How to quantify the performance of matrix-free operators and sparse arrays for matrix-vector products. Specifically, what is speed cost of having matrix-vector products dominated by function calls than by array addressing.
I'm solving a linear system arising from a PDE using iterative methods. Matrix-free methods have no memory issues and looks efficient for my case.
May be for a single matrixfree A wins(unless the function representing A is overkilled by internal computations), but how compositions and superposition of matrixfree operators behave.