Kaushik Kulkarni authored 4 months ago and Andreas Klöckner committed 4 months ago

Co-authored-by: Connor Ward <c.ward20@imperial.ac.uk>

Andreas Klöckner authored 1 year ago and Andreas Klöckner committed 7 months ago

On Python 3.12, this provokes a stack overflow in the scheduler. It is
not quite clear why that's the case; pure-Python recursion even with
generators seems to respond well to setrecursionlimit():

```py
def f(n):
    if n:
        yield from f(n-1)
    else:
        yield 5

import sys
sys.setrecursionlimit(3500)
print(list(f(3400)))
```

That said, there have been [behavior](https://github.com/python/cpython/pull/96510)
[changes](https://github.com/python/cpython/issues/112215)
in Py3.12 in this regard, but it is not clear what exactly
about Loopy's behavior makes it fall into the 'bad' case.

Andreas Klöckner authored 9 months ago and Andreas Klöckner committed 9 months ago

https://github.com/pocl/pocl/issues/1509

* Add UnrollPragmaTag for #pragma unroll

* add pragmas parameter to emit_sequential_loop

* pragma -> hint

* fix flake8

* Use Collection instead of Block

---------

Co-authored-by: Andreas Kloeckner <inform@tiker.net>

* Add within_inames when printing instructions

* add a test

* Also record empty within_inames

* Fix parsing of empty inames= option

Co-authored-by: Andreas Klöckner <inform@tiker.net>

Kaushik Kulkarni authored 2 years ago and Andreas Klöckner committed 2 years ago

Co-authored-by: Andreas Kloeckner <inform@tiker.net>

Andreas Klöckner authored 2 years ago and Andreas Klöckner committed 2 years ago

Co-authored-by: Alexandru Fikl <alexfikl@gmail.com>

Andreas Klöckner authored 2 years ago and Andreas Klöckner committed 2 years ago

- Refactor argument passing so that instead of implicitly-spawned
  `ImplementedDataInfo` objects, there are actual arguments (for
  automatic offsets and strides, base storage, and `sep`-tagged arrays).
  It also centralizes the logic for what goes into argument lists,
  instead of having various "filtered" versions scattered about.
- Get started on type-annotating a bit of loopy.
- Switch a not-small number of data structures to be dataclasses, notably `LoopKernel`.
- Drop OCCA support from the ISPC target. (I'm not aware of any users, ever.)
- Drop the Numba target outright. (I'm not aware of any users, ever.)
- Drop `LoopKernel.local_sizes`, which was usable to directly set the
  workgroup size. (I'm not aware of any users, ever.)
- Expire the deprecation for `iname_to_tags`.
- Bumps the Python compatibility target to 3.8, for `from __future__
  import annotations` and  `cached_property` (mypy does not support
  nested decorators)
- Bug fix: `tags` was not part of `LoopKernel.hash_fields`
- Bug fix: `InstructionBase.get_write_dependency_names()` was used to
  find written variables, `InstructionBase.assignee_var_names()` is
  correct
- Bug fix: KernelExecutorBase now uses linearize() so as to not bypass
  pre-linearization checks (cf. gh-639)