Author: Robert Hyatt
Date: 06:13:29 09/28/98
Go up one level in this thread
On September 28, 1998 at 03:09:32, Danniel Corbit wrote: >On September 27, 1998 at 18:15:36, Robert Hyatt wrote: >[snip] >>Depends on how far you take it. Cray Blitz... pure fortran (using the >>very best fortran compiler on the planet, too) is 5x *slower* than the >>hand-coded assembly language version. Different approach in the asm >>version, keeping things in registers longer, doing things that are >>architecture specific, etc... >> >>But it can make a big difference. >Do they have a KAP preprocessor for the CRAY? If so, was that used? I have >seen it make a 2 or 3x improvement in some VAX multi-cpu code. they were the first customer of Ken Kennedy's group. But KAP does *nothing* for a recursive tree search, because it doesn't recognize how to "unroll" the recursion to get parallel results. I tried this on Cray Blitz once when Dennis Moen (Cray compiler/library guru) asked me to, and on a 4 processor XMP,cray blitz ran 1.01 times faster. It found ways to parallelize the move input/output code because there are loops there. But it was totally lost in the alpha/beta search which is where *all* the work goes on.
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.