Backend-agnostic solver

MaterialPointSolver.jl is a high-performance solver that is backend-agnostic, meaning we only need to write a unified codebase that can run on different accelerator backends. The main computational process of the solver relies on kernel functions implemented by KernelAbstractions.jl, so users do not need to worry about the data migration process from the CPU to other hardware backends.

Users should use MaterialPointSolver.jl, and then selects the backend based on your hardware, for example:

julia> using MaterialPointSolver
julia> using CUDA

# using AMDGPU # for AMD   gpu
# using Metal  # for Apple gpu
# using oneAPI # for Intel gpu

Taking the CUDA backend as an example, we define automatic conversion methods in /ext/CUDAExt/devicehelpfunc_cuda.jl: host2device for transferring data to the device, and device2host! for downloading data back to the CPU from the backend.

Note

The reason this solution can succeed is that the structs' (including custom structs) fields only contain scalar and array types for the kernel function, and currently do not support other forms.

Although Julia has a garbage collection mechanism to automatically reclaim memory, we hope to clean up GPU memory usage promptly after each simulation. Currently, in /ext/CUDAExt/devicehelpfunc_cuda.jl, we use the function clean_device! to do this.

Note

For all functions related to the accelerated backend that are not part of the MPM process, a dedicated implementation may be required in the /ext folder.

Todo

Currently,

1. Memory clean can only be achieved on the Nvidia platform, and other platforms are waiting for it to provide the relevant API.

2. Metal on the Apple platform does not support FP64 computation precision.

3. Intel's GPU devices have not been tested, but the code implementation should be relatively easy. Ideally, it should only require filling in some parts of the /ext folder.