Merge branch 'develop' of github.com:baidu/Paddle into feature/fast_ci

emailweixu-patch-1
Yang Yu 7 years ago
commit 8d6f827a0d

@ -15,9 +15,9 @@
include(ExternalProject)
set(BOOST_PROJECT "extern_boost")
set(BOOST_VER "1.66.0")
set(BOOST_TAR "boost_1_66_0")
set(BOOST_URL "https://dl.bintray.com/boostorg/release/${BOOST_VER}/source/${BOOST_TAR}.tar.gz")
set(BOOST_VER "1.41.0")
set(BOOST_TAR "boost_1_41_0")
set(BOOST_URL "http://sourceforge.net/projects/boost/files/boost/${BOOST_VER}/${BOOST_TAR}.tar.gz")
set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)

@ -534,3 +534,8 @@ row_conv
--------
.. autofunction:: paddle.v2.fluid.layers.row_conv
:noindex:
multiplex
---------
.. autofunction:: paddle.v2.fluid.layers.multiplex
:noindex:

@ -2,9 +2,9 @@
## Background
Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries flexibly and efficiently.
Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries in a flexible and efficient manner.
On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example,Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example, Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
On the other hand, users usually do not want to care about the low-level hardware and computing libraries when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are hardware independent.
@ -17,7 +17,7 @@ For a general overview of fluid, please refer to the [overview doc](https://gith
There are mainly three parts that we have to consider while integrating a new device/library:
- Place and DeviceContext: indicates the device id and manages hardware resources
- Place and DeviceContext: indicate the device id and manage hardware resources
- Memory and Tensor: malloc/free data on certain device
@ -25,10 +25,10 @@ There are mainly three parts that we have to consider while integrating a new de
### Place and DeviceContext
Please remind that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
Please note that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
#### Place
Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add corresponding `DevicePlace`.
Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add the corresponding `DevicePlace`.
```
| CPUPlace
@ -144,7 +144,7 @@ class Tensor {
};
```
`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory.
`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configurate its shape, and then call `mutuable_data` to allocate the actual memory.
```cpp
paddle::framework::Tensor t;
@ -163,7 +163,7 @@ Fluid implements computing units based on different DeviceContexts. Some computi
Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example:
The interface is defined in header file.
The interface is defined in the header file.
```
template <typename DeviceContext, typename T>
@ -203,7 +203,7 @@ class MaxOutFunctor<platform::CUDADeviceContext, T> {
```
We get computing handle from a concrete DeviceContext, and make compution on tensors.
We first obtain the computing handle from a concrete DeviceContext, and then compute on tensors.
The implemention of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map.
@ -231,7 +231,7 @@ REGISTER_OP_CUDA_KERNEL(
## Advanced topics: How to switch between different Device/Library
Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
Generally, we will implement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not suitable on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run on GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
For more details, please refer to following docs:

@ -39,6 +39,7 @@ PaddlePaddle可以使用常用的Python包管理工具
"cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
"cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"

@ -42,6 +42,7 @@ If the links below shows up the login form, just click "Log in as guest" to star
"cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
"cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"

@ -74,8 +74,10 @@ cc_library(backward SRCS backward.cc DEPS net_op)
cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context fill_constant_op)
cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope
framework_proto backward glog lod_rank_table profiler)
framework_proto backward glog lod_rank_table profiler feed_fetch_method)
