Collinear spin GGA kernel #424
Conversation
| σ_real[2, :, :, :] .+= ∇ρ_real[1, :, :, :, α] .* ∇ρ_real[2, :, :, :, α] | ||
| σ_real[3, :, :, :] .+= ∇ρ_real[2, :, :, :, α] .* ∇ρ_real[2, :, :, :, α] | ||
| σ_real[tσ(1, 2), :, :, :] .+= ∇ρ_real[1, :, :, :, α] .* ∇ρ_real[2, :, :, :, α] | ||
| σ_real[tσ(2, 2), :, :, :] .+= ∇ρ_real[2, :, :, :, α] .* ∇ρ_real[2, :, :, :, α] | ||
| end |
There was a problem hiding this comment.
All this is quite messy due to the multiple formats used. It's a bit better in the densities branch, so we should really try to merge the branches, but I understand if you don't want to do it yet. Once it's done, it would be good to document more clearly (maybe with variable names) what arrays are in "DFTK format" [i,j,k,sigma] and which are in "libxc" format [sigmatau,i,j,k]
There was a problem hiding this comment.
Or maybe we should add one extra layer of indirection on top of the returned terms object, which does the mapping and adds the strange (1/2) factor automatically, only returning a view into the 3D real-space grid. That could help to make the contractions more readable (because it's just elementwise multiplications and sums so no further indexing needed), but it hides which indices are contiguous.
There was a problem hiding this comment.
I'll leave that for next time when I'll merge the branches, but I think this might be the way to go.
No description provided.