Reproducibility vs artifact-less images

[Reverend]

Shader precision is about producing artifact-less images, not about getting reproducable results up to the lsb.

Almost 4 years ago, someone I respect posted that here, as a result of some comments I made about precision and the importance of reproducibility.

4 years later, what are your thoughts on that comment?

 

[Tim Murray]

Real-time rendering is by its very definition a collection of tricks and corners that can be cut in order to produce reasonable images fast enough to appear to be in motion. The original statement is still true. If you get the same result 99.9% of the time when you use FP32 or FP64, and that .1% isn't noticeable to the eye, then it doesn't matter that you're not using FP64.

 

[Reverend]

If you get the same result 99.9% of the time when you use FP32 or FP64, and that .1% isn't noticeable to the eye, then it doesn't matter that you're not using FP64.

Yes, that's true and a bit of a "Duh". The problem is the first word above.

PS. I never mentioned any numbers, you did, which has very little to do with my topic.

 

[Tim Murray]

Yes, that's true and a bit of a "Duh". The problem is the first word above.

PS. I never mentioned any numbers, you did, which has very little to do with my topic.

Then your question makes no sense. It's not like using lower precision won't be deterministic and therefore produce the same output every time.

 

[KimB]

As far as I know, reproducibility is only a concern with regard to stochastic sampling patterns, and those don't even look better until you get to high numbers of samples (16-32, minimum).

 

[Jawed]

Then your question makes no sense. It's not like using lower precision won't be deterministic and therefore produce the same output every time.

I can only guess that the emphasis is placed upon "shader" meaning "algorithm", rather than ALU precision.

So some algorithms might have rather large "fuzzy regions" in their solution sets producing visible artefacts in corner cases as compared with other algorithms for notionally the same concept that are more robust.



Jawed

 

[Reverend]

Then your question makes no sense. It's not like using lower precision won't be deterministic and therefore produce the same output every time.

What I mean is that there isn't a need to specifically compare FP32 and FP64 (like you used as an example). That's why I said the problem stems from the word you used (i.e. "If you get the same result 99.9% of the time...")

I repeat the statement :

Shader precision is about producing artifact-less images, not about getting reproducable results up to the lsb.

If I said : Without reproducibility, there's no guarantee that an artifact-free image (as a result of using very high shader precision) on one machine is an artifact-free image on another machine, would I be correct?

Put it another way -- If I said : With reproducibility, we can guarantee that an artifact-free image (as a result of using very high shader precision) on one machine is an artifact-free image on another machine, would I be correct?

 

[KimB]

Ahhh, reproducibility across different machines. That makes a bit more sense now. I just don't think it's a realistic goal for many processes, unfortunately, as there are certain operations that are just ill-defined, such as anisotropic filtering, MIP mapping, and multisampling.

 

[Reverend]

Ahhh, reproducibility across different machines.

Was I being that cryptic in my original post? Is there really such a thing as real problems with reproducibility using various shader precisions on a single machine?

Just to remind you : I'm talking about shader precision.

Someone also mentioned something about "cutting corners"....

 

[KimB]

I was thinking about reproducibility in terms of re-rendering the same frame on the same hardware, which doesn't happen with stochastic sampling.

Anyway, I don't think reproducibility is typically going to be possible across different hardware, because texture usage is so common. I think there's also a bit of leeway in the IEEE spec such that we can't expect to see FP math results to be the same across different hardware, even if they largely comply to the IEEE spec.