2010.11.09_어셈 이미지 반전 1, 2, 4 픽셀

#include <stdio.h>
#include <windows.h>

const int image_Size = 256*256;
const unsigned char mask1 = 0xFF;
const unsigned short mask2 = 0xFFFF;
const unsigned int mask4 = 0xFFFFFFFF;

LARGE_INTEGER diff_ticks;
LARGE_INTEGER freq, proc_start, proc_end;

void invertImageC1byte(unsigned char *src, unsigned char *dst);
void invertImageC2byte(unsigned short *src, unsigned short *dst);
void invertImageC4byte(unsigned int *src, unsigned int *dst);

int main()
{
 char input[]="gLenna_256_256.raw";
 char output[]="result.raw";

 unsigned char *src, *dst;
 unsigned char *p1;
 unsigned short *p2;
 unsigned int *p4;
 double *p8;

 FILE *fin=0, *fout=0;
 
 fin = fopen(input, "rb");

 if( !fin )
 {
  printf("file open error\n");
  return 0;
 }

 src = (unsigned char*)calloc(sizeof(unsigned char)*image_Size, 1);
 dst = (unsigned char*)calloc(sizeof(unsigned char)*image_Size, 1);
 p1 = src;
 p2 = src;
 p4 = src;
 p8 = src;

 fread(src, sizeof(unsigned char), image_Size, fin);
 fclose(fin);

 QueryPerformanceCounter(&proc_start);
 //////////////////

 invertImageC1byte(p1, dst);
// invertImageC2byte(p2, dst);
// invertImageC4byte(p4, dst);
// invertImageC8byte(p8, dst);

 //////////////////
 QueryPerformanceCounter(&proc_end);
 QueryPerformanceFrequency(&freq);

 diff_ticks.QuadPart = (proc_end.QuadPart - proc_start.QuadPart);
 fprintf(stdout, "Elapsed CPU time : %.12f sec \n\n", ((double)diff_ticks.QuadPart/(double)freq.QuadPart));

 fout = fopen(output, "wb");
 fwrite(dst, sizeof(unsigned char), image_Size, fout);
 fclose(fout);

 return 0;
}

void invertImageC1byte(unsigned char *p1, unsigned char *dst)
{
 int i;

 printf("1byte \n");
 for(i=0; i<image_Size; i++)
  dst[i] = p1[i] ^ mask1;
}

void invertImageC2byte(unsigned short *p2, unsigned short *dst)
{
 int i;
 int end = image_Size >>1;

 printf("2byte \n");
 for(i=0; i<end; i++)
  dst[i] = p2[i] ^ mask2;
}

void invertImageC4byte(unsigned int *p4, unsigned int *dst)
{
 int i;
 int end = image_Size >>2;

 printf("4byte \n");
 for(i=0; i<end; i++)
  dst[i] = p4[i] ^ mask4;

}

void invertImageAsm1byte(unsigned char *src, unsigned char *dst)
{
 printf("1byte Asm\n");
 __asm
 {
  mov esi, src
  mov edi, dst
  mov ecx, image_Size
L1:
  mov al, [esi]
  xor al, mask1
  mov [edi], al
  add esi, 1
  add edi, 1
  loop L1
 }
}

void invertImageAsm2byte(unsigned short *src, unsigned short *dst)
{
  printf("2byte Asm\n");
 __asm
 {
  mov esi, src
  mov edi, dst
  mov ecx, image_Size
 
L1:
  mov ax, [esi]
  xor ax, mask2
  mov [edi], ax
  add esi, 1
  add edi, 1
  loop L1
 }

}
void invertImageAsm4byte(unsigned int *src, unsigned int *dst)
{
  printf("4byte Asm\n");
 __asm
 {
  mov esi, src
  mov edi, dst
  mov ecx, image_Size
  shr ecx, 2
 
L1:
  mov eax, [esi]
  xor eax, mask4
  mov [edi], eax
  add esi, 4
  add edi, 4
  loop L1
 }
}