[Pkg-opencl-devel] opencl-icd selection

[Giuseppe Bilotta]

On Mon, Nov 17, 2014 at 10:34 PM, Rebecca N. Palmer
<rebecca_palmer at zoho.com> wrote:
> This platform order is only meant to provide a sane default for applications
> that just want "a working, and preferably fast, device"; those that need
> something more specific will still need to do their own choosing or require
> manual intervention.
>
>> amd (fglrx) vs nvidia?
>
> I put mesa (=AMD hardware) in the middle because it can be either a discrete
> or an integrated GPU, with the intention of getting the right order for most
> machines.

And would you u amd-opencl-icd before or after nvidia?

> The obvious way to estimate speed from clinfo parameters (compute units *
> native vector width * max clock) doesn't work: it says my CPU
> (4*4float*3.2GHz) should be nearly the same speed as my GPU
> (16*4float*1GHz), but it is actually 8-16x slower.

Also, I don't think it's worth it to go this much into detail, we just
need to be "good enough" for most applications. (BTW What hardware are
you comparing? Do you manage to saturate the devices compute-wise or
are you memory-bound?)

>> (Yes, I have a couple of machines where I have discrete GPUs from both
>> vendors installed).
>
> Can you test:
> -does mesa-opencl-icd+beignet still break things (#768185) when mesa has
> hardware, and if so is mesa also broken by this

Can't test it in the immediate. The machine is in production and I
can't play around with its setup for the time being. I'll try to
gather the information as soon as possible.

> -which order (CPU or GPU first) amd-opencl-icd returns its devices in

GPU first.

>> Maybe this could be implemented at the distribution level
>
> Debian packaging can apply arbitrary source changes (debian/patches), so if
> we want to make some or all of this Debian-specific, we can do so without
> involving environment variables.

Right.

-- 
Giuseppe "Oblomov" Bilotta