CUDA库之nvjpeg(一)：入门介绍

---

• nvJPEG: 一个GPU加速的JPEG编解码库
• 官方文档： https://docs.nvidia.com/cuda/nvjpeg/index.html
• 官方教程：https://github.com/NVIDIA/CUDALibrarySamples/tree/master/nvJPEG

---

一、介绍

1.1 nvJPEG解码器

nvJPEG 库提供了高性能，GPU加速的JPEG图像格式的解码函数，并且普遍应用在深度学习领域和超大规模的多媒体应用中。

这个库提供单张图或者多张图同时解码的能力，可以充分利用GPU资源和优化效率，并且使用者也可以管理需要解码的内存，灵活性还是比较强的。

nvJPEG大体上依次使用以下函数：

• 使用JPEG图像数据流作为输入
• 从数据流中获取图像的宽和高
• 使用以上获取的信息来管理GPU内存并执行解码操作

nvJPEG提供了专用的API，用于从原始JPEG图像数据流中检索图像信息。

二、JPEG解码

nvJPEG库提供了单张图像解码和批量解码的函数

2.1 单张图像解码

先加载图像文件，以及图像文件的大小，输出就是解码后的buffer
以下是执行步骤

• step0. 加载图像数据

	std::string strFileName = "./input_images/img2.jpg";
	// Read an image from disk.
	std::ifstream input(strFileName.c_str(),
		std::ios::in | std::ios::binary | std::ios::ate);
	if (!(input.is_open())) {
		std::cerr << "Cannot open image: " << strFileName
			<< ", removing it from image list" << std::endl;
	}
	// Get the size
	std::streamsize file_size = input.tellg();
	input.seekg(0, std::ios::beg);

	 char* pchData = (char*)malloc(file_size);
	if (!input.read(pchData, file_size)) {
		std::cerr << "Cannot read from file: " << strFileName
			<< ", removing it from image list" << std::endl;

	}

• step1. 创建nvJPEG库的句柄（有两个函数nvjpegCreateSimple() or nvjpegCreateEx()）

	nvjpegHandle* nvjpegHandle = nullptr;
	CHECK_NVJPEG(nvjpegCreateSimple(&nvjpegHandle));

• step2. 调用nvjpegJpegStateCreate()函数，创建nvJPEG的状态

	nvjpegJpegState* nvjpegJpegState = nullptr;
	CHECK_NVJPEG(nvjpegJpegStateCreate(nvjpegHandle, &nvjpegJpegState));

除了以上外，还有其他的一些函数，也是有帮助的，这里先列举，可暂时忽略

nvjpegStatus_t nvjpegGetProperty(libraryPropertyType type, int *value);

[DEPRECATED] nvjpegStatus_t nvjpegCreate(nvjpegBackend_t backend, nvjpegHandle_t *handle , nvjpeg_dev_allocator allocator);

nvjpegStatus_t nvjpegCreateSimple(nvjpegHandle_t *handle);

nvjpegStatus_t nvjpegCreateEx(nvjpegBackend_t backend, nvjpegDevAllocator_t *dev_allocator, nvjpegPinnedAllocator_t *pinned_allocator, unsigned int flags, nvjpegHandle_t *handle);

nvjpegStatus_t nvjpegDestroy(nvjpegHandle_t handle);

nvjpegStatus_t nvjpegJpegStateCreate(nvjpegHandle_t handle, nvjpegJpegState_t *jpeg_handle);

nvjpegStatus_t nvjpegJpegStateDestroy(nvjpegJpegState handle);

• step3. 使用nvjpegGetImageInfo()从数据流中获取获取宽高
  函数的参数如下：

nvjpegStatus_t nvjpegGetImageInfo(
nvjpegHandle_t              handle, // 句柄
const unsigned char         *data,	// 文件数据
size_t                      length, // 文件长度
int                         *nComponents, // 通道
nvjpegChromaSubsampling_t   *subsampling, // 色度的二次采样信息
int                         *widths, // 宽
int                         *heights); // 高

调用案例如下：

	int channels = 0;
	int width = 0;
	int height = 0;
	nvjpegChromaSubsampling_t subsampling;
	CHECK_NVJPEG(nvjpegGetImageInfo(nvjpegHandle, (const unsigned char*)pchData, file_size, &channels, &subsampling, &width, &height));

• step4. 使用 nvjpegDecode()函数来解码单张JPEG图像，函数介绍如下：

nvjpegStatus_t nvjpegDecode(
nvjpegHandle_t          handle, // 库句柄
nvjpegJpegState_t       jpeg_handle, //状态句柄
const unsigned char     *data, // 输入的文件数据
size_t                  length, // 输入的文件数据长度
nvjpegOutputFormat_t    output_format,// 解码格式
nvjpegImage_t           *destination, // 输出解码后的buffer信息
cudaStream_t            stream); // 流

这里注意，nvjpegImage_t里的内存，需要自己管理和创建

	cudaStream_t stream;

