Re: [Algorithms] How to get 3dvector largest coordinate index?

[Alen L. <ale...@cr...>]

Jon wrote:
> I've never found this to show up in any profile, and on x86, I believe a
> competent compiler generates cmov.

I got burned with this some time ago, so I'll throw in a warning: A
competent compiler (e.g. icc) does indeed, but a very common one
(msvc) does not (at least not in VC8.0, which was the latest I tested
for that).

The compiler team says that it does in some cases, but I haven't been
able to make it do so. Ever. It certainly doesn't do so for x<y?x:y or
for x<y?0:1. If someone does know how to do that, please let me know.

Btw, I've actually had cases where the lack of cmov did show on the
profile (BIH raycaster), and manually rewritten code using SSE (one
float per register!) was significantly faster, so cache performance
was not an issue there. Interestingly enough, in many cases like this,
when I try to recompile with icc, it turns out its code is very
similar to my hand-optimized one.

And to add one more idea for the OP, if you feel confident in the
normality of the inputs, you can try using float-as-integer comparing
tricks, but the speed is better only if you don't generate the input
using float operations just before doing the compares (i.e. the input
is already in memory).

HTH,
Alen