Error Concealment | JPEG | MPEG-4 | Wavelet | YUV

Hello,

I am trying to optimize my code about RGB-Gray conversion on DM642.
I write this conversion as a function and insert it into the main function.
The main function deals with image capturing, YUV-RGB conversion, and output to VGA port.
Without the inserted function, everything works fine.

Therefore, the performance is obviously affected by the inserted function.
I've tried almost any optimizaiotn approach I know except packed-data
accessing.
Approaches I used are─
1.the restrict Keyword
2.pragma directive MUST_ITERATE, UNROLL
3.intrinsics _nassert, _extu
4.manually refine code to unroll the for loop instead of using UNROLL.

But it seems these approaches don't work fine.

However, according to the feedback from compiler, I find bit-shifting by
_extu is a major burden of performance, and word alignment is not achieved.
Besides, only few registers are used.

So, any advice on further optimizaiton?

What I come up with are─
1.packed-data accessing
2.DATA_ALIGN
3.another way to unroll the for loop
4.other possible intrinsics
5.task scheduling at dsp/bios level
6.Cache usage problem

Victor 

PS.
My code is as follows─
===========================================================================
void RGB565_RGB(unsigned short* restrict BMP, unsigned int ByteUsed)
      {
            unsigned int i;

            unsigned int R=0;
            unsigned int G=0;
            unsigned int B=0;

            _nassert(((int) BMP & 0x3) == 0);

      #pragma MUST_ITERATE(640,, 4);
      /*#pragma UNROLL(8);  */

      for(i=0; i<ByteUsed; i++)
            {

                  R= _extu(BMP[i], 16, 27);
                  G= _extu(BMP[i], 21, 26);
                  B= _extu(BMP[i], 27, 27);
                  R=G=B=((R+(G>>1)+B)*341)>>10;
                  BMP[i]=((R<<11)|(G*2<<5)|B);

            }

      }
===========================================================================

(You need to be a member of imagedsp -- send a blank email to imagedsp-subscribe@yahoogroups.com )