Diary of a Path Tracer – Sampling Strategy

Index: https://cranberryking.com/2020/08/22/diary-of-a-path-tracer-the-beginning/

Intro

In our last post, we looked at the SSE optimized BVH traversal of cranberray. BVH construction and traversal is a fairly complicated portion of cranberray; cranberray’s sampling strategy however is quite a bit simpler.

Sampling in cranberray is simple. We first split our image into a grid representing every pixel of our image. (Note that a rectangular pixel is a simplifying assumption [4]) Then we take our pixel box and sample it X number of times using X different sample points and average the results. Our sampling strategy determines how we spread out those X different sample points within our pixel rectangle.

The simplest strategy is simply to select a point randomly within our pixel box for every sample. However, this has the undesirable property of potentially introducing clumping causing us to miss some desirable information. [2][3]

Instead, cranberray makes use of a simple technique called N-Rooks sampling.

N-Rooks

N-Rooks is what’s known as a low discrepancy sequence. It has the desirable property of avoiding the clumping of samples that random sampling has and avoiding the aliasing issues of regular sampling. We won’t be explaining low discrepancy sequences here, but I refer the interested reader to [2] and [3] for excellent treatments on the subject.

N-Rooks sampling is very simple.

You first generate jittered samples along the diagonal of an X by X box.

And then you shuffle the columns of said box by starting from the leftmost column and randomly swapping with the columns to the right of our current column. (Including our current column). This is known as the Fisher-Yates shuffle. [6] It has the desirable property of uniformly selecting from any of the X! permutations of our columns.

And that’s it! We now sample our scene using these sample points and average their results to get our final output.

In our next post, we’ll be looking at BRDF multiple importance sampling using the balance heuristic as well as the GGX-Smith microfacet specular BRDF. Until then, happy coding!

Future Work

While writing this, I was considering the possibility of selecting the value of our pixel using a different strategy than the mean filter used with Monte Carlo integration. I would have to do some further digging to see if this has been presented anywhere. Something along the lines of a Gaussian filter or potentially something else entirely.

I would also like to implement other low discrepancy sequences in cranberray if ever time permits. N-Rooks was very simple, it would be interesting to be able to select between different sampling strategies.