How to compile SFMT

This document explains how to compile SFMT for users who are using UNIX like systems (for example Linux, Free BSD, cygwin, osx, etc) on terminal. I can't help those who use IDE (Integrated Development Environment,) please see your IDE's help to use SIMD feature of your CPU.

1. First Step: Compile test programs using Makefile.

1-1. Compile standard C test program.

Check if SFMT.c and Makefile are in your current directory. If not, cd to the directory where they exist. Then, type

make std

If it causes an error, try to type

cc -DSFMT_MEXP=19937 -o test-std-M19937 test.c SFMT.c

or try to type

gcc -DSFMT_MEXP=19937 -o test-std-M19937 test.c SFMT.c

If success, then check the test program. Type

./test-std-M19937 -b32

You will see many random numbers displayed on your screen. If you want to check these random numbers are correct output, redirect output to a file and diff it with SFMT.19937.out.txt, like this:

./test-std-M19937 -b32 > foo.txt
diff -w foo.txt SFMT.19937.out.txt

Silence means they are the same because diff reports the difference of two file.

If you want to know the generation speed of SFMT, type

./test-std-M19937 -s

It is very slow. To make it fast, compile it with -O3 option. If your compiler is gcc, you should specify -fno-strict-aliasing option with -O3. type

gcc -O3 -fno-strict-aliasing -DSFMT_MEXP=19937 -o test-std-M19937 test.c SFMT.c
./test-std-M19937 -s

1-2. Compile SSE2 test program.

If your CPU supports SSE2 and you can use gcc version 3.4 or later, you can make test-sse2-Mxxx. To do this, type

make sse2

or type

gcc -O3 -msse2 -fno-strict-aliasing -DHAVE_SSE2=1 -DSFMT_MEXP=19937 -o test-sse2-M19937 test.c SFMT.c

If everything works well,

./test-sse2-M19937 -s

will show much shorter time than test-std-M19937 -s.

1-4. Compile and check output automatically.

To make test program and check 32-bit output automatically for all supported MEXPs of SFMT, type

make std-check

To check test program optimized for SSE2, type

make sse2-check

These commands may take some time.

2. Second Step: Use SFMT pseudorandom number generator with your C program.

2-1. Use sequential call and static link.

Here is a very simple program sample1.c which calculates PI using Monte-Carlo method.

#include <stdio.h>
#include <stdlib.h>
#include "SFMT.h"

int main(int argc, char* argv[]) {
    int i, cnt, seed;
    double x, y, pi;
    const int NUM = 10000;
    sfmt_t sfmt;

    if (argc >= 2) {
	seed = strtol(argv[1], NULL, 10);
    } else {
	seed = 12345;
    }
    cnt = 0;
    sfmt_init_gen_rand(&sfmt, seed);
    for (i = 0; i < NUM; i++) {
	x = sfmt_genrand_res53(&sfmt);
	y = sfmt_genrand_res53(&sfmt);
	if (x * x + y * y < 1.0) {
	    cnt++;
	}
    }
    pi = (double)cnt / NUM * 4;
    printf("%lf\n", pi);
    return 0;
}
      

To compile sample1.c with SFMT.c with the period of 2⁶⁰⁷, type

gcc -O3 -DSFMT_MEXP=607 -o sample1 SFMT.c sample1.c

If your CPU supports SSE2 and you want to use optimized SFMT for SSE2, type

gcc -O3 -msse2 -DHAVE_SSE2 -DSFMT_MEXP=607 -o sample1 SFMT.c sample1.c

2-2. Use block call and static link.

Here is sample2.c which modifies sample1.c. The block call fill_array64 is much faster than sequential call, but it needs an aligned memory. The standard function to get an aligned memory is posix_memalign, but it isn't usable in every OS.

#include <stdio.h>
#define _XOPEN_SOURCE 600
#include <stdlib.h>
#include "SFMT.h"

int main(int argc, char* argv[]) {
    int i, j, cnt, seed;
    double x, y, pi;
    const int NUM = 10000;
    const int R_SIZE = 2 * NUM;
    int size;
    uint64_t *array;
    sfmt_t sfmt;

