#include <iostream>
#include <immintrin.h>

int main(int argc, const char *argv[])
{
	alignas(32) float a[8] = {1.0f, 2.0f, 3.0f, 4.0f, 1.5f, 2.5f, 3.5f, 4.5f};
	alignas(32) float b[8] = {5.0f, 6.0f, 7.0f, 8.0f, 5.5f, 6.5f, 7.5f, 8.5f};
	
	//load data in AVX registers
	__m256 vec_a = _mm256_load_ps(a);
	__m256 vec_b = _mm256_load_ps(b);
	
	//multiply elements
	__m256 vec_mul = _mm256_mul_ps(vec_a, vec_b);
	
	//horizontal add to compute the sum of all elements
	__m256 temp = _mm256_hadd_ps(vec_mul, vec_mul);
	temp = _mm256_hadd_ps(temp, temp);
	
	//extract 128 lower bits and sum
	__m128 low = _mm256_castps256_ps128 (temp);
	__m128 high = _mm256_extractf128_ps (temp,1);
	__m128 sum = _mm_add_ps (low, high);
	
	//extract the final result
	float result = _mm_cvtss_f32 (sum);
	
	std::cout << "Dot product: " << result << std::endl;
	
	return 0;
}