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

Merge remote-tracking branch 'ups/develop' into refine/jit

Browse Source
revert-16045-imperative_remove_desc
tensor-tang 6 years ago
parent 45bdd84dac a38db3cb99
commit 6ff230a624
79 changed files with 2290 additions and 658 deletions
Show Stats Download Patch File Download Diff File

5
Dockerfile
Unescape View File

@ -75,8 +75,9 @@ RUN curl -s -q https://glide.sh/get | sh
# and its size is only one-third of the official one.
# 2. Manually add ~IPluginFactory() in IPluginFactory class of NvInfer.h, otherwise, it couldn't work in paddle.
# See https://github.com/PaddlePaddle/Paddle/issues/10129 for details.
RUN wget -qO- http://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
tar -xz -C /usr/local && \
RUN wget -q https://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz --no-check-certificate && \
tar -zxf TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz -C /usr/local && \
cp -rf /usr/local/TensorRT/include /usr && \
cp -rf /usr/local/TensorRT/lib /usr

20
cmake/external/mkldnn.cmake vendored
Unescape View File

@ -31,9 +31,17 @@ IF(APPLE)
return()
ENDIF()
MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
# Introduce variables:
# * CMAKE_INSTALL_LIBDIR
INCLUDE(GNUInstallDirs)
SET(LIBDIR "lib")
if(CMAKE_INSTALL_LIBDIR MATCHES ".*lib64$")
SET(LIBDIR "lib64")
endif()
MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/l${LIBDIR} to runtime path")
SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/${LIBDIR}")
INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
@ -58,7 +66,7 @@ ExternalProject_Add(
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/intel/mkl-dnn.git"
GIT_TAG "830a10059a018cd2634d94195140cf2d8790a75a"
GIT_TAG "863ff6e7042cec7d2e29897fe9f0872e0888b0fc"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@ -79,9 +87,9 @@ ExternalProject_Add(
-DMKLROOT:PATH=${MKLML_ROOT}
)
if(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
else(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
endif(WIN32)
ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
@ -101,7 +109,7 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
# copy the real so.0 lib to install dir
# it can be directly contained in wheel or capi
if(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/bin/mkldnn.dll)
else(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}

10
paddle/contrib/float16/README.md
Unescape View File

@ -5,13 +5,13 @@ Kexin Zhao <zhaokexin01@baidu.com>
## Introduction
Deep learning is usually a two-stage work: training and inference. The training stage estimates model parameters (weights) from data. The inference stage loads the weights and uses them to interpret inputs. Typically, weights are 32-bit float values (float32). Some new devices, including NVIDIA Volta GPUs, support higher speed computation using 16-bit float values (float16).
This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
## What is float16?
float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float or float32 data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has a high tolerance for the loss of precision and range when using float16 to represent the weights, and the inference accuracy will only be minimally affected in most cases, which gives us the opportunity to use float16 data type to speed up the inference.
Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
## Why float16?
The trend in today's deep learning community is to use bigger and deeper model, which translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float32 are correspondingly three-fold:
@ -24,7 +24,7 @@ The trend in today's deep learning community is to use bigger and deeper model,
## Fluid implementation of float16 inference
### Overview
Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block.
Fluid use [Program](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block.
### Basic requirement
When an executor runs an operator, it uses a kernel to perform computations on tensors contained in the input variables, and then writes the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs.
@ -75,7 +75,7 @@ In this scenario, we already have a float32 inference program and some associate
We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. Consequently, we need to add cast operators in the float16 inference program for conversions between the float16 tensor and float32 tensor.
The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
### Experiment results
Simply running the following commands to reproduce the experiment results presented in this section:
@ -162,7 +162,7 @@ We find that the speedup provided by float16 inference starts relatively small a
We also did the same benchmark on a single NVIDIA GeForce GTX 1080 Ti GPU that does not support Tensor Core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then delivers around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that Tensor Core, which is specialized for float16 computations, is a critical component of high performance float16 inference.
Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for complete benchmark results.
Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/float16/float16_benchmark.md) for complete benchmark results.
### Summary
1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode.

