Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into grpc_benchmark
typhoonzero
7 years ago
commit
e834eb870a
@ -0,0 +1,81 @@
|
||||
/* 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 "paddle/framework/data_layout.h"
|
||||
#include "paddle/framework/data_type.h"
|
||||
#include "paddle/framework/library_type.h"
|
||||
#include "paddle/platform/device_context.h"
|
||||
#include "paddle/platform/place.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace framework {
|
||||
|
||||
struct OpKernelType {
|
||||
struct Hash {
|
||||
size_t operator()(const OpKernelType& key) const {
|
||||
int place = key.place_.which() + (1 << LEFT_SHIFT);
|
||||
int data_type =
|
||||
static_cast<int>(key.data_type_) + (1 << (LEFT_SHIFT + 1));
|
||||
int data_layout =
|
||||
static_cast<int>(key.data_layout_) + (1 << (LEFT_SHIFT + 2));
|
||||
int library_type =
|
||||
static_cast<int>(key.library_type_) + (1 << (LEFT_SHIFT + 3));
|
||||
std::hash<int> hasher;
|
||||
return hasher(place + data_type + data_layout + library_type);
|
||||
}
|
||||
};
|
||||
|
||||
// place, data_type, library_type kinds less than 2^8
|
||||
constexpr static int LEFT_SHIFT = 8;
|
||||
proto::DataType data_type_;
|
||||
DataLayout data_layout_;
|
||||
platform::Place place_;
|
||||
LibraryType library_type_;
|
||||
|
||||
OpKernelType(proto::DataType data_type, platform::Place place,
|
||||
DataLayout data_layout = DataLayout::kAnyLayout,
|
||||
LibraryType library_type = LibraryType::kPlain)
|
||||
: data_type_(data_type),
|
||||
data_layout_(data_layout),
|
||||
place_(place),
|
||||
library_type_(library_type) {}
|
||||
|
||||
OpKernelType(proto::DataType data_type,
|
||||
const platform::DeviceContext& dev_ctx,
|
||||
DataLayout data_layout = DataLayout::kAnyLayout,
|
||||
LibraryType library_type = LibraryType::kPlain)
|
||||
: data_type_(data_type),
|
||||
data_layout_(data_layout),
|
||||
place_(dev_ctx.GetPlace()),
|
||||
library_type_(library_type) {}
|
||||
|
||||
bool operator==(const OpKernelType& o) const {
|
||||
return platform::places_are_same_class(place_, o.place_) &&
|
||||
data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
|
||||
library_type_ == o.library_type_;
|
||||
}
|
||||
};
|
||||
|
||||
inline std::ostream& operator<<(std::ostream& os,
|
||||
const OpKernelType& kernel_key) {
|
||||
os << "data_type[" << kernel_key.data_type_ << "]:data_layout["
|
||||
<< kernel_key.data_layout_ << "]:place[" << kernel_key.place_
|
||||
<< "]:library_type[" << kernel_key.library_type_ << "]";
|
||||
return os;
|
||||
}
|
||||
|
||||
} // namespace framework
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,51 @@
|
||||
/* 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/op_kernel_type.h"
|
||||
#include <gtest/gtest.h>
|
||||
#include <iostream>
|
||||
|
||||
TEST(OpKernelType, ToString) {
|
||||
using OpKernelType = paddle::framework::OpKernelType;
|
||||
using DataType = paddle::framework::proto::DataType;
|
||||
using CPUPlace = paddle::platform::CPUPlace;
|
||||
using DataLayout = paddle::framework::DataLayout;
|
||||
using LibraryType = paddle::framework::LibraryType;
|
||||
|
||||
OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
|
||||
LibraryType::kCUDNN);
|
||||
|
||||
std::ostringstream stream;
|
||||
stream << op_kernel_type;
|
||||
ASSERT_EQ(
|
||||
stream.str(),
|
||||
"data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]");
|
||||
}
|
||||
|
||||
TEST(OpKernelType, Hash) {
|
||||
using OpKernelType = paddle::framework::OpKernelType;
|
||||
using DataType = paddle::framework::proto::DataType;
|
||||
using CPUPlace = paddle::platform::CPUPlace;
|
||||
using GPUPlace = paddle::platform::GPUPlace;
|
||||
using DataLayout = paddle::framework::DataLayout;
|
||||
using LibraryType = paddle::framework::LibraryType;
|
||||
|
||||
OpKernelType op_kernel_type_1(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
|
||||
LibraryType::kCUDNN);
|
||||
OpKernelType op_kernel_type_2(DataType::FP32, GPUPlace(0), DataLayout::kNCHW,
|
||||
LibraryType::kCUDNN);
|
||||
|
||||
OpKernelType::Hash hasher;
|
||||
ASSERT_NE(hasher(op_kernel_type_1), hasher(op_kernel_type_2));
|
||||
}
|
||||
@ -0,0 +1,161 @@
|
||||
/* 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 <cstdio>
|
||||
#include <functional>
|
||||
#include <iostream>
|
||||
#include <mutex>
|
||||
#include <queue>
|
||||
#include <thread>
|
||||
|
||||
#include "paddle/platform/call_once.h"
|
||||
#include "paddle/platform/enforce.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace framework {
|
||||
|
||||
typedef std::function<void()> Task;
|
||||
|
||||
class ThreadPool {
|
||||
public:
|
||||
/**
|
||||
* @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();
|
||||
}
|
||||
|
||||
~ThreadPool() {
|
||||
{
|
||||
// notify all threads to stop running
|
||||
running_ = false;
|
||||
scheduled_.notify_all();
|
||||
}
|
||||
|
||||
for (auto& t : threads_) {
|
||||
t->join();
|
||||
t.reset(nullptr);
|
||||
}
|
||||
}
|
||||
|
||||
int GetNumThreads() const { return num_threads_; }
|
||||
|
||||
int GetAvailable() {
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
return available_;
|
||||
}
|
||||
|
||||
/**
|
||||
* @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.
