Re: [gentoo-user] Re: tensorflow-2.5.0-r1 compilation failed

[Mark Knecht]

On Sat, Oct 16, 2021 at 11:06 AM gevisz <gevisz@×××××.com> wrote:
>
> сб, 16 окт. 2021 г. в 20:40, Mark Knecht <markknecht@×××××.com>:
> >
> > On Sat, Oct 16, 2021 at 9:50 AM gevisz <gevisz@×××××.com> wrote:
> > >
> > > To make things worse, I've got an "Illegal instruction (core dumped)"
> > > error after installing and trying to run tensorflow from Ubuntu 20.04
> > > which is installed on the same computer.
> > >
> >
> > What processor by chance are you using. Probably a year back
> > Google started requiring processors with AVX2 and FMA instructions.
> >
> > I can no longer run it on my Intel i7 980 Extreme unless I build from
> > source which is just too painful. It's the main reason I'm starting
> > to finally plan a new machine purchase.
> >
> > cat /proc/cpuinfo | grep flags
>
> I have googled and also think that the above error on Ubuntu 20.04 is
> due to the old processor.
> I have an AMD Phenom II  X4 processor on that computer. The main
> problem, however, is that
> I get an error when compiling tensorflow in Gentoo.
>
> The CPU flags are the following: fpu vme de pse tsc msr pae mce cx8
> apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht
> syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm 3dnowext 3dnow
> constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid pni monitor
> cx16 popcnt lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a
> misalignsse 3dnowprefetch osvw ibs skinit wdt hw_pstate vmmcall npt
> lbrv svm_lock nrip_save
>

So yes, in my experience your Ubuntu experiment using precompiled
tensorflow is probably due to failure due to missing instructions in that
processor.

My experience compiling tensorflow was that it's fit and miss with lots
of bazel problems. However, that's mostly from 2 years ago. (I bought
my current house 2 years ago and the work I did was at the previous
house.) There were a LOT of issues getting it compiled, and for
clarity I was compiling on Ubuntu not Gentoo. That said my best results
were using build instructions where you were insider of a Docker
instance but you had to match things like python revisions in the
Docker image with the one you were going to use in your environment
outside of Docker. Generally this meant running TF in a specific
python virtual environment and not just in your login.

I wish you the best of luck. In the end I was able to buy a reasonably
priced product that used TF but had the part I couldn't run in a
library I could just remove. I lose certain features in the product
but can reinsert that library when I get a new machine.

Best of luck,
Mark