+ support flow control in work units + pass single vars back from reductions + timing for each step of workload + multidimensional views (1, 2, and 3) + error messages using the spans from the AST + turn into proper python package + translate function calls (using whitelist of math related and utility functions) - local vars - dataclasses - inline functions - add option to not initialize views (in python) - don't transfer views that arent modified or at least used - Exec spaces: + OpenMP - GPU (CUDA) - PThreads - Serial + Python (for debugging, currently doesn't allow reduction) - test exec space functionality and performance: cuda/openmp/thread - Consider using Python lambdas instead of @work_unit. - We encountered an issue (on xps2) when protobuf was large. Need to understand if this is some inherent limitation or something that we can adjust as we go. - We start compiling units before we check that some methods actually exist. For example, we do not check if use_results exist and if it does not, then things fail after compilation (which takes long). It would be good first to check if all methods are present (that we expect) and then move forward. - DISCUSS: Rather than having specific method names for "run_work_units" and and "use_results" maybe we can let users write any methods they want, but they have to annotate them with "@kokkos_run" and "@kokkos_results". - LATER: Develop "translation" from C++ to Python (at least for some Kokkos kernel). - LATER: compilation can take substantial time, so do it just-in-time is costy. We want to think how to optimize/avoid this in the future, e.g., incremental compilation, generating many files/classes, etc.