[FFmpeg-devel] [PATCH v2 1/4] lavc/vaapi_encode_h265: Add P frame to GPB frame support for hevc_vaapi

Xiang, Haihao haihao.xiang at intel.com
Tue Feb 22 07:46:41 EET 2022


On Mon, 2022-02-21 at 12:06 +0000, Mark Thompson wrote:
> On 21/02/2022 02:13, Fei Wang wrote:
> > From: Linjie Fu <linjie.fu at intel.com>
> > 
> > Use GPB frames to replace regular P frames if backend driver does not
> > support it.
> > 
> > - GPB:
> >      Generalized P and B picture. P frames replaced by B frames with
> >      forward-predict only, L0 == L1. Normal B frames still have 2
> >      different ref_lists and allow bi-prediction
> > 
> > Signed-off-by: Linjie Fu <linjie.fu at intel.com>
> > Signed-off-by: Fei Wang <fei.w.wang at intel.com>
> > ---
> >   libavcodec/vaapi_encode.c      | 33 +++++++++++++++++++++++++++++++--
> >   libavcodec/vaapi_encode.h      |  1 +
> >   libavcodec/vaapi_encode_h265.c | 15 +++++++++++++++
> >   3 files changed, 47 insertions(+), 2 deletions(-)
> 
> This always fails immediately on current iHD
> (7e357b4bea76b2fe2522e6af41ae02ea69cec49e):
> 
> $ ./ffmpeg_g -v 44 -y -hwaccel vaapi -hwaccel_output_format vaapi -i in.mp4
> -an -c:v hevc_vaapi -low_power 1 out.mp4
> ...
> [hevc_vaapi @ 0x560e81d45e80] Using input frames context (format vaapi) with
> hevc_vaapi encoder.
> [hevc_vaapi @ 0x560e81d45e80] Input surface format is nv12.
> [hevc_vaapi @ 0x560e81d45e80] Using VAAPI profile VAProfileHEVCMain (17).
> [hevc_vaapi @ 0x560e81d45e80] Using VAAPI entrypoint VAEntrypointEncSliceLP
> (8).
> [hevc_vaapi @ 0x560e81d45e80] Using VAAPI render target format YUV420 (0x1).
> [hevc_vaapi @ 0x560e81d45e80] Using CTU size 64x64, min CB size 8x8.
> [hevc_vaapi @ 0x560e81d45e80] No quality level set; using default (25).
> [hevc_vaapi @ 0x560e81d45e80] RC mode: CQP.
> [hevc_vaapi @ 0x560e81d45e80] RC quality: 25.
> [hevc_vaapi @ 0x560e81d45e80] RC framerate: 30000/1001 (29.97 fps).
> [hevc_vaapi @ 0x560e81d45e80] Use GPB B frames to replace regular P frames.
> [hevc_vaapi @ 0x560e81d45e80] Using intra, GPB-B-frames and B-frames
> (supported references: 3 / 3).
> [hevc_vaapi @ 0x560e81d45e80] All wanted packed headers available (wanted 0xd,
> found 0x1f).
> [hevc_vaapi @ 0x560e81d45e80] Using level 4.
> ...
> [hevc_vaapi @ 0x560e81d45e80] Failed to end picture encode issue: 24 (internal
> encoding error).
> [hevc_vaapi @ 0x560e81d45e80] Encode failed: -5.
> Video encoding failed
> ...
> $ cat /proc/cpuinfo | grep 'model name' | head -1
> model name      : Intel(R) Core(TM) i7-1065G7 CPU @ 1.30GHz
> $ uname -v
> #1 SMP PREEMPT Debian 5.16.7-2 (2022-02-09)
> 
> Do you get this too, or is your setup different somehow?

Hi Mark,

I tested this patchset with iHD (7e357b4bea76b2fe2522e6af41ae02ea69cec49e) on
CFL (low_power=0), RKL and DG1, i965 on SKL, and didn't see this issue before.
This day I reproduced this issue on ICL. It seems iHD driver doesn't return
right values for ICL.  

