Minutes_2022_03_29

Numba Meeting: 2022-03-29

Attendees: Siu Kwan Lam, Guilherme Leobas, Nick Riasanovsky, Andre Masella, brandon willard, stuart, Travis Oliphant, Val, Benjamin Graham, Jim Pivarski, Ehsan Totoni, Todd A. Anderson, Graham Markall

NOTE: All communication is subject to the Numba Code of Conduct.

Please refer to this calendar for the next meeting date.

0. Feature Discussion

• Vision document update
  • Will be opening a PR for comments and suggestions
  • Follow up discussion next week
• PTX Caching:
  • PoC Gist
  • PR: #7944 - [WIP] Pathfinding for CUDA kernel caching
  • Discussion:
    • worry that new CUDA driver results in unsupported version of PTX/cubin
    • likely fixed by adding cuda driver version into the cache key
• CUDA discussion:
  • #7926 - CUDA: remove context query in launch configuration
  • #7927 - CUDA: discussion: retire the use of a separate launch configuration object, launch kernels directly!
    • What does the API for other libraries (CuPy, PyCUDA, etc.) look like?
      • CuPy: add_kernel((5,), (5,), (x1, x2, y)) # grid, block and arguments
      • PyCUDA: multiply_them(drv.Out(dest), drv.In(a), drv.In(b), block=(400,1,1), grid=(1,1))
• Moving NRT to LLVM IR so it is JIT-able
  • for use-cases that requires allocating small objects frequently (e.g. in a loop)
  • expose LLVM to the details of allocation thus allowing more aggressive optimization on the memory allocation routines
    • potentially turn heap-alloc to stack-alloc
  • Nick: how often do we see this?
    • Stuart: quite frequently asked
  • Stu: NRT can prevent LLVM code motion because of the opaqueness of the NRT calls. Often a problem in container types.
  • Ben: Will this apply to Python C-API?
  • Stu: Don't think it will apply to that code path.

1. New Issues

• #7928 - CUDA: functions can only be used as device functions or as kernels, not both in the same program.
• #7929 - CUDA: Using NUMBA_CUDA_USE_NVIDIA_BINDING with other GPU Libraries
• #7932 - llvmlite version issue while running tests
• #7933 - Checking *args is empty inside overloaded method
• #7935 - Experiment with adding a domain field to integer types
• #7936 - Add official support for thread IDs
• #7937 - enhancement for array_equal

Closed Issues

• #7930 - Numba usage and documentation
• #7931 - llvmlite too low version
• #7934 - Incorrect prange loop index typing, but only in parallel mode

2. New PRs

• #7938 - array_equal new impl
• #7940 - Implemented np.allclose in numba/np/arraymath.py
• #7941 - Remove debug dump output from closure inlining pass.

Closed PRs

• #7939 - Fix rendering in feature request template.

3. Next Release: Version 0.56.0/0.39.0, RC,