[I do not know what happens, but this is the second time the
computer-go archives miss the body. Maybe its the attachment. I
apologize for sending this a second time, this time without
attachment. If someone knows what the problem is ... I am using Gmail]
Antoine,
I apologize for misrepresenting your feelings. What I wanted to convey
is that your e-mail expressed that with the same amount of
computation, other implementation strategies provide more information,
so the information gained for the effort put in is disappointing. In
other words, bitmap-go had better be much faster to make up for the
lack of information gained ... and it wasn't. That's pretty much what
you are saying as well, I believe. As are the others in this thread. I
think we can all agree on this.
Going into this project, I was well aware that I was going to end up
with light playouts only, and that heavy playouts is the way to go
except perhaps for some special purposes such as ladder readouts. I
was also well aware of the scaling argument. But, I hadn't thought
that this would be the first thing thrown at my first post, as there
are many programmers that seem to be stuck at 9x9. Nonetheless,
bitmap-go as a topic keeps resurfacing on this mailing list every once
in a while and nobody ever put solid data and a reference
implementation behind it. That is what I wanted to accomplish with my
mockup.
Confession #2: I have been working on my own MCTS implementation using
the standard implementation methods that almost everybody else is
using. But there is a never-ending laundry-list of things that I would
like to include before I would reach reasonable strength (where
reasonable is a moving target). In the mean time, there are many
others that have demonstrated much more capable programmers than I
ever will be. So, by providing this bitmap-go mockup at least I had
the feeling that I was contributing something newsworthy to the
conversation. This may have never happened if I would wait until my
MCTS program is ready. I imagine that there are others on this list in
a similar situation. Besides, this was an interesting side-project and
perhaps someone else can benefit from it (go for it Brian!). And, yes,
it was fun.
Okay, enough mesmerizing, on to the main topic.
David, Lukasz,
I did modify my mockup to do 19x19 bitmap-go (attached). It is a
hardcoded solution using arrays of three 128-bit variables. I did not
even attempt to optimize this version, so this is not the best
possible solution. Nonetheless, here is a comparison:
Example output (9x9):
=====================
[game] = 30236.1, [moves] = 111.071
[game] = 30249.7, [moves] = 111.068
[game] = 30145.7, [moves] = 111.089
[game] = 30237.7, [moves] = 111.122
[game] = 30210.1, [moves] = 111.101
[game] = 78.0488 kpps, [moves] = 111.023
[game] = 78.0488 kpps, [moves] = 111.045
[game] = 78.0488 kpps, [moves] = 111.046
[game] = 79.0123 kpps, [moves] = 111.131
[game] = 78.0488 kpps, [moves] = 111.082
[legal] 110/51, [pick] 110/74, [play] 106/168, [score] 44, [win] 0.4187
[legal] 111/51, [pick] 111/74, [play] 106/168, [score] 40, [win] 0.4201
[legal] 111/52, [pick] 111/75, [play] 106/168, [score] 42, [win] 0.4276
[legal] 111/52, [pick] 111/74, [play] 106/169, [score] 41, [win] 0.4092
[legal] 111/52, [pick] 111/74, [play] 106/169, [score] 44, [win] 0.4221
Example output (19x19):
=======================
[game] = 316452, [moves] = 455.036
[game] = 316378, [moves] = 455.126
[game] = 316266, [moves] = 455.177
[game] = 316017, [moves] = 455.08
[game] = 316210, [moves] = 455.16
[game] = 7.45052 kpps, [moves] = 455.267
[game] = 7.45921 kpps, [moves] = 455.055
[game] = 7.45921 kpps, [moves] = 455.17
[game] = 7.45921 kpps, [moves] = 455.188
[game] = 7.47664 kpps, [moves] = 455.013
[legal] 454/144, [pick] 454/128, [play] 449/430, [score] 173, [win]
0.4592
[legal] 455/144, [pick] 455/128, [play] 449/431, [score] 173, [win]
0.4655
[legal] 454/144, [pick] 454/128, [play] 449/430, [score] 173, [win]
0.4611
[legal] 454/144, [pick] 454/128, [play] 449/431, [score] 173, [win]
0.4674
[legal] 455/144, [pick] 455/128, [play] 450/430, [score] 175, [win]
0.4661
Summary:
========
function 9x9 19x19 ratio
-----------------------------------
game [ cc ] 30216 316265 10.47
game [kpps] 78 7.46 10.45
moves 111 455 4.10
legal [ cc ] 52 144 2.77
pick [ cc ] 74 128 1.73
play [ cc ] 168 430 2.56
score [ cc ] 42 173 4.12
legal+pick+play 294 702 2.39
So, it looks that I was overly pessimistic in terms of performance
drop per move, which is a factor of 2.4x (with little effort to
optimize for 19x19). But, with 4.1x more moves, this still resulted in
a 10x speed penalty. With the provided reference numbers, Libego only
drops 4.5-5.0x, indicating that there is almost no performance loss
per move (1.2x). Is this difference roughly in line with others
expectation?
With that, I hope this at least provides some data for future
discussions, and reference implementation for others that may want to
pick up the ball from here. Someone could add a gtp-interface and make
it available as one of Don's reference implementations for starting
go-programmers. I would still be interested to improve this solution,
but I think I need a fresh set of eyes in order to make any more
progress. I would like to re-iterate that I expect significant gains
on a 64-bit OS as a result on the doubling of the number of available
registers. It would be nice if someone would be able to post some
results. Newer CPU's (will) also have instructions available that can
significantly speed things up (bit-test, compare, popcount, wider
variables, etc).
René van de Veerdonk
On Tue, Aug 25, 2009 at 12:42 PM, Antoine de Maricourt
<antoine.de-marico...@m4x.org <mailto:antoine.de-marico...@m4x.org>>
wrote:
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 <mailto:computer-go@computer-go.org>
http://www.computer-go.org/mailman/listinfo/computer-go/
------------------------------------------------------------------------
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
