[Arm-netbook] RedSleeve - New ARM Linux Distro

[Gordan Bobic]

Barry Kauler wrote:
>> Have you actually looked at what the realistic performance advantage is
>> on something like a Tegra2 between running armv5tel and armv7hl for a
>> typical desktop workload? It's not all that much. Not to mention that
>> the GPU drivers and proprietary DSP libraries for most ARMs (the closed
>> ones since there are no open ones) are only available for the soft-float
>> target.
> 
> I would have to chase up comparisons that I have read. Can't recall
> all the details off the top of my head...
> 
> I do recall that ARMv7 thumb-2 instructions offer significant
> reduction in executable size (and speed) -- unlike thumb-1 which isn't
> very usable.

Thumb2 isn't what most non-embedded things target, though. Everything 
targets the ARM instructions set. Thumb2 is mainly used for things like 
Java that execute intermediate byte code, IIRC. It replaces the java 
extensions that were available on the older ARMs.

> I should add that Ubuntu 12.04 will have ARMv7 hard-float binary packages.

Sure it will, but those aren't going to include proprietary binary 
drivers until the upstream vendors make them available in hard-float 
variants. Nvidia certainly don't do that yet, and I suspect most of the 
others don't either.

> I intend to compile some multimedia apps with NEON turned on,
> especially for my target hardware (A10).

You may be disappointed with the improvements. Last time I checked, gcc 
couldn't even leverage x86 SSE properly:

http://www.altechnative.net/2010/12/31/choice-of-compilers-part-1-x86/

Things may have improved slightly since then, but I'll believe it when I 
see it. If they hadn't had vectorization working properly over a decade 
after MMX and SSE instruction sets became available, what are the 
chances that another few years would have made all the difference? ;)

> Um, I found one of the reference docs that I had previously read:
> 
> http://wanderingcoder.net/2010/07/19/ought-arm/
> 
> ...they reckon that NEON is in all the ARMv7-based chips that they
> have encountered.
> ...no timing comparisons on that site.

Definitely not on all of them. Tegra2 is a Cortex A9 that doesn't have NEON.

> My thinking is that we have to try and extract maximum speed wherever
> we can with these chips, to get apps to have acceptable
> responsiveness. A little bit here, a little bit there, it adds up.
> Thumb-2 is one of those important "little bits".

I'm pretty sure it doesn't do what you think it does...

> I guess that I am being more hardware-specific than Red Sleeve. I am
> thinking in terms of ARMv7 +NEON and Mali GPU -- latter has drivers, a
> complete open source project for the GPU (kernel, openGL and
> X11drivers) is well underway.

And by the time it is stable enough for general consumption it'll be 
another 12-18 months, knowing how quickly these things progress. Nvidia 
have been working on their Tegra binary drivers for years and they are 
still unusably buggy.

> But perhaps that's what we need to be,
> fairly narrowly targeted, given the lack of standardisation (and so
> much closed up/secret) of ARM-based hardware.

That depends on your point of view. For a lot of things, targeting the 
widest range of devices is an advantage, especially since the software 
that benefits from the narrowly targeted improvements is not very 
numerous. Not to mention that the more narrowly you target, the more 
obscure the code paths are that you are using in your tool chain, so you 
are likely to trip over bugs that nobody found before you. It happens a 
lot once you stray off the straight and narrow.

Gordan