Minutes_2018_08_23

Numba Meeting: 2018-08-23

Attendees: Ehsan, Siu, Stan, Todd, Stuart

1. New issues

• #3255 - "Failed at convert to parfors" when a passed in function is used in the for parallelised loop
  • Check if fixed on dev builds already
  • if so, will be fixed in 0.40
• #3254 - Feature request: making it easier to profile nested jitted functions / methods
  • Fits well with profiling discussion we've been having
  • No good way to register dynamically created symbols with performance tools
  • May also align with tool Siu has been working on to get stack traces from a running thread
• #3253 - numba -s should show total cores and available cores
  • Good idea
  • A little fiddling to figure out cpu share from /proc
• #3252 - ppc64le sporadic fail for test_power_array_op
  • suspicious that we have test with undefined signed integer behavior
  • need to check
• #3251 - Importing ABCs from collections directly is deprecated
  • easy to fix
  • need to do to silence the warnings. not actual error until late 2019
• #3250 - Numba 0.39.0, use of unknown opcode FORMAT_VALUE, python 3.6
  • fixing opcode now doesn't do much because we can't lower this in nopython mode
  • Maybe work well with object mode blocks in future
  • add as opcode, but catch as error before lowering
• #3249 - Sync numpy random generator with numba random state
  • interesting idea
  • should we add a function for synchronizing state back and forth that can be used manually
• #3248 - import imp causing DeprecationWarning
  • already open PR to fix
• #3247 - interface for CUDA GPU memory manager
  • design in progress
  • thinking about how to enforce all usage of Numba have same allocator
• #3245 - parfors array analysis failure
  • fixed in PR
• #3243 - Tweak Sphinx
  • haven't figured out what magic CSS is needed
• #3240 - BUILD_LIST_UNPACK opcode is not supported.
  • only an issue in Python 3.5
  • Python 3.6 uses different opcode for this function syntax
• #3239 - numpy.array(...) LoweringError
  • Two problems
  • Need to check on 0.40 dev and clarify what current issue is
• #3238 - lowering error
  • user has gone back to construct a different minimal reproducer
• #3236 - Compatibility possible for numba jit compilation in Python runtime on GCP cloud functions?
  • no idea what the issue is here
  • hoping for feedback from someone who understands the GCP environment
• #3232 - Make ROCm device wavesize and wavebits into compile time constants
  • can lookup as attribute of agent?
• #3231 - consider caching for array-expr lowering
  • When profiling Numba, noticed that lowering of array expressions there are repeated expression that could be cache

2. Open PRs

New

• 3242 - Page with Numba talks and tutorial links
  • ready for review unless there are other talks people are aware of
• 3241 - Fix #3225. Normalize 0nd array to scalar in typing of indexing code.
  • has CI failure that needs to be addressed
  • possibly issue with test function itself
• 3237 - Fix DeprecationWarning about imp module
  • Stan will review
• 3234 - Fix types and layout for np.where.
  • Siu will review

Old

• 3230 - Fixes liveness analysis issue in looplifting
  • Stuart and Siu have both changed this PR
  • Needs a quick re-review
  • blocker for object mode block
• 3228 - Reduce redundant module linking
  • Stuart will review
• 3222 - Fix #3214. Mishandling of POP_BLOCK in while True loop.
  • Stuart will review
• 3212 - Support for np.vander
  • Stuart will review
• 3211 - Handle unpacking in building tuple (BUILD_TUPLE_UNPACK opcode)
  • Reviewed
  • Need to verify Mac works
  • Siu will merge
• 3209 - Support for np.tri, np.tril and np.triu
  • blocked on #3234
  • author on vacation
• 3202 - [WIP] TBB + backend refactor... DO NOT MERGE!
  • Currently chasing different bugs on different platforms
• 3186 Support Records in CUDA Const Memory
  • Stan needs to get back to user with PTX on failing systems
• 3172 Use float64 add Atomics, Where Available
  • needs re-review by Siu
• 3166 [WIP] Objmode with-block
  • Still in progress
• 3162 Support constant dtype string in nopython mode in functions like numpy.empty.
  • Need to resolve #3195
• 3160 First attempt at parallel diagnostics
  • Stuart will implement Todd's suggestion
• 3142 Issue3139
  • Blocked on 3127
• 3134 [WIP] Cfunc x86 abi
  • Needs re-review
• 3127 Support for reductions on arrays.
  • Ready to merge
• 3124 Fix 3119, raise for 0d arrays in reductions
  • Stuart needs to implement feedback
• 3093 [WIP] Singledispatch overload support for cuda array interface.
  • Needs review
• 3046 Pairwise sum implementation.
• #2999 Support LowLevelCallable
• #2983 [WIP] invert mapping b/w binop operators and the operator module
• #2950 Fix dispatcher to only consider contiguous-ness.
• #2942 Fix linkage nature (declspec(dllexport)) of some test functions
• #2894: [WIP] Implement jitclass default constructor arguments.
• #2817: [WIP] Emit LLVM optimization remarks

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

3. Feature Discussion

4. Next Release: Version 0.40, RC=Sept 3, 2018, Final=Sept 10, 2018

• Experimental python mode blocks
• Refactored threadpool interface
• AMD GPU backend
• Parallel diagnostics
• Usual collection of bug fixes