Problem with Lc0

[Libor Drozdek]

Hi. I recently reinstalled Windows on my computer and dnowloaded Lc0. Everything was fine before but now Leela makes completelly random moves. This is game against Stockfish:

[Event "?"]

[Site "?"]

[Date "2018.12.09"]

[Round "?"]

[White "lc0 11248"]

[Black "stockfish_10_x64"]

[Result "0-1"]

[ECO "A00"]

[Opening "Benko's Opening"]

[PlyCount "14"]

[TimeControl "120+6"]

1. g3 {+0.11/2 8.7s} d5 {+0.12/26 22s} 2. h3 {-0.06/3 8.8s} e5 {+1.14/19 4.2s}

3. a3 {+0.09/3 8.8s} Nf6 {+0.90/22 17s} 4. Ra2 {-0.09/3 7.1s} Be6 {+1.46/19 10s}

5. g4 {-0.06/3 10.0s} h5 {+2.94/20 4.9s} 6. f3 {-0.09/3 7.8s}

Ne4 {+11.52/23 3.9s} 7. d3 {-0.09/4 10s} Qh4# {+M1/127 0.022s, Black mates} 0-1

I can't understand what's happening. It's happening on both versions 19 and 19.2. Everything set to default. Do you have any ideas what may be wrong? I don't understand it because I used Lc0 for some time before an never had any problems. Thank you!

[Edward Panek]

Verify the network file

[Libor Drozdek]

How can I do that? I downloaded the network file from the same site as usual. It worked fine before. The same problem happens for every network file which I try

[Cor Hur]

I'm experiencing a similar problem. Took Lc0 over 100 moves to win a queen and king endgame. Kept shuffling it's queen and didn't even trying to cut off space from the enemy king. I also tried using multiple nets with the same results.

On Sunday, December 9, 2018 at 7:35:12 AM UTC-5, Libor Drozdek wrote:

[Libor Drozdek]

I can't even imagine that LC0 with play like on my computer would be able to win. It's favourite opening when playing against itself is 1. h3 h6 2. a3 a6 (On 10 000 nodes)

[Dietrich Kappe]

Hmmm, looks like maybe the random backend is selected. How are you executing lc0 and with which flags?

[Libor Drozdek]

I am running command .\lc0.exe --weights=weights/11248

No other flags. Everything set to default. It does the same for all networks. I absolutely don't have any idea what to try. How to solve tje problem. I tried to reinstall lc0, different weights, add some flags, etc. It does not help

[Dietrich Kappe]

If you run lc0 from the command line and type “go infinite”, what sort of output do you see?

[Libor Drozdek]

I am at work now so can't try. I will update when I get at home. But I tried command "go" without specifying depth yesterday. Is it the same? It starts evaluating moves as usual but the lines and evals are completely nonsensical.

[Tryfon Gavriel]

Hi all

I hope our Leela will not suffer if these kinds of seemingly odd and random games are sent - is that a risk to be concerned about ?

[Libor Drozdek]

These moves aren't sent. I use Leela only for analysis but now it's useless.

[Rudolf Posch]

Am Sonntag, 9. Dezember 2018 14:43:32 UTC+1 schrieb Cor Hur:

I'm experiencing a similar problem. Took Lc0 over 100 moves to win a queen and king endgame. Kept shuffling it's queen and didn't even trying to cut off space from the enemy king. I also tried using multiple nets with the same results.

This problem is well known, somebody in this forum  explained it with "For Lc0 a mate in 3 is as good as a mate in 10, Lc0 uses no distance to mate evaluation like AB engines". On my PC with Lc0 with 6 men syzygy TB Lc0 e.g. in a KR vs K endgame (and in all other similar positions) Lc0  shuffles the figure resp. its own king fully useless. Arena shows 2 table base hits for such senseless moves. The Lc0 evaluation decreases  slowly (I suppose because of the 50 move count)  an I am sitting  at my PC watching Lc0 with increasing anger and despair. But always suddenly  at some specific move Lc0 starts playing perfectly the endgame. And this at the right time, never losing because  of the 50 move count. And it plays the -for humans difficult- K+N+B vs. K endgame convincingly!  :=)

[Rudolf Posch]

Did you download the proper Lc0 package (blas, opencl, cuda ?). Do you have the newest windows (graphics) driver for the package you downloaded and use? 

