A way to short-circuit "any"?

Hello,

My code is bottlenecked by many calls of the following line:

if any(vectorA & vectorB)

where vectorA and vectorB are large, sparse, and logical.

Specifically, these vectors are 1.4M elements long and average 0.6% non-zeros. However, due to the previous calculations, they share a common non-zero quite often (say roughly half the time). As written, the & operation is performed before the any.

Since this is the natural bottleneck, I've been looking for ways to improve the computation, and short circuiting out comes to mind, but I haven't found an elegant way to do this in Matlab.

Does anyone know a slick optimization?

Thanks!

-Kyle

Answer 1

[Jan]

Is the multiplication faster?

if any(vectorA .* vectorB)

But I do not assume that this helps, because you still process more elements than needed. You need a Mex function for short-cutting. For full arrays you find in the FEX: anyExceed and anyEq. But substantial modifications are required for accessing sparse arrays.

[Kyle]

Thanks for the reply!

I should have mentioned this up front- I tried .* but didn't find it to be substantially faster.

I'll look into the file exchange functions- thanks for the tip. Trying to use full vector routines is disastrously slow, but I'll see if I can make modifications.

Answer 2

[Walter Roberson]

In previous discussions, it has been a bit of a mystery about whether any() short-circuits or not. If I recall correctly, timing tests (Jan's, probably) indicated that some short-circuiting was going on but not in strictly linear ways. If I recall, there is a FEX contribution that short-circuits logic operations.

However, the above was with respect to full arrays, not sparse.

[Jan]

any([1, zeros(1, 1000)]) is faster than any([zeros(1, 1000), 1]). But Kyle wants to shortcut the & actually, if any value triggers the any already.

[Walter Roberson]

I wonder... wasn't there something discussed a couple of months ago, about how "if" (and perhaps "while") now short-circuit & and | in some cases, but that other contexts use the full evaluation? Just maybe

if ~(A & B)

would short-circuit even for sparse.