Author: Stuart Cracraft
Date: 14:32:05 10/09/04
Go up one level in this thread
On October 09, 2004 at 13:05:32, Gerd Isenberg wrote: >On October 09, 2004 at 10:02:13, Stuart Cracraft wrote: > >>On October 09, 2004 at 09:58:38, Gerd Isenberg wrote: >> >>>On October 08, 2004 at 20:03:43, Zach Wegner wrote: >>> >>>>I don't know how to get WAC 141 in one second, but I know some things in your >>>>code to improve it. First, you should not dynamically allocate storage for moves >>>>(do not declare the array in search()), but rather have a large array of size >>>>MAXMOVES * MAXPLY (it can probably be less) that is indexed by a pointer array >>>>indexed by ply. >>> >>>I prefere your approach too, one huge move array and keeping index or pointer >>>per ply, but don't confuse Stuart's local move array on the stack with >>>dynamically allocated storage, which is usually got with alloc/malloc or >>>operator new[] and freed with operator delete[] in C++. >>> >>>Keeping things on the stack frame is only about sub esp,LOCAL_SIZE. >>>That's not that bad - at least there are no allocation/freeing costs. >>> >>>Gerd >>> >>> >>>>If this seems confusing, heres some pseudo-code: >>>> >>>>mv movestack[MAXMOVES * MAXPLY]; >>>>mv *firstmove[MAXPLY]; >>>>... >>>>firstmove[ply + 1] = gen(firstmove[ply]); >>>> >>>>This requires your gen() to return a pointer to the element after the last move >>>>used, which shouldn't be too difficult. >>>> >>>>Second, it seems you are not using fractional extensions the way most people do. >>>>The depth parameter is measured in some constant > 1 that is proportional to a >>>>ply. I suppose floats could be used, but is important that the depth parameter >>>>is not an int with 1 equal to a ply. The idea is that for some extensions you do >>>>not want to extend a whole ply for just one occurence, but add a little bit of a >>>>ply to the depth that could help trigger an extension later. As an example, at >>>>ply X you have a condition met and you want to extend a half of a ply. Then at >>>>ply X + 2(say) you have the same (or other) condition met, and you extend a half >>>>of a ply again. The net extension is then just one ply, while in your >>>>implementation it would not be extended because the half ply would be rounded >>>>off at each ply. Or maybe I just misunderstood your code... >>>> >>>>Regards, >>>>Zach >> >>Okay if I hear you right, I can get rid of memory allocation and deallocation >>calls by maintaining my own move stack. How much is that really worth? > >Difficult to say, probably negligible and not worth to change it. > >The advantage of an own move stack is that you can keep it gap-less and >therefore you may need less cachelines and memory resources. OTOH you need an >additional pointer/index per ply. > >Again, your local move[MAXMOVES] array on the stack is not as far expensive as >dynamic memory allocation on the heap via operator new[] and operator delete[]. > >But you allways reserve MAXMOVES for each ply - most often not needed at all. >Therefore you have some gaps on your stack, which makes it less efficient in >terms of cachelines and 4 KPages. > >Gerd Seems like this move stack would be best for the next version of the program in a year or two as it won't be me much now in clarity or speedup.
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.