On Sat, Jun 12, 2010 at 3:58 PM, Petr Baudis <[email protected]> wrote: > On Sat, Jun 12, 2010 at 08:57:11AM +0800, Fuming Wang wrote: >> My understandings is that light playout should implement Bruegmann's >> original proposal, which is Go rule plus basic patterns that avoid self-eye >> filling. Our FPGA implementation does this while cutting 2 corners, (1)ko >> violations are not checked; (2)suicides are allowed. We have tested in >> software,and found no significant difference in MC's ability to evaluate >> board positions with or without these 2 corners. I wonder what corners did >> libego cut? >> >> Gathering pieces of information in this thread, it seems that the light >> playout implementation on CPU is between 90k-170k/core on the latest >> hardware i7 3.2 GHz. >> >> (1) libego: 40k/GHz -> 40k x 3.2= 130k @ 3.2GHz >> (2) libego: 130k @ 2.5GHz -> 170k @ 3.2GHz >> (3) Mark boon's implementation: 70k @ 2.6 -> 90k @ 3.2GHz > > I just measured: > > $ ./bin/engine -b > Benchmarking, please wait ... > 200000 playouts in 1.74655 seconds > 114.511 kpps > 42.9669 kpps/GHz (clock independent) > 104827/94827 (black wins / white wins) > > on: Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz. > > This is single-threaded (libego can't do multiple threads AFAIK) with version > 0.1.127 - version 0.2.1 is much slower, at 27.5753 kpps. This is because > it does the CrazyStone-like gamma playouts, I assume - then, it probably > has to also handle the actual liberty counts in the library (I didn't > check that though).
That's slower than I expected. Maybe this version already has heavier playouts than the one I tried? If you like I could send you the version I used. I have the source and a working 32-bit binary, unfortunately it doesn't seem to compile on 64-bit linux (don't have time to find out why), but the old binary still works fine. Erik _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