Am Sonntag, 9. Dezember 2018 13:35:12 UTC+1 schrieb Libor Drozdek:

Hi. I recently reinstalled Windows on my computer and dowloaded Lc0. Everything was fine before but now Leela makes completelly random moves. This is game against Stockfish:

[Spliff Jiffer]

former v0.19 was able to finish KBNvsK correctly but v19.1rc2 is NOT anymore in my tests and will draw in the end (VSTC)...did u use v19.1rc2 ?

[Spliff Jiffer]

Am Montag, 10. Dezember 2018 17:50:26 UTC+1 schrieb Spliff Jiffer:

former v0.19 was able to finish KBNvsK correctly but v19.1rc2 is NOT anymore in my tests and will draw in the end (VSTC)...did u use v19.1rc2 ?...i had 6men TB installed

[Libor Drozdek]

This is the exact output that I see

PS C:\Users\libor> cd .\chess\engines\lc0-0.19.2\

PS C:\Users\libor\chess\engines\lc0-0.19.2> .\lc0.exe --weights=weights/11248

       _

|   _ | |

|_ |_ |_| v0.19.1-rc2 built Dec  7 2018

go infinite

Loading weights file from: weights/11248

Creating backend [opencl]...

OpenCL, maximum batch size set to 16.

Initializing OpenCL.

Detected 2 OpenCL platforms.

Platform version: OpenCL 2.0 AMD-APP (1800.11)

Platform profile: FULL_PROFILE

Platform name:    AMD Accelerated Parallel Processing

Platform vendor:  Advanced Micro Devices, Inc.

Device ID:      0

Device name:    Intel(R) Core(TM) i7-6500U CPU @ 2.50GHz

Device type:    CPU

Device vendor:  GenuineIntel

Device driver:  1800.11 (sse2,avx)

Device speed:   2592 MHZ

Device cores:   4 CU

Device score:   520

Platform version: OpenCL 2.1

Platform profile: FULL_PROFILE

Platform name:    Intel(R) OpenCL

Platform vendor:  Intel(R) Corporation

Device ID:      1

Device name:    Intel(R) HD Graphics 520

Device type:    GPU

Device vendor:  Intel(R) Corporation

Device driver:  24.20.100.6286

Device speed:   1050 MHZ

Device cores:   24 CU

Device score:   621

Device ID:      2

Device name:    Intel(R) Core(TM) i7-6500U CPU @ 2.50GHz

Device type:    CPU

Device vendor:  Intel(R) Corporation

Device driver:  7.6.0.716

Device speed:   2500 MHZ

Device cores:   4 CU

Device score:   521

Selected platform: Intel(R) OpenCL

Selected device: Intel(R) HD Graphics 520

with OpenCL 2.1 capability.

Loaded existing SGEMM tuning for batch size 16.

Wavefront/Warp size: 8

Max workgroup size: 256

Max workgroup dimensions: 256 256 256

info depth 1 seldepth 2 time 434 nodes 12 score cp -10 hashfull 0 nps 27 tbhits 0 pv g2g3 g7g6

info depth 1 seldepth 2 time 5454 nodes 21 score cp -10 hashfull 0 nps 3 tbhits 0 pv g2g3 g7g6 g3g4

info depth 2 seldepth 3 time 6324 nodes 198 score cp 10 hashfull 0 nps 31 tbhits 0 pv g2g4 f7f5 h2h3

info depth 2 seldepth 3 time 6465 nodes 398 score cp 11 hashfull 1 nps 61 tbhits 0 pv g2g3 c7c6 h2h3

info depth 2 seldepth 3 time 11468 nodes 398 score cp 11 hashfull 1 nps 34 tbhits 0 pv g2g3 c7c6 h2h3 c6c5

info depth 3 seldepth 4 time 13982 nodes 646 score cp -3 hashfull 3 nps 46 tbhits 0 pv h2h4 g7g5 g2g3 a7a6

info depth 3 seldepth 4 time 14233 nodes 908 score cp -4 hashfull 4 nps 63 tbhits 0 pv h2h3 h7h6 g2g4 g7g6 g4g5

info depth 3 seldepth 5 time 17744 nodes 1024 score cp -4 hashfull 4 nps 57 tbhits 0 pv h2h3 h7h6 g2g4 g7g6 a2a3

info depth 3 seldepth 5 time 22130 nodes 1280 score cp 2 hashfull 6 nps 57 tbhits 0 pv a2a3 g7g6 a1a2 a7a5 g2g3

[Libor Drozdek]

I download the same package as before reinstall Windows. Note that I already switched versions once to 19 and no problem there. I just checked my graphics card drivers and they are actual. My version is opencl. 

