Minutes_2018_10_02
Stan Seibert edited this page Oct 2, 2018
·
1 revision
Attendees: Stuart, Siu, Todd, Ehsan, Stan
-
0.40.1 patch release
- rebuild PyPI wheels only to enable TBB
- (Can't replace files on PyPI, so need to tag a new version)
- cherry pick contributor list fix #3338
- how to disable openmp in Linux wheels?
- detect in setup.py if building wheel ... or check environment variable
- Set existing flag to disable OpenMP
- TODO(Siu): PR to detect wheel building; e.g.
'bdist_wheel' in sys.argv
?
- rebuild PyPI wheels only to enable TBB
-
String status
- sidestepping const type issues for now by implementing for str subclass (will fix that before completion)
- Have:
- nopython mode data model + boxing / unboxing
- Logic for dealing with UCS1, 2, 4 unicode representations
==, <, <=, >, >=, len()
- Placeholder impl of
find
andin
, but using naive algorithm - Allocation and reference counting internals for creating strings in nopython mode and unboxing them to Python
- Need:
- Slicing (is there code for
ndarray
to interpret negative indices and strides we can reuse?) +, join()
- Python
FASTSEARCH
algorithm forin, find
- Refactoring to support
bytes
with similar implementations - Figure out what to do with
wchar
kind Unicode strings (??)- Maybe nothing? https://mail.python.org/pipermail/python-dev/2011-December/115044.html
- Only returned by unicode strings that have not been prepared by
PyUnicode_READY()
.
- Replace current
str
type with this type, but preserve the "const" use cases - Write new Numba extension tutorial using
str
as example. (Current interval example is too simple)
- Slicing (is there code for
-
3370 - Recursive functions will not serialize/pickle
- "known," but forgotten, so now we have an issue to remind us
-
3369 - iteration over multidimensional array
- known bug/missing feature
- May not have been possible to implement in the past, but now should be doable (but few people have used it before now)
-
3368 - lstsq sometimes overwrites its inputs
- confirmed bug
- aim to fix in 0.41
- asfortranarray() is not copying, but code assumed it was
-
3366 - Error message for unsupported comprehensions (e.g. generator) needs improving.
- need better message for this case
-
3364 - Error formatting text width is not padded to CPython/readline's 80chars
- low priority, but probably easy to fix
-
3361 - Fix warning from compiling C++ with a C compiler and invalid flags
- might be a distutils bug as well
- not harmful, just annoying
-
3360 - Fix
RuntimeWarning: 'numba.runtests' found in sys.modules
- has fix suggestion in PR
-
3359 - Unpacking of record fails in nopython mode
- unsupported numpy trick, but could be added
-
3358 - List modification is lost when list is passed with a tuple to a jitted function
- The reflection machinery is not recursing into nested types.
- Hard issue
- (Editor's note: Reflection needs to go away in the long term.)
-
3356 - With-objmode usability issue for long type string.
- the excuse of "let type detection figure out the signature" doesn't apply here because we have to be explicit about type for data leaving object mode block
- Hard to use other variable in nopython mode as "type example" because type inference hasn't finished when object mode block is being processed
- Two passes? (is two enough?)
-
3355 - Globals error in objmode context-manager for generated functions
- objmode is likely to be reporting the wrong global to be missing.
- Fix in PR (but may need more work?)
-
3354 - Support list as index type
- adds a dimension to the array
- more fancy indexing...
-
3353 - Implement
.reshape
to work on non-contiguous inputs- copies contents? (right??)
-
3350 - CUDA shared array size limitations
- should not be possible to request more than 48 kB for that hardware no matter the dimensionality
- need to reproduce and see why the error is not reporting in some cases
-
3349 - Document the container types
- good idea
- object mode blocks makes this more important
-
3347 - Objmode with lifting not using custom pipeline
- PR fix already merged
-
3344 - AWS Lambda OSError: [Errno 38] Function not implemented
- /dev/shm not writable on AWS Lambda, recent refactoring of thread backend causes problem with this due to RLock (which tries to use /dev/shm)
- Can we make Numba still work (but perhaps not be fork safe?) on this unusual platform?
-
3342 - Build against LLVM 7 stack
- New patches to work with LLVM 7
- Need to try them
-
3367 - Allow non-global reference to objmode() context-manager
- Siu needs to review
-
3363 - [WIP] string support
- see above discussion
-
3362 - Fix #3360:
RuntimeWarning: 'numba.runtests' found in sys.modules
- Touches a lot of files
- Needs review
-
3348 - Support user pipeline class in with lifting
- Merged this morning
-
3345 - [WIP] Support for np.cov
- still in progress
- need feedback from Stuart
-
3337 - [WIP] Support for np.ediff1d
- in progress
-
3335 - Fix memory management of
__cuda_array_interface__
views.- Needs review
-
3325 - [WIP] Add location information to exceptions.
- Fixed segfaults
- Doesn't reconstruct entire stack of nopython frames, just the immediate location of the exception
-
3320 - Support for np.partition
- author has follow-up questions
-
3307 - Adding NUMBA_ENABLE_PROFILING envvar, enabling jit event
- Siu will review
-
3228 - Reduce redundant module linking
- Stuart will review
-
3162 Support constant dtype string in nopython mode in functions like numpy.empty.
- Need to resolve #3195
-
3160 First attempt at parallel diagnostics
- Stuart will implement Todd's suggestion
-
3134 [WIP] Cfunc x86 abi
- Needs re-review
- 3046 Pairwise sum implementation.
- #2999 Support LowLevelCallable
- #2950 Fix dispatcher to only consider contiguous-ness.
- #2942 Fix linkage nature (declspec(dllexport)) of some test functions
- #2894: [WIP] Implement jitclass default constructor arguments.
- #2817: [WIP] Emit LLVM optimization remarks
===========================
- Type refactoring
- Parallel diagnostics
- LLVM 7
- Initial string support
- Finishing off stalled PRs
- Usual collection of bug fixes and small features