Iray needs to be double precision

[SunDevilDave]

I understand that speed improvements are gained through single precision, but it is imperative that ARC at least give us the option of turning on double precision rendering. There have been instances where the tessalation of the mesh is showing because of single precision accuracy. Regular lenses are fine...but I am talking about very long lenses combined and very small details. Also, this would match the quality of traditional Mental Ray rendering.

Tesla cards are built to handle double precision problems...much more so than GeForce cards. Give me double precision, and my community will buy more high-end cards

[dlevesque]

Hi Dave,

While these artifacts are suggestive of precision issue, we'd like to understand exactly what's causing the issue. It doesn't necessarily follow that transforming the entire iray ray-tracing pipeline to double precision would solve the issue; if it did, it might be overkill to have every be in double precision.

I've made attempts to reproduce artifacts similar to the ones you have. I can reproduce obvious precision artifacts by moving the scene very far from the origin, but those look very different from your image. Same with very long lenses. My current theory would be that the geometry pipeline is at cause, perhaps you are using a subdivision or displacement operator which itself introduces the precision issues?

It'd be great if you could post your scene.

cheers,
daniel
iray dev

[dlevesque]

PS: You want might to try your scene with the latest version of iray (2.1), we make some tweaks which might resolve your issue.

daniel
iray dev

[SunDevilDave]

Moving the scene far from orgin would start tearing the geometry I would think. These artifacts are from the rays entering the noise area of a very large number (rays being shot from a camera that is very far away from the origin). This is just my guess. What do you think?

Attached are the maya and mi scene files

Set up the following scene:
1cm cube at world origin
camera location 21.53 miles out
Lens 2000mm
Aperture .013 inches

As a side note, you'll notice the MRSA Standard render seems to block out the FG effect. That artifact disappears as the camera get closer to the origin. This does not happen on the MRFM render...they are supposed to be the same right?

Regardless, you can see there still is a precision issue. Much of what we do started on the foundation that Mental Ray was using double precision calculation for ray tracing. I know some of the pipe isn't double precise (I don't know which parts exactly, so tell me if you know).

I would really like a flag to turn up the precision at render time.

[SunDevilDave]

Is looking at a tiny cube more than twenty miles away extreme? Yes. Yes it is.

[dlevesque]

Hi Dave,

I tested with a >4000mm lens and didn't see any artifacts with iray (though plenty in the viewport). And the fact that you see artifacts with the standalone but not with the built-in version is somewhat revealing.

The artifacts you found are somehow related to shadowing: they have the same color and orientation as the cube's shadow. This suggests they're unrelated to the distance of the camera, and are probably caused by gaps between the triangles. If the triangles are far from the origin, then precision can't be the source of the gaps, and I would look at the origin of the data for the issue.

PS: While mental ray does use double precision in parts of its raytracing pipeline, iray uses geometry which is pre-processed by mental ray and already benefits from this pipeline.

daniel

[SunDevilDave]

hey Daniel,

I wouldn't put it past Autodesk to somehow screw up MRFM.
But just to clarify.

We have two artifacts.

Aritifact number 1
The triangulation of the mesh in MRSA iray mode:
Did you render my included scene? and if so with build 3.10.1.4? Can you ensure that it is being rendered in iray mode and not standard? Earlier this week we were convinced that assemblies were fixed for iray 2.1 when in fact an incorrect CUDA library was packaged with our MRSA install from Autodesk. MR was canceling iray mode and rendering the frame in standard mode instead. So you *might* have been seeing a standard render from MR, not iray.

Artifact number 2
The darkening of the FG after FG pass has finished and rendering has started in MRFM standard mode:
This is not just limited to the shadow but to the whole FG contribution. And fades off after the camera gets closer to the origin.

if you used your own file for the 4000mm test, could you attach please???
I am still trying to figure this out if you say theres hope.

Thanks,
Dave

[steve]

Aritifact number 1
The triangulation of the mesh in MRSA iray mode:

We have looked at it and could reproduce, but only when rendering the "orthogonal" camera of that scene with iray. The perspective camera rendered without trouble also in iray. Anyway, this seems to a bug rather than a general precision issue.

BTW, I exported the .mi file in "binary" mode to preserve full number precision. Your export was using "ascii" instead, which may lose some tiny bits of precision in the conversion number-string-number.

Best, Steve.

[SunDevilDave]

Steve,

By saying "orthogonal" camera, I assume you mean the type of camera I was using...not a true orthogonal camera, but pretty darn close. Can you confirm there is indeed a difference between what you are seeing in iray and what you are seeing from mental ray?

I tried binary format, and it didn't seem to make a difference. I would really like a list of which attributes will utilize doubles versus singles. In the mean time I will keep trying to test and get to the bottom of it. Expect more examples

Thanks,
Dave

[steve]

By saying "orthogonal" camera, I assume you mean the type of camera I was using...not a true orthogonal camera, but pretty darn close. Can you confirm there is indeed a difference between what you are seeing in iray and what you are seeing from mental ray?

Yes, confirmed. We are working on a fix for this specific problem in iray.

Expect more examples

That's great

Best, Steve.