These changes significantly reduce the GPU usage of smoke on my computer. The idea is rather simple: instead of scheduling compute for the whole window and then filtering out the border region only, directly schedule computes for each rectangle of the border.
The second commit improves the situation a little bit more by scheduling all rectangles in parallel.
These changes significantly reduce the GPU usage of smoke on my computer. The idea is rather simple: instead of scheduling compute for the whole window and then filtering out the border region only, directly schedule computes for each rectangle of the border.
The second commit improves the situation a little bit more by scheduling all rectangles in parallel.