Author: Dann Corbit
Date: 10:10:49 10/17/01
Go up one level in this thread
On October 15, 2001 at 14:23:46, Frank Phillips wrote: >On October 15, 2001 at 05:56:20, David Blackman wrote: >>> I am predicting double the speed. >>> >>>Note that I am not talking about the KAP compiler for Linux (but rather) the >>>Intel C++ compiler 6.0 for Linux (which is now in beta). >> >>Double would be bloody amazing for chess code. I could understand it some kinds >>of floating point on the grounds that: >> >>1. Floating point on the X87 is really ugly >>2. Gcc doesn't try very hard >>3. If anyone can figure out how to make this mess go fast, it should be Intel. >> >>For the integer stuff, gcc is not great, but it's not all that bad either. It's >>mostly about half way between Borland and Microsoft, though occasionally it >>manages to beat both. I haven't heard anyone claim the Intel compiler is twice >>as good as that. I've spent a fair while staring at inside loops of various >>stuff generated by gcc, and in most cases, there is no way anything is going to >>go twice as fast on the same chip. >> >>That said, i will probably try to get hold of this if it isn't to huge. > >After something of a struggle (my ISP automatically disconnects after two hours >and at less than 56k/s it takes longer) I finally got the Intel compiler down >and working. > >I am impressed. It is certainly not twice as fast but faster. On my program >and the little fixed ply search test I do to check that I still get the same >number of nodes when they should not have changed, the comparison is: >gcc2.96 403knps >intel 467knps >(MSVC++ 522knps) > >The executable has increased in size to 830kB. If I link it statically it goes >to 2.52MB. > >The executables use Intel libraries, so you have to set up the environment >variables. If anyone knows how to do this at boot, rather than through the >users.bashrc file, please let me know. > >Anyway, thanks for the tip Dan. If the Linux version of the C++ compiler works like the PC version, there are a lot of options to play with. Especially important are trying various combinations of target chip, iter-procedural optimization, and whole-program optimization. You can get quite a bit more out of it than is delivered by the default settings. Using profile guided optimization and a link-order list is also worth a try. By various combinations, you might be able to get a much larger boost.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.