[FFmpeg-devel] [PATCH, v2] lavc/vaapi_encode: add support for AVC Trellis

Fu, Linjie linjie.fu at intel.com
Mon Jun 17 08:47:24 EEST 2019



> -----Original Message-----
> From: mypopy at gmail.com [mailto:mypopy at gmail.com]
> Sent: Wednesday, June 12, 2019 16:14
> To: FFmpeg development discussions and patches <ffmpeg-
> devel at ffmpeg.org>
> Cc: Fu, Linjie <linjie.fu at intel.com>
> Subject: Re: [FFmpeg-devel] [PATCH, v2] lavc/vaapi_encode: add support for
> AVC Trellis
> 
> On Wed, Jun 12, 2019 at 3:28 PM Linjie Fu <linjie.fu at intel.com> wrote:
> >
> > Add support for VAAPI AVC Trellis Quantization with limitation:
> >     - VA-API version >= (1, 0, 0)
> >
> > Use option "-trellis off/I/P/B" to disable or enable Trellis
> > quantization for I/P/B frames.
> >
> > Signed-off-by: Linjie Fu <linjie.fu at intel.com>
> > ---
> > [v2]: Since nonstandard struct for VAEncMiscParameterQuantization is
> > fixed: https://github.com/intel/libva/issues/265
> > update patch based on:
> >
> http://git.ffmpeg.org/gitweb/ffmpeg.git/commit/2880a32c668023bfee47450
> 95c885450d547ae45
> >  libavcodec/vaapi_encode.c      | 48
> ++++++++++++++++++++++++++++++++++
> >  libavcodec/vaapi_encode.h      |  9 +++++--
> >  libavcodec/vaapi_encode_h264.c |  9 +++++++
> >  3 files changed, 64 insertions(+), 2 deletions(-)
> >
> > diff --git a/libavcodec/vaapi_encode.c b/libavcodec/vaapi_encode.c
> > index dd2a24de04..fbfbe78c6b 100644
> > --- a/libavcodec/vaapi_encode.c
> > +++ b/libavcodec/vaapi_encode.c
> > @@ -1671,6 +1671,48 @@ rc_mode_found:
> >      return 0;
> >  }
> >
> > +static av_cold int vaapi_encode_init_quantization(AVCodecContext
> *avctx)
> > +{
> > +#if VA_CHECK_VERSION(1, 0, 0)
> > +    VAAPIEncodeContext *ctx = avctx->priv_data;
> > +    VAStatus vas;
> > +    VAConfigAttrib attr = { VAConfigAttribEncQuantization };
> > +    int trellis = ctx->trellis;
> > +
> > +    vas = vaGetConfigAttributes(ctx->hwctx->display,
> > +                                ctx->va_profile,
> > +                                ctx->va_entrypoint,
> > +                                &attr, 1);
> > +    if (vas != VA_STATUS_SUCCESS) {
> > +        av_log(avctx, AV_LOG_ERROR, "Failed to query quantization "
> > +               "config attribute: %d (%s).\n", vas, vaErrorStr(vas));
> > +        return AVERROR_EXTERNAL;
> > +    }
> > +
> > +    if (attr.value == VA_ATTRIB_NOT_SUPPORTED ||
> > +        attr.value == VA_ENC_QUANTIZATION_NONE) {
> > +        av_log(avctx, AV_LOG_WARNING, "Special Quantization attribute is
> not "
> > +                "supported: will use default quantization.\n");
> > +    } else if (attr.value == VA_ENC_QUANTIZATION_TRELLIS_SUPPORTED){
> > +        av_log(avctx, AV_LOG_VERBOSE, "Quantization Trellis supported.\n");
> > +
> > +        ctx->quantization_params = (VAEncMiscParameterQuantization) {
> > +            .quantization_flags.value = trellis,
> > +        };
> > +
> > +        vaapi_encode_add_global_param(avctx,
> > +                                      VAEncMiscParameterTypeQuantization,
> > +                                      &ctx->quantization_params,
> > +                                      sizeof(ctx->quantization_params));
> > +    }
> > +#else
> > +    av_log(avctx, AV_LOG_WARNING, "The encode quantization option
> (Trellis) is "
> > +           "not supported with this VAAPI version.\n");
> > +#endif
> > +
> > +    return 0;
> > +}
> > +
> >  static av_cold int vaapi_encode_init_gop_structure(AVCodecContext
> *avctx)
> >  {
> >      VAAPIEncodeContext *ctx = avctx->priv_data;
> > @@ -2132,6 +2174,12 @@ av_cold int
> ff_vaapi_encode_init(AVCodecContext *avctx)
> >      if (err < 0)
> >          goto fail;
> >
> > +    if (ctx->trellis) {
> > +        err = vaapi_encode_init_quantization(avctx);
> > +        if (err < 0)
> > +            goto fail;
> > +    }
> > +
> >      if (avctx->compression_level >= 0) {
> >          err = vaapi_encode_init_quality(avctx);
> >          if (err < 0)
> > diff --git a/libavcodec/vaapi_encode.h b/libavcodec/vaapi_encode.h
> > index eeec06036b..b24735da59 100644
> > --- a/libavcodec/vaapi_encode.h
> > +++ b/libavcodec/vaapi_encode.h
> > @@ -37,7 +37,7 @@ struct VAAPIEncodePicture;
