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

[gevisz <gevisz@gmail.com>]

пн, 10 янв. 2022 г. в 18:16, Mark Knecht <markknecht@gmail.com>:
>
> <SNIP>
> >
> > Thank you for your reply, Mark.
> >
> > Unfortunately, you missed my previous message in this thread
> > where I wrote that I do have Ubuntu 20.04 on the same computer.
> > However, tensorflow fails to run on it because it is not compiled
> > to be inconsistent with my videocard. So, Gentoo is my only option
> > for this hardware.
> >
> <SNIP>
>
> My apologies. This thread has gone on for a while and I had to review
> to get caught up.
>
> OK, so assuming I understand correctly (please correct me if I do
> not) then you are talking about ONE computer that uses an AMD
> Phenom II  X4 processor. This computer dual boots, or is a Gentoo
> machine with an Ubuntu VM.

It is a dual boot. (Sorry for the late reply.)

> In an earlier response to this thread you showed the flags
> supported by this processor which did not include
> the AVX, AVX2 or the FMA3/FMA4 flags. It is my
> understanding that this processor cannot run the
> current versions of tensorflow whether you compile
> it yourself or not, at least in the non-GPU version.

My hardware cannot run tensorflow precompiled by Ubuntu.
However, some versions of tensorflow successfully compile
and run in Gentoo:
- all versions of tensorflow-2.5 fail to compile in Gentoo;
- tensorflow-2.7.0 successfully compiled on 21-11-2021;
- on 30-12-2021 recompilation of the same tensorflow-2.7.0
  failed because of some changed dependencies;
- I have not yet tried to compile tensorflow-2.8.0.

Actually, I gave up trying to compile it and deleted it from my system.
Moreover, I am not going to compile it in the near future as I am now
in the war zone.

> WRT to your video card, tensorflow does not require the use
> of a GPU. There are two versions, tensorflow-cpu and
> tensorflow-gpu. If you were to build the -cpu version then it
> is my understanding it would run an a headless machine,
> presuming the processor has AVX/AVX2/FMA hardware
> support.
>
> If the processor DOES have AVX/FMA support but you were having
> problems emerging TF in Gentoo then a virtual machine running
> Ubuntu might have helped you as you could use a precompiled
> apt or snap package. However I don't think anything gets you
> past not having AVX/FMA hardware support.
>
> I am in the same situation. My big machine is an Intel i7 980
> Extreme. I used to be able to run TF but have not been able
> to since Google raised the CPU requirements.
>
> If I am not understanding your hardware setup, or you think
> there is a path around the AVX/FMA hardware problem
> please let me know and I'll explore it more deeply with you.

In addition to the old CPU, I have quite an old video card,
namely, ATI R4770. However, I still believe that it is possible
to compile tensorflow so that it could run on my hardware.
At least, I did it for tensorflow-2.7.0 on 21-11-2021.