OpenCV中parallel_for 和 parallel_for_学习笔记

OpenCV 从2.4.3开始加入了并行计算的函数parallel_for和parallel_for_（更准确地讲，parallel_for以前就存在于tbb模块中，但是OpenCV官网将其列在2.4.3.的New Features中，应该是重新改写过的）。

2.4.3中自带的calcOpticalFlowPyrLK函数也用parallel_for重写过了，之前我一直认为parallel_for就是用来并行计算的，之前也自己写了一些用parallel_for实现的算法。直到今天在opencv官网中看到别人的提问，才发现parallel_for实际上是serial loop，而parallel_for_才是parallel loop（OpenCV官网answer）。

为了比较for循环，parallel_for和parallel_for_ 三者的差异，下面做了一个简单的测试，对一个Mat中所有的元素（按列为单位）做立方操作。

test.hpp

/**@ Test parallel_for and parallel_for_
/**@ Author: chouclee
/**@ 03/17/2013*/
#include <opencv2/core/internal.hpp>
namespace cv
{
namespace test
{
	class parallelTestBody : public ParallelLoopBody//参考官方给出的answer，构造一个并行的循环体类
	{
	public:
		parallelTestBody(Mat& _src)//class constructor
		{
			src = &_src;
		}
		void operator()(const Range& range) const//重载操作符（）
		{
			Mat& srcMat = *src;
			int stepSrc = (int)(srcMat.step/srcMat.elemSize1());//获取每一行的元素总个数（相当于cols*channels，等同于step1)
			for (int colIdx = range.start; colIdx < range.end; ++colIdx)
			{
				float* pData = (float*)srcMat.col(colIdx).data;
				for (int i = 0; i < srcMat.rows; ++i)
					pData[i*stepSrc] = std::pow(pData[i*stepSrc],3);
			}	
		}

	private:
		Mat* src;
	};

	struct parallelTestInvoker//构造一个供parallel_for使用的循环结构体
	{
		parallelTestInvoker(Mat& _src)//struct constructor
		{
			src = &_src;
		}
		void operator()(const BlockedRange& range) const//使用BlockedRange需要包含opencv2/core/internal.hpp
		{
			Mat& srcMat = *src;
			int stepSrc = (int)(srcMat.step/srcMat.elemSize1());
			for (int colIdx = range.begin(); colIdx < range.end(); ++colIdx)
			{
				float* pData = (float*)srcMat.col(colIdx).data;
				for (int i = 0; i < srcMat.rows; ++i)
					pData[i*stepSrc] = std::pow(pData[i*stepSrc],3);
			}
		}
		Mat* src;
	};
}//namesapce test
void parallelTestWithFor(InputArray _src)//'for' loop
{
	CV_Assert(_src.kind() == _InputArray::MAT);
	Mat src = _src.getMat();
	CV_Assert(src.isContinuous());
	int stepSrc = (int)(src.step/src.elemSize1());
	for (int x = 0; x < src.cols; ++x)
	{
		float* pData = (float*)src.col(x).data;
		for (int y = 0; y < src.rows; ++y)
			pData[y*stepSrc] = std::pow(pData[y*stepSrc], 3);
	}
};

void parallelTestWithParallel_for(InputArray _src)//'parallel_for' loop
{
	CV_Assert(_src.kind() == _InputArray::MAT);
	Mat src = _src.getMat();
	int totalCols = src.cols;
	typedef test::parallelTestInvoker parallelTestInvoker;
	parallel_for(BlockedRange(0, totalCols), parallelTestInvoker(src));
};

void parallelTestWithParallel_for_(InputArray _src)//'parallel_for_' loop
{
	CV_Assert(_src.kind() == _InputArray::MAT);
	Mat src = _src.getMat();
	int totalCols = src.cols;
	typedef test::parallelTestBody parallelTestBody;
	parallel_for_(Range(0, totalCols), parallelTestBody(src));
};
}//namespace cv

main.cpp

/**@ Test parallel_for and parallel_for_
/**@ Author: chouclee
/**@ 03/17/2013*/
#include <opencv2/opencv.hpp>
#include <time.h>
#include "test.hpp"
using namespace cv;
using namespace std;

int main(int argc, char* argv[])
{
	Mat testInput = Mat::ones(40,400000, CV_32F);
	clock_t start, stop;

	start = clock();
	parallelTestWithFor(testInput);
	stop = clock();
	cout<<"Running time using \'for\':"<<(double)(stop - start)/CLOCKS_PER_SEC*1000<<"ms"<<endl;

	start = clock();
	parallelTestWithParallel_for(testInput);
	stop = clock();
	cout<<"Running time using \'parallel_for\':"<<(double)(stop - start)/CLOCKS_PER_SEC*1000<<"ms"<<endl;

	start = clock();
	parallelTestWithParallel_for_(testInput);
	stop = clock();
	cout<<"Running time using \'parallel_for_\':"<<(double)(stop - start)/CLOCKS_PER_SEC*1000<<"ms"<<endl;

	system("pause");
}

输入为400000*40时，结果如下：
Debug模式
Running time using 'for': 1376ms
Running time using 'parallel_for': 1316ms
Running time using 'parallel_for_': 553ms
Release模式
Running time using 'for': 463ms
Running time using 'parallel_for': 475ms
Running time using 'parallel_for_': 301ms

输入改为40*400000
Debug模式
Running time using 'for': 1005ms
Running time using 'parallel_for': 1013ms
Running time using 'parallel_for_': 526ms
Release模式
Running time using 'for': 105ms
Running time using 'parallel_for': 106ms
Running time using 'parallel_for_': 81ms

输入改为4000*4000
Debug模式
Running time using 'for': 1138ms
Running time using 'parallel_for': 1136ms
Running time using 'parallel_for_': 411ms
Release模式
Running time using 'for': 234ms
Running time using 'parallel_for': 239ms
Running time using 'parallel_for_': 130ms

大多数情况下，parallel_for比for循环慢那么一丁丁点儿，有时甚至会比for循环快一些，总体上两者差不多，parallel_for_一直都是最快的。但上面的代码只是做测试使用（因此强制按列进行操作），实际上，像上面这种简单的操作，直接对Mat使用for循环和指针递增操作，只需要几十毫秒。但是，对于复杂算法，比如光流或之类的，使用parallel_for（虽然不是并行操作，但代码简洁易于维护，且速度和for循环差不多）或者parallel_for_将是不错的选择。

Reference:

http://answers.opencv.org/question/3730/how-to-use-parallel_for/