[FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V mspel_pixels

[Rémi Denis-Courmont]

Le sunnuntaina 5. toukokuuta 2024, 12.18.56 EEST flow gg a écrit :
> > Does MF2 actually improve perfs over M1 here?
> 
> The difference here seems very small, but when both mf2 and m1 are correct,
> the test results have only shown mf2 to be better, so I want to use mf2.

I can live with that. But this is a slippery slope because large vector sizes 
would involve even smaller fractions. Then we would need to compute the value 
which might negate the performance gains from fractional multipliers.

The fastest approach that I can think of is a symbolic LA (which expands to 
1xAUIPC + 1xLA) to load a precomputed VTYPE value from a static variable. 
Furthermore, this requires VSETVL, which precludes immediate constant VL 
Indeed, the VSETIVL instruction does not exist.

AFAIU, BananaPi F3 has 256-bit vectors already now.

-- 
Rémi Denis-Courmont
http://www.remlab.net/