Dev Builds » 20210820-0650

Use this dev build

NCM plays each Stockfish dev build 20,000 times against Stockfish 14. This yields an approximate Elo difference and establishes confidence in the strength of the dev builds.

Summary

Host Duration Avg Base NPS Games WLD Standard Elo Ptnml(0-2) Gamepair Elo
ncm-dbt-01 10:01:58 1281373 3354 967 692 1695 +28.55 +/- 5.97 5 285 828 548 11 +56.22 +/- 11.84
ncm-dbt-02 09:55:32 1309735 3300 948 616 1736 +35.07 +/- 5.71 5 221 868 549 7 +70.44 +/- 11.51
ncm-dbt-03 10:01:27 1313353 3362 971 690 1701 +29.11 +/- 5.92 4 277 849 536 15 +56.29 +/- 11.68
ncm-dbt-04 09:58:25 1309010 3332 961 654 1717 +32.1 +/- 5.79 1 256 854 545 10 +62.82 +/- 11.64
ncm-dbt-05 09:58:24 1313495 3310 973 680 1657 +30.84 +/- 6.04 3 278 813 545 16 +59.35 +/- 11.95
ncm-dbt-06 10:01:37 1306233 3342 1000 706 1636 +30.64 +/- 5.98 3 274 840 534 20 +58.13 +/- 11.75
20000 5820 4038 10142 +31.04 +/- 2.41 21 1591 5052 3257 79 +60.5 +/- 4.79

