caffe网络在多线程中无法使用GPU的解决方案 | cpp caffe net run in multiple threads

本文首发于个人博客https://kezunlin.me/post/8d877e63/，欢迎阅读!

cpp caffe net run in multiple threads

Guide

set_mode

Caffe fails to use GPU in a new thread ？？？
see here

the `Caffe::mode_` variable that controls this is thread-local,
so ensure you’re calling `caffe.set_mode_gpu()` in each thread
before running any Caffe functions. That should solve your issue.

Caffe set_mode GPU 在多线程下失效
在main thread中设置GPU模式，在worker thread中调用网络进行检测，
GPU模式不起效，默认仍然使用CPU模式，所以速度很慢，和GPU相比慢了
10倍左右。

解决方案：在子线程中set_mode,然后调用网络进行检测。
(1)创建网络在main thread。static 网络存储在全局静态数据区。
worker thread可以直接使用。
(2) 在worker thread中检测，需要在子线程中set_mode,然后调用网络进行检测。

结论：
(1)caffe的set_mode所在的线程必须和使用nets进行forward的线程相同。否则默认使用CPU模式，速度会很慢。
(2)caffe的nets初始化可以在main thread也可以在worker thread。

code example

#include <iostream>
#include <string>
#include <thread>

#include <gtest/gtest.h>
#include <glog/logging.h>

#include <boost/date_time/posix_time/posix_time.hpp>  

// opencv
#include <opencv2/core.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc.hpp>

using namespace std;

#include "algorithm/algorithm.h"
using namespace kezunlin::algorithm;

#pragma region net-demo

void topwire_demo(bool run_in_worker_thread)
{
	if (run_in_worker_thread) {
		CaffeApi::set_mode(true, 0, 1234);// set in worker thread-1, use GPU-0
	}
    
    // do net detect 
    // ...
}

void railway_demo(bool run_in_worker_thread)
{
	if (run_in_worker_thread) {
		CaffeApi::set_mode(true, 0, 1234);// set in worker thread-1, use GPU-0
	}
    
    // do net detect 
    // ...
}

void sidewall_demo(bool run_in_worker_thread)
{
	if (run_in_worker_thread) {
		CaffeApi::set_mode(true, 0, 1234);// set in worker thread-1, use GPU-0
	}

    // do net detect 
    // ...
}

void lockcatch_demo(bool run_in_worker_thread)
{
	if (run_in_worker_thread) {
		CaffeApi::set_mode(true, 0, 1234);// set in worker thread-1, use GPU-0
	}

	// do net detect 
    // ...
}

#pragma endregion


#pragma region worker-thread-demo

void worker_thread_topwire_demo(bool run_in_worker_thread)
{
	std::thread thr(topwire_demo, run_in_worker_thread);
	thr.join();
}

void worker_thread_railway_demo(bool run_in_worker_thread)
{
	std::thread thr(railway_demo, run_in_worker_thread);
	thr.join();
}

void worker_thread_sidewall_demo(bool run_in_worker_thread)
{
	std::thread thr(sidewall_demo, run_in_worker_thread);
	thr.join();
}

void worker_thread_lockcatch_demo(bool run_in_worker_thread)
{
	std::thread thr(lockcatch_demo, run_in_worker_thread);
	thr.join();
}

#pragma endregion

enum DETECT_TYPE {
	SET_IN_MAIN_DETECT_IN_MAIN, // 主线程set_mode，主线程检测，40ms左右，使用GPU
	SET_IN_WORKER_DETECT_IN_WORKER, // 子线程set_mode，子线程检测，40ms左右，使用GPU
	SET_IN_MAIN_DETECT_IN_WORKER // 主线程set_mode，子线程检测，400ms左右，慢了10倍左右，没有使用GPU
};

void thread_demo()
{
	DETECT_TYPE detect_type = SET_IN_MAIN_DETECT_IN_MAIN;
	detect_type = SET_IN_WORKER_DETECT_IN_WORKER;
	detect_type = SET_IN_MAIN_DETECT_IN_WORKER;

	init_algorithm_api();

	switch (detect_type)
	{
	case SET_IN_MAIN_DETECT_IN_MAIN:
		topwire_demo(false);
		railway_demo(false);
		sidewall_demo(false);
		lockcatch_demo(false);
		break;
	case SET_IN_WORKER_DETECT_IN_WORKER:
		worker_thread_topwire_demo(true);
		worker_thread_railway_demo(true);
		worker_thread_sidewall_demo(true);
		worker_thread_lockcatch_demo(true);
		break;
	case SET_IN_MAIN_DETECT_IN_WORKER:
		worker_thread_topwire_demo(false);
		worker_thread_railway_demo(false);
		worker_thread_sidewall_demo(false);
		worker_thread_lockcatch_demo(false);
		break;
	default:
		break;
	}

	free_algorithm_api();
}

void test_algorithm_api()
{
	thread_demo();
}

TEST(algorithn_test, test_algorithm_api) {
	test_algorithm_api();
}
  • SET_IN_MAIN_DETECT_IN_MAIN, // 主线程set_mode，主线程检测，40ms左右，使用GPU
  • SET_IN_WORKER_DETECT_IN_WORKER, // 子线程set_mode，子线程检测，40ms左右，使用GPU
  • SET_IN_MAIN_DETECT_IN_WORKER // 主线程set_mode，子线程检测，400ms左右，慢了10倍左右，没有使用GPU

Reference

History

  • 20180712: created.

Copyright

posted @ 2019-11-13 08:55  kezunlin  阅读(520)  评论(0)  编辑  收藏  举报