02-11-2017, 08:14 PM
Their code is not too optimized for our use-case. We get a very high speedup from the fact that we query a lot of noise values in a uniformly-spaced grid, which means a lot of the calculations can be cached. That's actually the magic behind why our noise implementation is (much) faster than the Simplex noise. I'm not sure if such optimization is possible on the GPU, where memory-access (caching) is an expensive operation, unlike calculating from scratch.