|
||||
*/
|
||||
void Run(const Task& fn) {
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
tasks_.push(fn);
|
||||
lock.unlock();
|
||||
scheduled_.notify_one();
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Wait until all the tasks are completed.
|
||||
*/
|
||||
void Wait() {
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
completed_.wait(lock, [=] { return Done() == true; });
|
||||
}
|
||||
|
||||
private:
|
||||
ThreadPool& operator=(const ThreadPool&) = delete;
|
||||
ThreadPool(const ThreadPool&) = delete;
|
||||
|
||||
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)));
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* @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 = tasks_.front();
|
||||
tasks_.pop();
|
||||
|
||||
--available_;
|
||||
lock.unlock();
|
||||
|
||||
// run the task
|
||||
task();
|
||||
|
||||
{
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
++available_;
|
||||
if (Done()) {
|
||||
completed_.notify_all();
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
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));
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
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_;
|
||||
std::mutex mutex_;
|
||||
std::condition_variable scheduled_;
|
||||
std::condition_variable completed_;
|
||||
};
|
||||
|
||||
std::unique_ptr<ThreadPool> ThreadPool::threadpool(nullptr);
|
||||
std::once_flag ThreadPool::init_flag;
|
||||
} // namespace framework
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,58 @@
|
||||
/* 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 "threadpool.h"
|
||||
#include <gtest/gtest.h>
|
||||
#include <atomic>
|
||||
#include <chrono>
|
||||
#include <map>
|
||||
#include <thread>
|
||||
|
||||
namespace framework = paddle::framework;
|
||||
|
||||
void do_sum(framework::ThreadPool* pool, std::atomic<int>& sum, int cnt) {
|
||||
for (int i = 0; i < cnt; ++i) {
|
||||
pool->Run([&sum]() { sum.fetch_add(1); });
|
||||
}
|
||||
}
|
||||
|
||||
TEST(ThreadPool, ConcurrentInit) {
|
||||
framework::ThreadPool* pool;
|
||||
int concurrent_cnt = 50;
|
||||
std::vector<std::thread> threads;
|
||||
for (int i = 0; i < concurrent_cnt; ++i) {
|
||||
std::thread t([&pool]() { pool = framework::ThreadPool::GetInstance(); });
|
||||
threads.push_back(std::move(t));
|
||||
}
|
||||
for (auto& t : threads) {
|
||||
t.join();
|
||||
}
|
||||
}
|
||||
|
||||
TEST(ThreadPool, ConcurrentStart) {
|
||||
framework::ThreadPool* pool = framework::ThreadPool::GetInstance();
|
||||
std::atomic<int> sum(0);
|
||||
std::vector<std::thread> threads;
|
||||
int concurrent_cnt = 50;
|
||||
// sum = (n * (n + 1)) / 2
|
||||
for (int i = 1; i <= concurrent_cnt; ++i) {
|
||||
std::thread t(do_sum, pool, std::ref(sum), i);
|
||||
threads.push_back(std::move(t));
|
||||
}
|
||||
for (auto& t : threads) {
|
||||
t.join();
|
||||
}
|
||||
pool->Wait();
|
||||
EXPECT_EQ(sum, ((concurrent_cnt + 1) * concurrent_cnt) / 2);
|
||||
}
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue