[FFmpeg-devel] Patch review request: 10bit DNxHD decoding support

[Maxim]

Maxim schrieb:
> Joseph Artsimovich schrieb:
>   
>> Hi,
>>
>> At this point I am happy with my patch adding support for decoding 10bit DNxHD, apart from poor performance due to the lack of assembly optimized IDCT code that can handle 10bit samples.  Please review and give feedback.
>>   
>>     
>
> I'm currently working on a support for the 10bit IDCT (both SSE2 and
> Altivec). Some minor modifications (regarding rounding constants) of the
> existing functions (idct_altivec.c and idct_sse2_xvid.c) are required.
> There are two minor issues:
>
> * it will be difficult to obtain the same transform precision using the
> functions mentioned above because of the higher data pass. A proprietary
> 10bit video codec I'm currently working on produces 14bits DCT
> coefficients instead of 12bits for the 8bit IDCT. I think this is a
> minor issue because all codes currently supporting 10bits are I-Frames
> only and doesn't rely on very exact IDCT.
>
> * the SSE2 engine is slow as hell on early Intel CPUs like mine (Intel
> Core Duo "Yonah") so a good-performance MMX2 version will be appreciated too
>   

Regarding 10bit IDCT: what does the VC-3 spec say about transform
precision? Does it give some metrics like mean error, mean square error
etc. ?

Best regards
Maxim P.