Currently, the GPU eigensolver is hardcoded for 30 nonzero entries per row (the Hessian for the Pedersen force field). However, we may want to diagonalize adjacency matrices (3 nz per row) and the force field including "flatness" (>30 nz per row). What is the performance hit for making it dynamic? Is it bad enough to warrant templating?
Currently, the GPU eigensolver is hardcoded for 30 nonzero entries per row (the Hessian for the Pedersen force field). However, we may want to diagonalize adjacency matrices (3 nz per row) and the force field including "flatness" (>30 nz per row). What is the performance hit for making it dynamic? Is it bad enough to warrant templating?