Hi René,
David,
Confession: I have not tested 19x19. As you have noted, and others
before you over the years, a 19x19 board does not fit in one but three
128-bit registers and there would be a rather big penalty as a result,
perhaps (likely?) wiping out all of the benefits of bitmaps. Antoine
voiced his disappointment about the speed advantage even on 9x9 in the
e-mail to the list that served as my basis. My hope, as Hideko Kato
points out, is in longer registers or perhaps being able to port this
to a GPU. A 64-bit OS with twice the number registers would also
surely help out. Nevertheless, I was surprised that I was able to get
to almost the same level of speed as Libego provides.
As far as I remember, I was not disappointed by the speed itself on 9x9
boards, but mainly with the following 2 points:
a) my feeling is that, as you say, it does not scale very well on 19x19
(on current hardware).
I don't think other "classical" implementations suffer such a big
penalty when scaling from 9x9 to 19x19.
b) I also felt that this kind of implementation was not very flexible.
For instance, I had another classical implementation, running at
equivalent speed, but in which local 3x3 pattern matching was almost for
free, as well as other more elaborated information.
When I started to think about introducing 3x3 local patterns in the
bitmap only implementation, it was clear it would not be for free.
At that time, my conclusion was that if one only needs pure random play
with no intelligence at all during playouts, then bitmap implementation
could compete (at least on 9x9).
If one needs more elaborated knowledge (such as small patterns matching,
or even knowledge about blocks of stones), then pure bitmap
implementation is probably not so competitive.
I thus gave up with the idea and jumped to more promising ones.
Anyway, I'm glad my post has been usefull to you. And I encourage you to
improve your implementation and let us know, especially if you have fun.
Starting with something and playing with it is a good way to have new
ideas, even if this makes your initial ones look less interesting a
while after.
Best regards,
Antoine
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
