Minutes_2023_05_02
esc edited this page May 3, 2023
·
1 revision
Attendees: Siu Kwan Lam, Val, Andre Masella, collison, dipto, Graham Markall, Guilherme, Kaustubh, Minjie Wang, Neo, stuart, Todd A. Anderson FPOC (last week): Val FPOC (incoming): Graham
NOTE: All communication is subject to the Numba Code of Conduct.
Please refer to this calendar for the next meeting date.
- numba 0.57.0 and llvmlite 0.40.0 release finalizing
- Projects have been tagged
- Waiting on artifacts
- Tagged versions of conda packages uploaded to anaconda.org
- Wheels still missing
- Need uploading after the meeting
- conda-forge not sure about SVML patches, but yes they need to be applied
- Updates from the subprojects
- Numba RVSDG
- Siu working on interface between Numba-RVSDG library and Numba itself
- Good progress in general in last 2 weeks
- NIL
- Not much progress due to release
- Priority is to focus on RVSDG first
- PIXIE
- Output format and tooling for self-describing libraries including bitcode and specific versions of functions
- Target for the new AoT compiler and Numba internal use
- Also no recent progress due to release process
- Numba RVSDG
- [#7980] https://github.com/numba/numba/pull/7980 (Siu)
- How to spell signature with default argument?
- Asking uses to spell with Omitted / Optional in a string signature is unusual, so some design / thought is required.
- PR got stuck at review due to other overriding priorities.
- If there is interest in this an offline discussion / PR discussion would be doable.
- Discourse thread forthcoming.
- 0.57Rc1 milestone still has 34 open things in it.
- https://github.com/numba/numba/milestone/63
- Graham will move the CUDA issues to their own board
- Document how many Python / NumPy releases we are going to support at a time.
- Needs a better policy, because it's random right now
- In practice, 4 Pythons (3.8-3.11 now) and 4 NumPys (1.21-1.24 now)
- Follow NEP29 which dicates which NumPy versions are supported and when versions go EOL
- Numpy 2.0 is coming, so this needs an answer soonish
- There may be more subtle details, such as type-systems, RVSDG uses, compiler pipeline changes, do we backport fixes for older NumPys, is compiler support uniform?
- Example: split between Python and NumPy scalars may be implemented only in the new pipeline, so perhaps NumPy 2.0 could only be supported in the new pipeline (hypothetically)
- Requires further consideration / iteration / discussion.
- Needs a better policy, because it's random right now
- How to do runtime array aliasing check? (Todd)
- Parfors currently has an issue where multiple slices can alias and inplace updates on one of them has unintended side effects on another
- Could a runtime alias check be inserted to catch this scenario and abort / error?
- Siu: there is no whole internal alias check
- Stu: there are some parts of the solution, (
extents_may_overlap()
) innumba/numpy/arrayobj.py
. - Todd: is this an implementation of NumPy's
may_share_memory()
?- No, it's an independent more naive implementation
- PR #3908 has a copy of an implementation of
may_share_memory()
: https://github.com/numba/numba/pull/3908.
- Merge [#8854]? https://github.com/numba/numba/pull/8854 (Mingjie)
- Some docs for ParallelAccelerator development are needed, but not necessary to hold up this PR
- Do we need a test for this?
- No test at the moment, it would be difficult to test correctness.
- To be marked as Ready to Merge
- Neo:
- Previously requested assistance / mentoring on Numba
- Assistance still required
- Could post on Numba Discourse Group: https://numba.discourse.group
- Numba on Mastodon: https://fosstodon.org/@numba
- Expose TargetLibraryInfo pass https://github.com/numba/llvmlite/pull/934
- numba#8919 - Make is so that nothing is compiled in the process of listing the available tests.
- numba#8920 - Migrate code base towards objmode fallback removal
- numba#8921 - Numba 0.57.0 Checklist
- numba#8923 - CUDA Vectorize appears to induce spurious device-to-host copies even with all inputs on device
-
numba#8926 - Numba cache files don't get invalidated with Python cache files
- Suggest to turn off the caching during development
- Read the docs on limitations of caching, both user-docs and developer docs
- suggest to use luk's patch https://github.com/numba/numba/pull/8396
- Suggest to use a git-hook during development to prune the cache
- Ask luk-f-a to perhaps comment on this issue
-
numba#8931 - non-deterministic
NaN
values inscipy.integrate.solve_ivp
when compiling function withnumba
- Seems an odd issue
- It may be due to pip packages because of the load chain, might be worth trying anaconda
-
numba#8933 - Replace deprecated imp module with importlib
- Will need to be done for 3.12
- Need a label on GH for
3.12
-
numba#8934 - jitclass reflected dunder methods do not work in a njit function
- Update docs with limitation
- Proper fix needs 6 months of full-time engagement
- llvmlite#937 - llvmlite 0.40.0 Checklist
- numba#8928 - guvectorize raises deprecation warning even if nopython=True is passed
- numba#8927 - Add a Numba power-on-self-test script and use in CI.
- merged - numba#8922 - Cherrypicks for 0.57 final
- merged - numba#8924 - Change llvmlite version pinning.
- merged - numba#8925 - Update release checklist template
- merged - numba#8929 - Update CHANGE_LOG for 0.57.0 final.
- merged - numba#8930 - Fix year in change log
- merged - numba#8932 - Fix 0.57 release changelog
- merged - numba#8935 - Cherrypick for 0.57 final release
- merged - llvmlite#938 - Cherrypicks for release 0.40.0 final
- merged - llvmlite#939 - Bump llvmdev build number to include the nozlib change for windows
- merged - llvmlite#940 - Update CHANGE_LOG for 0.40.0 final.