Single-node asynchronous GCP using Kokkos
This is the issue for tracking the development of a single-node asynchronous GCP algorithm using Kokkos
- Uses HogWild!-esque approach with a single copy of GCP factor matrices shared between all threads
- Builds on "fused" approach combining semi-stratified sampling and MTTKRP evaluation for gradient
- Further incorporates SGD "step" method for updating factor matrices with each sample gradient
- Relies on Kokkos atomics API for such updates