bumblebee-nvidia stuff

[Luca Boccassi]

On 20 October 2015 at 00:00, Andreas Beckmann <anbe at debian.org> wrote:
> On 2015-10-19 22:48, Luca Boccassi wrote:
>> On Mon, 2015-10-19 at 16:32 +0200, Andreas Beckmann wrote:
>
>>> Also I don't understand why there is a list of kernel module packages
>>> here, don't you need both
>>>  * the nvidia.ko kernel module (or however it is named currently) and
>>>  * the nvidia lib*GL* libraries (but not in the default searchpath)
>>> to run bumblebee stuff with the proprietary nvidia driver backend?
>>
>> Maybe for Cuda? You don't need the full driver and the GL libraries for
>> that, right?
>
> That should work ... that's what we have the cudaonly variant of the glx
> nvidia alternative. Maybe we don't cover the case of having
> nvidia-driver installed but wanting to use nvidia only for cuda and xorg
> with vesa only ... but we could make this possible, too
>
>> In that case, one could have just bumblebee and the kernel
>> module. But I've never tried such a setup.
>
> Did I get this right: in this scenario one uses the nvidia gpu for cuda
> and bumblebee only to turn the nvidia gpu off if one doesn't run cuda.
> And there is either no X (on a notebook?) or nothing uses acceleration
> (otherwise that would need the nvidia GL libs via 'primusrun
> glx-hungry-stuff'?)

Bumblebee has a --no-xorg option specifically for Cuda/OpenCL, which
as the name implies doesn't create an X session, but just starts the
card (and overloads the LD_LIBRARY_PATH). So, in theory, I think one
could have the notebook running X with mesa and use bumblebee without
the nvidia-driver and all the GL libs, but just with the module (and
relative dependencies) and Cuda/OpenCL. This is just in theory as I
have not tried bumblebee without the full nvidia-driver installed. I
can add it to the list if it could be useful.

Kind regards,
Luca Boccassi