4
paddle/fluid/API.spec
Unescape View File

@ -144,7 +144,7 @@ paddle.fluid.layers.label_smooth (ArgSpec(args=['label', 'prior_dist', 'epsilon'
paddle.fluid.layers.roi_pool (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)), ('document', 'c317aa595deb31649083c8faa91cdb97'))
paddle.fluid.layers.roi_align (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)), ('document', '12c5bbb8b38c42e623fbc47611d766e1'))
paddle.fluid.layers.dice_loss (ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)), ('document', '1ba0508d573f65feecf3564dce22aa1d'))
paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', 'b3ecb819454832885c1f0f3ab9a5b938'))
paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', '7a1966d7c3a48f1fc0881cdaf5d83b0b'))
paddle.fluid.layers.image_resize_short (ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)), ('document', '06211aefc50c5a3e940d7204d859cdf7'))
paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'e4fb4ed511b2293b8f04f7e872afbfd7'))
paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', '735fa9758a6d7ff3b47d7b827f961c1d'))
@ -221,6 +221,7 @@ paddle.fluid.layers.psroi_pool (ArgSpec(args=['input', 'rois', 'output_channels'
paddle.fluid.layers.teacher_student_sigmoid_loss (ArgSpec(args=['input', 'label', 'soft_max_up_bound', 'soft_max_lower_bound'], varargs=None, keywords=None, defaults=(15.0, -15.0)), ('document', '2f6ff96864054a31aa4bb659c6722c99'))
paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None), ('document', '431a4301c35032166ec029f7432c80a7'))
paddle.fluid.layers.tree_conv (ArgSpec(args=['nodes_vector', 'edge_set', 'output_size', 'num_filters', 'max_depth', 'act', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(1, 2, 'tanh', None, None, None)), ('document', '34ea12ac9f10a65dccbc50100d12e607'))
paddle.fluid.layers.npair_loss (ArgSpec(args=['anchor', 'positive', 'labels', 'l2_reg'], varargs=None, keywords=None, defaults=(0.002,)), ('document', '46994d10276dd4cb803b4062b5d14329'))
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)), ('document', '33bbd42027d872b3818b3d64ec52e139'))
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)), ('document', 'b1ae2e1cc0750e58726374061ea90ecc'))
paddle.fluid.layers.read_file (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', 'b0a1c2fc51c27a106da28f3308c41f5e'))
@ -329,6 +330,7 @@ paddle.fluid.layers.polygon_box_transform (ArgSpec(args=['input', 'name'], varar
paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '991e934c3e09abf0edec7c9c978b4691'))
paddle.fluid.layers.box_clip (ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '397e9e02b451d99c56e20f268fa03f2e'))
paddle.fluid.layers.multiclass_nms (ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)), ('document', 'ca7d1107b6c5d2d6d8221039a220fde0'))
paddle.fluid.layers.distribute_fpn_proposals (ArgSpec(args=['fpn_rois', 'min_level', 'max_level', 'refer_level', 'refer_scale', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '7bb011ec26bace2bc23235aa4a17647d'))
paddle.fluid.layers.box_decoder_and_assign (ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'box_score', 'box_clip', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '005a5ae47d6c8fff721931d69d072b9f'))
paddle.fluid.layers.accuracy (ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None)), ('document', '9808534c12c5e739a10f73ebb0b4eafd'))
paddle.fluid.layers.auc (ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1)), ('document', 'e0e95334fce92d16c2d9db6e7caffc47'))

5
paddle/fluid/framework/ir/fuse_pass_base.h
Unescape View File

@ -14,6 +14,7 @@
#pragma once
#include <string>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/framework/scope.h"
@ -24,6 +25,10 @@ namespace ir {
static const char kParamScopeAttr[] = "__param_scope__";
static const char kFuseStatisAttr[] = "__fuse_statis__";
// When we use trt or other third_party lib, the parameters are managed by
// the lib, but not the fluid. So we need to record them to avoid duplicate
// allocation.
static const char kRepetitiveParamAttr[] = "__repetitive_param__";
enum FuseOptions {
DO_NOT_FUSE, // fusing will not be done

