Minutes_2018_02_15

Numba Meeting: 2018-02-15

Attendees: Stan, Siu, Stuart, Ehsan, Todd, Stefan, Oscar

New:

numba#2736: Using numpy functions in @guvectorize-d block
- Add to FAQ
numba#2742: Support ndarray.fill under parfors
- feature request
- Stuart will try this to see how easy it is
- Is there a way to streamline these ndarray methods with no equivalent numpy module function?
numba#2743: Create better error message for testing POST failure
- This confuses contributors when they add an invalid test
numba#2745: Support axes argument in transpose
- feature request
- New contributor is working on this right now
numba#2747: Wrong indexing of numpy arrays while using multiprocessing
- needtriage
numba#2749: Using optional type as parameter to jitted function is troublesome
- enhancement
- not just jitclass limitation: general issue with autodetecting optional types
- not easy to fix
numba#2750: Support np.random.permutation
- feature request
- also may be possibly done by #2745 contributor
numba#2751: Support for Numpy.ufunc.reduce
- feature request
- currently rely on NumPy to implement reduce, so can't call from nopython mode
numba#2752: 2+D indexing with arrays or list not supported
- feature request
- fancy indexing in ndarray not fully supported
- not too hard to implement
numba#2753: Casting to int inside loop yields numpy array indexing error
- Related to numba#2719 (variable scoping)
- requires a more deep refactoring of Numba IR
numba#2754: Support generated_jit for cuda target
- feature request
- not hard

Outstanding:

numba#2660: Support bools from cffi in nopython
- Very close
- What size are bools on windows?
numba#2734: More Constants From cuda.h
- Ready to go
numba#2727: Changes to enable vectorization in ParallelAccelerator
- In review now
- Stuart adding more tests
numba#2737: Fix #2007 (part 1). Empty array handling in np.linalg.
- NumPy behavior is inconsistent release to release
- Current behavior in NumPy is "mostly correct"
- We will aim for NumPy 1.14 behavior for now
numba#2739: Explicitly state default value of error_model in docstring
- Looks good
numba#2740: Fix 2208. Generate better error message
- Ready to merge when dev cycle opens
numba#2741: Enhance error message for undefined variables
- Ready to merge
numba#2744: Add diagnostic error message to test suite discovery failure
- Addresses issue 2743
- Ready to merge
numba#2748: Added Intel SVML optimizations as opt-out choice working by default
- Reviewing now, need to test further
- Will need to refactor and add libsvml detection
- Can we also enable similar functionality with Apple Accelerate?
- what is enabled fastmath
WIP: np.correlate, np.convolve

Gufunc rewrite:
- Detail writeup: https://gist.github.com/sklam/9fe5431672441e6a689e3ec2a121d104
- Next steps
  - Stuart will take a look at ndtypes and how to integrate into Numba
  - Stefan starting on gumath with new function signatures
Pending Python 3.7 breakage:
- New LOAD_METHOD / CALL_METHOD bytecodes
  - https://bugs.python.org/issue29263
- Python 3.7 release schedule:
  - Currently on 3.7.0 beta 1
  - Release candidate 1: 2018-05-21
  - Final: 2018-06-15
- Add support in 0.39?
Community:
- Wikipedia page
- CI badges for travis/appveyor/codecov etc
- Bug report template
- New docs:
  - "Cool stuff" page
  - Explain our DSO infrastructure especially with respect to LLVM
  - "how @jit works" page, high level with pictures etc
- numba annotation tool needs updating, demonstrating and advertising cf. cython's tool (jupyter integration?)
- exception handling, catch all and rewrite with links to explanatory pages
- should we do a llc, IR -> C tool
Lowering extensions:
- How do we get data out of them and back up the stack
- How do we get more information to them, closures ?