forked from LeelaChessZero/lc0
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathchangelog.txt
145 lines (112 loc) · 5.31 KB
/
changelog.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
v0.16.0 (2018-07-20)
~~~~~~~
* Fully switched to official releases! No more https://crem.xyz/lc0/
* Fixed a bug when pv display and smart pruning didn't sometimes work properly
after tree reuse.
* Format of protobuf network files was changed.
* Autodiscovery of protobuf based network files works now.
lc0-win-20180715-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Support of new format of network files (needed for lc0 launch on main
training server)
* Fixed hang/poor performance in the beginning of search when there are many
threads. (Happened on linux only though).
* Memory footprint is reduced a bit. (~-60 bytes per node)
lc0-win-20180711-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Edge-node separation introduced a bug that smart pruning didn't work. That's
fixed.
* Changed options parsing so that --backend-opts=cudnn-fp16 is now possible.
* Performance fixes (mostly for slowness introduced by edge-node separation).
lc0-win-20180708-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Mutex contention has been reduced (but locking mutex more rarely).
Helps a lot with many threads running. Especially recommended to check with
multi-GPU configuration.
* Memory usage reduced at least 2x (probably more).
* cudnn backend crashed on large batches (>800) that's fixed.
There is still a limit of batch size 1024 though.
* (not in cudnn build, but for completeness)
Fixed NN computation with BLAS backend, it had up to 5% error before that.
* Default time budgeting params have been changed again! (not by mach this time)
--slowmover=1.95
--time-curve-peak=26.2
--time-curve-left-width=82
--time-curve-right-width=74
lc0-win-20180701-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* fp16-based computation for very modern NVidia GPUs!
May reduce precision a bit, but should be compensated by nps boost.
Enable with --backend=cudnn-fp16 flag
* V is now not stored in nodes (a bit less RAM used while thinking)
* (not in cudnn build, but listing for completeness) blas batching support.
lc0-win-20180629-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Default time budgeting parameters have been changed (again!):
--slowmover=1.93
--time-curve-peak=26
--time-curve-left-width=67
--time-curve-right-width=76
* When generating training games, the engine could confuse client by sending
corrupted output. That's fixed.
lc0-win-20180624-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Default time budgeting parameters have been changed:
--slowmover=2.13 (was 1.8)
--time-curve-peak=22.0 (was 41.0)
--time-curve-left-width=450.0 (was 1000.0)
--time-curve-right-width=30.0 (was 39.5)
* During training game generation, the engine is able to send resign statistics.
lc0-win-20180622-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Time budged allocation has been changed, it allocates more time to early
stages of the game.
Graphs are here: https://github.com/LeelaChessZero/lc0/pull/59
Slowmover value has so be recalibrated, and default value was changed from 2.2 to 1.8.
* Fixed a race condition in cache prefetch code. Realistically it hardly every
occured before though.
lc0-win-20180619-cuda92-cudnn714-00
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* Fix a bug instroduced in version 20180609 which caused the engine to miss checkmates sometimes.
lc0-win-20180614-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* "go searchmoves" uci command is now supported
* It's possible now to disable tree reuse in training games
* Few improvements for random backend
* Lc0 now shows version in uci response
* Analyzer mode has been removed
* extra-virtual-loss has been removed
* Implemented resign (for training games)
lc0-win-20180609-cuda92-cudnn714-01
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* In addition to --backpropagate-gamma, there is also --backpropagate-beta!
Default is 1.0.
lc0-win-20180609-cuda92-cudnn714-00
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Visible changes:
* Experimental changes from 20180604 are now default.
* Memory footprint is reduced by 8 bytes per visible node (+ ~240 bytes in
invisible nodes per visible)
* Introduced --backpropagate-gamma flag.
Default is 1.0. There are rumours that reducing it to 0.75 improves play.
* Extra-virtual-loss parameter has been removed.
* Quotes in backend-opts parameter were not parsed properly, that's fixed.
lc0-win-20180604-cuda92-cudnn714-experimental
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Visible changes:
* Experimental default settings:
cPUCT: 3.4
FPU reduction: 0.9
policy Softmax: 2.2
* Fix memory leak when GUI doesn't ever issue `isready` uci command.
lc0-win-20180602-cuda92-cudnn714-00
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Visible changes:
* cPUCT is now 3.1 by default instead of 1.2 (or what it was before)
* Fixed Batch normalization epsilon in tensorflow backend (but noone uses tensorflow anyway)
* Periodically (every 5 seconds) output "uci info" even if bestmove/depth doesn't change.
* Memory management is redone so that node release happens after "bestmove" and "isready", rather than after "position" uci command.
That garbage collection could take tens of milliseconds and chess GUI already started timer at that point.
Memory management is always fragile, so fresh crashes and memory leaks are possible.
Invisible changes:
* Store castlings again as e1g1 and not e1h1. Fixes a bug that tree was not reused after castling.