Quadratic Memory Usage (mk2) about enzyme HOT 4 CLOSED

So once again this code should not require any cache to be created (let alone an O(n^2) cache).

It turns out that the relevant optimization ('even if this is memory is written over, you don't need to cache it unless the potential write occurs between the place you would need to cache and its use') was already implemented in my dev branch (fortranrb) which is currently in the process of being cleaned up and merge [as of ~1.5 weeks ago it had ~40 commits of various interesting optimizations].

If you don't mind living on the bleeding edge (which by virtue of needing cleanup may be a bit more unstable) checking out the fortranrb branch should reduce addditional memory to zero. Hopefully the remainder of these optimizations should be cleaned up and merged into master within a few days.

I'm going to leave this issue open until everything is merged to master.

from enzyme.

Thanks. Ok, I'll try to live on the bleeding edge.

from enzyme.

I tried it on your branch and it works 👍

from enzyme.

Merged into master by #120

from enzyme.