Minutes_2024_02_06

Numba Meeting: 2024-02-06

Attendees: FPOC (last week): Val FPOC (incoming): Kaustubh

NOTE: All communication is subject to the Numba Code of Conduct.

Please refer to this calendar for the next meeting date.

The Numba user survey is coming up:
- https://hackmd.io/qDUGn56yQvykJGLpP3dzNA?view
- Will go online soon
- Will stay open for a month
- Results and analysis will be made available via Discourse
Jim reports back on some data that they collected about dependent projects. Data is in a public S3 bucket. Plan is to discuss the data and the results during the next office hours. That is next week.
Discussion on NumPy 2.0 changes.
- NEP-50 requires a type system with separate Python and NumPy scalar types.
- Numba currently supports coercion and casting rules that allows code to be less rigidly type stable.
- With the arrival of a new type system in Numba, should type stability be more strongly enforced? Example:
```
@jit
def foo(predicate):
    if predicate:
        x = np.int8(123)
    else:
        x = np.int16(456)
# What is the type of `x`?
# Current type system unifies types.int8 and types.int16 to a types.int16
# Is this behaviour type stable?
```
- Is it ok to not be able to cast between e.g. Python and NumPy type systems (even though the representation in LLVM IR is likely the same).
- What if casting was just removed and everything had to be made stable and "cast" was achieved through constructors like int or np.int<>()? (it's currently done at lowering time by indirect lookup).

numba#9410 - Numba 0.59.0 Final Checklist
numba#9412 - debuginfo issues with parallel functions: file names are not identified
numba#9413 - Python 3.13
numba#9414 - Inconsistent support of log2
numba#9415 - module 'numpy' has no attribute 'typeDict' with numpy 1.24.4
numba#9418 - Structured array scalar slices: iadd, isub, etc, not able to compile (simple example code)
numba#9419 - Outdated Python support list on numba.pydata.org
numba#9420 - Tuple of functions (with identical signature) does not compile for CUDA
numba#9421 - Support for Definition of Numpy Arrays of Complex Type Within njit-ed Functions?
- Probably the issue is trying to create arrays from lists, looks like a feature request
numba#9423 - guvectorize does not check contiguity, breaks indexing
- Needs a guard to prevent people from doing this.
llvmlite#1028 - llvmlite 0.42.0 Checklist
llvmlite#1030 - Symbol not found when building on arm64 platform macOSX
- Incompatible C++ runtime
- Homebrew LLVM is a mess
- Maybe try unsetting the flags
llvmlite#1031 - Add flag to IRBuilder.if_else to control whether an "endif" block should be created

numba#9416 - Add math.log2 support
numba#9417 - Add np.log* bindings for CUDA
numba#9422 - Implement ndarray.byteswap
- Contribution looks good, should probably be reviewed