Zygote performances for simple function

At least for those two examples it works, but perhaps you have other constraints that weren’t expressed in the original post.

Back to the allocations, I had a look at both the generated Julia and LLVM IR with Cthulhu. It appears that replacing -x with a different operation (e.g. sin(x)) has the same outcome. Yet I’m only able to see one allocation at either the Julia or LLVM level. It feels like this is somehow hitting some compiler heuristic, but I can’t figure out what.

1 Like