|
下面是GTP客户端解压得的的readme文件,
his is the commandline version of Leela. You likely will want to use
this together with a GTP-capable GUI, or some other tool that speaks
the GTP protocol.
Leela is compatible with GTP version 2, but note that you must add the
"--gtp" option to enable GTP mode.
Allowed options:
-h [ --help ] Show commandline options.
-g [ --gtp ] Enable GTP mode.
-t [ --threads ] arg (=4) Number of threads to use.
-p [ --playouts ] arg Weaken engine by limiting the number of
playouts.
-b [ --lagbuffer ] arg (=100) Safety margin for time usage in centiseconds.
-l [ --logfile ] arg File to log input/output to.
-q [ --quiet ] Disable all diagnostic output.
-k [ --komiadjust ] Adjust komi one point in my disadvantage
(territory scoring).
--noponder Disable thinking on opponent's time.
--nonets Disable use of neural networks.
--nobook Disable use of the fuseki library.
--gpu arg ID of the OpenCL device(s) to use (disables
autodetection).
Note that although the engine can be handicapped by limiting the playout
count, playouts are not standardized so this is not a suitable setting
for comparisons between engines. The OpenCL version is stronger at a
given playout count.
Ruleset
-------
Leela plays according to Chinese rules: area scoring with positional superko.
Compensation for handicap stones is given.
Linux
-----
To get the OpenCL version working, you will have to to install
both the OpenCL libraries for your GPU (e.g. nvidia-libopencl1) as
well as an OpenCL ICD loader (e.g. ocl-icd-libopencl1, nvidia-opencl-icd,
amd-opencl-icd or mesa-opencl-icd - the ICD loader is not card specific and
they are supposed to be interchangeable).
You can see the installed OpenCL drivers in /etc/OpenCL/vendors/
Benchmarking
------------
Leela contains a number of internal benchmarks that can be used
to judge the speed (and resulting strength) of a system:
- CPU, integer operations: type "benchmark".
Reference score, Intel Core i5-6600 on Windows = ~14000 g/s (3500 g/s per thread)
- CPU and GPU, floating point: type "netbench". The OpenCL version
will run this computation on the GPU, but CPU speed and core count still
matters for keeping the GPU fed.
Reference score, Intel Core i5-6600 on Windows = ~100 p/s (first/predictions)
Reference score, Intel Core i5-6600 on Windows = ~265 p/s (second/evaluations)
Reference score, NVIDIA GTX 1070 on Windows = ~475 p/s (first/predictions)
Reference score, NVIDIA GTX 1070 on Windows = ~1850 p/s (second/evaluations) |
|