[Bf-committers] BIH progress update

[Yves Poissant]

----- Original Message ----- 
From: "Joe Eagar" <joeedh at gmail.com>
To: "Yves Poissant" <ypoissant2 at videotron.ca>; "bf-blender developers" 
<bf-committers at blender.org>
Sent: Thursday, March 20, 2008 10:31 AM
Subject: Re: [Bf-committers] BIH progress update


> Code specific to only 1 architecture isn't allowed.  However, I think if 
> you wrote a generic SIMD macro library and used *it* then it'd be ok.  At 
> least that was what I was planning on doing back when I was considering 
> using SSE for something (using SSE by itself had already been vetoed at 
> that point, but I figured a generic macro library could support Altivec 
> too, so it wouldn't really be architecture-specific anymore).

Thinking about, designing and developing a generic SIMD macro library or set 
of classes would be a sidetrack that is way beyond my commitment. Could that 
be an interesting SoC project? The main hurdle is not so much abstracting 
the operations in an efficient way but the fact that the data layout for 
SIMD operation is also specific plus memory alignment issues. This said, 
abstracting the SIMD operations in an ultra optimal way is also a primary 
requirement. My experience, so far, with SSE is that it is very easy to 
loose any gain with just a little sloppy implementation.

> BTW, do you have a link to that forum post? It sounds cool.

http://ompf.org/forum/viewtopic.php?f=4&t=781

Look at post #4 for the c source file. It is cool indeed. I may try it just 
for my own education to see how well it can really accelerate the ray to 
scene intersection process.

If you are interested in SSE implementations, you should take a look at 
Arauna : http://igad.nhtv.nl/~bikker/ and look at his kdtree.c source code.

Yves