[Dietrich Kappe]

Even with your low nps, you shouldn’t be seeing these moves. Can you pass it the verbose move stats flag, perform a ‘go movetime 60000’ and share the output? Also, there isn’t a confit file laying about?

[Libor Drozdek]

Max workgroup size: 256

Max workgroup dimensions: 256 256 256

info depth 1 seldepth 2 time 532 nodes 21 score cp -10 hashfull 0 nps 39 tbhits 0 pv g2g3 g7g6

info depth 1 seldepth 2 time 5583 nodes 21 score cp -10 hashfull 0 nps 3 tbhits 0 pv g2g3 g7g6 g3g4

info depth 2 seldepth 3 time 6129 nodes 198 score cp 10 hashfull 0 nps 32 tbhits 0 pv d2d3 b8a6 h2h3

info depth 2 seldepth 3 time 6829 nodes 402 score cp 11 hashfull 2 nps 58 tbhits 0 pv g2g3 c7c6 a2a3

info depth 2 seldepth 3 time 11921 nodes 402 score cp 11 hashfull 2 nps 33 tbhits 0 pv g2g3 c7c6 a2a3 c6c5

info depth 3 seldepth 4 time 14281 nodes 658 score cp 0 hashfull 3 nps 46 tbhits 0 pv h2h3 b8a6 g2g3 a8b8

info depth 3 seldepth 4 time 19361 nodes 919 score cp -4 hashfull 4 nps 47 tbhits 0 pv h2h3 a7a5 a2a3 a8a7 h3h4

info depth 3 seldepth 5 time 22269 nodes 1175 score cp -4 hashfull 5 nps 52 tbhits 0 pv h2h3 h7h5 a2a3 a7a6 a1a2

info depth 3 seldepth 5 time 27326 nodes 1440 score cp -1 hashfull 6 nps 52 tbhits 0 pv h2h3 h7h6 g2g3 b7b6 a2a3

info depth 4 seldepth 6 time 30146 nodes 1696 score cp 4 hashfull 8 nps 56 tbhits 0 pv b1c3 g7g6 a1b1 a7a6 g2g3

info depth 4 seldepth 6 time 35172 nodes 1960 score cp 2 hashfull 9 nps 55 tbhits 0 pv b1c3 g7g6 a1b1 h7h6 b1a1 a7a6

info depth 4 seldepth 7 time 38284 nodes 2216 score cp 2 hashfull 10 nps 57 tbhits 0 pv b1c3 a7a5 a1b1 a8a7 b1a1 h7h6 a1b1

info depth 4 seldepth 7 time 39059 nodes 2482 score cp -1 hashfull 11 nps 63 tbhits 0 pv h2h3 b8c6 g2g4 a8b8 a2a3 b8a8

info depth 4 seldepth 7 time 41151 nodes 2490 score cp -1 hashfull 11 nps 60 tbhits 0 pv h2h3 b8c6 g2g4 a8b8 a2a3 b8a8

info string g1h3  (161 ) N:      76 (+33) (P:  3.49%) (Q:  0.00110) (U: 0.05128) (Q+U:  0.05238) (V: -0.0266)

info string g1f3  (159 ) N:      97 (+23) (P:  3.90%) (Q:  0.00056) (U: 0.05214) (Q+U:  0.05270) (V: -0.0260)

info string d2d3  (288 ) N:     100 (+25) (P:  4.09%) (Q:  0.00021) (U: 0.05240) (Q+U:  0.05261) (V: -0.0261)

