Minutes_2022_04_05

Numba Meeting: 2022-04-05

Attendees: Siu Kwan Lam, Andre Masella, Benjamin Graham, Ehsan Totoni, Graham Markall, Nick Riasanovsky, Shannon Quinn, stuart, Todd A. Anderson

NOTE: All communication is subject to the Numba Code of Conduct.

Please refer to this calendar for the next meeting date.

Release 0.56 next steps
- Agree on an ETA (probably Mid/End May)
- Decide and discuss what will make it and what will not
- Do this next week
Wheels announced - https://numba.discourse.group/t/wheels-for-apple-silicon-m1/1282
- check discourse and test wheels on M1 if possible
CUDA Kernel caching (#7944, see below)
Which LLVM(s?) to support
- HPC may be more constrained and would like older versions too (llvmlite: #802)
- The LLVM release cycle also seems to be accelerating
- Depends on how "different" new LLVM versions are
- Concern about testing different LLVMs
- Only llvmlite binding? Or should this work with Numba too?
- Suggest to keep older code paths, but stop testing them
- 12 and 13 API compatbile
- Shouldn't acquire "many" older releases, only keep the last one or two, part of the update process should be to remove "very old" versions.
- This policy will need to be documented in the appropriate place
- Will be a "best effort"
RFC on the mission statement: https://github.com/numba/numba/pull/7950
- "Scientific Computing" <-- the NumPy tagline
- Perhaps start by quoting the NumPy tagline: "The fundamental package for scientific computing in Python"
- Also mention terms like "number crunching"
- What about the non-goals, like webservers and such?
#7960 - discussion
(If time allows) How much magic is too much?
- Todd: limits by compile time overhead
- Nick: cannot guess user intent; e.g. using a HW accelerator. debuggability?
- Siu: fusing compiler and API?

#7942 - Issue using with objmode with custom output type
#7943 - CUDA Vector Types
#7945 - Debugging: optimization issue with parallel:True
#7948 - No variable slicing with None, as optional type not inferred
#7951 - Wheel Support for Mac Arm / Apple Silicon / M1
#7952 - Cache temp filename not unique across nodes
#7953 - insert() casting entire list
#7954 - Provide a way for lowering to know what Compute Capability is being targeted
#7955 - Add support for numpy.ndarrays in numpy functions in ALL parameter positions where numpy supports them

#7947 - Support @overload(intrinsic), i.e. extension API used within the API codegen?!

#7944 - [WIP] Pathfinding for CUDA kernel caching
#7946 - instructions for creating a build environment were outdated
#7949 - Add Cuda Vector Types
#7950 - mission statement
#7956 - Stop using pip for 3.10 on public ci (Revert "start testing Python 3.10 on public CI")