+/pykokkos

git clone https://git.julia.blue/+/pykokkos

Joseph RyanImplement basic support for Debug ExecSpace1a80dfe

master
1.8 KiB45 linesraw
1  + support flow control in work units
2  + pass single vars back from reductions
3  + timing for each step of workload
4  + multidimensional views (1, 2, and 3)
5  + error messages using the spans from the AST
6  + turn into proper python package
7  + translate function calls (using whitelist of math related and utility functions)
8
9  - local vars
10  - dataclasses
11  - inline functions
12  - add option to not initialize views (in python)
13  - don't transfer views that arent modified or at least used
14  - Exec spaces:
15    + OpenMP
16    - GPU (CUDA)
17    - PThreads
18    - Serial
19    + Python (for debugging, currently doesn't allow reduction)
20  - test exec space functionality and performance: cuda/openmp/thread
21
22  - Consider using Python lambdas instead of @work_unit.
23
24  - We encountered an issue (on xps2) when protobuf was large.  Need
25    to understand if this is some inherent limitation or something
26    that we can adjust as we go.
27
28  - We start compiling units before we check that some methods
29    actually exist.  For example, we do not check if use_results exist
30    and if it does not, then things fail after compilation (which
31    takes long).  It would be good first to check if all methods are
32    present (that we expect) and then move forward.
33
34  - DISCUSS: Rather than having specific method names for
35    "run_work_units" and and "use_results" maybe we can let users
36    write any methods they want, but they have to annotate them with
37    "@kokkos_run" and "@kokkos_results".
38
39  - LATER: Develop "translation" from C++ to Python (at least for some
40    Kokkos kernel).
41
42  - LATER: compilation can take substantial time, so do it
43    just-in-time is costy.  We want to think how to optimize/avoid
44    this in the future, e.g., incremental compilation, generating many
45    files/classes, etc.