeof(T) * data.size()) {
buffer.Resize(sizeof(T) * data.size());
}
std::memcpy(buffer.data(), data.data(), sizeof(T) * data.size());
// copy LoD
for (const auto &level : fetch.lod()) {
output->lod.emplace_back(level);
// set shape.
auto shape = framework::vectorize(fetch.dims());
output->shape.assign(shape.begin(), shape.end());
// set data.
const T *data = fetch.data<T>();
int num_elems = inference::VecReduceToInt(shape);
output->data.Resize(num_elems * sizeof(T));
// The fetched tensor output by fetch op, should always in CPU memory, so just
// copy.
memcpy(output->data.data(), data, num_elems * sizeof(T));
// set lod
output->lod.clear();
for (auto &level : fetch.lod()) {
output->lod.emplace_back(level.begin(), level.end());
}
}
@ -330,4 +291,10 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
#endif
}
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<NativeConfig>(
const NativeConfig &config) {
return CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
}
} // namespace paddle

9
paddle/fluid/inference/api/api_impl.h
Unescape View File

@ -14,15 +14,22 @@
#pragma once
// logging.h and windows.h conflict
#define GLOG_NO_ABBREVIATED_SEVERITIES
// solve static linking error in windows
// https://github.com/google/glog/issues/301
#define GOOGLE_GLOG_DLL_DECL
#include <glog/logging.h>
#include <map>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/profiler.h"

16
paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc
Unescape View File

@ -25,10 +25,11 @@ using inference::analysis::Argument;
using inference::Singleton;
using inference::analysis::Analyzer;
using framework::proto::ProgramDesc;
using paddle::contrib::MixedRTConfig;
class TensorRTSubgraphPredictor : public NativePaddlePredictor {
public:
explicit TensorRTSubgraphPredictor(const TensorRTConfig& config)
explicit TensorRTSubgraphPredictor(const MixedRTConfig& config)
: NativePaddlePredictor(config), config_(config) {}
bool Init(const std::shared_ptr<framework::Scope>& parent_scope) {
@ -121,13 +122,13 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor {
}
private:
TensorRTConfig config_;
MixedRTConfig config_;
};
template <>
std::unique_ptr<PaddlePredictor>
CreatePaddlePredictor<TensorRTConfig, PaddleEngineKind::kAutoMixedTensorRT>(
const TensorRTConfig& config) {
CreatePaddlePredictor<MixedRTConfig, PaddleEngineKind::kAutoMixedTensorRT>(
const MixedRTConfig& config) {
VLOG(3) << "create TensorRTSubgraphPredictor";
if (config.use_gpu) {
// 1. GPU memeroy
@ -156,6 +157,13 @@ CreatePaddlePredictor<TensorRTConfig, PaddleEngineKind::kAutoMixedTensorRT>(
return std::move(predictor);
}
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<MixedRTConfig>(
const MixedRTConfig& config) {
return CreatePaddlePredictor<MixedRTConfig,
PaddleEngineKind::kAutoMixedTensorRT>(config);
}
} // namespace paddle
USE_TRT_CONVERTER(elementwise_add_weight);

6
paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc
Unescape View File

@ -20,6 +20,8 @@
namespace paddle {
using contrib::MixedRTConfig;
DEFINE_string(dirname, "", "Directory of the inference model.");
void CompareTensorRTWithFluid(bool enable_tensorrt) {
@ -32,7 +34,7 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
config0.fraction_of_gpu_memory = 0.3;
config0.device = 0;
TensorRTConfig config1;
MixedRTConfig config1;
config1.model_dir = FLAGS_dirname + "word2vec.inference.model";
config1.use_gpu = true;
config1.fraction_of_gpu_memory = 0.3;
@ -42,7 +44,7 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
auto predictor0 =
CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config0);
auto predictor1 =
CreatePaddlePredictor<TensorRTConfig,
CreatePaddlePredictor<MixedRTConfig,
PaddleEngineKind::kAutoMixedTensorRT>(config1);
for (int batch_id = 0; batch_id < 1; batch_id++) {

