Minutes_2018_02_08

Numba Meeting: 2018-02-08

Attendees: Siu, Stuart, Todd A, Oscar V, Stefan K, Ehsan

numba#2735:CUDA: Stream Should Be Used for Debug Transfers
- Can we fix this between RC and final 0.37?
- Siu: not sure if it is really a bug
numba#2729: Default integer size in numpy and numba on Windows (part 2)
- Siu: it's a limitation of our type-inference
numba#2723: Plan for Python 2.7 Deprecation
- Dec 2018: Last Python 2.7 release, make python27 branch
- Jan 2019: Drop all Python 2.7 support from master, do another release
- Jan-Dec 2019: only backport critical bug fixes to python27 branch
- Jan 2020: Python 2.7 support ends
- So far only feedback has been 👍
- If no other changes, should we put statement into Numba 0.37 or 0.38 docs?
numba#2719: enhance variable scoping to allow refining types in branches
numba#2718: setup_requires + distutils_extension
numba#2717: reproducible error with numpy import when profiling
numba#2713: Adjusting threadpool size at runtime
numba#2705: prange fails when returning tuple of numpy arrays
- related to numba#2537.
- can't reproduce on master.
- should we close?

numba#2660: Support bools from cffi in nopython
- More tests added by PR submitter
numba#2734: More Constants From cuda.h
numba#2727: Changes to enable vectorization in ParallelAccelerator
- Is this ready for review after the 0.37 release?
- Other changes needed?
numba#2737: Fix #2007 (part 1). Empty array handling in np.linalg.

Gufunc rewrite proposal:
- Detail writeup: https://gist.github.com/sklam/9fe5431672441e6a689e3ec2a121d104
- What is next step for 0.38 development cycle?
  - StefanK is working on the gufunc loop in gumath.
(from last week) OpenCL + parallel accelerator
- OpenCL issues:
  - Historically inconsistent support by major hardware vendors (compiler bugs, runtimes limited to specific operating systems, long delays in updates)
  - Numba code generation needs generic pointer address spaces, which only arrived with OpenCL 2.x
  - OpenCL 2.x switched to SPIR-V as the IR, which is not based on LLVM IR
  - In order to reuse Numba's lowering code, we want to target LLVM IR.
  - The current state of LLVM IR -> SPIR-V translation tools is not good. Several attempts seem experimental or unsupported.