Question about the CUDA acceleration module in the planned SCHISM 's code development ?

[lj-cug]

GPU accelerated SCHISM development is practical now. There are several ways to implement GPU acceleration in ocean models:
(1) OpenACC accelerated Jacobian iterative solver module (Zhang Hongchun, et al. 2025)
(2) total CUDA C porting (POM.gpu-v1.0: Xu S., et al. 2015; Xie Jianbin, et al. 2025)
(3) external libraries like PETSc, AmgX, Ginkgo (SHYFEM team, 2022)
Method 1 is limited by flexibility and compiler (NVHPC must be used); Method 2 has problem with respects of integration with another modeling components and code sustainability. Method 3 is most practical and also easy to integrated with other advanced external libraries.
So could you tell me some basic information about the CUDA accelerated 3D barotropical module in SCHISM if convenient before publication and code release ? This will help me to adjust my code developing plan and avoid duplicated code developing. What's the implementation ways used for the 3D barotropical module in SCHISM as above-mentioned ?
Thanks!
LI Jian

[josephzhang8]

Do you have trouble accessing github? If not, I can ask Jose to grant you access to gSCHISM repository...

Yes we made the decision to go all the way using CUDA C++ and Jose put a lot of careful thoughts into efficient memory layout to take advantage of latest NVIDIA GPUs (Hopper, Blackwell). It can also take >1 GPUs.

[lj-cug]

My problem for accessing github was from the limitation in China's network management policy. I can access to github in every afternoon. (^^)
Now I only use the some codes from SHYFEM model to integrate the AmgX wrapper with the PETSc module in SCHISM_v5.13.0. The non-hydrostatic pressure part was testing. I think solving 3D Poission for GPU memory is more challenging.
Considering the interface unification and communication efficiency among nodes, PETSc library was necessary and powerful.