[FFmpeg-devel] MPEG-2 Acceleration Refactor

[Ivan Kalvachev]

2007/6/22, Michael Niedermayer <michaelni at gmx.at>:
[...]
>
> i am not saying how to solve this, the original intent was some hardware
> acceleration support for mpeg1/2 with specific hardware _IIRC_, iam just
>
> saying how not to do it, which is
> * no slowdown for the normal case
> * no messy code which also means not replacing large functions by large
>   macros

How about placing the new code under #ifdef HAVE_XVMC #endif and not
bother refactoring the coefficient reading functions at all?
In normal case it would not be defined, in other case hw acceleration
would be used and it should be way faster:)
Afaik they are going to do their own version of the key xvmc functions.

On another hand, I wonder if software idct routine that takes the
coefficients in the same form would be faster. I think of something
with lookup table and SIMD. Well not entire idct function, but at
least the first half  (row stage).