@JoOkuma brought up a strong possibility for computing midband spatial frequencies more quickly: multiplying by a mask (fast on a GPU) instead of indexing the mask. See more discussion here.
I'm breaking this out into it's own issue because in my brief trial couldn't recreate a speedup, but in further discussion with @JoOkuma he convinced me to keep trying.
@JoOkuma brought up a strong possibility for computing midband spatial frequencies more quickly: multiplying by a mask (fast on a GPU) instead of indexing the mask. See more discussion here.
I'm breaking this out into it's own issue because in my brief trial couldn't recreate a speedup, but in further discussion with @JoOkuma he convinced me to keep trying.