loop-optimization.c

// gcc -Wall -std=c99 -Ofast -march=native loop-optimization.c -o loop
// http://www.lshift.net/blog/2013/11/27/optimizing-loops-in-c-for-higher-numerical-throughput-and-for-fun

#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>

#define COLS (16)
#define ROWS (1000)
#define COUNT (100000)

float globalSrc[ROWS][COLS];
float globalDst[ROWS];


// sum[0] = sum(vec[0][0..COLS-1]^2)
// sum[1] = sum(vec[1][0..COLS-1]^2)
// sum[2] = sum(vec[2][0..COLS-1]^2)
// ...
// sum[ROWS] = sum(vec[ROWS][0..COLS-1]^2)
void sum_of_squares_0(float dst[ROWS], float src[ROWS][COLS]) {
    for (size_t x = 0; x < COLS; x++) {
        for (size_t y = 0; y < ROWS; y++) {
            dst[y] += src[y][x] * src[y][x];
        }
    }
}


void sum_of_squares_1(float dst[ROWS], float src[ROWS][COLS]) {
    for (size_t y = 0; y < ROWS; y++) {
        for (size_t x = 0; x < COLS; x++) {
            dst[y] += src[y][x] * src[y][x];
        }
    }
}

// total = vec[0..ROWS]
double sum_of_sums(float sum[ROWS]) {
    double total = 0.0;
    for (int y = 0; y < ROWS; y++) {
        //        printf("%d: %f\n", y, sum[y]);
        total += sum[y];
    }
    return total;
}

// vec[0..COLS] = [0..COLS]
void init_vector(float vec[COLS]) {
    for (int i = 0; i < COLS; i++) {
        vec[i] = (float) i;
    }
}

typedef void (*sum_of_squares_t)(float dst[ROWS], float src[ROWS][COLS]);

int main(int argc, char **argv) {

    if (argc <= 1) {
        printf("Usage: %s num\n", argv[0]);
        return 1;
    }

    sum_of_squares_t sum_func;
    if (*argv[1] == '0') sum_func = sum_of_squares_0;
    else if (*argv[1] == '1') sum_func = sum_of_squares_1;
    else {
        printf("Function '%s' not recognized.\n", argv[1]);
        return 1;
    }


    for (size_t y = 0; y < ROWS; y++) {
        init_vector(&globalSrc[y][0]);
    }

    for (size_t n = 0; n < COUNT; n++) {
        memset(globalDst, 0, ROWS * sizeof(*globalDst));
        sum_func(globalDst, globalSrc);
    }

    double total = sum_of_sums(globalDst);
    printf("Total: %lf (%d iterations)\n", total, COUNT);

    return 0;
}