	CHECK_CUDA(cudaStreamCreate(&stream));

	nvjpegImage_t dstImage;
	memset(&dstImage, 0, sizeof(nvjpegImage_t));
	// realloc output buffer if required
	int mul = 3;
	channels = 1;
	for (int c = 0; c < channels; c++) {
		int aw = mul * width;
		int ah = height;
		int sz = aw * ah;
		dstImage.pitch[c] = aw;
		//if (sz > dstImage.pitch[c]) {
		/*	if (dstImage.channel[c]) {
				CHECK_CUDA(cudaFree(dstImage.channel[c]));
			}*/
			CHECK_CUDA(cudaMalloc((void**)&dstImage.channel[c], sz));
			//dstImage.pitch[c] = sz;
		//}
	}

	nvjpegOutputFormat_t outFormat = nvjpegOutputFormat_t::NVJPEG_OUTPUT_BGRI;
	CHECK_NVJPEG(nvjpegDecode(nvjpegHandle, nvjpegJpegState, (const unsigned char*)pchData, file_size,
		outFormat, &dstImage, stream));
	//CHECK_CUDA(cudaStreamSynchronize(stream));
	int sz = height * width * sizeof(unsigned char);
	unsigned char* pvB = (unsigned char*)malloc(sz *3);
	/*unsigned char* pvG = (unsigned char*)malloc(sz);
	unsigned char* pvR = (unsigned char*)malloc(sz);*/
	CHECK_CUDA(cudaMemcpy(pvB, dstImage.channel[0], sz*3, cudaMemcpyDeviceToHost));
	//CHECK_CUDA(cudaMemcpy(pvG, dstImage.channel[1], sz, cudaMemcpyDeviceToHost));
	//CHECK_CUDA(cudaMemcpy(pvR, dstImage.channel[2], sz, cudaMemcpyDeviceToHost));
	cv::Mat B = cv::Mat(cv::Size(width, height), CV_8UC3, pvB);
	//cv::Mat G = cv::Mat(cv::Size(width, height), CV_8UC1, pvG);
	//cv::Mat R = cv::Mat(cv::Size(width, height), CV_8UC1, pvR);
	cv::imwrite("H:/temp/rgb.bmp", B);

三、完整代码

注意，没有考虑现存的释放等操作，仅参考用

#include <iostream>
#include <fstream>
#include "opencv2/opencv.hpp"
#include "nvjpegDecoder.h"
int main()
{
	nvjpegHandle* nvjpegHandle = nullptr;
	CHECK_NVJPEG(nvjpegCreateSimple(&nvjpegHandle));

	nvjpegJpegState* nvjpegJpegState = nullptr;
	CHECK_NVJPEG(nvjpegJpegStateCreate(nvjpegHandle, &nvjpegJpegState));

	std::string strFileName = "H:/github/CUDALibrarySamples/nvJPEG/nvJPEG-Decoder/input_images/img2.jpg";
	// Read an image from disk.
	std::ifstream input(strFileName.c_str(),
		std::ios::in | std::ios::binary | std::ios::ate);
	if (!(input.is_open())) {
		std::cerr << "Cannot open image: " << strFileName
			<< ", removing it from image list" << std::endl;
	}
	// Get the size
	std::streamsize file_size = input.tellg();
	input.seekg(0, std::ios::beg);

	 char* pchData = (char*)malloc(file_size);
	if (!input.read(pchData, file_size)) {
		std::cerr << "Cannot read from file: " << strFileName
			<< ", removing it from image list" << std::endl;

	}

	int channels = 0;
	int width = 0;
	int height = 0;
	nvjpegChromaSubsampling_t subsampling;
	CHECK_NVJPEG(nvjpegGetImageInfo(nvjpegHandle, (const unsigned char*)pchData, file_size, &channels, &subsampling, &width, &height));
	cudaStream_t stream;

	CHECK_CUDA(cudaStreamCreate(&stream));

	nvjpegImage_t dstImage;
	memset(&dstImage, 0, sizeof(nvjpegImage_t));
	// realloc output buffer if required
	int mul = 3;
	channels = 1;
	for (int c = 0; c < channels; c++) {
		int aw = mul * width;
		int ah = height;
		int sz = aw * ah;
		dstImage.pitch[c] = aw;
		//if (sz > dstImage.pitch[c]) {
		/*	if (dstImage.channel[c]) {
				CHECK_CUDA(cudaFree(dstImage.channel[c]));
			}*/
			CHECK_CUDA(cudaMalloc((void**)&dstImage.channel[c], sz));
			//dstImage.pitch[c] = sz;
		//}
	}

	nvjpegOutputFormat_t outFormat = nvjpegOutputFormat_t::NVJPEG_OUTPUT_BGRI;
	CHECK_NVJPEG(nvjpegDecode(nvjpegHandle, nvjpegJpegState, (const unsigned char*)pchData, file_size,
		outFormat, &dstImage, stream));
	//CHECK_CUDA(cudaStreamSynchronize(stream));
	int sz = height * width * sizeof(unsigned char);
	unsigned char* pvB = (unsigned char*)malloc(sz *3);
	/*unsigned char* pvG = (unsigned char*)malloc(sz);
	unsigned char* pvR = (unsigned char*)malloc(sz);*/
	CHECK_CUDA(cudaMemcpy(pvB, dstImage.channel[0], sz*3, cudaMemcpyDeviceToHost));
	//CHECK_CUDA(cudaMemcpy(pvG, dstImage.channel[1], sz, cudaMemcpyDeviceToHost));
	//CHECK_CUDA(cudaMemcpy(pvR, dstImage.channel[2], sz, cudaMemcpyDeviceToHost));
	cv::Mat B = cv::Mat(cv::Size(width, height), CV_8UC3, pvB);
	//cv::Mat G = cv::Mat(cv::Size(width, height), CV_8UC1, pvG);
	//cv::Mat R = cv::Mat(cv::Size(width, height), CV_8UC1, pvR);
	cv::imwrite("H:/temp/rgb.bmp", B);
	//cv::imwrite("H:/temp/g.bmp", G);
	//cv::imwrite("H:/temp/b.bmp", B);

	printf("hello world！");
	return 0;
}