GPU TODO list

[GVigne]

Following @antoine-levitt's advice, here is a short TODO list of additional features worth implementing for GPUs.

Short term (see also #794)

• CI (see Add GPU CI via gitlab #796 )
• Add some more developer information to the docs. In particular add that CI is run automatically only when on:
  • master
  • a branch with a name ending in gpu (e.g. lobpcg-gpu)
  • when manually triggered
  • on the DFTK.jl main repo (i.e. no forks)
• Find out what is happening with the drop! function and speed up the mapreduce part
• Speed up the view when doing an FFT. Removing the bounds check doesn't seem to do much, so maybe there is no "trick" to make it go faster. Instead, we could work on the mapping. For example, Antoine pointed out that if basis.variational is false, the mapping is the identity, so we could do:

if basis.variational 
    f_fourier .= view(f_real, kpt.mapping)
else
    f_fourier .= f_real

Medium/long term

• Xc term: I think a good way to start with this is by working on the DftFunctionals package
• Compatibility with automatic differentiation
• Get stresses to work

Miscellaneous

• Support for an other architecture: in theory, one only has to define the corresponding DFTK.Architecture. However, it is highly likely that there will be a few bugs or missing features (for example linear algebra functions) from the corresponding GPU package.
• Add support for more solvers (NLsolve, CROP). There have been discussions of removing the dependency in NLsolve from DFTK: the replacement solvers should keep in mind GPU programming and try to be compatible.
• Investigate on the gmres! function from IterativeSolvers: it doesn't seem to be CUDA-compatible for now. Making it CUDA compatible would allow us to use the χ0Mixing.