Some Questions

[liujiangling]

Hello author team, I am a graduate student specializing in SWE-GPU. Regarding the cuda part in your beta_version, I have some questions, including but not limited to GPU interface design, domain manager, and CUDA kernel:

1. Is the attribute design of cpu_to_gpu or gpu_to_cpu in various functions in the GPU interface file for testing purposes? If we use the GPU to obtain an acceleration ratio, then during the use of each kernel function, data is constantly transferred between CPU and GPU, greatly wasting the acceleration advantage of the GPU. Why not load all the data required for simulation into the GPU at once after GPU initialization or warm-up?
2. During the design process of the CUDA kernel, some kernels utilize Vertex data for computation, while others utilize Centroid data. Is there any new design idea for future optimization?
3. Regarding boundary conditions, I believe they should also be handled as CUDA kernels. In cases where the boundary conditions are complex, it is also crucial to use GPU batch assignment processing.

[stoiver]

@liujiangling Happy to discuss. But development of anuga is now undertaken at https://github.com/anuga-community/anuga_core and it would be great if you created an issue over there so the anuga development team can respond to your questions.

By the way, we are now concentrating on using openmp offloading to generate gpu code for anuga (have a look at the develop branch.