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

fix_gru_py
yuyang18 7 years ago
commit 7c777dd549

@ -25,7 +25,6 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
"${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
find_package(Sphinx)
if(NOT CMAKE_CROSSCOMPILING)
find_package(CUDA QUIET)
endif(NOT CMAKE_CROSSCOMPILING)
@ -226,5 +225,7 @@ if(WITH_PYTHON)
endif()
if(WITH_DOC)
find_package(Sphinx REQUIRED)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()

@ -19,8 +19,9 @@
----------------
PaddlePaddle需要使用Docker环境完成编译这样可以免去单独安装编译依赖的步骤可选的不同编译环境Docker镜像
可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到。或者
参考下述可选步骤从源码中构建用于编译PaddlePaddle的Docker镜像。
可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到,您也可以
`这里 <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`_ 找到 paddle_manylinux_devel
镜像的编译以及使用方法。或者参考下述可选步骤从源码中构建用于编译PaddlePaddle的Docker镜像。
如果您选择不使用Docker镜像则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。

@ -22,6 +22,8 @@ How To Build
You need to use Docker to build PaddlePaddle
to avoid installing dependencies by yourself. We have several pre-built
Docker images `here <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ ,
you can also find how to build and use paddle_manylinux_devel Docker image from
`here <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`_
Or you can build your own image from source as the optional step below:
.. code-block:: bash

@ -192,6 +192,10 @@ class ExecutionContext {
return op_.Attr<T>(name);
}
bool HasInput(const std::string& name) const { return op_.HasInputs(name); }
bool HasOutput(const std::string& name) const { return op_.HasOutputs(name); }
size_t InputSize(const std::string& name) const {
return op_.Inputs(name).size();
}

@ -57,7 +57,8 @@ ParallelExecutor::ParallelExecutor(
const std::unordered_set<std::string> &bcast_vars,
const ProgramDesc &main_program, const std::string &loss_var_name,
Scope *scope, const std::vector<Scope *> &local_scopes,
const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy)
const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
size_t num_trainers, size_t trainer_id)
: member_(new ParallelExecutorPrivate(places)) {
member_->global_scope_ = scope;
@ -79,7 +80,13 @@ ParallelExecutor::ParallelExecutor(
// Bcast Parameters to all GPUs
#ifdef PADDLE_WITH_CUDA
member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_));
auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
ncclUniqueId *nccl_id = nullptr;
if (nccl_id_var != nullptr) {
nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
}
member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
member_->places_, nccl_id, num_trainers, trainer_id));
#endif
if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 &&
local_scopes.empty()) { // Is CUDA

@ -44,7 +44,8 @@ class ParallelExecutor {
const std::string &loss_var_name, Scope *scope,
const std::vector<Scope *> &local_scopes,
const ExecutionStrategy &exec_strategy,
const BuildStrategy &build_strategy);
const BuildStrategy &build_strategy,
size_t num_trainers = 1, size_t trainer_id = 0);
~ParallelExecutor();

@ -1 +1,2 @@
cc_library(dot SRCS dot.cc)
cc_library(analysis SRCS dot.cc node.cc node.h)
cc_test(test_node SRCS node_tester.cc DEPS analysis)

@ -0,0 +1,24 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
namespace paddle {
namespace inference {
namespace analysis {
enum class Device { CPU, GPU };
} // namespace analysis
} // namespace inference
} // namespace paddle

@ -21,6 +21,7 @@
#include <glog/logging.h>
#include <sstream>
#include <string>
#include <unordered_map>
#include <vector>

