Author: Robert Hyatt
Date: 02:38:33 08/13/98
Go up one level in this thread
On August 13, 1998 at 01:35:03, Bela Andrew Evans wrote: >Hi Bob -- > >some quick questions for you. If a good assembly language programmer took your >crafty chess engine and rewrote it in machine language, what kind of speedup >would you estimate on different platforms, or specifically, the PCs of today? >50% faster, 100%, 200%? Or would the speedup be minimal? > >And how long would you estimate such a project would take? > >Thanks in advance, Bela Very difficult to answer. With Cray Blitz, Harry Nelson and I did this. But we did two things simltaneously: we converted fortran to assembly, but we also modified algorithms and ideas globally to fit the cray architecture. We ended up with a program 4.5 times faster. Doing this on the PC would take a lot of time, because it isn't just an issue of converting to assembly, it is also an issue of studying the X86 platforms and modifying the algorithms and data structures to better fit that machine. And that would be difficult when you look at how frequently the processor changes, from basic L2 cache speed/size, to the pipeline approach, to superscalar execution, even to some supporting register renaming to give you effectively a large group of registers... The speedup would depend on how "deep" someone wanted to go. IE, on the Cray, we moved the chess board into a single vector register and *never* touched that so the board was always instantly available. There were other things done like this (the repetition list and some things like that). IE this is *really* algorithm-specific/architecture-specific matching. And when Merced drops out year after next, the architecture will be so different this would get thrown away...
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.