Re: [FFmpeg-devel] [PATCH v7 3/4] avcodec/hevc: Add alpha layer support

[Andreas Rheinhardt <andreas.rheinhardt@outlook.com>]

Zhao Zhili:
> 
> 
>> On Feb 12, 2025, at 00:10, Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
>>
>> Zhao Zhili:
>>>
>>>
>>>> On Feb 11, 2025, at 23:37, Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
>>>>
>>>> Zhao Zhili:
>>>>> From: Zhao Zhili <zhilizhao@tencent.com>
>>>>>
>>>>> Signed-off-by: Zhao Zhili <zhilizhao@tencent.com>
>>>>> ---
>>>>> libavcodec/hevc/hevcdec.c | 73 ++++++++++++++++++++++++++++++++++++++-
>>>>> libavcodec/hevc/hevcdec.h |  2 ++
>>>>> libavcodec/hevc/refs.c    | 35 ++++++++++++++++++-
>>>>> 3 files changed, 108 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/libavcodec/hevc/hevcdec.c b/libavcodec/hevc/hevcdec.c
>>>>> index e9c045f7a1..f71edf213b 100644
>>>>> --- a/libavcodec/hevc/hevcdec.c
>>>>> +++ b/libavcodec/hevc/hevcdec.c
>>>>> @@ -466,6 +466,25 @@ static int export_multilayer(HEVCContext *s, const HEVCVPS *vps)
>>>>>    return 0;
>>>>> }
>>>>>
>>>>> +int ff_hevc_is_alpha_video(const HEVCContext *s)
>>>>> +{
>>>>> +    const HEVCVPS *vps = s->vps;
>>>>> +    int ret = 0;
>>>>> +
>>>>> +    if (vps->nb_layers != 2 || !vps->layer_id_in_nuh[1])
>>>>> +        return 0;
>>>>> +
>>>>> +    /* decode_vps_ext() guarantees that SCALABILITY_AUXILIARY with AuxId other
>>>>> +     * than alpha cannot reach here.
>>>>> +     */
>>>>> +    ret = (s->vps->scalability_mask_flag & HEVC_SCALABILITY_AUXILIARY);
>>>>> +
>>>>> +    av_log(s->avctx, AV_LOG_DEBUG, "Multi layer video, %s alpha video\n",
>>>>> +           ret ? "is" : "not");
>>>>> +
>>>>> +    return ret;
>>>>> +}
>>>>> +
>>>>> static int setup_multilayer(HEVCContext *s, const HEVCVPS *vps)
>>>>> {
>>>>>    unsigned layers_active_output = 0, highest_layer;
>>>>> @@ -473,6 +492,18 @@ static int setup_multilayer(HEVCContext *s, const HEVCVPS *vps)
>>>>>    s->layers_active_output = 1;
>>>>>    s->layers_active_decode = 1;
>>>>>
>>>>> +    if (ff_hevc_is_alpha_video(s)) {
>>>>> +        const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(s->avctx->pix_fmt);
>>>>> +
>>>>> +        if (!(desc->flags & AV_PIX_FMT_FLAG_ALPHA))
>>>>> +            return 0;
>>>>> +
>>>>> +        s->layers_active_decode = (1 << vps->nb_layers) - 1;
>>>>> +        s->layers_active_output = 1;
>>>>> +
>>>>> +        return 0;
>>>>> +    }
>>>>> +
>>>>>    // nothing requested - decode base layer only
>>>>>    if (!s->nb_view_ids)
>>>>>        return 0;
>>>>> @@ -530,6 +561,34 @@ static int setup_multilayer(HEVCContext *s, const HEVCVPS *vps)
>>>>>    return 0;
>>>>> }
>>>>>
>>>>> +static enum AVPixelFormat map_to_alpha_format(HEVCContext *s,
>>>>> +                                              enum AVPixelFormat pix_fmt)
>>>>> +{
>>>>> +    switch (pix_fmt) {
>>>>> +    case AV_PIX_FMT_YUV420P:
>>>>> +    case AV_PIX_FMT_YUVJ420P:
>>>>> +        return AV_PIX_FMT_YUVA420P;
>>>>> +    case AV_PIX_FMT_YUV420P10:
>>>>> +        return AV_PIX_FMT_YUVA420P10;