info string e2e3  (317 ) N:     103 (+26) (P:  4.18%) (Q:  0.00070) (U: 0.05191) (Q+U:  0.05260) (V: -0.0264)

info string d2d4  (293 ) N:     104 (+22) (P:  4.08%) (Q:  0.00068) (U: 0.05189) (Q+U:  0.05257) (V: -0.0243)

info string e2e4  (322 ) N:     106 (+18) (P:  4.11%) (Q: -0.00057) (U: 0.05313) (Q+U:  0.05256) (V: -0.0256)

info string c2c3  (259 ) N:     109 (+24) (P:  4.31%) (Q:  0.00079) (U: 0.05194) (Q+U:  0.05273) (V: -0.0248)

info string c2c4  (264 ) N:     111 (+24) (P:  4.38%) (Q:  0.00072) (U: 0.05206) (Q+U:  0.05278) (V: -0.0251)

info string b1a3  (34  ) N:     126 (+29) (P:  4.86%) (Q:  0.00243) (U: 0.05032) (Q+U:  0.05275) (V: -0.0262)

info string b2b4  (234 ) N:     129 (+27) (P:  5.27%) (Q: -0.00172) (U: 0.05422) (Q+U:  0.05250) (V: -0.0260)

info string b2b3  (230 ) N:     130 (+29) (P:  5.33%) (Q: -0.00103) (U: 0.05378) (Q+U:  0.05274) (V: -0.0270)

info string f2f4  (351 ) N:     134 (+23) (P:  5.27%) (Q: -0.00122) (U: 0.05391) (Q+U:  0.05269) (V: -0.0257)

info string f2f3  (346 ) N:     134 (+26) (P:  5.41%) (Q: -0.00152) (U: 0.05427) (Q+U:  0.05275) (V: -0.0271)

info string a2a4  (207 ) N:     138 (+30) (P:  5.83%) (Q: -0.00290) (U: 0.05573) (Q+U:  0.05283) (V: -0.0262)

info string h2h4  (403 ) N:     140 (+35) (P:  5.87%) (Q: -0.00117) (U: 0.05386) (Q+U:  0.05269) (V: -0.0264)

info string g2g4  (378 ) N:     141 (+28) (P:  5.61%) (Q: -0.00065) (U: 0.05334) (Q+U:  0.05269) (V: -0.0265)

info string a2a3  (204 ) N:     145 (+29) (P:  6.02%) (Q: -0.00276) (U: 0.05558) (Q+U:  0.05282) (V: -0.0263)

info string b1c3  (36  ) N:     149 (+40) (P:  5.71%) (Q:  0.00423) (U: 0.04858) (Q+U:  0.05281) (V: -0.0256)

info string g2g3  (374 ) N:     152 (+23) (P:  5.93%) (Q: -0.00195) (U: 0.05448) (Q+U:  0.05252) (V: -0.0234)

info string h2h3  (400 ) N:     165 (+19) (P:  6.35%) (Q: -0.00291) (U: 0.05547) (Q+U:  0.05255) (V: -0.0260)

bestmove h2h3 ponder b8c6

[Libor Drozdek]

And no config file around. 

On Monday, December 10, 2018 at 8:08:05 PM UTC+1, Dietrich Kappe wrote:

[Dietrich Kappe]

You wanted —verbose-move-stats

[Dietrich Kappe]

Sorry, you had the verbose move stats on. I don’t know what net you’re running, but it ain’t 11248. Try downloading the latest test30 net and see if it makes a difference.

[Libor Drozdek]

It does the same for every net I download. I always delete all of them to make sure there is not some old network. No difference. Evaluation and the suggested moves are the same every time

[Libor Drozdek]

I did this before many times. That's not that I downloaded Lc0 for the first time and don't know how to use network files

[Libor Drozdek]

