Minutes_2020_09_01
Valentin Haenel edited this page Sep 1, 2020
·
1 revision
Attendees: Siu, Graham, Hameer, Guilherme, Mathieu, Todd, Val, Stuart, Juan
- 0.51.2 fixes
- critical bug on caused by certain CFG
- in the future, we need to be more careful with algorithm changes or other big changes in a point release.
- Longer RC, request for testers and signoff.
- TexasBBQ (integration tester)
- suffers from package conflict and unstable-ness due to changing versions
- We should be like other compiler projects with long RC and community testers to signoff for a official release
- TexasBBQ (integration tester)
- State of mainline.
- TBB failure affecting mainline CI
- symptom: load TBB, exec parfor loop, fork process from CPython, retire fork, rerun parfor loop.
- we should workaround this and notify TBB devs
- Czars - Split "performance" label into "performance - compilation", "performance - runtime".
- Open meeting next week
- Longer RC period + process for releasing, tagging, testing and getting signoff from the community?
- Juan: update on issue to caching cfunc
- has potential fix
- will open PR
- Stuart: discourse topic from stumpy author (Sean Law)
- Topic: trigger eager compilation using caching at install/init
- doable by custom decorator to trigger eager compilation
- there's probably needs to it inside numba; e.g. typed containers, str library
- Stuart: exploring if Numba can compile itself
-
#6191 - nested dictionary as attribute of jitclass failed
- user question
- spelling type is unintuitive
-
#6190 - Memory leak !!! Please give me a hand
- generator leak
- search for dups + close
-
#6189 - Integer to string conversion is extremely slow
- int->str impl is fresh and needs work to improve
-
#6188 - Failed in nopython mode pipeline (step: nopython mode backend)
- user implemented their own import
- #6186 - DFS-based backedge algorithm can be slow for pathological cases
-
#6184 - Improve error message for NumPy alias type used as dtype in ArrayNdCtors
- A starter patch exists, any takers?
-
#6183 - CUDA fastmath: use fast math trig / exp / log / fdivide functions
- flag may not be turned on for anything other than CPU?
- #6182 - NumPy 1.20 support
- #6181 - NumPy 1.19 support
- #6180 - Type inference recursion error for self referencing types
- #6179 - No module named 'numba.targets'
- #6178 - umap.UMAP. error
-
#6177 - Importing AOT compiled function in 0.51.0 still requires numba installed
retrospective deprecation?- it should work as the NRT is statically compiled into the AOT compiled BLOB?
perhaps just fix the docs?- Test in subprocess, check if numba imported?
- #6175 - ir_utils causes a DeprecationWarning in numpy 1.20
-
#6173 - After the first function with default arguments is compilled, all subsequent functions with default arguments that get compilled will be very slow to run
- Fingerprinting / dispatch logic?
- #6172 - Release 0.51.1 checklist
-
#6170 - ImportError: cannot import name 'npyimpl'
- pip can have multiple package versions installed
- can we detect this?
- why isn't the llvmlite version mismatch detected, occasionally?
-
#6169 - Compilation is slower with overload_method than jitclasses for similar functionality
- Not high priority
- #6164 - Assign np.number its own type for comparisons
-
#6163 - Better error message in
with ... as ...
- PR #6185
- #6174 - Cannot install numba - failing to build wheel for "llvmlite" - please help!
- #6167 - #6147 broke CUDASIM
- #6195 - PR 6187 Continue. Don't visit already checked successors
- #6187 - [WIP] Don't visit already checked successors
- #6185 - [WIP] Better error message on NotDefinedError
- #6176 - Remove references to deprecated numpy globals
- #6166 - Remove Python 2 compatibility from numba.core.utils
- #6165 - #5425 continued
- Requests for 0.52
- Fast(er)
typed.List
/typed.Dict
? Doesn't have to be in 0.52, next 3-5 months is fine.
- Fast(er)