[FFmpeg-devel] [PATCH] avcodec/libx264: fix extradata when config annexb=0
Zhao Zhili
quinkblack at foxmail.com
Wed Mar 6 18:46:39 EET 2024
> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Andreas Rheinhardt
> Sent: 2024年3月7日 0:24
> To: ffmpeg-devel at ffmpeg.org
> Subject: Re: [FFmpeg-devel] [PATCH] avcodec/libx264: fix extradata when config annexb=0
>
> Zhao Zhili:
> >
> >
> >> -----Original Message-----
> >> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Andreas Rheinhardt
> >> Sent: 2024年3月6日 22:19
> >> To: ffmpeg-devel at ffmpeg.org
> >> Subject: Re: [FFmpeg-devel] [PATCH] avcodec/libx264: fix extradata when config annexb=0
> >>
> >> Zhao Zhili:
> >>> From: Zhao Zhili <zhilizhao at tencent.com>
> >>>
> >>> ---
> >>> configure | 2 +-
> >>> libavcodec/libx264.c | 148 ++++++++++++++++++++++++++++++++++++-------
> >>> 2 files changed, 125 insertions(+), 25 deletions(-)
> >>>
> >>> diff --git a/configure b/configure
> >>> index bbf1a70731..ef58c3fe97 100755
> >>> --- a/configure
> >>> +++ b/configure
> >>> @@ -3491,7 +3491,7 @@ libwebp_encoder_deps="libwebp"
> >>> libwebp_anim_encoder_deps="libwebp"
> >>> libx262_encoder_deps="libx262"
> >>> libx264_encoder_deps="libx264"
> >>> -libx264_encoder_select="atsc_a53"
> >>> +libx264_encoder_select="atsc_a53 h264parse"
> >>> libx264rgb_encoder_deps="libx264"
> >>> libx264rgb_encoder_select="libx264_encoder"
> >>> libx265_encoder_deps="libx265"
> >>> diff --git a/libavcodec/libx264.c b/libavcodec/libx264.c
> >>> index 10d646bd76..10dfda125d 100644
> >>> --- a/libavcodec/libx264.c
> >>> +++ b/libavcodec/libx264.c
> >>> @@ -34,9 +34,12 @@
> >>> #include "avcodec.h"
> >>> #include "codec_internal.h"
> >>> #include "encode.h"
> >>> +#include "get_bits.h"
> >>> +#include "h264_ps.h"
> >>> #include "internal.h"
> >>> #include "packet_internal.h"
> >>> #include "atsc_a53.h"
> >>> +#include "put_bits.h"
> >>> #include "sei.h"
> >>>
> >>> #include <x264.h>
> >>> @@ -865,6 +868,124 @@ static int convert_pix_fmt(enum AVPixelFormat pix_fmt)
> >>> return 0;
> >>> }
> >>>
> >>> +static int set_avcc_extradata(AVCodecContext *avctx, x264_nal_t *nal, int nnal)
> >>> +{
> >>> + x264_nal_t *sps_nal = NULL;
> >>> + x264_nal_t *pps_nal = NULL;
> >>> + GetBitContext gbc;
> >>> + PutBitContext pbc;
> >>> + H264ParamSets ps = { 0 };
> >>> + uint8_t *p;
> >>> + int ret, profile;
> >>> +
> >>> + /* We know it's in the order of SPS/PPS/SEI, but it's not documented in x264 API.
> >>> + * The x264 param i_sps_id implies there is a single pair of SPS/PPS.
> >>> + */
> >>> + for (int i = 0; i < nnal; i++) {
> >>> + if (nal[i].i_type == NAL_SPS)
> >>> + sps_nal = &nal[i];
> >>> + else if (nal[i].i_type == NAL_PPS)
> >>> + pps_nal = &nal[i];
> >>> + }
> >>> + av_assert0(sps_nal);
> >>> + av_assert0(pps_nal);
> >>
> >> We should never assert on the output of external libraries.
> >
> > Ok, will just remove the assert. I don't think a check is needed.
> >
> >>
> >>> +
> >>> + // +11 for AVCDecoderConfigurationRecord, will shrink to the real size finally
> >>> + avctx->extradata_size = sps_nal->i_payload + pps_nal->i_payload + 11;
> >>> + avctx->extradata = av_mallocz(avctx->extradata_size + AV_INPUT_BUFFER_MIN_SIZE);
> >>
> >> AV_INPUT_BUFFER_PADDING_SIZE
> >
> > Oops.
