Disallow Accelerate as a LAPACK backend for NumPy
Apple has dropped support for Accelerate. It has bugs that have not been fixed, and is closed source so we cannot fix them ourselves. We have been getting a handful of reports from users who end up building NumPy on macOS, and inadvertently link to Accelerate, then end up with wrong linalg results. In PR 14880 https://github.com/numpy/numpy/pull/14880 I propose to disallow finding it when building NumPy. At this time it will remain in distutils as one of the backends to support users, but how do people feel about a future PR to totally remove it? Matti
Extra data point here: SciPy already dropped support for Accelerate as of version 1.2.0. Best, Ian Henriksen On Mon, Nov 11, 2019 at 6:40 PM Matti Picus <matti.picus@gmail.com> wrote:
Apple has dropped support for Accelerate. It has bugs that have not been fixed, and is closed source so we cannot fix them ourselves. We have been getting a handful of reports from users who end up building NumPy on macOS, and inadvertently link to Accelerate, then end up with wrong linalg results. In PR 14880 https://github.com/numpy/numpy/pull/14880 I propose to disallow finding it when building NumPy. At this time it will remain in distutils as one of the backends to support users, but how do people feel about a future PR to totally remove it?
Matti
_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion
On 12 Nov 2019, at 3:27 am, Ian Henriksen <insertinterestingnamehere@gmail.com> wrote:
Extra data point here: SciPy already dropped support for Accelerate as of version 1.2.0.
Best,
Ian Henriksen
On Mon, Nov 11, 2019 at 6:40 PM Matti Picus <matti.picus@gmail.com> wrote: Apple has dropped support for Accelerate. It has bugs that have not been fixed, and is closed source so we cannot fix them ourselves. We have been getting a handful of reports from users who end up building NumPy on macOS, and inadvertently link to Accelerate, then end up with wrong linalg results. In PR 14880 https://github.com/numpy/numpy/pull/14880 I propose to disallow finding it when building NumPy. At this time it will remain in distutils as one of the backends to support users, but how do people feel about a future PR to totally remove it?
+1 from this side - when switching the packaged version of Scipy to OpenBLAS (already taking Numpy along the way) I noticed barely any performance penalties (if - with some benchmarks - it wasn’t actually faster than Accelerate). Cheers, Derek
participants (3)
-
Derek Homeier -
Ian Henriksen -
Matti Picus