Minutes_2020_07_28

Numba Meeting: 2020-07-28

Attendees: Siu, Graham, Guilherme, Hameer, Todd, Val

0. Feature Discussion

• 0.51.0rc1
  • llvm 10 issues
    • LLVM 10.0.1
    • unable to use 10.0.0 with aarch64 due to a sigabrt
  • literal over-specialize
    • WIP #6035
    • 1.5x slow testsuite

1. New Issues

• #6037 - FreeBSD: numba/np/ufunc/workqueue.c:30:10: fatal error: 'alloca.h' file not found
• #6033 - Feature request jitclass inherting from Callable and class member support
• #6029 - Numba error while import umap
• #6026 - Floating-point bit twiddling in CUDA
• #6023 - Tuple getitem cannot compile with IntegerLiterals as index
• #6022 - Compilation with cuda.jit randomly fails with segfault
• #6019 - Asking about the dict in numba
• #6018 - Cache Fails When Cannot Find Unrelated Custom Definitions
  • problems with dynamic type definition.
  • can python/pickle even do that?
• #6014 - suggestion For Parallel loop and atomic container
  • Potentially a duplicate, find the Bag features request
  • Limited by gufunc requirement on ndarray
• ** #6041 - Passing in dtype to array constructor fails with pip-installed Numba.
  • numpy 1.20
  • ABI change
  • numba-wheel with numpy-built from source
  • Theoretically, only subclassing from PyArray_DType in C should break.
  • Only in Pip version, doesn't repro in conda.

Closed Issues

• #6034 - NUMBA_DISABLE_JIT disables typed.List constructor
• #6032 - Failed to pickle because of PicklingError: Can't pickle class.. it's not found as numba.experimental.jitclass

2. New PRs

• #6038 - Closes #6037, fixing FreeBSD compilation
• #6036 - Update deprecation doc from numba.errors to numba.core.errors
• #6035 - Fix overspecialize of literal
• #6031 - Continue 5821
• **** #6030 - CUDA: Don't optimize IR before sending it to NVVM
  • Before, Numba generated LLVM IR is pre-optimized by LLVM before sending to NVVM
  • There are more & more problems caused by the LLVM pre-optimization; i.e. memset, syntax fixup
  • The PR removes the pre-optimization
  • Works; just tests failure caused by difference in optimization.
  • May slowdown a few cases (e.g. as seen in record use-cases)
• #6028 - Fix #6023
• #6027 - [WIP] Support of inlining of binop Exprs in overloads
• **** #6025 - Initial support for selecting the chunk size for parallel regions.
  • Currently
    • thread scheduling base on number of cores
    • work is front-loaded (early threads has more work)
  • OpenMP like dynamic sheduling using chunksize
  • How to test?
    • maybe exposing the scheduling
  • Would prevent fusion
• #6024 - CI/check5922
• #6020 - CUDA: Fix #5820, adding atomic nanmin / nanmax
• #6017 - Fixing up numba_do_raise
• #6016 - Fixes change of list spelling in a cuda test.
• #6015 - Adding BooleanLiteral type
• #6013 - emphasize cuda kernel functions are asynchronous

Closed PRs

• #6021 - [WIP] Does the c parameter to typeof_impl serve a purpose?

3. Next Release: Version 0.51.0, RC=22 July, Final 29 July?

• Requests for 0.51

• high risk stuff for 0.51.

• 0.51 potential tasks (To be updated)

4. Upcoming tasks