2019-07-11 13:56:52 +08:00
|
|
|
//
|
|
|
|
// ThreadPool.cpp
|
|
|
|
// MNN
|
|
|
|
//
|
|
|
|
// Created by MNN on 2019/06/30.
|
|
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
|
|
//
|
|
|
|
#ifdef MNN_USE_THREAD_POOL
|
2019-12-27 22:16:57 +08:00
|
|
|
#include "backend/cpu/ThreadPool.hpp"
|
2019-07-11 13:56:52 +08:00
|
|
|
#include <string.h>
|
2025-07-16 13:11:57 +08:00
|
|
|
#include <unordered_map>
|
2019-12-27 22:16:57 +08:00
|
|
|
#include <MNN/MNNDefine.h>
|
2025-07-16 13:11:57 +08:00
|
|
|
#include "ThreadPool.hpp"
|
2022-08-25 12:25:30 +08:00
|
|
|
|
2019-07-19 17:09:09 +08:00
|
|
|
#define MNN_THREAD_POOL_MAX_TASKS 2
|
2019-07-11 13:56:52 +08:00
|
|
|
namespace MNN {
|
2025-07-16 13:11:57 +08:00
|
|
|
static std::unordered_map<long int, ThreadPool*> gInstances;
|
2019-07-11 13:56:52 +08:00
|
|
|
static std::mutex gInitMutex;
|
2025-07-16 13:11:57 +08:00
|
|
|
int ThreadPool::init(int numberThread, unsigned long cpuMask, ThreadPool*& threadPool) {
|
|
|
|
if (1 >= numberThread) {
|
|
|
|
numberThread = 1;
|
2019-07-19 17:09:09 +08:00
|
|
|
}
|
2019-07-30 17:12:52 +08:00
|
|
|
std::lock_guard<std::mutex> _l(gInitMutex);
|
2025-07-16 13:11:57 +08:00
|
|
|
|
|
|
|
if (gInstances.find(cpuMask) == gInstances.end()){
|
|
|
|
gInstances[cpuMask] = new ThreadPool(numberThread);
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
threadPool = gInstances[cpuMask];
|
|
|
|
if (gInstances[cpuMask]->numberThread() < numberThread){
|
|
|
|
return gInstances[cpuMask]->numberThread();
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
return numberThread;
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
|
2019-07-11 13:56:52 +08:00
|
|
|
void ThreadPool::destroy() {
|
2019-07-30 17:12:52 +08:00
|
|
|
std::lock_guard<std::mutex> _l(gInitMutex);
|
2025-07-16 13:11:57 +08:00
|
|
|
for (auto i= gInstances.begin(); i != gInstances.end(); i++){
|
|
|
|
if (i->second){
|
|
|
|
delete i->second;
|
|
|
|
}
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
gInstances.clear();
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
ThreadPool::ThreadPool(int numberThread) {
|
|
|
|
mNumberThread = numberThread;
|
2025-07-16 13:11:57 +08:00
|
|
|
mActiveCount = 0;
|
2020-11-05 16:41:56 +08:00
|
|
|
mTaskAvailable.resize(MNN_THREAD_POOL_MAX_TASKS);
|
2020-11-04 14:47:02 +08:00
|
|
|
mTasks.resize(MNN_THREAD_POOL_MAX_TASKS);
|
2019-07-19 17:09:09 +08:00
|
|
|
for (int t = 0; t < mTasks.size(); ++t) {
|
|
|
|
mTaskAvailable[t] = true;
|
|
|
|
for (int i = 0; i < mNumberThread; ++i) {
|
|
|
|
mTasks[t].second.emplace_back(new std::atomic_bool{false});
|
|
|
|
}
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
|
|
|
for (int i = 1; i < mNumberThread; ++i) {
|
2019-07-19 17:09:09 +08:00
|
|
|
int threadIndex = i;
|
2020-11-04 14:47:02 +08:00
|
|
|
mWorkers.emplace_back([this, threadIndex]() {
|
2019-07-11 13:56:52 +08:00
|
|
|
while (!mStop) {
|
2025-07-16 13:11:57 +08:00
|
|
|
while (mActiveCount > 0) {
|
2019-07-19 17:09:09 +08:00
|
|
|
for (int i = 0; i < MNN_THREAD_POOL_MAX_TASKS; ++i) {
|
|
|
|
if (*mTasks[i].second[threadIndex]) {
|
|
|
|
mTasks[i].first.first(threadIndex);
|
|
|
|
