|
Uni20
Asynchronous tensor and runtime library.
|
Optional customization point for backend-owned metric inner products. More...
#include <matrix_free.hpp>
Optional customization point for backend-owned metric inner products.
Implementations may provide this operation to compute x^T B y directly, using backend-owned scratch storage, cached B*y values, fused GPU kernels, or MPI collectives. Generic solvers should use metric_inner_product_or_apply so backends without this customization still get a correct default through metric.apply or B.matvec.