Minutes_2020_03_31

Numba Meeting: 2020-03-31

Attendees: Aaron, Ehsan, Graham, Guilherme, Pearu, Stuart, Todd

0.49.0RC1 issues and notes:
- a few issues but nothing major
- Awkward successfully ported to this release
- Some project depends on specific import behavior
People to declare 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)
Keep roadmap more up-to-date
czar role update:
- Adding group
- Adding docs
Questions
- compilation time
  - up to several minutes
  - Ehsan ask if we have thought about writing the IR & passes in C++
  - Todd & Siu suggest MLIR as a potential IR alternative
  - Todd suggests looking at compilation times for each Numba passes and subpasses.
  - Ehsan suggests per-compilation unit optimization level
  - Stuart suggests adding clock to llvmlite to track time in LLVM
  - Stuart suggests improvement to analysis pass to better track mutation

#5449 - bump max llvmlite version to accept 0.32.0
- Patch: https://github.com/numba/numba/pull/5455
#5448 - 0.49rc1 fail 1 test on windows
- easy to fix
- not a blocker
#5444 - Numba doesnt recognize another classes
#5442 - prange working with numpy?
- Patch: https://github.com/numba/numba/pull/5447
#5439 - Array analysis failure cannot unpack non-iterable NoneType object
- parfor bug
#5435 - BoundFunction(array.reshape for array(int64, 1d, C))
- lack info
#5434 - All templates without literals error
- 1element array scalar-degrade
#5433 - Incompatible shape for array
- lack info
#5429 - Exception using .get_ipc_handle(...) on array from `as_cuda_array(...)
- CUDA
- Patch: https://github.com/numba/numba/pull/5437
#5427 - Memory leak in JIT function with numpy arrays
- memory leak
- confirmed bug
- bad generator
#5424 - parfors copy propagation hanging
- parfor bug
#5421 - CUDA printing is unreliable
- CUDA. doc.
** #5419 - Reference leak for arguments to njit functions if partial type inference is used
- memory leak in compiler
#5416 - Type code attribute for TypeRef
- interesting discussion
#5412 - literal_unroll problem unpacking induction variable
- bug in passes
** #5411 - Avoid calling typers/overloads without literal types
- discussion
- error messages being hidden
- suggest use of RequireLiteralValue
#5410 - Numba 0.49.0rc1 Checklist
- TODO: tag dev
#5409 - numba.jit does not work with static class methods?

#5465 - Remove macro expansion and replace uses with FE typing + BE lowering
#5457 - implement repr for untyped lists
#5455 - Move to llvmlite 0.32.*
- 0.49.0rc1 fix
#5454 - Don't warn for numba.experimental.jitclass
- 0.49.0rc1 fix
#5453 - fix building docs on windows
#5452 - Add version information
#5451 - [WIP] instance cache for UndefinedFunction
#5450 - Docs: remove numbers and add titles to TOC
- TODO: try it
- 0.49.0rc1 fix
#5447 - Handle idx and size unification in wrap_index manually.
#5437 - Fix #5429: Exception using .get_ipc_handle(...) on array from `as_cuda_array(...)
#5436 - Fix #5408: test_set_registers_57 fails on Maxwell
- cuda
#5426 - Fix a SystemError if unboxing fails in a ufunc
#5425 - Improve missing type annotation errmsg
#5423 - Fix #5421: Add notes on printing in CUDA kernels
** #5422 - Replace NativeValue.is_error with NativeValue.error_blk
- discussion: What if @unbox is making a new function? think caching unboxer.
** #5418 - Add np.asfarray impl
- discussion: float_
#5417 - Ensure that tuple unboxing cleans up early items if later ones fail