search

xarray-contrib / flox

Fast & furious GroupBy operations for dask.array
https://flox.readthedocs.io
Apache License 2.0
124 stars 18 forks source link

Vectorized grouped (nan)quantile #329

Closed dcherian closed 9 months ago

max-sixty commented 9 months ago

Just saw this in the release notes — very nice!

Always feel free to ping numbagg if you want anything like this in there. (also nice to have it without numba though...)

dcherian commented 9 months ago

It's all vectorized so I'm not sure there's that much benefit. There might be some for the lexsort along an axis though

max-sixty commented 9 months ago

It's all vectorized so I'm not sure there's that much benefit. There might be some for the lexsort along an axis though

Yes agree...

And numba is slower than numpy on sorting!

dcherian commented 9 months ago

And numba is slower than numpy on sorting!

I have learned that intel engineers show up to optimize numpy's sorting with all kinds of SIMD tricks, so that's perhaps not so surprising: e.g. https://github.com/numpy/numpy/pull/25045

max-sixty commented 9 months ago

Nice!!