图像旋转--利用opencv，cuda-npp的实现。

一，软件开发中，某些场景需要旋转图像，本文注意处理的是旋转图像90°，180°，270°。任意角度的旋转不涉及。注意选择观看。

图像的旋转可以利用opencv实现cpu完成，也可以通过cuda中的npp工具实现gpu完成，gpu的效率要比cpu搞很多。具体代码片段如下：

二，opencv代码实现：

1，实现如下：

void rotateImage(int degree) {

	cv::Mat srcCopy;
	cv::Mat src_img = imread("./bird.jpg", IMREAD_UNCHANGED);
	if (degree == 90) {
		cv::Mat srcCopy = cv::Mat(src_img.rows, src_img.cols, src_img.depth());
		cv::transpose(src_img, srcCopy);
		cv::flip(srcCopy, srcCopy, 1);
	}else if (degree == 180) {
		cv::Mat srcCopy = cv::Mat(src_img.rows, src_img.cols, src_img.depth());
		cv::flip(src_img, srcCopy, -1);
	}else if (degree == 270) {
		cv::Mat srcCopy = cv::Mat(src_img.rows, src_img.cols, src_img.depth());
		cv::transpose(src_img, srcCopy);
		cv::flip(srcCopy, srcCopy, 0);
	}
	imwrite("./retOpenCV.bmp", srcCopy);	
}

有的时候，是yuv数据，需要用yuv来构造Mat对象，代码如下：

Mat YuvToMat (unsigned char * pYuvData, int & width, int & height)
{
	cv::Mat yuvImg;  
    yuvImg.create(height*3/2, width, CV_8UC1);  
    memcpy(yuvImg.data, pYuvData, framesize*sizeof(unsigned char));  
    cv::Mat rgbImg;   
    cv::cvtColor(yuvImg, rgbImg, CV_YUV2BGR_I420);  
}

二，利用cuda的npp库实现，代码如下：

1, cuda samples里给的例子，需要用到的是freeImge第三方库：

FIBITMAP* LoadImg(const char* szFile)
{
	FREE_IMAGE_FORMAT nFif;

	if (szFile == NULL || *szFile == 0)
	{
		return NULL;
	}

	if ((nFif = FreeImage_GetFileType(szFile, 0)) == FIF_UNKNOWN)
	{
		if ((nFif = FreeImage_GetFIFFromFilename(szFile)) == FIF_UNKNOWN)
		{
			return NULL;
		}
	}

	if (!FreeImage_FIFSupportsReading(nFif))
	{
		return NULL;
	}

	return FreeImage_Load(nFif, szFile);
}

void  NppbyFreeImage() {
	cudaError_t cuRet;
	NppStatus nppRet;
	BOOL fiRet = false;
	FIBITMAP* pSrcBmp = NULL;
	FIBITMAP* pDstBmp = NULL;
	unsigned char* pSrcData = NULL;
	unsigned char* pDstData = NULL;
	Npp8u* pSrcDataCUDA = NULL;
	Npp8u* pDstDataCUDA = NULL;
	NppiSize oSrcSize = { 0 };
	NppiSize oDstSize = { 0 };
	NppiRect oSrcROI = { 0 };
	NppiRect oDstROI = { 0 };
	int nImgBpp = 0;
	int nSrcPitch = 0;
	int nDstPitch = 0;
	int nSrcPitchCUDA = 0;
	int nDstPitchCUDA = 0;
	double aBoundingBox[2][2] = { 0 };
	double nAngle = 0;

	FreeImage_Initialise(0);

	/* 载入文件 */
	pSrcBmp = LoadImg("./bird.jpg");
	assert(pSrcBmp != NULL);

	//获取图像深度
	nImgBpp = (FreeImage_GetBPP(pSrcBmp) >> 3);
	cout << "freeImg nImgBpp :" << nImgBpp << endl;
	//获取数据
	pSrcData = FreeImage_GetBits(pSrcBmp);

	oSrcSize.width = (int)FreeImage_GetWidth(pSrcBmp);
	cout << "freeImg oSrcSize.width :" << oSrcSize.width << endl;
	oSrcSize.height = (int)FreeImage_GetHeight(pSrcBmp);
	cout << "freeImg oSrcSize.height :" << oSrcSize.height << endl;
	nSrcPitch = (int)FreeImage_GetPitch(pSrcBmp);
	cout << "freeImg nSrcPitch :" << nSrcPitch << endl;

	oSrcROI.x = oSrcROI.y = 0;
	oSrcROI.width = oSrcSize.width;
	oSrcROI.height = oSrcSize.height;

	nAngle = atof("90");

	/* 设置显卡,构建上下文 */
	cuRet = cudaSetDevice(0);
	assert(cuRet == cudaSuccess);