> >
> >  enum {
> >      MAX_CONFIG_ATTRIBUTES  = 4,
> > -    MAX_GLOBAL_PARAMS      = 4,
> > +    MAX_GLOBAL_PARAMS      = 5,
> >      MAX_DPB_SIZE           = 16,
> >      MAX_PICTURE_REFERENCES = 2,
> >      MAX_REORDER_DELAY      = 16,
> > @@ -220,6 +220,9 @@ typedef struct VAAPIEncodeContext {
> >      // Packed headers which will actually be sent.
> >      unsigned int    va_packed_headers;
> >
> > +    // Quantization mode
> > +    int trellis;
> > +
> >      // Configuration attributes to use when creating va_config.
> >      VAConfigAttrib  config_attributes[MAX_CONFIG_ATTRIBUTES];
> >      int          nb_config_attributes;
> > @@ -256,7 +259,9 @@ typedef struct VAAPIEncodeContext {
> >  #if VA_CHECK_VERSION(0, 36, 0)
> >      VAEncMiscParameterBufferQualityLevel quality_params;
> >  #endif
> > -
> > +#if VA_CHECK_VERSION(1, 0, 0)
> > +    VAEncMiscParameterQuantization quantization_params;
> > +#endif
> >      // Per-sequence parameter structure
> (VAEncSequenceParameterBuffer*).
> >      void           *codec_sequence_params;
> >
> > diff --git a/libavcodec/vaapi_encode_h264.c
> b/libavcodec/vaapi_encode_h264.c
> > index d1427112ea..427fb6320e 100644
> > --- a/libavcodec/vaapi_encode_h264.c
> > +++ b/libavcodec/vaapi_encode_h264.c
> > @@ -72,6 +72,7 @@ typedef struct VAAPIEncodeH264Context {
> >      int sei;
> >      int profile;
> >      int level;
> > +    int trellis;
> >
> >      // Derived settings.
> >      int mb_width;
> > @@ -1233,6 +1234,8 @@ static av_cold int
> vaapi_encode_h264_init(AVCodecContext *avctx)
> >      if (priv->qp > 0)
> >          ctx->explicit_qp = priv->qp;
> >
> > +    ctx->trellis = priv->trellis;
> > +
> >      return ff_vaapi_encode_init(avctx);
> >  }
> >
> > @@ -1263,6 +1266,12 @@ static const AVOption
> vaapi_encode_h264_options[] = {
> >          { "cabac", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN,
> INT_MAX, FLAGS, "coder" },
> >          { "vlc",   NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN,
> INT_MAX, FLAGS, "coder" },
> >          { "ac",    NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN,
> INT_MAX, FLAGS, "coder" },
> > +    { "trellis", "Trellis Quantization",
> > +      OFFSET(trellis), AV_OPT_TYPE_FLAGS, { .i64 = 0}, 0, INT_MAX, FLAGS,
> "trellis"},
> > +        { "off",   NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN,
> INT_MAX, FLAGS, "trellis"},
> > +        { "I",     NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 2 }, INT_MIN,
> INT_MAX, FLAGS, "trellis"},
> > +        { "P",     NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 4 }, INT_MIN,
> INT_MAX, FLAGS, "trellis"},
> > +        { "B",     NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 8 }, INT_MIN,
> INT_MAX, FLAGS, "trellis"},
> >
> >      { "aud", "Include AUD",
> >        OFFSET(aud), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
> > --
> > 2.17.1
> >
> 
> Do you test this function for this feature with the other
> options/features, as I know, this feature have some limitation, I
> believe is a hardware or driver  limitation, and we enable this
> features about 2 years ago, but didn't submit the patch because the
> limitation

Thanks for the remind, I’ve no idea about previous patches on trellis, which is
not mentioned in the forked ffmpeg repo when I take this task.

> Some question:
> 1. How is the impact on encoder performance and image quality?

Enabling trellis will gain better quality in the low QP range.
During my test, enable trellis can increase ~1% quality (PSNR).
Please feel free let me know if you have any concerns to offer this option within spec and increasing the quality.

> 2. What's the limitation with the other VA-API features?
As far as I know, Quality gain is QP dependent, with lower QP value, we could see higher quality gain.
In BRC mode, it might be disabled, but I haven't noticed this during my test. 
We have the daily CI to ensure ffmpeg-vaaapi/qsv mainline quality, and pre-patch check-in service in github to ensure our patches quality. And I already pass the pre-checkin

> 3. Missed the docs
Will add docs.


Thanks,
Linjie


More information about the ffmpeg-devel mailing list