CrazyAra, ClassicAra, MultiAra 0.9.5 release

[carldaman]

My folder lacks the .trt files. Anyway, I didn't want to use the gpu version, but the cpu MKL package which worked for prior versions.

Hello carldaman,
yes, it looks like the MXNet library provided in the Windows version is unable to load the new ClassicAra-sl-model-wdlp-rise3.3-input3.0 model.
I am sorry for the inconvenience. All other models (MultiAra models, and CrazyAra model) should load fine in the Windows CPU version.
However, the Mac and Linux CPU version should be able to load the new ClassicAra model.
I tried to build a newer MXNet backend library on Windows but it failed due to several out of memory compiler errors for me, despite having 32 GiB RAM).

As an alternative, I started to add a OpenVino backend.
The OpenVino API can be used with CPUs (AMD & Intel), the Intel Neural Compute Stick and Intel GPUs.

• https://github.com/QueensGambit/CrazyAra/pull/153

So there will probably be a release 0.9.5.post0 soon to provide a working ClassicAra CPU Windows version.

Hi Johannes,
OK, thanks for looking into it. A working cpu version would be awesome.

[IQ_QI]

My folder lacks the .trt files. Anyway, I didn't want to use the gpu version, but the cpu MKL package which worked for prior versions.

Hello carldaman,
yes, it looks like the MXNet library provided in the Windows version is unable to load the new ClassicAra-sl-model-wdlp-rise3.3-input3.0 model.
I am sorry for the inconvenience. All other models (MultiAra models, and CrazyAra model) should load fine in the Windows CPU version.
However, the Mac and Linux CPU version should be able to load the new ClassicAra model.
I tried to build a newer MXNet backend library on Windows but it failed due to several out of memory compiler errors for me, despite having 32 GiB RAM).

As an alternative, I started to add a OpenVino backend.
The OpenVino API can be used with CPUs (AMD & Intel), the Intel Neural Compute Stick and Intel GPUs.

• https://github.com/QueensGambit/CrazyAra/pull/153

So there will probably be a release 0.9.5.post0 soon to provide a working ClassicAra CPU Windows version.

Hi Johannes,
OK, thanks for looking into it. A working cpu version would be awesome.

The OpenVino Backend has been integrated and a new release is available:

• https://github.com/QueensGambit/CrazyAr ... .9.5.post0

It features a new UCI-option Threads_NN_Inference which defines how many threads to use for neural network inference. This no longer requires setting up an environment variable called OMP_NUM_THREADS.

The latest ClassicAra model is included within each release package.

Current limitation for OpenVino backend:

• No Int8 inference support enabled yet.

• Only tested on regular CPU and not with Intel GPUs or the Intel Neural Compute Stick.

Regression test

The new OpenVino backend is about 100 - 150 nps faster on CPU and much easier to install than the MXNetMKL backend.

Code: Select all

[TimeControl "7+0.1"]
Score of ClassicAra - 0.9.5.post0 OpenVino vs ClassicAra 0.9.5 - MXNetMKL: 82 - 17 - 55 [0.711]
Elo difference: 156.4 +/- 45.9, LOS: 100.0 %, DrawRatio: 35.7 %
154 games finished.

[AdminX]

My folder lacks the .trt files. Anyway, I didn't want to use the gpu version, but the cpu MKL package which worked for prior versions.

Hello carldaman,
yes, it looks like the MXNet library provided in the Windows version is unable to load the new ClassicAra-sl-model-wdlp-rise3.3-input3.0 model.
I am sorry for the inconvenience. All other models (MultiAra models, and CrazyAra model) should load fine in the Windows CPU version.
However, the Mac and Linux CPU version should be able to load the new ClassicAra model.
I tried to build a newer MXNet backend library on Windows but it failed due to several out of memory compiler errors for me, despite having 32 GiB RAM).

As an alternative, I started to add a OpenVino backend.
The OpenVino API can be used with CPUs (AMD & Intel), the Intel Neural Compute Stick and Intel GPUs.

• https://github.com/QueensGambit/CrazyAra/pull/153

So there will probably be a release 0.9.5.post0 soon to provide a working ClassicAra CPU Windows version.

