Hi all,
I would like to let you know that in https://github.com/numpy/numpy/pull/32231,
we are implementing a fused np.minmax implementation that uses the mechanism to
register reduction loops to ufuncs added in
https://github.com/numpy/numpy/pull/31816. The implementation includes
optimizations such as loop unrolling and SIMD (similar to the ones np.min/max
were already using). This obviously does not remove any actual computation as
both min and max need to be computed but it does them in a single pass over the
input array. We saw a 1.25x to a 2x improvement over consecutive min/max calls
depending on the axis being reduced and the dtype (the 2x comes in the cases
where the computation is mostly limited by reading from memory).
This is a long requested feature for NumPy (see
https://github.com/numpy/numpy/issues/9836 and
https://stackoverflow.com/questions/12200580/numpy-function-for-simultaneous-max-and-min
for example)
There are ~15 sites in NumPy and ~70 sites in SciPy where both min and max are
computed consecutively and np.minmax can replace those.
I'm just sending this email out to inform the community and to also welcome
feedback if people think this is a bad idea or have thoughts to share.
Cheers,
Iason.
_______________________________________________
NumPy-Discussion mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3//lists/numpy-discussion.python.org
Member address: [email protected]
