Minutes_2022_08_09

Numba Meeting: 2022-08-09

Attendees: brandon willard, Graham Markall, Kaustubh Chaudhari, LI Da, Luk, Shannon Quinn, Todd A. Anderson, stuart, Siu Kwan Lam, Val, Jim Pivarski

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

Please refer to this calendar for the next meeting date.

0. Discussion

• Dropping support for old Python versions:
  • https://github.com/numba/numba/pull/8318
  • https://github.com/numba/numba/pull/8319
  • https://github.com/numba/numba/pull/7920
  • Question: what is the minimum supported Python for 0.57 and when can/should we merge the patches to remove support?
  • What implications does this have for CI??
  • Stencil tests rely on 3.7 AST, do they need an upgrade? #8325 needs to be merged first.
  • Consensus: #8325 needs to be merged first then we can merge the other PRs
  • Consensus: Python 3.8 will be the oldest version of Python supported by 0.57
• Dropping support for win-32 and linux-32
  • What does Anaconda distro support look like?
  • How does this co-incide with dropping of Python support
  • What about ARMv7 / AArch32? (Support statement needs an update, this is no longer supported)
  • https://github.com/numba/numba/issues/7452
  • Consensus: linux-32 to be removed from code immediately and from the Anaconda internal farm after 0.56.1
  • Consensus: win-32 to removed also as per https://www.anaconda.com/blog/new-release-anaconda-distribution-now-supporting-m1
• Dropping support for old NumPy versions: https://github.com/numba/numba/pull/8283#issuecomment-1197505150
  • Only 1.19 and upwards?
  • Need to check on the platform support
  • Will this have any implications
  • Consensus: Runtime: 0.57 will support only 1.19 and above
  • Consensus: Build time: probably ok to use 1.19 to build
• Issue #8309 / PR #8310 - CUDA: Atomic addition on complex components does not work in 0.56:
  • This is a regression due to changes in PR #7999, which made array.real and array.imag overloads
  • This is a problem because:
    • The overloads return arrays (not permitted in CUDA device functions)
    • Overload can't be resolved from the low-level API (used to implement CUDA atomics)
  • One fix (in PR #8310) is to revert this change and use the low-level API for .real and .imag again
    • Pros: it works, and is a relatively simple fix
    • Cons: it is a step backwards in the core implementation
  • Another fix is to make overloads returning arrays work on CUDA
    • Pros: Retains forward direction in the core, a general CUDA improvement
    • Cons: Lots of work - needs NRT, change to call convention, and rewriting of many CUDA implementations as overloads
    • Siu tried doing this but got stuck on (1) target_extension problems (active context not CUDA) and (2) _generate_real_imag_attr.<locals>.intrin_typing.<locals>.codegen (an @intrinsic) not found in lowering registry.
  • Does a third way exist?
• Resolution: Merge #8310 after review for 0.56.1.
  • Look into issues preventing it working with overloads afterwards
• PR #8294: CUDA: Add trig ufunc support
  • Works for the CUDA target, but adding simulator support is difficult.
  • Does it need simulator support?
    • Option 1: use __array_ufunc__ to fix implementation.
    • Option 2: Try making the "Fake within CUDA kernel array" an actual ndarray subclass.
    • Option 3: If neither of the above approaches succeed, discuss whether simulator support is required again next week.
• Compilation speed benchmark - Luk

1. New Issues

• #8303 - Allow multiple outputs for guvectorize on CUDA target
• #8304 - Python 3.11
• #8305 - Len of two concatenated Bytes objects is 0
• #8307 - Ambiguous overloads are allowed for calls to jitted functions inside other jitted functions
  • Needs poking, else it is operating in "surprise mode".
• #8309 - Numba 0.56 does not support atomic.add on arrays of complexs anymore
• #8311 - presence of while loop breaks literal_unroll compilation
• #8314 - Numba not vectorizing 2d copy loop
• #8317 - LoweringError
• #8322 - Obsolete pycc script/code
  • Answer is yes: please remove.

Closed Issues

• #8312 - numba's np.full_like is inconsistent with numpy's
• #8313 - Segfault in nested loop with aliased array

2. New PRs

• #8306 - Fix len of two concatenated Bytes objects bug
• #8308 - CUDA: Support for multiple signatures
• #8310 - CUDA: Fix Issue #8309 - atomics don't work on complex components
• #8315 - Add get_local_mem_per_thread method to Dispatcher
• #8316 - Fix error handling in float unboxing
• #8318 - Cleanup old support for Python 3.6 and earlier code
• #8319 - Bump minimum supported Python version to 3.8
• #8320 - Add name support for GUFuncs
• #8321 - Fix literal_unroll pass erroneously exiting on non-conformant loop.
• #8323 - Remove mk_unique_var from array_analysis.py
• #8324 - Remove use of mk_unique_var in untyped_passes.py
• #8325 - Remove use of mk_unique_var in stencil.py
• #8326 - Remove mk_unique_var from parfor_lowering.py

Closed PRs

3. Next Release: Version 0.57.0/0.40.0, RC Jan 2023