英伟达CUVID硬解，并通过FFmpeg读取文件
虽然FFmpeg本身有cuvid硬解，但是找不到什么好的资料，英伟达的SDK比较容易懂，参考FFmpeg源码，将NVIDIA VIDEO CODEC SDK的数据获取改为FFmpeg获取，弥补原生SDK不能以流作为数据源的不足。所用SDK版本为Video_Codec_SDK_7.1.9，英伟达官网可下载。
1.修改数据源
首先是FFmpeg的一些常规的初始化
bool VideoSource::init(const std::string sFileName, FrameQueue *pFrameQueue) { assert(0 != pFrameQueue); oSourceData_.hVideoParser = 0; oSourceData_.pFrameQueue = pFrameQueue; int i; AVCodec *pCodec; av_register_all(); avformat_network_init(); pFormatCtx = avformat_alloc_context(); if (avformat_open_input(&pFormatCtx, sFileName.c_str(), NULL, NULL) != 0){ printf("Couldn't open input stream.\n"); return false; } if (avformat_find_stream_info(pFormatCtx, NULL)<0){ printf("Couldn't find stream information.\n"); return false; } videoindex = -1; for (i = 0; i<pFormatCtx->nb_streams; i++) if (pFormatCtx->streams[i]->codec->codec_type == AVMEDIA_TYPE_VIDEO){ videoindex = i; break; } if (videoindex == -1){ printf("Didn't find a video stream.\n"); return false; } pCodecCtx = pFormatCtx->streams[videoindex]->codec; pCodec = avcodec_find_decoder(pCodecCtx->codec_id); if (pCodec == NULL){ printf("Codec not found.\n"); return false; } //Output Info----------------------------- printf("--------------- File Information ----------------\n"); av_dump_format(pFormatCtx, 0, sFileName.c_str(), 0); printf("-------------------------------------------------\n"); memset(&g_stFormat, 0, sizeof(CUVIDEOFORMAT)); switch (pCodecCtx->codec_id) { case AV_CODEC_ID_H263: g_stFormat.codec = cudaVideoCodec_MPEG4; break; case AV_CODEC_ID_H264: g_stFormat.codec = cudaVideoCodec_H264; break; case AV_CODEC_ID_HEVC: g_stFormat.codec = cudaVideoCodec_HEVC; break; case AV_CODEC_ID_MJPEG: g_stFormat.codec = cudaVideoCodec_JPEG; break; case AV_CODEC_ID_MPEG1VIDEO: g_stFormat.codec = cudaVideoCodec_MPEG1; break; case AV_CODEC_ID_MPEG2VIDEO: g_stFormat.codec = cudaVideoCodec_MPEG2; break; case AV_CODEC_ID_MPEG4: g_stFormat.codec = cudaVideoCodec_MPEG4; break; case AV_CODEC_ID_VP8: g_stFormat.codec = cudaVideoCodec_VP8; break; case AV_CODEC_ID_VP9: g_stFormat.codec = cudaVideoCodec_VP9; break; case AV_CODEC_ID_VC1: g_stFormat.codec = cudaVideoCodec_VC1; break; default: return false; } //这个地方的FFmoeg与cuvid的对应关系不是很确定，不过用这个参数似乎最靠谱 switch (pCodecCtx->sw_pix_fmt) { case AV_PIX_FMT_YUV420P: g_stFormat.chroma_format = cudaVideoChromaFormat_420; break; case AV_PIX_FMT_YUV422P: g_stFormat.chroma_format = cudaVideoChromaFormat_422; break; case AV_PIX_FMT_YUV444P: g_stFormat.chroma_format = cudaVideoChromaFormat_444; break; default: g_stFormat.chroma_format = cudaVideoChromaFormat_420; break; } //找了好久，总算是找到了FFmpeg中标识场格式和帧格式的标识位 //场格式是隔行扫描的，需要做去隔行处理 switch (pCodecCtx->field_order) { case AV_FIELD_PROGRESSIVE: case AV_FIELD_UNKNOWN: g_stFormat.progressive_sequence = true; break; default: g_stFormat.progressive_sequence = false; break; } pCodecCtx->thread_safe_callbacks = 1; g_stFormat.coded_width = pCodecCtx->coded_width; g_stFormat.coded_height = pCodecCtx->coded_height; g_stFormat.display_area.right = pCodecCtx->width; g_stFormat.display_area.left = 0; g_stFormat.display_area.bottom = pCodecCtx->height; g_stFormat.display_area.top = 0; if (pCodecCtx->codec_id == AV_CODEC_ID_H264 || pCodecCtx->codec_id == AV_CODEC_ID_HEVC) { if (pCodecCtx->codec_id == AV_CODEC_ID_H264) h264bsfc = av_bitstream_filter_init("h264_mp4toannexb"); else h264bsfc = av_bitstream_filter_init("hevc_mp4toannexb"); } return true; }