cc_library(prune SRCS prune.cc DEPS framework_proto)
cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

@ -0,0 +1,87 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 <condition_variable>
#include <mutex>
#include <queue>
namespace paddle {
namespace framework {
template <typename T>
class Channel {
public:
explicit Channel(std::size_t capacity) : capacity_(capacity) {}
void Send(T* channel_element) {
std::unique_lock<std::mutex> lock(mu_);
if (IsBounded()) {
full_cond_var_.wait(lock, [this]() {
bool capacity_valid = capacity_ > 0 ? !IsCapacityFull() : true;
return capacity_valid;
});
}
channel_.push_back(std::move(*channel_element));
lock.unlock();
empty_cond_var_.notify_one();
}
T* Receive() {
std::unique_lock<std::mutex> lock(mu_);
empty_cond_var_.wait(lock, [this]() { return !channel_.empty(); });
T* channel_element = std::move(channel_.front());
channel_.pop_front();
NotifyAllSenders(&lock);
return channel_element;
}
size_t Size() {
std::unique_lock<std::mutex> lock(mu_);
return channel_.size();
}
void Clear() {
std::unique_lock<std::mutex> lock(mu_);
channel_.clear();
NotifyAllSenders(&lock);
}
private:
std::size_t capacity_;
std::mutex mu_;
std::condition_variable empty_cond_var_;
std::condition_variable full_cond_var_;
std::deque<T> channel_;
private:
void NotifyAllSenders(std::unique_lock<std::mutex>* lock) {
if (IsBounded()) {
lock->unlock();
full_cond_var_.notify_one();
}
}
bool IsBounded() const { return capacity_ > 0; }
bool IsCapacityFull() const { return channel_.size() >= capacity_; }
};
} // namespace operator
} // namespace paddle

@ -79,5 +79,33 @@ inline void VisitDataType(proto::DataType type, Visitor visitor) {
}
}
inline std::string DataTypeToString(const proto::DataType type) {
using namespace paddle::framework::proto;
switch (type) {
case DataType::FP16:
return "float16";
case DataType::FP32:
return "float32";
case DataType::FP64:
return "float64";
case DataType::INT16:
return "int16";
case DataType::INT32:
return "int32";
case DataType::INT64:
return "int64";
case DataType::BOOL:
return "bool";
default:
PADDLE_THROW("Not support type %d", type);
}
}
inline std::ostream& operator<<(std::ostream& out,
const proto::DataType& type) {
out << DataTypeToString(type);
return out;
}
} // namespace framework
} // namespace paddle

@ -17,6 +17,7 @@ limitations under the License. */
#include <set>
#include "gflags/gflags.h"
#include "paddle/framework/feed_fetch_method.h"
#include "paddle/framework/feed_fetch_type.h"
#include "paddle/framework/lod_rank_table.h"
#include "paddle/framework/lod_tensor_array.h"
@ -149,5 +150,164 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
}
}
// Check whether the block already has feed operators and feed_holder.
// Return false if the block does not have any feed operators.
// If some feed operators have been prepended to the block, check that
// the info contained in these feed operators matches the feed_targets
// and feed_holder_name. Raise exception when any mismatch is found.
// Return true if the block has feed operators and holder of matching info.
static bool has_feed_operators(
BlockDesc* block, std::map<std::string, const LoDTensor*>& feed_targets,
const std::string& feed_holder_name) {
size_t feed_count = 0;
for (auto* op : block->AllOps()) {
if (op->Type() == kFeedOpType) {
feed_count++;
PADDLE_ENFORCE_EQ(op->Input("X")[0], feed_holder_name,
"Input to feed op should be '%s'", feed_holder_name);
std::string feed_target_name = op->Output("Out")[0];
PADDLE_ENFORCE(
feed_targets.find(feed_target_name) != feed_targets.end(),
"Feed operator output name '%s' cannot be found in 'feed_targets'",
feed_target_name);
}
}
if (feed_count > 0) {
PADDLE_ENFORCE_EQ(
feed_count, feed_targets.size(),
"The number of feed operators should match 'feed_targets'");
// When feed operator are present, so should be feed_holder
auto var = block->FindVar(feed_holder_name);
PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable",
feed_holder_name);
PADDLE_ENFORCE_EQ(var->GetType(), proto::VarDesc::FEED_MINIBATCH,
"'%s' variable should be 'FEED_MINIBATCH' type",
feed_holder_name);
}
return feed_count > 0;
}
// Check whether the block already has fetch operators and fetch_holder.
// Return false if the block does not have any fetch operators.