Hi Johannes,
OK, thanks for looking into it. A working cpu version would be awesome.

The OpenVino Backend has been integrated and a new release is available:

• https://github.com/QueensGambit/CrazyAr ... .9.5.post0

It features a new UCI-option Threads_NN_Inference which defines how many threads to use for neural network inference. This no longer requires setting up an environment variable called OMP_NUM_THREADS.

The latest ClassicAra model is included within each release package.

Current limitation for OpenVino backend:

• No Int8 inference support enabled yet.

• Only tested on regular CPU and not with Intel GPUs or the Intel Neural Compute Stick.

Regression test

The new OpenVino backend is about 100 - 150 nps faster on CPU and much easier to install than the MXNetMKL backend.

Code: Select all

[TimeControl "7+0.1"]
Score of ClassicAra - 0.9.5.post0 OpenVino vs ClassicAra 0.9.5 - MXNetMKL: 82 - 17 - 55 [0.711]
Elo difference: 156.4 +/- 45.9, LOS: 100.0 %, DrawRatio: 35.7 %
154 games finished.

So to be clear, the OpenVino version should be used on CPU's and Intel GPU's. So I should stick with TensorRT for Nvidia GPU's? Is that correct?

Thanks!

[IQ_QI]

Yes, the OpenVino back-end

supports heterogeneous execution across an Intel® CPU, Intel® Integrated Graphics, Intel® Neural Compute Stick 2 and Intel® Vision Accelerator Design with Intel® Movidius™ VPUs

https://docs.openvinotoolkit.org/latest/index.html

If you have an Nvidia GPU, I still recommend to use the TensorRT back-end instead.

For now the device "CPU" is hard-coded in the OpenVino back-end, but could be made available via the "Context" UCI-Option:
https://github.com/QueensGambit/CrazyAr ... pi.cpp#L91

[Joerg Oster]

Yes, the OpenVino back-end

supports heterogeneous execution across an Intel® CPU, Intel® Integrated Graphics, Intel® Neural Compute Stick 2 and Intel® Vision Accelerator Design with Intel® Movidius™ VPUs

https://docs.openvinotoolkit.org/latest/index.html

If you have an Nvidia GPU, I still recommend to use the TensorRT back-end instead.

For now the device "CPU" is hard-coded in the OpenVino back-end, but could be made available via the "Context" UCI-Option:
https://github.com/QueensGambit/CrazyAr ... pi.cpp#L91

Does this only run on Intel CPUs?
When running ./ClassicAra on Linux I get the following error:

Code: Select all

error while loading shared libraries: libinference_engine_transformations.so: cannot open shared object file: No such file or directory

[IQ_QI]

Yes, the OpenVino back-end

supports heterogeneous execution across an Intel® CPU, Intel® Integrated Graphics, Intel® Neural Compute Stick 2 and Intel® Vision Accelerator Design with Intel® Movidius™ VPUs

https://docs.openvinotoolkit.org/latest/index.html

If you have an Nvidia GPU, I still recommend to use the TensorRT back-end instead.

For now the device "CPU" is hard-coded in the OpenVino back-end, but could be made available via the "Context" UCI-Option:
https://github.com/QueensGambit/CrazyAr ... pi.cpp#L91

Does this only run on Intel CPUs?
When running ./ClassicAra on Linux I get the following error:

Code: Select all

error while loading shared libraries: libinference_engine_transformations.so: cannot open shared object file: No such file or directory

No, it also runs on AMD CPUs, but is not officially supported by OpenVino.
The file

Code: Select all

libinference_engine_transformations.so

should be included in the Linux package.
To make sure it finds the shared library you can add the install directory to your LD_LIBRARY_PATH.

Code: Select all

export LD_LIBRARY_PATH=<install-dir>/CrazyAra_ClassicAra_MultiAra_0.9.5.post0_Linux_OpenVino

You can add this line to e.g. your ~/.bashrc file and activate it using

Code: Select all

source ~/.bashrc

Alternatively, you can modify the rpath for the binary, to include the current working directory.
I did add a line to CMakeList.txt to change the rpath, but apparently that had no effect here.

Code: Select all

set_target_properties(${PROJECT_NAME} PROPERTIES LINK_FLAGS "-Wl,-rpath,./")

https://github.com/QueensGambit/CrazyAr ... s.txt#L462

[Joerg Oster]

Yes, the OpenVino back-end

supports heterogeneous execution across an Intel® CPU, Intel® Integrated Graphics, Intel® Neural Compute Stick 2 and Intel® Vision Accelerator Design with Intel® Movidius™ VPUs

https://docs.openvinotoolkit.org/latest/index.html

If you have an Nvidia GPU, I still recommend to use the TensorRT back-end instead.

For now the device "CPU" is hard-coded in the OpenVino back-end, but could be made available via the "Context" UCI-Option:
https://github.com/QueensGambit/CrazyAr ... pi.cpp#L91

Does this only run on Intel CPUs?
When running ./ClassicAra on Linux I get the following error:

Code: Select all

error while loading shared libraries: libinference_engine_transformations.so: cannot open shared object file: No such file or directory

No, it also runs on AMD CPUs, but is not officially supported by OpenVino.
The file

Code: Select all

libinference_engine_transformations.so

should be included in the Linux package.
To make sure it finds the shared library you can add the install directory to your LD_LIBRARY_PATH.

Code: Select all

export LD_LIBRARY_PATH=<install-dir>/CrazyAra_ClassicAra_MultiAra_0.9.5.post0_Linux_OpenVino

You can add this line to e.g. your ~/.bashrc file and activate it using

Code: Select all

source ~/.bashrc

Alternatively, you can modify the rpath for the binary, to include the current working directory.
I did add a line to CMakeList.txt to change the rpath, but apparently that had no effect here.

Code: Select all

set_target_properties(${PROJECT_NAME} PROPERTIES LINK_FLAGS "-Wl,-rpath,./")

https://github.com/QueensGambit/CrazyAr ... s.txt#L462

Working now!
Thanks for your help, really appreciated.

[AdminX]

Finally after after allot of GUI issues with Chessbase Cloud (Cloud Server would stop responding during LTC game), and Install issues with Banksia (BSG not detecting CrazyAra or ClassicAra, only MultiAra was detected). I was able to complete a game at LTC, using ClassicAra vs Slowchess in the latest Banksia GUI (0.50) over my home network.

ClassicAra:
Processor Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz 3.60 GHz
Nvidia RTX 2070 Super
Installed RAM 32.0 GB
Windows 10 64-bit operating system, x64-based processor
2 Threads

Slowchess:
Processor Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz 1.99 GHz
Intel UHD 620
Nvidia GeForce MX150
Installed RAM 32.0 GB (31.8 GB usable)
Windows 10 64-bit operating system, x64-based processor
8 Threads



I found it interesting looking at the above graph and seeing when ClassicAra saw that it was falling behind, how it would search at greater depths to get out of it's hole.

[pgn]
[Event "banksia game"]
[Date "2021.09.05"]
[White "rm::LTSlowChess Blitz 2.7 avx"]
[Black "ClassicAra 0.9.5 (Aug 20 2021) "]
[Result "1/2-1/2"]
[TimeControl "40/7200+0"]
[Time "02:01:13"]
[Termination "repetition"]
[ECO "D04"]
[Opening "Queen's pawn game"]