	/* 分配显存 */
	int type = 0;
	switch (nImgBpp)
	{
	case 1:
		pSrcDataCUDA = nppiMalloc_8u_C1(oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA);
		break;
	case 3:
		pSrcDataCUDA = nppiMalloc_8u_C3(oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA);
		break;
	case 4:
		pSrcDataCUDA = nppiMalloc_8u_C4(oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA);
		break;
	default:
		assert(0);
		break;
	}
	assert(pSrcDataCUDA != NULL);

	/* 将原图传入显存 */
	cudaMemcpy2D(pSrcDataCUDA, nSrcPitchCUDA, pSrcData, nSrcPitch, oSrcSize.width * nImgBpp, oSrcSize.height, cudaMemcpyHostToDevice);

	/* 计算旋转后长宽 */
	nppiGetRotateBound(oSrcROI, aBoundingBox, nAngle, 0, 0);
	oDstSize.width = (int)ceil(fabs(aBoundingBox[1][0] - aBoundingBox[0][0]));
	oDstSize.height = (int)ceil(fabs(aBoundingBox[1][1] - aBoundingBox[0][1]));

	/* 建目标图 */
	pDstBmp = FreeImage_Allocate(oDstSize.width, oDstSize.height, nImgBpp << 3);
	assert(pDstBmp != NULL);

	pDstData = FreeImage_GetBits(pDstBmp);

	nDstPitch = (int)FreeImage_GetPitch(pDstBmp);
	cout << "freeImg nDstPitch :" << nDstPitch << endl;
	oDstROI.x = oDstROI.y = 0;
	oDstROI.width = oDstSize.width;
	cout << "freeImg  oDstSize.width :" << oDstSize.width << endl;
	oDstROI.height = oDstSize.height;
	cout << "freeImg oDstSize.height :" << oDstSize.height << endl;

	/* 分配显存 */
	switch (nImgBpp)
	{
	case 1:
		pDstDataCUDA = nppiMalloc_8u_C1(oDstSize.width, oDstSize.height, &nDstPitchCUDA);
		break;
	case 3:
		pDstDataCUDA = nppiMalloc_8u_C3(oDstSize.width, oDstSize.height, &nDstPitchCUDA);
		break;
	case 4:
		pDstDataCUDA = nppiMalloc_8u_C4(oDstSize.width, oDstSize.height, &nDstPitchCUDA);
		break;
	}
	assert(pDstDataCUDA != NULL);
	cudaMemset2D(pDstDataCUDA, nDstPitchCUDA, 0, oDstSize.width * nImgBpp, oDstSize.height);

	/* 处理 */
	switch (nImgBpp)
	{
	case 1:
		nppRet = nppiRotate_8u_C1R(pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
			pDstDataCUDA, nDstPitchCUDA, oDstROI,
			nAngle, -aBoundingBox[0][0], -aBoundingBox[0][1], NPPI_INTER_CUBIC);
		break;
	case 3:
		nppRet = nppiRotate_8u_C3R(pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
			pDstDataCUDA, nDstPitchCUDA, oDstROI,
			nAngle, -aBoundingBox[0][0], -aBoundingBox[0][1], NPPI_INTER_CUBIC);
		break;
	case 4:
		nppRet = nppiRotate_8u_C4R(pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
			pDstDataCUDA, nDstPitchCUDA, oDstROI,
			nAngle, -aBoundingBox[0][0], -aBoundingBox[0][1], NPPI_INTER_CUBIC);
		break;
	}
	assert(nppRet == NPP_NO_ERROR);

	cudaMemcpy2D(pDstData, nDstPitch, pDstDataCUDA, nDstPitchCUDA, oDstSize.width * nImgBpp, oDstSize.height, cudaMemcpyDeviceToHost);

	fiRet = FreeImage_Save(FIF_BMP, pDstBmp, "./retFreeImage.bmp");
	assert(fiRet);

	nppiFree(pSrcDataCUDA);
	nppiFree(pDstDataCUDA);

	cudaDeviceReset();

	FreeImage_Unload(pSrcBmp);
	FreeImage_Unload(pDstBmp);
}

2，opencv + cuda npp，代码如下：

void  NppbyOpenCV2() {
	cudaError_t cuRet;
	NppStatus nppRet;
	BOOL fiRet = false;
	unsigned char* pSrcData = NULL;
	unsigned char* pDstData = NULL;
	Npp8u* pSrcDataCUDA = NULL;
	Npp8u* pDstDataCUDA = NULL;
	NppiSize oSrcSize = { 0 };
	NppiSize oDstSize = { 0 };
	NppiRect oSrcROI = { 0 };
	NppiRect oDstROI = { 0 };
	int nImgBpp = 0;
	int nSrcPitch = 0;
	int nDstPitch = 0;
	int nSrcPitchCUDA = 0;
	int nDstPitchCUDA = 0;
	double aBoundingBox[2][2] = { 0 };
	double nAngle = 0;

