Cycles-X

I have an idea regarding accelerating indoor or many bounces scenes.
We usually have bigger surfaces that contribute to bounces, now, couldnt we tell cycles this:

  1. Analyze the scene for the biggest normal surfaces that would bounce light heavily around the scene at ONLY 2 bounces and a quarter of resolution. Basically similarly to an irradiance cache that gets blurred and extremely smoothed out but not only for the camera, but for the entire scene.
  2. Store that info as a rough GI emissive cache.
  3. When a ray hits now, calculate the coefficient with that stored value in mind, while at the same time, that cache now makes the whole scene emissive to some degree, emitting more rays from primary surfaces!
  4. Now all rays should be nearly noise free even after many bounces in the most remote corners.

The point is that if that cache makes primary hit surfaces emissive, the scene would converge insanely fast. Of course that emission would have to store the normals of the emission rays itself too. A mirror will have paralell rays, while the white surface of a wall a random one. That info has to flow into the coefficient value. This is different from other irradiance cache and radiosity stuff, because it would take all materials into account in the prepass, regarding their normals and roughnesses.

1 Like

Looks like after the July14th build, CyclesX no longer recognizes my GTX980 and reports no available GPUs. It reverts to CPU-only rendering.

There’s no bug report active yet, but still you can report in the cycles x feedback thread on BDT

1 Like

Note that Brecht clarified that the only patch they looked at was the animation denoising patch. They only discussed the idea about Apple Metal support and Path Guiding. So it’s not that there are some actual developments, it is just that they’re discussing things.
Looking forward to path guiding. My dream for 3.0 is Path Guiding and Spectral Cycles both implemented, so we can have both caustics and dispersion (the initial implementation will probably not include dispersion though), it would be fantastic.

2 Likes

You might be hitting the minimum OPTIX drivers version. Your card needs to support the latest version of the OPTIX drivers.

Yep, I think you may be right. I updated my drivers to the very latest, and still no luck. Hopefully, I can get my hands on an upgrade soon.

Probably https://developer.blender.org/D11868

Hi, Optix 7.3 support Maxwell cores, this is GTX 900 and up with Cuda Compute Capability 5.0 and up.
Your card should be supported.

No idea why it not work, you post on Blender Developer, I hope one of the developer steps by and give us more information.

Cheers, mib

The shadow catcher functionality (which was the most complex thing missing from the original Cycles X builds) has now been tied up and completed.
:anchor: T87837 Cycles X - Restoring Features (blender.org)

I would think the remaining items shouldn’t take near as long to complete, which means there is only a matter of weeks before progress gets made on the big new sampling enhancements.

2 Likes

Does ambient occlusion work now? I mean not in-shader

https://developer.blender.org/T87837

I have a GTX 960 on Linux, 470.57.02 driver.
That’s right, Cycles-X is acting strange. If your GPU is not detected, you open the scene and then in Edit > Preferences > System tab, you do some modification there, for example choosing Optix and then CUDA again. In this way there is some kind of refresh and GPU will be available in render tab.
Regarding OptiX denoiser, it gives corrupt results if you choose Color+Albedo+Normal (Normal problem apparently).

1 Like

Thanks, I forgot that page…

Some news for AMD GPU owners.

In last Rendering Meeting Notes Brecht mentioned AMD’s HIP implementation of Cycles-X:

If I understand it correctly HIP (Heterogeneous Interface for Portability) is a compatibility layer that can take CUDA code and run it with negligible performance penalty. So optimistic assumption would be that AMD GPU rendering speed not only could match CUDA but it can stay on that level at least for some time.

On the other side HIP is not a drop-in replecement for CUDA and some initial tweaking and optimisations are required to achieve same level of performance.

12 Likes

This scene have a lot of glass meshes, with 32 bounces in all paths, lit with a hdri only and with high DOF.

On a 3060ti with optix with yesterday build at 128samples

Cycles 60sec
Cycles X 22sec

Thats about x3 faster !

2 Likes

A test like that is hard to judge.

To be fair…

Show both the cycles render and the cycles X render results.

But make an effort to resolve it clean with minimal noise. Denoise or samples.

Show two clean renders and post the times

MMm… kinda but not exactly. https://github.com/ROCm-Developer-Tools/HIP

Basically at a high level HIP and CUDA allow you to compile C++ code to run on a GPU. HIP also includes some tools to translate CUDA library calls to AMD specific ones.

4 Likes

Oh, thank you for clarification @bsavery.
I took the description at the beggining of the README.md of the HIP project quite literally.
Anyway its good to know that you are making progress. Looking forward to test CyclesX in AMD flavour!

Refraction has always been problematic. I’m glad it has improved so much. I would have to wait something like an hour on my current rig for that kind of scene to converge (gtx970).

Do you think it is doable with Apple Metal for AMD/Apple Mx graphics?

2 Likes