Test Detail

ID Host Base NPS Games WLD Standard Elo Ptnml(0-2) Gamepair Elo CLI PGN
186146 ncm-dbt-02 1221348 142 42 29 71 +31.89 +/- 29.34 0 12 35 23 1 +59.29 +/- 58.13
186145 ncm-dbt-05 1222461 152 41 29 82 +27.48 +/- 28.62 0 14 37 24 1 +50.64 +/- 56.5
186144 ncm-dbt-04 1233107 162 52 32 78 +43.11 +/- 23.83 0 7 48 25 1 +83.04 +/- 47.88
186143 ncm-dbt-06 1234002 170 54 37 79 +34.86 +/- 24.86 0 11 47 26 1 +66.19 +/- 49.5
186142 ncm-dbt-03 1203088 188 48 36 104 +22.2 +/- 27.39 0 21 42 29 2 +37.1 +/- 52.71
186141 ncm-dbt-01 1198235 186 53 39 94 +26.2 +/- 24.44 0 16 47 30 0 +52.7 +/- 50.0
186140 ncm-dbt-02 1227570 500 152 95 253 +39.78 +/- 15.5 1 38 114 97 0 +82.1 +/- 31.97
186139 ncm-dbt-05 1238498 500 131 91 278 +27.85 +/- 15.12 1 40 128 80 1 +56.07 +/- 30.14
186138 ncm-dbt-04 1232488 500 144 95 261 +34.16 +/- 14.78 0 38 125 87 0 +68.99 +/- 30.52
186137 ncm-dbt-06 1215818 500 144 100 256 +30.65 +/- 16.27 1 44 120 80 5 +56.07 +/- 31.17
186136 ncm-dbt-03 1224502 500 146 98 256 +33.46 +/- 16.09 1 43 116 87 3 +64.66 +/- 31.68
186135 ncm-dbt-01 1176345 500 142 102 256 +27.85 +/- 15.37 1 42 124 82 1 +56.07 +/- 30.66
186134 ncm-dbt-02 1233361 500 146 105 249 +28.55 +/- 14.92 1 38 131 79 1 +57.5 +/- 29.74
186133 ncm-dbt-05 1220979 500 168 103 229 +45.42 +/- 15.07 0 32 124 91 3 +88.0 +/- 30.62
186132 ncm-dbt-04 1253342 500 142 103 255 +27.15 +/- 14.95 0 41 131 76 2 +51.8 +/- 29.75
186131 ncm-dbt-06 1225337 500 151 95 254 +39.08 +/- 15.7 0 38 123 84 5 +71.89 +/- 30.78
186130 ncm-dbt-03 1228641 500 151 110 239 +28.56 +/- 15.3 1 42 122 85 0 +58.93 +/- 30.91
186129 ncm-dbt-01 1197894 500 158 109 233 +34.16 +/- 15.77 1 39 124 82 4 +64.66 +/- 30.65
186128 ncm-dbt-02 1220060 500 133 89 278 +30.65 +/- 14.16 2 29 142 77 0 +64.66 +/- 28.2
186127 ncm-dbt-05 1233954 500 149 94 257 +38.37 +/- 16.02 0 45 106 98 1 +76.25 +/- 32.98
186126 ncm-dbt-04 1209092 500 144 92 264 +36.26 +/- 15.29 0 39 122 87 2 +70.44 +/- 30.91
186125 ncm-dbt-03 1226913 500 150 96 254 +37.67 +/- 15.13 0 38 121 90 1 +74.79 +/- 31.04
186124 ncm-dbt-06 1221521 500 148 116 236 +22.26 +/- 15.72 0 50 120 78 2 +41.89 +/- 31.16
186123 ncm-dbt-01 1205514 500 147 92 261 +38.37 +/- 14.79 0 36 123 91 0 +77.71 +/- 30.77
186122 ncm-dbt-02 1221519 500 139 83 278 +39.08 +/- 15.08 0 34 130 82 4 +73.34 +/- 29.83
186121 ncm-dbt-05 1229569 500 145 116 239 +20.17 +/- 15.45 0 48 128 71 3 +36.26 +/- 30.15
186120 ncm-dbt-04 1237804 500 127 97 276 +20.87 +/- 14.5 0 42 137 70 1 +40.48 +/- 28.98
186119 ncm-dbt-06 1194149 500 159 105 236 +37.67 +/- 14.87 1 31 134 81 3 +73.34 +/- 29.28
186118 ncm-dbt-03 1252546 500 143 106 251 +25.76 +/- 14.6 1 36 140 71 2 +50.38 +/- 28.55
186117 ncm-dbt-01 1172887 500 147 104 249 +29.95 +/- 16.11 1 46 114 87 2 +58.93 +/- 31.93
186116 ncm-dbt-04 1218478 500 150 95 255 +38.37 +/- 15.17 1 35 123 90 1 +77.71 +/- 30.77
186115 ncm-dbt-02 1230481 500 150 90 260 +41.89 +/- 13.8 0 26 139 84 1 +83.57 +/- 28.49
186114 ncm-dbt-05 1225852 500 144 103 253 +28.55 +/- 16.14 1 47 114 86 2 +56.07 +/- 31.92
186113 ncm-dbt-03 1238030 500 140 103 257 +25.76 +/- 14.6 0 40 134 75 1 +50.38 +/- 29.35
186112 ncm-dbt-01 1162635 500 145 108 247 +25.76 +/- 14.47 1 37 136 76 0 +53.22 +/- 29.08
186111 ncm-dbt-06 1213954 500 153 101 246 +36.26 +/- 14.91 0 38 122 90 0 +73.34 +/- 30.91
186110 ncm-dbt-04 1218663 500 148 107 245 +28.55 +/- 15.05 0 41 129 78 2 +54.65 +/- 30.01
186109 ncm-dbt-05 1246470 500 149 109 242 +27.85 +/- 15.25 1 38 135 72 4 +51.8 +/- 29.22
186108 ncm-dbt-02 1222517 500 135 96 269 +27.16 +/- 14.31 1 35 138 76 0 +56.07 +/- 28.8
186107 ncm-dbt-03 1249167 500 150 99 251 +35.56 +/- 15.37 1 34 133 77 5 +66.1 +/- 29.44
186106 ncm-dbt-01 1206991 500 127 101 272 +18.08 +/- 16.13 1 52 120 74 3 +33.46 +/- 31.15
186105 ncm-dbt-06 1249110 500 139 117 244 +15.3 +/- 15.57 1 49 130 67 3 +27.85 +/- 29.9
173506 ncm-dbt-05 1459894 8 2 1 5 +43.64 +/- 153.21 0 1 1 2 0 +88.74 +/- 710.5
173505 ncm-dbt-02 1457374 8 3 2 3 +43.58 +/- 76.58 0 0 3 1 0 +88.62 +/- 172.94
173504 ncm-dbt-01 1441967 18 5 5 8 0.0 +/- 76.96 0 2 5 2 0 -0.0 +/- 162.16
173503 ncm-dbt-04 1447639 20 8 1 11 +126.91 +/- 99.48 0 1 2 6 1 +240.82 +/- 566.7
173502 ncm-dbt-06 1452929 22 7 4 11 +47.65 +/- 46.98 0 0 8 3 0 +97.17 +/- 102.32
173501 ncm-dbt-03 1457820 24 7 6 11 +14.48 +/- 63.77 0 2 7 3 0 +29.02 +/- 132.98
173500 ncm-dbt-05 1457368 50 13 9 28 +27.85 +/- 50.47 0 5 11 9 0 +56.07 +/- 105.56
173499 ncm-dbt-02 1457222 50 19 6 25 +92.45 +/- 42.2 0 1 10 14 0 +200.24 +/- 114.26
173498 ncm-dbt-06 1452486 50 13 7 30 +41.89 +/- 49.33 0 4 11 10 0 +85.04 +/- 106.19
173497 ncm-dbt-04 1446176 50 16 11 23 +34.86 +/- 51.88 0 5 10 10 0 +70.44 +/- 110.15
173496 ncm-dbt-01 1444448 50 16 11 23 +34.86 +/- 47.98 0 4 12 9 0 +70.44 +/- 101.42
173495 ncm-dbt-03 1456478 50 8 11 31 -20.87 +/- 52.71 0 9 10 6 0 -41.89 +/- 109.37
173494 ncm-dbt-05 1456924 50 14 11 25 +20.87 +/- 52.71 0 6 10 9 0 +41.89 +/- 109.37
173493 ncm-dbt-02 1458555 50 16 9 25 +48.96 +/- 46.46 0 3 12 10 0 +99.95 +/- 101.68
173492 ncm-dbt-06 1452635 50 15 12 23 +20.86 +/- 52.71 0 5 13 6 1 +27.85 +/- 96.81
173491 ncm-dbt-01 1442391 50 14 11 25 +20.87 +/- 48.93 0 5 12 8 0 +41.89 +/- 101.14
173490 ncm-dbt-04 1449843 50 14 8 28 +41.89 +/- 53.17 0 5 9 11 0 +85.04 +/- 115.03
173489 ncm-dbt-03 1454704 50 14 12 24 +13.89 +/- 51.12 0 5 14 5 1 +13.9 +/- 92.45
173488 ncm-dbt-02 1457081 50 13 12 25 +6.95 +/- 45.38 0 5 14 6 0 +13.9 +/- 92.45
173487 ncm-dbt-04 1452487 50 16 13 21 +20.87 +/- 35.36 0 2 18 5 0 +41.89 +/- 72.28
173486 ncm-dbt-01 1445806 50 13 10 27 +20.86 +/- 56.25 0 6 11 7 1 +27.85 +/- 105.14
173485 ncm-dbt-03 1455004 50 14 13 23 +6.95 +/- 53.12 0 7 10 8 0 +13.9 +/- 108.98
173484 ncm-dbt-05 1456477 50 17 14 19 +20.86 +/- 40.37 0 2 19 3 1 +27.85 +/- 67.09
173483 ncm-dbt-06 1456627 50 17 12 21 +34.86 +/- 47.98 0 4 12 9 0 +70.44 +/- 101.42

