Stockfish Testing Queue

Pending - 0 tests 0.0 hrs

None

Active - 0 tests

Finished - 14 tests

19-02-15 Cof noskip2 diff
LLR: 2.95 (-2.94,2.94) [0.00,3.50]
Total: 30243 W: 4052 L: 3815 D: 22376
sprt @ 120+1.2 th 4 Show Elo gain of noskip (derived from latest master) on VLTC 4 threads
19-01-29 Cof novoting diff
LLR: 0.33 (-2.94,2.94) [-3.00,1.00]
Total: 157875 W: 20295 L: 20483 D: 117097
sprt @ 60+0.6 th 8 quick scaling check for no thread voting on 60+0.6 8 threads. Should fail quickly. No functional change on 1 thread.
19-01-25 Cof noskip diff
ELO: -1.56 +-3.2 (95%) LOS: 17.1%
Total: 10000 W: 1097 L: 1142 D: 7761
10000 @ 60+0.6 th 30 Try to measure performance of noskip on 30 threads. I hope 2GB hash is OK for this test. We can stop this test, when it takes too much time.
19-01-25 Cof noskip diff
Pending...
sprt @ 60+0.6 th 8 2nd run of noskip at 8 threads and 60+0.06, just to avoid that first run was a fluke.
19-01-20 Cof noskip diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 53093 W: 8198 L: 8431 D: 36464
sprt @ 20+0.2 th 8 LTC 20+0.2 8 thread for removing thread skipping scheme. No functional change for 1 thread. Well, 5+0.05 test failed quickly, skipping scheme seems to be ~4 Elo at 5+0.05 8 threads. Reason for this test is, that the skipping scheme does NOT scale well with time, so I expect between 0 and -2 Elo. In my local tests (Amd FX CPU) it was 0 Elo, I would like to see it run on fishtest noob hardware. I will limit max games to 20000, to limit the resource usage. If I am wrong that scaling of the skipping scheme is bad, this test will fail quickly (like 5+0.05 run before).
19-01-23 Cof noskip diff
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 15138 W: 2005 L: 1877 D: 11256
sprt @ 60+0.6 th 8 Remove thread skipping scheme: For 8 threads we have now 2 data points on fishtest for "noskip" 20000 games 5+0.05: -4.93 +-3.0 21500 games 20+0.2: -0.05 [-2.66,2.56] So what happens on longer time control??? let's try 60+0.6. It will take a lot of resources, so I will set a limit of 20000 games.
19-01-19 Cof noskip diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 5042 W: 880 L: 1047 D: 3115
sprt @ 5+0.05 th 8 Remove skipping scheme, no functional change on 1 thread
18-12-04 Cof master diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 18796 W: 3002 L: 3096 D: 12698
sprt @ 60+0.6 th 1 Do the same test for LTC, too.I think 64MB is sufficient for 60+0.6, but let's see.
18-12-04 Cof master diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 45589 W: 9878 L: 9864 D: 25847
sprt @ 10+0.1 th 1 The last test clearly showed that 8MB is better for STC. Let's double the hash size one more time and check, if there is still an Elo gain [0,4].
18-12-04 Cof master diff
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 18116 W: 4089 L: 3853 D: 10174
sprt @ 10+0.1 th 1 According to my local tests 8MB TT is stronger then 4MB on 10+0.1. Test on fishtest with [0,4], if it passes, we should change the default TT size for STC.
18-09-06 Cof pullreq1663 diff
LLR: -0.16 (-2.94,2.94) [0.00,4.00]
Total: 10085 W: 1618 L: 1595 D: 6872
sprt @ 5+0.05 th 63 Test for pullreq1663 for speed gain. 2nd try, I wanted to test with [0,4] only. target machine: AMD 64 core computer.
18-02-07 Cof O2 diff
LLR: -3.88 (-2.94,2.94) [0.00,4.00]
Total: 22746 W: 4918 L: 5044 D: 12784
sprt @ 10+0.1 th 1 Switch to O2 in general, enable lto for Windows-mingw. Huge gain on my Windows PC with gcc 7.3. I am not sure, if it will work on every Windows machine and hopefully no regression under Linux. Test for Elo gain.
17-12-17 Cof NumaBug diff
ELO: 9.24 +-4.6 (95%) LOS: 100.0%
Total: 5000 W: 634 L: 501 D: 3865
5000 @ 10+0.1 th 39 No functional change for Linux! Only relevant for Windows AND multi-Numa-node machines. I will put a 40 Core machine to fishtest, when test is approved. please note the Priority +1
17-06-17 Cof rev_pref_earlier diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 71365 W: 13006 L: 12919 D: 45440
sprt @ 10+0.1 th 1 sorry, it's my first test, wrote wrong numbers for signature. Non functional change revert prefetch earlier as parameter patch [0,4]