libavcodec/libx264 do not produce B-frames

Question

I am writing an application in C++ that uses libavcodec with libx264 to encode video. However, the encoded data ended up being much larger than I expected. I analyzed the results and discovered that my encoding never produced B-frames, only I- and P-frames.

I created a standalone utility based on the ffmpeg source code and examples to test my encoder setup. It reads in an H.264 file, re-encodes the decoded frames, and outputs the result to a file using the ITU H.264 Annex B format. I also used ffmpeg to perform the same operation so I could compare against a known good implementation. My utility never outputs B-frames whereas ffmpeg does.

I have since tried to figure out what ffmpeg does that my code doesn't. I first tried manually specifying encoder settings related to B-frames. This had no effect.

I then tried running both ffmpeg and my utility under gdb and comparing the contents of the AVStream, AVCodecContext, and X264Context prior to opening the encoder and manually setting any fields that appeared different. Even with identical settings, I still only produce I- and P-frames.

Finally, I thought that perhaps the problem was with my timestamp handling. I reworked my test utility to mimic the pipeline used by ffmpeg and to output timestamp debugging output like ffmpeg does. Even with my timestamps identical to ffmpeg's I still get no B-frames.

At this point I don't know what else to try. When I run ffmpeg, I run it with the command line below. Note that aside from the "superfast" preset, I pretty much use the default values.

ffmpeg -v debug -i ~/annexb.264 -codec:v libx264 -preset superfast -g 30 -f h264 ./out.264

The code that configures the encoder is listed below. It specifies the "superfast" preset too.

static AVStream *add_video_stream(AVFormatContext *output_ctx, AVCodec **output_codec, enum AVCodecID codec_id)
{
    *output_codec = avcodec_find_encoder(codec_id);
    if (*output_codec == NULL) {
        printf("Could not find encoder for '%s' (%d)\n", avcodec_get_name(codec_id), codec_id);
        return NULL;
    }

    AVStream *output_stream = avformat_new_stream(output_ctx, *output_codec);
    if (output_stream == NULL) {
        printf("Could not create video stream.\n");
        return NULL;
    }
    output_stream->id = output_ctx->nb_streams - 1;
    AVCodecContext *codec_ctx = output_stream->codec;

    avcodec_get_context_defaults3(codec_ctx, *output_codec);

    codec_ctx->width = 1280;
    codec_ctx->height = 720;

    codec_ctx->time_base.den = 15000;
    codec_ctx->time_base.num = 1001;

/*    codec_ctx->gop_size = 30;*/
    codec_ctx->pix_fmt = AV_PIX_FMT_YUVJ420P;

    // try to force B-frame output
/*    codec_ctx->max_b_frames = 3;*/
/*    codec_ctx->b_frame_strategy = 2;*/

    output_stream->sample_aspect_ratio.num = 1;
    output_stream->sample_aspect_ratio.den = 1;

    codec_ctx->sample_aspect_ratio.num = 1;
    codec_ctx->sample_aspect_ratio.den = 1;

    codec_ctx->chroma_sample_location = AVCHROMA_LOC_LEFT;

    codec_ctx->bits_per_raw_sample = 8;

    if ((output_ctx->oformat->flags & AVFMT_GLOBALHEADER) != 0) {
        codec_ctx->flags |= CODEC_FLAG_GLOBAL_HEADER;
    }

    return output_stream;
}


int main(int argc, char **argv)
{
    // ... open input file 

    avformat_alloc_output_context2(&output_ctx, NULL, "h264", output_path);
    if (output_ctx == NULL) {
        fprintf(stderr, "Unable to allocate output context.\n");
        return 1;
    } 

    AVCodec *output_codec = NULL;
    output_stream = add_video_stream(output_ctx, &output_codec, output_ctx->oformat->video_codec);
    if (output_stream == NULL) {
        fprintf(stderr, "Error adding video stream to output context.\n");
        return 1;
    }
    encode_ctx = output_stream->codec;

    // seems to have no effect
#if 0
    if (decode_ctx->extradata_size != 0) {
        size_t extradata_size = decode_ctx->extradata_size;
        printf("extradata_size: %zu\n", extradata_size);
        encode_ctx->extradata = av_mallocz(extradata_size + FF_INPUT_BUFFER_PADDING_SIZE);
        memcpy(encode_ctx->extradata, decode_ctx->extradata, extradata_size);
        encode_ctx->extradata_size = extradata_size;
    }
#endif // 0

    AVDictionary *opts = NULL;
    av_dict_set(&opts, "preset", "superfast", 0);
    // av_dict_set(&opts, "threads", "auto", 0); // seems to have no effect

    ret = avcodec_open2(encode_ctx, output_codec, &opts);
    if (ret < 0) {
        fprintf(stderr, "Unable to open output video cocec: %s\n", av_err2str(ret));
        return 1;
    }

    // ... decoding/encoding loop, clean up, etc.

    return 0;
}

My test utility produces the following debug output in which you can see there are no B-frames produced:

[libx264 @ 0x1b8c9c0] using mv_range_thread = 56
[libx264 @ 0x1b8c9c0] using SAR=1/1
[libx264 @ 0x1b8c9c0] using cpu capabilities: MMX2 SSE2Fast SSSE3 SSE4.2 AVX
[libx264 @ 0x1b8c9c0] profile High, level 3.1
Output #0, h264, to './out.264':
    Stream #0:0, 0, 1/90000: Video: h264, yuvj420p, 1280x720 [SAR 1:1 DAR 16:9], 1001/15000, q=-1--1, 90k tbn, 14.99 tbc

<SNIP>

[libx264 @ 0x1b8c9c0] frame=   0 QP=17.22 NAL=3 Slice:I Poc:0   I:3600 P:0    SKIP:0    size=122837 bytes
[libx264 @ 0x1b8c9c0] frame=   1 QP=18.03 NAL=2 Slice:P Poc:2   I:411  P:1825 SKIP:1364 size=25863 bytes
[libx264 @ 0x1b8c9c0] frame=   2 QP=17.03 NAL=2 Slice:P Poc:4   I:369  P:2159 SKIP:1072 size=37880 bytes
[libx264 @ 0x1b8c9c0] frame=   3 QP=16.90 NAL=2 Slice:P Poc:6   I:498  P:2330 SKIP:772  size=50509 bytes
[libx264 @ 0x1b8c9c0] frame=   4 QP=16.68 NAL=2 Slice:P Poc:8   I:504  P:2233 SKIP:863  size=50791 bytes
[libx264 @ 0x1b8c9c0] frame=   5 QP=16.52 NAL=2 Slice:P Poc:10  I:513  P:2286 SKIP:801  size=51820 bytes
[libx264 @ 0x1b8c9c0] frame=   6 QP=16.49 NAL=2 Slice:P Poc:12  I:461  P:2293 SKIP:846  size=51311 bytes
[libx264 @ 0x1b8c9c0] frame=   7 QP=16.65 NAL=2 Slice:P Poc:14  I:476  P:2287 SKIP:837  size=51196 bytes
[libx264 @ 0x1b8c9c0] frame=   8 QP=16.66 NAL=2 Slice:P Poc:16  I:508  P:2240 SKIP:852  size=51577 bytes
[libx264 @ 0x1b8c9c0] frame=   9 QP=16.55 NAL=2 Slice:P Poc:18  I:477  P:2278 SKIP:845  size=51531 bytes
[libx264 @ 0x1b8c9c0] frame=  10 QP=16.67 NAL=2 Slice:P Poc:20  I:517  P:2233 SKIP:850  size=51946 bytes

<SNIP>

[libx264 @ 0x1b8c9c0] frame I:7     Avg QP:13.71  size:152207
[libx264 @ 0x1b8c9c0] frame P:190   Avg QP:16.66  size: 50949
[libx264 @ 0x1b8c9c0] mb I  I16..4: 27.1% 30.8% 42.1%
[libx264 @ 0x1b8c9c0] mb P  I16..4:  6.8%  6.0%  0.8%  P16..4: 61.8%  0.0%  0.0%  0.0%  0.0%    skip:24.7%
[libx264 @ 0x1b8c9c0] 8x8 transform intra:41.2% inter:86.9%
[libx264 @ 0x1b8c9c0] coded y,uvDC,uvAC intra: 92.2% 28.3% 5.4% inter: 50.3% 1.9% 0.0%
[libx264 @ 0x1b8c9c0] i16 v,h,dc,p:  7%  7% 77%  8%
[libx264 @ 0x1b8c9c0] i8 v,h,dc,ddl,ddr,vr,hd,vl,hu:  7% 15% 49%  6%  4%  3%  5%  3%  8%
[libx264 @ 0x1b8c9c0] i4 v,h,dc,ddl,ddr,vr,hd,vl,hu: 19% 25% 24%  6%  7%  4%  6%  3%  6%
[libx264 @ 0x1b8c9c0] i8c dc,h,v,p: 72% 14% 10%  4%
[libx264 @ 0x1b8c9c0] Weighted P-Frames: Y:0.0% UV:0.0%
[libx264 @ 0x1b8c9c0] kb/s:6539.11

ffmpeg, on the other hand, produces the following output that is almost identical but includes B-frames:

[libx264 @ 0x20b9c40] using mv_range_thread = 56
[libx264 @ 0x20b9c40] using SAR=1/1
[libx264 @ 0x20b9c40] using cpu capabilities: MMX2 SSE2Fast SSSE3 SSE4.2 AVX
[libx264 @ 0x20b9c40] profile High, level 3.1
[h264 @ 0x20b8160] detected 4 logical cores
Output #0, h264, to './out.264':
  Metadata:
    encoder         : Lavf54.63.104
    Stream #0:0, 0, 1/90000: Video: h264, yuvj420p, 1280x720 [SAR 1:1 DAR 16:9], 1001/15000, q=-1--1, 90k tbn, 14.99 tbc
Stream mapping:
  Stream #0:0 -> #0:0 (h264 -> libx264)

