Hi,

On Tue, Jun 10, 2014 at 1:52 AM, Sturla Molden sturla.molden@gmail.com wrote:

Frédéric Bastien nouiz@nouiz.org wrote: OSes

...

Good luck, and if they act on this, I would be happy to know how you did.

If we report a segfault Apple will probably think we are to blame. 99.9 % of bug reports come from idiots, and those who screen bug reports are basically payed to hit the ignore button. Indeed, good luck on reporting it.

I think in the short run we will be better off re-implementing cblas_sgemv with cblas_sgemm on Apple OSes. Luckily the error in cblas_sgemv does not silently produce garbage data but segfaults the process, so anyone affected will see it.

Would you consider doing a PR for that?

I believe I can build a 32 / 64 bit numpy binary against ATLAS, but it's certainly much less straightforward.

Cheers,

Matthew

On 10/06/14 14:57, Matthew Brett wrote:

Would you consider doing a PR for that?

Here is a patch you can try before I post a PR.

It can also be build independently of NumPy, so you don't need to rebuild NumPy just for testing it. (The change to numpy is in a different folder.)

I decided against using cblas_sgemm. Instead it just enforces alignment to 32 byte boundaries. Because cblas_sgemm would require a copy if the vector is strided, it didn't matter.

I have tested with Accelerate, OpenBLAS and MKL, clang and icc. From what I can tell it works correctly and does not segfault on misalignment.

Sturla

On Mon, Jun 9, 2014 at 8:52 PM, Sturla Molden sturla.molden@gmail.com wrote:

If we report a segfault Apple will probably think we are to blame. 99.9 % of bug reports come from idiots, and those who screen bug reports are basically payed to hit the ignore button. Indeed, good luck on reporting it.

I've reported bugs to Apple and seen them fixed in major releases of the operating system (though not necessarily minor releases). Unless things have changed considerably since my last bug report, I don't think it's fair to characterize their quality assurance team this way, particularly without evidence to back up your point.

A

I also have filed a bug report with apple in the past to fix an accelerate bug. The key I think was to clearly demonstrate and isolate the problem like any good bug report should have. In my case, they got back to me within a week and the bug was fixed a week after that. The release schedule is a bit difficult however but we should file one in any case, especially with Yosemite being under active testing and 10.9.4 in the testing stages.

Kyle

On Tue, Jun 10, 2014 at 12:02 PM, Julian Taylor < jtaylor.debian@googlemail.com> wrote:

On 10.06.2014 15:46, Aron Ahmadia wrote:

...

On Mon, Jun 9, 2014 at 8:52 PM, Sturla Molden <sturla.molden@gmail.com mailto:sturla.molden@gmail.com> wrote:

If we report a segfault Apple will probably think we are to blame.
99.9 %
of bug reports come from idiots, and those who screen bug reports are
basically payed to hit the ignore button. Indeed, good luck on

reporting

...

it.

I've reported bugs to Apple and seen them fixed in major releases of the operating system (though not necessarily minor releases). Unless things have changed considerably since my last bug report, I don't think it's fair to characterize their quality assurance team this way, particularly without evidence to back up your point.

I'm also optimistic they'll fix it. E.g. they also fixed the -mno-fused-madd screwup they did with their python installation in 10.9 in the minor release 10.9.3. So with a little luck 10.9.4 will already have the fix.

The bug has now been filed by Matthew, thanks for that. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion