Author: Vincent Diepeveen
Date: 02:57:37 12/08/00
Go up one level in this thread
On December 07, 2000 at 15:20:02, Roy Eassa wrote: >But what impact, if any, does floating point have on such chess apps as Crafty, >Chess Tiger, Fritz, Junior, Hiarcs, Shredder, et. al.? (E.g., which would help >them more: a 50% boost in integer speed only, or a 200% boost in floating speed >only?) First of all you need to be completely assembly to make usage of floating point. secondly what INSTRUCTIONS from FPU do you plan to use? I guess doing everything in FPU in C would slow down a program 4 times or so? Just keeping hashing in the FPU casted to MMX is of course a possibility. That would be the only possible usage of MMX/FPU as far as i can see. But does 0.5% speedup makeup for rewriting your entire program? Other instructions don't make sense simply from MMX. Some values can sure be stored there, but that's basically interesting for assembly programs, NOT for C compiled programs. I am writing in C, because i want to be able to easily read my code. I could already speedup my code 5 to 10% by rewriting some 32 bits datastructures to 8 bits, i get again huge casting problems etcetera etcetera, but i speed it up bigtime. But i chose to make my program 32 bits completely to get rid of all those casting problems and just be busy with the most important: improving the quality of the program, being busy algorithmically and being busy with its evaluation, without worrying always about the casting problems which sure get there if you have 2 mb of C source code! > >On December 07, 2000 at 15:11:55, Robert Hyatt wrote: > >>On December 07, 2000 at 14:53:11, Roy Eassa wrote: >> >>>5+ years ago, in another chess computer forum, I recall posting that computer >>>chess makes essentially no use of floating point arithmetic. I.e., the >>>performance of the CPU's integer arithmetic was overwhelmingly more important to >>>its performance running chess software than its floating point performance. >>> >>>Is that correct in today's top programs? Was I even correct back then? >> >> >>back then it depended on the architecture. IE Cray was no slouch at FP >>arithmetic, and in many cases FP was a fast or faster than int math. In >>today's PC, FP is slower, but with the multiple-pipe superscalar approach, >>some FP could be beneficial as those operations could be done in parallel >>with int operations.
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.