Re: multiplication enhancement

[Kevin Ryde]

"Louis Helm" <address@hidden> writes:
>
> While using GMP in my project, Seventeen or Bust (http://sb.pns.net/), I
> noticed that the mul_basecase.asm used for K7s in v4.0.1 is faster on P6
> class processors than the generic mul_basecase.asm which is normally
> linked when compiling for pentium III procs.

Thanks, I'll look into that.  Not sure why there isn't already a p6
specific mul_basecase, probably just never got around to it.

> I am not an assembly expert so I'm assuming it is benefiting from SIMD
> instructions that are common to both and it's a significant speed
> improvement for our project (~20%) on pentium processors.

That k7 code doesn't use mmx actually, but it does have more loop
unrolling than the generic version.

> For reference
> the common multiplication done in SB involves the squaring of numbers
> with about 18,750 limbs so I'm assuming most of the basecase
> multiplication is coming from pointwise multiplications from large FFT
> multiplications.

Squaring should end up calling sqr_basecase rather than mul_basecase,
unless I'm horribly mistaken.

There's both p6 and k7 specific sqr_basecase routines, though not a
generic x86 one as it happens.