Re: [gentoo-embedded] ARMs w/ Floating Point

[M. Edward (Ed) Borasky]

Jack Poulson wrote:
> Unluckily, state-space calculations are usually chalk-full of all those
> guys.
>
> Thankfully, C is my language of choice, and I'd be more than happy with
> any references at all.
>
> On 7/14/07, M. Edward (Ed) Borasky <znmeb@×××××××.net> wrote:
>>
>> Jack Poulson wrote:
>> > There flight tests proved the current system inefficient for all of the
>> > planned computation and communication with the groundserver. They
>> > claimed to be using a 416 Mhz xscale, which seems fairly beefy to me,
>> > and I was under the impression that ramming floating point instructions
>> > through a fixed-point register was grossly inefficient.
>> >
>> > However, I am very new to embedded systems, and your question
>> > makes me think that the control algorithm itself is the culprit.
>>
>> In all likelihood you are emulating IEEE floating point in software,
>> most likely because fixed point algorithms aren't all that well known
>> outside of old fogeys like myself and people with a lust for squeezing
>> the ultimate out of embedded hardware.
>>
>> If you're lucky, it's only 32-bit arithmetic. If you're *real* lucky,
>> there are no logs, exponentials or trig functions in the code. If you're
>> really really lucky, there aren't even any divides or square roots. :)
>>
>> By the way, it's been so long since I did any of this stuff that I don't
>> know where one can find *current* references on the subject. But if
>> you're willing to get your hands dirty in C and/or Forth code, there are
>> some rewards waiting for you. You'd be amazed how much interesting stuff
>> you can do with table lookup.
>> --
>> gentoo-embedded@g.o mailing list
>>
>>
>
The general rule is, "if the hardware ALU can't do it, use table lookup." :)

Another general rule is "16-bit arithmetic is your friend"

But just getting rid of floating point emulation might be all you need
at 416 MHz.
-- 
gentoo-embedded@g.o mailing list