see gh-8910 for a masked array change that makes use of these loops.

How do timings compare for cases that do and don't take advantage of this?

it should easily be a factor 8-16, a SSE unit can perform 16 boolean ops at once.

I'm more concerned about the impact of the branching for small loop sizes when we aren't in the special case this is optimizing for

What determines the number of elements in the inner loop? Does the ufunc always pick a contiguous dimension, or the longest one, or what?

The ufunc machinery overhead is an order of magnitude higher than the cost of the inner loops. The code inside them only starts mattering when the arrays become larger than a few thousand elements.

The inner loop size is the full array if it can be expressed via the strides without buffering.
With buffering (casting, broadcasting, reductions) the inner loop size is 8192 contiguous elements (np.setbufsize)

This helps, but it's still slower than working with the full arrays (on my MSVC build, anyway).

Setup:

In [1]: b = np.ma.make_mask_none((100, 100), writeable=False)
In [2]: f = np.ma.make_mask_none((100, 100), writeable=True)

Before:

In [3]: %timeit b | b
100000 loops, best of 3: 12.7 µs per loop

In [4]: %timeit b | f
100000 loops, best of 3: 12.5 µs per loop

In [5]: %timeit f | f
1000000 loops, best of 3: 1.86 µs per loop

After

In [3]: %timeit b | b
100000 loops, best of 3: 8.83 µs per loop

In [4]: %timeit b | f
100000 loops, best of 3: 2.59 µs per loop

In [5]: %timeit f | f
1000000 loops, best of 3: 1.82 µs per loop

I didn't add code for scalar | scalar as that probably doesn't happen in practice, the masked code usually checks for double nomask before doing a masked operation.
But I added it now and cleaned up the code a bit.

That it is still slower than full array is actually due to the iterator. In the full array case it uses a fastpath to skip the nditer setup. Apparently this does not trigger when a zero-d array is involved, so it sets up the expensive iterator.
It should still be good enough, the mask operation cost is only a fraction of the full masked array operation.

Does the SIMD work supercede this? @Qiyu8 @seiko2plus Thoughts?

IMO, A memcpy&memset method is provided to replace SIMD/uloop method for special cases such as readonly masked arrays, but I don't see enough evidence to proves the benefits here. if It's shows better performance than SIMD/uloop, then we should accept it. OTOH, this PR reminds me that #16960 is extending current sse2 functions to universal intrinsics, we can have a bench test after #16960 is merged.

@juliantaylor @eric-wieser I'm inclined to close this, thoughts?

I'm going to close this. Thanks Julian.