Dear NGSolve community,
in the current release we have redesigned GPU offloading in NGSolve. Kernels are JIT-compiled from strings.
We have merged CUDA and Metal to a common syntax, such that the same kernels compile for NVIDIA GPUs and Apple GPUs.
At the moment we have ported our existing CUDA kernels to commonGPU, and reworked the matrix-free operator application from scratch.
Examples are in
and updated CUDA examples
We hope it is a useful interface for your GPU experiments, and it will speed up your simulations.
We are happy to get your feedback.
Joachim