[FFmpeg-devel] [PATCH 2/7] lavc/vaapi_encode_h264: Enable macro block level bit rate control.

[Mark Thompson]

On 03/11/18 02:52, Jun Zhao wrote:
> Enables macro block level bit rate control, that generally improves
> subjective visual quality. It may have a negative impact on
> performance and objective visual quality metrics. Default is off
> and can't compatible with Constant QP.
> 
> Signed-off-by: Jun Zhao <jun.zhao at intel.com>
> ---
>  libavcodec/vaapi_encode_h264.c |   21 +++++++++++++++++++++
>  1 files changed, 21 insertions(+), 0 deletions(-)
> 
> diff --git a/libavcodec/vaapi_encode_h264.c b/libavcodec/vaapi_encode_h264.c
> index 7bb77cf..060277b 100644
> --- a/libavcodec/vaapi_encode_h264.c
> +++ b/libavcodec/vaapi_encode_h264.c
> @@ -58,6 +58,7 @@ typedef struct VAAPIEncodeH264Context {
>      int sei;
>      int profile;
>      int level;
> +    int mb_rate_control;
>  
>      // Derived settings.
>      int mb_width;
> @@ -889,6 +890,22 @@ static av_cold int vaapi_encode_h264_configure(AVCodecContext *avctx)
>      return 0;
>  }
>  
> +static av_cold int vaapi_encode_h264_bit_rate_control(AVCodecContext *avctx)
> +{
> +    VAAPIEncodeContext      *ctx = avctx->priv_data;
> +    VAAPIEncodeH264Context *priv = avctx->priv_data;
> +
> +    if (priv->mb_rate_control) {
> + #if VA_CHECK_VERSION(0, 39, 2)
> +        ctx->rc_params.rc.rc_flags.bits.mb_rate_control = priv->mb_rate_control;
> + #else
> +        av_log(avctx, AV_LOG_WARNING, "The MB rate control option is not "
> +               "supported with this VAAPI version.\n");
> + #endif
> +    }
> +    return 0;
> +}
> +
>  static const VAAPIEncodeProfile vaapi_encode_h264_profiles[] = {
>      { FF_PROFILE_H264_HIGH, 8, 3, 1, 1, VAProfileH264High },
>      { FF_PROFILE_H264_MAIN, 8, 3, 1, 1, VAProfileH264Main },
> @@ -904,6 +921,8 @@ static const VAAPIEncodeType vaapi_encode_type_h264 = {
>  
>      .configure             = &vaapi_encode_h264_configure,
>  
> +    .bit_rate_control      = &vaapi_encode_h264_bit_rate_control,
> +
>      .sequence_params_size  = sizeof(VAEncSequenceParameterBufferH264),
>      .init_sequence_params  = &vaapi_encode_h264_init_sequence_params,
>  
> @@ -1001,6 +1020,8 @@ static const AVOption vaapi_encode_h264_options[] = {
>        OFFSET(qp), AV_OPT_TYPE_INT, { .i64 = 20 }, 0, 52, FLAGS },
>      { "quality", "Set encode quality (trades off against speed, higher is faster)",
>        OFFSET(quality), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, INT_MAX, FLAGS },
> +    { "mb_rate_control", "MB level bitrate control (only supported on GEN9+)",
> +      OFFSET(mb_rate_control), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS, "mb_rate_control" },
>      { "coder", "Entropy coder type",
>        OFFSET(coder), AV_OPT_TYPE_INT, { .i64 = 1 }, 0, 1, FLAGS, "coder" },
>          { "cavlc", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS, "coder" },
> 

Enabling this option still completely kills the GPU on my Skylake (6300) test machine (in I think older kernel versions this also caused the whole machine to hard-reset immediately, in current ones it only kills the GPU requiring a reboot to fix).  See <https://github.com/intel/intel-vaapi-driver/issues/172>.

- Mark


E.g., from 4.18.10:

[  138.818477] [drm] GPU HANG: ecode 9:0:0x8fd8ffff, in ffmpeg_g [595], reason: hang on rcs0, action: reset
[  138.819029] [drm] GPU hangs can indicate a bug anywhere in the entire gfx stack, including userspace.
[  138.819604] [drm] Please file a _new_ bug report on bugs.freedesktop.org against DRI -> DRM/Intel
[  138.820190] [drm] drm/i915 developers can then reassign to the right component if it's not a kernel issue.
[  138.820787] [drm] The gpu crash dump is required to analyze gpu hangs, so please always attach it.
[  138.821394] [drm] GPU crash dump saved to /sys/class/drm/card0/error
[  138.822022] i915 0000:00:02.0: Resetting rcs0 for hang on rcs0
[  146.811837] i915 0000:00:02.0: Resetting rcs0 for hang on rcs0
[  146.815773] [drm:gen8_reset_engines [i915]] *ERROR* rcs0: reset request timeout
[  146.817686] i915 0000:00:02.0: Resetting chip for hang on rcs0
[  146.819565] [drm:gen8_reset_engines [i915]] *ERROR* rcs0: reset request timeout
[  146.929115] [drm:gen8_reset_engines [i915]] *ERROR* rcs0: reset request timeout
[  147.041196] [drm:gen8_reset_engines [i915]] *ERROR* rcs0: reset request timeout
[  147.151935] i915 0000:00:02.0: Failed to reset chip
[  147.156055] [drm:gen8_reset_engines [i915]] *ERROR* rcs0: reset request timeout