This is output for last test 30 network (31951):

info depth 1 seldepth 2 time 426 nodes 13 score cp -4 hashfull 0 nps 30 tbhits 0 pv b1a3 h7h6

info depth 1 seldepth 2 time 494 nodes 21 score cp -3 hashfull 0 nps 42 tbhits 0 pv g2g4 h7h6

info depth 2 seldepth 3 time 5222 nodes 179 score cp 4 hashfull 0 nps 34 tbhits 0 pv g2g3 g7g5 h2h3

info depth 2 seldepth 3 time 7291 nodes 413 score cp 7 hashfull 2 nps 56 tbhits 0 pv b1c3 b8a6 a1b1

info depth 3 seldepth 4 time 9368 nodes 550 score cp 0 hashfull 2 nps 58 tbhits 0 pv g2g3 h7h6 a2a3 a7a6

info depth 3 seldepth 4 time 14412 nodes 556 score cp 0 hashfull 2 nps 38 tbhits 0 pv g2g3 h7h6 a2a3 a7a6

info depth 3 seldepth 4 time 14831 nodes 812 score cp 0 hashfull 3 nps 54 tbhits 0 pv h2h3 b8c6 a2a3 a8b8

info depth 3 seldepth 5 time 16954 nodes 1068 score cp -1 hashfull 5 nps 62 tbhits 0 pv h2h3 b8c6 a2a4 a8b8

info depth 3 seldepth 5 time 22058 nodes 1076 score cp -1 hashfull 5 nps 48 tbhits 0 pv h2h3 g7g5 a2a3 h7h6 a1a2

info depth 4 seldepth 6 time 24594 nodes 1593 score cp 0 hashfull 7 nps 64 tbhits 0 pv h2h3 g7g5 g2g3 h7h6 a2a3

info depth 4 seldepth 6 time 29603 nodes 1603 score cp 0 hashfull 7 nps 54 tbhits 0 pv h2h3 h7h6 g2g3 a7a6 a2a3 a8a7

info depth 4 seldepth 7 time 33219 nodes 2135 score cp 0 hashfull 9 nps 64 tbhits 0 pv h2h3 a7a5 a2a3 h7h6 g2g3 a8a7

info depth 4 seldepth 7 time 38241 nodes 2156 score cp 0 hashfull 9 nps 56 tbhits 0 pv h2h3 h7h6 a2a3 a7a6 a1a2 a8a7 g2g3

info depth 4 seldepth 8 time 42192 nodes 2705 score cp 0 hashfull 11 nps 64 tbhits 0 pv h2h3 h7h6 g2g3 a7a6 a2a3 a8a7 a1a2

info depth 4 seldepth 8 time 47005 nodes 2729 score cp 0 hashfull 11 nps 58 tbhits 0 pv h2h3 g7g6 g2g3 h7h6 a2a3 a7a6 a1a2

info string g1h3  (161 ) N:     101 (+20) (P:  3.73%) (Q: -0.00002) (U: 0.05211) (Q+U:  0.05209) (V: -0.0093)

info string b1a3  (34  ) N:     105 (+27) (P:  3.92%) (Q:  0.00185) (U: 0.05015) (Q+U:  0.05199) (V: -0.0089)

info string d2d3  (288 ) N:     114 (+22) (P:  4.17%) (Q:  0.00031) (U: 0.05183) (Q+U:  0.05214) (V: -0.0112)

info string e2e3  (317 ) N:     116 (+26) (P:  4.22%) (Q:  0.00153) (U: 0.05029) (Q+U:  0.05182) (V: -0.0114)

info string c2c4  (264 ) N:     122 (+23) (P:  4.51%) (Q: -0.00055) (U: 0.05264) (Q+U:  0.05209) (V: -0.0113)

info string b2b4  (234 ) N:     124 (+21) (P:  4.94%) (Q: -0.00564) (U: 0.05762) (Q+U:  0.05198) (V: -0.0104)

info string c2c3  (259 ) N:     125 (+23) (P:  4.54%) (Q: -0.00003) (U: 0.05189) (Q+U:  0.05187) (V: -0.0101)

info string e2e4  (322 ) N:     129 (+29) (P:  4.69%) (Q:  0.00165) (U: 0.05025) (Q+U:  0.05190) (V: -0.0108)

info string b2b3  (230 ) N:     134 (+22) (P:  5.15%) (Q: -0.00391) (U: 0.05590) (Q+U:  0.05199) (V: -0.0117)

info string d2d4  (293 ) N:     134 (+30) (P:  4.97%) (Q:  0.00081) (U: 0.05130) (Q+U:  0.05212) (V: -0.0107)

