I think a word of caution should be provided. The sort of spin-lock used here while may appear to work in the vast majority of scenarios, the forward progress guarantee required doesn't exists on all (or even many) modern GPUs. It's entirely possible for the thread holding the lock to get suspended, and never resume due to another thread spinning while waiting on the lock.
I really like Tobias, and think they are doing some great stuff on The Machinery, but I think he should (and eventually will be forced to) reconsider how he is implementing picking.
Indeed. I discussed this with a bunch of my GPU buddies, and we came to the conclusion that this code is not safe in the face of lack of forward progress guarantees. Also, shameless self-promotion, I have an extensive discussion of GPU forward progress in my recent blog: https://raphlinus.github.io/gpu/2020/04/30/prefix-sum.html
Nice that it works with various deformations, but because of the latency and more shaders i'l skip this. I have still learned some things from this post.
I really like Tobias, and think they are doing some great stuff on The Machinery, but I think he should (and eventually will be forced to) reconsider how he is implementing picking.