[FFmpeg-devel] IDCT permutation (was: pre discussion around Blackfin dct_quantize_bfin routine)

[Marc Hoffman]

On 6/14/07, M?ns Rullg?rd <mans at mansr.com> wrote:
>
>
> Marc Hoffman wrote:
>
> > Is the ARM dsp optimizations complete from your point of view and this
>
> Far from it, I'd say.
>
> > is now where your planning to optimizing the code to get the biggest
> > bang for the buck?  The arm you are using what is the memory hierarchy
> > configuration?
>
> The ARM I'm using is an ARMv6 core.  I don't recall the exact specs off
> the top of my head.


So this change is for something that appears to be be ~2% of something that
probably has things which could improve much more in other place?

Would it be better to have the address calculation done as a post operation?
I guess what I mean is instead of computing the base+offset*2 for every
lookup would it be more efficient to store the displacements between
successive elements?

So the scan table would become something like

  2,  14,  16, -14, -14,   2,  14,  14,
 14,  16, -14, -14, -14, -14,   2,  14,
 14,  14,  14,  14,  16, -14, -14, -14,
-14, -14, -14,   2,  14,  14,  14,  14,
 14,  14,  14,   2, -14, -14, -14, -14,
-14, -14,  16,  14,  14,  14,  14,  14,
  2, -14, -14, -14, -14,  16,  14,  14,
 14,   2, -14, -14,  16,  14,   2,   0,

allowing direct access to the element and then modify the base as you walk
through the data set?

Marc