5 ms·

I previously read this: https://cachemiss.xyz/blog/parallel-reduce-and-scan-on-the-GPU https://cachemiss.xyz/blog/parallel-reduce-and-scan-on-the-G... whose im

I previously read this: https://cachemiss.xyz/blog/parallel-reduce-and-scan-on-the-GPU https://cachemiss.xyz/blog/parallel-reduce-and-scan-on-the-G... whose implementation seems a lot simpler than what is presented here. What is the main difference here? I've only dabbled a bit in Vulkan. (I wrote a pathtracer that I wanted to speed up using compaction; for which I needed a prefix sum)

raphlinus 6y ago
It's a good blog, but only solves a much simpler version of the problem - it only computes prefix sum of a workgroup worth of data. It does not carry the sums forward from workgroup to workgroup, which is the entire point of the decoupled look-back approach. The per-workgroup implementation in that blog is basically similar to mine, though I've put more effort into maxing out the partition size (using as many registers as available).