Finch.jl icon indicating copy to clipboard operation
Finch.jl copied to clipboard

Published 20 hours ago • verified publisher icon willow-ahrens

SDDMM speedup goal

Open willow-ahrens opened this issue 9 months ago • 2 comments

Lets get a real speedup in https://github.com/pydata/sparse/pull/674

willow-ahrens avatar May 09 '24 18:05 willow-ahrens

@willow-ahrens We merged https://github.com/pydata/sparse/pull/674 and with Finch v0.6.26 we have x2.5 speedup (It's still away from x100) for NxN arrays where N=10000 and density is 0.0001 compared to both Numba and SciPy! :tada: Great progress! Thank you!

Finch
Took 8.787564675013224 s.

Numba
Took 22.904020706812542 s.

SciPy
Took 22.59452811876933 s.

https://github.com/pydata/sparse/actions/runs/9078498356/job/24945650514#step:5:299

mtsokol avatar May 14 '24 12:05 mtsokol

That's great! I'll try to run it in debug mode to see if the autoscheduler can be improved

willow-ahrens avatar May 14 '24 14:05 willow-ahrens

Finch Took 0.009948968887329102 s.

Numba Took 0.44066866238911945 s.

SciPy Took 0.2068446477254232 s.

willow-ahrens avatar May 21 '24 21:05 willow-ahrens

Wow! This looks amazing!

Did you change any parameters/code in https://github.com/pydata/sparse/blob/main/examples/sddmm_example.py?

mtsokol avatar May 21 '24 21:05 mtsokol

https://gist.github.com/willow-ahrens/94f35177fa163cc78ed3d0242b6051ef

willow-ahrens avatar May 21 '24 22:05 willow-ahrens

I changed some of the code. The issue right now is that the performance of dense matmul is pretty good, so I changed the density and dimensions

willow-ahrens avatar May 21 '24 22:05 willow-ahrens

this is a 43x improvement, I think this is good enough to close the issue 🎉 . I am curious what kind of matmul is being used in numba.

willow-ahrens avatar May 21 '24 22:05 willow-ahrens

x43 is what we needed for sure! Thank you!!!

mtsokol avatar May 21 '24 22:05 mtsokol