Minutes_2020_05_19

Jump to bottom Edit New page

Valentin Haenel edited this page May 19, 2020 · 2 revisions

Numba Meeting: 2020-05-19

Attendees: Siu, Stu, Todd, Pearu, Aaron, Hameer, Guilherme, Val

0. Feature Discussion

#5609 Jitclass infer spec from type annotations
- Need to form opinions in next few weeks.
- PR changes:
  - extended typeof to support typing generics
  - jitclass can do:
```
@jitclass
class Counter:
    x: int

    def __init__(self):
        self.x = 0

    @property
    def count(self) -> int:
        return self.x

    def inc(self) -> int:
        self.x += 1
        return self.x
```
  - Note:
    - extension to regular function
    - cartesian product problem for describing types in signatures
    - can it be inferred at the call site? (separate problem)
    - would make writing jitclass spec easier
- TODO slightly unrelated. Jitclass ctor infer from callsite
  - start discussion thread with Hameer
- TODO cont' discuss
CoC Draft discussion:
- Current draft: https://github.com/numba/numba-governance/blob/draft/code-of-conduct.md
- Adapted from https://github.com/pandas-dev/pandas-governance/blob/master/code-of-conduct.md
- NOTE:
  - make sure we have a private googlegroup
- TODO:
  - ~~make PR for CoC document~~
  - review https://github.com/numba/numba-governance/pull/1
  - Email core devs to get approval.
Potential way through TBB nested threading problem
- complicated, worth a try

1. New Issues

#5715 - Implement in/operator.contains consistently.
** #5713 - List() and Dict() always recompile _getitem,_setitem, _length, etc. every time; maybe should cache?
- important performance issue
#5712 - Greater support of thread specific arrays
- could be quick fix to remove guard that prevent returning arrays in device function
#5708 - libnvvm segfault compiling this kernel
- 2 args round() not supported in NVVM
- we should stop emitting on 2 args round()
#5706 - Please add support for Metal as GPU accelerator
**** #5703 - Nested tuples with numba.prange cause LoweringError
- discussion:
  - should not continue to patch for nested tuples
  - plans to remove gufunc
#5698 - BUG: AssertionError: Failed in nopython mode pipeline (step: convert to parfors). Invalid reduction format. (Error with parallel=True)
**** #5696 - Too slow jit compilation of a function
- parfor takes >50% of time in ~3 min compilation
- how to speed it up?
- complex CFG graph; time spent a lot in computing CFG; parfor passes also does a lot of CFG computing.
- discussed caching analysis passes
- discussed "enforcing"/"strict" mode to do expensive checks in Numba IR integrity.
#5693 - Issue boxing tuple of dispatcher

Closed Issues

#5710 - CUDA Matrix Multiplication Code failing with ConstantInferenceError
#5691 - Doesn’t work outright on latest Centos from Google cloud compute

2. New PRs

#5718 - CUDA lib path tests: unset CUDA_PATH when CUDA_HOME unset
** #5717 - Initialise Numba extensions ahead of any compilation starting.
#5716 - Docs: Add example usage of List.empty_list
- prefer to make List(dtype=intp)
- List.empty_list will be deprecated soon
#5714 - Fix bug in test
#5711 - Division with timedelta input
- bigger picture problem: unit logic in numpy have changed in time
#5709 - [WIP] Refactoring of cuda.jit and kernel / dispatcher
#5707 - [WIP] refactoring np.argmax with overload
#5705 - A fix for cases where SSA looks like a reduction variable.
#5702 - Fixes for LLVM 9
#5701 - Wip/declarative typing 2
**** #5700 - Make it easier to understand type inference: add SSA dump, use for DEBUG_TYPEINFER
- discussion: getting too many flags?
#5699 - Implement np.split and np.array_split
- TODO: comment on use of typedlist; prefer reflected list for now

Closed PRs

#5719 - [WIP] Fix #5713: cache typed dict and list functions
#5704 - CUDA Eager compilation: Fix max_registers kwarg
#5697 - Fix usedef issue with parfor loopnest variables.
#5695 - Fix usedef issue with parfor loopnest variables.
#5694 - Skip part of slice array analysis if any part is not analyzable.

3. Next Release: Version 0.50.0, RC=May 28th

Requests for 0.50
high risk stuff for 0.50.
- Declarative typing (Siu and Stu)
  - declaring accepted types in things like @overload
  - so we can have better errmsg
  - allow overloading "contextual information"; i.e. targets (CPU, GPU), fastmath flag, exception handling
  - make sure we don't break existing @overload/@lower_builtin use cases
- Remove macro expansion (Graham)
  - replace "macros" with standard typing/lowering; i.e. @register / @lower (@overload not yet working with CUDA)
llvm 9
plan to have first open meeting on June 4th.
0.51 potential tasks
- TypedSet
  - base on TypedDict
  - Related:
  https://github.com/numba/numba/pull/5639
- Nested modes
- Move SSA pass up the pipeline
  - make sure more passes can work in SSA form
- CUDA Dispatcher / kernel interface:
  - Share machinery with CPU target (e.g. for multiple signatures, typeconv, etc.)

4. Upcoming tasks

Opening up the numba meeting
- needs to approve a CoC
- plan to have first open meeting on June 2nd.