[Libre-soc-dev] v3.1B prefix

Luke Kenneth Casson Leighton lkcl at lkcl.net
Mon Dec 21 17:53:19 GMT 2020


On Mon, Dec 21, 2020 at 5:18 PM Alexandre Oliva <oliva at gnu.org> wrote:

> > are you *sure* that's right?
>
> Well, yeah, I was comparing apples to apples.  The numbers from my Dec
> 13 post that I quoted and compared were from before your cregs* changes,
> so the comparison with the alternate mode-switching logic used the same
> cregs* from before your change.

ok got it.  i'm running on an objdump of /bin/bash, the number drops
by *4*% down to 0.86 when using these:
cregs3 = { 0, 9, 1, 2, 3, 31, 10, 30, 4 }

and if it's extended to 16 regs it drops to 83.7%.
cregs3 = { 0, 9, 1, 2, 3, 31, 10, 30, 4, 29,8,5,27,26,12,7 }

assuming that cregs2 can be extended by 1 extra bit as well, that
drops to 82.6%.

due to the square-law effect that means a 31% reduction in cache power
consumption is achievable.  that's... significant.

i'll try on /usr/bin/gcc and perhaps libc6 as well to see if the same
sorts of numbers stack up.  nice idea, with the bitpattern, alexandre.

l.



More information about the Libre-soc-dev mailing list