[FFmpeg-devel] [PATCH 2/2] lavc/hevc_refs: Fix the logic of find_ref_idx()

[Fu, Linjie]

> From: Fu, Linjie <linjie.fu at intel.com>
> Sent: Tuesday, May 12, 2020 21:44
> To: ffmpeg-devel at ffmpeg.org
> Cc: Xu, Guangxin <guangxin.xu at intel.com>; Fu, Linjie <linjie.fu at intel.com>
> Subject: [PATCH 2/2] lavc/hevc_refs: Fix the logic of find_ref_idx()
> 
> From: Xu Guangxin <guangxin.xu at intel.com>
> 
> Currently find_ref_idx() would trigger 2 scans in DPB to find the
> requested POC:
> 1. Firstly, ignore MSB of ref->poc and search for the requested POC;
> 2. Secondly, compare the entire ref->poc with requested POC;
> 
> For long term reference, we are able to only check LSB if MSB is not
> presented(e.g. delta_poc_msb_present_flag == 0). However, for short
> term reference, we should never ignore poc's MSB and it should be
> kind of bit-exact. (Details in 8.3.2)
> 
> Otherwise this leads to decoding failures like:
> [hevc @ 0x5638f4328600] Error constructing the frame RPS.
> [hevc @ 0x5638f4328600] Error parsing NAL unit #2.
> [hevc @ 0x5638f4338a80] Could not find ref with POC 21
> Error while decoding stream #0:0: Invalid data found when processing input
> 
> Search the requested POC based on whether MSB is used, and avoid
> the 2-times scan for DPB buffer. This benefits both native HEVC
> decoder and integrated HW decoders.
> 
> Signed-off-by: Linjie Fu <linjie.fu at intel.com>
> ---
> 
> Since it's kind of difficult to generate an identical bitstream for
> fate or test, I'd like to elaborate more about one of the failures:
> 
>     requested POC = 5;
>     LtMask = (1 << 4) - 1 = 15;
>     ref[0]->poc = 21; // unexpected ref for poc = 5 (short term)
>     ref[1]->poc = 5;  // expected ref for poc = 5 (short term)
> 
> Hence find_ref_idx() would wrongly return a ref with poc = 21, which
> leads to the decoding error.
> 
>  libavcodec/hevc_refs.c | 38 ++++++++++++++++++++------------------
>  1 file changed, 20 insertions(+), 18 deletions(-)
> 
> diff --git a/libavcodec/hevc_refs.c b/libavcodec/hevc_refs.c
> index 7870a72..73aa6d8 100644
> --- a/libavcodec/hevc_refs.c
> +++ b/libavcodec/hevc_refs.c
> @@ -358,24 +358,26 @@ int ff_hevc_slice_rpl(HEVCContext *s)
>      return 0;
>  }
> 
> -static HEVCFrame *find_ref_idx(HEVCContext *s, int poc)
> +static HEVCFrame *find_ref_idx(HEVCContext *s, int poc, uint8_t use_msb)
>  {
>      int i;
> -    int LtMask = (1 << s->ps.sps->log2_max_poc_lsb) - 1;
> 
> -    for (i = 0; i < FF_ARRAY_ELEMS(s->DPB); i++) {
> -        HEVCFrame *ref = &s->DPB[i];
> -        if (ref->frame->buf[0] && (ref->sequence == s->seq_decode)) {
> -            if ((ref->poc & LtMask) == poc)
> -                return ref;
> +    if (use_msb) {
> +        for (i = 0; i < FF_ARRAY_ELEMS(s->DPB); i++) {
> +            HEVCFrame *ref = &s->DPB[i];
> +            if (ref->frame->buf[0] && (ref->sequence == s->seq_decode)) {
> +                if (ref->poc == poc)
> +                    return ref;
> +            }
>          }
> -    }
> -
> -    for (i = 0; i < FF_ARRAY_ELEMS(s->DPB); i++) {
> -        HEVCFrame *ref = &s->DPB[i];
> -        if (ref->frame->buf[0] && ref->sequence == s->seq_decode) {
> -            if (ref->poc == poc || (ref->poc & LtMask) == poc)
> -                return ref;
> +    } else {
> +        int LtMask = (1 << s->ps.sps->log2_max_poc_lsb) - 1;
> +        for (i = 0; i < FF_ARRAY_ELEMS(s->DPB); i++) {
> +            HEVCFrame *ref = &s->DPB[i];
> +            if (ref->frame->buf[0] && ref->sequence == s->seq_decode) {
> +                if ((ref->poc & LtMask) == poc)
> +                    return ref;
> +            }
>          }
>      }
> 
> @@ -427,9 +429,9 @@ static HEVCFrame
> *generate_missing_ref(HEVCContext *s, int poc)
> 
>  /* add a reference with the given poc to the list and mark it as used in DPB
> */
>  static int add_candidate_ref(HEVCContext *s, RefPicList *list,
> -                             int poc, int ref_flag)
> +                             int poc, int ref_flag, uint8_t use_msb)
>  {
> -    HEVCFrame *ref = find_ref_idx(s, poc);
> +    HEVCFrame *ref = find_ref_idx(s, poc, use_msb);
> 
>      if (ref == s->ref || list->nb_refs >= HEVC_MAX_REFS)
>          return AVERROR_INVALIDDATA;
> @@ -485,7 +487,7 @@ int ff_hevc_frame_rps(HEVCContext *s)
>          else
>              list = ST_CURR_AFT;
> 
> -        ret = add_candidate_ref(s, &rps[list], poc,
> HEVC_FRAME_FLAG_SHORT_REF);
> +        ret = add_candidate_ref(s, &rps[list], poc,
> HEVC_FRAME_FLAG_SHORT_REF, 1);
>          if (ret < 0)
>              goto fail;
>      }
> @@ -495,7 +497,7 @@ int ff_hevc_frame_rps(HEVCContext *s)
>          int poc  = long_rps->poc[i];
>          int list = long_rps->used[i] ? LT_CURR : LT_FOLL;
> 
> -        ret = add_candidate_ref(s, &rps[list], poc,
> HEVC_FRAME_FLAG_LONG_REF);
> +        ret = add_candidate_ref(s, &rps[list], poc,
> HEVC_FRAME_FLAG_LONG_REF, long_rps->poc_msb_present[i]);
>          if (ret < 0)
>              goto fail;
>      }
> --
> 2.7.4

Ping for this patch set.
Verified with 1000 + decoding conformance cases including 8/10 bit, no regression observed.

 - Linjie