Dynamical matrix for local term

[epolack]

Open points that can be clarified after:

• Docs of new sternheimer
• Docs of compute_δHψ_sα

[mfherbst]

Timings are great, I agree. No need for the wrapper.

[mfherbst]

I'd be curious why @antoine-levitt , but not blocking for the merge.

[mfherbst]

Final nits

[epolack]

Actually, no it does not work. If x is an array, then this will fail (e.g., ERROR: LoadError: MethodError: Cannot convert an object of type Array{Float64, 3} to an object of type Float64). I tried to play around to change the function to something similar and as general, but I could not think of anything.

[mfherbst]

Yeah, but then just add another method on AbstractVector{<:Real} or does that not do the trick ?

[epolack]

No because of ForwardDiff. The encapsulation of complex is a bit tricky. I am looking if I manage to solve this.

[epolack]

Ok, seems to work with promote_type.

[mfherbst]

I found a parallelisation issue, that I need to fix and actually a way to completely avoid the convert_enforced.

We also need to check GPU support (will do once the CPU CI is green).

[antoine-levitt]

How is real_enforced better than enforce_real? Also beware that lowpass_for_symmetry is possibly slow

[epolack]

How is real_enforced better than enforce_real? Also beware that lowpass_for_symmetry is possibly slow

real_enforced was to match convert_enforced, but as it was removed, I will revert back the rename.

[mfherbst]

GPU CI did not go through cleanly (precompilation failed, so I have no proper output from the run). Do you guys have access to a CUDA GPU or should I debug ?

[epolack]

I ran the two GPU tests fine locally, so not sure what to debug. The iron one is a bit strange because of error LOBPCG is badly failing to keep the vectors normalized; this should never happen that pops-up at random.
The only think we touch GPU-wise on this PR is the removal of to_cpu functions in local.jl because [f for _ in stuff] should take care of this.

(Also, maybe it should be written somewhere that CUDA.set_runtime_version!(v"11.x") is mandatory per Libxc requirement?)

[mfherbst]

Ok, so I trust the GPU CI goes through. Let's see.