@ -0,0 +1,62 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/inference/analysis/dot.h"
#include <gtest/gtest.h>
#include <memory>
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
namespace paddle {
namespace inference {
namespace analysis {
class DotTester : public ::testing::Test {
protected:
void SetUp() override {
std::vector<Dot::Attr> attrs({{"title", "hello"}});
dot.reset(new Dot(attrs));
dot->AddNode("a", {Dot::Attr{"shape", "box"}, Dot::Attr("color", "blue")});
dot->AddNode("b", {});
dot->AddNode("c", {});
dot->AddEdge("a", "b", {});
dot->AddEdge("b", "c", {});
dot->AddEdge("a", "c", {});
}
std::unique_ptr<Dot> dot;
};
TEST_F(DotTester, Build) {
auto codes = dot->Build();
// Output the DOT language code, the generated codes are too long to compare
// the string.
//
// The output is
//
// digraph G {
// title="hello"
// node_1
// node_2
// node_0[label="a" shape="box" color="blue"]
// node_0->node_1
// node_1->node_2
// node_0->node_2
// } // end G
LOG(INFO) << '\n' << codes;
}
} // namespace analysis
} // namespace inference
} // namespace paddle

@ -0,0 +1,74 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace inference {
namespace analysis {
template <typename IteratorT>
class iterator_range {
IteratorT begin_, end_;
public:
template <typename Container>
explicit iterator_range(Container &&c) : begin_(c.begin()), end_(c.end()) {}
iterator_range(const IteratorT &begin, const IteratorT &end)
: begin_(begin), end_(end) {}
const IteratorT &begin() const { return begin_; }
const IteratorT &end() const { return end_; }
};
/*
* An registry helper class, with its records keeps the order they registers.
*/
template <typename T>
class OrderedRegistry {
public:
T *Register(const std::string &name, T *x) {
PADDLE_ENFORCE(!dic_.count(name));
dic_[name] = data_.size();
data_.emplace_back(std::unique_ptr<T>(x));
return data_.back().get();
}
T *Lookup(const std::string &name) {
auto it = dic_.find(name);
if (it == dic_.end()) return nullptr;
return data_[it->second].get();
}
protected:
std::unordered_map<std::string, int> dic_;
std::vector<std::unique_ptr<T>> data_;
};
} // namespace analysis
} // namespace inference
} // namespace paddle
#define PADDLE_DISALLOW_COPY_AND_ASSIGN(type__) \
\
type__(const type__ &) = delete; \
\
void operator=(const type__ &) = delete;

@ -0,0 +1,67 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/inference/analysis/node.h"
#include "glog/logging.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace inference {
namespace analysis {
std::vector<Dot::Attr> Value::dot_attrs() const {
return std::vector<Dot::Attr>({Dot::Attr("style", "filled,rounded"),
Dot::Attr("shape", "box"),
Dot::Attr("fillcolor", "red")});
}
std::vector<Dot::Attr> Function::dot_attrs() const {
return std::vector<Dot::Attr>({Dot::Attr("style", "filled,rounded"),
Dot::Attr("shape", "diamond"),
Dot::Attr("fillcolor", "yellow")});
}
Node *NodeMap::Create(Node::Type type) {
switch (type) {
case Node::Type::kFunction:
nodes_.emplace_back(new Function);
break;
case Node::Type::kValue:
nodes_.emplace_back(new Value);
break;
default:
PADDLE_THROW("Not supported node type.");
}
nodes_.back()->id_ = size() - 1;
return nodes_.back().get();
}
Node *NodeMap::GetMutable(size_t id) {
PADDLE_ENFORCE_GT(size(), id);
return nodes_[id].get();
}
const Node &NodeMap::Get(size_t id) const {
PADDLE_ENFORCE_GT(size(), id);
return *nodes_[id].get();
}
void NodeMap::Delete(size_t id) {
PADDLE_ENFORCE_LT(id, size());
nodes_[id]->SetDeleted();
}
} // namespace analysis
} // namespace inference
} // namespace paddle

