Re: [Chicken-users] performance of bignums

[Stephen Eilert]

On Thu, Jun 25, 2015 at 5:53 PM, Peter Bex <address@hidden> wrote:

On Thu, Jun 25, 2015 at 10:26:53PM +0200, Peter Bex wrote:
> On Thu, Jun 25, 2015 at 05:04:17PM -0300, Stephen Eilert wrote:
> > Not sure about the status of this particular GSOC, but SBCL could also be
> > "cheating".
> >
> > http://www.sbcl.org/gsoc2013/ideas/#sec-1.2
> >
> > Now, I thought GMP were GPL'd and SBCL not, so I'm unsure about the legal
> > implications, if it does bundle GMP now.
>
> Thanks for posting that; I hadn't looked at the sbcl implementation
> of the pi-digits benchmark, but indeed that page says "Pi digits on
> the benchmarks game is currently a bunch of calls to GMP".  This would
> *really* be cheating in my opinion, as it's not testing the language
> implementation but using nonstandard extensions that replace stuff
> that's already part of the core.
>
> I had just assumed it would be using the sbcl native implementation
> (which indeed is pretty damn good and almost 100% pure Lisp code).

In fact, there are two implementations, one of which uses native
operations.  It is still faster than CHICKEN and very close to Guile/GMP:

http://benchmarksgame.alioth.debian.org/u32/program.php?test=pidigits&lang=sbcl&id=2#log

It does use some type annotations, though.  Adding some to the
CHICKEN version makes it go a tiny bit faster, but not by much.

Cheers,
Peter

That compilation output is amazing!

; note: forced to do GENERIC-* (cost 30)
;       unable to do inline fixnum arithmetic (cost 4) because:
;       The first argument is a INTEGER, not a FIXNUM.
;       The result is a (VALUES INTEGER &OPTIONAL), not a (VALUES FIXNUM &REST T).
;       unable to do inline (signed-byte 32) arithmetic (cost 5) because:
;       The first argument is a INTEGER, not a (SIGNED-BYTE 32).
;       The result is a (VALUES INTEGER &OPTIONAL), not a (VALUES
;                                                          (SIGNED-BYTE 32) &REST
;                                                          T).

From the benchmark page (http://benchmarksgame.alioth.debian.org/u32/performance.php?test=pidigits#about):

"In addition to language specific multiprecision arithmetic, we will accept programs that use GMP."

So, the "cheating" is official. Let's cheat too!

I fail to see the point though. If it's a programming language shootout, does it make sense to offload the meat of the processing to GMP? Do they intend to measure the FFI performance?

— Stephen