> 
> > diff --git a/libavcodec/vaapi_encode.c b/libavcodec/vaapi_encode.c
> > index 3bf379b1a0..95eca7c288 100644
> > --- a/libavcodec/vaapi_encode.c
> > +++ b/libavcodec/vaapi_encode.c
> > @@ -1845,6 +1845,30 @@ static av_cold int
> > vaapi_encode_init_gop_structure(AVCodecContext *avctx)
> >           ref_l1 = attr.value >> 16 & 0xffff;
> >       }
> >   
> > +    ctx->p_to_gpb = 0;
> > +
> > +#if VA_CHECK_VERSION(1, 9, 0)
> > +    attr = (VAConfigAttrib) { VAConfigAttribPredictionDirection };
> > +    vas = vaGetConfigAttributes(ctx->hwctx->display,
> > +                                ctx->va_profile,
> > +                                ctx->va_entrypoint,
> > +                                &attr, 1);
> 
> This probably shouldn't be done at all if the user has selected a codec
> without B-frames or a configuration which is intra-only, because the log
> message is confusing:
> 
> [mjpeg_vaapi @ 0x55b90d72ee00] Driver does not report whether support GPB, use
> regular P frames.
> [mjpeg_vaapi @ 0x55b90d72ee00] Using intra frames only.
> 
> > +    if (vas != VA_STATUS_SUCCESS) {
> > +        av_log(avctx, AV_LOG_WARNING, "Failed to query prediction direction
> > "
> > +               "attribute: %d (%s).\n", vas, vaErrorStr(vas));
> 
> And fail?

4/4 also ignores the error. It would be better to handle the error in the same
way and update 4/4 too. 

