[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Fsuk-manchester] Open ARM GPU drivers
From: |
Dave Love |
Subject: |
Re: [Fsuk-manchester] Open ARM GPU drivers |
Date: |
Tue, 05 Feb 2013 17:42:02 +0000 |
User-agent: |
Gnus/5.110011 (No Gnus v0.11) Emacs/22.2 (gnu/linux) |
Bob Ham <address@hidden> writes:
>> I'd be interested if anyone knows about the possibility of using the
>> drivers for floating point work, presumably via OpenCL.
>
> Non-existant I'd expect. The leading (only?) free software
> implementation of OpenCL is Clover¹
[There are others, for some value of "implementation", but I don't know
how useful/available they are; e.g an open64 front/back end.]
> As far as I know, there are three main ARM graphics cores with projects
> to develop a free software driver: the Mali, Adreno and PowerVR SGX.
Thanks for the info.
>> If so, what
>> sort of performance might Mali 400, for instance, provide?
>
> Good question. I can't imagine you'd get anywhere near the kind of
> performance of a desktop graphics chip due to the power constraints.
Yes, but the comparison wouldn't be with other GPUs, but with SIMD
(SSE-type) units. For what it's worth, linear algebra with a free
library using Neon (1 GHz A9) is ~10 times slower then, say, typical AMD
Barcelona-type cores, but can't tackle big enough problems anyway.
> I'm not surprised; the focus of most efforts are intended to support
> things like simulations, computational biology and other scientific
> computing, often in supercomputers.
[That's what I do, for some value of "supercomputers".]
> In that kind of context it doesn't make much sense to target low
> power, performance-constrained chips like the Mali.
It might, if you have them anyway in a well-ARMed system intended for
specific sorts of jobs, even if you wouldn't use them for HPC.
> (Unless, of course, you're able to engineer your own floating-point
> applications directly for the hardware.
The appeal of OpenCL is in avoiding some level of hardware specifics
generally, though you typically do have to write hardware-specific
assembler to get performance in the kernels :-(.