> >
> >>
> >>> + if (!avctx->extradata)
> >>> + return AVERROR(ENOMEM);
> >>> +
> >>> + // Skip size part
> >>> + p = sps_nal->p_payload + 4;
> >>> + init_get_bits8(&gbc, p, sps_nal->i_payload - 4);
> >>> + skip_bits(&gbc, 8);
> >>> + ret = ff_h264_decode_seq_parameter_set(&gbc, avctx, &ps, 0);
> >>
> >> This is completely overblown. You only read three fields: profile_idc,
> >> the constraint fields and level_idc. All three are at the start of the
> >> SPS and occupy a whole byte, so there is no need for a GetBitContext at all.
> >> (Yes, I am aware of the hypothetical scenario in which we could have a
> >> 0x03 escape byte before level_idc (in case of a hypothetical future
> >> profile zero), but even then using ff_h264_decode_seq_parameter_set() is
> >> overblown.)
> >
> > You missed chroma_format_idc, bit_depth_luma/bit_depth_chroma.
> > I can get that from pixel format, but there are a lot of them.
> >
>
> Ok, but even then I fail to see a convenience advantage here. After all,
> using it e.g. forces the following loop onto you. And it forces cleanup
> code here. Also, there is a gotcha here. See below.
I didn't use ff_h264_decode_seq_parameter_set at first, until met
chroma_format_idc. Get it from multiple pixel format doesn't like
the right way to go. Manually decode golomb is code duplication.
The following loop can be removed by using i_sps_id from x264
params. Loop over sps_list is more self-contained.
For hevc, it will be more complex and unlikely to avoid ff_hevc_decode_nal_.
>
> >>
> >>> + if (ret < 0)
> >>> + return ret;
> >>> +
> >>> + for (int i = 0; i < FF_ARRAY_ELEMS(ps.sps_list); i++) {
> >>> + if (ps.sps_list[i]) {
> >>> + ps.sps = ps.sps_list[i];
> >>> + break;
> >>> + }
> >>> + }
> >>> +
> >>> + // Now create AVCDecoderConfigurationRecord
> >>> + init_put_bits(&pbc, avctx->extradata, avctx->extradata_size);
> >>> + put_bits(&pbc, 8, 1); // version
> >>> + put_bits(&pbc, 8, ps.sps->profile_idc); // AVCProfileIndication
> >>> + put_bits(&pbc, 8, ps.sps->constraint_set_flags); // profile_compatibility
>
> Sure this is correct at all? It seems to me that h264_ps.c reverses the
> order of the constraint_set_flags (constraint_set0_flag (which is the
> 0x80 bit in a file) will be mapped to 0x01 etc.) and it discards
> reserved_zero_2bits. You should instead simply copy the relevant byte
> from the SPS.
>
> >>> + put_bits(&pbc, 8, ps.sps->level_idc); // AVCLevelIndication
> >>> + put_bits(&pbc, 11, 0x7FF);
> >>> +
> >>> + p = sps_nal->p_payload + 4;
> >>> + put_bits(&pbc, 5, 1); // numOfSequenceParameterSets
> >>> + put_bits(&pbc, 16, sps_nal->i_payload - 4); // sequenceParameterSetLength
> >>
> >> I also think that using a PutBitContext is overblown here (most fields
> >> don't cross a byte-boundary, the bit-alignment of every field is fixed),
> >> in particular if you copy the NALUs eight bits at a time:
> >
> > It's readable, close related to the spec, and it counts how many bytes actually used
> > automatically. I choose it for these convenience over performance.