info string f2f4  (351 ) N:     136 (+23) (P:  4.90%) (Q: -0.00009) (U: 0.05214) (Q+U:  0.05205) (V: -0.0098)

info string f2f3  (346 ) N:     142 (+26) (P:  5.12%) (Q:  0.00028) (U: 0.05158) (Q+U:  0.05186) (V: -0.0091)

info string g2g4  (378 ) N:     142 (+30) (P:  5.22%) (Q:  0.00068) (U: 0.05135) (Q+U:  0.05203) (V: -0.0084)

info string g1f3  (159 ) N:     145 (+31) (P:  5.39%) (Q:  0.00004) (U: 0.05185) (Q+U:  0.05188) (V: -0.0111)

info string b1c3  (36  ) N:     148 (+30) (P:  5.22%) (Q:  0.00233) (U: 0.04964) (Q+U:  0.05197) (V: -0.0108)

info string a2a4  (207 ) N:     149 (+30) (P:  5.65%) (Q: -0.00149) (U: 0.05344) (Q+U:  0.05194) (V: -0.0103)

info string h2h4  (403 ) N:     155 (+28) (P:  5.69%) (Q: -0.00061) (U: 0.05263) (Q+U:  0.05202) (V: -0.0098)

info string g2g3  (374 ) N:     157 (+31) (P:  5.67%) (Q:  0.00107) (U: 0.05107) (Q+U:  0.05214) (V: -0.0098)

info string a2a3  (204 ) N:     160 (+24) (P:  5.87%) (Q: -0.00199) (U: 0.05406) (Q+U:  0.05206) (V: -0.0110)

info string h2h3  (400 ) N:     190 (+17) (P:  6.42%) (Q: -0.00066) (U: 0.05256) (Q+U:  0.05190) (V: -0.0104)

bestmove h2h3 ponder g7g6

[Art Shoe]

The only thing that I can think of is that your Lc0 engine never sees the network files you specify and so always defaults to something else. Make sure that you include the correct network file name in full as a command-line parameter. Double-check the syntax. On my machine it is --weights D:\Chess\Lc0\Networks\weights_11260_4773.pb.gz. Notice the full name with the extension. Sorry if this doesn't help.

[Art Shoe]

Obviously, --weights=D:\Chess\Lc0\Networks\weights_11260_4773.pb.gz with the equal sign.

[gvergh...@gmail.com]

On Sunday, December 9, 2018 at 6:35:12 AM UTC-6, Libor Drozdek wrote:

I can't understand what's happening. It's happening on both versions 19 and 19.2.

Just to make sure it's not your pc at fault, try v18 and post the result...

https://github.com/LeelaChessZero/lc0/releases/tag/v0.18.0

[Spliff Jiffer]

i reported that v19.1rc2 wasnt able to mate in KBNvsK anymore (VSTC)...this issue is resolved again with v19.1.1 according my quick tests, thank you !

still 1 annoing issue left though:

in arena-gui the engine is "marked" as ardy "loaded" when the initialisation-process hasnt even started and therefore the engine looses some seconds at the start of every new game (if Leela has to make first move at least) whithin a tournament !...this was introduced AFTER v0.19...thx for consideriation in advance.

[Libor Drozdek]

While I was testing another version I relized than opencl tuning does not complete. This is output:

Will try 578 valid configurations.

(1/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=16 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 8354.0 us (64.3 GFLOPS)

(2/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 7413.7 us (72.4 GFLOPS)

(4/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=16 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 6383.5 us (84.1 GFLOPS)

(5/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 4898.6 us (109.6 GFLOPS)

(69/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 4616.7 us (116.3 GFLOPS)

(77/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 4560.5 us (117.7 GFLOPS)

(97/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=16 NDIMC=16 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 4487.6 us (119.6 GFLOPS)

(115/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 4038.4 us (132.9 GFLOPS)

(120/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 3738.8 us (143.6 GFLOPS)

(130/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=16 NDIMC=16 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 3719.5 us (144.3 GFLOPS)

(192/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=2 3481.0 us (154.2 GFLOPS)

(199/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=2 3379.0 us (158.9 GFLOPS)

(225/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=2 3071.5 us (174.8 GFLOPS)

(229/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=2 3048.5 us (176.1 GFLOPS)

Wavefront/Warp size: 8

After 229 it just stops. Could it be the problem? Anyway, the problem is for all opencl versions which I tried. 

