[LAD] opencl [was Re: [LAU] ANN: CMKeyboard]

[Patrick Shirkey]

On Tue, December 11, 2012 3:33 am, David Robillard wrote:
> On Mon, 2012-12-10 at 11:12 -0500, Paul Davis wrote:
>> On Mon, Dec 10, 2012 at 10:56 AM, Patrick Shirkey <
>> pshirkey at boosthardware.com> wrote:
>>
>> >
>> > >
>> >
>> > We have several headless machines running GPU's with thousands of
>> > processing units available.  Much more power than the first "Lord of
>> the
>> > Rings" movie was made with.
>> >
>> > > Still worth exploring though, and a "cl~" processor for my system is
>> > > definitely on the todo list.
>> > >
>> >
>> > We are exploring the possibilities here too. Essentially a library
>> that
>> > allows sending specific operations across a netjack cluster for
>> realtime
>> > multimedia processing.
>> >
>>
>> you might want to check the latency before you get too far into plans
>> like
>> this. i've heard that it is improving, but still not really what one
>> would
>> expect of a "realtime FX processor".
>
> Indeed.  Crazy throughput, but transferring to and from the GPU kills
> you.  Worth investigating anyway since many-core is probably going to
> stick around and become faster, but I doubt current GPUs can achieve
> reasonable real-time audio latency.
>

This is the part we are working on. It requires building the software and
then discussing the issues with the driver developers directly. Takes time
and effort in more ways than one. Plus there are certain forces actively
working against us to put as many barriers in the way as possible.

> I think the programmable GPUs on recent Intel chips (Ivy Bridge) is an
> interesting development; though much less powerful and far less cores,
> they have full memory bandwidth (the other thing that kills you), and
> presumably minimal latency since it's on-die.  Adding 8 or so cores may
> not be in the same realm as adding hundreds, but for many things the
> latency and memory bandwidth dominates anyway so a billion cores on a
> GPU would still be slower.  Anything memory bound is much slower even on
> the best GPUs.  I'll happily take 8 extra cores on the CPU...
>

AMD Fusion chips are CPU/GPU combos so you get the best of both in that
regard. The next gen HSA platform promises to be even more powerful but
AMD will have to survive long enough for that promise to be delivered.
They have made some recent progress with opencl support for Linux so that
has been well received.

> Unfortunately they're not programmable on Linux yet, only on Windows.

I have been assured that the brand new NUC platform from Intel is fully
Linux compatible with opengl/cl support too. What specific issues are you
seeing?

> A  complete joke if they're targeting scientific GPGPU, and useless for LAD
> too.  Intel seriously needs to get on this and fix their drivers...
>

Compared to AMD, Intel are actually making a lot of headway with their
Linux drivers. At least they actually release the specs these days and
have a full department for Linux development. AMD are having real trouble
getting their heads around the concept of open source *and* multimedia. My
dealings with Intel so far suggest they are light years ahead in terms of
understanding the potential. Compared to AMD it's almost night and day.



--
Patrick Shirkey
Boost Hardware Ltd