Minutes_2019_07_30
Attendees: Ehsan, Pearu, Stuart, Val, Stan, Siu, Todd, James
- 0.45.1 patch release: https://github.com/numba/numba/milestone/29
- almost ready for tagging
- numpy1.17 + cuda fixes
- Python 3.x retirement schedule proposal:
- Discussion on closure inlining
- #4380 - numba fails to be imported in Termux
-
#4378 - CUDA flag to emulate
-fdefault-real-*
/force e.g.float32
throughout- good idea. Not sure how to implement this.
- what should the scope of this be? Change typing of constants and getitem on containers?
- maybe most of the problem is literals?
- omg numpy cast rules 🤯
(float32 * 2 -> float64, float32[:] * 2 -> float32??)
- **** #4377 - Update Py2.7 EOL statement.
- Stan will update based on discussion
- Numba 0.47 will be last Python 2.7 release
- Need to update docs and sunsetting website PR
- **** #4370 - Appending a Python int to a Numba typed list (of type np.intp) raises an error.
- need to reject unsupported types (numpy dtypes not currently accepted)
- should we accept NumPy dtypes as if they were Numba scalar types?
- not yet
- advertise public APIs for simplified creation of Numba types
- should make
numba.typeof
more fancy to construct nested container types?- yes, type by example
- TODO: work on delcarative type first
-
#4369 -
NamedTuple.__init__
hangs with second argument as literal- need to strengthen type checking so that it rejects incorrect usage
-
#4368 - alloc-dealloc-mismatch on import
- likely know cause
- open PR on llvmlite to fix
-
#4367 - Numba could not be imported
- still working with user
-
#4365 - Header in documentation cuts off the top of the page sometimes
- open PR to fix
-
#4364 - typed Dict slower than array with custom dtype
- hash table lookup are slow
- user workaround is to manually cache dict lookup outside loop if that is appropriate to their problem
-
#4362 - [FEA] NVIDIA Jetson SoC (TX2, Xavier, Nano) Failure to Zero-Copy from CPU to GPU with numba.cuda.to_device()
- direct people to pluggable allocator issue
-
#4358 - Meta issue: np1.17 fixes
- review in progress
-
#4356 - getting the IR for library functions
- answered user clarifying what Numba does
-
#4355 - typed-list: expose a view for numpy arrays
- interesting idea, not high priority
-
#4352 - RAPIDS RMM + Numba Integration:
context.deallocations
isNoneType
- PR ready to merge
-
#4351 - Errors with Cuda Memory Management (I suppose)
- Still suspicious this is a watchdog problem
- Maybe out of bounds error?
- #4379 - windows 10, Ryzen1800X cpu, python ver 3.7 numba error
- #4360 - Unsupported context manager in use
- #4359 - Spurious logging messages
- #4376 - [WIP] add np's flip functionality
-
#4375 - Add docstrings to inspect methods
- needs review
- #4374 - np.alen
- #4373 - Set maximum name size to maximum allowable value
- #4372 - Replace all "except BaseException" with "except Exception".
- #4371 - Fix nump1.17 random function non-aliasing
- #4366 - Offset search box to avoid wrapping on some pages with Safari. Fixes #4365.
-
#4363 - Fix np.interp for np1.17 nan handling
- minor review suggestions to fix
-
#4361 - Add allocation hoisting info to LICM section at diagnostic L4
- ready for review (5 seconds or less)
-
#4357 - Fix np1.17 isnan, isinf, isfinite ufuncs
- minor review suggestions to fix
-
#4354 - Try to fix #4352.
- ready to merge
- #4353 - [WIP] Inspection tool to check what numba supports
-
bug fixes
-
Priorities for 0.46 - Target late Sept - 2) Rewrite passes - 3) Fix exception handling - (as time allows) Start work on new Numba IR implementation - Historical IR implementation is an object graph that do not make creating rewrite passes easy, and easy to create an inconsistent state. No "rollback" option. - Goal is create a new data structure that can solve these limitations. - Need to be able to round-trip between new IR container and old-style container so that existing rewrite passes will continue to work. - Need to prepare by refactoring existing rewrite passes to make them more amenable to a new IR representation. - Also need to document "best practices" for creating IR so that contributors will avoid known bad practices. (i.e Inline-closure-pass) - experimental repo for the new container: https://github.com/sklam/etude-okvmap - 1) Continuing on caching: - Catching transitive dependencies - More unusual cases (mixed mode, function parameters, etc) - New CI system!!!!!!!! - (proposal) Declarative typing option for
@overload/@overload_method