[Pkg-opencl-devel] adding more virtual packages libopencl-1.X-1 ?

[Vincent Danjean]

Using pkg-opencl-devel at lists.alioth.debian.org that I created but that
was not used yet, CCing Tomasz and Simon that can also be interesting
into subscribing for this ML.

  Hi,

Le 24/07/2013 01:36, Andreas Beckmann a écrit :
> Hi Vincent,
> 
> #717606 shows some problem with the current libopencl1 approach:
> pyopencl wants to use some OpenCL 1.2 features and therefore fails to
> run with nvidia-libopencl1.
> 
> What do you think about the following approach:
> 
> * each vendor-libopencl1 package adds
>     Provides: libopencl-1.1-1, libopencl-1.2-1, ...
>   for *all* supported versions of the standard
>   (the name here is libopencl-${standardversion}-${soversion})
>   (i.e. nvidia only 1.1, amd and ocl-icd both 1.1 and 1.2)
>   (we probably don't need 1.0, do we?)
> 
> a) every package needing something newer than the baseline (which
> currently is 1.1) manually adds Depends: libopencl-X.Y-1
> that would make apt choose a sufficient alternative
> 
> and/or
> b) the shlibs files add a dependency on libopencl-X.Y-1 (where X.Y is
> the newest standard supported by this vendor implementation) instead of
> libopencl1.

The b approach would be more transparent for users. [and the a option
can be used for applications that require a specific OpenCL version
that is not reflected by their use of symbols].

  I will try to implement it quickly in ocl-icd.

  As a minor detail, I would keep libopencl-... package name for libraries
and instead use directly opencl-1.1 and opencl-1.2 (the name of the norm
that the application requires). What do you think of that ?

  Regards,
    Vincent

> Andreas
> 
> PS: I recently introduced a virtual package libcuda-5.0-1, mainly for
> use by Ubuntu that don't have a separate libcuda1 package but instead
> bundle libcuda.so.1 in each of their nvidia driver packages
> Updating the Build-Depends/Depends of nvidia-cuda-toolkit for each new
> nvidia driver package was quite cumbersome, so we implemented this
> "versioned" virtual package approach (using variant (b) from above) -
> the first test for thsi approach will be cuda 5.5
> 


-- 
Vincent Danjean          Adresse: Laboratoire LIG - Bât. INRIA Rhône-Alpes
Téléphone:  +33 4 76 61 55 10            655 avenue de l'Europe
Fax:        +33 4 76 61 52 52            Montbonnot Saint Martin
Email: Vincent.Danjean at imag.fr           38334 Saint-Ismier cedex