[ptx] x64 benchmarks

[JD Smith]

On Tue, 2006-01-17 at 09:27 +0100, Pablo d'Angelo wrote:
> Marek Januszewski schrieb:
> > Hello,
> > 
> >  So how about we set up some big benchmarking project and we all post
> > results here?
> 
> I always had the feeling that the AMD64 bit version run a bit faster.
> 
> While I haven't set up a complete benchmark project, but I have remapped a
> single image from a project (with lots of black, empty space). No crop
> (nona) or acceleration (PTmender) was used.
> 
> This are the user times:
>              x86 32 bit (s)    AMD64 (s)      speedup
> PTStitcher	35.495
> PTmender	35.675		28.995		23%
> nona            27.588		22.179		24%
> 
> So at least for panoramic stitching, actually using the 64 bit instructions
> on that the AMD64 provides seems to improve remapping speed by ~ 20%
> 
> Note: if fast transform or cropped tiff output would have been used, the
> times would differ a lot. For smaller panos, the fast transform code used by
> PTStitcher and PTmender would blow nona away, while for larger panos, nona
> would be faster

Is that because of the "calculating the black region" issue with
PTStitcher vs. nona?

The fast transform stuff is actually fairly simple: just evaluate the
true transform coarsely, interpolate for points in between, and do error
checks to see whether to add more real mappings in each region.
PTStitcher does this a row at a time, i.e. in 1D.  I think this is worth
putting in nona, possibly as a full 2D sparse sampling.  That, said the
slowest parts of panorama creation are now, for me, SIFT point
detection, and enblend, so it may be worth putting emphasis on those
components instead.

JD