Skip to content

Minutes_2023_06_13

Jump to bottom Edit New page
esc edited this page Jun 14, 2023 · 1 revision

Numba Meeting: 2023-06-13

Attendees: Siu Kwan Lam, Andre Masella, collison, Da Li, Graham Markall, Guilherme, Jim Pivarksi, Kaustubh, Mingjie Wang, stuart, Todd A. Anderson, Matthew Murray, Luk, collison, Dipto FPOC (last week): Siu FPOC (incoming): Andre

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

Please refer to this calendar for the next meeting date.

0. Discussion

  • numba 0.57.1rc1 llvmlite 0.40.1rc1 releases -- both are live
  • RE: custom LLVM pass
    • analysis info does not have serialization so it must be redone in the isolated custom pass
    • cost of redoing analysis is not (might be) signifiant for PyOMP usecase
    • Action: Andre will look at this again and get a prototype ready for testing
  • How to recommend Numba extensions? e.g. PyOMP support OMP contruct
    • Should there be a way for Numba to communicate that extensions are available (e.g., if you try to access a scipy function, recommend numba-scipy)
    • Giving good information about what compilation failed for non-JITted user functions would be a necessary prerequisitie with value on its own; also for non-supported built-in functions
    • Could be integrated with overloading in a way to provide alternatives for specific type/built-in errors; limited due to code bloat in Numba
    • Relatively to implement if desirable
    • Part of the difficulty is the complexity of error messages that mostly provide developer-relevant error information
      • Not finding error information is the same as not having it
    • The official list would have to be in Numba, so there would have to be a process to add extensions and recognize the correct situations (more than just functions and datatypes)
    • Future actions unclear
    • Possible plan would be to register hooks on types for providing extra information?
    • Problematic as this would only be useful for the case where there is a Python API that requires a separate overloads (e.g., scipy and numba-scipy)
    • This only handles the case where the community is separate, so more of a social issue than technical
  • How can we help research/academic project to justify using or working on Numba?
  • Multiple array parameters that are views of some base array and effect on performance (e.g., vectorization). (Todd)
    • Should this be a Numba optimization pass
    • Facilitated by metadata that NumPy doesn't have and inference might be hard
    • This is attractive; if the IR is the problem, we should extend the IR
    • LLVM metadata might be enough to help the vectorizer figure it out; might require two paths for the case where it's provably safe

New "Ready for Review" PRs

1. New Issues

  • numba#9008 - Memory leak when passing string literal as an argument to a jitted function
  • numba#9010 - CUDA: support array copy similar to np.copy

Closed Issues

  • numba#9009 - problemme d executiion
    • warning: the issue is suspcisous; do not click links in it.

2. New PRs

  • numba#9007 - CUDA: Add comparison ufunc support
  • numba#9011 - Implementation and test updates to support NumPy 1.25
  • numba#9012 - RVSDG-frontend
    • pending numba-rvsdg package to re-enable CI
  • llvmlite#958 - fixup LLVM versions in version compat table
  • llvmlite#959 - Remove support for LLVM < 14
  • llvmlite#960 - add various bullets to release checklists and sync

Closed PRs

  • merged - numba#9004 - Skip MVC test when libraries unavailable
  • merged - numba#9005 - Fix: Issue #8923 - avoid spurious device-to-host transfers in CUDA ufuncs
  • merged - numba#9006 - link to version support table instead of using explicit versions

3. Short-term Roadmap

gantt: https://github.com/numba/numba/issues/8971

Add a custom footer
Add a custom sidebar
Clone this wiki locally