Author: Dann Corbit
Date: 11:07:50 08/30/01
Go up one level in this thread
On August 30, 2001 at 13:56:53, Scott Gasch wrote: >I'm moving around data structures in my engine to consolodate things that are >going to be needed on a per-cpu basis if/when I go parallel. > >One such structure is my move stack. It's a big array of moves with a start and >end index per ply. So for example it might look like this: > >start[0] = 0 ... end[0] = 32 [array entries 0..32 hold moves at ply 0] >start[1] = 33 ... end[1] = 60 [array entries 33..60 hold moves at ply 1] >... > >Well if more than one thread is searching at once I will need more than one move >stack and more than one ply counter. So I kept the same move stack struct and >made g_MoveStack an array: > >MOVE_STACK g_MoveStack[NUM_CPU]; > >The code to access the move stack goes from this: > >g_MoveStack.iStart[g_iPly] = 0; > >to this: > >g_MoveStack[iCpuId].iStart[g_iPly[iCpuId]] = 0; > >Talk about a huge impact -- move move generator benchmark literally is 4x >slower! These dereferences are damn expensive. There has to be a better way, >can one of you assembly gurus give me a clue? > >Here is a solution I am thinking about -- have a struct per-thread that houses >the ply and a pointer to the start of the right move stack entry. Then do >something like this: > >THREAD_INFO *pThreadInfo = &(g_ThreadInfo[iCurrentThreadId]); >(pThreadInfo->pMoveStack)->iStart[sThreadInfo->iCpuId] = 0; > >I bet this is just as slow though... Any advice? Dumb question: Why not let them all have their own move generator and just share the hash table?
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.