Subject: Re: [boost] non-branching min()
From: DE (satan66613_at_[hidden])
Date: 2009-09-10 14:00:11
on 10.09.2009 at 21:41
Simonson, Lucanus J wrote :
> Hmm, I brought this exact idea to boost about three years ago. I had min, abs and median of three.
> [code here]
> You can improve the efficiency for larger types by using arrays of
> pointer to argument instead of arrays of argument value.
> In my own testing the speedup for prescott was 25% for min and 45%
> for median of three. I would expect less speedup than that in core2
> processors because the pipeline is shorter.
pipeline is shorter for a reason
it sould not degrade overall performance (on the contrary it should
> Also there is a new
> instruction in core2 that allows the min to compile to a branchless
> instruction stream. You need to use the proper flags with the
> compiler to enable the use of new instructions. If you are
> compiling in max compatibility mode you are dropping performance on
> the floor. In our testing branch free min (as you propose) was no
> faster than a < b ? a:b; when compiling with the new insturctions
> enabled for a core2 machine. Median of three was still 25% faster though.
> The response from boost three years ago was that such low level
> architecture specific optimizations have no place in portable libraries.
i wasn't here 3 years ago
however like stepanov said once "great minds think alike"
anyway thanks, i'll think of it and sertainly try it out
-- Pavel ps good luck with your polygon lib
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk