Hi Greg,

about 2 weeks ago, I reported a very confusing color value clipping problem here. (See list archive.) The occurence of clipping is dependent on the use of certain GLSL operations. My problem is still unsolved. Generally, GPUs tend to have a fixed-point architecture, which naturally implies clipping. (Someone correct me if I'm wrong, please.) CPU implementations of non-GPU-supported operations might use floating- point, right? Is it possible that there is some implicit clipping done when the data pipeline leaves the CPU domain, but _not_ between operations in the same CI filter if it is executed on the CPU? On a fixed-point GPU I would expect clipping of intermediate values.
I hope to find this non-transparent behaviour documented some day.
I will be watching your thread on quartz-dev, and I hope both of us will get help by someone who knows. Good luck, and thank you for notifying this group.

Jens Groh


Am 04.10.2008 um 10:24 schrieb Greg Best:

I sent this to the Quartz-dev list, but I'm going to try here too-- apologies if you're getting this a second time. I got a response that the clipping doesn't happen on the GPU, so operations within the kernel are safe-- but no luck finding a way around this limitation. This is starting to drive me crazy and I would love to find a way around it short of going to full blown GLSL.


It looks like Core Image actively clips pixel values to be in the [0,1] range. At what stage does that happen?

The CI API provides a lot of functionality that permits results outside the [0,1] range-- for example in calculating gradients or transcendentals. Is the limited range a function of the GPU, or can these computations be chained within a kernel routine without risk of clipping? Is there any way of coercing the kernel routine to return values outside the [0,1] range, or do all returned values need to be re-encoded into this range?

Thanks--
Greg

_______________________________________________
Do not post admin requests to the list. They will be ignored.
Quartzcomposer-dev mailing list      ([email protected])
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/quartzcomposer-dev/archive%40mail-archive.com

This email sent to [EMAIL PROTECTED]

Reply via email to