Floating Point Support

[lpcware]

Content originally posted in LPCWare by Handi on Sat Jan 28 15:40:14 MST 2012
Hello folks,

I've recently discovered that the Cortex M3 doesn't support hardware floating point. I've never worked with 32-bit mcus and I am wondering how fast would the software fp operations be...

A quick search seems to indicate that Cortex M4 does support it, but I had troubles finding dev boards for it and from some comments I read it seems like it still lacks proper documentation.

Ideally, I'm looking for a hardware fp supporting 32-bit mcu that costs under 20$ and has available dev boards under 50$. Preferably available on digikey.

Any help would be very welcome since I planned ordering components on monday and I am on a relatively tight schedule.

Thanks
pav

[lpcware]

Content originally posted in LPCWare by Handi on Wed Feb 01 23:27:40 MST 2012
Hello,
Thanks for your input!

I've bought an stm32f4 dev board, which has a cortex m4 core. They offered a very good price for the board (20$).

I am not sure yet if single precision will be sufficient, but I will give it a try. I am planning to do some Kalman filtering with many variables, which will require floating point operations on large matrices. I am still at a relatively early stage of the process, so I don't know exactly the computational power that I'll need. In any case, since my matrices should be well defined, and a high degree of convergence is not required, I think single point will suffice.

In any case, I've bought an lpc1769 cortex m3 dev board at the same time to compare the two.

Now I just have to choose a good IDE.

[lpcware]

Content originally posted in LPCWare by TheFallGuy on Sun Jan 29 02:44:11 MST 2012
Don't forget that cortex-m4 CPU is single precision (floats or 32 bits) only. If you use doubles it will use software -same as for cortex-m3.

The question is, how critical is floating point performance? If you just need it occasionally, the software emulation will be ok. I suggest you try your algorithm on lpc1769 and see how close it is to meeting you needs, or not.

There is a Hitex lpc4350 board available now, but it is a lot more than $50. As lpc4350 is flash less it has lots of external flash and peripherals. Lpc4350 also has a 2nd CPU, so probably overkill for what you want.

[lpcware]

Content originally posted in LPCWare by Rob65 on Sat Jan 28 23:31:18 MST 2012
Pav,

a forum search learns that there is a different thread about the Cortex M4 [B][U]here[/U][/B].
The funny search is that a second search delivered [U][B]this page[/B][/U] about floating point support.

I thought I had seen a thread mentioning development boards but cannot find it anymore.

And then there is of course always the nxp.com website, this gives me [U][B]ics.nxp.com/products/lpc4000/lpc43xx/[/B][/U]. A nice page with a description of the different MCUs that are part of NXP's Cortex M4 family.

But I hope you do realize that the M4 is still quite new and that part will not be all be available in large quaitities from day 1.

I can only find the lpc4310/20/30/50 available right now (ROM/Flash less parts) with prices in the € 9,- .. € 13,- range but the development board is much more expensive than your 50$ price limit. But it should not be too hard to create your own board with a part like the 4310.

The question of course is why you need hardware FP support.
I don't have any need for FP numbers at all, I am able to solve all my problems using fixed point math and 32 bits fixed point deliver higher accuracy than 32 bits floats. But then not all problems can be solved using fixed point - a fixed/float software combination could still be fast enough.
But it all depends on the speed and your problem.

Current lpc1769 boards run up to 120 MHz.
The new lpc43xx family runs up to 204 MHz - which is the fastest M4 currently available.

Hope this helps.

Regards,[INDENT]Rob
[/INDENT]