> 
> > +    } else if (attr.value == VA_ATTRIB_NOT_SUPPORTED) {
> > +        av_log(avctx, AV_LOG_VERBOSE, "Driver does not report whether "
> > +               "support GPB, use regular P frames.\n");
> 
> "support GPB" is a strange thing to say.  It's a constraint - any driver which
> supports B-frames will let you have the same thing in both RefPicLists, but
> some require it because they don't support P-frames.
> 
> So maybe something like "Driver does not report any additional prediction
> constraints, using P-frames."?
> 
> > +    } else {
> > +        if (attr.value & VA_PREDICTION_DIRECTION_BI_NOT_EMPTY) {
> > +            ctx->p_to_gpb = 1;
> > +            av_log(avctx, AV_LOG_VERBOSE, "Use GPB B frames to replace "
> > +                   "regular P frames.\n");
> 
> Maybe "Driver does not support P-frames, replacing them with B-frames."?
> 
> > +        } else
> > +            av_log(avctx, AV_LOG_VERBOSE, "Use regular P frames.\n");
> > +    }
> > +#endif
> > +
> >       if (ctx->codec->flags & FLAG_INTRA_ONLY ||
> >           avctx->gop_size <= 1) {
> >           av_log(avctx, AV_LOG_VERBOSE, "Using intra frames only.\n");
> > @@ -1861,8 +1885,13 @@ static av_cold int
> > vaapi_encode_init_gop_structure(AVCodecContext *avctx)
> >           ctx->p_per_i  = INT_MAX;
> >           ctx->b_per_p  = 0;
> >       } else {
> > -        av_log(avctx, AV_LOG_VERBOSE, "Using intra, P- and B-frames "
> > -               "(supported references: %d / %d).\n", ref_l0, ref_l1);
> > +       if (ctx->p_to_gpb)
> > +           av_log(avctx, AV_LOG_VERBOSE, "Using intra, GPB-B-frames and "
> > +                  "B-frames (supported references: %d / %d).\n",
> > +                  ref_l0, ref_l1);
> 
> Seems easier to just say intra and B-frames (though this isn't really adding
> anything to the message above - it's still effectively a P-frame referring to
> one previous frame, just using B-slices to do it in a strange way).
> 
> > +       else
> > +           av_log(avctx, AV_LOG_VERBOSE, "Using intra, P- and B-frames "
> > +                  "(supported references: %d / %d).\n", ref_l0, ref_l1);
> >           ctx->gop_size = avctx->gop_size;
> >           ctx->p_per_i  = INT_MAX;
> >           ctx->b_per_p  = avctx->max_b_frames;
> > diff --git a/libavcodec/vaapi_encode.h b/libavcodec/vaapi_encode.h
> > index b41604a883..61c5615eb8 100644
> > --- a/libavcodec/vaapi_encode.h
> > +++ b/libavcodec/vaapi_encode.h
> > @@ -328,6 +328,7 @@ typedef struct VAAPIEncodeContext {
> >       int idr_counter;
> >       int gop_counter;
> >       int end_of_stream;
> > +    int p_to_gpb;
> >   
> >       // Whether the driver supports ROI at all.
> >       int             roi_allowed;
> > diff --git a/libavcodec/vaapi_encode_h265.c b/libavcodec/vaapi_encode_h265.c
> > index e1dc53dfa9..ea45893508 100644
> > --- a/libavcodec/vaapi_encode_h265.c
> > +++ b/libavcodec/vaapi_encode_h265.c
> > @@ -886,6 +886,7 @@ static int
> > vaapi_encode_h265_init_slice_params(AVCodecContext *avctx,
> >                                                  VAAPIEncodePicture *pic,
> >                                                  VAAPIEncodeSlice *slice)
> >   {
> > +    VAAPIEncodeContext                *ctx = avctx->priv_data;
> >       VAAPIEncodeH265Context           *priv = avctx->priv_data;
> >       VAAPIEncodeH265Picture           *hpic = pic->priv_data;
> >       const H265RawSPS                  *sps = &priv->raw_sps;
> > @@ -908,6 +909,9 @@ static int
> > vaapi_encode_h265_init_slice_params(AVCodecContext *avctx,
> >   
> >       sh->slice_type = hpic->slice_type;
> >   
> > +    if (sh->slice_type == HEVC_SLICE_P && ctx->p_to_gpb)
> > +        sh->slice_type = HEVC_SLICE_B;
> > +
> >       sh->slice_pic_order_cnt_lsb = hpic->pic_order_cnt &
> >           (1 << (sps->log2_max_pic_order_cnt_lsb_minus4 + 4)) - 1;
> >   
> > @@ -1066,6 +1070,9 @@ static int
> > vaapi_encode_h265_init_slice_params(AVCodecContext *avctx,
> >           av_assert0(pic->type == PICTURE_TYPE_P ||
> >                      pic->type == PICTURE_TYPE_B);
> >           vslice->ref_pic_list0[0] = vpic->reference_frames[0];
> > +        if (ctx->p_to_gpb && pic->type == PICTURE_TYPE_P)
> > +            // Reference for GPB B-frame, L0 == L1
> > +            vslice->ref_pic_list1[0] = vpic->reference_frames[0];
> >       }
> >       if (pic->nb_refs >= 2) {
> >           // Forward reference for B-frame.
> > @@ -1073,6 +1080,14 @@ static int
> > vaapi_encode_h265_init_slice_params(AVCodecContext *avctx,
> >           vslice->ref_pic_list1[0] = vpic->reference_frames[1];
> >       }
> >   
> > +    if (pic->type == PICTURE_TYPE_P && ctx->p_to_gpb) {
> > +        vslice->slice_type = HEVC_SLICE_B;
> > +        for (i = 0; i < FF_ARRAY_ELEMS(vslice->ref_pic_list0); i++) {
> > +            vslice->ref_pic_list1[i].picture_id = vslice-
> > >ref_pic_list0[i].picture_id;
> > +            vslice->ref_pic_list1[i].flags      = vslice-
> > >ref_pic_list0[i].flags;
> > +        }
> > +    }
> 
> Is it sensible to implement this properly in the generic code rather than
> having special ad-hoc code here?
> 
> (Is there any chance this will also be a thing for e.g. H.264 or H.266 in
> future?)

Considering that VAConfigAttribPredictionDirection is not a specific attribute
for hevc in VA-API, I think it is acceptable to implement this in the generic
code. 

Thanks
Haihao

> 
> > +
> >       return 0;
> >   }
> >   
> 
> - Mark
> _______________________________________________
> 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