Commit

Commit ID 18dcf1f09754284325157f2d270df10a09297958
Author Tomasz Sobczyk
Date 2021-08-20 06:50:25 UTC
Optimize and tidy up affine transform code. The new network caused some issues initially due to the very narrow neuron set between the first two FC layers. Necessary changes were hacked together to make it work. This patch is a mature approach to make the affine transform code faster, more readable, and easier to maintain should the layer sizes change again. The following changes were made: * ClippedReLU always produces a multiple of 32 outputs. This is about as good of a solution for AffineTransform's SIMD requirements as it can get without a bigger rewrite. * All self-contained simd helpers are moved to a separate file (simd.h). Inline asm is utilized to work around GCC's issues with code generation and register assignment. See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101693, https://godbolt.org/z/da76fY1n7 * AffineTransform has 2 specializations. While it's more lines of code due to the boilerplate, the logic in both is significantly reduced, as these two are impossible to nicely combine into one. 1) The first specialization is for cases when there's >=128 inputs. It uses a different approach to perform the affine transform and can make full use of AVX512 without any edge cases. Furthermore, it has higher theoretical throughput because less loads are needed in the hot path, requiring only a fixed amount of instructions for horizontal additions at the end, which are amortized by the large number of inputs. 2) The second specialization is made to handle smaller layers where performance is still necessary but edge cases need to be handled. AVX512 implementation for this was ommited by mistake, a remnant from the temporary implementation for the new... This could be easily reintroduced if needed. A slightly more detailed description of both implementations is in the code. Overall it should be a minor speedup, as shown on fishtest: passed STC: LLR: 2.96 (-2.94,2.94) <-0.50,2.50> Total: 51520 W: 4074 L: 3888 D: 43558 Ptnml(0-2): 111, 3136, 19097, 3288, 128 and various tests shown in the pull request closes https://github.com/official-stockfish/Stockfish/pull/3663 No functional change
Copyright 2011–2024 Next Chess Move LLC