Spin wave calculations require solving a generalized Hermitian-definite eigenvalue equation. On the CPU Sunny.jl uses the LinearAlgebra package. Rocsolver provides a strided-batched version for AMD accelerators. Using CUDA.jl, we created our own implementation for NVIDIA accelerators.
We would like to move these implementation differences to a portability layer such as JACC.BLAS.
Keita suggested extending EigenG-Batched to work with Hermitian matrices and work with the generalized eigenvalue problem. This would require significantly more work and coordination, but may also bring additional users to JACC.