Le 14/02/2014 07:45, jhu a écrit :
> Apparently it does by about 4%. Not quite sure how or why though.
> 
> Intel Core i5 3317U (Ivy Bridge) @ 1.7 GHz (turbo 2.4 GHz), Ubuntu 12.04
> 
> gcc 4.8, -march=core-avix-i -mavx
>   Photon Time:      0 hours  0 minutes  2 seconds (2.558 seconds)
>               using 7 thread(s) with 2.981 CPU-seconds total
>   Radiosity Time:   No radiosity
>   Trace Time:       0 hours  7 minutes 58 seconds (478.870 seconds)
>               using 4 thread(s) with 1909.554 CPU-seconds total
> 
> gcc 4.8, -march=core-avx-i -msse2
>   Photon Time:      0 hours  0 minutes  2 seconds (2.607 seconds)
>               using 7 thread(s) with 3.055 CPU-seconds total
>   Radiosity Time:   No radiosity
>   Trace Time:       0 hours  8 minutes 18 seconds (498.217 seconds)
>               using 4 thread(s) with 1986.597 CPU-seconds total

There is a few operations (functions in povray) applied to vector (3 or
more double/float) which might enjoy such optimisation.

SSE2 is using only 128 bits of data (2 doubles), whereas AVX is able to
use 256 bits (4 doubles).

The operations are simple, but repeated a lot of time so grabbing a few
cycles per operation gives that change.


-- 
Just because nobody complains does not mean all parachutes are perfect.

Post a reply to this message