Advanced Micro Devices Inc (AMD)

[mmoy]

For the person that said that the simple approach is usually best:

Old code:

 
 	     for (i = 0, rp = row; i < row_width; i++, rp += 3) 
 	       { 
 		 png_byte save = *rp; 
 		 *rp = *(rp + 2); 
 		 *(rp + 2) = save; 
 	       } 
 

New code:

 
              int alignment = ((unsigned int) rp) & 3; 
  
              rp = row; 
              printf("png_do_bgr\n"); 
  
              /* Align on quadword */ 
  
              for (i = 0; i < alignment; i++, rp += 3) { 
 	       png_byte save = *rp; 
                *rp = *(rp + 2); 
 	       *(rp + 2) = save; 
 	     } 
              row_width -= alignment; 
  
              __asm{ 
                  mov    edi, DWORD PTR row_width 
                  mov    esi, DWORD PTR rp 
                  shr    edi, 2                  ; Div by 4 as couplet=4 
  
                do_couplet: 
                  mov    eax, DWORD PTR [esi] 
                  mov    ecx, DWORD PTR [esi+4] 
                  bswap  eax 
                  mov    edx, ecx 
                  mov    ch, al 
                  mov    al, dh 
                  ror    eax, 8 
                  mov    DWORD PTR [esi], eax 
                  bswap  ecx 
                  bswap  edx 
                  mov    eax, DWORD PTR [esi+8] 
                  mov    ch, al 
                  mov    al, dh 
                  bswap  ecx 
                  bswap  eax 
                  mov    DWORD PTR [esi+4], ecx 
                  rol    eax, 8 
                  mov    DWORD PTR [esi+8], eax 
  
                  add    esi, 12 
                  sub    edi, 1 
                  jne    do_couplet 
  
                  mov    DWORD PTR rp, esi 
              } 
              row_width &= 3;           /* Do the tail */ 
              for (i = 0; i < row_width; i++, rp += 3) { 
 	       png_byte save = *rp; 
                *rp = *(rp + 2); 
 	       *(rp + 2) = save; 
 	     }