Minutes_2018_08_09

Numba Meeting: 2018-08-09

Attendees: Siu, Stan,

1. New issues

• #3208 - Inconsistent behavior for objects allocated in intrinsics used in two JITed functions
  • Usage error, Siu has responded
• #3207 - Pass c_char or c_char_p to ctypes function in numba
  • c_char is straightforward and would simplify Numba internals
  • c_char_p has potential nul-termination and lifetime questions
• #3206 - @guvectorize copy_to_host() wasting too much time
  • Stan can respond
  • Timing kernel execution, not copy time
• #3205 - Performance case study
  • Likely culprit is Numba logic to handle negative indexing
  • Try different approaches to create a best-practice for identifying non-negative indices to compiler to avoid if-statements.
  • This could be parallelized as well, possibly with some changes to ParallelAccelerator.
• #3204 - AssertionError while looking up variables
  • Duplicate of loop-lifting bug reported periodically
  • Stuart has figured out problem and has most of a fix
• #3203 - numba.cuda.detect() feature request
  • numba -s diagnostics would be helpful
• #3201 - Instantiate default arguments for extension functions early in the pipeline
  • Useful, how to implement?
  • Normalization pass to populate these defaults?
• #3200 - NVIDIA cub support
  • Could be big request
  • Needs thinking and response
• #3196 - datetime64, timedelta64 support on GPU
  • See PR
• #3195 - Problems with support_literals and the Const type
  • This issue is notes from discussion last meeting

2. Open PRs

New

• 3202 - [WIP] TBB + backend refactor... DO NOT MERGE!
  • Going to replace current thread backend due to various hard to fix issues in the implementation
  • two backends: TBB (one known issue), OpenMP
  • Comparing both to each other
  • Will check platform compatibility
  • Nested parallelism?
• 3199 Support inferring stencil index as constant in simple unary expressions
  • Ready to merge
  • Stuart will give quick look and merge
• 3198 Fix GPU datetime timedelta types usage
  • Ready to merge

Old

• 3186 Support Records in CUDA Const Memory
  • needs Siu to check something
• 3181 Support function extension in alias analysis
  • Todd to review
• 3172 Use float64 add Atomics, Where Available
  • needs review by Siu
  • Does this impact distributed usage when cluster is not homogenous
• 3171 Fix #3146. Fix CFUNCTYPE void* return-type handling
  • Stuart will review
• 3166 [WIP] Objmode with-block
  • Still in progress
• 3162 Support constant dtype string in nopython mode in functions like numpy.empty.
  • Blocked until we figure out 3195 or limit the scope.
• 3160 First attempt at parallel diagnostics
  • Looking for feedback on utility of output, impl is a bit hacky
• 3153 Fix canonicalize_array_math typing for calls with kw args
  • Siu will review
• 3145 support for np.fill_diagonal
  • Stuart will review and decide how much strange NumPy behavior to replicate
• 3142 Issue3139
  • Stuart will run through build farm
• 3137 Fix for issue3103
  • Stuart will run through build farm
• 3134 [WIP] Cfunc x86 abi
  • Needs re-review
• 3128 WIP: Fix recipe for jetson tx2/ARM
  • Will merge when ready
• 3127 Support for reductions on arrays.
  • Stuart will run through build farm
• 3124 Fix 3119, raise for 0d arrays in reductions
  • Stuart needs to fix.
• 3093 [WIP] Singledispatch overload support for cuda array interface.
  • Needs review
• 3046 Pairwise sum implementation.
• 3017 Add facility to support with-contexts
• #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

• buildbot talk to slack?
• ask users to disable gitter.im pushes from travisci from private forks?
  • Stan will fix

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