/**************************************************\
*    Functions developed by zhong yunde.           *
*   filename    : test_readfile.cpp                  *
*    creat time    : 2008/05/08                      *
*   功能：测试文件读取模式(用于程序优化)          *
\**************************************************/

#include <stdio.h>
#include <assert.h>
#include <stdlib.h>
#include <time.h>

#define XSIZE 2592
#define YSIZE 1944
#define IMGSIZE XSIZE*YSIZE
typedef struct RGB {
    unsigned char r;
    unsigned char g;
    unsigned char b;
} RGB;

struct RGB in[IMGSIZE];
unsigned char out[IMGSIZE];

// 彩色图象转黑白图象
void calc_lum1(void)
{
    for(unsigned long i=0;i<IMGSIZE;i++)
    {
        double r,g,b,y;
        r=in[i].r; g=in[i].g; b=in[i].b;
        y=0.299*r+0.587*g+0.114*b;
        out[i]=(unsigned char)y;    
    }
}

void calc_lum2(void)
{
    for(unsigned long i=0;i<IMGSIZE;i++)
    {
        double r,g,b,y;
        r=in[i].r; g=in[i].g; b=in[i].b;
        y=(long)(r*1224+g*2404+b*467)>>12;
        out[i]=(unsigned char)y;    
    }
}

// 查表数组
register unsigned char D[256],F[256],E[256];
void table_init()
{
    for(short i=0;i<256;i++)
    {
        D[i]=(i*1224)>>12;
        E[i]=(i*2404)>>12;
        F[i]=(i*467)>>12;  
    }
}
void calc_lum3(void)
{
    for(long i=0;i<IMGSIZE;i++)
    {
        unsigned char r,g,b;
        r=D[in[i].r];  g=E[in[i].g]; b=F[in[i].b];  //查表
        out[i] = r+g+b ;
    }
}

inline void calc_lum5(void)
{
    for(long i=0;i<IMGSIZE;i++)
    {
        unsigned char r,g,b;
        r=D[in[i].r];  g=E[in[i].g]; b=F[in[i].b];  //查表
        out[i] = r+g+b;    
    }
}

// 双核处理
void calc_lum4(void)
{
    for(long i=0;i<IMGSIZE;i += 2)
    {
        double r,g,b,y,  r1,g1,b1,y1;
        // ALU1
        r=D[in[i].r];  g=E[in[i].g]; b=F[in[i].b];  //查表
        y=(long)(r+g+b)>>12;
        out[i]=(unsigned char)y;
        // ALU2
        r1=D[in[i+1].r];  g1=E[in[i+1].g]; b1=F[in[i+1].b];  //查表
        y1=(long)(r1+g1+b1);
        out[i]=(unsigned char)y1;
    }
}



int main()
{

    char *filein = "Bretagneg.raw";
    char *fileout = "Black.raw";
    clock_t start_time,end_time;
    FILE *IN=NULL, *OUT=NULL;
    unsigned long i=0;

    if ( NULL==(IN=fopen(filein, "rb"))) {
        fprintf(stderr, "Failed to open %s for reading !!\n", filein);
        return -1;
    }
    
    if ( NULL==(OUT=fopen(fileout, "wb")) ) {
        fprintf(stderr, "Failed to open %s for reading !!\n", fileout);
        return -1;
    }

    // 方式一 : 整体读取 ,指针赋值
    fseek(IN,0,SEEK_SET);
    start_time=clock();
    unsigned char *image=(unsigned char *)malloc(sizeof(RGB)*IMGSIZE);
    fread(image,sizeof(RGB)*IMGSIZE,1,IN);
    for (i=0;i<IMGSIZE;i++)
    {
        in[i].r = *(image++);//image[3*i];
        in[i].g = *(image++);//image[3*i+1];
        in[i].b = *(image++);//image[3*i+2];
    }
    image -= sizeof(RGB)*IMGSIZE;
    free(image); image =NULL;
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);

    // 方式二 : 整体读取 ,数组赋值(竟然比上面指针方式好!)
    fseek(IN,0,SEEK_SET);
    start_time=clock();
    image=(unsigned char *)malloc(sizeof(RGB)*IMGSIZE);
    fread(image,sizeof(RGB)*IMGSIZE,1,IN);
    for (i=0;i<IMGSIZE;i++)
    {
        in[i].r = image[3*i];
        in[i].g = image[3*i+1];
        in[i].b = image[3*i+2];
    }
    free(image); image =NULL;
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    // 方式三 : 一个个读取,效果最差
    fseek(IN,0,SEEK_SET);
    start_time=clock();
    for (i=0;i<IMGSIZE;i++)
    {
        in[i].r = fgetc(IN);
        in[i].g = fgetc(IN);
        in[i].b = fgetc(IN);
    }
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    // 方式一 ：直接套用浮点公式
    start_time=clock();
    calc_lum1();
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    // 方式二 ：转化为定点公式
    start_time=clock();
    calc_lum2();
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    table_init(); // 初始化表格

    start_time=clock();
    calc_lum3();
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    start_time=clock();
    calc_lum5(); // 只有一个ALU 却使用双核模式则可能反而降速
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    start_time=clock();
    calc_lum4(); // 只有一个ALU 却使用双核模式则可能反而降速
    end_time=clock();
    printf("the Time of Transform=%6.5f(sec)\n",(float)(end_time-start_time)/CLOCKS_PER_SEC);
    
    
    fwrite(out, 1, IMGSIZE, OUT);
    printf("\n completely \n");

    return 0;
}


//! 结论 :了解机器结构，编写和硬件对应的程序
/************************************************************
* 1 :文件读取建议整体读取，即使这样需要额外数据重排操作,
*    但是内存处理相对硬盘文件读写速度很快
* 2 :并非所有浮点操作转成定点操作，速度就能提高，如上由于PC机
*    有存在硬件浮点处理器，因此浮点操作不一定比定点慢
*************************************************************/