Slate: Transpose(Tensor(form)) -> Tensor(adjoint(form))

[sv2518]

This PR addresses #2262. It's a small codegen improvement which inlines transposes into the local assembly kernels generated by TSFC. I have done this on the highest level (in Slate) via def __new__.

[wence-]

Looks like there's an actual failure in the hybridisation notebook?

[wence-]

Other than the actual test failure in the build, I think this is good.

[sv2518]

Ok I think some tests were failing, because I took the ufl adjoint and not the one from firedrake.ufl_expr.

[wence-]

Now it looks like we have many fails because adjoint(linear_form) isn't defined, whereas Tensor(linear_form).T does work, I think?

[sv2518]

Mhh, maybe this is not the right way to do it after all. Maybe I need to do it at a lower level. The leftover issues we have are that a) M.T.arguments() != M.T.arguments[::-1] because M.T generates a new Tensor with new arguments and b) e.g. that M.T.T != M.

[wence-]

Mhh, maybe this is not the right way to do it after all. Maybe I need to do it at a lower level. The leftover issues we have are that a) M.T.arguments() != M.T.arguments[::-1] because M.T generates a new Tensor with new arguments

I think the difference is that:

adjoint(a)

Turns the argument with count 0 into the one with count 1, and vice versa

whereas

Tensor(a).arguments()[::-1]

just reverses the argument list.

b) e.g. that M.T.T != M.

This one should be the case, unless I misunderstand?

[sv2518]

Sorry I think I explained b) wrong actually. In the Slate optimiser we drop double transposes so that M.T.T is just M. If we introduce the change on Slate level we won't know when we have double transposes on terminal tensors anymore, so we can't do that optimisation. But in fact actually we don't need this optimisation on terminal tensors anymore if we go over the adjoint, I think, because we just do Tensor(adjoint(adjoint(form)) so that no extra operations in the Slate kernel are needed.
I guess I could work around a) by specifying reordered_arguments [EDIT: ok that does not help because I still need to switch the count argument].

[wence-]

Do you want to squash everything nicely, or shall I just do it in the merge commit? Thanks!

[sv2518]

Sorry I had to force-push twice because I screwed up the squashing the first way round. If you give me an approving review, I will merge tomorrow myself as soon as the tests are passing. :)
Thanks for feedback!

[wence-]

Thanks!

[wence-]

I will merge tomorrow myself as soon as the tests are passing.

I turned on auto-merge, so it might be done by the time you get up!