视频多线程解码

最近在做视频解码相关的工作，有一个功能需要将整个视频解码之后放到内存里面，经测试，一分钟的视频解码需要20s，不算太理想，考虑用多线程来实现。

基本思路是获取所有关键帧信息，然后开启不同的线程来从不同的关键帧开始解码。

1，获取所有关键帧信息，获取所有的关键帧时间戳，大约会花费0.2s：

+ (NSMutableArray *)keyFramePtsWithM3u8Path:(NSString *)path{
 NSMutableArray *timestamp_array = [[NSMutableArray alloc] init];
 canceled = false;
 AVFormatContext *qFormatCtx = NULL;
 AVCodecContext *qCodeCtx = NULL;
 AVCodec *qCodec = NULL;
 AVPacket qPacket;
 struct SwsContext *qSwsCtx;
 const char *filePath = [path UTF8String];
 if (filePath == nil) {
 return nil;
 }
 av_register_all();
 if (avformat_open_input(&qFormatCtx, filePath, NULL, NULL) != 0) {
 return nil;
 }
 if (avformat_find_stream_info(qFormatCtx, NULL) < 0) {
 return nil;
 }
 av_dump_format(qFormatCtx, 0, filePath, 0);
 int videoStream = -1;
 for (int i = 0;i < qFormatCtx->nb_streams;i++) {
 if (qFormatCtx->streams[i]->codecpar->codec_type == AVMEDIA_TYPE_VIDEO) {
 videoStream = i;
 break;
 }
 }
 if (videoStream == -1) {
 return nil;
 }
 qCodeCtx = avcodec_alloc_context3(NULL);
 if (avcodec_parameters_to_context(qCodeCtx, qFormatCtx->streams[videoStream]->codecpar) < 0) {
 return nil;
 }
 qCodec = avcodec_find_decoder(qCodeCtx->codec_id);
 if (qCodeCtx == NULL) {
 return nil;
 }
 if (avcodec_open2(qCodeCtx, qCodec, NULL) < 0) {
 return nil;
 }
 qSwsCtx = sws_getContext(qCodeCtx->width,
 qCodeCtx->height,
 qCodeCtx->pix_fmt,
 qCodeCtx->width,
 qCodeCtx->height,
 AV_PIX_FMT_RGBA,
 SWS_BICUBIC, NULL, NULL, NULL);
 while (av_read_frame(qFormatCtx, &qPacket) >= 0) {
 if (qPacket.stream_index == videoStream) {
 if(canceled){
 av_packet_unref(&qPacket);
 break;
 }
 if(qPacket.flags==1){
 [timestamp_array addObject:[NSString stringWithFormat:@"%lld",qPacket.pts]];
 }
 }
 av_packet_unref(&qPacket);
 }
 sws_freeContext(qSwsCtx);
 avcodec_close(qCodeCtx);
 avformat_close_input(&qFormatCtx);
 return timestamp_array;
}

2，根据关键帧信息开启多个队列（这里为每5个关键帧开启一个队列，可以适当调整）：

- (void)parseVideoToImagesMultiThreadWithM3u8Path:(NSString *)path andError:(NSError *__autoreleasing *)error{
 NSMutableArray *timestamps = [Decoder keyFramePtsWithM3u8Path:path andError:nil];
 int i=0;
 int gap = 5;//每五个关键帧开启一个队列
 __block long thread_count = ceil([timestamps count]/(float)gap);
 __block long complete_count = 0;
 __block NSMutableDictionary *imgs = [[NSMutableDictionary alloc] init];
 UInt64 t = [[NSDate date] timeIntervalSince1970]* 1000;
 for(NSString *time in timestamps){
 if(i%gap==0){
 NSString *start = time;
 NSString *end = (i+gap)<=([timestamps count]-1)?timestamps[i+gap]:@"1000000000";
 dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
 [Decoder decodeWithM3u8Path:path andStart:start andEnd:end andBlock:^(unsigned char *buffer, int width, int height, NSString *timestamp) {
 const size_t bufferLength = width * height * 3;
 NSData *data = [NSData dataWithBytes:buffer length:bufferLength];
 [imgs setObject:data forKey:timestamp];
 } andError:nil];
 complete_count++;
 if(complete_count==thread_count){
 if ([self.delegate respondsToSelector:@selector(finished:)]) {
 [self.delegate finished:imgs];
 }
 NSLog(@"Cutter---:%fms",[[NSDate date] timeIntervalSince1970]* 1000 - t);
 }
 });
 }
 i++;
 }
}