	FreeImage_Initialise(0);

	/* 载入文件 */
	Mat pSrcMat = imread("./bird.jpg", IMREAD_UNCHANGED);

	//获取图像深度
	nImgBpp = pSrcMat.channels();
	cout << "freeImg nImgBpp :" << nImgBpp << endl;
	//获取数据
	pSrcData = pSrcMat.data;

	oSrcSize.width = pSrcMat.cols;
	oSrcSize.height = pSrcMat.rows;
	nSrcPitch = pSrcMat.step;

	oSrcROI.x = oSrcROI.y = 0;
	oSrcROI.width = oSrcSize.width;
	oSrcROI.height = oSrcSize.height;

	nAngle = atof("90");

	/* 设置显卡,构建上下文 */
	cuRet = cudaSetDevice(0);
	assert(cuRet == cudaSuccess);

	/* 分配显存 */
	int type = 0;
	switch (nImgBpp)
	{
	case 1:
		pSrcDataCUDA = nppiMalloc_8u_C1(oSrcSize.width*nImgBpp, oSrcSize.height, &nSrcPitchCUDA);
		break;
	case 3:
		pSrcDataCUDA = nppiMalloc_8u_C3(oSrcSize.width*nImgBpp, oSrcSize.height, &nSrcPitchCUDA);
		break;
	case 4:
		pSrcDataCUDA = nppiMalloc_8u_C4(oSrcSize.width*nImgBpp, oSrcSize.height, &nSrcPitchCUDA);
		break;
	default:
		assert(0);
		break;
	}
	assert(pSrcDataCUDA != NULL);

	/* 将原图传入显存 */
	cudaMemcpy2D(pSrcDataCUDA, nSrcPitchCUDA, pSrcData, nSrcPitch, oSrcSize.width*nImgBpp, oSrcSize.height, cudaMemcpyHostToDevice);

	/* 计算旋转后长宽 */
	nppiGetRotateBound(oSrcROI, aBoundingBox, nAngle, 0, 0);
	oDstSize.width = (int)ceil(fabs(aBoundingBox[1][0] - aBoundingBox[0][0]));
	oDstSize.height = (int)ceil(fabs(aBoundingBox[1][1] - aBoundingBox[0][1]));

	/* 建目标图 */
	Mat pDstMat(oDstSize.height, oDstSize.width, pSrcMat.type());

	pDstData = pDstMat.data;

	nDstPitch = pDstMat.step;
	oDstROI.x = oDstROI.y = 0;
	oDstROI.width = oDstSize.width;
	oDstROI.height = oDstSize.height;

	/* 分配显存 */
	switch (nImgBpp)
	{
	case 1:
		pDstDataCUDA = nppiMalloc_8u_C1(oDstSize.width, oDstSize.height, &nDstPitchCUDA);
		break;
	case 3:
		pDstDataCUDA = nppiMalloc_8u_C3(oDstSize.width, oDstSize.height, &nDstPitchCUDA);
		break;
	case 4:
		pDstDataCUDA = nppiMalloc_8u_C4(oDstSize.width, oDstSize.height, &nDstPitchCUDA);
		break;
	}
	assert(pDstDataCUDA != NULL);
	cudaMemset2D(pDstDataCUDA, nDstPitchCUDA, 0, oDstSize.width * nDstPitch, oDstSize.height);

	/* 处理 */
	switch (nImgBpp)
	{
	case 1:
		nppRet = nppiRotate_8u_C1R(pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
			pDstDataCUDA, nDstPitchCUDA, oDstROI,
			nAngle, -aBoundingBox[0][0], -aBoundingBox[0][1], NPPI_INTER_CUBIC);
		break;
	case 3:
		nppRet = nppiRotate_8u_C3R(pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
			pDstDataCUDA, nDstPitchCUDA, oDstROI,
			nAngle, -aBoundingBox[0][0], -aBoundingBox[0][1], NPPI_INTER_CUBIC);
		break;
	case 4:
		nppRet = nppiRotate_8u_C4R(pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
			pDstDataCUDA, nDstPitchCUDA, oDstROI,
			nAngle, -aBoundingBox[0][0], -aBoundingBox[0][1], NPPI_INTER_CUBIC);
		break;
	}
	assert(nppRet == NPP_NO_ERROR);

	cudaMemcpy2D(pDstData, nDstPitch, pDstDataCUDA, nDstPitchCUDA, oDstSize.width*nImgBpp, oDstSize.height, cudaMemcpyDeviceToHost);

	imwrite("./retOpenCV2_90.bmp", pDstMat);
	assert(fiRet);

	nppiFree(pSrcDataCUDA);
	nppiFree(pDstDataCUDA);

	cudaDeviceReset();
}