Author: Robert Hyatt
Date: 11:05:27 09/03/03
Go up one level in this thread
On September 03, 2003 at 13:30:26, Gian-Carlo Pascutto wrote: >On September 03, 2003 at 13:18:49, Robert Hyatt wrote: > >>>Each time you get a subtree score, you must send out the score update to >>>all processors, or store it locally and rely on remote processors to check >>>it in your memory. >>> >>>Either way, you need remote accesses. >>> >>>Got it now? >>> >>>-- >>>GCP >> >>No, because you don't have to do that. You _can_ ask "did my score appear >>to change alpha/beta bounds at the split point?" If the answer is no, you >>do nothing. That is the case 99% of the time. > >Yeah yeah, that case was implicit. The point of the thread is that it's >not a simple matter of start thread, copy date....end thread, write data >back. You must do some remote access always. Yes, but the "quantity" can be limited. IE I copy more data _to_ a thread than I copy back when I finish the thread. you have to make sure the search work done far outweighs the time spent in starting/ending things. But that is not hard to do. > >In the example above, you'd still need to get the next move or put yourself >in the idle list, even though updating the score wouldn't be needed. Getting the next move doesn't count. Given there are 35 moves in a normal position, even a 2 ply search beyond that move dwarfs the time needed to access a non-local memory area to get the next move. > >Depending on what you do, you may want to check if alpha didn't change >as well (not needed if you abort such trees). > >-- >GCP
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.