3，那么我们就需要一个有开始和结束时间戳的方法来实现区间解码：

+ (void)decodeWithM3u8Path:(NSString *)path andStart:(NSString *)start andEnd:(NSString *)end andBlock:(void (^)(unsigned char *, int, int, NSString *))block andError:(NSError *__autoreleasing *)error{
 canceled = false;
 AVFormatContext *qFormatCtx = NULL;
 AVCodecContext *qCodeCtx = NULL;
 AVCodec *qCodec = NULL;
 AVPacket qPacket;
 AVFrame *qFrame = NULL;
 AVFrame *qFrameRGB = NULL;
 struct SwsContext *qSwsCtx;
 uint8_t *buffer = NULL;
 int64_t start_timestamp = [start longLongValue];
 int64_t end_timestamp = [end longLongValue];
 const char *filePath = [path UTF8String];
 if (filePath == nil) {
 return;
 }
 av_register_all();
 if (avformat_open_input(&qFormatCtx, filePath, NULL, NULL) != 0) {
 return;
 }
 if (avformat_find_stream_info(qFormatCtx, NULL) < 0) {
 return;
 }
 av_dump_format(qFormatCtx, 0, filePath, 0);
 int videoStream = -1;
 for (int i = 0;i < qFormatCtx->nb_streams;i++) {
 if (qFormatCtx->streams[i]->codecpar->codec_type == AVMEDIA_TYPE_VIDEO) {
 videoStream = i;
 break;
 }
 }
 if (videoStream == -1) {
 return;
 }
 qCodeCtx = avcodec_alloc_context3(NULL);
 if (avcodec_parameters_to_context(qCodeCtx, qFormatCtx->streams[videoStream]->codecpar) < 0) {
 return;
 }
 qCodec = avcodec_find_decoder(qCodeCtx->codec_id);
 if (qCodeCtx == NULL) {
 return;
 }
 if (avcodec_open2(qCodeCtx, qCodec, NULL) < 0) {
 return;
 }
 qFrame = av_frame_alloc();
 qFrameRGB = av_frame_alloc();
 if (qFrame == NULL || qFrameRGB == NULL) {
 return;
 }
 int heigth = qCodeCtx->height>>2;
 int width = qCodeCtx->width>>2;
 int numBytes = av_image_get_buffer_size(AV_PIX_FMT_RGB24,
 width,
 heigth, 1);
 buffer = (uint8_t *)av_malloc(numBytes * sizeof(uint8_t));
 av_image_fill_arrays(qFrameRGB->data,
 qFrameRGB->linesize,
 buffer,
 AV_PIX_FMT_RGB24,
 width,
 heigth, 1);
 qSwsCtx = sws_getContext(qCodeCtx->width,
 qCodeCtx->height,
 qCodeCtx->pix_fmt,
 width,
 heigth,
 AV_PIX_FMT_RGB24,
 SWS_FAST_BILINEAR, NULL, NULL, NULL);
 while (av_read_frame(qFormatCtx, &qPacket) >= 0) {
 if (qPacket.stream_index == videoStream) {
 if(canceled){
 av_packet_unref(&qPacket);
 break;
 }
 if(qPacket.pts>=start_timestamp){
 if(qPacket.pts<end_timestamp){
 int ret = avcodec_send_packet(qCodeCtx, &qPacket);
 if (ret < 0) {
 return;
 }
 ret = avcodec_receive_frame(qCodeCtx, qFrame);
 if (ret ==0) {
 sws_scale(qSwsCtx,
 (uint8_t const * const *)qFrame->data,
 qFrame->linesize, 0,
 qCodeCtx->height,
 qFrameRGB->data,
 qFrameRGB->linesize);
 block(buffer,width,heigth,[NSString stringWithFormat:@"%lld",qFrame->pts]);
 }
 }
 }
 }
 av_packet_unref(&qPacket);
 }
 av_free(buffer);
 av_frame_free(&qFrameRGB);
 av_frame_free(&qFrame);
 sws_freeContext(qSwsCtx);
 avcodec_close(qCodeCtx);
 avformat_close_input(&qFormatCtx);
}

这样就可以实现多线程的解码了，由于机器为双核，经测最后多线程解码的时间缩减为11s，还是很可观的。鉴于大部分视频长度都在30s左右，解码时间可以控制在五六秒的范围能，对于用户体验有不少的提升。

或者直接：av_dict_set(&opt, “threads”, “auto”, 0);

也可以实现多线程解码，当然这个是ffmpeg内部实现的。

文章首发于我的博客(yangf.vip)