‘-march=native’ causes compute_extrema to fail on some architectures

[drew-parsons]

PR #187 introduced default cffi compile options -march=native when into compute_extrema (eval.py).

This causes test_eval.py to fail on some architectures, since -march=native is not a universal flag. In particular powerpc (ppc64el) and sparc64,
https://buildd.debian.org/status/fetch.php?pkg=scifem&arch=ppc64el&ver=0.19.0-1&stamp=1779876989&raw=0
https://buildd.debian.org/status/fetch.php?pkg=scifem&arch=ppc64&ver=0.19.0-1&stamp=1779877925&raw=0
https://buildd.debian.org/status/fetch.php?pkg=scifem&arch=sparc64&ver=0.19.0-1&stamp=1779878936&raw=0

The error from ppc64el includes

        except DistutilsExecError as msg:
>           raise CompileError(msg)
E           distutils.compilers.C.errors.CompileError: command '/usr/bin/powerpc64le-linux-gnu-gcc' failed with exit code 1

/usr/lib/python3/dist-packages/setuptools/_distutils/compilers/C/unix.py:223: CompileError

During handling of the above exception, another exception occurred:
...
    def test_extrema_func(cell_type, extrema, dtype, degree: int):
...    
>       u_ex, _X_ex = compute_extrema(u, extrema, tol=tol)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

tests/test_eval.py:147: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
scifem/eval.py:350: in compute_extrema
    X_c, extrema = find_cell_extrema(
----------------------------- Captured stderr call -----------------------------
powerpc64le-linux-gnu-gcc: error: unrecognized command-line option ‘-march=native’; did you mean ‘-mcpu=native’?
____________ test_extrema_func[float64-CellType.tetrahedron-max-5] _____________
=========================== short test summary info ============================
FAILED tests/test_eval.py::test_extrema_func[float64-CellType.tetrahedron-min-5]
FAILED tests/test_eval.py::test_extrema_func[float64-CellType.tetrahedron-max-5]
FAILED tests/test_eval.py::test_extrema_func[float64-CellType.hexahedron-min-5]
FAILED tests/test_eval.py::test_extrema_func[float64-CellType.hexahedron-max-5]
FAILED tests/test_eval.py::test_extrema_func[float32-CellType.tetrahedron-min-5]
FAILED tests/test_eval.py::test_extrema_func[float32-CellType.tetrahedron-max-5]
FAILED tests/test_eval.py::test_extrema_func[float32-CellType.hexahedron-min-5]
FAILED tests/test_eval.py::test_extrema_func[float32-CellType.hexahedron-max-5]
===== 8 failed, 1699 passed, 6 xfailed, 257 warnings in 424.03s (0:07:04) ======

For end-users it can be managed manually by controlling the jit_options given to test_extrema_func. But that leaves the question of how to manage test_eval.py in scifem CI testing.

There could be a number of different ways to manage it

• use a try: catch: block in test_eval.py::test_extrema_func to catch CompileError at the level of the tests, and then skip or rerun with alternative jit_options
• refactor

  scifem/src/scifem/eval.py

  Line 339 in ea5953a

  "cffi_extra_compile_args": ["-march=native", "-O3"],

  to set -march=native (either testing the flag at runtime, or taking the value from a grey-list dict and identifying which value to use via platform.machine())

In regards to setting alternative jit_options for the failing case, the error message suggests that -mcpu=native might work in these architectures.

What's your preferred solution?

[jorgensd]

Let’s make a patch release where we remove -march-native as a default option.

[drew-parsons]

This bug is still live in 20.0.0
https://buildd.debian.org/status/fetch.php?pkg=scifem&arch=ppc64el&ver=0.20.0-1&stamp=1782573591&raw=0
https://buildd.debian.org/status/fetch.php?pkg=scifem&arch=riscv64&ver=0.19.0-1&stamp=1780307421&raw=0
etc

Should we go ahead with the removing of -march=native ?

[jorgensd]

Sorry, I forgot about this. Will fix.

[finsberg]

@drew-parsons, we fixed this in #235, and this fix should be in v0.20.1. Could you check if tests are passing now?

[drew-parsons]

Looking fine in 0.20.2, ppc64el is happy
https://buildd.debian.org/status/package.php?p=scifem