Question related to GPU/CPU section in ReactorCvode

[sachdevakunal]

Hi. For solving the reaction equations, ReactorCvode is generally used in PeleLMeX. The react function in ReactorCvode is divided into two regions, GPU and CPU. What exactly is the difference between the two regions? As far as I understand, the reaction integration takes place cell by cell in the CPU region, and in the GPU region, the state vector has all the information at all the cells at the current AMR level, and reaction equations are solved all together at once. Is that correct?

[baperry2]

That's basically the gist of it. On GPU integrating all cells at once (over a box, not the entire level) minimizes thread divergence. There are some other differences between CPU and GPU code paths in terms of how memory has to be allocated and what internal algorithmic options are available.