Question: batched FE_Values reinit?

[GrahamBenHarper]

I have a brief question related to FE_Values and GPUs. Usually there is some notion of "worksets" or batched computations, but when I look at step-64, it seems the main assembly routine consists of doing a reinit on each cell

for (const auto &cell : dof_handler.active_cell_iterators)
    if (cell->is_locally_owned())
      {
        cell_rhs = 0;
 
        fe_values.reinit(cell);
...

Have things like worksets or batched cell reinits been considered, or are these technical details hidden away under the interface? Thanks for your time :)

[masterleinad]

The point of step-64 is the matrix-free interface on GPUs and assembling the right-hand side is not the focus but we could likely also use a matrix-free approach for that similar to what step-37 is doing, see https://www.dealii.org/current/doxygen/deal.II/step_37.html#LaplaceProblemassemble_rhs.

[GrahamBenHarper]

Ah, that's exactly what I had in mind! Is that capability on GPUs currently possible, but not currently demonstrated in a tutorial program?

[masterleinad]

Ah, that's exactly what I had in mind! Is that capability on GPUs currently possible, but not currently demonstrated in a tutorial program?

In the actual matrix-free operator, the cells are batched (avoiding write conflicts between dofs) internally already.

[GrahamBenHarper]

Got it; that's what I assumed, but wanted to double-check. Thanks @masterleinad!