Minutes_2019_06_25

Attendees: Val, Stuart, Siu, Ehsan,

0. Feature Discussion

• Cache write atomicity discussion.
  • allow process-wide config to enable cache=True and use it on our testsuite.
• #3908 - [WIP] Duplicate code with a dynamic aliasing check
  • Ehsan is worried about the current approach; i.e. code-duplication that will increase complexity in compiler-passes and debugging.
  • There are example code that doesn't work. Needs Todd to provide reproducer (#TODO).
  • pull out IR copying code into a separate PR. (#TODO)

1. New issues

• **** llvmlite#495 - Include compiler-rt in llvmlite
  • proposal 1 is probably easiest.
  • check if the compiler-rt separates builtins and the sanitizers.
  • we can put builtins into llvmlite
  • and sanitizers in separate packages
• #4221 - Make lowering reveal which overload is picked
• **** #4220 - support for 'dtype' keyword argument in np.sum()
  • Semantic rewrite #4113 will work
  • reduce-like methods can reuse logic for dtype and axis
  • Suggestions:
    • A sum module for collections of sum algorithms (#TODO open issue)
    • refactor common code to support dtype and axis (#TODO open issue)
• #4218 - Feature Request: parallel mode for map
  • Needs support for map in sequential code first
• #4217 - not helping for the np.argsort()
  • TODO: sanity check this; why numba is 20% slower
• **** #4215 - Broadcasting error with sliced array and ParallelAccelerator
  • The patch has issue with lowering.
  • We'd want to refactor the related code to use the compiler
• **** #4214 - Array analysis support for user defined arrays
  • May need revisit ArrayCompatible.array_priority (#TODO)
• #4213 - Problem with Numba's internals
  • No info, closed.
• **** #4212 - Parfors corrupted typemap for type_annotation
• #4207 - using numba with scipy.integrate.quad
  • Wishlist for Scumba
• **** #4206 - Python crashes for numba 0.44.1 (list.append not threadsafe)
  • How would we start marking methods as thread safe/unsafe?
  • Can we enumerate everything that is safe in a parallel environment?
  • Shortterm: blacklist mutable containers (#TODO)
• #4203 - Upgrading numba 0.43.1 to 0.44.0 breaks my code
  • Object mode with liftable loops sometimes rely on IR mutations from nopython mode path.
  • May let this one ride until we drop object mode fallback entirely
• #4202 - nopython mode doesn't support numpy ops with scalar input
  • Duplicate of older code
  • Caused by numpy functions written in the old builder style
  • Fix is to replace most of the NumPy impls with @overload*
• #4201 - numba.cuda.cudadrv.devicearray.DeviceNDArray is not iterable
  • Easy to fix
• **** #4198 - Auto-registration of custom IR nodes in all extension APIs.
  • Need to discuss best way to approach this
  • Can we collapse this list of extensions?
  • Is single dispatch a solution?
  • Scalability question for having both custom IR-Nodes and custom-Pass.
  • What does LLVM do?
  • Discussion notes:
    • likely require a ABC and Mixins with concrete implementation. (#TODO)
• #4194 - how can we append np array in jit
  • Waiting for user response about whether np.append is needed, or can be done with other methods.
  • feature request

Already Closed

• #4219 - Memory leak with python 3
• #4195 - Untyped global name 'prange': cannot determine Numba type of <class 'numba.ir.UndefinedType'>
• #4193 - Question: vecotrize vmap in Jax

2. Open PRs

• #4216 - Add dominator tree
• #4211 - Developer docs for caching
  • Updated, ready for re-review
  • Should wait for #4155 review?
  • TODO: check thread/process torture tests
• #4210 - Update min/argmin and max/argmax to handle non-leading nans
  • needs review, also waiting for author comment
• #4209 - Revert previous network tol patch and try with conda config
  • Stan will review and mark for merge
• **** #4208 - R&D inlining, jitted and overloaded.
  • Discuss when / how this will be merged
  • method / attribute overloading will come in future PR
• #4205 - Update return value order in normalize_signature docstring
  • Ready to merge
• **** #4204 - [WIP]Enable support for count() method for unicode string
  • git history fun
  • Discuss string strategy
    • copy unicode database from CPython
    • create unicode type function and expose in jitcode as intrinsics
    • important to optimization on ASCII
• #4200 - Add signature parameter to inspect_types
• #4199 - Extend __cuda_array_interface__ with optional mask attribute, bump version to 1
• #4197 - hyperlink repomap

Already merged

• #4196 - Catch missing imports of internal functions.
• #4189 - Changelog for 0.44.1
• #4192 - Extend __cuda_array_interface__ doc to support masked arrays

Old Active

4. Next Release: Version 0.45, RC=July 5, Final=July 12

• Type List
• Caching improvements requirement gathering
  • transitive dependency
  • bug related to storing runtime-only info
  • may punt on jitclasses
  • may punt on issues of (g)ufunc
  • future: pre-filled/package-able cache
    • open questions:
      • numba version pinning(?)
      • regen cache for different numba versions
      • regen at post-install