[Tinyos-devel] Crc interface buffer size restriction

[Andreas Köpke]

Philip Levis wrote:
> On Jan 14, 2009, at 1:07 AM, Andreas Köpke wrote:
> > Vlado Handziski wrote:
> >> On Tue, Jan 13, 2009 at 11:07 PM, Philip Levis
> >>
> >> <pal at cs.stanford.edu> wrote:
> >>> On Jan 13, 2009, at 1:41 PM, Philip Levis wrote:
> >>>
> >>> One other thought: if you want a 16-bit length CRC interface,
> >>> chances
> >>> are you want it to be split-phase, due to the possible time duration
> >>> of the operation.
> >>
> >> There are also alternative table look up versions in the eyesIFX
> >> folder
> >> that can be useful.
> >
> > However, the table lookup is only twice as fast as the current default
> > implementation, which itself is an 10 times improvement over the
> > original bit
> > shift for loop. You might want to do time measurements before
> > deciding whether
> > it should be split phase: the LPL channel check for telosb takes
> > 11ms in an
> > atomic block
>
> This is totally broken and should be fixed.

Well -- tough task. It would require to enable the interrupt on the CCA pin, 
but currently there is a single component that is coded around the assumption 
that it is the only receiver of such an interrupt. 

> > and the 64bit multiplication of the default random number
> > generator takes about 1ms in an atomic block.
>
> It's doing the multiplication in software because the codegen for
> multiply is broken in the gcc version we use. IIRC, boomerang solved
> this by doing a little bit of assembly for a hw multiply. We might
> want to do the same thing.

Question is rather: why does it have to do this in an atomic block? Did anyone 
check lately whether the HW multiplier still does not work for the latest 
msp430-gcc?

> > So I guess we can spare the 1ms
> > for 16bit CRC in an non-atomic block easily.
>
> My concern was doing a CRC over 65,535 bytes. That would take much
> more than 1ms, right?

On an eyesIFX it would take about 0.9s -- slightly more than 1ms ;-)
For telosb it should be around 0.45s, due to the fact that it uses the table 
lookup from Ross' CRC guide. Well, not many nodes do have 64k RAM.

> Phil