Bug#795262: nvidia-cuda-toolkit: Reboot, after package installation, do not unload the Nouveau drivers

[François Legendre]

Hello,

1) Nvidia proprietary drivers are not used to build CUDA applications ;
2) Nvcc (Nvidia compiler for CUDA) generates a so called "ptx code" ;
3) It is the job of the Nvidia proprietary drivers to JIT compile the
ptx code with respect to the current GPU.

Then, when I installed the nvidia-cuda-toolkit, I was able to build
the CUDA application but unable to run the application. So, as already
told, I blacklisted the nouveau drivers.

Regards,
----
François


2015-08-21 10:16 GMT+02:00 Vincent Cheng <vcheng at debian.org>:
> On Fri, Aug 21, 2015 at 1:07 AM, Luca Boccassi <luca.boccassi at gmail.com> wrote:
>> On Aug 21, 2015 08:55, "Vincent Cheng" <vcheng at debian.org> wrote:
>>>
>>> Hi François and Luca,
>>>
>>> On Sat, Aug 15, 2015 at 4:07 PM, Luca Boccassi <luca.boccassi at gmail.com>
>>> wrote:
>>> > On Wed, 2015-08-12 at 13:58 +0200, François Legendre wrote:
>>> >> Package: nvidia-cuda-toolkit
>>> >> Version: 6.0.37-5
>>> >> Severity: normal
>>> >>
>>> >> Dear Maintainer,
>>> >>
>>> >> Many thanks for your work.
>>> >>
>>> >> I install the nvidia-cuda-toolkit using synaptic. After reboot, the
>>> >> Nouveau
>>> >> drivers are still loaded. The command
>>> >>
>>> >> # lsmod | grep nouveau
>>> >>
>>> >> prints some lines.
>>> >>
>>> >> I follow the indications found at
>>> >> http://docs.nvidia.com/cuda/cuda-getting-
>>> >> started-guide-for-linux/#axzz3iaadD0fW for Ubuntu distribution :
>>> >>
>>> >> 1. Create a file at /etc/modprobe.d/blacklist-nouveau.conf with the
>>> >> following
>>> >> contents:
>>> >> blacklist nouveau
>>> >> options nouveau modeset=0
>>> >> 2. Regenerate the kernel initramfs:
>>> >> # update-initramfs -u
>>> >>
>>> >> Now, it works.
>>> >
>>> > Hello François,
>>> >
>>> > Thank you for your report.
>>> >
>>> > I'm not sure we should blacklist nouveau by default in the nvidia-cuda
>>> > packages, since they do not depend on the nvidia-driver packages (I
>>> > believe Cuda can be used without the latter). It could leave users with
>>> > a broken system.
>>>
>>> AFAIK you need to install the proprietary driver in order to use cuda
>>> (but I could be mistaken)?
>>>
>>> I don't think nvidia-cuda-toolkit should ship its own modprobe.d conf
>>> file; the proprietary nvidia driver packages already do, and end users
>>> should already have that installed.
>>
>> Hello,
>>
>> I agree, I don't think the toolkit should ship a modprobe file.
>>
>> I thought the drivers were mandatory to run Cuda programs, but not to build
>> them. Unless I'm mistaken, the toolkit is used to build.
>
> Ah, that would make sense...I don't know a thing about how cuda works,
> to be honest.
>
>> If this is not the case, should then the toolkit package depend on the
>> drivers?
>
> That's probably a question that's best answered by either Andreas or
> Graham. Thoughts?
>
> Regards,
> Vincent