dubik

const int N_VECTORS = 1<<20;
const int VECTOR_COMPONENTS = 512;
const int KBlockSize = 256;

__global__
void calculateDistances(const float v[VECTOR_COMPONENTS], float *vectors, float *distances)
{
    int vectorIndex = threadIdx.x + threadIdx.y * KBlockSize;
    if (vectorIndex < N_VECTORS) {
        float *vector = vectors + vectorIndex * VECTOR_COMPONENTS;

dubik

// CacheSmth.cpp : Defines the entry point for the console application.
//

#include "stdafx.h"
#include <chrono>
#include <iostream>

struct Timer
{
    Timer()