Minutes_2020_09_22

Numba Meeting: 2020-09-22

Attendees: Guilherme, Hameer, Stuart, Todd, Siu, Val, Juan

0. Feature Discussion

• Reviving ASV benchmark
  • our benchmark: https://github.com/numba/numba-benchmark
  • historically published to: http://numba.pydata.org/numba-benchmark/
  • for benchmarking LLVM C++ refprune pass and typedlist enhancement
  • to prevent future pref regression in release
    • there were a few recent hotfix needed
  • Hameer tips:
    • suggest run the benchmark long enough so the runtime is larger than the standard deviation.
  • Stuarts Tip:
    • make sure we are not benchmarking the dispatcher
• SLP vectorize
  • SLP: superword level parallelism
  • "local" (in CFG---i.e per-BasicBlock) optimization
  • Add some things that aren't loops to the benchmarks
  • Likely won't involve SVML
  • Turn it on, try it out
• How many optimisation passes should we run?
  • Stuart made experiment to optimize until it stops changing
  • What is the trade-off?
  • What heuristic do we need?
    • Get feedback from the vectorizer
  • Guilherme suggested: https://homepages.dcc.ufmg.br/~fernando/publications/papers/ZanellaSBLP20.pdf
  • Hameer: chrome is doing something similar
• constexpr

class Clazz(object):
    x = 7

clazz = Clazz()

@njit
def foo():
    f = constexpr(lambda : 3 + clazz.x)
    return f * 2

• example usecase #5339
• Siu suggest to take type of local variables (incl. the arguments) so it can do "type-based" operations at compile time.
• Hameer is interested to test it.

1. New Issues

• #6255 - Eager compilation at global scope in CUDA tests doesn't work
  • likely caused by CUDAContext resetting
  • Close issue with explanation
• #6253 - Recompilation of bound methods templates due to omitted args in signature
• #6251 - Cannot cache functions with callable arguments
  • First class functions may be an alternative to making it work
• #6249 - LoweringError assigning array to .real part of NumPy array
• #6247 - use of @jit and @njit makes function execute with false output
• #6246 - AttributeError: module 'numba' has no attribute 'core' #5675

Closed Issues

2. New PRs

• #6252 - Show channel URLs
• #6245 - CUDA Cooperative grid groups
• #6244 - CUDA Docs: Make workflow using simulator more explicit

Closed PRs

• #6254 - Fix win-32 hashing bug (from Stuart Archibald)
• #6250 - Skip some failing-on-32-bit tests (may be windows only)
• #6248 - Add support for CUDA atomic subtract operations

3. Next Release: Version 0.52.0, RC=7th Oct, Final=RC+=~3weeks?

• Requests for 0.52
  • Fast(er) typed.List/typed.Dict? Doesn't have to be in 0.52, next 3-5 months is fine.

4. Upcoming tasks