Minutes_2021_04_27

Numba Meeting: 2021-04-27

Attendees: Brandon, Caleb, Hernan, Leo, Luk, Todd, Siu, Stu, Val

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

0. Feature Discussion/admin

• Python 3.6 removal.
  • following NEP 29
• NumPy 1.20 and minimum version move to 1.17?
  • still have a few regressions
    • numpy changes behavior
    • NaN cast to int
• No JIT on import.
• Vendoring Cloudpickle 1.60 PR 6977
  • Already have parts of this in the code base, this just gets everything else.
• LLVM 11 update.
  • llvm 11 builds are done as as llvmlite builds. All under llvm11 label on the Numba channel.
  • Numba currently failing, mostly SVML tests, also some issue in aarch64 that need investigation.
• Brandon talks about Aesara from PyMC team
  • Theano used to produce C files to be compiled by a C compiler.
  • plan to use Numba to replace the C backend as a default backend
  • already done a JAX backend
  • Questions:
    • Q: Stuart asks about how not JAX as the default
    • A:
      • JAX needs as many workaround
      • limitation from tracing
      • shape cannot be symbolic in tracing
• Brandon also talks about http://minikanren.org/
  • A DSL for logic programming
  • python binding: https://github.com/pythological/kanren

1. New Issues

• #6976 - parallel mode ignores divide by zero errors, sometimes fills with nonsense
• #6975 - Heap slower than without numba??
• #6974 - Error while import
• #6973 - "cannot import name '_typeconv' from 'numba.core.typeconv'"
• *** #6972 - Wrapper or type to avoid inlinng
  • caching function argument vs inlining
  • user want noinline, function argument be passed as pointer
  • TODO: check first class function is working and its effect with cache
  • Luk comments:
    • workaround using a container of function pointer
    • a partial signature; i.e. mark an argument to use function pointer
• #6969 - Parfor reductions not working when overload uses inline='always' and parallel=True flags
• #6967 - CUDA vectorize returns array from scalar input
• #6965 - CUDA: Atomics tests fail with NumPy 1.20
• #6962 - Parallelism that worked in 0.52 no longer works in 0.53.1
• #6960 - Regression with parfor in numba=0.53.1 when aliased arrays are struct attributes
• #6959 - cannot import numba inside ipykernel: OSError: Could not load shared object file: llvmlite.dll
• #6957 - Caching of functions with keyword arguments
• #6956 - Excessive recompilation due to use of literally and potentially unaware in memory cache.
• #6955 - CUDA: Checklist of features required for Awkward Array extensions
• #6954 - @vectorize with *args failing
• #6952 - Vectorized ufuncs don't respect casting keyword argument
• #6951 - Support for array of objects / jitclasses
• #6950 - Numba master breaks cuDF
• #6949 - x += x.T and x = x + x.T yield different result.
• #6947 - Losing local struct variable pointer in jitted function
  • another struct scalar out live the array because of legacy code
• #6946 - Performance issue: conditional inline array allocation (not called) or different order of conditions -> unexpectedly slow
• #6943 - Error storing record view
• #6942 - Inconsistent self-values assignment within 2-dimensional array

Closed Issues

• #6970 - @intrinsic not working with CUDA
• #6968 - CUDA: Implement printf()
• #6963 - TBB test_fork_from_non_main_thread failing intermittently

2. New PRs

• #6964 - Move minimum supported Python version to 3.7
• #6961 - Update overload glue to deal with typing_key
• #6953 - CUDA: Fix and deprecate inspect_ptx(), fix NVVM option setup for device functions
• #6948 - Refactor registry init.

Closed PRs

• #6971 - Fix CUDA @intrinsic use
• #6966 - Fix issue with TBB test detecting forks from incorrect state.
• #6958 - Inconsistent behavior of reshape between numpy and numba/cuda device array
• #6945 - Fix issue with array analysis tests needing scipy.
• #6944 - CUDA: Support for @overload
• #6941 - ABC the target descriptor and make consistent throughout.

3. Next Release: Version 0.54.0/0.37.0, RC=May 2021