Multiple ray generation points (camera)

Is it possible to run ray tracing from multiple locations for the same object simultaneously

Hi @EvanYL, yes it is possible. You can decide in your raygen program how to map your launch index into a ray.

This means it is very easy to (for example) render an output image consisting of image tiles, where you choose a different camera based on which tile the launch index is in.

Or, sometimes people will pass in a 1D buffer of rays to the OptiX kernel, and every single ray might have a different origin. You could look at the optixRaycasting sample in the SDK for an example of this.

You have complete control over how rays and camera views are generated and organized, to allow for workflows like tiled rendering or texture baking or stereo or any number of scenarios.


Is there an upper limit of a total number of ray can be processed at the same time? Is it relate to the model of the GPU? Is there a simple estimate equation to calculate like the X number of RT core can process Y number of ray (no matter it is from 1 or multiple origin)

Please have a look into this thread and follow the links in there:

In theory you can have a as many optixTrace() call invocations inside your ray tracing kernel as you want when staying under the maximum launch dimension.
But you will compete against the operating system for the GPU resources.
If you are running on a GPU dedicated to compute tasks only (Tesla Compute Cluster (TCC) driver mode), there is no OS timeout like under the Windows Display Driver Model (WDDM since Vista, it got better in Windows 10 with WDDM2).

The RT cores on the top end Turing GPUs could already handle over 11 GRays/second. With complex scenes or shaders you will usually run into memory bandwidth limits first.

It’s usually the simplest approach to partition work into reasonably sized chunks and launch less work more often instead of using some super expensive kernel which does it all.
There are plenty of options in computer graphics to split work, e.g. into individual samples in a progressive Monte Carlo algorithm, into tiles of an image, into an image per light source etc.