Minutes_2021_10_05

Numba Meeting: 2021-10-05

Attendees: Siu Kwan Lam, Graham Markall, Guilherme Leobas, Luk, Nick Riasanovsky, stuart, Todd Anderson, Valentin Haenel, Sebastian Berg, Jim Pivarski, brandon willard

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

0. Feature Discussion

• tuple_setitem negative indexing support
  • Add warnings in the docstring for lack of negative indexing support
  • Note to reviewers: pay special attention to use of unsafe intrinsics
  • Safe mode?
    • A lot of work to add checks
    • false sense of security
    • any low-level tools to detect OOB memory store?
• #7359: Extend support for nested arrays inside numpy records
  • Graham will help on the CUDA failures
• #6972: Wrapper or type to avoid inlining
  • Perhaps, WAP should be pickled
• 0.54.1 - release by the end of the week
• [NumPy] New DTYPE API with runtime switch to enable: https://github.com/numpy/numpy/pull/19919
  • once merged, we can test by setting an env-var/

1. New Issues

• #7436 - Suppress fprintf(stderr, ...) from tbb
• #7445 - Compile Slowdown 0.50 -> 0.55, CUDA and NJIT
• #7446 - Support arbitrary fixed-size integers
• #7448 - Allow mixing types (C or Python) within single jit-function
• #7451 - Corruption on ppc64le/arm64 (potential codegen/LLVM opt bug)
  • Problem coming from OOB -1; not actually arch
• #7452 - LLVM ERROR: Symbol not found: __sync_fetch_and_add_4 running tests on armv7l (armel)
• #7455 - Non-deterministic hang when testing set membership

Closed Issues

• #7450 - https://github.com/numba/numba/issues/new
• #7454 - Add support for looping over empty tuples

2. New PRs

• #7437 - Add support for np.broadcast_shapes
• #7438 - eliminate lowering of Dispatcher as dynamic address
• #7440 - Refactor TargetConfig naming.
• #7441 - Permit any string as a key in literalstrkeydict type.
• #7442 - Add some diagnostics to SVML test failures.
• #7443 - Refactor template selection logic for targets.
• #7444 - use correct variable name in closure
• #7447 - cleanup Numba metadata
• #7449 - BUG: Fastcall fix ufunc methods (NumPy 1.21 support)
• #7453 - CUDA: Provide stream in async_done result

Closed PRs

• #7439 - [do not merge] debug

3. Next Release: Version 0.55.0/0.38.0, RC=Early Nov

• Request for 0.55
  • broadcast_to https://github.com/numba/numba/pull/7119
    • driven by need for vectorize support in aesara
    • Now merged