Stockfish 17 on a snapdragon processor

[bobbobbey]

I learned how to compile this over the weekend. Stockfish instructions are pretty good.
But when you let the program figure out the best version for your computer, it generates an x86 version. So I had to skip that feature.

Seems to be working just fine in chessbase

Thought I'd try sharing it.
https://www.dropbox.com/scl/fi/w7mcrsig ... yi28z&dl=0

[Hai]

Because it's not a feature. It's a BUG.
Congrats and thx.
I wonder how long it will take for the Stockfish developers to speed up Stockfish on Snapdragon CPUs.

[bobbobbey]

I would say it's less a bug and more that I'm learning as I go along.
The compiler that I used is MSYS2. It runs on x86 chips, which means that it's using that PRISM system to talk to my chip. I don't know how to tell it otherwise. Even the ARM version of the app is telling me that my chip is x86.

I'll keep playing with it over the next few weeks as I find the time.

[mpvank]

Thx! Will give it a try.

[bobbobbey]

I'm still running tests, but it seems like the 16.1 neon engine that stockfish compiled is performing better than the 17 ARMv8 that I compiled.

I can't find the older engine on the internet, so here is a link.

https://www.dropbox.com/scl/fi/gplmwwo2 ... yyzjq&dl=0

Please let me know your results.

[Hai]

I'm still running tests, but it seems like the 16.1 neon engine that stockfish compiled is performing better than the 17 ARMv8 that I compiled.

I can't find the older engine on the internet, so here is a link.

https://www.dropbox.com/scl/fi/gplmwwo2 ... yyzjq&dl=0

Please let me know your results.

Do you compile neon only or pop-neon-dotprod?

[Ciekce]

I'm still running tests, but it seems like the 16.1 neon engine that stockfish compiled is performing better than the 17 ARMv8 that I compiled.

if you're directly comparing NPS, note that SF 17 has a significantly larger net than SF 16.1 and will naturally be slower

[bobbobbey]

I didn’t know that.

The 16.1 version was also better with the tactics marathon. Was slower on average, but it answered more of the questions. I had the programs give up after 20 seconds. The default was 10 minutes but I didn’t have patience for that.

The match was even except the 16.1 win in time on the ultra fast games. I ran 30 or so 10 minute games overnight.

[bobbobbey]

I'm still running tests, but it seems like the 16.1 neon engine that stockfish compiled is performing better than the 17 ARMv8 that I compiled.

I can't find the older engine on the internet, so here is a link.

https://www.dropbox.com/scl/fi/gplmwwo2 ... yyzjq&dl=0

Please let me know your results.

Do you compile neon only or pop-neon-dotprod?

The 16.1 version is what I found last year in the stockfish website. This is the name listed. I didn’t compile it. The 17 was compiled as ARMv8. The stockfish manual website has the details. I don’t remember the details.

[Ciekce]

I didn’t know that.

The 16.1 version was also better with the tactics marathon. Was slower on average, but it answered more of the questions. I had the programs give up after 20 seconds. The default was 10 minutes but I didn’t have patience for that.

The match was even except the 16.1 win in time on the ultra fast games. I ran 30 or so 10 minute games overnight.

this sounds like very far from statistically significant