Looking for 0.4.0 mojo users .... can you test this ? (0.6.0 results)

[manatlan]

Hi ...

$ wget https://raw.githubusercontent.com/manatlan/sudoku_resolver/master/grids.txt
$ wget https://raw.githubusercontent.com/manatlan/sudoku_resolver/master/sudoku.mojo
$ mojo run sudoku.mojo

and give back : the time (in seconds) which is displayed at the end of the treatments ?

On my computer, the "0.4.0" was 6.65 seconds (median) ....
with the newer "0.5.0"; it's a lot slower, it takes more than 50seconds ( a lot slower than the python3 version ! )

I have opened #1216 ... but it was closed ;-(

[Sharktheone]

I have a mojo 0.4.0 install in my manjaro test VM.

After the upgrade to mojo 0.5.0:

I've made a snapshot so I could test it again.

[jackos]

Thanks for the info @manatlan might be some CPU features got turned off in for your machine in 0.5.0 what is the output if you run this file with mojo: https://github.com/modularml/mojo/blob/main/examples/deviceinfo.mojo

[jackos]

Would you mind running that file as well @Sharktheone with both 0.4.0 and 0.5.0?

[Sharktheone]

System information: 
    OS          :  linux
    CPU         :  znver4
    Arch        :  x86_64-unknown-linux-gnu
    Num Cores   :  16
    CPU Features:  sse4 avx avx2 avx512f avx512_vnni

With mojo 0.4.0 it just fails with error: module 'info' does not contain 'has_vnni'
When i remove these lines it shows the exact same output but without avx512_vnni

Note: This is the output from my host, in my VM the only difference is Num Cores. The VM only has 8.

[jackos]

I'm able to reproduce on remote Linux machine, now just trying to isolate which section of code has regressed

[manatlan]

@jackos
sorry for the delay, I was AFK

Would you mind running that file as well @Sharktheone with both 0.4.0 and 0.5.0?

I haven't the 0.4.0 anymore ;-( ... so I can only run it on 0.5.0 :

System information: 
    OS          :  linux
    CPU         :  alderlake
    Arch        :  x86_64-unknown-linux-gnu
    Num Cores   :  4
    CPU Features:  sse4 avx avx2 avx_vnni

[richardkiss]

I ran it on 0.4.0 on x86_64 ubuntu and got 8.48 s.

I ran it on 0.5.0 on Mac M2 and got 41.79 s.

So this is consistent with your suggestion that 0.5.0 has a regression.

[soraros]

@jackos Should we create a separate issue tracking the slowness on macOS? I know there is no regression, however the performance certainly isn't ideal.

[richardkiss]

I ran it on 0.5.0 on the same x86_64 ubuntu box and got 56.9 s.

So yeah, 0.5.0 has regression, probably an optimizing problem. It's also much larger:

280K    sudoku
168K    sudoku.4

[manatlan]

mojo 0.6.0 is still the slower (compared to rust/java/js/nim/pypy/go) ;-(
https://github.com/manatlan/sudoku_resolver/blob/master/RESULTS.md