// If some fetch operators have been appended to the block, check that
// the info contained in these fetch operators matches the fetch_targets
// and fetch_holder_name. Raise exception when any mismatch is found.
// Return true if the block has fetch operators and holder of matching info.
static bool has_fetch_operators(
BlockDesc* block, std::map<std::string, LoDTensor*>& fetch_targets,
const std::string& fetch_holder_name) {
size_t fetch_count = 0;
for (auto* op : block->AllOps()) {
if (op->Type() == kFetchOpType) {
fetch_count++;
PADDLE_ENFORCE_EQ(op->Output("Out")[0], fetch_holder_name,
"Output of fetch op should be '%s'", fetch_holder_name);
std::string fetch_target_name = op->Input("X")[0];
PADDLE_ENFORCE(
fetch_targets.find(fetch_target_name) != fetch_targets.end(),
"Fetch operator input name '%s' cannot be found in 'fetch_targets'",
fetch_target_name);
}
}
if (fetch_count > 0) {
PADDLE_ENFORCE_EQ(
fetch_count, fetch_targets.size(),
"The number of fetch operators should match 'fetch_targets'");
// When fetch operator are present, so should be fetch_holder
auto var = block->FindVar(fetch_holder_name);
PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable",
fetch_holder_name);
PADDLE_ENFORCE_EQ(var->GetType(), proto::VarDesc::FETCH_LIST,
"'%s' variable should be 'FETCH_LIST' type",
fetch_holder_name);
}
return fetch_count > 0;
}
void Executor::Run(const ProgramDesc& program, Scope* scope,
std::map<std::string, const LoDTensor*>& feed_targets,
std::map<std::string, LoDTensor*>& fetch_targets,
const std::string& feed_holder_name,
const std::string& fetch_holder_name) {
auto* copy_program = new ProgramDesc(program);
auto* global_block = copy_program->MutableBlock(0);
if (!has_feed_operators(global_block, feed_targets, feed_holder_name)) {
// create feed_holder variable
auto* feed_holder = global_block->Var(feed_holder_name);
feed_holder->SetType(proto::VarDesc::FEED_MINIBATCH);
feed_holder->SetPersistable(true);
int i = 0;
for (auto& feed_target : feed_targets) {
std::string var_name = feed_target.first;
VLOG(3) << "feed target's name: " << var_name;
// prepend feed op
auto* op = global_block->PrependOp();
op->SetType(kFeedOpType);
op->SetInput("X", {feed_holder_name});
op->SetOutput("Out", {var_name});
op->SetAttr("col", {static_cast<int>(i)});
op->CheckAttrs();
i++;
}
}
// map the data of feed_targets to feed_holder
for (auto* op : global_block->AllOps()) {
if (op->Type() == kFeedOpType) {
std::string feed_target_name = op->Output("Out")[0];
int idx = boost::get<int>(op->GetAttr("col"));
SetFeedVariable(scope, *feed_targets[feed_target_name], feed_holder_name,
idx);
}
}
if (!has_fetch_operators(global_block, fetch_targets, fetch_holder_name)) {
// create fetch_holder variable
auto* fetch_holder = global_block->Var(fetch_holder_name);
fetch_holder->SetType(proto::VarDesc::FETCH_LIST);
fetch_holder->SetPersistable(true);
int i = 0;
for (auto& fetch_target : fetch_targets) {
std::string var_name = fetch_target.first;
VLOG(3) << "fetch target's name: " << var_name;
// append fetch op
auto* op = global_block->AppendOp();
op->SetType(kFetchOpType);
op->SetInput("X", {var_name});
op->SetOutput("Out", {fetch_holder_name});
op->SetAttr("col", {static_cast<int>(i)});
op->CheckAttrs();
i++;
}
}
Run(*copy_program, scope, 0, true, true);
// obtain the data of fetch_targets from fetch_holder
for (auto* op : global_block->AllOps()) {
if (op->Type() == kFetchOpType) {
std::string fetch_target_name = op->Input("X")[0];
int idx = boost::get<int>(op->GetAttr("col"));
*fetch_targets[fetch_target_name] =
GetFetchVariable(*scope, fetch_holder_name, idx);
}
}
delete copy_program;
}
} // namespace framework
} // namespace paddle

@ -41,6 +41,12 @@ class Executor {
void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true,
bool create_vars = true);
void Run(const ProgramDesc& program, Scope* scope,
std::map<std::string, const LoDTensor*>& feed_targets,
std::map<std::string, LoDTensor*>& fetch_targets,
const std::string& feed_holder_name = "feed",
const std::string& fetch_holder_name = "fetch");
private:
const platform::Place place_;
};

@ -0,0 +1,56 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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/framework/feed_fetch_method.h"
#include "glog/logging.h"
#include "paddle/framework/variable.h"
namespace paddle {
namespace framework {
void SetFeedVariable(Scope* scope, const LoDTensor& input,
const std::string& var_name, size_t index) {
// If var_name Variable is not found in GlobalScope, a new variable will
// be created.
VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
Variable* g_feed_value = scope->Var(var_name);
auto& feed_inputs =
*(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
if (index >= feed_inputs.size()) {
feed_inputs.resize(index + 1);
}
// shared data with input tensor
feed_inputs[index].ShareDataWith(input);
// set lod
feed_inputs[index].set_lod(input.lod());
}
LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
size_t index) {
// Since we want to fetch LodTensor from a variable, the variable must
// be created alreadly.
Variable* g_fetch_value = scope.FindVar(var_name);
PADDLE_ENFORCE(g_fetch_value->IsType<FeedFetchList>(),
"Only %s can be invoked by GetFetchVariable",
typeid(FeedFetchList).name());
auto& fetch_outputs = *g_fetch_value->GetMutable<FeedFetchList>();
auto& tensor = fetch_outputs[index];
VLOG(3) << "Fetch " << var_name << " with index " << index
<< " shape= " << tensor.dims();
PADDLE_ENFORCE_LT(index, fetch_outputs.size());
return tensor;
}
} // namespace framework
} // namespace paddle

@ -13,46 +13,18 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "glog/logging.h"
#include "paddle/framework/feed_fetch_type.h"
#include "paddle/framework/scope.h"
#include "paddle/framework/variable.h"
namespace paddle {
namespace framework {
void SetFeedVariable(Scope* scope, const LoDTensor& input,
const std::string& var_name, size_t index) {
// If var_name Variable is not found in GlobalScope, a new variable will
// be created.
VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
Variable* g_feed_value = scope->Var(var_name);
auto& feed_inputs =
*(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
if (index >= feed_inputs.size()) {
feed_inputs.resize(index + 1);
}
// shared data with input tensor
feed_inputs[index].ShareDataWith(input);
// set lod
feed_inputs[index].set_lod(input.lod());
}
const std::string& var_name, size_t index);
LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
size_t index) {
// Since we want to fetch LodTensor from a variable, the variable must
// be created alreadly.
