#include <iostream>
#include <ctime>
#include <cstdio>
#include <xmmintrin.h>
using namespace std;

#define N (1 << 20)

float sum1 = 0, sum2 = 0, sum3 = 0, sum4 = 0, sum5 = 0, sum6 = 0;
alignas(16) float a[N], b[N];
void init() {
    for (int i = 0; i < N; i++) {
        a[i] = rand() % 4;
        b[i] = rand() % 4;
    }
}

void slow() {
    for (int i = 0; i < N; i++) {
        sum1 += a[i];
    }
    for (int i = 0; i < N; i++) {
        sum1 += b[i];
    }
}

void faster() {
    for (int i = 0; i < N; i++) {
        sum2 += a[i] + b[i];
    }
}

void FASTER() {
    int i, to = N - 1;
    float s0 = 0, s1 = 0;
    for (i = 0; i < to; i += 2) {
        s0 = a[i] + a[i + 1] + s0;
        s1 = b[i] + b[i + 1] + s1;
    }
    while (i < N) sum3 += a[i] + b[i], i++;
    sum3 = s0 + s1;
}

void BIGFUCKINGOPTIMIZATION() {
    __m128 sse_s0 = _mm_setr_ps(0, 0, 0, 0);
    __m128 sse_s1 = _mm_setr_ps(0, 0, 0, 0);
    float s[4];
    for (int i = 0; i < N; i += 8) {
        __m128 sse_a0 = _mm_load_ps(a + i);
        __m128 sse_a1 = _mm_load_ps(a + i + 4);
        __m128 sse_b0 = _mm_load_ps(b + i);
        __m128 sse_b1 = _mm_load_ps(b + i + 4);
         sse_s0 = _mm_add_ps(sse_s0, _mm_add_ps(sse_a0, sse_a1));
         sse_s1 = _mm_add_ps(sse_s1, _mm_add_ps(sse_b0, sse_b1));
    }
    __m128 sse_s = _mm_add_ps(sse_s0, sse_s1);
    _mm_store_ps(s, sse_s);
    sum4 = s[0] + s[1] + s[2] + s[3];
}

void KILLME() {
    __m128 sse_s0 = _mm_setr_ps(0, 0, 0, 0);
    __m128 sse_s1 = _mm_setr_ps(0, 0, 0, 0);
    float s[4];
    for (int i = 0; i < N; i += 16) {
        __m128 sse_a0 = _mm_load_ps(a + i);
        __m128 sse_a1 = _mm_load_ps(a + i + 4);
        __m128 sse_a2 = _mm_load_ps(a + i + 8);
        __m128 sse_a3 = _mm_load_ps(a + i + 12);
        __m128 sse_b0 = _mm_load_ps(b + i);
        __m128 sse_b1 = _mm_load_ps(b + i + 4);
        __m128 sse_b2 = _mm_load_ps(b + i + 8);
        __m128 sse_b3 = _mm_load_ps(b + i + 12);
         sse_s0 = _mm_add_ps(sse_s0, _mm_add_ps(_mm_add_ps(sse_a0, sse_a1), _mm_add_ps(sse_a2, sse_a3)));
         sse_s1 = _mm_add_ps(sse_s1, _mm_add_ps(_mm_add_ps(sse_b0, sse_b1), _mm_add_ps(sse_b2, sse_b3)));
    }
    __m128 sse_s = _mm_add_ps(sse_s0, sse_s1);
    _mm_store_ps(s, sse_s);
    sum5 = s[0] + s[1] + s[2] + s[3];
}

void NIGHTMARE() {
    __m128 sse_s0 = _mm_setr_ps(0, 0, 0, 0);
    __m128 sse_s1 = _mm_setr_ps(0, 0, 0, 0);
    __m128 sse_s2 = _mm_setr_ps(0, 0, 0, 0);
    __m128 sse_s3 = _mm_setr_ps(0, 0, 0, 0);
    float s[4];
    for (int i = 0; i < N; i += 16) {
        __m128 sse_a0 = _mm_load_ps(a + i);
        __m128 sse_a1 = _mm_load_ps(a + i + 4);
        __m128 sse_a2 = _mm_load_ps(a + i + 8);
        __m128 sse_a3 = _mm_load_ps(a + i + 12);
        __m128 sse_b0 = _mm_load_ps(b + i);
        __m128 sse_b1 = _mm_load_ps(b + i + 4);
        __m128 sse_b2 = _mm_load_ps(b + i + 8);
        __m128 sse_b3 = _mm_load_ps(b + i + 12);
         sse_s0 = _mm_add_ps(sse_s0, _mm_add_ps(sse_a0, sse_a1));
         sse_s1 = _mm_add_ps(sse_s1, _mm_add_ps(sse_a2, sse_a3));
         sse_s2 = _mm_add_ps(sse_s2, _mm_add_ps(sse_b0, sse_b1));
         sse_s3 = _mm_add_ps(sse_s3, _mm_add_ps(sse_b2, sse_b3));
    }
    __m128 sse_s = _mm_add_ps(_mm_add_ps(sse_s0, sse_s1), _mm_add_ps(sse_s2, sse_s3));
    _mm_store_ps(s, sse_s);
    sum6 = s[0] + s[1] + s[2] + s[3];
}

int main() {
    srand(time(0));
    init();
    auto sp = clock();
    slow();
    auto fp = clock();
    printf("----Slow sum----\nTime: %.6lf\nResult: %d\n", ((float)fp - sp) / CLOCKS_PER_SEC, (int)sum1);

    sp = clock();
    faster();
    fp = clock();
    printf("----Fast sum----\nTime: %.6lf\nResult: %d\n", ((float)fp - sp) / CLOCKS_PER_SEC, (int)sum2);

    sp = clock();
    FASTER();
    fp = clock();
    printf("----FASTER sum----\nTime: %.6lf\nResult: %d\n", ((float)fp - sp) / CLOCKS_PER_SEC, (int)sum3);

    sp = clock();
    BIGFUCKINGOPTIMIZATION();
    fp = clock();
    printf("----BIGFUCKINGOPTIMIZATION sum----\nTime: %.6lf\nResult: %d\n", ((float)fp - sp) / CLOCKS_PER_SEC, (int)sum4);

    sp = clock();
    KILLME();
    fp = clock();
    printf("----KILLME sum----\nTime: %.6lf\nResult: %d\n", ((float)fp - sp) / CLOCKS_PER_SEC, (int)sum5);

    sp = clock();
    NIGHTMARE();
    fp = clock();
    printf("----NIGHTMARE sum----\nTime: %.6lf\nResult: %d\n", ((float)fp - sp) / CLOCKS_PER_SEC, (int)sum6);
    return 0;
}