2012/2/28 Kostya Shishkov <[email protected]>:
> also lms_predict is really straightforward - you just obtain a scalar product
> and shift it, lms_update() is a bit harder but also quite trivial

The part of lms_predict that would be dspified is 2.5%. Still worth it
but not as important. mclms_update also, but again the loops are of
quite lesser importance.

>> Most operations seems to involve int16_t, so dspizing should not be
>> very difficult (famous last words).
>
> that's because it's 16-bit only for now :P
> in reality even reference decoder uses two sets of functions - with 16- and
> 32-bit precision (which of course can be DSPised).

16x16->32 multiplies available for x86 since first MMX set, but the
only multiply that takes a 32bits input on x86 is there only since
SSE4, which is quite limiting.

Anyway, I indeed haven't be cautious in formulating this report, as I
think indeed dsp discussions should not slow the code getting in.

Christophe
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to