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

@ -130,15 +130,21 @@ std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
if (adj_list.find(n) == adj_list.end()) {
adj_list[n] = std::unordered_set<ir::Node *>();
}
std::vector<ir::Node *> nodes;
for (auto &var : n->inputs) {
for (auto &adj_n : var->inputs) {
PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
<< " -> " << n->Name() << reinterpret_cast<void *>(n)
<< " via " << var->Name() << reinterpret_cast<void *>(var);
adj_list[n].insert(adj_n);
nodes.push_back(adj_n);
}
}
std::sort(nodes.begin(), nodes.end(), [](ir::Node *node1, ir::Node *node2) {
return node1->id() > node2->id();
});
adj_list[n].insert(std::make_move_iterator(nodes.begin()),
std::make_move_iterator(nodes.end()));
}
return adj_list;
}

6
paddle/fluid/inference/analysis/argument.h
Unescape View File

@ -23,8 +23,12 @@
#pragma once
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
@ -133,6 +137,8 @@ struct Argument {
DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int);
DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode,
AnalysisConfig::Precision);
DECL_ARGUMENT_FIELD(tensorrt_use_static_engine, TensorRtUseStaticEngine,
bool);
// Memory optimized related.
DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool);

31
paddle/fluid/inference/analysis/helper.h
Unescape View File

@ -17,10 +17,12 @@ limitations under the License. */
#include <sys/stat.h>
#include <cstdio>
#include <fstream>
#include <memory>
#include <set>
#include <string>
#include <typeindex>
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/framework.pb.h"
@ -217,6 +219,35 @@ static std::string GetTrtCalibTableData(const std::string &model_opt_cache_dir,
return "";
}
static std::string GetTrtEngineSerializedPath(const std::string &model_root,
const std::string &engine_key) {
return model_root + "/trt_serialized_" + engine_key;
}
static std::string GetTrtEngineSerializedData(
const std::string &model_opt_cache_dir, const std::string &engine_key) {
std::string trt_serialized_path =
GetTrtEngineSerializedPath(model_opt_cache_dir, engine_key);
if (FileExists(trt_serialized_path)) {
VLOG(3) << "Trt serialized file: " << trt_serialized_path
<< "is found here";
std::ifstream infile(trt_serialized_path, std::ios::in);
std::stringstream buffer;
buffer << infile.rdbuf();
std::string trt_engine_serialized_data(buffer.str());
return trt_engine_serialized_data;
}
return "";
}
static void SaveTrtEngineSerializedDataToFile(
const std::string &trt_serialized_path,
const std::string &engine_serialized_data) {
std::ofstream outfile(trt_serialized_path);
outfile << engine_serialized_data;
outfile.close();
}
} // namespace analysis
} // namespace inference
} // namespace paddle

3
paddle/fluid/inference/analysis/ir_pass_manager.cc
Unescape View File

@ -81,6 +81,9 @@ void IRPassManager::CreatePasses(Argument *argument,
pass->Set(
"model_opt_cache_dir",
new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
pass->Set("gpu_device_id", new int(argument->gpu_device_id()));
pass->Set("use_static_engine",
new bool(argument->tensorrt_use_static_engine()));
}
pre_pass = pass_name;

3
paddle/fluid/inference/analysis/ir_pass_manager.h
Unescape View File

@ -22,7 +22,10 @@
#pragma once
#include <memory>
#include <string>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"

267
paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
View File

File diff suppressed because it is too large Load Diff

12
paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.h
Unescape View File

@ -13,7 +13,12 @@
// limitations under the License.
#pragma once
#include <paddle/fluid/framework/ir/fuse_pass_base.h>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
@ -26,8 +31,9 @@ class TensorRtSubgraphPass : public framework::ir::FusePassBase {
std::unique_ptr<framework::ir::Graph> graph) const override;
private:
void CreateTensorRTOp(framework::ir::Node *x,
framework::ir::Graph *graph) const;
void CreateTensorRTOp(framework::ir::Node *x, framework::ir::Graph *graph,
const std::vector<std::string> &graph_params,
std::vector<std::string> *repetitive_params) const;
void CleanIntermediateOutputs(framework::ir::Node *node);
};