@ -0,0 +1,235 @@
/* 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. */
/*
* This file defines the Node class and its subclasses. A Node is the basis
* analysis element in a computation graph.
* There are basically two kinds of nodes, the function node and value node.
*/
#pragma once
#include <limits>
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/inference/analysis/device.h"
#include "paddle/fluid/inference/analysis/dot.h"
#include "paddle/fluid/inference/analysis/helper.h"
namespace paddle {
namespace inference {
namespace analysis {
class NodeMap;
/*
* Node Representation.
*
* This is a very important class for analysis. It is the base class of all
* nodes computed by a program that may be used as operands to other nodes.
* Node is the super class of other important classes such as Function and
* Value, some nodes can have a name.
*/
class Node {
public:
// Node type. NOTE the new node types should add here.
enum class Type { kNone = -1, kFunction, kValue, kFunctionBlock };
Node() = default;
struct Attr;
// Cast to a subclass type, Function for example.
template <typename Subclass>
Subclass &As() {
return *dynamic_cast<Subclass *>(this);
}
// Formatted representation of this Node.
virtual std::string repr() const {
return name() + "(" + std::to_string(id()) + ")";
}
// DOT node representation. One Node type can customize its own node
// representation.
virtual std::vector<Dot::Attr> dot_attrs() const {
return std::vector<Dot::Attr>({Dot::Attr("style", "filled")});
}
// Get an additional attribute and convert it to T data type. NOTE this will
// silently create a new attribute if not exists.
Attr &attr(const std::string &name) { return attrs_[name]; }
int id() const { return id_; }
bool deleted() const { return deleted_; }
void SetDeleted() { deleted_ = true; }
void SetName(const std::string &name) { name_ = name; }
const std::string &name() const { return name_; }
void SetType(Type type) { type_ = type; }
Type type() const { return type_; }
void *extra_info() const { return extra_info_; }
void SetExtraInfo(void *extra_info) { extra_info_ = extra_info; }
// Input links.
std::vector<Node *> inlinks;
// Output links.
std::vector<Node *> outlinks;
// A helper class to maintain the status from Pass.
// TODO(superjomn) add a checker here to ensure the T is primary.
struct Attr {
// NOTE T should be a primary type or a struct combined by several primary
// types.
// NOTE the STL containers should not use here.
// Some usages
// Attr attr;
// T data;
// attr.data.assign((char*)data, sizeof(data));
bool &Bool() { return As<bool>(); }
float &Float() { return As<float>(); }
int32_t &Int32() { return As<int32_t>(); }
int64_t &Int64() { return As<int64_t>(); }
private:
template <typename T>
T &As() {
// init storage in the first usage.
if (data_.empty()) {
VLOG(4) << "resize data to " << sizeof(T);
type_hash_ = typeid(T).hash_code();
data_.resize(sizeof(T));
}
PADDLE_ENFORCE(type_hash_ == typeid(T).hash_code(), "type not matched");
PADDLE_ENFORCE_EQ(data_.size(), sizeof(T), "Node attr type recast error");
return *reinterpret_cast<T *>(&data_[0]);
}
private:
std::string data_;
size_t type_hash_{std::numeric_limits<size_t>::max()};
};
virtual ~Node() {}
friend class NodeMap;
PADDLE_DISALLOW_COPY_AND_ASSIGN(Node);
protected:
// The id number not the name is a node's unique identifier in the computation
// graph.
int id_{-1};
std::string name_;
Type type_{Type::kNone};
// Mark this node is deleted by some pass.
bool deleted_{false};
void *extra_info_;
mutable std::unordered_map<std::string, Attr> attrs_;
};
class Function;
/*
* Value represents a value node, it has some attributes including dims, data
* type and so on.
*/
class Value : public Node {
public:
enum class DataType { kInt32, kInt64, kFloat32, kFloat64 };
using Dims = std::vector<int>;
void SetDataType(DataType data_type) { data_type_ = data_type; }
DataType data_type() const { return data_type_; }
void SetDims(const Dims &dims) { dims_ = dims; }
const Dims &dims() const { return dims_; }
Device device() const { return device_; }
void SetDevice(Device device) { device_ = device; }
std::vector<Dot::Attr> dot_attrs() const override;
PADDLE_DISALLOW_COPY_AND_ASSIGN(Value);
protected:
Value() { SetType(Node::Type::kValue); }
friend class NodeMap;
private:
DataType data_type_;
Dims dims_;
Device device_;
};
/*
* Function represents any kind of executable concepts that takes several Values
* as input, and outputs several Values.
*/
class Function : public Node {
public:
std::vector<Dot::Attr> dot_attrs() const override;
// Get the operator's type from Desc.
const std::string &func_type() const { return func_type_; }
// Set the operator's type.
void SetFuncType(const std::string &func_type) { func_type_ = func_type; }
PADDLE_DISALLOW_COPY_AND_ASSIGN(Function);
protected:
std::string func_type_;
Function() { SetType(Node::Type::kFunction); }
friend class NodeMap;
};
/*
* FunctionBlock is a Node that contains a sub-graph multiple Node.
*/
struct FunctionBlock : public Node {
std::string repr() const override { return "block-" + std::to_string(id()); }
std::vector<Node *> subgraph;
};
class NodeMap {
public:
// Create a new node with type.
Node *Create(Node::Type type);
// Get a node by its id.
Node *GetMutable(size_t id);
const Node &Get(size_t id) const;
void Delete(size_t id);
const std::vector<std::unique_ptr<Node>> &nodes() { return nodes_; }
size_t size() const { return nodes_.size(); }
private:
std::vector<std::unique_ptr<Node>> nodes_;
std::unordered_map<std::string, Node *> map_;
};
} // namespace analysis
} // namespace inference
} // namespace paddle

