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




#define sizeTable 2048
#define repeatNumber 100

//
clock_t startTime = 0;
clock_t start = 0;
clock_t end = 0;

struct vector
    {
    float a;
    float b;
    float c;
    float d;
    };


struct vector a[sizeTable];
struct vector b[sizeTable];
struct vector sum[sizeTable];
struct vector sub[sizeTable];
struct vector mul[sizeTable];
struct vector div_res[sizeTable];


//SIMD------------------

void sum_simd(struct vector a, struct vector b, struct vector * wynik){

asm(
    "movaps %1, %%xmm0\n\t"
        "movaps %2, %%xmm1\n\t"
        "addps %%xmm1, %%xmm0\n\t"
        "movaps %%xmm0, %0\n\t"
   : "=m" (*wynik)
   : "m" (a),"m" (b)
    );

}

void sub_simd(struct vector a, struct vector b, struct vector * wynik){
asm(
    "movaps %1, %%xmm0\n\t"
    "movaps %2, %%xmm1\n\t"
    "subps %%xmm1, %%xmm0\n\t"
    "movaps %%xmm0, %0\n\t"
   : "=m" (*wynik)
   : "m" (a),"m" (b)
    );
}



void div_simd(struct vector a, struct vector b, struct vector * wynik){
asm(
    "movaps %1, %%xmm0\n\t"
    "movaps %2, %%xmm1\n\t"
    "divps %%xmm1, %%xmm0\n\t"
    "movaps %%xmm0, %0\n\t"
   : "=m" (*wynik)
   : "m" (a),"m" (b)
    );
}

void mul_simd(struct vector a, struct vector b, struct vector * wynik){
asm(
    "movaps %1, %%xmm0\n\t"
    "movaps %2, %%xmm1\n\t"
    "mulps %%xmm1, %%xmm0\n\t"
    "movaps %%xmm0, %0\n\t"
    : "=m" (*wynik)
   : "m" (a),"m" (b)
    );
}



//SISD-------------------------------------------------

void sum_sisd(struct vector a, struct vector b, struct vector * wynik){

asm(
    "fld %4\n\t"    //wrzuc a.a na stos FPU
    "fadd %8\n\t"   //dodaj b.a do st0
    "fstp %0\n\t"   //przenies st0 do wynik->1

    "fld %5\n\t"
    "fadd %9\n\t"
    "fstp %1\n\t"

    "fld %6\n\t"
    "fadd %10\n\t"
    "fstp %2\n\t"

    "fld %7\n\t"
    "fadd %11\n\t"
    "fstp %3\n\t"

   :    "=m" (wynik->a), //0
        "=m" (wynik->b), //1
        "=m" (wynik->c), //2
        "=m" (wynik->d)  //3
   :    "m" (a.a), //4
        "m" (a.b), //5
        "m" (a.c), //6
        "m" (a.d), //7
        "m" (b.a), //8
        "m" (b.b), //9
        "m" (b.c), //10
        "m" (b.d) //11
    );

}
void sub_sisd(struct vector a, struct vector b, struct vector * wynik){

asm(
    "fld %4\n\t"
    "fsub %8 \n\t"
    "fstp %0\n\t"

    "fld %5\n\t"
    "fsub %9\n\t"
    "fstp %1\n\t"

    "fld %6\n\t"
    "fsub %10\n\t"
    "fstp %2\n\t"

    "fld %7\n\t"
    "fsub %11\n\t"
    "fstp %3\n\t"

   :    "=m" (wynik->a), //0
        "=m" (wynik->b), //1
        "=m" (wynik->c), //2
        "=m" (wynik->d)  //3
   :    "m" (a.a), //4
        "m" (a.b), //5
        "m" (a.c), //6
        "m" (a.d), //7
        "m" (b.a), //8
        "m" (b.b), //9
        "m" (b.c), //10
        "m" (b.d) //11
    );

}
void mul_sisd(struct vector a, struct vector b, struct vector * wynik){

asm(
    "fld %4\n\t"
    "fmul %8 \n\t"
    "fstp %0\n\t"

    "fld %5\n\t"
    "fmul %9\n\t"
    "fstp %1\n\t"

    "fld %6\n\t"
    "fmul %10\n\t"
    "fstp %2\n\t"

    "fld %7\n\t"
    "fmul %11\n\t"
    "fstp %3\n\t"

   :    "=m" (wynik->a), //0
        "=m" (wynik->b), //1
        "=m" (wynik->c), //2
        "=m" (wynik->d)  //3
   :    "m" (a.a), //4
        "m" (a.b), //5
        "m" (a.c), //6
        "m" (a.d), //7
        "m" (b.a), //8
        "m" (b.b), //9
        "m" (b.c), //10
        "m" (b.d) //11
    );

}

void div_sisd(struct vector a, struct vector b, struct vector * wynik){

asm(
    "fld %4\n\t"
    "fdiv %8 \n\t"
    "fstp %0\n\t"

    "fld %5\n\t"
    "fdiv %9\n\t"
    "fstp %1\n\t"

