Stockfish NN release (NNUE)

[Joerg Oster]

I need a non-bmi2 compile (but with popc), for AMD Ryzen. BMI2 is terrible slow on AMD

Hi Stefan,
try this one https://www.file-upload.net/download-14 ... 2.exe.html
It's a modern build but still with AVX2 instructions.
No guarantee it works.

[Joerg Oster]

Thanks. In particular, I'm wondering what makes it "efficiently updatable" in a way that, say, LCZero is not.

Efficiency is promoted by a special structure of a neural network.

Almost all the weights of the neural network are located on the input layer, and much less in the intermediate layers (whereas Leela has a lot of calculations in the intermediate layers).

Most of the input features are zero. Accordingly, the weights by which they are multiplied in the input layer do not need to be calculated.

When a move is made, the state of only two input features changes (except when the king moves). Therefore, you only need to subtract the weight of the old position of the piece and add a new one. All intermediate layers will need to be recalculated, but there will be few calculations.

Thanks for the helpful info!

So in a case of a capture, do we have three input changes then,
one for the side of the captured piece, and two for the side to move?

[Rom77]

Thanks for the helpful info!

So in a case of a capture, do we have three input changes then,
one for the side of the captured piece, and two for the side to move?

Sorry, I don't know. Perhaps the answer is in the docs in the links in my posts above.

[pohl4711]

I need a non-bmi2 compile (but with popc), for AMD Ryzen. BMI2 is terrible slow on AMD

Hi Stefan,
try this one https://www.file-upload.net/download-14 ... 2.exe.html
It's a modern build but still with AVX2 instructions.
No guarantee it works.

It works! But it runs only +7.8% faster on my AMD Ryzen 12core, than the BMI2-version. Strange. The "normal" Stockfish are much, much faster with popcount compared to bmi2. But 7.8% are better than nothing, of course.

[Joerg Oster]

I need a non-bmi2 compile (but with popc), for AMD Ryzen. BMI2 is terrible slow on AMD

Hi Stefan,
try this one https://www.file-upload.net/download-14 ... 2.exe.html
It's a modern build but still with AVX2 instructions.
No guarantee it works.

It works! But it runs only +7.8% faster on my AMD Ryzen 12core, than the BMI2-version. Strange. The "normal" Stockfish are much, much faster with popcount compared to bmi2. But 7.8% are better than nothing, of course.

Glad it even works at all!

[jjoshua2]

Can you make a stockfish.nnue.halfkp_384x2-32-32-modern?

[jjoshua2]

It's a modern build but still with AVX2 instructions.

nodchip posted a new release. https://github.com/nodchip/Stockfish/re ... 2020-06-22
Some popcnt/avx2 builds for all these binaries would be useful I'm sure! And nodchip on these ryzen builds says "Could you please send a pull request? I will merge it."

[Damir]

Sorry but I can not support this engine in any way. It is not user friendly. When I try and load it in Infinity GUI it stops responding and crashes after few moves when still in book.. .

[Jesse Gersenson]

Does anyone have a reliable method for building this project on Linux?

[peter]

It's a modern build but still with AVX2 instructions.

nodchip posted a new release. https://github.com/nodchip/Stockfish/re ... 2020-06-22
Some popcnt/avx2 builds for all these binaries would be useful I'm sure! And nodchip on these ryzen builds says "Could you please send a pull request? I will merge it."

These popcnt compiles from Norman work fine for me
https://github.com/FireFather/sf-nnue/releases