49
paddle/fluid/inference/api/demo_ci/CMakeLists.txt
Unescape View File

@ -1,13 +1,32 @@
cmake_minimum_required(VERSION 3.0)
project(cpp_inference_demo CXX C)
option(WITH_MKL "Compile demo with MKL/OpenBlas support, default use MKL." ON)
option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." OFF)
option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON)
macro(safe_set_static_flag)
foreach(flag_var
CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
if(${flag_var} MATCHES "/MD")
string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
endif(${flag_var} MATCHES "/MD")
endforeach(flag_var)
endmacro()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
if (WIN32)
set(CMAKE_STATIC_LIBRARY_PREFIX "lib")
if (WITH_STATIC_LIB)
safe_set_static_flag()
add_definitions(-DSTATIC_LIB)
set(CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS} "/w")
set(CMAKE_CXX_FLAGS_RELEASE ${CMAKE_CXX_FLAGS_RELEASE} "/w")
endif()
set(CMAKE_STATIC_LIBRARY_PREFIX "lib")
else()
set(CMAKE_STATIC_LIBRARY_PREFIX "")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
set(CMAKE_STATIC_LIBRARY_PREFIX "")
endif()
message("flags" ${CMAKE_CXX_FLAGS})
if(NOT DEFINED PADDLE_LIB)
message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib")
@ -16,14 +35,18 @@ if(NOT DEFINED DEMO_NAME)
message(FATAL_ERROR "please set DEMO_NAME with -DDEMO_NAME=demo_name")
endif()
option(WITH_MKL "Compile demo with MKL/OpenBlas support, default use MKL." ON)
option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." OFF)
option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON)
if(WITH_GPU)
set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library")
if(NOT WIN32)
set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library")
else()
if(CUDA_LIB STREQUAL "")
set(CUDA_LIB "C:\\Program\ Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\lib\\x64")
endif()
endif(NOT WIN32)
endif()
include_directories("D:/Paddle/")
include_directories("${PADDLE_LIB}")
include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
include_directories("${PADDLE_LIB}/third_party/install/glog/include")
@ -83,10 +106,18 @@ set(DEPS ${DEPS}
${MATH_LIB} ${MKLDNN_LIB}
${CMAKE_STATIC_LIBRARY_PREFIX}glog ${CMAKE_STATIC_LIBRARY_PREFIX}gflags ${CMAKE_STATIC_LIBRARY_PREFIX}protobuf
${EXTERNAL_LIB})
# NOTE(dzhwinter) shlwapi is deprecated.
set(DEPS ${DEPS} libcmt shlwapi)
endif(NOT WIN32)
if(WITH_GPU)
set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX})
if(NOT WIN32)
set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX})
else()
set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} )
set(DEPS ${DEPS} ${CUDA_LIB}/cublas${CMAKE_STATIC_LIBRARY_SUFFIX} )
set(DEPS ${DEPS} ${CUDA_LIB}/cudnn${CMAKE_STATIC_LIBRARY_SUFFIX} )
endif()
endif()
target_link_libraries(${DEMO_NAME} ${DEPS})

8
paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc
Unescape View File

@ -18,6 +18,8 @@ limitations under the License. */
#include <gflags/gflags.h>
#include <glog/logging.h>
#include <algorithm>
#include <memory>
#include <thread> //NOLINT
#include "paddle/fluid/inference/paddle_inference_api.h"
@ -67,7 +69,8 @@ void Main(bool use_gpu) {
0.000932706};
const size_t num_elements = outputs.front().data.length() / sizeof(float);
// The outputs' buffers are in CPU memory.
for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
for (size_t i = 0; i < std::min(static_cast<size_t>(5), num_elements);
i++) {
PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],
result[i]);
}
@ -113,7 +116,8 @@ void MainThreads(int num_threads, bool use_gpu) {
const size_t num_elements =
outputs.front().data.length() / sizeof(float);
// The outputs' buffers are in CPU memory.
for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
for (size_t i = 0; i < std::min(static_cast<size_t>(5), num_elements);
i++) {
PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],
result[i]);
}

19
paddle/fluid/inference/api/demo_ci/windows_inference.md
Unescape View File

@ -0,0 +1,19 @@
# windows inference
本文介绍windows inference，目前只提供了静态编译，编译出paddle_fluid.lib，包含了除openblas.dll之外的所有第三方依赖库。
1. 下载最新的paddle_fluid.lib和openblas.dll，并把它们放在同一个目录下。
2. 准备预训练好的模型文件，例如models中的模型，可以将模型用safe_inference_model接口保存下来。将模型文件放到该目录下
3. 进入Paddle/paddle/fluid/inference/api/demo_ci目录，新建build目录，然后使用cmake生成vs2015的solution文件。
其中PADDLE_LIB是前面的paddle_fluid.lib对应文件夹, CUDA_LIB指定为x64格式下的cuda系统库目录文件夹。
```shell
cmake .. -G "Visual Studio 14 2015 Win64" -DWITH_GPU=ON -DWITH_MKL=OFF -DWITH_STATIC_LIB=ON -DCMAKE_BUILD_TYPE=Release -DDEMO_NAME=inference_icnet -DPADDLE_LIB=D:\to_the_paddle_fluid.lib -DCUDA_LIB=D:\tools\v8.0\lib\x64
```
然后用vs2015打开对应的项目文件，注意使用静态链接 "/MT"，生成对应的exe。将openblas.dll放到exe所在目录。
4. 该exe即为项目生成文件，可绑定运行。
## FAQ
1. cmake需要您手动下载，并添加到系统路径里
2. 路径中的不要包含空格，例如发现CUDA_LIB路径是Program Files(x86)可能会出错。可以将CUDA拷贝到一个新位置。

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

Write Preview
Loading…
Cancel
Save