Device-resident and CPU-only implementation performance

Determine the performance and implementation behavior of Togeari in CPU-only and fully device-resident realizations.

Background

The reported prototype uses a heterogeneous implementation: sparse operator products run on a GPU, cuDSS factorization and solves use hybrid memory, and reduced eigensolves and FGMRES control run on the host. The measured timings therefore do not characterize either a CPU-only implementation or an entirely device-resident implementation.

The authors explicitly identify both alternative execution models as unresolved implementation cases, leaving their setup costs, memory behavior, and solve performance undetermined.