1. Nf3 d5 {+0.27/36 187171 2776880} 2. d4 Nf6 {+0.28/35 183002 3428602}
3. e3 {D04: Queen's pawn game} c5 {-0.02/37 62399 973707} 4. c4 cxd4 {-0.08/37 191176 2847860}
5. exd4 {+0.00/28 191904 131159286} g6 {-0.07/35 188202 4707000} 6. Qb3 Bg7 {-0.20/44 187462 2874833}
7. cxd5 O-O {-0.21/54 194278 3673725} 8. Nc3 {-0.01/33 202964 135726049} Nbd7 {-0.22/64 128795 5213446}
9. a4 {-0.01/34 202962 110566503} a5 {-0.20/51 189551 2731231} 10. Be2 {-0.04/35 202959 122707270} Nb6 {-0.13/19 188141 2631807}
11. O-O {-0.02/32 202957 53698967} Nfxd5 {-0.11/46 188006 5255467} 12. Nxd5 {-0.08/34 202956 126341831} Nxd5 {-0.10/56 186115 7301981}
13. Bd2 {+0.00/33 202954 113769838} Be6 {-0.10/49 185714 9260393} 14. Bc4 {+0.08/31 202954 75328348} Nc7 {-0.07/44 187590 6061434}
15. Rfc1 {+0.08/34 202951 91169139} Bxc4 {-0.06/23 186362 8438340} 16. Qxc4 {+0.04/32 81278 58626648} Nd5 {-0.04/21 187084 10589436}
17. Qb5 {+0.08/35 207691 124222174} b6 {-0.05/34 188326 12889661} 18. Rc4 {+0.02/35 278719 188278495} e6 {-0.18/40 187838 2510032}
19. Rac1 {+0.02/34 204645 75085499} Qf6 {-0.15/45 186738 5141260} 20. Rc6 {+0.01/32 204627 90044932} Qf5 {-0.28/21 186488 2670150}
21. Qe2 {+0.02/33 204606 104050457} Rad8 {-0.23/38 184502 2443995} 22. g3 {+0.01/34 204593 127372042} h6 {-0.20/16 183865 2723792}
23. h4 {+0.01/34 204567 120679942} Qg4 {-0.23/41 184630 2647689} 24. Kh2 {+0.02/34 204549 138277070} Rd7 {-0.43/28 100205 1473764}
25. R6c4 {+0.01/33 204520 135887600} Rfd8 {-0.28/32 191161 2661911} 26. Kg2 {-0.01/30 230802 156883328} Rd6 {-0.17/46 188379 3337031}
27. Qe1 {+0.16/31 202779 124705024} Qf5 {-0.06/47 187931 2778050} 28. Ne5 {+0.32/28 195652 132620063} Re8 {-0.01/16 188567 3374961}
29. Ra1 {+0.32/31 203266 136435792} Qh5 {+0.19/36 189549 2390344} 30. Ra3 {+0.36/31 203202 137250817} Rf8 {+0.35/32 188096 2608075}
31. Rf3 {+0.40/31 416907 281072725} Ne7 {+0.10/23 59443 935903} 32. Qc1 {+0.76/31 182768 125088543} Rdd8 {+0.21/46 199184 2677528}
33. b4 {+0.77/33 182740 117494844} Nf5 {+0.69/41 202004 2554890} 34. g4 {+0.84/32 216692 163823654} Nxh4+ {+0.64/52 199742 4965939}
35. Kg3 {+0.99/35 303296 236624128} Bxe5+ {+0.62/48 198002 6861936} 36. Kh3 {+0.52/37 431169 339141796} Nxf3+ {+0.62/46 416 6776430}
37. gxh5 {+0.36/36 69571 54394141} Nxd2 {+0.60/52 70556 6477048} 38. Qxd2 {+0.40/36 47742 36975100} g5 {+0.60/50 318 6370210}
39. bxa5 {+0.32/37 137717 108083603} bxa5 {+0.57/67 459766 11461747} 40. Qxa5 {+0.32/38 137604 111022988} Bxd4 {+0.53/50 410650 11259846}
41. Kg2 {+0.32/38 144979 132163005} Bf6 {+0.54/48 788 10743024} 42. Qc5 {+0.60/36 179286 111895320} Rd5 {+0.42/43 189734 3256433}
43. Qa7 {+0.59/40 179275 130193767} Rfd8 {+0.45/41 482 1850639} 44. Qb7 {+0.20/35 243343 209688354} Rf5 {+0.26/26 59540 906020}
45. Rc8 {+0.20/38 177530 121354343} Kg7 {+0.13/18 62834 1173257} 46. Rxd8 {+0.20/37 93757 94261531} Bxd8 {+0.13/16 202628 3872118}
47. Qb4 {+0.20/41 179862 128800335} Ba5 {+0.08/26 201557 2971203} 48. Qa3 {+0.20/39 179860 177107286} g4 {+0.04/20 63213 932260}
49. Qb2+ {+0.09/37 179854 170844953} Kg8 {+0.05/18 53107 1546736} 50. Qb8+ {+0.08/41 179857 149329647} Kh7 {+0.05/19 45776 1710972}
51. Kf1 {+0.24/42 179845 168071928} Kg7 {+0.04/18 217670 4308225} 52. Qg3 {+0.24/44 179838 139398613} Rg5 {+0.05/28 174902 5383726}
53. Ke2 {+0.24/41 111696 108713534} Kg8 {+0.05/10 49159 1678487} 54. Qb8+ {+0.24/42 182215 116382164} Kg7 {+0.02/16 224095 3698139}
55. Qf4 {+0.24/44 182185 150999386} Kg8 {+0.03/17 224583 4439434} 56. Qd4 {+0.24/42 182155 153318015} g3 {-0.30/29 69594 1130526}
57. fxg3 {+0.24/43 182121 117771096} Rxg3 {-0.29/20 23397 1341729} 58. Qh4 {+0.24/45 182085 141202789} Rg5 {-0.24/39 236989 3208343}
59. Kd3 {+0.24/46 182044 101977668} Rc5 {-0.24/24 59814 2940980} 60. Qh2 {+0.24/45 182003 182345663} Bc7 {-0.25/23 247180 3660841}
61. Qg2+ {+0.24/44 181958 183090407} Rg5 {-0.23/27 246750 5127945} 62. Qa8+ {+0.24/46 181954 122957722} Kg7 {-0.23/29 246907 3861607}
63. Qb7 {+0.24/48 181951 180828336} Ba5 {-0.22/28 165568 6233911} 64. Qh1 {+0.24/46 174413 187887224} Rc5 {-0.28/16 170145 2451355}
65. Qg1+ {+0.24/46 182345 110186206} Rg5 {-0.23/18 252208 3869998} 66. Qh2 {+0.24/47 174089 191723555} Rc5 {-0.24/20 113654 1706154}
67. Qb2+ {+0.24/44 182778 183973840} Kh7 {-0.25/24 264749 4180804} 68. Qe2 {+0.24/45 182763 143273071} Kg8 {-0.27/23 102549 1637495}
69. Qh2 {+0.18/40 182746 180103293} Bc7 {-0.27/21 118483 2508183} 70. Qg1+ {+0.24/44 182725 104186242} Rg5 {-0.24/22 292397 4389856}
71. Qa7 {+0.24/47 182698 202954208} Ba5 {-0.22/27 241574 7457715} 72. Qa8+ {+0.24/46 116683 129993448} Kg7 {-0.25/21 295721 3916692}
73. Qf3 {+0.24/44 161560 189284530} Rc5 {-0.25/20 51109 4625275} 74. Qg3+ {+0.24/45 192826 193887165} Kh7 {-0.27/21 156044 2187442}
75. Qg4 {+0.24/46 114288 130122410} Rg5 {-0.27/28 140851 2243785} 76. Qe4+ {+0.24/44 166124 192361557} Kg7 {-0.27/35 394011 5504638}
77. Qe2 {+0.24/44 211640 183441140} Rc5 {-0.29/16 271845 3687303} 78. Qh2 {+0.24/44 211559 231076872} Bc7 {-0.28/33 436272 6140548}
79. Qh4 {+0.24/43 211432 240374197} Kg8 {-0.28/27 278775 4317062} 80. Qe7 {+0.24/38 174267 200059044} Rd5+ {-0.22/20 325396 5968446}
81. Kc4 {+0.24/44 174922 210069235} Ba5 {-0.23/18 6919 2945716} 82. Qh4 {+0.24/41 182176 140975097} Rf5 {-0.24/21 114997 3624345}
83. Qh2 {+0.08/38 182154 179834605} Kh7 {-0.23/23 197371 3200838} 84. Kb3 {+0.08/40 182132 113166198} Kg8 {-0.23/14 136971 2116127}
85. Kc4 {+0.00/39 386096 476587855} Rd5 {-0.21/23 112291 1793173} 86. Qe2 {+0.00/43 330964 426200618} Rg5 {-0.23/18 199127 3013422}
87. Qf3 {+0.00/44 172145 181662382} Rf5 {-0.22/25 148041 2530813} 88. Qa8+ {+0.00/42 172118 107866860} Kg7 {-0.18/24 202682 3232462}
