Minutes_2018_05_31

Numba Meeting: 2018-05-31

Attendees: Stuart, Ehsan, Todd, Siu, Stan

1. New issues

• #3006: Detect whether LLVM has SVML enabled and override SVML setting if not present
  • Stuart working on this now
• #3005: 'str' object has no attribute 'name'
  • Side effect of not supporting two *args arguments
• #3003: Reduce CUDA launch overhead
  • Just FYI
• #3001: Reduce traceback size in user-facing errors
  • Already have a PR to address this
• #2998: Wrong results when calculating power with an array
  • Root cause determined, see #3006
• #2997: Private variables with prange
  • Should be hoisting allocation out
  • But not happening because the np.empty is not marked as a pure function
  • PR open to address this
  • Should the "pure" and "no side effect" lists be shared?
• #2996: Consider adding polly to LLVM builds
  • Inspired by user wanting loop nesting to automatically be reordered for optimal performance
  • Needs investigation
  • Are there good use cases?
• #2995: Support passing in LowLevelCallable as argument to jitted function
  • Already has PR in progress
  • No PyCapsule support
• #2994: Support for PEP 3102 Keyword-only arguments in nopython-mode in nested functions
  • Needs investigating. Maybe not a huge change required to support this
• #2993: Invalid use of make_function needs catching before lowering feature_request
  • Usually caused by unsupported Python constructs
  • Need to emit more friendly error message before lowering
• #2992: NotImplementedError/ LoweringError for any map(lambda ... )
  • map not supported
  • not hard to do in principle with ParallelAccelerate
  • Returning a generator in nopython mode is a challenge
• #2991: numba.vectorize does not preserve function attributes
  • "Easy" to fix some attributes, like name
  • Do we wait for vectorize replacement to fix?
• #2988: Support For CPU Atomics
  • Maybe follow CUDA API at first?
  • Are there more efficient ways to do this on Intel?
• #2987: cuda.mapped_array slices don't have a gpu_data attribute
  • Siu will take a look
• #2982: How best to parallelize numba.stencil operations
  • Could improve docs to make it clear when parallelism is helpful
  • Can we simplify the API for calling parallel stencil from interpreter
• #2981: Numba.stencil lacks a docstring
  • Easy
• #2980: PR breaking code cleanup
  • Are there smaller changes that can be done?
• #2979: Support duck arrays in numba gufuncs
  • Straightforward to do, assuming there are not some gotchas
  • Needs some investigation about immutable vs non-immutable NumPy views
• #2978: Support numpy.cross feature_request
  • Stuart wants to do this
• #2975: Planning for end of NVVM text IR support
  • Know what to do, reminder for now
• #2974: Autogenerate sphinx autodoc docs on Master
  • This might not work very well with Python and our code base
• #2973: Issue in unpacking a list of mixed type tuples
  • Need to explain typing heuristic for passing lists into functions

2. Open PRs

New

• #3004
  Add numpy.empty to the list of pure functions.
  • Will merge
• #3002 Second pass at error rewrites (interpreter errors).
• #3000 TESTING Enh/refactorenv
• #2999 Support LowLevelCallable
• #2990 Improve function call matching for intrinsics
• #2986 Fix environment propagation
• #2984 Support for namedtuples in array analysis
  • Merged this morning
• #2983 [WIP] invert mapping b/w binop operators and the operator module
• #2977 Added docstring coding convention and specified line terminator to developer docs
• #2976 [WIP] Expansion of developer documentation.

Old

• #2968 WIP: Adding xnd kernels
  • Siu giving author feedback on approach
  • This might work for ragged arrays now??
• #2959 np.unique return_counts support
  • Doesn't work as is
  • Provided feedback on how to fix
• #2957 Add new 3.7 opcode support.
  • Being tested using beta 3.7 packages
  • Stuart will add object mode test that produces these opcodes
• #2953 percentile and nanpercentile
  • reviewed, waiting on author fixes
• #2950 Fix dispatcher to only consider contiguous-ness.
  • Stuart to talk to Siu about read-only array case
• #2942 Fix linkage nature (declspec(dllexport)) of some test functions
  • Should maybe generalize this to cover other places
• #2929 Add Flag To Prevent Unneccessary D->H Copies
  • Need to implement decorator to show user how to solve their problem without putting a global default copy strategy into code
• #2910: [WIP] More CUDA intrinsics
  • Check if it works on GPU
  • Make sure it goes thru the buildfarm before merging
  • Blocked on how to simulate in CUDA simulator. See recent njwhite patches?
• #2905: Fix for #2862
  • need to replicate old incorrect behavior in NumPy <= 1.11
• #2894: [WIP] Implement jitclass default constructor arguments.
  • Need to give some feedback to contributor on approach
• #2869: WIP: use pairwise summation in sum
  • Producing wrong results in some cases
  • Open new PR with Stuart's non-recursive implementation
• #2840: Support list of refcounted types
  • Problems found in review that need to be fixed
  • Siu will take a look
• #2817: [WIP] Emit LLVM optimization remarks
  • Want to refactor to not require temp file on disk
• #2793: Simplify and remove javascript from html_annotate templates.
  • Need to push some refactoring onto the PR
  • Interest from scikit-image developers

===========================

3. Feature Discussion

• Patch release plan for Numba/llvmlite to disable SVML when not supported by LLVM linked against llvmlite

4. Next Release: Version 0.39, RC=June 18, 2018, Final=June 25, 2018

• Python 3.7 support
• ARM fixes
• Reference-counted items (arrays lists) in lists
• Python mode blocks in nopython mode
• HTML annotation improvements