Minutes_2020_05_12

Numba Meeting: 2020-05-12

Attendees: Aaron, Guilherme, Graham, Todd, Hameer, Pearu, Val, Stuart

0. Feature Discussion

• numba 0.49.1 and llvmlite 0.32.1

  • Released
  • problems:
    • import umap
      • likely messed up installation

• May 14 is start of Burndown

  • major changes
    • prelude to decl-typing: error handling improvements
    • marco expansion removal; mostly cuda target
    • llvm9
      • first build (QEMU based) of power is segfaulting all the time

• conda-forge build surprise

  • SVML patches are different but we are not sure how it is working

• (Cont') discussion on function subtyping

  • part discussion on merits of producer/consumer declaring
  • also on whether something more dynamic could be made work
  • also on motivation
  • TODO: outcome, discuss again next week, more people to investigate:
    • gist: https://gist.github.com/luk-f-a/9063ee1eea50166141a21ee852fb4812
    • PR: https://github.com/numba/numba/pull/5579
  • new gist: https://gist.github.com/luk-f-a/77fb61b25cadbc497337ab9b53b1d0c9
  • Stuart notes:
    • need to restrict the type-conversion/semantic behavior w.r.t new issues with first-class function i.e. tuple of function types
    • need defining the semantic first-class function type
  • TODO: in reviewing the PR, need to explore new usecases; what it should do? what it shouldn't do?

• Opening up the numba meeting

  • audience size limitation?
  • time for core developers to discuss
  • Hameer mentions numpy has alternating community-focused meeting vs triage/internal-task meeting.
  • maybe start with a once a month community-focused meeting
  • needs to approve a CoC
  • plan to have first open meeting on June 4th.

• nesting modes

  • nested parallel=True is very common
    • i.e. datashader
  • now with true nested parallel region, performance degradation will happen.

1. New Issues

• #5692 - issue with >1d array iteration
  • produces wrong answer!!!!
• #5691 - Doesn’t work outright on latest Centos from Google cloud compute
• #5690 - Support __len__ in jitclass
• #5688 - Segfault when executing a cached lexsort implementation
• #5687 - Segfault in jitted class when deleting and appending list elements
• **** #5685 - Problem with literal unroll of nested function tuples
  • PR #5689
  • already discussed in function subtyping (see above)
• #5683 - Numba 0.49.1 checklist
• #5681 - Avoid _raise_errors class attribute in BaseTypeInference
• #5680 - 'key already in dictionary' while analyzing bytecode
  • byteflow & SSA related
• #5679 - printf that can be called from Numba jit-decorated function
• #5678 - bug in analysis NumbaParallelSafetyWarning
• #5677 - umap import failure
• #5675 - AttributeError: module 'numba' has no attribute 'core'
• #5670 - AOT compilation fails with "unresolved external symbol __svml_log2_ha"
  • check at burndown
• #5667 - Numba leaks compilation objects when parallel=True
• #5666 - Keep numba.types shim

Closed Issues

• #5684 - Iterating empty array shape tuples fails
• #5672 - Need a way to construct a string from bytes or decode bytes to string in nopython mode
• #5671 - Instantiating a jitclass with unicode attributes fails
• #5668 - All example code fails in the 5 minute guide to Numba
• #5663 - TypeError vectorized assignment in ´prange` loop with if statements

2. New PRs

• #5689 - Fix bug in unified_function_type (issue 5685)
• **** #5686 - Fix issue with SSA not minimal
• #5674 - Fix #3955. Allow with objmode to be cached
• #5673 - Fix inliners to run all passes on IR and clean up correctly.

Closed PRs

• #5682 - Update CHANGE_LOG for 0.49.1
• #5676 - Restore numba.types as public API
• #5669 - Update example in 5m guide WRT SSA type stability.

3. Next Release: Version 0.50.0, RC=May 28th

• Requests for 0.50

• high risk stuff for 0.50.

  • Declarative typing (Siu and Stu)
    • declaring accepted types in things like @overload
    • so we can have better errmsg
    • allow overloading "contextual information"; i.e. targets (CPU, GPU), fastmath flag, exception handling
    • make sure we don't break existing @overload/@lower_builtin use cases
  • Remove macro expansion (Graham)
    • replace "macros" with standard typing/lowering; i.e. @register / @lower (@overload not yet working with CUDA)

• llvm 9

• plan to have first open meeting on June 4th.

• 0.51 potential tasks

  • TypedSet
    • base on TypedDict

    • Related:

    https://github.com/numba/numba/pull/5639
  • Nested modes
  • Move SSA pass up the pipeline
    • make sure more passes can work in SSA form