Variable* g_fetch_value = scope.FindVar(var_name);
PADDLE_ENFORCE(g_fetch_value->IsType<FeedFetchList>(),
"Only %s can be invoked by GetFetchVariable",
typeid(FeedFetchList).name());
auto& fetch_outputs = *g_fetch_value->GetMutable<FeedFetchList>();
auto& tensor = fetch_outputs[index];
VLOG(3) << "Fetch " << var_name << " with index " << index
<< " shape= " << tensor.dims();
PADDLE_ENFORCE_LT(index, fetch_outputs.size());
return tensor;
}
size_t index);
} // namespace framework
} // namespace paddle

@ -26,9 +26,9 @@ TEST(OpKernelType, ToString) {
OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
LibraryType::kCUDNN);
ASSERT_EQ(
paddle::framework::KernelTypeToString(op_kernel_type),
"data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]");
ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type),
"data_type[float32]:data_layout[NCHW]:place[CPUPlace]:library_type["
"CUDNN]");
}
TEST(OpKernelType, Hash) {

@ -1,24 +1,93 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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
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
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. */
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/framework/threadpool.h"
namespace paddle {
namespace framework {
std::unique_ptr<ThreadPool> ThreadPool::threadpool(nullptr);
std::once_flag ThreadPool::init_flag;
std::unique_ptr<ThreadPool> ThreadPool::threadpool_(nullptr);
std::once_flag ThreadPool::init_flag_;
ThreadPool* ThreadPool::GetInstance() {
std::call_once(init_flag_, &ThreadPool::Init);
return threadpool_.get();
}
void ThreadPool::Init() {
if (threadpool_.get() == nullptr) {
// TODO(Yancey1989): specify the max threads number
int num_threads = std::thread::hardware_concurrency();
PADDLE_ENFORCE_GT(num_threads, 0);
threadpool_.reset(new ThreadPool(num_threads));
}
}
ThreadPool::ThreadPool(int num_threads)
: total_threads_(num_threads), idle_threads_(num_threads), running_(true) {
threads_.resize(num_threads);
for (auto& thread : threads_) {
// TODO(Yancey1989): binding the thread on the specify CPU number
thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this)));
}
}
ThreadPool::~ThreadPool() {
{
// notify all threads to stop running
running_ = false;
scheduled_.notify_all();
}
for (auto& t : threads_) {
t->join();
t.reset(nullptr);
}
}
void ThreadPool::Wait() {
std::unique_lock<std::mutex> lock(mutex_);
completed_.wait(lock, [=] { return Done() == true; });
}
void ThreadPool::TaskLoop() {
while (running_) {
std::unique_lock<std::mutex> lock(mutex_);
scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
if (!running_) {
break;
}
// pop a task from the task queue
auto task = std::move(tasks_.front());
tasks_.pop();
--idle_threads_;
lock.unlock();
// run the task
task();
{
std::unique_lock<std::mutex> lock(mutex_);
++idle_threads_;
if (Done()) {
completed_.notify_all();
}
}
}
}
} // namespace framework
} // namespace paddle

@ -20,52 +20,36 @@ limitations under the License. */
#include <mutex>
#include <queue>
#include <thread>
#include <vector>
#include "paddle/platform/enforce.h"
namespace paddle {
namespace framework {
// ThreadPool maintains a queue of tasks, and runs them using a fixed
// number of threads.
class ThreadPool {
public:
typedef std::packaged_task<void()> Task;
/**
* @brief Get a instance of threadpool, the thread number will
* be specified as the number of hardware thread contexts
*/
static ThreadPool* GetInstance() {
std::call_once(init_flag, &ThreadPool::Init);
return threadpool.get();
}
// Returns the singleton of ThreadPool.
static ThreadPool* GetInstance();
~ThreadPool() {
{
// notify all threads to stop running
running_ = false;
scheduled_.notify_all();
}
for (auto& t : threads_) {
t->join();
t.reset(nullptr);
}
}
~ThreadPool();
int GetNumThreads() const { return num_threads_; }
// Returns the number of threads created by the constructor.
size_t Threads() const { return total_threads_; }
int GetAvailable() {
// Returns the number of currently idle threads.
size_t IdleThreads() {
std::unique_lock<std::mutex> lock(mutex_);
return available_;
return idle_threads_;
}
/**
* @brief Push a function to the queue, and will be scheduled and
* executed if a thread is available.
* @param[in] Task, will be pushed to the task queue.
* @return std::future<void>, we could wait for the task finished by
* f.wait().
*/
// Run pushes a function to the task queue and returns a std::future
// object. To wait for the completion of the task, call
// std::future::wait().
template <typename Callback>
std::future<void> Run(Callback fn) {
std::unique_lock<std::mutex> lock(mutex_);
@ -77,84 +61,40 @@ class ThreadPool {
return f;
}
/**
* @brief Wait until all the tasks are completed.
