GMP integration

Scott wrote:
> On Mar 27, 8:48 pm, Brad Lucier <luc...@math.purdue.edu> wrote: 

> I'm not saying you can't get it right.  All I claim is that a naive
> implementation by someone using an FFT to get fast convolution while
> ignoring the details that FFTs have noise from floating point error
> can get you an algorithm that is wrong occasionally.

I have faith in Lucier. He is everything but naive when it
comes to floating points.

"You're an evil man."
Linus Thorvalds on Lucier

The comment followed
http://gcc.gnu.org/ml/gcc-patches/2001-03/msg00463.html
which indeed is evil.

--
Jens Axel Søgaard

Brad Lucier skrev:
> On Mar 27, 5:26 pm, Scott <xsco...@gmail.com> wrote:
> 
>
> Error bounds that are good enough to multiply half-billion bit numbers
> using floating-point fft were given by Colin Percival around 2002.
> Gambit uses FFT multiplication with floating-point arithmetic, written
> in Scheme.  For multiplication, division, and square roots of numbers
> of about a million bits it's about half as fast as GMP, which is
> written in C and assembler.

Impressive!

--
Jens Axel Søgaard

On Mar 28, 1:58 pm, Jens Axel Soegaard <inva...@soegaard.net> wrote:
>
> I have faith in Lucier. He is everything but naive when it
> comes to floating points.
>

Yup, looks like I was teaching Gramma how to suck eggs on that
one...  :-)

Jens Axel Soegaard wrote:
> If the 1998 note still is correct, then Larceny has its own
> bignum implementation.
>
> http://www.ccs.neu.edu/home/lth/lar...arithmetic.html

Right.  Larceny does not use GMP.

Reliance on GMP would create portability problems, especially
for Common Larceny.  There wouldn't be anything wrong with
changing native or Petit Larceny so they could use GMP on
machines for which GMP has been installed, but I spent
a couple of days looking into that and found that Larceny's
bignums have other performance problems that would have to
be fixed first.  Once they're fixed, it looks like an
implementation of the Karatsuba algorithm in Scheme (which
I've already prototyped) would probably perform well enough
for most applications.  We'd like to continue to distribute
ready-to-run binaries that don't depend on GMP, even if we
develop a version that can use GMP.

Will

William D Clinger skrev:
> Jens Axel Soegaard wrote: 
>
> Right.  Larceny does not use GMP.
>
> Reliance on GMP would create portability problems, especially
> for Common Larceny.  There wouldn't be anything wrong with
> changing native or Petit Larceny so they could use GMP on
> machines for which GMP has been installed, but I spent
> a couple of days looking into that and found that Larceny's
> bignums have other performance problems that would have to
> be fixed first.  Once they're fixed, it looks like an
> implementation of the Karatsuba algorithm in Scheme (which
> I've already prototyped) would probably perform well enough
> for most applications.  We'd like to continue to distribute
> ready-to-run binaries that don't depend on GMP, even if we
> develop a version that can use GMP.

That makes a lot of sense. Thanks for the explanation.

--
Jens Axel Søgaard

William D Clinger wrote:

> We'd like to continue to distribute
> ready-to-run binaries that don't depend on GMP, even if we
> develop a version that can use GMP.

Wasn't there like a -static something flag for gcc that kind of
intended to solve that problem?  (I haven't used it myself)

Abdulaziz Ghuloum wrote:
> Wasn't there like a -static something flag for gcc that kind of
> intended to solve that problem?  (I haven't used it myself)

It's more of a licensing issue than a technical issue.
Some of Larceny's binary distributions are unencumbered
by LGPL, and that matters to some users.

Will

Hi,

William D Clinger <cesura17@yahoo.com> writes:

> It's more of a licensing issue than a technical issue.
> Some of Larceny's binary distributions are unencumbered
> by LGPL, and that matters to some users.

What do you mean by "unencumbered by LGPL"?  LGPL is *not* copyleft.

Thanks,
Ludovic.

Ludovic Court=E8s wrote:
> What do you mean by "unencumbered by LGPL"?

Read http://www.gnu.org/licenses/lgpl.html

> LGPL is *not* copyleft.

Understood.  No part of Larceny is encumbered by GPL.
Only one part is even encumbered by LGPL.  Two of the
binary distributions are encumbered by LGPL.  The
others aren't.  If this matters to you, which seems
unlikely considering your unfamiliarity with LGPL,
then you should read the COPYRIGHT file.

Will

Ludovic Courtès wrote:

> William D Clinger <cesura17@yahoo.com> writes:
> 
>
> What do you mean by "unencumbered by LGPL"?  LGPL is *not* copyleft.

You don't have to be encumbered by copyleft.  My brain can be
enough encumbrance at times.  :-)

Anyways, in a previous message, Will Clinger stated that:

"Petit Larceny is not encumbered by GPL
or LGPL or BSD or apache licenses."

Given the number of licenses and people in the COPYRIGHT file
that comes with the larceny distribution (501 lines), it makes
me wonder what the implications of these licenses are.  But as
per Will's comments, I'm sure had thought it all out.

Aziz,,,