> >
>
> I regard a simple memcpy() as more convenient than the following loop:
>
> >>
> >>> + for (int i = 0; i < sps_nal->i_payload - 4; i++)
> >>> + put_bits(&pbc, 8, p[i]);
> >>> +
> >>> + p = pps_nal->p_payload + 4;
> >>> + put_bits(&pbc, 8, 1); // numOfPictureParameterSets
> >>> + put_bits(&pbc, 16, pps_nal->i_payload - 4); // pictureParameterSetLength
> >>> + for (int i = 0; i < pps_nal->i_payload - 4; i++)
> >>> + put_bits(&pbc, 8, p[i]);
> >>> +
> >>> + profile = ps.sps->profile_idc;
> >>> + if (profile != 66 && profile != 77 && profile != 88) {
> >>> + put_bits(&pbc, 6, 0x3F);
> >>> + put_bits(&pbc, 2, ps.sps->chroma_format_idc);
> >>> + put_bits(&pbc, 5, 0x1F);
> >>> + put_bits(&pbc, 3, ps.sps->bit_depth_luma - 8);
> >>> + put_bits(&pbc, 5, 0x1F);
> >>> + put_bits(&pbc, 3, ps.sps->bit_depth_chroma - 8);
> >>> + put_bits(&pbc, 8, 0);
> >>> + }
> >>> + flush_put_bits(&pbc);
> >>> + avctx->extradata_size = put_bytes_output(&pbc);
> >>> +
> >>> + ff_h264_ps_uninit(&ps);
> >>> +
> >>> + return 0;
> >>> +}
> >>> +
> >>> +static int set_extradata(AVCodecContext *avctx)
> >>> +{
> >>> + X264Context *x4 = avctx->priv_data;
> >>> + x264_nal_t *nal;
> >>> + uint8_t *p;
> >>> + int nnal, s;
> >>> +
> >>> + s = x264_encoder_headers(x4->enc, &nal, &nnal);
> >>> + if (s < 0)
> >>> + return AVERROR_EXTERNAL;
> >>> +
> >>> + if (!x4->params.b_annexb)
> >>> + return set_avcc_extradata(avctx, nal, nnal);
> >>> +
> >>> + avctx->extradata = p = av_mallocz(s + AV_INPUT_BUFFER_PADDING_SIZE);
> >>> + if (!p)
> >>> + return AVERROR(ENOMEM);
> >>> +
> >>> + for (int i = 0; i < nnal; i++) {
> >>> + /* Don't put the SEI in extradata. */
> >>> + if (nal[i].i_type == NAL_SEI) {
> >>> + av_log(avctx, AV_LOG_INFO, "%s\n", nal[i].p_payload + 25);
> >>> + x4->sei_size = nal[i].i_payload;
> >>> + x4->sei = av_malloc(x4->sei_size);
> >>> + if (!x4->sei)
> >>> + return AVERROR(ENOMEM);
> >>> + memcpy(x4->sei, nal[i].p_payload, nal[i].i_payload);
> >>> + continue;
> >>> + }
> >>> + memcpy(p, nal[i].p_payload, nal[i].i_payload);
> >>> + p += nal[i].i_payload;
> >>> + }
> >>> + avctx->extradata_size = p - avctx->extradata;
> >>> +
> >>> + return 0;
> >>> +}
> >>> +
> >>> #define PARSE_X264_OPT(name, var)\
> >>> if (x4->var && x264_param_parse(&x4->params, name, x4->var) < 0) {\
> >>> av_log(avctx, AV_LOG_ERROR, "Error parsing option '%s' with value '%s'.\n", name, x4->var);\
> >>> @@ -1233,30 +1354,9 @@ FF_ENABLE_DEPRECATION_WARNINGS
> >>> return AVERROR_EXTERNAL;
> >>>
> >>> if (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) {
> >>> - x264_nal_t *nal;
> >>> - uint8_t *p;
> >>> - int nnal, s, i;
> >>> -
> >>> - s = x264_encoder_headers(x4->enc, &nal, &nnal);
> >>> - avctx->extradata = p = av_mallocz(s + AV_INPUT_BUFFER_PADDING_SIZE);
> >>> - if (!p)
> >>> - return AVERROR(ENOMEM);
> >>> -
> >>> - for (i = 0; i < nnal; i++) {
> >>> - /* Don't put the SEI in extradata. */
> >>> - if (nal[i].i_type == NAL_SEI) {
> >>> - av_log(avctx, AV_LOG_INFO, "%s\n", nal[i].p_payload+25);
> >>> - x4->sei_size = nal[i].i_payload;
> >>> - x4->sei = av_malloc(x4->sei_size);
> >>> - if (!x4->sei)
> >>> - return AVERROR(ENOMEM);
> >>> - memcpy(x4->sei, nal[i].p_payload, nal[i].i_payload);
> >>> - continue;
> >>> - }
> >>> - memcpy(p, nal[i].p_payload, nal[i].i_payload);
> >>> - p += nal[i].i_payload;
> >>> - }
> >>> - avctx->extradata_size = p - avctx->extradata;
> >>> + ret = set_extradata(avctx);
> >>> + if (ret < 0)
> >>> + return ret;
> >>> }
> >>>
> >>> cpb_props = ff_encode_add_cpb_side_data(avctx);
> >>
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
More information about the ffmpeg-devel
mailing list