[FFmpeg-devel] [PATCH 2/3] avcodec/cbs_av1: infer segmentation parameters from reference frames

[Mark Thompson]

On 21/10/2020 01:11, James Almer wrote:
> Partially implements of setup_past_independence() and load_previous().
> These ensures they are always set, even if the values were not coded
> in the input bitstream and will not be coded in the output bitstream.
> 
> Signed-off-by: James Almer <jamrial at gmail.com>
> ---
>   libavcodec/cbs_av1.h                 |  2 ++
>   libavcodec/cbs_av1_syntax_template.c | 43 +++++++++++++++++++++-------
>   2 files changed, 35 insertions(+), 10 deletions(-)
> 
> diff --git a/libavcodec/cbs_av1.h b/libavcodec/cbs_av1.h
> index 97aeee9795..a2d78e736f 100644
> --- a/libavcodec/cbs_av1.h
> +++ b/libavcodec/cbs_av1.h
> @@ -416,6 +416,8 @@ typedef struct AV1ReferenceFrameState {
>   
>       int8_t  loop_filter_ref_deltas[AV1_TOTAL_REFS_PER_FRAME];
>       int8_t  loop_filter_mode_deltas[2];
> +    uint8_t feature_enabled[AV1_MAX_SEGMENTS][AV1_SEG_LVL_MAX];
> +    int16_t feature_value[AV1_MAX_SEGMENTS][AV1_SEG_LVL_MAX];
>   } AV1ReferenceFrameState;
>   
>   typedef struct CodedBitstreamAV1Context {
> diff --git a/libavcodec/cbs_av1_syntax_template.c b/libavcodec/cbs_av1_syntax_template.c
> index 4edf4fd47c..2df5619279 100644
> --- a/libavcodec/cbs_av1_syntax_template.c
> +++ b/libavcodec/cbs_av1_syntax_template.c
> @@ -743,8 +743,11 @@ static int FUNC(quantization_params)(CodedBitstreamContext *ctx, RWContext *rw,
>   static int FUNC(segmentation_params)(CodedBitstreamContext *ctx, RWContext *rw,
>                                        AV1RawFrameHeader *current)
>   {
> +    CodedBitstreamAV1Context  *priv = ctx->priv_data;
>       static const uint8_t bits[AV1_SEG_LVL_MAX] = { 8, 6, 6, 6, 6, 3, 0, 0 };
>       static const uint8_t sign[AV1_SEG_LVL_MAX] = { 1, 1, 1, 1, 1, 0, 0, 0 };
> +    static const uint8_t default_feature_enabled[AV1_SEG_LVL_MAX] = { 0 };
> +    static const int16_t default_feature_value[AV1_SEG_LVL_MAX] = { 0 };
>       int i, j, err;
>   
>       flag(segmentation_enabled);
> @@ -763,22 +766,38 @@ static int FUNC(segmentation_params)(CodedBitstreamContext *ctx, RWContext *rw,
>               flag(segmentation_update_data);
>           }
>   
> -        if (current->segmentation_update_data) {
>               for (i = 0; i < AV1_MAX_SEGMENTS; i++) {
> -                for (j = 0; j < AV1_SEG_LVL_MAX; j++) {
> -                    flags(feature_enabled[i][j], 2, i, j);
> +                const uint8_t *feature_enabled;
> +                const int16_t *feature_value;

Same comment as previous about ref_feature_enabled etc.

> +
> +                if (current->primary_ref_frame == AV1_PRIMARY_REF_NONE) {
> +                    feature_enabled = default_feature_enabled;
> +                    feature_value = default_feature_value;
> +                } else {
> +                    feature_enabled =
> +                        priv->ref[current->ref_frame_idx[current->primary_ref_frame]].feature_enabled[i];
> +                    feature_value =
> +                        priv->ref[current->ref_frame_idx[current->primary_ref_frame]].feature_value[i];
> +                }
>   
> -                    if (current->feature_enabled[i][j] && bits[j] > 0) {
> -                        if (sign[j])
> -                            sus(1 + bits[j], feature_value[i][j], 2, i, j);
> -                        else
> -                            fbs(bits[j], feature_value[i][j], 2, i, j);
> +                for (j = 0; j < AV1_SEG_LVL_MAX; j++) {
> +                    if (current->segmentation_update_data) {
> +                        flags(feature_enabled[i][j], 2, i, j);
> +
> +                        if (current->feature_enabled[i][j] && bits[j] > 0) {
> +                            if (sign[j])
> +                                sus(1 + bits[j], feature_value[i][j], 2, i, j);
> +                            else
> +                                fbs(bits[j], feature_value[i][j], 2, i, j);
> +                        } else {
> +                            infer(feature_value[i][j], 0);
> +                        }
>                       } else {
> -                        infer(feature_value[i][j], 0);
> +                        infer(feature_enabled[i][j], feature_enabled[j]);
> +                        infer(feature_value[i][j], feature_value[j]);
>                       }
>                   }
>               }
> -        }
>       } else {
>           for (i = 0; i < AV1_MAX_SEGMENTS; i++) {
>               for (j = 0; j < AV1_SEG_LVL_MAX; j++) {
> @@ -1645,6 +1664,10 @@ update_refs:
>                      sizeof(current->loop_filter_ref_deltas));
>               memcpy(priv->ref[i].loop_filter_mode_deltas, current->loop_filter_mode_deltas,
>                      sizeof(current->loop_filter_mode_deltas));
> +            memcpy(priv->ref[i].feature_enabled, current->feature_enabled,
> +                   sizeof(current->feature_enabled));
> +            memcpy(priv->ref[i].feature_value, current->feature_value,
> +                   sizeof(current->feature_value));
>           }
>       }
>   
> 

Also sensible.  (Needs the reindent as well, not much value in it being separate given that this diff has most lines changed anyway.)

Thanks,

- Mark