    "fld %6\n\t"
    "fdiv %10\n\t"
    "fstp %2\n\t"

    "fld %7\n\t"
    "fdiv %11\n\t"
    "fstp %3\n\t"

   :    "=m" (wynik->a), //0
        "=m" (wynik->b), //1
        "=m" (wynik->c), //2
        "=m" (wynik->d)  //3
   :    "m" (a.a), //4
        "m" (a.b), //5
        "m" (a.c), //6
        "m" (a.d), //7
        "m" (b.a), //8
        "m" (b.b), //9
        "m" (b.c), //10
        "m" (b.d) //11
    );

}

void SISD(){
    double timeSum = 0,timeSub = 0,timeMul = 0,timeDiv = 0;
    int j, i;
//usrednianie wyniku----------
    for(j = 0;j<repeatNumber;j++){

//dodawanie-----------
for(i = 0;i<sizeTable;i++){
            startTime = clock();
            sum_sisd(a[i],b[i],&sum[i]);
            timeSum += ((double)clock() - startTime)/CLOCKS_PER_SEC*1000.0;
        }

//odjemowanie---------
for( i = 0;i<sizeTable;i++){
            startTime = clock();
            sub_sisd(a[i],b[i],&sub[i]);
            timeSub += ((double)clock() - startTime)/CLOCKS_PER_SEC*1000.0;
        }

//mnozenie-------------
for(i = 0;i<sizeTable;i++){
            startTime = clock();
            mul_sisd(a[i],b[i],&mul[i]);
            timeMul += ((double)clock() - startTime)/CLOCKS_PER_SEC*1000.0;
        }
//dzielnie------------

for(i = 0;i<sizeTable;i++){
            startTime = clock();
            div_sisd(a[i],b[i],&div_res[i]);
            timeDiv += ((double)clock() - startTime)/CLOCKS_PER_SEC*1000.0;
        }

    }

   timeSum = timeSum/(float)(repeatNumber*sizeTable);
    timeDiv = timeDiv/(float)(repeatNumber*sizeTable);
    timeMul = timeMul/(float)(repeatNumber*sizeTable);
    timeSub = timeSub/(float)(repeatNumber*sizeTable);

printf("Liczba liczb: %d\nSredni czas [ms]: \n+ %lf \n- %lf \n* %lf \n/ %lf \n",(4*sizeTable),timeSum,timeSub,timeMul,timeDiv);
}

//generator-------------------------------------------------
void randomGenerator()
{
int i=0;
    for ( i; i < sizeTable;i++){
        a[i].a = (float)(rand()%1000000)/100.0f;
        a[i].b = (float)(rand()%1000000)/100.0f;
        a[i].c = (float)(rand()%1000000)/100.0f;
        a[i].d = (float)(rand()%1000000)/100.0f;
        b[i].a = (float)(rand()%1000000)/100.0f;
        b[i].b = (float)(rand()%1000000)/100.0f;
        b[i].c = (float)(rand()%1000000)/100.0f;
        b[i].d = (float)(rand()%1000000)/100.0f;
    }
}


//SIMD_cout---------------------------------------------------
void SIMD_test(){

double timeSum = 0,timeSub = 0,timeMul = 0,timeDiv = 0;
int j, i;
for (j = 0;j<repeatNumber;j++){
//dodawanie

        for (i = 0;i<sizeTable;i++){
            start=clock();
            sum_simd(a[i],b[i],&sum[i]);
            end=clock();
    timeSum+= ((double) (end - start)) / CLOCKS_PER_SEC*1000.0;
        }

//odejmowanie

        for (i = 0;i<sizeTable;i++){
        start=clock();
            sub_simd(a[i],b[i],&sub[i]);

    timeSub+= ((double)clock() - start)/CLOCKS_PER_SEC*1000.0;
        }


//mnozenie

        for (i = 0;i<sizeTable;i++){
            start=clock();
            mul_simd(a[i],b[i],&mul[i]);
            end=clock();
    timeMul+= ((double) (end - start)) / CLOCKS_PER_SEC*1000.0;
        }

//dzielenie

        for (i = 0;i<sizeTable;i++){
    start=clock();
            div_simd(a[i],b[i],&div_res[i]);
    end=clock();
    timeDiv+= ((double) (end - start)) / CLOCKS_PER_SEC*1000.0;
        }

  }
    timeSum=timeSum/(float)(repeatNumber);
    timeSub=timeSub/(float)(repeatNumber);
    timeMul=timeMul/(float)(repeatNumber);
    timeDiv=timeDiv/(float)(repeatNumber);

printf("Liczba liczb: %d\nSredni czas [ms]: \n+ %lf \n- %lf \n* %lf \n/ %lf \n",(4*sizeTable),timeSum,timeSub,timeMul,timeDiv);
}

int main(void){

srand(time(NULL));
randomGenerator();

printf("Typ obliczen: SIMD\n");
SIMD_test();
printf("Typ obliczen: SISD\n");
SISD();


return 0;
}