Performance bencmarks comparison with scipy.sparse

[rth]

I see there is an asv setup in the repo but I couldn't find any results.

How does pydata/sparse currently compare to scipy.sparse (e.g. for dot product)? Or is most of linalg operations expected to use scipy.sparse.linlag in the long run and so the performance would be equivalent?

[hameerabbasi]

Tl;dr: It's bad. Sometimes 10 times worse, even for element-wise operations.

We're looking at approaches that essentially guarantee optimal performance and GPU support in all cases (#326). However, we don't have a ETA for this.

[hameerabbasi]

In [3]: x = sparse.random((5000, 5000), density=0.001)

In [4]: y = sparse.random((5000, 5000), density=0.001)

In [5]: %timeit x + y  # Numba compilation
5.86 ms ± 190 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [6]: %timeit x + y
5.97 ms ± 70.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [7]: xs = x.tocsr()

In [8]: ys = y.tocsr()

In [9]: %timeit xs + ys
408 µs ± 8.44 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

[hameerabbasi]

For __matmul__:

In [11]: %timeit x @ y
1.05 s ± 18.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [12]: %timeit xs @ ys
1.09 ms ± 17.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

[rth]

Thanks for the explanations @hameerabbasi !

[rgommers]

It seems like matmul could easily be improved though right? I just had a peek at the SciPy implementation (csr_matmul in sparsetools/csr.h) and it's a relatively straightforward triple for-loop in C++. So I guess no one has put much effort into a similar Numba implementation here, otherwise it wouldn't be 1000x slower.

[rgommers]

Not that I'm suggesting to optimize that right now, I just want to say there's no fundamental reason why it's slow, just no one got around to it since this package is still new.

[hameerabbasi]

dot is implemented in Numba here:

sparse/sparse/_coo/common.py

Line 1150 in 98a39cd

def _dot_coo_coo(coords1, data1, coords2, data2): # pragma: no cover

Perhaps the memoization isn't working.

[hameerabbasi]

#350 makes this a bit better.