@ -0,0 +1,34 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/inference/analysis/node.h"
#include <gtest/gtest.h>
namespace paddle {
namespace inference {
namespace analysis {
TEST(Node, Attr) {
// Node is an abstract class, use Value instead for they share the same Attr
// logic.
NodeMap nodes;
auto* node = nodes.Create(Node::Type::kValue);
node->attr("v0").Int32() = 2008;
ASSERT_EQ(node->attr("v0").Int32(), 2008);
}
} // namespace analysis
} // namespace inference
} // namespace paddle

@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/framework.pb.h"
namespace paddle {
@ -58,8 +59,8 @@ class EngineBase {
struct Buffer {
void* buffer{nullptr}; // buffer should be allocated only once.
int max_size; // buffer allocated space.
int size; // data size.
size_t max_size; // buffer allocated space.
size_t size; // data size.
DeviceType device{DeviceType::UNK}; // tells which device this buffer is on.
};

@ -1,5 +1,4 @@
nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto)
nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
add_subdirectory(convert)

@ -1,4 +1,4 @@
nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine)
nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)

@ -21,15 +21,18 @@ namespace tensorrt {
class ReluOpConverter : public OpConverter {
public:
ReluOpConverter() {}
void operator()(const framework::OpDesc& op) override {
void operator()(const framework::proto::OpDesc& op) override {
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework::OpDesc op_desc(op, nullptr, nullptr);
LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose "
"type is Relu";
const nvinfer1::ITensor* input_tensor =
engine_->GetITensor(op.Input("X")[0]);
engine_->GetITensor(op_desc.Input("X")[0]);
nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER(
engine_, Activation, *const_cast<nvinfer1::ITensor*>(input_tensor),
nvinfer1::ActivationType::kRELU);
engine_->SetITensor(op.Output("Out")[0], layer->getOutput(0));
engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0));
}
};

@ -21,7 +21,7 @@ namespace tensorrt {
class Conv2dOpConverter : public OpConverter {
public:
Conv2dOpConverter() {}
void operator()(const framework::OpDesc& op) override {
void operator()(const framework::proto::OpDesc& op) override {
LOG(INFO)
<< "convert a fluid conv2d op to tensorrt conv layer without bias";
}

