Author: Robert Hyatt
Date: 11:04:39 10/01/98
Go up one level in this thread
On October 01, 1998 at 13:14:36, Roberto Waldteufel wrote: > >On October 01, 1998 at 08:12:57, Robert Hyatt wrote: > >>On October 01, 1998 at 01:58:58, Georg Langrath wrote: >> >>>If 32 bits are more effective than 16, does this go on? I mean are 64 bits >>>better than 32, 128 bits better then 64 and so on. What is the optimum? >>> >>>Georg >> >> >>Easy question. The optimum word size is the word size that moves around >>the exact number of bits you need. IE in crafty, everything is based on >>64 bit words, so 64 bit machines are more effecient for me. 128 would >>buy nothing for me without some re-design... > >Hi Bob, > >Would it not be possible to improve efficiency a lot by doing two 64-bit >operations in a single instruction, like MMX does two instructions at once on >pentiums? Or you could use all sorts of new data structures, like an "etended" >bitboard that allows four different settings per square. Certainly you would >need to rethink some design issues, but I bet you would find ways to make use of >all those extra bits, and even if you didn't, your compiler might find ways to >optimise for 128 bits all on its own. I think the more bits, the better the >possibilities, but it needs careful design to make the most of what you have >available. Look how long it took for software to really start using the full >32-bit instruction set on the intel processors, and even now there is a lag with >respect to MMX usage on pentiums. > >I find the most difficult aspect of code optimisation has been the operating >system (Windows), which I do not properly understand, so I find it hard to >figure out efficient ways to use memory, but for actual nuts-and-bolts stuff >like writing an evaluation function for example, it pays to use assembler and >make maximum use of the (few) registers available. When I program in assembler I >have dreams about bigger registers...:-) > >Roberto all I can say is that if you have 64 bits of data, and you have a processor that can move 64 bits around easily, it is a win. If you take the "commercial" engines and port them to the alpha, they won't get much of anything from the port, other than the alpha is a better overall processor (speed-wise) than the intel chips. If DEC released a 128 bit processor (really not very likely for lots of reasons) it would not help me a bit at present. I would probably start by finding ways to combine two 64 bit bitmaps into one 128 bit word, for example. IE the occupied squares bitmap + the 3 rotated versions would likely end up in two words, which would win lots of ways... but it would take some re-designing...
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.