*/
void Wait() {
std::unique_lock<std::mutex> lock(mutex_);
completed_.wait(lock, [=] { return Done() == true; });
}
// Wait until all the tasks are completed.
void Wait();
private:
DISABLE_COPY_AND_ASSIGN(ThreadPool);
explicit ThreadPool(int num_threads)
: num_threads_(num_threads), available_(num_threads), running_(true) {
threads_.resize(num_threads);
for (auto& thread : threads_) {
// TODO(Yancey1989): binding the thread on the specify CPU number
thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this)));
}
}
explicit ThreadPool(int num_threads);
/**
* @brief If the task queue is empty and avaialbe
* is equal to the number of threads, means that
* all tasks are completed.
*
* Note: this function is not thread-safe.
*
* @return true if all tasks are completed.
*/
bool Done() { return tasks_.empty() && available_ == num_threads_; }
void TaskLoop() {
while (running_) {
std::unique_lock<std::mutex> lock(mutex_);
scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
if (!running_) {
break;
}
// pop a task from the task queue
auto task = std::move(tasks_.front());
tasks_.pop();
--available_;
lock.unlock();
// run the task
task();
{
std::unique_lock<std::mutex> lock(mutex_);
++available_;
if (Done()) {
completed_.notify_all();
}
}
}
}
// If the task queue is empty and avaialbe is equal to the number of
// threads, means that all tasks are completed. Note: this function
// is not thread-safe. Returns true if all tasks are completed.
// Note: don't delete the data member total_threads_ and use
// threads_.size() instead; because you'd need to lock the mutex
// before accessing threads_.
bool Done() { return tasks_.empty() && idle_threads_ == total_threads_; }
static void Init() {
if (threadpool.get() == nullptr) {
// TODO(Yancey1989): specify the max threads number
int num_threads = std::thread::hardware_concurrency();
PADDLE_ENFORCE_GT(num_threads, 0);
threadpool.reset(new ThreadPool(num_threads));
}
}
// The constructor starts threads to run TaskLoop, which retrieves
// and runs tasks from the queue.
void TaskLoop();
// Init is called by GetInstance.
static void Init();
private:
static std::unique_ptr<ThreadPool> threadpool;
static std::once_flag init_flag;
static std::unique_ptr<ThreadPool> threadpool_;
static std::once_flag init_flag_;
int num_threads_;
int available_;
bool running_;
std::queue<Task> tasks_;
std::vector<std::unique_ptr<std::thread>> threads_;
const size_t total_threads_;
size_t idle_threads_;
std::queue<Task> tasks_;
std::mutex mutex_;
bool running_;
std::condition_variable scheduled_;
std::condition_variable completed_;
};

@ -22,11 +22,7 @@ namespace framework = paddle::framework;
void do_sum(framework::ThreadPool* pool, std::atomic<int>& sum, int cnt) {
std::vector<std::future<void>> fs;
for (int i = 0; i < cnt; ++i) {
auto f = pool->Run([&sum]() { sum.fetch_add(1); });
fs.push_back(std::move(f));
}
for (auto& f : fs) {
f.wait();
fs.push_back(framework::Async([&sum]() { sum.fetch_add(1); }));
}
}

@ -15,7 +15,6 @@ limitations under the License. */
#include "inference.h"
#include <fstream>
#include "paddle/framework/executor.h"
#include "paddle/framework/feed_fetch_method.h"
#include "paddle/framework/init.h"
#include "paddle/framework/scope.h"
@ -154,7 +153,7 @@ void InferenceEngine::Execute(const std::vector<framework::LoDTensor>& feeds,
LOG(FATAL) << "Please initialize the program_ and load_program_ first.";
}
if (feeds.size() < feed_var_names_.size()) {
if (feeds.size() != feed_var_names_.size()) {
LOG(FATAL) << "Please feed " << feed_var_names_.size() << " input Tensors.";
}
@ -165,19 +164,22 @@ void InferenceEngine::Execute(const std::vector<framework::LoDTensor>& feeds,
executor->Run(*load_program_, scope, 0, true, true);
std::map<std::string, const framework::LoDTensor*> feed_targets;
std::map<std::string, framework::LoDTensor*> fetch_targets;
// set_feed_variable
for (size_t i = 0; i < feed_var_names_.size(); ++i) {
framework::SetFeedVariable(scope, feeds[i], "feed", i);
