Author: Bo Persson
Date: 03:11:31 05/17/98
Go up one level in this thread
On May 15, 1998 at 01:51:18, Ivan Tirado wrote: [snip] > Thanks for the reply. Yes, you're right. I have seen this behaviour >myself with the add - substract on floating point numbers. Remember that >multiplication is basically fast addition, so an algirithim may still >benefit from FP multiply, although not from FP divide. This, of course >from a mathematical standpoint, since the FP internals may alter the >outcome ( may or may not be faster, platform dependent ). Now, since >you've stated that chess programs use a lot of add-substracts with >integers, wouldn't MMX give a BIG speedup to the eval? Or is coding for >MMX too hard/inconvenient? I believe that an eval function optimized for >MMX would vastly outperform a similar algorithm in a non MMX setup. >Would you please comment on this? The problem with using MMX instructions is that they are not yet supported by the compiler optimizer. I have tried using inlined asm MMX-instructions with MSVC. The result was that although the bitmap ops ran faster, the other instructions were less well optimized by the compiler. Particularly register allocations were a lot more conservative. This led to additional load/store of variables, negating the effect of the MMX ops. Bo Persson bo.persson@mbox3.swipnet.se
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.