这里面非常重要的一段代码是
if (pCodecCtx->codec_id == AV_CODEC_ID_H264 || pCodecCtx->codec_id == AV_CODEC_ID_HEVC) { if (pCodecCtx->codec_id == AV_CODEC_ID_H264) h264bsfc = av_bitstream_filter_init("h264_mp4toannexb"); else h264bsfc = av_bitstream_filter_init("hevc_mp4toannexb"); }
网上有许多代码和伪代码都说实现了把数据源修改为FFmpeg，但我在尝试的时候发现cuvidCreateVideoParser创建的Parser的回调函数都没有调用。经过一番折腾，综合英伟达网站、stackoverflow和FFmpeg源码，才发现对H264数据要做一个处理才能把AVPacket有效的转为CUVIDSOURCEDATAPACKET。其中h264bsfc的定义为AVBitStreamFilterContext* h264bsfc = NULL;
2.AVPacket转CUVIDSOURCEDATAPACKET,并交给cuvidParseVideoData
void VideoSource::play_thread(LPVOID lpParam) { AVPacket *avpkt; avpkt = (AVPacket *)av_malloc(sizeof(AVPacket)); CUVIDSOURCEDATAPACKET cupkt; int iPkt = 0; CUresult oResult; while (av_read_frame(pFormatCtx, avpkt) >= 0){ if (bThreadExit){ break; } bStarted = true; if (avpkt->stream_index == videoindex){ cuCtxPushCurrent(g_oContext); if (avpkt && avpkt->size) { if (h264bsfc) { av_bitstream_filter_filter(h264bsfc, pFormatCtx->streams[videoindex]->codec, NULL, &avpkt->data, &avpkt->size, avpkt->data, avpkt->size, 0); } cupkt.payload_size = (unsigned long)avpkt->size; cupkt.payload = (const unsigned char*)avpkt->data; if (avpkt->pts != AV_NOPTS_VALUE) { cupkt.flags = CUVID_PKT_TIMESTAMP; if (pCodecCtx->pkt_timebase.num && pCodecCtx->pkt_timebase.den){ AVRational tb; tb.num = 1; tb.den = AV_TIME_BASE; cupkt.timestamp = av_rescale_q(avpkt->pts, pCodecCtx->pkt_timebase, tb); } else cupkt.timestamp = avpkt->pts; } } else { cupkt.flags = CUVID_PKT_ENDOFSTREAM; } oResult = cuvidParseVideoData(oSourceData_.hVideoParser, &cupkt); if ((cupkt.flags & CUVID_PKT_ENDOFSTREAM) || (oResult != CUDA_SUCCESS)){ break; } iPkt++; //printf("Succeed to read avpkt %d !\n", iPkt); checkCudaErrors(cuCtxPopCurrent(NULL)); } av_free_packet(avpkt); } oSourceData_.pFrameQueue->endDecode(); bStarted = false; }
这里FFmpeg读取数据包后，对H264和HEVC格式，有一个重要的处理，就是前面提到的，
if (h264bsfc) { av_bitstream_filter_filter(h264bsfc, pFormatCtx->streams[videoindex]->codec, NULL, &avpkt->data, &avpkt->size, avpkt->data, avpkt->size, 0); }
这个处理的含义见雷霄华的博客http://blog.csdn.net/leixiaohua1020/article/details/39767055。
这样，通过FFmpeg，CUVID就可以对流进行处理了。个人尝试过读取本地文件和rtsp流。FFmpeg读取rtsp流的方式竟然只需要把文件改为rtsp流的地址就可以，以前没做过流式的，我还以为会很复杂的。
3.一点数据
这是在GTX 1080上把解码进程（没做显示）开了20路解码得到的数据。20路1920X1080解码还能到平局37fps，这显卡也是6得不行。
工程源码：http://download.csdn.net/download/qq_33892166/9792997
源码遇到了一个问题，没找到原因。代码在GTX 1080和Tesla P4上的解码效果很好。P4由于驱动模式是TCC模式，所以只能解码，不能显示；1080上可解码，可显示。但是在我自己电脑上的GT940M上，即时是原生SDK在cuvidCreateDecoder的时候也总是报错CUDA_ERROR_NO_DEVICE。驱动似乎没问题，试了CUDA的demo,CUDA运算也是正常的，查的资料表明GT940M应该是支持CUVID的。希望知道原因的朋友能指教一二。
-----------------------------------------2017.7.7更新----------------------------------------
修改代码中的一处内存泄漏问题：
把play_thread()中的
