[FFmpeg-devel] [PATCH 1/2] lavc/vaapi_encode: correct the HRD buffer size calculate.

[Mark Thompson]

On 31/10/17 16:00, Jun Zhao wrote:
> On 2017/10/31 23:45, Mark Thompson wrote:
>> On 31/10/17 15:11, Jun Zhao wrote:> On 2017/10/31 18:19, Mark Thompson wrote:
>>>> On 31/10/17 02:37, Jun Zhao wrote:
>>>>> From 7eef9be1c8a92bf625d62a0f97f762f1342c6d78 Mon Sep 17 00:00:00 2001
>>>>> From: Jun Zhao <jun.zhao at intel.com>
>>>>> Date: Tue, 31 Oct 2017 10:13:42 +0800
>>>>> Subject: [PATCH 1/2] lavc/vaapi_encode: correct the HRD buffer size calculate.
>>>>>
>>>>> when rc_buffer_size didn't setting, always use the max bit rate
>>>>> per second as HRD buffer size.
>>>>>
>>>>> Signed-off-by: Jun Zhao <jun.zhao at intel.com>
>>>>> Signed-off-by: Wang, Yi A <yi.a.wang at intel.com>
>>>>> ---
>>>>>  libavcodec/vaapi_encode.c | 21 ++++++++++-----------
>>>>>  1 file changed, 10 insertions(+), 11 deletions(-)
>>>>>
>>>>> diff --git a/libavcodec/vaapi_encode.c b/libavcodec/vaapi_encode.c
>>>>> index 590f4be4ed..d5f89ef346 100644
>>>>> --- a/libavcodec/vaapi_encode.c
>>>>> +++ b/libavcodec/vaapi_encode.c
>>>>> @@ -1144,19 +1144,9 @@ static av_cold int vaapi_encode_init_rate_control(AVCodecContext *avctx)
>>>>>          return AVERROR(EINVAL);
>>>>>      }
>>>>>  
>>>>> -    if (avctx->rc_buffer_size)
>>>>> -        hrd_buffer_size = avctx->rc_buffer_size;
>>>>> -    else
>>>>> -        hrd_buffer_size = avctx->bit_rate;
>>>>> -    if (avctx->rc_initial_buffer_occupancy)
>>>>> -        hrd_initial_buffer_fullness = avctx->rc_initial_buffer_occupancy;
>>>>> -    else
>>>>> -        hrd_initial_buffer_fullness = hrd_buffer_size * 3 / 4;
>>>>> -
>>>>>      if (ctx->va_rc_mode == VA_RC_CBR) {
>>>>>          rc_bits_per_second   = avctx->bit_rate;
>>>>>          rc_target_percentage = 100;
>>>>> -        rc_window_size       = 1000;
>>>>>      } else {
>>>>>          if (avctx->rc_max_rate < avctx->bit_rate) {
>>>>>              // Max rate is unset or invalid, just use the normal bitrate.
>>>>> @@ -1166,8 +1156,17 @@ static av_cold int vaapi_encode_init_rate_control(AVCodecContext *avctx)
>>>>>              rc_bits_per_second   = avctx->rc_max_rate;
>>>>>              rc_target_percentage = (avctx->bit_rate * 100) / rc_bits_per_second;
>>>>>          }
>>>>> -        rc_window_size = (hrd_buffer_size * 1000) / avctx->bit_rate;
>>>>>      }
>>>>> +    rc_window_size = (rc_bits_per_second * 1000) / avctx->bit_rate;
>>>>> +
>>>>> +    if (avctx->rc_buffer_size)
>>>>> +        hrd_buffer_size = avctx->rc_buffer_size;
>>>>> +    else
>>>>> +        hrd_buffer_size = rc_bits_per_second;
>>>>> +    if (avctx->rc_initial_buffer_occupancy)
>>>>> +        hrd_initial_buffer_fullness = avctx->rc_initial_buffer_occupancy;
>>>>> +    else
>>>>> +        hrd_initial_buffer_fullness = hrd_buffer_size * 3 / 4;
>>>>>  
>>>>>      ctx->rc_params.misc.type = VAEncMiscParameterTypeRateControl;
>>>>>      ctx->rc_params.rc = (VAEncMiscParameterRateControl) {
>>>>> -- 
>>>>> 2.14.1
>>>>>
>>>> Why?  If the RC buffer size is unset it currently guesses one second of the target bitrate - in what way is the peak bitrate any more appropriate as a guess?
>>> In VBR mode, when rc_max_rate more than bit_rate, I think HDR size need
>>> to use rc_max_rate, not the bit_rate.
>> Ok, why do you think that?  I'm not necessarily against this change (in cases where it matters the user will provide rc_buffer_size explicitly), but please provide some reasoning for it.
> I think the code logic have considered the explicit rc_buffer_size
> setting, and this patch
> correct the HRD buffer size with MAX(rc_max_rate, bit_rate) when no
> setting rc_buffer_size in
> VBR mode.

I agree that this is what your patch does.

>           In the old logic, in VBR mode, if user don't setting
> rc_buffer_size
> explicitly, the HRD buffer size always use avctx->bit_rate, it does not
> make sense.
Why not?

- Mark