Minutes_2021_03_30

Numba Meeting: 2021-03-30

Attendees:

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

0. Feature Discussion/admin

• 0.53.1 status
  • Need to do a GH release
  • Release notebooks are reviewed, need minor updates then can be promoted.
• Mentorship program
  • 5 participants at present
• Should Numba behave more like C-extension? #6862
• Graham proposed API for adding call site attributes: numba/llvmlite#702
• Related to #6862: noop_ufunc(inp, out=outarray)
  • if outarray is initialized to np.arange(N)
  • noop_ufunc does not write to the output array.
  • would outarray retain the same values after the call?

1. New Issues

• #6872 - try: if array.shape[1] - tuple index out of range. works without numba, doesn't work with @njit(fastmath=True, nogil=True, cache=True)
• #6864 - guvectorize returns different output after upgrading to Numpy 1.20
  • TODO: Ask numpy devs
• **** #6862 - Program crashes with ACCESS_VIOLATION. Windows only.
  • Should Numba behave more like C-extension?
  • Python 3.9 interpreter loading a Python 2.7 DLL into memory. Numba via llvm picked up symbols from Python 2.7 lib instead of 3.9 and as a result, segfault.
  • LLVM symbol look up is differnt to what OS does.
  • More general issue, could well occur with e.g. BLAS.
  • Ignore for now. Document the problem as known limitation. TODO Add task issue for this.
• #6859 - Parallel prange fails on higher number of iterations and is slower
  • TODO: need a quick doc update to mention list.append (any container mutation) is not thread-safe
• #6857 - Implement np.nan_to_num
• #6855 - Rollaxis
• #6853 - Error "The use of yield in a closure is unsupported"
  • Make better error msg... mention "generator"

Closed Issues

• #6868 - CUDA: Eagerly compiling kernels with tuple parameters causes a segfault
• #6865 - 0.53.1 Checklist
• #6863 - Try command
• #6860 - Doc error: old reference to hsa
• #6858 - ImportError: DLL load failed

2. New PRs

• #6873 - docs: Update reference to @jitclass location
• #6871 - Force text-align:left when using Annotate
• #6870 - Alternative to #6770 for more flexibility
• #6869 - Implement builtin sum()
• ** #6856 - Avoid recompiling inlined overload impl
  • need reproducer/benchmark
• #6852 - nanvar with ddof-v1

Closed PRs

• #6867 - Update changelog for 0.53.1
• #6866 - Changes for 0.53.1
• #6861 - updated reference to hsa with roc
• #6854 - PR 6096 continued
• #6851 - set non-reported llvm timing values to 0.0

3. Next Release: Version 0.54.0/0.37.0, RC=May 2021

4. Upcoming tasks

llvmlite call site attributes

PR: https://github.com/numba/llvmlite/pull/702

Existing: attributes on declaration arguments

Example IR:

declare noalias i32 @"my_func"(i32 zeroext %".1", i32 dereferenceable(5) dereferenceable_or_null(10) %".2", double %".3", i3     2* nonnull align 4 %".4")

This is generated with:

func = self.function()
func.args[0].add_attribute("zeroext")
func.args[1].attributes.dereferenceable = 5
func.args[1].attributes.dereferenceable_or_null = 10
func.args[3].attributes.align = 4
func.args[3].add_attribute("nonnull")

Proposed: attributes on call site arguments

Example IR:

call void @"fun"(i32* noalias sret %"retval", i32 42, i32* noalias %"other")

Example API:

fun = ir.Function(builder.function.module, fun_ty, 'fun')
fun.args[0].add_attribute('sret')
retval = builder.alloca(int32, name='retval')
other = builder.alloca(int32, name='other')
builder.call(
    fun,
    (retval, ir.Constant(int32, 42), other),
    arg_attrs={
        0: ir.ArgumentAttributes(('sret', 'noalias')),
        2: ir.ArgumentAttributes('noalias'),
    }
)

Alternatives?

Create an actual arguments class, similar to the formal arguments used in declarations, so that attributes can be added later. Example of declaration from test_function_attribute in llvmlite/tests/test_ir.py, line 141:

func = self.function()
func.args[0].add_attribute("zeroext")
func.args[1].attributes.dereferenceable = 5
func.args[1].attributes.dereferenceable_or_null = 10
func.args[3].attributes.align = 4
func.args[3].add_attribute("nonnull")

Thoughts

• PR #702's proposed API works, but is asymmetric with attributes on arguments in declarations.
• The alternative is more work.
• Accept the PR #702 API, noting that we might want to expand on it with the alternative proposal later? (or not expand on it later if no need...)

Notes

Stuart:

• Additive, not invasive
• Doesn't preclude doing it the other way later
• Better than maintaining a hack

Siu:

a = ir.ArgumentAttributes()
a.align = 4

On removal of tail:

• Siu: tail is a bug not a feature, so OK to remove
• Stuart: Put a notice in the release notes that it's gone.