if (avpkt && avpkt->size) { if (h264bsfc) { av_bitstream_filter_filter(h264bsfc, pFormatCtx->streams[videoindex]->codec, NULL, &avpkt->data, &avpkt->size, avpkt->data, avpkt->size, 0); } cupkt.payload_size = (unsigned long)avpkt->size; cupkt.payload = (const unsigned char*)avpkt->data; if (avpkt->pts != AV_NOPTS_VALUE) { cupkt.flags = CUVID_PKT_TIMESTAMP; if (pCodecCtx->pkt_timebase.num && pCodecCtx->pkt_timebase.den){ AVRational tb; tb.num = 1; tb.den = AV_TIME_BASE; cupkt.timestamp = av_rescale_q(avpkt->pts, pCodecCtx->pkt_timebase, tb); } else cupkt.timestamp = avpkt->pts; } } else { cupkt.flags = CUVID_PKT_ENDOFSTREAM; } oResult = cuvidParseVideoData(oSourceData_.hVideoParser, &cupkt); if ((cupkt.flags & CUVID_PKT_ENDOFSTREAM) || (oResult != CUDA_SUCCESS)){ break; } iPkt++;
为
AVPacket new_pkt = *avpkt; if (avpkt && avpkt->size) { if (h264bsfc){ int a = av_bitstream_filter_filter(h264bsfc, pFormatCtx->streams[videoindex]->codec, NULL, &new_pkt.data, &new_pkt.size, avpkt->data, avpkt->size, avpkt->flags & AV_PKT_FLAG_KEY); if (a>0){ if (new_pkt.data != avpkt->data)//-added this { av_free_packet(avpkt); avpkt->data = new_pkt.data; avpkt->size = new_pkt.size; } } else if (a<0){ goto LOOP0; } *avpkt = new_pkt; } cupkt.payload_size = (unsigned long)avpkt->size; cupkt.payload = (const unsigned char*)avpkt->data; if (avpkt->pts != AV_NOPTS_VALUE) { cupkt.flags = CUVID_PKT_TIMESTAMP; if (pCodecCtx->pkt_timebase.num && pCodecCtx->pkt_timebase.den) { AVRational tb; tb.num = 1; tb.den = AV_TIME_BASE; cupkt.timestamp = av_rescale_q(avpkt->pts, pCodecCtx->pkt_timebase, tb); } else cupkt.timestamp = avpkt->pts; } } else { cupkt.flags = CUVID_PKT_ENDOFSTREAM; } oResult = cuvidParseVideoData(oSourceData_.hVideoParser, &cupkt); if ((cupkt.flags & CUVID_PKT_ENDOFSTREAM) || (oResult != CUDA_SUCCESS)) { break; } av_free(new_pkt.data);
这个泄漏是av_bitstream_filter_filter造成的，解决办法参考http://blog.csdn.net/lg1259156776/article/details/73283920 。
-----------------------------2017.8.30 补-----------------------------------
貌似还是有小伙伴被内存泄漏难住了，这里我给出我最新的读取数据包线程函数的完整代码，希望有所帮助
void VideoSource::play_thread(LPVOID lpParam) { AVPacket *avpkt; avpkt = (AVPacket *)av_malloc(sizeof(AVPacket)); CUVIDSOURCEDATAPACKET cupkt; CUresult oResult; while (av_read_frame(pFormatCtx, avpkt) >= 0){ LOOP0: if (bThreadExit){ break; } if (avpkt->stream_index == videoindex) { AVPacket new_pkt = *avpkt; if (avpkt && avpkt->size) { if (h264bsfc){ int a = av_bitstream_filter_filter(h264bsfc, pFormatCtx->streams[videoindex]->codec, NULL, &new_pkt.data, &new_pkt.size, avpkt->data, avpkt->size, avpkt->flags & AV_PKT_FLAG_KEY); if (a>0){ if (new_pkt.data != avpkt->data)//-added this { av_free_packet(avpkt); avpkt->data = new_pkt.data; avpkt->size = new_pkt.size; } } else if (a<0){ goto LOOP0; } *avpkt = new_pkt; } cupkt.payload_size = (unsigned long)avpkt->size; cupkt.payload = (const unsigned char*)avpkt->data; if (avpkt->pts != AV_NOPTS_VALUE) { cupkt.flags = CUVID_PKT_TIMESTAMP; if (pCodecCtx->pkt_timebase.num && pCodecCtx->pkt_timebase.den) { AVRational tb; tb.num = 1; tb.den = AV_TIME_BASE; cupkt.timestamp = av_rescale_q(avpkt->pts, pCodecCtx->pkt_timebase, tb); } else cupkt.timestamp = avpkt->pts; } } else { cupkt.flags = CUVID_PKT_ENDOFSTREAM; } oResult = cuvidParseVideoData(oSourceData_.hVideoParser, &cupkt); if ((cupkt.flags & CUVID_PKT_ENDOFSTREAM) || (oResult != CUDA_SUCCESS)) { break; } av_free(new_pkt.data); } else av_free_packet(avpkt); } oSourceData_.pFrameQueue->endDecode(); bStarted = false; if (pCodecCtx->codec_id == AV_CODEC_ID_H264 || pCodecCtx->codec_id == AV_CODEC_ID_HEVC) { av_bitstream_filter_close(h264bsfc); } }
