[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: xsimd portability to other archs?



Hi Nilesh,

Le 13/10/2021 à 21:42, Nilesh Patra a écrit :
Hi Julien,

Very sorry for not responding earlier.
On Fri, Oct 08, 2021 at 11:47:32AM +0200, Julien Lamy wrote:
Hi Nilesh,

Le 07/10/2021 à 07:48, Nilesh Patra a écrit :
Hi Julien,

xsimd was accepted into the archive many weeks back[1]
Since you had access to the porter boxes for debugging this,
did you have any luck/any progress debugging it and getting this
built on the archs it doesn't go well on?

Using porterboxes and a RPi, I could get the unit tests to run on armv7l,
aarch64, mips64 and ppc64le. However, this required downloading a specific
version of the gtest framework and not using the one packaged in Debian
(DOWNLOAD_GTEST=ON).

Ah, are the test failures due to gtest instead of the package itself?

My guess exactly.

Would you know what exactly might be the delta between the gtest you fetched and the
one in the archive?

Sadly no. If I understand correctly, the test fetch from a clone of gtest [1], which, at the time I tested on porterboxes was not on par with its upstream. I'm not familiar enough with gtest to pinpoint where the error could stem from (although I strongly suspect something related to the template machinery of xsimd).

Moreover, the two other sample programs (a benchmark
and a generator of the Mandelbrot set) do not both compile on all
aforementioned arches. It seems that for now the support of non-x86 arches
is not fully there.

Okay, I see. Maybe before release we could limit the archs then?
Or is it only due to gtest and we simply disable tests on those few archs?

I don't think it's purely related to gtest: the test programs (benchmark and Mandelbrot set) don't use it, and also fail to compile. Regarding the arch limitation, Drew Parsons suggested earlier [2] that build failures on a never-uploaded package wasn't really a problem: we could keep the failing architectures so that, when the situation improves with a future upstream version, they would be packaged automatically. I've noticed that upstream uses non-x86 architectures in their CI framework, so things are moving in the right direction.

On a related note, I've updated the repository on Salsa to the latest
upstream release (7.6.0): CI passes (with the exception of the crossbuild to
arm64), could I get a sponsored upload for xsimd?

Ofcourse, I just uploaded now. Thank you very much for your work!

Thanks for the upload!

[1] https://github.com/xtensor-stack/xsimd/blob/master/test/downloadGTest.cmake.in#L17
[2] https://lists.debian.org/debian-science/2021/04/msg00017.html

Cheers,
--
Julien


Reply to: