4.2 running slower than 4.1 on Pentium4, ideas?

Brian Gladman brg at gladman.plus.com
Thu Apr 24 11:45:53 CEST 2008

----- Original Message ----- 
From: "Rev. Chris Korda" <victimofleisure at gmail.com>
To: <gmp-discuss-owner at swox.com>; <gmp-discuss at swox.com>
Sent: Wednesday, April 23, 2008 11:12 PM
Subject: 4.2 running slower than 4.1 on Pentium4, ideas?

My tests show that for x86/Pentium4, 4.2 runs slower than 4.1, and
presumably this is not the expected result. There are lots of differences so
it could take me a long time to discover which ones are causing the problem.
Perhaps others on this list who are more familiar with GMP will be able to
suggest some likely causes.

You cannot expect GMP to run at its best speed unless you tune it. There are 
quiite a few configuration parameters that are used to set the points at 
which algorithms are switched and these depend on the particular processor 
configuration in use if optimum performance is to be achieved.

If you take one of my ports from several years ago and run it on a modern 
architecture without tuning its default parameters you cannot then expect to 
get optimum performance.  This does not mean that there are errors in the 
port, it simply means that you have not tuned it properly.

In fact I only develop the Windows related 64-bit GMP assembler code now so 
the 32-bit assembler code has not been touched for several years and this 
means that the default tuning that I have left in place is very outdated and 
virtually certain to be wrong for current 32-bit processor configurations.

This is the GMP discussion list, yes? If my post is inappropriate for some
reason it would be helpful to at least say so, or better yet explain *why*
it's inappropriate. Otherwise I have no idea what's wrong and no way of
fixing it. I'm trying to help improve GMP, isn't that how open source
projects are supposed to work?

My guess is that you have not adjusted the GMP algorithm thresholds to match 
your processor configuration.

    Brian Gladman

More information about the gmp-discuss mailing list