    if (argc >= 2) {
	seed = strtol(argv[1], NULL, 10);
    } else {
	seed = 12345;
    }
    size = sfmt_get_min_array_size64(&sfmt);
    if (size < R_SIZE) {
	size = R_SIZE;
    }
#if defined(__APPLE__) || \
    (defined(__FreeBSD__) && __FreeBSD__ >= 3 && __FreeBSD__ <= 6)
    printf("malloc used\n");
    array = malloc(sizeof(double) * size);
    if (array == NULL) {
	printf("can't allocate memory.\n");
	return 1;
    }
#elif defined(_POSIX_C_SOURCE)
    printf("posix_memalign used\n");
    if (posix_memalign((void **)&array, 16, sizeof(double) * size) != 0) {
	printf("can't allocate memory.\n");
	return 1;
    }
#elif defined(__GNUC__) && (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 3))
    printf("memalign used\n");
    array = memalign(16, sizeof(double) * size);
    if (array == NULL) {
	printf("can't allocate memory.\n");
	return 1;
    }
#else /* in this case, gcc doesn't support SSE2 */
    printf("malloc used\n");
    array = malloc(sizeof(double) * size);
    if (array == NULL) {
	printf("can't allocate memory.\n");
	return 1;
    }
#endif
    cnt = 0;
    j = 0;
    sfmt_init_gen_rand(&sfmt, seed);
    sfmt_fill_array64(&sfmt, array, size);
    for (i = 0; i < NUM; i++) {
	x = sfmt_to_res53(array[j++]);
	y = sfmt_to_res53(array[j++]);
	if (x * x + y * y < 1.0) {
	    cnt++;
	}
    }
    free(array);
    pi = (double)cnt / NUM * 4;
    printf("%lf\n", pi);
    return 0;
}
      

To compile sample2.c with SFMT.c with the period of 2²²⁸¹, type

gcc -O3 -DSFMT_MEXP=2281 -o sample2 SFMT.c sample2.c

If your CPU supports SSE2 and you want to use optimized SFMT for SSE2, type

gcc -O3 -msse2 -DHAVE_SSE2 -DSFMT_MEXP=2281 -o sample2 SFMT.c sample2.c

2-4. Initialize SFMT using sfmt_init_by_array function.

Here is sample4.c which modifies sample1.c. The 32-bit integer seed can only make 2³² kinds of initial state, to avoid this problem, SFMT provides sfmt_init_by_array function. This sample uses sfmt_init_by_array function which initialize the internal state array with an array of 32-bit. The size of an array can be larger than the internal state array and all elements of the array are used for initialization, but too large array is wasteful.

#include <stdio.h>
#include <string.h>
#include "SFMT.h"

int main(int argc, char* argv[]) {
    int i, cnt, seed_cnt;
    double x, y, pi;
    const int NUM = 10000;
    uint32_t seeds[100];
    sfmt_t sfmt;

    if (argc >= 2) {
	seed_cnt = 0;
	for (i = 0; (i < 100) && (i < strlen(argv[1])); i++) {
	    seeds[i] = argv[1][i];
	    seed_cnt++;
	}
    } else {
	seeds[0] = 12345;
	seed_cnt = 1;
    }
    cnt = 0;
    sfmt_init_by_array(&sfmt, seeds, seed_cnt);
    for (i = 0; i < NUM; i++) {
	x = sfmt_genrand_res53(&sfmt);
	y = sfmt_genrand_res53(&sfmt);
	if (x * x + y * y < 1.0) {
	    cnt++;
	}
    }
    pi = (double)cnt / NUM * 4;
    printf("%lf\n", pi);
    return 0;
}
      

To compile sample4.c, type

gcc -O3 -DSFMT_MEXP=19937 -o sample4 SFMT.c sample4.c

Now, seed can be a string. Like this:

./sample4 your-full-name

Appendix: C preprocessor definitions

Here is a list of C preprocessor definitions that users can specify to control code generation. These macros must be set just after -D compiler option.

SFMT_MEXP

This macro is required. This macro means Mersenne exponent and the period of generated code will be 2^SFMT_MEXP-1. SFMT_MEXP must be one of 607, 1279, 2281, 4253, 11213, 19937, 44497, 86243, 132049, 216091.

HAVE_SSE2

This is optional. If this macro is specified, optimized code for SSE2 will be generated.

HAVE_ALTIVEC

This is optional. If this macro is specified, optimized code for AltiVec will be generated. This macro automatically turns on BIG_ENDIAN64 macro. This macro of SFMT ver. 1.4 is not tested at all.

BIG_ENDIAN64

This macro is required when your CPU is BIG ENDIAN and you use 64-bit output. If __BIG_ENDIAN__ macro is defined, this macro is automatically turned on. GCC defines __BIG_ENDIAN__ macro on BIG ENDIAN CPUs. This macro of SFMT ver. 1.4 is not tested at all.

ONLY64

This macro is optional. If this macro is specified, optimized code for 64-bit output for BIG ENDIAN CPUs will be generated and code for 32-bit output won't be generated. BIG_ENDIAN64 macro must be specified with this macro by user or automatically. This macro of SFMT ver. 1.4 is not tested at all.