Ethereal Pawn-King NN

[xr_a_y]

Does you training pytorch code calling Ethereal search/qsearch/eval from outside or do you managed to use only pre-computed data for learning ?

[AndrewGrant]

Does you training pytorch code calling Ethereal search/qsearch/eval from outside or do you managed to use only pre-computed data for learning ?

Pre computed data only. I store off a half dozen values about the position's evaluation. There is no interfacing between Ethereal and the Pytorch code. Which is why I'm not planning on sharing it at this time.

[pohl4711]

If there is interest, I can generate some compilations for Windows

Yes, please. I would like to do a 7000 games testrun for my ratinglist (need compile for AMD Ryzen). Because, at the moment there is no other test running, I could start the test immediately and the result would go online 2 days later.

https://www.sp-cc.de

[chysiddh14]

lots of user waiting for android verison ,kindly share it also.

[xr_a_y]

Does you training pytorch code calling Ethereal search/qsearch/eval from outside or do you managed to use only pre-computed data for learning ?

Pre computed data only. I store off a half dozen values about the position's evaluation. There is no interfacing between Ethereal and the Pytorch code. Which is why I'm not planning on sharing it at this time.

I kinda think that is the way to go indeed. And this allows using python frameworks and gpu things, ...

[AndrewGrant]

So I kinda went AWOL in this thread, since I discovered a disconnect between training and actual implementation. I thought surely I could fix that quickly and gain a ton more elo. Turns out, not so much, at least for Standard. I did sort out FRC though.

Here are a bunch of binaries:

We have the usual NOPOPCNT, POPCNT, and PEXT, and Android ones. As well as SEE4.1 versions of both POPCNT and PEXT. I'm not explicitly using any fancy vector operations, but the compiler does do them for me in some cases. I suspect the SEE4.1s should be faster for those with modern architectures, but I do not know truthfully. This is something I will have to sort out before Ethereal 12.75, since I want to make sure I provide the (Windows) testers with the best binaries possible.

[pohl4711]

Cool! Thank you very much. Testrun will start tomorrow, when the Igel 2.8.0 testrun is finished.

I tried both popcount compiles on my AMD Ryzen 3900 CPU and can report, that the normal popcount compile is a little bit faster, than the popcount SEE4.1. But the difference is very small (below 2%). I will use the popcount compile for my testrun.

[AndrewGrant]

Cool! Thank you very much. Testrun will start tomorrow, when the Igel 2.8.0 testrun is finished.

I tried both popcount compiles on my AMD Ryzen 3900 CPU and can report, that the normal popcount compile is a little bit faster, than the popcount SEE4.1. But the difference is very small (below 2%). I will use the popcount compile for my testrun.

Thanks for the pending test, and thanks for the speed test.

[MikeB]

If there is interest, I can generate some compilations for Windows / Android users.

Great news, thank!
Windows SSE4.1-popcnt would be great.

Looking forward regards

Awesome , Windows with with above plus -mavx2 please. Thanks!

[AndrewGrant]

Awesome , Windows with with above plus -mavx2 please. Thanks!

I'm assuming you have a Ryzen? If not, it should still work I imagine.

Thats compiled with march=znver1 as the target architecture.