Are Infs still faster than NaNs?

Years ago, I was advised that the processing of Inf elements was much faster than the processing of NaN elements, and was advised to use Infs instead of NaNs unless there was a compelling reason. When I replaced NaNs with Infs, I did indeed find that Infs crunched much faster. In recent days, though (R2013), I'm finding that the playing field is pretty even among Infs, NaNs, and finite array elements, e.g.,

>> A=nan(3e3); tic;fft2(A);toc
Elapsed time is 0.192480 seconds.

>> A=inf(3e3); tic;fft2(A);toc
Elapsed time is 0.200329 seconds.

>> A=rand(3e3); tic;fft2(A);toc
Elapsed time is 0.211940 seconds.

Was I hallucinating back then, or has there been an improvement in the hardware/software that handles NaNs and Infs?

[dpb]

Don't recall ever hearing it; could be.

Earliest version installed here is R12; essentially same result as yours above with it (altho precedes time at which nan and inf had arguments). So, if were so, would seem to have been even earlier.

[Matt J]

Here's the link to the original Newsgroup discussion.

http://www.mathworks.com/matlabcentral/newsreader/view_thread/256238#665904

I knew I didn't imagine it. And I'm definitely not the only one in that thread that witnessed slow NaNs.

Answer 1

[Matt J]

Hmmm, possibly a false alarm. In the tests below, minimization operations involving NaNs are still noticeably slower,

    n=1e4;k=1e4;
    A=rand(n,k);b=randi(k,n,1);
    idx=bsxfun(@lt,b,(1:k));

    A(idx)=nan;

    tic
    c=min(A,[],2);
    toc
    %Elapsed time is 0.348245 seconds.

    A(idx)=inf;

    tic
    c=min(A,[],2);
    toc
    %Elapsed time is 0.126220 seconds.

Still, I seem to remember that a lot of other operations were slower, too.

Answer 2

[Jan]

It depends on the processor. AMD CPUs and SSE engines of Intel cpus process NaNs efficiently.

[Matt J]

I've always used Intel CPUs. Can't imagine that accounts for what I saw, unless the efficiencies came only recently.