Ron Parker wrote:
> 
> On Thu, 16 Sep 1999 18:37:29 +0200, Tomas Plachetka wrote:
> >however, i'm more interested in what would happen if the
> >mosaic preview were not used. would the radiosity code
> >still work? would it just run slower? would the quality
> >of the rendering suffer?
> 
> I think it would just run more slowly.  As far as I can tell, the
> main purpose of the mosaic preview is to seed the cache with some
> good candidates for later lookups.

I also have such a feeling. I'll try skipping the mosaic 
preview in the sequential POV-Ray version soon.

> However, I'm betting that removing the mosaic preview step would
> still not solve the discontinuities you get when you render in
> parallel.  I'm not sure I see any way to eliminate those.  You
> could eliminate half of them by making every other band render
> from bottom to top, but that would probably make the other half
> worse.

I agree, this does not sound well at all.

> Using a mosaic preview of the entire scene on one machine to build
> a cache which you would then send to all the rendering processes
> when they start might help, but it still wouldn't be perfect.

This would be just postponing the bad things.

> The problem is that each process builds its own cache over time,
> and they tend to diverge over time if the processes are doing work
> that varies too much.  The discontinuities at band edges are thus
> due to the fact that the process that rendered the first band had a
> much larger cache to work from when it rendered the last row of that
> band.  One possible solution is to use much smaller bands, on the
> order of a single row of pixels.  Better would be to do that, plus
> make all of the running processes share a single cache somehow.
> I suspect you'd still get discontinuities that got worse toward the
> right edge of the image, though, particularly if you used vastly
> different processors or if one or more tasks had much less work to
> do.

Smaller bands increase communication, as well as a shared
cache. (I was/am actually thinking about sharing the cache.) 
The problem of ray tracing parallelisation seems to nested 
when it comes to Monte Carlo. The troubles with a shared 
cache are not unlike the troubles with sharing the workload 
between slaves. Even worse, there is a correlation between
them.

Btw, are there any more intelligent parallel implementations 
of POV-Ray which tackle the "radiosity" (and/or antialiasing, 
etc.) problem?

	y.

"Nothing is perfect," sighed the fox.
  -- A. de Saint-Exupery

Post a reply to this message