11
paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc
Unescape View File

@ -31,6 +31,13 @@ void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
// The parameters are on the cpu, therefore, synchronization is not necessary.
if (!argument->use_gpu()) return;
auto &graph = argument->main_graph();
std::vector<std::string> repetitive_params;
if (graph.Has(framework::ir::kRepetitiveParamAttr))
repetitive_params = graph.Get<std::vector<std::string>>(
framework::ir::kRepetitiveParamAttr);
LOG(INFO) << "Sync params from CPU to GPU";
PADDLE_ENFORCE(argument->gpu_device_id_valid());
@ -43,6 +50,10 @@ void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
// Because there exists the case that new parameter variables are not added to
// the program in the analysis pass.
for (auto &var_name : all_vars) {
if (std::count(repetitive_params.begin(), repetitive_params.end(),
var_name)) {
continue;
}
auto *var = scope->FindLocalVar(var_name);
PADDLE_ENFORCE(var != nullptr);
if (var->IsType<framework::LoDTensor>() ||

1
paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h
Unescape View File

@ -17,6 +17,7 @@
#include <string>
#include <vector>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/platform/place.h"

4
paddle/fluid/inference/api/analysis_config.cc
Unescape View File

@ -103,6 +103,7 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
CP_MEMBER(tensorrt_max_batchsize_);
CP_MEMBER(tensorrt_min_subgraph_size_);
CP_MEMBER(tensorrt_precision_mode_);
CP_MEMBER(trt_use_static_engine_);
// MKLDNN related.
CP_MEMBER(use_mkldnn_);
CP_MEMBER(mkldnn_enabled_op_types_);
@ -144,7 +145,7 @@ void AnalysisConfig::EnableMKLDNN() {
void AnalysisConfig::EnableTensorRtEngine(
int workspace_size, int max_batch_size, int min_subgraph_size,
AnalysisConfig::Precision precision_mode) {
AnalysisConfig::Precision precision_mode, bool use_static) {
#ifdef PADDLE_WITH_CUDA
if (!use_gpu()) {
LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
@ -156,6 +157,7 @@ void AnalysisConfig::EnableTensorRtEngine(
tensorrt_max_batchsize_ = max_batch_size;
tensorrt_min_subgraph_size_ = min_subgraph_size;
tensorrt_precision_mode_ = precision_mode;
trt_use_static_engine_ = use_static;
Update();
#else