@ -23,26 +23,42 @@ namespace tensorrt {
using platform::is_gpu_place;
using platform::is_cpu_place;
class DefaultInputConverter : public EngineInputConverter {
class DefaultIOConverter : public EngineIOConverter {
public:
DefaultInputConverter() {}
DefaultIOConverter() {}
// NOTE out is GPU memory.
virtual void operator()(const LoDTensor& in, void* out,
size_t max_size) override {
PADDLE_ENFORCE(out != nullptr);
PADDLE_ENFORCE_LE(in.memory_size(), max_size);
PADDLE_ENFORCE(stream_ != nullptr);
const auto& place = in.place();
size_t size = in.memory_size();
PADDLE_ENFORCE_LE(size, max_size);
if (is_cpu_place(place)) {
PADDLE_ENFORCE(stream_ != nullptr);
PADDLE_ENFORCE_EQ(0,
cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
cudaMemcpyHostToDevice, *stream_));
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
cudaMemcpyHostToDevice, *stream_));
} else if (is_gpu_place(place)) {
PADDLE_ENFORCE_EQ(0,
cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
cudaMemcpyHostToHost, *stream_));
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
cudaMemcpyDeviceToDevice, *stream_));
} else {
PADDLE_THROW("Unknown device for converter");
}
cudaStreamSynchronize(*stream_);
}
// NOTE in is GPU memory.
virtual void operator()(const void* in, LoDTensor* out,
size_t max_size) override {
PADDLE_ENFORCE(in != nullptr);
PADDLE_ENFORCE(stream_ != nullptr);
const auto& place = out->place();
size_t size = out->memory_size();
PADDLE_ENFORCE_LE(size, max_size);
if (is_cpu_place(place)) {
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
cudaMemcpyDeviceToHost, *stream_));
} else if (is_gpu_place(place)) {
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
cudaMemcpyDeviceToDevice, *stream_));
} else {
PADDLE_THROW("Unknown device for converter");
}
@ -50,7 +66,8 @@ class DefaultInputConverter : public EngineInputConverter {
}
};
REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter);
// fluid LodTensor <-> tensorrt ITensor
REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter);
} // namespace tensorrt
} // namespace inference

@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <string>
#include <unordered_map>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/utils/singleton.h"
@ -25,43 +26,57 @@ namespace tensorrt {
using framework::LoDTensor;
/*
* Convert Input from Fluid to an Engine.
* TensorRT's ITensor follows row major, NCHW. Fluid is also row major, so in
* most cases just need to copy the data.
* Convert Input from Fluid to TensorRT Engine.
* Convert Output from TensorRT Engine to Fluid.
*
* Note that TensorRT's ITensor follows row major, NCHW. Fluid is also row
* major,
* so in the default case just need to copy the data.
*/
class EngineInputConverter {
class EngineIOConverter {
public:
EngineInputConverter() {}
EngineIOConverter() {}
virtual void operator()(const LoDTensor& in, void* out, size_t max_size) {}
virtual void operator()(const void* in, LoDTensor* out, size_t max_size) {}
void SetStream(cudaStream_t* stream) { stream_ = stream; }
static void Run(const std::string& in_op_type, const LoDTensor& in, void* out,
size_t max_size, cudaStream_t* stream) {
static void ConvertInput(const std::string& op_type, const LoDTensor& in,
void* out, size_t max_size, cudaStream_t* stream) {
PADDLE_ENFORCE(stream != nullptr);
auto* converter = Registry<EngineInputConverter>::Lookup(
in_op_type, "default" /* default_type */);
auto* converter = Registry<EngineIOConverter>::Lookup(
op_type, "default" /* default_type */);
PADDLE_ENFORCE_NOT_NULL(converter);
converter->SetStream(stream);
(*converter)(in, out, max_size);
}
virtual ~EngineInputConverter() {}
static void ConvertOutput(const std::string& op_type, const void* in,
LoDTensor* out, size_t max_size,
cudaStream_t* stream) {
PADDLE_ENFORCE(stream != nullptr);
auto* converter = Registry<EngineIOConverter>::Lookup(
op_type, "default" /* default_type */);
PADDLE_ENFORCE_NOT_NULL(converter);
converter->SetStream(stream);
(*converter)(in, out, max_size);
}
virtual ~EngineIOConverter() {}
protected:
cudaStream_t* stream_{nullptr};
};
#define REGISTER_TENSORRT_IO_CONVERTER(op_type__, Converter__) \
struct trt_io_##op_type__##_converter { \
trt_io_##op_type__##_converter() { \
Registry<EngineIOConverter>::Register<Converter__>(#op_type__); \
} \
}; \
trt_io_##op_type__##_converter trt_io_##op_type__##_converter__;
} // namespace tensorrt
} // namespace inference
} // namespace paddle
#define REGISTER_TENSORRT_INPUT_CONVERTER(in_op_type__, Converter__) \
struct trt_input_##in_op_type__##_converter { \
trt_input_##in_op_type__##_converter() { \
::paddle::inference::Registry<EngineInputConverter>::Register< \
Converter__>(#in_op_type__); \
} \
}; \
trt_input_##in_op_type__##_converter trt_input_##in_op_type__##_converter__;

