[FFmpeg-devel] [PATCH] MMX implementation of VC-1 inverse transforms

[Christophe GISQUET]

Hi,

Michael Niedermayer a ?crit :
[...]
> for the first idct yes it should fit in 16bit without tricks

I'm not sure it is worth to have 2 macros for dct4 and 2 functions for
dct8 using that fact to reduce computations. Or maybe as an incremental
patch if it's worth it.

[...]
> before >>7:
> -65535 -65535  -65535  -65535
> -65535 -65535  -65535  -65535
> -65535 -65535  -65535  -65535
> -65535 -65535  -65535  -65535

It looks safe to only do one preshift of value 1 for every 1d transform,
like what you proposed for the dct4. This greatly simplifies what should
be done on the 1d dct8

I'll spawn another subthread about the 1d dct8 in reply to your very
first mail, also replying your other comments I haven't addressed yet.

Best regards,
-- 
Christophe GISQUET