Error sparsifying an icmp about enzyme HOT 14 CLOSED

@martinjm97 this PR #1712 tells Enzyme to act conservative on non-affine things which hopefully fixes your issue here (aka hopefully having the effect of saying non-differentiable vars can be non-affine)

from enzyme.

@martinjm97 I don't see the relevant code, can you upload to enzyme explorer (enzyme.mit.edu/explorer)

from enzyme.

ah I see, the code is not the same as in the CI test harness

e.g. see https://github.com/martinjm97/sparse_deriv_enzyme/blob/0650345099790f8ac4e3471de1d7330c543fd9b5/eigenanalysis/eigen_analysis.cpp#L75

vs

Can you just use the version from CI (e.g. https://github.com/EnzymeAD/Enzyme/blob/main/enzyme/test/Integration/Sparse/eigen_analysis.cpp) which should have the relevant setup / code, and then add the harness onto it?

from enzyme.

Sorry, the code I'm referring to isn't: https://github.com/martinjm97/sparse_deriv_enzyme/blob/0650345099790f8ac4e3471de1d7330c543fd9b5/eigenanalysis/eigen_analysis.cpp#L75

It's https://github.com/martinjm97/sparse_deriv_enzyme/blob/cpp-viz/harness/eigenanalysis.cpp. Does the link not work for you?

Here I use matrix.h as in CI. Everything is the same modulo some minor changes to types.

from enzyme.

So I don't see any autodiff call in that file? Nor any todense

from enzyme.

Sorry, just edited! I realized I was on main instead of cpp-viz. The link should now be correct.

from enzyme.

Do you see it here: https://github.com/martinjm97/sparse_deriv_enzyme/blob/cpp-viz/harness/eigenanalysis.cpp.

from enzyme.

Yeah so there's still indirect indexing there via faces (https://github.com/martinjm97/sparse_deriv_enzyme/blob/bba6af8b79d7aa34bd6deceb32f775a884a2c5a2/harness/eigenanalysis.cpp#L103)

from enzyme.

That the same as in

I think it's okay because we don't differentiate wrt that part of the computation.

from enzyme.

I mean I think that's what it is yelling about. We can extend it to add support for that [and be conservative and assume it is always the worst case condition].

from enzyme.

It's weird that it's working in CI, but not in the version in harness.

I'll can try again from scratch and see if there's some mistake that popped in somewhere.

from enzyme.

The CI one does a different L vs M, and the one in CI had the affine-ification, the other didn't

from enzyme.

Yeah. You're right. I'll revert and debug.

from enzyme.

Good point. I'll try getting L working

from enzyme.