@ -21,7 +21,7 @@ namespace tensorrt {
class MulOpConverter : public OpConverter {
public:
MulOpConverter() {}
void operator()(const framework::OpDesc& op) override {
void operator()(const framework::proto::OpDesc& op) override {
LOG(INFO) << "convert a fluid mul op to tensorrt fc layer without bias";
}
};

@ -31,10 +31,10 @@ namespace tensorrt {
class OpConverter {
public:
OpConverter() {}
virtual void operator()(const framework::OpDesc& op) {}
virtual void operator()(const framework::proto::OpDesc& op) {}
void Run(const framework::OpDesc& op, TensorRTEngine* engine) {
std::string type = op.Type();
void Run(const framework::proto::OpDesc& op, TensorRTEngine* engine) {
std::string type = op.type();
auto* it = Registry<OpConverter>::Lookup(type);
PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type);
it->SetEngine(engine);
@ -42,14 +42,16 @@ class OpConverter {
}
// convert fluid op to tensorrt layer
void ConvertOp(const framework::OpDesc& op, TensorRTEngine* engine) {
void ConvertOp(const framework::proto::OpDesc& op, TensorRTEngine* engine) {
OpConverter::Run(op, engine);
}
// convert fluid block to tensorrt network
void ConvertBlock(const framework::BlockDesc& block, TensorRTEngine* engine) {
for (auto op : block.AllOps()) {
OpConverter::Run(*op, engine);
void ConvertBlock(const framework::proto::BlockDesc& block,
TensorRTEngine* engine) {
for (size_t i = 0; i < block.ops_size(); i++) {
const auto& op = block.ops(i);
OpConverter::Run(op, engine);
}
}

@ -16,6 +16,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/place.h"
@ -26,7 +27,7 @@ namespace paddle {
namespace inference {
namespace tensorrt {
void Compare(float input, float expect) {
void Compare(const std::string op_type, float input, float expect) {
framework::Scope scope;
platform::CUDAPlace place;
platform::CUDADeviceContext ctx(place);
@ -35,6 +36,7 @@ void Compare(float input, float expect) {
auto x_var = scope.Var("X");
auto x_tensor = x_var->GetMutable<framework::LoDTensor>();
x_tensor->Resize({1, 1});
x_tensor->mutable_data<float>(place);
std::vector<float> init;
init.push_back(input);
framework::TensorFromVector(init, ctx, x_tensor);
@ -45,14 +47,15 @@ void Compare(float input, float expect) {
out_tensor->mutable_data<float>(place);
framework::OpDesc op_desc;
op_desc.SetType("relu");
op_desc.SetType(op_type);
op_desc.SetInput("X", {"X"});
op_desc.SetOutput("Out", {"Out"});
auto relu_op = framework::OpRegistry::CreateOp(op_desc);
auto op = framework::OpRegistry::CreateOp(*op_desc.Proto());
// run fluid op
relu_op->Run(scope, place);
op->Run(scope, place);
// get fluid output
std::vector<float> out1;
framework::TensorToVector(*out_tensor, ctx, &out1);
@ -63,21 +66,28 @@ void Compare(float input, float expect) {
engine->InitNetwork();
engine->DeclareInput("X", nvinfer1::DataType::kFLOAT,
nvinfer1::DimsCHW{1, 1, 1});
// convert op
OpConverter op_converter;
op_converter.ConvertOp(op_desc, engine);
op_converter.ConvertOp(*op_desc.Proto(), engine);
engine->DeclareOutput("Out");
engine->FreezeNetwork();
engine->SetInputFromCPU("X", &input, 1 * sizeof(float));
// run tensorrt op
// convert LoDTensor to ITensor
size_t size = x_tensor->memory_size();
EngineIOConverter::ConvertInput(op_type, *x_tensor,
engine->buffer("X").buffer, size, &stream);
// run tensorrt Outp
engine->Execute(1);
float out2;
engine->GetOutputInCPU("Out", &out2, 1 * sizeof(float));
ASSERT_EQ(out1[0], out2);
// convert ITensor to LoDTensor
EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer,
out_tensor, size, &stream);
// get tensorrt output
std::vector<float> out2;
framework::TensorToVector(*out_tensor, ctx, &out2);
// compare
ASSERT_EQ(out1[0], out2[0]);
ASSERT_EQ(out1[0], expect);
delete engine;
@ -85,8 +95,8 @@ void Compare(float input, float expect) {
}
TEST(OpConverter, ConvertRelu) {
Compare(1, 1); // relu(1) = 1
Compare(-5, 0); // relu(-5) = 0
Compare("relu", 1, 1); // relu(1) = 1
Compare("relu", -5, 0); // relu(-5) = 0
}
} // namespace tensorrt