<SNIP>

[libx264 @ 0x20b9c40] frame=   0 QP=17.22 NAL=3 Slice:I Poc:0   I:3600 P:0    SKIP:0    size=122835 bytes
[libx264 @ 0x20b9c40] frame=   1 QP=18.75 NAL=2 Slice:P Poc:8   I:984  P:2045 SKIP:571  size=54208 bytes
[libx264 @ 0x20b9c40] frame=   2 QP=19.40 NAL=2 Slice:B Poc:4   I:447  P:1581 SKIP:1572 size=24930 bytes
[libx264 @ 0x20b9c40] frame=   3 QP=19.78 NAL=0 Slice:B Poc:2   I:199  P:1002 SKIP:2399 size=10717 bytes
[libx264 @ 0x20b9c40] frame=   4 QP=20.19 NAL=0 Slice:B Poc:6   I:204  P:1155 SKIP:2241 size=15937 bytes
[libx264 @ 0x20b9c40] frame=   5 QP=18.11 NAL=2 Slice:P Poc:16  I:990  P:2221 SKIP:389  size=64240 bytes
[libx264 @ 0x20b9c40] frame=   6 QP=19.35 NAL=2 Slice:B Poc:12  I:439  P:1784 SKIP:1377 size=34048 bytes
[libx264 @ 0x20b9c40] frame=   7 QP=19.88 NAL=0 Slice:B Poc:10  I:275  P:1035 SKIP:2290 size=16911 bytes
[libx264 @ 0x20b9c40] frame=   8 QP=19.91 NAL=0 Slice:B Poc:14  I:257  P:1270 SKIP:2073 size=19172 bytes
[libx264 @ 0x20b9c40] frame=   9 QP=17.90 NAL=2 Slice:P Poc:24  I:962  P:2204 SKIP:434  size=67439 bytes
[libx264 @ 0x20b9c40] frame=  10 QP=18.84 NAL=2 Slice:B Poc:20  I:474  P:1911 SKIP:1215 size=37742 bytes

<SNIP>

[libx264 @ 0x20b9c40] frame I:7     Avg QP:15.95  size:130124
[libx264 @ 0x20b9c40] frame P:52    Avg QP:17.78  size: 64787
[libx264 @ 0x20b9c40] frame B:138   Avg QP:19.32  size: 26231
[libx264 @ 0x20b9c40] consecutive B-frames:  6.6%  0.0%  0.0% 93.4%
[libx264 @ 0x20b9c40] mb I  I16..4: 30.2% 35.2% 34.6%
[libx264 @ 0x20b9c40] mb P  I16..4: 13.9% 11.4%  0.3%  P16..4: 60.4%  0.0%  0.0%  0.0%  0.0%    skip:13.9%
[libx264 @ 0x20b9c40] mb B  I16..4:  5.7%  3.3%  0.0%  B16..8: 15.8%  0.0%  0.0%  direct:25.7%  skip:49.5%  L0:43.2% L1:37.3% BI:19.5%
[libx264 @ 0x20b9c40] 8x8 transform intra:39.4% inter:77.2%
[libx264 @ 0x20b9c40] coded y,uvDC,uvAC intra: 90.7% 26.6% 3.0% inter: 34.0% 4.1% 0.0%
[libx264 @ 0x20b9c40] i16 v,h,dc,p:  7%  7% 77%  9%
[libx264 @ 0x20b9c40] i8 v,h,dc,ddl,ddr,vr,hd,vl,hu:  7% 16% 51%  5%  4%  3%  5%  3%  7%
[libx264 @ 0x20b9c40] i4 v,h,dc,ddl,ddr,vr,hd,vl,hu: 22% 27% 20%  6%  6%  3%  6%  3%  6%
[libx264 @ 0x20b9c40] i8c dc,h,v,p: 71% 15% 11%  3%
[libx264 @ 0x20b9c40] Weighted P-Frames: Y:0.0% UV:0.0%
[libx264 @ 0x20b9c40] kb/s:4807.16

I'm sure I'm missing something simple, but I can't for the life of me see what it is. Any assistance would be greatly appreciated.

Answer 1

After much debugging, I finally found the solution. The issue was not in the stream and encoder setup listed above, but in how I was passing frames to the encoder.

Since I was just trying to transcode the data, I was simply taking pictures that had been emitted by the decoder and passing them to the encoder. However, the decoder sets the pict_type value for the AVFrame structures it produces to AV_PICTURE_TYPE_P. The encoder, in turn, uses the pict_type value to determine the type of frames to generate. Hence, the encoder only ever produces P-frames since that is what is always specified in the input.

Resetting the pict_type field to AV_PICTURE_TYPE_NONE allows the encoder to produce whatever type of frame it thinks is best. Once I updated my code to do this, my test utility generated B-frames and produced a file identical to the one output by ffmpeg itself.