feed_targets[feed_var_names_[i]] = &feeds[i];
}
executor->Run(*program_, scope, 0, true, true);
// get_fetch_variable
fetchs.resize(fetch_var_names_.size());
for (size_t i = 0; i < fetch_var_names_.size(); ++i) {
fetchs[i] = framework::GetFetchVariable(*scope, "fetch", i);
fetch_targets[fetch_var_names_[i]] = &fetchs[i];
}
executor->Run(*program_, scope, feed_targets, fetch_targets);
delete place;
delete scope;
delete executor;

@ -323,7 +323,7 @@ template <typename T>
struct FloorFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out>
void operator()(Device d, X x, Out out) const {
out.device(d) = x.ceil();
out.device(d) = x.floor();
}
};

@ -101,8 +101,8 @@ bool RPCClient::Wait() {
if (req_count_ <= 0) {
return true;
}
std::vector<bool> a(req_count_);
const size_t kReqCnt = req_count_;
bool a[kReqCnt];
std::vector<std::future<void>> waits(req_count_);
for (int i = 0; i < req_count_; i++) {

@ -135,14 +135,14 @@ class GRUOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
GRU Operator implements part calculations of the complete GRU as following:
\f[
update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r) \\
output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
$$
update\_gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
reset\_gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r) \\
output\_candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)
\f]
$$
@note To implement the complete GRU, fully-connected operator must be used
@note To implement the complete GRU, fully-connected operator must be used
before to feed xu, xr and xc as the Input of GRU operator.
)DOC");
}

File diff suppressed because it is too large Load Diff

@ -0,0 +1,24 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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/operators/lstmp_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
lstmp, ops::LSTMPKernel<paddle::platform::CUDADeviceContext, float>,
ops::LSTMPKernel<paddle::platform::CUDADeviceContext, double>);
REGISTER_OP_CUDA_KERNEL(
lstmp_grad,
ops::LSTMPGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::LSTMPGradKernel<paddle::platform::CUDADeviceContext, double>);

File diff suppressed because it is too large Load Diff

@ -119,7 +119,13 @@ REGISTER_OPERATOR(multiplex, ops::MultiplexOp, ops::MultiplexOpMaker,
REGISTER_OPERATOR(multiplex_grad, ops::MultiplexGradOp);
REGISTER_OP_CPU_KERNEL(
multiplex,
ops::MultiplexCPUKernel<paddle::platform::CPUDeviceContext, float>);
ops::MultiplexCPUKernel<paddle::platform::CPUDeviceContext, float>,
ops::MultiplexCPUKernel<paddle::platform::CPUDeviceContext, double>,
ops::MultiplexCPUKernel<paddle::platform::CPUDeviceContext, int>,
ops::MultiplexCPUKernel<paddle::platform::CPUDeviceContext, int64_t>);
REGISTER_OP_CPU_KERNEL(
multiplex_grad,
ops::MultiplexGradCPUKernel<paddle::platform::CPUDeviceContext, float>);
ops::MultiplexGradCPUKernel<paddle::platform::CPUDeviceContext, float>,
ops::MultiplexGradCPUKernel<paddle::platform::CPUDeviceContext, double>,
ops::MultiplexGradCPUKernel<paddle::platform::CPUDeviceContext, int>,
ops::MultiplexGradCPUKernel<paddle::platform::CPUDeviceContext, int64_t>);

@ -90,7 +90,13 @@ namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
multiplex,
ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, float>);
ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, float>,
ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, double>,
ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, int>,
ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
multiplex_grad,
ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, float>);
ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, float>,
ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, double>,
ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, int>,
ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, int64_t>);

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

Loading…
Cancel
Save