>>>>> +    case AV_PIX_FMT_YUV444P:
>>>>> +        return AV_PIX_FMT_YUVA444P;
>>>>> +    case AV_PIX_FMT_YUV422P:
>>>>> +        return AV_PIX_FMT_YUVA422P;
>>>>> +    case AV_PIX_FMT_YUV422P10LE:
>>>>> +        return AV_PIX_FMT_YUVA422P10LE;
>>>>> +    case AV_PIX_FMT_YUV444P10:
>>>>> +        return AV_PIX_FMT_YUVA444P10;
>>>>> +    case AV_PIX_FMT_YUV444P12:
>>>>> +        return AV_PIX_FMT_YUVA444P12;
>>>>> +    case AV_PIX_FMT_YUV422P12:
>>>>> +        return AV_PIX_FMT_YUVA422P12;
>>>>> +    default:
>>>>> +        av_log(s->avctx, AV_LOG_WARNING, "No alpha pixel format map for %s\n",
>>>>> +               av_get_pix_fmt_name(pix_fmt));
>>>>> +        return AV_PIX_FMT_NONE;
>>>>> +    }
>>>>> +}
>>>>> +
>>>>> static enum AVPixelFormat get_format(HEVCContext *s, const HEVCSPS *sps)
>>>>> {
>>>>> #define HWACCEL_MAX (CONFIG_HEVC_DXVA2_HWACCEL + \
>>>>> @@ -540,9 +599,13 @@ static enum AVPixelFormat get_format(HEVCContext *s, const HEVCSPS *sps)
>>>>>                     CONFIG_HEVC_VIDEOTOOLBOX_HWACCEL + \
>>>>>                     CONFIG_HEVC_VDPAU_HWACCEL + \
>>>>>                     CONFIG_HEVC_VULKAN_HWACCEL)
>>>>> -    enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmt = pix_fmts;
>>>>> +    enum AVPixelFormat pix_fmts[HWACCEL_MAX + 3], *fmt = pix_fmts;
>>>>> +    enum AVPixelFormat alpha_fmt = AV_PIX_FMT_NONE;
>>>>>    int ret;
>>>>>
>>>>> +    if (ff_hevc_is_alpha_video(s))
>>>>> +        alpha_fmt = map_to_alpha_format(s, sps->pix_fmt);
>>>>> +
>>>>>    switch (sps->pix_fmt) {
>>>>>    case AV_PIX_FMT_YUV420P:
>>>>>    case AV_PIX_FMT_YUVJ420P:
>>>>> @@ -664,6 +727,8 @@ static enum AVPixelFormat get_format(HEVCContext *s, const HEVCSPS *sps)
>>>>>        break;
>>>>>    }
>>>>>
>>>>> +    if (alpha_fmt != AV_PIX_FMT_NONE)
>>>>> +        *fmt++ = alpha_fmt;
>>>>>    *fmt++ = sps->pix_fmt;
>>>>>    *fmt = AV_PIX_FMT_NONE;
>>>>>
>>>>> @@ -3194,6 +3259,12 @@ static int hevc_frame_start(HEVCContext *s, HEVCLayerContext *l,
>>>>>                !sps->vui.common.video_signal_type_present_flag)
>>>>>                pix_fmt = sps_base->pix_fmt;
>>>>>
>>>>> +            // Ignore range mismatch between base layer and alpha layer
>>>>> +            if (ff_hevc_is_alpha_video(s) &&
>>>>> +                sps_base->pix_fmt == AV_PIX_FMT_YUV420P &&
>>>>> +                pix_fmt == AV_PIX_FMT_YUVJ420P)
>>>>> +                pix_fmt = sps_base->pix_fmt;
>>>>> +
>>>>>            if (pix_fmt     != sps_base->pix_fmt ||
>>>>>                sps->width  != sps_base->width   ||
>>>>>                sps->height != sps_base->height) {
>>>>> diff --git a/libavcodec/hevc/hevcdec.h b/libavcodec/hevc/hevcdec.h
>>>>> index 4e95035688..b2b725b5cd 100644
>>>>> --- a/libavcodec/hevc/hevcdec.h
>>>>> +++ b/libavcodec/hevc/hevcdec.h
>>>>> @@ -714,6 +714,8 @@ void ff_hevc_hls_residual_coding(HEVCLocalContext *lc, const HEVCPPS *pps,
>>>>>
>>>>> void ff_hevc_hls_mvd_coding(HEVCLocalContext *lc, int x0, int y0, int log2_cb_size);
>>>>>