{ *mTasks[i].second[threadIndex] = false; }
|
|
|
|
}
|
|
|
|
}
|
|
|
|
std::this_thread::yield();
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2020-11-04 14:47:02 +08:00
|
|
|
std::unique_lock<std::mutex> _l(mQueueMutex);
|
2025-07-16 13:11:57 +08:00
|
|
|
mCondition.wait(_l, [this] { return mStop || mActiveCount > 0; });
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
|
|
|
});
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ThreadPool::~ThreadPool() {
|
2020-12-03 15:25:45 +08:00
|
|
|
{
|
|
|
|
std::lock_guard<std::mutex> _l(mQueueMutex);
|
|
|
|
mStop = true;
|
|
|
|
}
|
2020-11-04 14:47:02 +08:00
|
|
|
mCondition.notify_all();
|
2019-07-11 13:56:52 +08:00
|
|
|
for (auto& worker : mWorkers) {
|
2020-11-04 14:47:02 +08:00
|
|
|
worker.join();
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2019-07-19 17:09:09 +08:00
|
|
|
for (auto& task : mTasks) {
|
|
|
|
for (auto c : task.second) {
|
|
|
|
delete c;
|
|
|
|
}
|
|
|
|
}
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
|
|
|
|
2019-07-19 17:09:09 +08:00
|
|
|
int ThreadPool::acquireWorkIndex() {
|
2025-07-16 13:11:57 +08:00
|
|
|
std::lock_guard<std::mutex> _l(mQueueMutex);
|
2019-07-19 17:09:09 +08:00
|
|
|
for (int i = 0; i < MNN_THREAD_POOL_MAX_TASKS; ++i) {
|
2025-07-16 13:11:57 +08:00
|
|
|
if (mTaskAvailable[i]) {
|
|
|
|
mTaskAvailable[i] = false;
|
2019-07-19 17:09:09 +08:00
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
void ThreadPool::releaseWorkIndex(int index) {
|
|
|
|
if (index < 0 || index >= MNN_THREAD_POOL_MAX_TASKS) {
|
|
|
|
return;
|
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
std::lock_guard<std::mutex> _l(mQueueMutex);
|
|
|
|
mTaskAvailable[index] = true;
|
2019-07-19 17:09:09 +08:00
|
|
|
}
|
|
|
|
|
2025-07-16 13:11:57 +08:00
|
|
|
void ThreadPool::active() {
|
2020-12-03 15:25:45 +08:00
|
|
|
{
|
2025-07-16 13:11:57 +08:00
|
|
|
std::lock_guard<std::mutex> _l(mQueueMutex);
|
|
|
|
mActiveCount++;
|
2020-12-03 15:25:45 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
mCondition.notify_all();
|
2019-07-19 17:09:09 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
void ThreadPool::deactive() {
|
|
|
|
mActiveCount--;
|
2019-07-19 17:09:09 +08:00
|
|
|
}
|
|
|
|
|
2025-07-16 13:11:57 +08:00
|
|
|
void ThreadPool::enqueue(TASK&& task, int index) {
|
2019-07-19 17:09:09 +08:00
|
|
|
if (1 >= task.second || 0 > index) {
|
|
|
|
for (int i = 0; i < task.second; ++i) {
|
2019-07-11 13:56:52 +08:00
|
|
|
task.first(i);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
enqueueInternal(std::move(task), index);
|
2019-07-19 17:09:09 +08:00
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
void ThreadPool::enqueueInternal(TASK&& task, int index) {
|
|
|
|
if (mActiveCount == 0) {
|
- build:
- unify schema building in core and converter;
- add more build script for android;
- add linux build script for python;
- ops impl:
- add floor mod support in binary;
- use eltwise impl in add/max/sub/mul binary for optimization;
- remove fake double support in cast;
- fix 5d support for concat;
- add adjX and adjY support for batch matmul;
- optimize conv2d back prop filter;
- add pad mode support for conv3d;
