[Computer-go] Hardware pattern matching acceleration

Detlef Schmicker ds2 at physik.de
Wed May 29 11:02:03 PDT 2013


Hi,

sounds interesting. I am not to experienced with large patterns at the
moment, but we (oakfoam) are using >1000000 (circular) patterns up to 15
(as the size is counted in several papers), so 30000 might be not enough
for a strong program. Usually we need all pattern matches at the board
at the same time, is your 2ms this time?

I am very interested in GPU acceleration, as this might be very
interesting in mobile devices. They usually have a strong GPU and I was
thinking about accelerating exactly what you are talking about. It is
somewhat difficult to calculate the performance this approach would
offer.

Another interesting acceleration might be real liberties (instead of
pseudo liberties) in the playout moves. This might help to use more
heavy playouts. But I do not have data, if this would help, as we do
only have pseudo liberties.

Detlef


Am Mittwoch, den 29.05.2013, 21:38 +0400 schrieb Рождественский Дмитрий:
> Hi,
> 
> after thinking over your advices in the previous thread and making some investigation I have figured out two options of working on the hardware accelerator. One is to develop a totaly new algorithm that fits for a hardware acceleration better than current ones. Or to find what can be improved in current algorithms moreorless revolutionary, because just acceleration of an algorithm part is not a solution,
> 
> I thought that maybe it will be interesting to improve pattern matching. Current programs can massively match relatively small patterns. Hardware may have the following parametres:
> - pattern size up to 9x9 with wildcards ("does not matter" field state to eleminate influence of insignificant peripheral stones' positions);
> - additional attributes as usual (liberties, ko, distance to an edge)
> - internal position calculator with pattern extractor (just send a cell position and receive its belonging to a pattren back)
> - several patterns' evaluation at a time (should further specify how much)
> - about 30000 7*7 patterns in an $200 device
> - about 2 microseconds time
> 
> Does anyone have an idea will it be valuable?
> 
> Dmitry
> _______________________________________________
> Computer-go mailing list
> Computer-go at dvandva.org
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
> 





More information about the Computer-go mailing list