>>>>> +int ff_hevc_is_alpha_video(const HEVCContext *s);
>>>>> +
>>>>> extern const uint8_t ff_hevc_qpel_extra_before[4];
>>>>> extern const uint8_t ff_hevc_qpel_extra_after[4];
>>>>> extern const uint8_t ff_hevc_qpel_extra[4];
>>>>> diff --git a/libavcodec/hevc/refs.c b/libavcodec/hevc/refs.c
>>>>> index dd7f7f95a8..6f10efd0ac 100644
>>>>> --- a/libavcodec/hevc/refs.c
>>>>> +++ b/libavcodec/hevc/refs.c
>>>>> @@ -79,6 +79,31 @@ void ff_hevc_flush_dpb(HEVCContext *s)
>>>>>    }
>>>>> }
>>>>>
>>>>> +static int replace_alpha_plane(AVFrame *alpha, AVFrame *base)
>>>>> +{
>>>>> +    AVBufferRef *base_a = av_frame_get_plane_buffer(base, 3);
>>>>> +    uintptr_t data = (uintptr_t)alpha->data[0];
>>>>> +    int ret;
>>>>> +
>>>>> +    for (int i = 0; i < FF_ARRAY_ELEMS(alpha->buf) && alpha->buf[i]; i++) {
>>>>> +        AVBufferRef *buf = alpha->buf[i];
>>>>> +        uintptr_t buf_begin = (uintptr_t)buf->data;
>>>>> +
>>>>> +        if (data >=  buf_begin && data < buf_begin + buf->size) {
>>>>> +            ret = av_buffer_replace(&alpha->buf[i], base_a);
>>>>> +            if (ret < 0)
>>>>> +                return ret;
>>>>> +
>>>>> +            alpha->linesize[0] = base->linesize[3];
>>>>> +            alpha->data[0] = base->data[3];
>>>>> +
>>>>> +            return 0;
>>>>> +        }
>>>>> +    }
>>>>
>>>> Why does the decoding process actually need multiple references to the
>>>> buffer of the alpha plane?
>>>
>>> I’m not sure if I understand your question correctly.
>>>
>>> 1. Both the base layer and the alpha layer are decoded as yuv420p.
>>> 2. The y plane of alpha layer is actually alpha.
>>> 3. The decoder finally output yuva, not two yuv420p.
>>>
>>> So here make y plane of alpha frame reference the alpha plane of base frame.
>>> When output, the base frame will contain YUV and alpha from two layers.
>>>
>>
>> The base layer AVFrame already contains a reference to the alpha plane;
>> so why do we need another one? Does the base AVFrame get unreferenced
>> before the alpha plane is finished decoding?
> 
> In the decoding process, it’s almost the same as MV-HEVC:
> 1. The base layer is decoded to the YUV of base frame, without touching the alpha plane.
> 2. The alpha layer is decoded to YUV of alpha frame.
> 
> We can output the two frames like MV-HEVC, but it’s user friendly to output in YUVA format.
> So I’m letting the alpha layer decoded to the alpha plane of the base layer. Or I can do the
> reverse. It’s the same.
> 
> We can skip the alpha frame totally. Then the decoding process and DPB buffer management
> will be different. Current method is simple but use more memory. I can work on that lately.
> 
It seems like you still don't get my question. Let me give it another
try: What would happen if you simply did not set alpha->buf[0] (but
still data[0] and linesize[0])? Would it crash? Or would it work because
the actual alpha plane is always referenced by the base layer AVFrame?

- Andreas

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".