Jukka Larja (roskakori2006.delete@this.gmail.com) on October 30, 2020 11:29 am wrote:
> Mark Roulo (nothanks.delete@this.xxx.com) on October 30, 2020 10:21 am wrote:
> > Jukka Larja (roskakori2006.delete@this.gmail.com) on October 30, 2020 9:52 am wrote:
> > > me (me.delete@this.me.com) on October 30, 2020 9:47 am wrote:
> > > > > And I presume you are talking about AVX2. Both PS4 and Xbox One support AVX,
> > > > > and that alone obviously didn't gather any significant support (even though
> > > > > consoles have much more benefit from CPU optimizations compared to PC).
> > > > >
> > > > > -JLarja
> > > >
> > > > IIRC Jaguar's AVX isn't any faster than using SSE so devs mostly ignored it I guess.
> > >
> > > Is any AVX faster than SSE?
> > >
> > > -JLarja
> >
> > For my employer’s loads, AVX2 runs about 1.7x faster than SSE for compute bound loads.
>
> Yes, but that's AVX2, which is not available on Jaguar.
>
> -JLarja

It is AVX2 because many of our loads are short integer loads.

AVX won’t be faster on Jaguar than SSE because the Jaguar vector units are 128 bits wide so the AVX instructions needs roughly 2x the clocks per instruction which cancels out doing 2x more work per op.


I would expect AVX implementations that are 256 bits wide (eg. Sandy Bridge) to run ~1.7x faster than SSE on floating point code.

The slowdown is AMD specific (and only on some cores), not inherent with AVX.