Minutes_2018_01_11

Numba Meeting: 2018-01-11

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

llvmlite#320 : libz/libc++ linkage issue
- Deep rabbit hole: Apple-shipped Python appears to not be respecting @rpath or DYLD_LIBRARY_PATH. The llvmlite (and numba) wheels have linkage to @rpath/libz.1.dylib and @rpath/libc++.1.dylib, which are present in /usr/lib. Need to use install_name_tool to remap the library dependencies to absolute paths when making wheels. No changes needed for conda packages.
- There is an alternative, ship the libs, which also guarantees compatibility and symbol res?
#2680: numba.prange breaks autovectorization
- Same basic issue as numba#2679
- Workaround is iscontinguous() PR, real fix is redoing gufuncs
numba#2677: user request for containers of functions
- Put on feature backlog
numba#2431: Weird multiprocessing block importing Numba function.
- Looks very suspicious: segfaults like we see in incorrect multithreaded use of LLVM, but after a fork?
- More strange since happens if LLVM (via llvmlite) is loaded after a fork.
- Workaround seems to be changing multiprocessing to use spawn or forkserver
- Need some outside input from people who have dealt with strange multiprocessing issues

numba#2678: Add Gitter badge
- Easy merge (during meeting)
numba#2684: Fix #2340, generator error yielding bool
- @sklam to review
numba#2681: Allow Assigning to DeviceNDArrays
- User contribution based on some previous PRs fixing CUDA device arrays. @sklam to review
numba#2674: ParallelAccelerator Refactor
- Ready for review. Section of disabled code needs additional work in future PR
- @stuartarchibald to review next week
numba#2672: Add ascontinguousarray
- Still needs more work
numba#2668: Fix empty_like from readonly array
- Needs tests, not sure if submitter will add them, so @stuartarchibald will add commits to PR
numba#2660: Support bools from cffi in nopython
- Needs additional testing. @stuartarchibald will give submitter feedback for additional test scenarios.
numba#2627: Remove hacks to make llvmlite threadsafe
- Ready to go. Merged during meeting

CUDA 9.0/9.1 status:
- Need to fix new CUDA 9.x cudatoolkit recipe to work again on CUDA 8.
- Builders ready to go once this is sorted
Dask GDF: Apply transformations to groups
- Highlighted need to improve gufunc infrastructure to make group transformations on the GPU easier for users to write
Gufunc rewrite (aiming to start landing improvements in 0.38)
- Plan draft/notes: https://gist.github.com/sklam/6245c0b98159cf306bd527174827ba4a
- General agreement that externalizing the broadcast implementation into a C library (as Travis has been talking about) is the best approach to avoid slowing down code generation
- Inlining broadcast loops could be useful in some cases, but need some empirical evidence
- gufunc extensions may be post Numba 1.0