@ -12,40 +12,63 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <gtest/gtest.h>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
#include <gtest/gtest.h>
namespace paddle {
namespace inference {
namespace tensorrt {
class EngineInputConverterTester : public ::testing::Test {
public:
void SetUp() override { tensor.Resize({10, 10}); }
void IOConverterTester(const platform::DeviceContext& ctx) {
cudaStream_t stream;
ASSERT_EQ(0, cudaStreamCreate(&stream));
framework::LoDTensor tensor;
};
// init fluid in_tensor
framework::LoDTensor in_tensor;
in_tensor.Resize({10, 10});
auto place = ctx.GetPlace();
in_tensor.mutable_data<float>(place);
std::vector<float> init;
for (int64_t i = 0; i < 10 * 10; ++i) {
init.push_back(i);
}
framework::TensorFromVector(init, ctx, &in_tensor);
TEST_F(EngineInputConverterTester, DefaultCPU) {
// init tensorrt buffer
void* buffer;
tensor.mutable_data<float>(platform::CPUPlace());
ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
size_t size = in_tensor.memory_size();
ASSERT_EQ(cudaMalloc(&buffer, size), 0);
cudaStream_t stream;
EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
&stream);
// convert fluid in_tensor to tensorrt buffer
EngineIOConverter::ConvertInput("test", in_tensor, buffer, size, &stream);
// convert tensorrt buffer to fluid out_tensor
framework::LoDTensor out_tensor;
out_tensor.Resize({10, 10});
out_tensor.mutable_data<float>(place);
EngineIOConverter::ConvertOutput("test", buffer, &out_tensor, size, &stream);
// compare in_tensor and out_tensor
std::vector<float> result;
framework::TensorToVector(out_tensor, ctx, &result);
EXPECT_EQ(init.size(), result.size());
for (size_t i = 0; i < init.size(); i++) {
EXPECT_EQ(init[i], result[i]);
}
cudaStreamDestroy(stream);
}
TEST_F(EngineInputConverterTester, DefaultGPU) {
void* buffer;
tensor.mutable_data<float>(platform::CUDAPlace());
ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
TEST(EngineIOConverterTester, DefaultCPU) {
platform::CPUPlace place;
platform::CPUDeviceContext ctx(place);
IOConverterTester(ctx);
}
cudaStream_t stream;
EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
&stream);
TEST(EngineIOConverterTester, DefaultGPU) {
platform::CUDAPlace place;
platform::CUDADeviceContext ctx(place);
IOConverterTester(ctx);
}
} // namespace tensorrt

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

Loading…
Cancel
Save