Any chance of seeing a more compute like version of "CrestFillMaskArtefacts"?

[dkyeck]

Please describe the desired feature/behaviour. CrestFillMaskArtefacts right now uses a lot of load texture instructions.

Supporting images/mockups, or example reference videos are very helpful to communicate ideas. using a more compute like approach utilizing group shared memory should reduce the number of load texture instructions significantly.

[magenta404]

Tried it and performance was worse when measured using Nsight. But maybe I was doing something wrong. You can see it for yourself here: perf/fill-mask-shared-memory.