Re: [Raypp-general] Re: C++ performance optimizations

[Andreas K. <and...@we...>]

Hi!

> I agree that the technique is really exciting, but I don't see many places
> where it can be used in Ray++. For example, all vector and matrix
operations
> are already unrolled by hand and inlined where I thought that it is
necessary.
> But if you have a counter-example, where expression templates make a
noticeable
> difference, I would certainly like to see it!

I'm not really an expert on the subject as I just discovered this technique
recently. I thought that you could get some performance optimizations
whenever more vector or matrix operations are combined to a long expression.
But you are probably right, I just skimmed through some of the code
and there's actually nothing like that...


> [...] Ray++ uses virtual functions only where this is absolutely
> necessary. If we tried to remove them, all of the flexibility would be
lost.

That's true. There's probably no better way to do it without sacrificing
the fexibility.


> The most promising area of speedup is IMO the bounding box hierarchy
> generation. The current algorithm is very ad-hoc and could certainly be
improved.

You're talking about POV_HMAKER, right? From that name I was assuming
it is the same technique POVRay uses (which can't be that bad, although I
don't
know it). Could you give a short intro how it works?

Andreas.