[munch117]

It's not the same algorithm: That code completes all 200 iterations unconditionally. For some parts of the input domain, the numpy speed-up is enough to compensate and make it faster anyway, but for other parts it will be slower.

[kzrdude]

Agree, even if it compensates by supplying a gradually shrinking compute mask, but not to all operations. But an effect is there, less populated chunks compute faster.