Minutes_2018_03_15

Numba Meeting: 2018-03-15

Attendees: Siu, Stuart, Todd, Ehsan

1. New/outstanding issues

• #2820: Allow setting error model in `vectorize`
  • can numpy.seterr provide what the user want already?
• #2818: Exception TypeError: "'NoneType' object is not callable" in <bound method ModuleRef.__del__ of <llvmlite.binding.module.ModuleRef object at 0x7f85e655ddd0>> ignored
  • Is this due to the dtor ordering bug in Py2.7/Py3.5
• #2815: `vectorize` doesn't work with `generated_jit` when jitting is disabled
• #2812: CUDA Gufunc error with scalar output
  • Fixed by #2813
• #2811: Exception support in prange body
  • Related to stalled patch #2307
• #2810: nopython array creation in C code
  • TODO: send pointers to where things are
• #2809: CUDA unit test expected to fail no longer fails on Volta
  • Related to new warp sync behavior?
  • Is this a good thing? Siu: Yes
• #2808 CUDA autotune table needs updating for CC 7.0 (Volta)
  • Looks pretty easy to fix
  • We can drop this. There's version in the cuda driver.
• #2806: numpy.array() not handling an array type
  • we should fix this
  • also np.asarray?
• #2804: Wrong results and deadlocks when using jit(parallel=True)
  • need to fix. but difficult
• #2799: ndarray.transpose() and and ndarray.T do not set data ownership correctly
  • new issue araised from a patch
• llvmlite#336 : LLVM viewCFG maybe broken
  • OSX build may be broken.
  • LLVM 6 has the same issue. Likely problem with our buildscripts.

3. Open PRs

• llvmlite#337: Add support for optimization remarks in pass managers.
• llvmlite#330: Add support for LLVM fence instruction
• #2819: Support for np.reshape(array, shape), including tests.
• #2817: [WIP] Emit LLVM optimization remarks
• #2816: Array analysis for transpose with arbitrary arguments
• #2813: Fix #2812: gufunc scalar output bug.
• #2807: Expose `import_cython_function`
• #2801: Passing jitted functions as args
  • Siu needs to follow up
• #2793: Simplify and remove javascript from html_annotate templates.
  • Needs some tests and help on other Python versions
  • Waiting on us
• #2780 PowerPC reference counting fences & minor fixes
  • Related: llvmlite#330 Add support for LLVM fence instruction
  • Wait until LLVM 6.0
  • Waiting on us
• #2779 Implement np.random.permutation
  • Works, but uses Python style RNG, so hard to compare to NumPy
• #2778 Add More Device Array API Functions to CUDA Simulator
  • merged
• #2777 Add support for np.correlate and np.convolve
  • Needs review: Siu
• #2748 Added Intel SVML optimizations as opt-out choice working by default
  • Stuart has built. Needs us to test further.
• #2822: [WIP] Adds stable argsort
  • Needs tests. Adds mergesort.

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

4. Feature Discussion

• Gumath update
  • implemented broadcasting
  • WIP ufunc loops to work with numba generated kernels
  • there are works to put XND arrays in Plasma
• Ehsan question about returning tuple from C code as a struct
  • Answer: numba cannot handle aggregate types as arguments or return type. Only simple scalar types and pointer types are supported. Thus, the tuple must be returned as a pointer to the struct.

5. Next Release: Version 0.38, RC=April 4, 2018, Final=April 11, 2018

• First gufunc improvements
• Better SIMD generation (SVML + parfor fixes)
• LLVM 6.0
• Better debug/troubleshooting tools
• Improve docs and information as per community feedback
• Keep working through backlog of bugs and minor feature requests
• Pipeline manipulation (sklam)