- fix bug in conv2d & conv depthwise with very small feature map;
- optimize binary without broacast;
- add data types support for gather;
- add gather ND support;
- use uint8 data type in gather v2;
- add transpose support for matmul;
- add matrix band part;
- add dim != 4 support for padding, reshape & tensor convert;
- add pad type support for pool3d;
- make ops based on TensorFlow Lite quantization optional;
- add all & any support for reduction;
- use type in parameter as output type in reduction;
- add int support for unary;
- add variable weight support for conv2d;
- fix conv2d depthwise weights initialization;
- fix type support for transpose;
- fix grad outputs count for reduce grad and reshape grad;
- fix priorbox & detection output;
- fix metal softmax error;
- python:
- add runSessionWithCallBackInfo interface;
- add max nodes limit (1400) for visualization tool;
- fix save error in python3;
- align default dim;
- convert:
- add extra design for optimization;
- add more post converting optimizers;
- add caffe v1 weights blob support;
- add cast, unary, conv transpose support for onnx model;
- optimize batchnorm, conv with variable weights, prelu, reshape, slice, upsample for onnx model;
- add cos/sin/atan/tan support for unary for tensorflow model;
- add any/all support for reduction for tensorflow model;
- add elu, conv3d, pool3d support for tensorflow model;
- optimize argmax, batchnorm, concat, batch to space, conv with variable weights, prelu, slice for tensorflow model;
- others:
- fix size computer lock;
- fix thread pool deadlock;
- add express & parameters in express;
- rewrite blitter chooser without static map;
- add tests for expr;
2019-10-29 13:37:26 +08:00
|
|
|
for (int i = 0; i < task.second; ++i) {
|
|
|
|
task.first(i);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
2019-07-19 17:09:09 +08:00
|
|
|
int workSize = task.second;
|
2025-07-16 13:11:57 +08:00
|
|
|
if (workSize > mNumberThread) {
|
2019-07-19 17:09:09 +08:00
|
|
|
mTasks[index].first = std::make_pair(
|
2025-07-16 13:11:57 +08:00
|
|
|
[workSize, &task, this](int tId) {
|
|
|
|
for (int v = tId; v < workSize; v += mNumberThread) {
|
2019-07-19 17:09:09 +08:00
|
|
|
task.first(v);
|
|
|
|
}
|
2025-07-16 13:11:57 +08:00
|
|
|
},
|
|
|
|
mNumberThread);
|
|
|
|
workSize = mNumberThread;
|
2019-07-19 17:09:09 +08:00
|
|
|
} else {
|
|
|
|
mTasks[index].first = std::move(task);
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
|
|
|
{
|
2019-07-19 17:09:09 +08:00
|
|
|
for (int i = 1; i < workSize; ++i) {
|
|
|
|
*mTasks[index].second[i] = true;
|
|
|
|
}
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
2019-07-19 17:09:09 +08:00
|
|
|
mTasks[index].first.first(0);
|
|
|
|
bool complete = true;
|
|
|
|
do {
|
|
|
|
complete = true;
|
|
|
|
for (int i = 1; i < workSize; ++i) {
|
|
|
|
if (*mTasks[index].second[i]) {
|
|
|
|
complete = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2023-04-14 21:45:47 +08:00
|
|
|
std::this_thread::yield();
|
2019-07-19 17:09:09 +08:00
|
|
|
// FUNC_PRINT(notComplete);
|
|
|
|
} while (!complete);
|
2019-07-11 13:56:52 +08:00
|
|
|
}
|
|
|
|
} // namespace MNN
|
|
|
|
#endif
|