39
paddle/fluid/inference/api/analysis_predictor.cc
Unescape View File

@ -243,6 +243,8 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
input_ptr = input.mutable_data<int64_t>(ddim, place_);
} else if (inputs[i].dtype == PaddleDType::FLOAT32) {
input_ptr = input.mutable_data<float>(ddim, place_);
} else if (inputs[i].dtype == PaddleDType::INT32) {
input_ptr = input.mutable_data<int32_t>(ddim, place_);
} else {
LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
return false;
@ -326,8 +328,11 @@ bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
} else if (type == framework::proto::VarType::INT64) {
GetFetchOne<int64_t>(fetch, output);
output->dtype = PaddleDType::INT64;
} else if (type == framework::proto::VarType::INT32) {
GetFetchOne<int32_t>(fetch, output);
output->dtype = PaddleDType::INT32;
} else {
LOG(ERROR) << "unknown type, only support float32 and int64 now.";
LOG(ERROR) << "unknown type, only support float32, int64 and int32 now.";
}
}
return true;
@ -365,6 +370,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
argument_.SetTensorRtPrecisionMode(config_.tensorrt_precision_mode_);
argument_.SetTensorRtUseStaticEngine(config_.trt_use_static_engine_);
}
if (config_.use_mkldnn_) {
@ -438,12 +444,14 @@ void AnalysisPredictor::PrepareFeedFetch() {
}
feeds_[idx] = op;
feed_names_[op->Output("Out")[0]] = idx;
idx2feeds_[idx] = op->Output("Out")[0];
} else if (op->Type() == "fetch") {
int idx = boost::get<int>(op->GetAttr("col"));
if (fetches_.size() <= static_cast<size_t>(idx)) {
fetches_.resize(idx + 1);
}
fetches_[idx] = op;
idx2fetches_[idx] = op->Input("X")[0];
}
}
}
@ -456,6 +464,22 @@ void AnalysisPredictor::CreateFeedFetchVar(framework::Scope *scope) {
var->GetMutable<framework::FeedFetchList>();
}
std::vector<std::string> AnalysisPredictor::GetInputNames() {
std::vector<std::string> input_names;
for (auto &item : idx2feeds_) {
input_names.push_back(item.second);
}
return input_names;
}
std::vector<std::string> AnalysisPredictor::GetOutputNames() {
std::vector<std::string> output_names;
for (auto &item : idx2fetches_) {
output_names.push_back(item.second);
}
return output_names;
}
std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
const std::string &name) {
PADDLE_ENFORCE(executor_->scope()->FindVar(name), "no name called %s", name);
@ -463,6 +487,13 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
res->input_or_output_ = true;
res->SetName(name);
if (platform::is_cpu_place(place_)) {
res->SetPlace(PaddlePlace::kCPU);
} else {
auto gpu_place = boost::get<platform::CUDAPlace>(place_);
res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
}
return res;
}
@ -473,6 +504,12 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
res->input_or_output_ = false;
res->SetName(name);
if (platform::is_cpu_place(place_)) {
res->SetPlace(PaddlePlace::kCPU);
} else {
auto gpu_place = boost::get<platform::CUDAPlace>(place_);
res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
}
return res;
}

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

@ -15,12 +15,14 @@
#pragma once
#include <algorithm>
#include <map>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/api/api_impl.h"
#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/string/printf.h"
#ifdef PADDLE_WITH_TESTING
@ -53,6 +55,9 @@ class AnalysisPredictor : public PaddlePredictor {
std::vector<PaddleTensor> *output_data,
int batch_size = -1) override;
std::vector<std::string> GetInputNames();
std::vector<std::string> GetOutputNames();
std::unique_ptr<ZeroCopyTensor> GetInputTensor(
const std::string &name) override;
std::unique_ptr<ZeroCopyTensor> GetOutputTensor(
@ -131,7 +136,11 @@ class AnalysisPredictor : public PaddlePredictor {
std::shared_ptr<framework::ProgramDesc> inference_program_;
std::vector<framework::OpDesc *> feeds_;
std::map<std::string, size_t> feed_names_;
// Sorted according to the idx.
std::map<size_t, std::string> idx2feeds_;
std::vector<framework::OpDesc *> fetches_;
std::map<size_t, std::string> idx2fetches_;
// Memory buffer for feed inputs. The temporary LoDTensor will cause serious
// concurrency problems, wrong results and memory leak, so cache them.
std::vector<framework::LoDTensor> feed_tensors_;

2
paddle/fluid/inference/api/api.cc
Unescape View File

@ -28,6 +28,8 @@ int PaddleDtypeSize(PaddleDType dtype) {
return sizeof(float);
case PaddleDType::INT64:
return sizeof(int64_t);
case PaddleDType::INT32:
return sizeof(int32_t);
default:
assert(false);
return -1;