[gvergh...@gmail.com]

To me, it's obviously the drivers on your pc. After reinstalling Windows, your pc would have installed updates and somehow in the process, corruption has occured.

I'd advise a run of Display Device Uninstaller and an install of the latest gpu drivers from the company, and not Microsoft.

Watch a youtube video on DDU first, to familiarize yourself.

[gvergh...@gmail.com]

It's Display Driver Uninstaller

[Rudolf Posch]

I have set up a random white KNB vs. black king position in Arena and played from this position  a test mating match between Lc0 V0.19.1 net Id 31956 with 6 men syzygy table bases  against  Stockfish-x64 V9 with also 6 men syzygy TB.

SF9 wins with KNB  straight with less moves against the lone king Lc0 defender.

Lc0 as white shuffles -as seen in other endgames- aimless for some time around. But -fortunately for the peace in my mind- it gets enlightened

 and mates convincingly the perfect with the KNB.k syzygy table bases defending SF9,  and this shortly before reaching the 50 move count which would mean a draw :=)

Look into the attached tournament .pgn file, Leela mated SF9 in the 5 games as white in the 97th,  99th!, 97th, 95th and 99th ply!!! 

Because of the 50 move rule a black to mate sequence must be ended with the 100th ply. The shortest distance to mate is 33 plies  for positions with the 4 pieces K,N, B und  k  

http://www.maths.usyd.edu.au/u/thiel/misc/endgames/  . Captures in KNBk endgames lead to a draw, so in the worst case Leela as attacker may shuffle until the 100 -33 = 67 ply of the game (34th move of white). But after that Leela must play perfect against a perfect defending Stockfish!  :=)

[evalon32]

On Wednesday, December 12, 2018 at 4:02:15 PM UTC-5, Rudolf Posch wrote:

I have set up a random white KNB vs. black king position in Arena and played from this position  a test mating match between Lc0 V0.19.1 net Id 31956 with 6 men syzygy table bases  against  Stockfish-x64 V9 with also 6 men syzygy TB.

SF9 wins with KNB  straight with less moves against the lone king Lc0 defender.

Lc0 as white shuffles -as seen in other endgames- aimless for some time around. But -fortunately for the peace in my mind- it gets enlightened

 and mates convincingly the perfect with the KNB.k syzygy table bases defending SF9,  and this shortly before reaching the 50 move count which would mean a draw :=)

Look into the attached tournament .pgn file, Leela mated SF9 in the 5 games as white in the 97th,  99th!, 97th, 95th and 99th ply!!! 

Because of the 50 move rule a black to mate sequence must be ended with the 100th ply. The shortest distance to mate is 33 plies  for positions with the 4 pieces K,N, B und  k  

http://www.maths.usyd.edu.au/u/thiel/misc/endgames/  .

That table shows moves, not plies. The longest DTM for KBNk is 33 moves (65 plies). (The shortest is 1 ply, of course.)

[Tony Mars Rover]

Are you running the regular version of Lc0 or the Cuda version?

[Tony Mars Rover]

[Rudolf Posch]

@evalon32 Thank you very much for your correction, I apologize for sloppily  confusing the 33 move value for the longest DTM in KBNk with half moves.

In my random test position (see below and https://syzygy-tables.info/?fen=8/2N5/2B5/2K1k3/8/8/8/8_w_-_-_0_1)  the DTM with Bd5 is 52 plies.

So  Lc0 and SF9 should win this position as white with best play at the 26th move (26. xx # mate).

 -   SF9 mates as white in its 5 test games (see attach) with 53, 53, 57, 57 and 57 plies (53 looks perfect, 57 is odd?).   

 -   But Lc0 plays at the beginning aimless moves, which could lead inadvertently in the worst case to a position with DTM 33 moves = 66 plies. 
So  Leela as attacker may shuffle at most until the 100 - 66 = 34th ply of the game. With the 17th move of white  (17. xxx) Leela must play perfectly in the worst case.

Am I right? Maybe there is a +/- 1 error in my little calculations :=(