[Radiance-general] Re: Physically-based rendering using OpenGL?

[Greg Ward]

The basic problem is that you need to have the entire scene  
description available to every process, and GPU's don't work that  
way.  They are optimized for SIMD (single-instruction, multiple-data)  
processing, meaning that you run the same operation on many operands  
simultaneously.  This is great if you are multiplying huge matrices  
together, but not much use in a ray-tracing context.

Even if the GPU were optimized for this problem, much of what goes on  
in Radiance is ray tree evaluation.  It all sort of happens together  
-- you need to trace rays, figure out what to do at surfaces based on  
material properties, evaluate .cal files, share ambient data (reading  
and writing a shared cache) and all sorts of things that would have to  
be completely rethought for a GPU implementation.

-Greg

> From: Ignacio Munarriz <info at aisarquitectura.com>
> Date: July 12, 2010 11:05:27 AM PDT
>
> I'm not aware of any physical simulation of light software using  
> gpu, but probably it would be interesting to externalize ray- 
> geometry intersections in radiance. I think there is already gpu  
> code(as libraries or open source) that makes that intersections.  
> Radiance instances, meshes, memory management ... would be some  
> issues to solve, maybe waiting a bit the ray-geometry intersections  
> will be normalized in gpus and it will be easier to send rays and  
> geometry from radiance to the cards
>
> El 12/07/2010 19:05, R Fritz escribió:
>> Thank you. That makes it clear that something is possible. But  
>> what? As of version 3.5, Mental Ray (the base of iray) had some  
>> serious limitations in physically based rendering. In Labayrade and  
>> Fontoynont's 2009 CIE 171:2006 testing, it could not, for instance,  
>> simulate a clear sky. Iray is based on Mental Ray 3.8, and it is  
>> possible that some of those problems have been resolved since  
>> version 3.5, but there does not seem to have been any testing  
>> published that says so. Mental Image's iray site doesn't mention  
>> CIE 171. The Luxrender site doesn't mention any testing at all. Are  
>> you aware of any testing that shows accuracy of either of these  
>> tools in the physical simulation of light?
>>
>> Greg, have you looked at this at all?
>>
>> Randolph
>>