Stockfish NNUE Questions

[Dann Corbit]

The AVX build option is new. I need to try it, because I am just running 'Modern' right now.

[Stephen Ham]

Hi Dann,

I was like you in running only Modern. But Corres is correct that the AVX build seems a bit faster.

Regards,
-Steve-

[zullil]

The AVX build option is new. I need to try it, because I am just running 'Modern' right now.

Modern uses popcnt, but not the AVX2 instructions.

ifeq ($(ARCH),$(filter $(ARCH),x86-64-sse41-popcnt x86-64-modern))
arch = x86_64
prefetch = yes
popcnt = yes
sse = yes
ssse3 = yes
sse41 = yes
endif

ifeq ($(ARCH),x86-64-avx2)
arch = x86_64
prefetch = yes
popcnt = yes
sse = yes
ssse3 = yes
sse41 = yes
avx2 = yes
endif

[Stephen Ham]

Hi Louis,

I'm TalkChess' resident computer dummy. Nonetheless in your display, we see: popcnt = yes. Doesn't that mean popcnt is still used by the AVX build?

ifeq ($(ARCH),x86-64-avx2)
arch = x86_64
prefetch = yes
popcnt = yes
sse = yes
ssse3 = yes
sse41 = yes
avx2 = yes
endif

Regards,
-Steve-

[Milos]

I have a Ryzen 9 3950x CPU and I found the most faster SF+NNUE binary (.exe) is the AVX2 optimized one.

Yes, same with my Ryzen 7 laptop,

I suppose this is because the bmi2 binaries include the PEXT instruction, which is very slow on Ryzen?

I never understood all the fuss with that PEXT instruction. I don't even know where it is used in SF source but judging by what it does it seems can be totally avoided by carefully coding the stuff, and I can't believe it would bring more than a mare 1-2% speedup in the very best case.
AVX is on the other hand a proper SIMD thing and brings up a lot.
The AVX compile switch is anyway not for SF but for NNUE part compilation where SIMD insteuctions are handcoded in updating NN score computation.

[zullil]

Hi Louis,

I'm TalkChess' resident computer dummy. Nonetheless in your display, we see: popcnt = yes. Doesn't that mean popcnt is still used by the AVX build?

ifeq ($(ARCH),x86-64-avx2)
arch = x86_64
prefetch = yes
popcnt = yes
sse = yes
ssse3 = yes
sse41 = yes
avx2 = yes
endif

Regards,
-Steve-

Yes. The avx2 build uses popcnt in addition to the avx2 instructions. But unlike the bmi2 build, it doesn't use the pext instruction. On my Intel system, the bmi2 build is fastest; it uses popcnt, avx2 and pext.

[Dann Corbit]

Thanks to all for the AVX2 hints, it was an excellent gain for my threadripper.

If you have a threadripper or other modern AMD, and you like to use my builds, try the stockfish-x64-avx2.exe binary.
I saw a tangible benefit

[h1a8]

Three Questions:

1. In engine parameters, For the "use nnue" box selected is it best to select max threads (8 in my case) or use lower threads (like 6 or 2)?

2. Is there an option to choose hybrid and pure?

3. in android (or ios) for the latest version of stockfish which is stronger hybrid or pure?

[h1a8]

I have an 8 core machine. When I install stockfish nnue I notice that, in engine parameters, the default number of threads is 6. Is this best? Or should I change it to 8 threads?

[bastiball]

I have an 8 core machine. When I install stockfish nnue I notice that, in engine parameters, the default number of threads is 6. Is this best? Or should I change it to 8 threads?

If you want the best performance, you could use 8cores and it is much faster than 6 threads.