Minutes_2020_10_27

Numba Meeting: 2020-10-27

Attendees: Stuart, Graham, Todd, Aaron, Hameer, Siu, Juan

0. Feature Discussion

• Conclude type annotations with way forward.
  • Requirements ahead of accepting PRs:
    • (first thing) typeguard (Siu suggestion)
    • (third thing) sphinx type annotation extension (Siu suggestion)
    • (second thing, following typeguard) Style guide
      • cross reference the discourse thread
      • start with a minimum guide
    • (zero thing) "concept" owners
• Aaron's question(s):
  • AoT compilation (pycc) parallelizing compilation
  • parallel limiting things:
    • llvm context are not threadsafe
    • numba inserting into typing/target contexts
  • granularity
    • coarse grain module level will be easier
    • fine grain per-function per-signature will be more challenging
  • Call graph shape?
  • Compile time information: foo.overloads[foo.signatures[0]].metadata ( CompileResult.metadata)
  • Aaron will open a issue for this
• Finalizing 0.52.0RC2
  • Waiting on llvmlite fix.
• Juan's question(s)
  • how to best get compilation timings?
    • the CompileResult.metadata as mentioned above
    • https://numba.readthedocs.io/en/stable/reference/jit-compilation.html?highlight=metadata#Dispatcher.get_metadata

1. New Issues

• #6417 - Issue creating list of dictionaries with constant keys and heterogenous data
  • Case 1 is extra work to make that sort of thing work
  • Case 2 is a bug in type inference that needs sorting
  • Make new issue for second bug
• #6416 - np.zeroes/ones unable to be resolved
  • user problem with 0-d ndarray scalar
• #6415 - Numpy array indexing lead to an error in numba
  • user questions/unsupported python features
• #6414 - After updating the Numba from 0.43 to 0.51.2, the parallel code is displaying the error
  • waiting for additional info
• #6413 - Make parfor array analysis store information only for integer types.
  • parfors
• #6411 - Compilation time regression in 0.52rc1
  • Fix in llvmlite PR
• #6405 - unicode.join typing fails with unichr array
  • Good first issue to fix
• #6404 - Constant lowering compliant dictionary types
  • Closed
• #6401 - Crash involving typed.Dict, caching, and parametrized type
  • PR https://github.com/numba/numba/pull/6402
• #6399 - 0.52.0rc1 bug in array analysis for tuples of multi-dimensional arrays
  • parfors tuple issue; to be fix in RC2
• #6397 - ListType needs to implement key
  • PR #6403 fix proposed
• #6394 - triangular solver
  • Closed - no more info recently
• #6393 - Support __hash__ for numpy.datetime64
  • feature request
    • np.<datetim64>.view should work
    • cast dt64 to int should work
• #6392 - Strange TBB constraints in 0.52.0rc1 conda package
  • PR #6396 to fix
• #6388 - NamedTuple fails if type name doesn't match variable name
  • Closed
• #6387 - NumPy ndarray expression with broadcast is slower when not use local variable.
• #6386 - Can't AOT String Keyed Dictionaries: undefined symbol: _numba_hashsecret_siphash_k1
  • Try debugging on Windows as it's more strict about linkage
  • May need stub values in C code that's included

Closed Issues

• #6412 - array_dot operator failing in nopython mode
• #6389 - graphviz/dot internal buffer truncation causes issues for inspect_cfg.
• #6366 - "Numba requires at least version 0.31.0 of llvmlite."
• #6356 - Enumerate decrefs too eagerly

2. New PRs

• #6419 - Update CHANGE_LOG for 0.52.0rc2
• #6418 - docs: use https for homepage
• #6410 - Fixes issue #6386
• #6409 - Implement cuda atomic xor
  • Graham to review
  • Based on atomic and
• #6408 - Don't do array analysis for tuples that contain arrays.
• #6406 - Enable caching functions arising from exec of strings.
• #6403 - Add types.ListType.key
• #6402 - Fix typed.Dict and typed.List crashing on parametrized types
• #6400 - Testhound/cuda atomic or
  • Graham to re-review
• #6398 - Fixed Error

Closed PRs

• #6407 - MAINT: Switch to np.compat for long
• #6395 - Fix/052 rc1 octomerge3
• #6391 - Fix/052 rc1 octomerge2
• #6390 - Updates inspect_cfg
• #6385 - Fix/052 rc1 octomerge1
• #6384 - Fix failing tests to match on platform invariant int spelling.
• #6380 - Only test managed allocations on Linux
• #6379 - Use O0 to enable inline and not affect loop-vectorization by later O3…
• #6378 - Fix incorrect CPU override in vectorization test.
• #6375 - CUDA tests: reduce managed memory allocation to 0.1 * GPU RAM
• #6372 - Skips failing tests if SVML is present.
• #6369 - disable refprune pass
• #6365 - Update CHANGE_LOG for 0.52.0
• #6364 - Move deprecation schedules for 0.52
• #6363 - Note that reference counted types are not permitted in set().
• #6362 - Move llvmlite support in README to 0.35
• #6361 - relax the llvmlite requirements for ASV and testing purposes.
• #6359 - Fixes typed list indexing on 32bit
• #6357 - Fix enumerate invalid decref

3. Next Release: Version 0.53.0, RC=Q1 2021

4. Upcoming tasks