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

@ -203,6 +203,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
input_ptr = input.mutable_data<int64_t>(ddim, place_);
} else if (inputs[i].dtype == PaddleDType::FLOAT32) {
input_ptr = input.mutable_data<float>(ddim, place_);
} else if (inputs[i].dtype == PaddleDType::INT32) {
input_ptr = input.mutable_data<int32_t>(ddim, place_);
} else {
LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
return false;
@ -281,8 +283,11 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
} else if (type == framework::DataTypeTrait<int64_t>::DataType) {
GetFetchOne<int64_t>(fetch, output);
output->dtype = PaddleDType::INT64;
} else if (type == framework::DataTypeTrait<int32_t>::DataType) {
GetFetchOne<int32_t>(fetch, output);
output->dtype = PaddleDType::INT32;
} else {
LOG(ERROR) << "unknown type, only support float32 and int64 now.";
LOG(ERROR) << "unknown type, only support float32, int64 and int32 now.";
}
}
return true;

3
paddle/fluid/inference/api/api_impl_tester.cc
Unescape View File

@ -42,6 +42,9 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
} else if (t->type() == framework::proto::VarType::FP32) {
pt.data.Reset(t->data<void>(), t->numel() * sizeof(float));
pt.dtype = PaddleDType::FLOAT32;
} else if (t->type() == framework::proto::VarType::INT32) {
pt.data.Reset(t->data<void>(), t->numel() * sizeof(int32_t));
pt.dtype = PaddleDType::INT32;
} else {
LOG(FATAL) << "unsupported type.";
}

18
paddle/fluid/inference/api/demo_ci/utils.h
Unescape View File

@ -88,7 +88,7 @@ void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
}
break;
}
case PaddleDType::FLOAT32:
case PaddleDType::FLOAT32: {
for (size_t i = 0; i < numel; ++i) {
CHECK_LT(
fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
@ -96,6 +96,13 @@ void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
}
break;
}
case PaddleDType::INT32: {
for (size_t i = 0; i < numel; ++i) {
CHECK_EQ(static_cast<int32_t*>(output.data.data())[i], refer.data[i]);
}
break;
}
}
}
/*
@ -113,12 +120,19 @@ static std::string SummaryTensor(const PaddleTensor& tensor) {
}
break;
}
case PaddleDType::FLOAT32:
case PaddleDType::FLOAT32: {
for (int i = 0; i < std::min(num_elems, 10); i++) {
ss << static_cast<float*>(tensor.data.data())[i] << " ";
}
break;
}
case PaddleDType::INT32: {
for (int i = 0; i < std::min(num_elems, 10); i++) {
ss << static_cast<int32_t*>(tensor.data.data())[i] << " ";
}
break;
}
}
return ss.str();
}

60
paddle/fluid/inference/api/details/zero_copy_tensor.cc
Unescape View File

@ -15,6 +15,7 @@
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
@ -73,6 +74,61 @@ T *ZeroCopyTensor::data(PaddlePlace *place, int *size) const {
return res;
}
template <typename T>
void ZeroCopyTensor::copy_from_cpu(const T *data) {
EAGER_GET_TENSOR;
PADDLE_ENFORCE_GE(
tensor->numel(), 0,
"You should call ZeroCopyTensor::Reshape(const std::vector<int> &shape)"
"function before copy data from cpu.");
size_t ele_size = tensor->numel() * sizeof(T);
if (place_ == PaddlePlace::kCPU) {
auto *t_data = tensor->mutable_data<T>(platform::CPUPlace());
std::memcpy(static_cast<void *>(t_data), data, ele_size);
} else {
#ifdef PADDLE_WITH_CUDA
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
platform::CUDAPlace gpu_place(device_);
auto *t_data = tensor->mutable_data<T>(gpu_place);
auto *dev_ctx =
static_cast<const platform::CUDADeviceContext *>(pool.Get(gpu_place));
memory::Copy(gpu_place, static_cast<void *>(t_data), platform::CPUPlace(),
data, ele_size, dev_ctx->stream());
#else
PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif
}
}
template <typename T>
void ZeroCopyTensor::copy_to_cpu(T *data) {
EAGER_GET_TENSOR;
auto ele_num = tensor->numel();
auto *t_data = tensor->data<T>();
auto t_place = tensor->place();
if (platform::is_cpu_place(t_place)) {
std::memcpy(static_cast<void *>(data), t_data, ele_num * sizeof(T));
} else {
#ifdef PADDLE_WITH_CUDA
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto gpu_place = boost::get<platform::CUDAPlace>(t_place);
auto *dev_ctx =
static_cast<const platform::CUDADeviceContext *>(pool.Get(gpu_place));
memory::Copy(platform::CPUPlace(), static_cast<void *>(data), gpu_place,
t_data, ele_num * sizeof(T), dev_ctx->stream());
#else
PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif
}
}
template void ZeroCopyTensor::copy_from_cpu<float>(const float *data);
template void ZeroCopyTensor::copy_from_cpu<int64_t>(const int64_t *data);
template void ZeroCopyTensor::copy_to_cpu<float>(float *data);
template void ZeroCopyTensor::copy_to_cpu<int64_t>(int64_t *data);
template float *ZeroCopyTensor::data<float>(PaddlePlace *place,
int *size) const;
template int64_t *ZeroCopyTensor::data<int64_t>(PaddlePlace *place,
@ -92,10 +148,10 @@ void *ZeroCopyTensor::FindTensor() const {
return tensor;
}
std::vector<int64_t> ZeroCopyTensor::shape() const {
std::vector<int> ZeroCopyTensor::shape() const {
EAGER_GET_TENSOR;
PADDLE_ENFORCE(tensor_, "not found tensor called %s in the scope", name_);
return framework::vectorize(tensor->dims());
return framework::vectorize2int(tensor->dims());
}
void ZeroCopyTensor::SetLoD(const std::vector<std::vector<size_t>> &x) {