89. Qg2+ {+0.00/43 172093 95135506} Rg5 {-0.17/34 200185 5789607} 90. Qf3 {+0.00/44 172064 45241786} Kf8 {-0.18/15 156646 2833886}
91. Kb3 {+0.00/40 172034 148000374} Rf5 {-0.22/22 138806 2154160} 92. Qe2 {+0.00/42 172004 197804993} Kg8 {-0.22/21 140934 2232958}
93. Qh2 {+0.00/40 129371 147341448} Rg5 {-0.22/21 162897 2227848} 94. Ka2 {+0.00/41 173442 136708635} Rd5 {-0.20/21 110220 1721768}
95. Qh1 {+0.00/40 173422 81152185} Kf8 {-0.21/25 211186 3144728} 96. Kb3 {+0.00/41 173400 136638321} Kg8 {-0.19/18 140563 2266598}
97. Ka2 {+0.00/42 281733 260686209} Rf5 {-0.19/16 161681 2666674} 98. Kb2 {+0.00/41 168942 73855486} Rg5 {-0.22/17 161787 2193792}
99. Qa8+ {+0.00/41 168942 122107910} Kg7 {-0.16/17 220795 3277912} 100. Qf3 {+0.00/42 168936 186987203} Rd5 {-0.18/15 221382 6166111}
101. Kb3 {+0.00/39 168935 146822732} Kg8 {-0.16/13 177685 2475609} 102. Qg2+ {+0.00/39 168928 151808356} Kf8 {-0.09/11 223536 3007164}
103. Qf2 {+0.00/39 167840 187897948} Rxh5 {-0.27/30 219844 2949549} 104. Qf6 {+0.00/46 168977 182117400} Kg8 {-0.27/31 219884 5453252}
105. Kb2 {+0.00/37 114981 128518490} Bc7 {-0.28/27 97214 1485115} 106. Ka3 {+0.00/42 172247 172751160} Bb6 {-0.28/23 127743 1718057}
107. Ka2 {+0.00/45 172196 169192031} Kf8 {-0.30/29 119860 1882284} 108. Qh8+ {+0.00/41 172140 173097966} Ke7 {-0.29/22 1662 113574}
109. Kb3 {+0.00/41 140496 147853243} Rh3+ {-0.20/30 123346 1899861} 110. Kb4 {+0.00/45 174562 183490375} Rh5 {-0.09/32 274458 3858370}
111. Kb3 {+0.00/45 174524 165367321} Ba5 {-0.08/30 4949 3630393} 112. Qd4 {+0.00/41 174480 59676091} Be1 {-0.07/14 157544 3970914}
113. Qf4 {+0.00/38 174425 176151678} Ke8 {-0.04/12 181595 2988797} 114. Qb8+ {+0.00/41 174356 144032634} Ke7 {-0.03/10 339906 2874662}
115. Qf4 {+0.00/41 283187 166934583} Ke8 {-0.03/4 342981 3445523} 116. Qb8+ {+0.00/41 217842 221549382} Ke7 {-0.03/2 337351 755285}
117. Qc7+ {+0.00/43 146365 131303532} Ke8 {-0.05/2 335241 1294251} 118. Qc8+ {+0.00/39 146226 50177476} Ke7 {-0.02/4 3730 104084}
119. Qc7+ {+0.00/40 132106 127164534} Ke8 {-0.04/2 497443 966324} 120. Qc8+ {+0.00/38 160664 30089414} Ke7 {+0.00/2 3159 50422}
121. Qc7+ {+0.00/43 178727 126824680} 1/2-1/2
[/pgn]

[Gabor Szots]

Do I understand well that to run ClassicAra you need at least 2 cores, one for the search, one for the network?

[IQ_QI]

I found it interesting looking at the above graph and seeing when ClassicAra saw that it was falling behind, how it would search at greater depths to get out of it's hole.

ClassicAra uses a simple heuristic for time management to search longer if the current evaluation is worse than the previous one.
https://github.com/QueensGambit/CrazyAr ... r.cpp#L157
This could be the reason why it searched for higher depth in these positions.