loop_unrolling_example.c

#include <stdio.h>

#ifdef NOINLINE
#define NOINLINE_SYM __attribute__((noinline)) 
#else
#define NOINLINE_SYM
#endif

NOINLINE_SYM int calc_orignal(int n) {
    int fact = 1;
    for (int i = 1; i < n; i++) {
        fact *= i;
    }
    return fact;
}

NOINLINE_SYM int calc_using_pipeline4(int n) {
    int fact0 = 1, fact1 = 1, fact2 = 1, fact3 = 1;
    // 为方便起见，这里假设n是4的倍数
    for (int i = 1; i < n; i += 4) {
        fact0 *= i;
        fact1 *= i + 1;
        fact2 *= i + 2;
        fact3 *= i + 3;
    }
    return fact0 * fact1 * fact2 * fact3;
}

NOINLINE_SYM int calc_using_pipeline8(int n) {
    int fact0 = 1, fact1 = 1, fact2 = 1, fact3 = 1;
    int fact4 = 1, fact5 = 1, fact6 = 1, fact7 = 1;
    for (int i = 1; i < n; i += 8) {
        fact0 *= i;
        fact1 *= i + 1;
        fact2 *= i + 2;
        fact3 *= i + 3;
        fact4 *= i + 4;
        fact5 *= i + 5;
        fact6 *= i + 6;
        fact7 *= i + 7;
    }
    return fact0 * fact1 * fact2 * fact3 * fact4 * fact5 * fact6 * fact7;
}

NOINLINE_SYM int calc_using_fake_pipeline(int n) {
    int fact = 1;
    for (int i = 1; i < n; i += 4) {
        fact *= i;
        fact *= i + 1;
        fact *= i + 2;
        fact *= i + 3;
    }
    return fact;
}

int main(int argc, char const *argv[])
{
    char c = argv[1][0];
    int n = 1000000000;
    int ret;
    if (c == '1') {
        ret = calc_orignal(n);
    } else if (c == '2') {
        ret = calc_using_pipeline4(n);
    } else if (c == '3') {
        ret = calc_using_fake_pipeline(n);
    } else if (c == '4') {
        ret = calc_using_pipeline8(n);
    }
    return ret;
}

// Here is the script that run the test conveniently
/*
#!/bin/bash

set -e

# just for making output prettier
echo_red() {
    echo -e "\033[31m$1\033[0m"
}

echo_green() {
    echo -e "\033[32m$1\033[0m"
}

echo_blue() {
    echo -e "\033[34m$1\033[0m"
}

echo_cyan() {
    echo -e "\033[36m$1\033[0m"
}

build_factorial() {
    build_args="-DNOINLINE"
    echo "---------- using compiler $1"

    f=loop_unrolling_example.c
    $1 $build_args $f -o a.O0
    echo_red "running non-optimized original version ..."
    time ./a.O0 1
    echo_green "running non-optimized using pipeline4 ..."
    time ./a.O0 2
    echo_blue "running non-optimized using fake pipeline ..."
    time ./a.O0 3
    echo_cyan "running non-optimized using pipeline8 ..."
    time ./a.O0 4

    $1 $build_args $f -O3 -o a.O3
    echo_red "running optimized original version ..."
    time ./a.O3 1
    echo_green "running optimized using pipeline4 ..."
    time ./a.O3 2
    echo_blue "running optimized using fake pipeline ..."
    time ./a.O3 3
    echo_cyan "running optimized using pipeline8 ..."
    time ./a.O3 4
}

CC=$1
build_factorial $CC
*/