2
paddle/fluid/inference/api/details/zero_copy_tensor_dummy.cc
Unescape View File

@ -37,7 +37,7 @@ template int64_t *ZeroCopyTensor::mutable_data(PaddlePlace place);
void *ZeroCopyTensor::FindTensor() const { return nullptr; }
std::vector<int64_t> ZeroCopyTensor::shape() const { return {}; }
std::vector<int> ZeroCopyTensor::shape() const { return {}; }
void ZeroCopyTensor::SetLoD(const std::vector<std::vector<size_t>> &x) {}

15
paddle/fluid/inference/api/helper.h
Unescape View File

@ -50,6 +50,11 @@ class Timer {
}
};
static int GetUniqueId() {
static int id = 0;
return id++;
}
static void split(const std::string &str, char sep,
std::vector<std::string> *pieces) {
pieces->clear();
@ -81,6 +86,13 @@ static void split_to_int64(const std::string &str, char sep,
std::transform(pieces.begin(), pieces.end(), std::back_inserter(*is),
[](const std::string &v) { return std::stoi(v); });
}
static void split_to_int(const std::string &str, char sep,
std::vector<int> *is) {
std::vector<std::string> pieces;
split(str, sep, &pieces);
std::transform(pieces.begin(), pieces.end(), std::back_inserter(*is),
[](const std::string &v) { return std::stoi(v); });
}
template <typename T>
std::string to_string(const std::vector<T> &vec) {
std::stringstream ss;
@ -197,6 +209,9 @@ static std::string DescribeTensor(const PaddleTensor &tensor,
case PaddleDType::INT64:
os << "int64";
break;
case PaddleDType::INT32:
os << "int32";
break;
default:
os << "unset";
}

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

@ -135,7 +135,8 @@ struct AnalysisConfig {
*/
void EnableTensorRtEngine(int workspace_size = 1 << 20,
int max_batch_size = 1, int min_subgraph_size = 3,
Precision precision = Precision::kFloat32);
Precision precision = Precision::kFloat32,
bool use_static = true);
/** A boolean state telling whether the TensorRT engine is used.
*/
bool tensorrt_engine_enabled() const { return use_tensorrt_; }
@ -233,6 +234,7 @@ struct AnalysisConfig {
// subgraph, 3 as default value.
int tensorrt_min_subgraph_size_{3};
Precision tensorrt_precision_mode_;
bool trt_use_static_engine_;
// memory reuse related.
bool enable_memory_optim_{false};

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

Write Preview
Loading…
Cancel
Save