High performance computing framework for image processing

frame		Introduce
GPU	NPP	NVIDIA Performance Primitives,NVIDIA The company aims at GPU Developed GPU Accelerated images 、 video 、 Signal processing library , After installation CUDA Environment will be automatically installed . By calling NPP function , You can dispense with handwriting CUDA Kernel functions , Fast development .
	CUDA	Compute Unified Device Architecture： NVIDIA Developed a parallel computing platform and programming model . It uses a graphics processor (GPU) Handling capacity of , Can greatly improve computing performance . You need to know its development language first CUDA C, And then develop .
CPU	IPP	Integrated Performance Primitives,Intel High performance multimedia function library , Contains many functions optimized from the bottom , It covers a variety of applications including image processing , Its interface form is the same as NPP The library is similar to . For image processing ,IPP The library functions provided are introduced and referenced Blog .
	TBB	Threading Building Blocks,Intel Developed for parallel programming based on C++ The framework of language , It's a set C++ Template library . It provides a lot of features , Have a higher level of abstraction than threads , Mainly used for multi-core CPU Multi thread processing acceleration under the platform .

Summarized below ：

1）NPP and IPP Are provided by the encapsulated library functions , It mainly provides general algorithms . For example, filtering in image processing 、 Color space conversion, etc .NPP be used for GPU Parallel acceleration of the platform ,IPP be used for CPU Multi thread parallel acceleration of the platform .

2）CUDA and TBB Namely GPU and CPU Parallel development framework under the platform . Typically , For image processing for loop （ Pixel by pixel ） Handle ,CUDA You can accomplish multiple tasks by writing kernel functions CUDA Parallel acceleration of cores , and TBB You can accomplish multiple tasks through its specific interface CPU Parallel processing acceleration .

3） In the process of development , use first OpenMP Conduct CPU Image processing acceleration of the platform , But found CPU High occupancy , And the processing speed has not improved . Subsequent use TBB Development , Achieved the expected goal .IPP and TBB You can go from Intel Download on , see here .

Finally, the author uses TBB Accelerated critical code snippets , It mainly completes the color correction of color image , stay Xeon E3-1230 v2 platform （4 The core 8 Threads ） On , The execution speed of the algorithm is significantly improved . The code is as follows ：

// Color image color correction 
void ColorCorrect8UC3(Mat source, Mat& dst, int nR, int nG, int nB)
{
	dst = source.clone();
	if ((nR == 100) && (nG == 100) && (nB == 100))
		return;

	Mat src = source.clone();
	if (nR < 0)
		nR = 0;
	if (nR > 100)
		nR = 100;

	if (nG < 0)
		nG = 0;
	if (nG > 100)
		nG = 100;

	if (nB < 0)
		nB = 0;
	if (nB > 100)
		nB = 100;

	int width = src.cols;
	int height = src.rows;

	unsigned char* pSrc = src.ptr();
	unsigned char* pDst = dst.ptr();

	//parallel_for coordination blocked_range2d It will be of great help to image processing 
	//blocked_range2d Parameter description of ：
	//(y Starting value ,y End value ,y Step value ,x Starting value ,x End value ,x Step value )
	tbb::parallel_for(tbb::blocked_range2d<int>(0, height, 1, 0, width, 1),
		[&](const tbb::blocked_range2d<int>& r)
	{
		for (int i = r.rows().begin(); i < r.rows().end(); ++i) 
		{
			for (int j = r.cols().begin(); j < r.cols().end(); ++j) 
			{
				pDst[i*width * 3 + j * 3 + 2] = (unsigned char)(pSrc[i*width * 3 + j * 3 + 2] * nR / 100.0);
				pDst[i*width * 3 + j * 3 + 1] = (unsigned char)(pSrc[i*width * 3 + j * 3 + 1] * nG / 100.0);
				pDst[i*width * 3 + j * 3 + 0] = (unsigned char)(pSrc[i*width * 3 + j * 3 + 0] * nB / 100.0);
			}
		}
	});
}