Minutes_2023_01_17
Siu Kwan Lam edited this page Jan 17, 2023
·
1 revision
Attendees: Siu Kwan Lam, Andre Masella, brandon willard, Graham Markall, Luk, stuart, Todd A. Anderson, Guilherme, Ianna Osborne, FPOC (last week): Siu FPOC (incoming): Stuart
NOTE: All communication is subject to the Numba Code of Conduct.
Please refer to this calendar for the next meeting date.
- (discussed in triage) Add in a GHA to lock closed issues after ~ a week
- conda/conda-build may have a lock bot for reference
- Graham to follow up with a PR sometime in future
- (discussed in triage) OrcJITv2 status
- https://github.com/gmarkall/llvmlite/tree/orcjit-tidied
- Based on a version of Andre's LLVM 14 branch:
- Why do we use the C API?
- Original intention was to reduce maintenance burden?
- In practice it comes with its own overheads
- Extra notes:
- Going to try and get this into llvmlite soon.
- Should be on par with MCJIT binding.
- (discussed in triage) Remove pinned message on Discourse
- should be done
- (discussed in triage) Gitter switching to Matrix:
- (stuart) Need to decide whether to try making dynamic exceptions fully dynamic now or defer this xref #8134.
- review as is.
- make change for full dynamic + boxer for 0.58
- (Guilherme) Fixed length strings in Numba
- Numba compile a new function for each fixed length string
@njit def f(a): return len(a) In [1]: f(np.array(['foo', 'bar', 'bar'], dtype='U3')) Out[1]: 3 In [2]: f(np.array(['foo', 'bar', 'bar'], dtype='U4')) Out[2]: 3 In [3]: f.signatures Out[3]: [(array([unichr x 3], 1d, C)), (array([unichr x 4], 1d, C))]
- New dtype work in NumPy may allow for this
- Codebase assumes type size to be known
- Question about use case. Large functions that end up being recompiled.
- (Guilherme) Update numba-scipy CI config
- Deprecate Azure CI in favor of GitHub Actions?
- maybe; wait for Val
- Bump Numba version? Current numba-scipy is being tested with Numba 0.47.
- Yes. latest released numba
- Similarly, use a matrix combination of the last X versions of SciPy and Python
- Yes. x=3, perhaps use oldest and newest versions of python supported by Numba.
- Deprecate Azure CI in favor of GitHub Actions?
- (Luk) Question on Roadmap in gitter/Discourse.
- In the coming weeks should set this out for 2023 (near term).
- (Luk) Cache invalidation PR (#8396), 0.57, 0.58 or PR backlog?
- Q. Was this waiting for splitting up lowering and typing? A. No!
- Added to 0.58 milestone.
- Math RFC numba#8700.
- numba#8699 - Cannot create datetime64 array spec for jitclass
- *** numba#8700 - RFC: Separate NumPy and CPython implementions/functionality.
- numba#8703 - Numba and Debian Bookworm
-
numba#8705 -
np.cos
,np.sin
,np.sqrt
returning incorrect values in very specific scenarios- libSVML is probably correct
- But, LLVM SVML passes might have issues
- numba#8707 - Move baseline supported NumPy version to 1.21.
-
numba#8708 - Raise a warning if fastmath options nnan/ninf are in use with
isnan
orisinf
- numba#8709 - numba stencil too slow
-
numba#8712 -
LoweringError
when usingis
with jitclass - numba#8713 - Allow lazy compilation with explicit signature
- numba#8715 - Support for CUDA 12.0
- numba#8710 - numpy array layout is not determined correctly when passed as jitclass property
-
numba#8714 - NvvmError: Failed to compile IR version 1.6 incompatible with current version 2.0
<unnamed>: error: incompatible IR detected
. Possible mix of compiler/IR from different releases. NVVM_ERROR_IR_VERSION_MISMATCH
- numba#8702 - Supply concrete timeline for objmode fallback deprecation/removal.
- numba#8704 - Added traceback printing capability while NRT debugging
-
numba#8706 - Fix doctest for
@vectorize
- numba#8711 - [CI only] testing patch for #8670
- llvmlite#898 - Wrap LLD into python
- merged - numba#8701 - Relaxed ULP testing precision for NumPy Generator tests across all systems
- merged - numba#8716 - CI: Use
set -e
in "Before Install" step and fix install - merged - llvmlite#899 - Run Minconda install with bash