Bug#895729: RFS: mkl-dnn/0.15+git20180803.3f58c1 [ITP] -- tensorflow dependency (amd64 specific)
control: tag -1 +moreinfo
On Thu, Aug 09, 2018 at 08:01:10PM +0200, Adam Borowski wrote:
> On Thu, Aug 09, 2018 at 10:16:17AM +0000, Lumin wrote:
> > * Package name : mkl-dnn
> > Version : 0.15+git20180803.3f58c16-1
> > Upstream Author : intel
> Alas, the build flags use -march=native -mtune=native which is a big no-no.
> The first makes the package crash on any processor lacking an extension that
> was present on the build machine and was used by the compiler; unless some
> kind of runtime detection is used, packages are allowed only the baseline
> ISA for the architecture. As for -mtune=native, it makes the package build
> unreproducibly, differing based on where it was compiled.
My bad, I overlooked the two flags. The cmake files have been patched
in master branch of packaging repo.
> The second problem is that in the testsuite, test_convolution_format_any
> fails (0/5 sub-tests). This might be related to my machine being:
> vendor_id : AuthenticAMD
> model name : AMD Phenom(tm) II X6 1055T Processor
Well, I have been waiting for intel to fix test failures for a long
time. Finally the snapshot 0.15+git20180803.3f58c16 doesn't fail
any test on dom-amd64 (E5 2699v?) and my I5-7440HQ, but now it failed
on AMD cpu ...
> Log of the FTBFS attached.
Thanks for the log, I've forwarded it to upstream.
I shouldn't let any test failure from mkl-dnn pass, so we have to wait
for upstream to fix the problem. Fortunately, TensorFlow can be compiled
with or without mkl-dnn. It doesn't matter if the initial upload of
TensorFlow is not linked against mkl-dnn. The difference that mkl-dnn
would bring to TensorFlow is computation speed-up.