Calvin <nomail@nomail> wrote:
> Thorsten Froehlich <tho### [at] trfde> wrote:
> > Calvin wrote:
> > > I'm very interested in porting the parallelizing code to Cell, and possibly also
> > > to CUDA.

> Are you saying it's unfeasible because of the lack of double floating point
> support, or is there some architectural reason? [Cell/SPU has doubles and the
> soon-to-be-released G92/9800 have doubles]

  Don't know about Cell, but in CUDA it's probably just impossible. For one,
CUDA doesn't support recursive function calls, which is more or less a must
in raytracing. (POV-Ray uses recursion not only for tracing rays but other
things as well.) Also single-precision floating point (with division and
square root of even lower quality compared to single-precision math done
with the FPU) is also a rather big wall. Not to talk about the CPU/GPU
bottleneck. As for multithreading, threads must run in groups of at least
32 threads that execute identical instructions simultaneously. Not very
likely in raytracing (again, because of all those pesky random recursions...)

-- 
                                                          - Warp

Post a reply to this message