[Openmcl-devel] Bignum speed

[R. Matthew Emerson]

On Oct 17, 2008, at 1:22 AM, R. Matthew Emerson wrote:
>
> On my machine, I get 2.450994 seconds on the x86-64 lisp, and on the
> experimental x86-32 lisp, I get 0.344 seconds.  (Just for grins, I can
> report that an old PowerBook G4 gets 0.909 seconds.)  So, yes,  the
> x86-64 lisp looks pretty bad.
[...]
>
> Anyway, that's just additional evidence to support gb's prior message
> that there's plenty of room for some straightforward tuning.  If
> tuning the lisp code doesn't provide good enough results, it might not
> be too hard to adapt the x8632 assembly code for x8664.  Maybe if I
> get some free time...


This bugged me enough that I went ahead and improved the x8664 bignum  
multiply performance by replacing the inner loop of the multiply- 
bignums routine with a lap function that I adapted from the x8632  
version.

This improves performance on that mini-benchmark to 0.33 seconds,  
basically at parity with the experimental 32-bit lisp.  Not  
outstanding, but at least we won't be laughed at by all the other kids.

(This is in the trunk only, by the way.  The changeset is http://trac.clozure.com/openmcl/changeset/11111 
  for anyone following along at home.)