|
Uni20
Asynchronous tensor and runtime library.
|
Optional customization point for direct vector norms. More...
#include <matrix_free.hpp>
Optional customization point for direct vector norms.
A backend may compute norms more accurately or efficiently than a generic sqrt(real(inner_product(x, x))), for example by using a fused device reduction or an MPI norm collective. Solvers should call norm_or_inner_product so backends without this customization still get a correct default from the inner product.