Minutes_2020_08_04

Numba Meeting: 2020-08-04

Attendees: Graham, Guilherme, Hameer, Stuart, Todd, Val, Siu, Pearu, Ehsan

0. Feature Discussion

• 0.51.0rc1

  • llvm 10 status:
    • removing NVPTX specific patches and rebuild
    • LLVM 10.0.1 build and test underway
    • RC1 nonblocking:
      • to look at SVML issues
      • figure out icc_rt compatibility
    • aarch64 stay on LLVM 9 due to sigabrt in reg to reg copy
  • discuss whether we release if known slow down with literal-list/dict.
    • Likely fine/fixed by #6070

• {Optional, Compiler} {feature, extensions}

  • explicit option to turn on
  • examples:
    • https://doc.rust-lang.org/rustdoc/unstable-features.html
    • https://wiki.haskell.org/Language_extensions
  • from __future__ import ... is reserved for Python
    • a per module configuration
    • what happens when different libraries use different feature sets?
  • configuration problem is very hard for the situation in Numba
  • background: features like Record subtyping, first-class function touches core compiler.

• Announce public meeting

  • 5 minutes on what's new in 0.51
  • to open discourse topic to get ideas for discussion topics

• Ehsan roadmap for things wanted to contribute

  • e.g. replace lower_builtin w/ overload is a good idea

1. New Issues

• #6067 - Failure to spot runaway recursion
• #6065 - Size Match AssertionError in Numba parallelization
• #6063 - TypingError while using StructRef in typed list
  • Need to pass in an instance of a type and not a class
• #6061 - Setting cache=True with a recursive function results in Segmentation fault
  • Crashes even when loading cached function
• #6060 - Direct construction of typed list from an existing Python list is undocumented
• #6055 - Cudasim acting differently than Cuda (when allocating)
• #6053 - Numba disregards the first element of the output list, using simple python range
  • patched by https://github.com/numba/numba/pull/6056
• #6051 - Cube Root Intrinsic for CUDA
• #6050 - Support zero-length arrays in numba.cuda
• #6049 - *** Cudasim jit does not support bind=
  • considering deprecate bind
  • PR that deprecate this is already merged
  • closed
• #6048 - Cudasim does not support device_memset (this may be too much to ask?)
• #6047 - Cudasim does not support mapped_array
• #6045 - Passing an unusual slice into a jitted function gives a strange error
• #6044 - int(inf) does not raise OverflowError
• #6041 - *** Passing in dtype to array fails on NumPy master in pip-installed version
  • NumPy ABI/API split at np1.18
  • Ask conda folks about this
  • Should we encourage Numpy to backport all the way back to Numpy 1.11?
  • [name=Hameer Abbasi] Shouldn't be necessary, if we build with 1.18, should be compatible all the way back until the last API change.
    • alternatively, change PyArray_DescrCheck to what is expand to.
• #6040 - CUDA: Add runtime version to numba -s

Closed Issues

• #6059 - ESDA Local Moran Error
• #6054 - numba error when using scanpy
• #6039 - CPU vs GPU performane comparison for distance calc

2. New PRs

• #6066 - Lower math.isfinite in numba.cuda
• #6064 - Lower math.frexp and math.ldexp in numba.cuda
• #6062 - Change references to numba.pydata.org to https
• #6058 - *** Add prefer_literal option to overload API
• #6056 - Fix issue on invalid inlining of non-empty build_list by inline_arraycall
• #6046 - Fixes statement reordering bug in maximize fusion step.
• #6043 - *** [WIP] Add no_unliteral flag to overloads
  • update #6058 post 0.51 to use enum for at least 4 options:
    • prefer literal, prefer nonliteral, no literal, force literal
• #6042 - CUDA: Lower launch overhead by launching kernel directly
• #6038 - Closes #6037, fixing FreeBSD compilation

Closed PRs

• #6057 - fix aarch64/python_3.8 failure on master
• #6052 - Fix dtype for atomic_add_double tests

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