Re: GPU tests on master next - core-image-x11 - sfp

From: Alexandre Belloni <alexandre.belloni@free-electrons.com>
To: "Eric Bénard" <eric@eukrea.com>
Cc: meta-freescale@yoctoproject.org
Subject: Re: GPU tests on master next - core-image-x11 - sfp
Date: Thu, 29 Aug 2013 10:01:43 +0200	[thread overview]
Message-ID: <521EFFE7.8020304@free-electrons.com> (raw)
In-Reply-To: <20130828222721.07cd5727@e6520eb>

Hi,

On 28/08/2013 22:27, Eric Bénard wrote:
>
> on the GPU point of view no.
>
> I find interesting that the hardfp gives lower results than the
> softfp : maybe that would be interesting to have true real life
> benchmarks on this side (not GPU centric) before defaulting hardfp in
> meta-fsl-arm for i.MX6 platforms.

I'm not sure we can actually find real life benchmarks, it always
depends on how intensive are the floating point operation compared to
the rest of the system.

Also, I remember seeing that the difference between softfp and hardfp is
not as big on cortex-a9 as on cortex-a8. Actually, softfp is already
using the VFP but using the soft ABI. Then, it has to copy values from
integer registers to float registers and that is the overhead. This
overhead is lower on cortex-a9 because the vfp is pipelined, this was
not the case on cortex -a8.

You can find interesting benchs here:

https://wiki.linaro.org/OfficeofCTO/HardFloat/Benchmarks
https://wiki.linaro.org/OfficeofCTO/HardFloat/Benchmarks201205

-- 
Alexandre Belloni, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com