I wondering what the potential and limits are to writing to shared arrays on the GPU in julia, and generally.
My WIP spatial simulation package DynamicGrids.jl seems like a prime candidate fro running on the GPU, as the majority of model components simply do some transformation of a single array index. But there are some components that need to write to arbitrary points in the array based on the current cell value, such as random jumps in organism dispersal models.
This means that multiple processes would occasionally want to write to the same cell at the same time, say adding to an existing population value in that cell.
Is that something that is possible to do with CUDAnative?