[Dspforum] Double word data aligning

[Doug Wenstrand]

When you declare an array of floats:float samples[1024]

and then use the DATA_ALIGN(samples,8) -- note the 8, not 64 as it is in
bytes, not bits

it does exactly as you guessed :  samples[0] is on an 8 byte boundary, and
every next element comes 4 bytes later since float is a 4-byte type.

dspf_sp_dotprod doesn't do any skipping though.  It presumes that the data
it is operating on is aligned at a double word boundary.  It doesn't presume
that every element is on a double word boundary.  So, as long as the first
element that it gets is on a double word boundary, it works as expected on
single precision floating point data.  The reason it requires this
restriction is that it uses a load double-word instruction to load two
pieces of data at once.  That instruction needs to have a double-word
address to load.

Note though that aligning samples[1024] on a double-word boundary allows you
to call dspf_sp_dotprod(samples,...)
but this would cause a problem :
dspf_sp_dotprod(&samples[1],...)

what I'm saying is that the aligment that is important is the alignment of
the pointer that it gets.  It doesn't necessarily know anything about the
beginning of your array

does that make sense?


On Thu, Oct 8, 2009 at 6:19 PM, O'Connor, Sean R.
<Sean.OConnor at jhuapl.edu>wrote:

>  I have a few questions about data aligning.  For the DSP Lib function
> dspf_sp_dotprod, it requires pointers to two single precision floating point
> arrays which must be double word aligned.  So when you declare an array of
> floats, call it samples[], and then use DATA_ALIGN(samples, 64) to align it
> on a double word boundary (I think that is correct), will it place the first
> element on the boundary and then every 32 bits after place another array
> element?  If this is the case, will dspf_sp_dotprod skip over every other
> element in the array?
>
> Thanks,
> Sean
>
> _______________________________________________
> Dspforum mailing list
> Dspforum at echelonembedded.com
> http://echelonembedded.com/mailman/listinfo/dspforum_echelonembedded.com
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://echelonembedded.com/pipermail/dspforum_echelonembedded.com/attachments/20091008/3fa60a90/attachment.html>