[maemo-developers] [maemo-developers] OGG support

[From: Paul Mundt]

On Sun, Oct 30, 2005 at 07:59:46PM +0100, Nils Faerber wrote:
> I am quite sure that the ARM9 does not support FP at all. What is IMHO
> supported are some FP traps that would trap into kernel or other FP
> software implementations. But that dies not speed up anything. There is
> to my knowledge no embedded CPU that has hw FP built in.
> 
Some versions of the ARM9 support VFP, which allows for partial hardware
FP support for some basic single and double precision opcodes, and traps
for the rest (it however does not offer a IEEE754 compliant interface in
hardware, and requires quite a bit of help from software to do so). OMAP
2420 supports this, OMAP 1710 does not. As such, we presently use
in-kernel NWFPE.

Additionally, since VFP still uses this antiquated co-processor
load/store scheme, it's pretty tough to get instruction level
parallelism, or decent performance out of the pipelining (unless you can
come up with some carefully crafted set of instructions that you can keep
in the icache long enough for it to matter). This is just speculation
though, someone with more direct VFP experience might have some better
insight in to this.

As far as embedded CPUs supporting hardware FP, this is a ridiculous
statement. Almost every other embedded architecture out there besides ARM
supports IEEE754 in hardware. Some opt for something more like the VFP
approach, where most of the heavy and frequently used opcodes are
handled, and the rest is left to software.

> All software FP implementations are of course very CPU intensive. You
> have to calculate the FP result in software which takes several hundred
> cycles while in hardware it would only take just a few cycles. So FP on
> emebdded devices will probably always be dead slow - at least for the
> next future.
> 
This is not necessarily true either, and is one of the bigger reasons for
pushing EABI. It makes sense to use VFP for what it supports natively,
most of the rest of it is better left to something like soft-float.
Kernel FP emulation is slow by definition.

> Concerning the 770 the integer version could be a first start for having
> OGG capabilities at all. The OMAP DSP device is open to a certain
> extend. The tools to develop code for it are freely available, some
> documentation is too. The DSP in the 770 is controlled by the also free
> DSP gateway project. So there is no big reason why there cannot be a
> free DSP OGG code any time in the near future. I guess on of the major
> problems here is that developing code for DSPs is not very well known in
> the open source community... on the other that could be a very
> interesting software project ;)
> 
An integer based decoder on MPU side would certainly be the most sensible
approach.