Numba Meeting: 2019-12-10

Attendees: Graham, Pearu, Ehsan, Todd, Stuart, Siu, Stan, Val, Aaron, James

0. Feature Discussion

• from future import disable reflected list PR
  • ready for review, PRS: #4883 #4937 and #4941
• Functions as first class types progress
  • Added FunctionType
  • Can't call them yet.
  • working on lowering of calls
• #4903: Multiple reductions
  • Todd will be opening a PR to fix this
• Should we have an anti-aliasing check in testing to verify that objects are not being reused in the IR.
  • Aliasing being introduced in parfor creation. (new var name = expression for numpy exp, and expression was being reused)
  • Make IR nodes immutable?
  • Stick aliasing check in the legalizer?
  • This will uncover a lot of potential issues
• what are the requirements for refcounting in boxing functions?
  • not sure, also should add to dev docs
• review what PRs are in and out of release

1. New issues

• #4939 - comparing ndarray with boolean fails in lowering error
  • Stuart has PR (https://github.com/numba/numba/pull/3901) with fix for all of these, except those that require SSA
• #4933 - Discussion: should __cuda_array_interface__ include a stream pointer?
  • out of scope, IMO
• #4930 - Cannot create jitclass with python list
  • type spelling not obvious
• #4928 - min/max on typed lists extremely slower than on python lists
  • user education on the cost of boxing/unboxing from interpreter in typed list. Seems to be resolved
• #4927 - Importing numba is slow
  • easy task
• #4926 - heapq doesn't work with numba typed lists
  • due to hardcoded C impl in CPython's heapq
  • need to fix Numba's heapq w/ typed list
• #4925 - Make the "not" supported part of the python language in the docs more clear.
  • Siu will fix formatting to be more shouty when he updates docs for try/except support.
• #4923 - Cannot use 1d array containing a single value as a predicate.
  • TODO:
    • allow typing to pass
    • but raise in runtime
• #4922 - prange reduction incorrect when operation happens inside a conditional block
  • PR already opened
• #4919 - object of type Set has no len() when using it in condition
  • See previous predicate discussion

Already Closed

• #4936 - Numba 0.47 Python 3.8 on pypi and anaconda cloud
• #4924 - @jit
• #4917 - reshape() supports contiguous array only
• #4915 - The value of the variable does not change in the while True: loops

2. New Open PRs

• #4938 - [WIP] Improve errors / warnings for cuda vectorize decorator
• #4937 - Support None as a type for the typed-list
• **** #4935 - Better handling of prange with multiple reductions on the same variable.
• #4934 - Fix fails in TestCudaWarpOperations
  • Need to add V100 to our test matrix
• #4932 - Fixup AUTHORS file
  • Going to remove AUTHORS file as redundant to Github history
• #4931 - Add physical limits for CC 7.0 / 7.5 to autotune
• #4929 - [Experiment] Using env_manager.read_const(env_manager.add_const(...)) in boxers
• #4921 - fix creating dict with JIT disabled
• #4920 - fix creating list with JIT disabled

Already merged/closed

• #4916 - fix Azure test failure detection on windows
• #4918 - Release0.46
  • Closed as seems like a random PR

4. Next Release: Version 0.47.0, RC=December 19

• CPython 3.8

• Requests for 0.47 (last release for the year) - jitclass performance issues - llvm 9 trial - CTK libcudadevrt.a - CI needs to take 50% of current time - Val & Stu already looking at this - also checking Azure CI config to avoid wasting compute time - Caching: - transitive dependency - other issues: i.e. function as argument, with objmode - distributing cache - Immutable list and deprecating reflected list - Switch to pytest (see above) - Using Numba to generate LLVM/NVVM IR for different targets https://github.com/numba/numba/issues/4546 - @overload for gpu