double randmtzig_randn (void)
{
while (1)
{
//#ifdef __LP64__
#if 1
/* arbitrary mantissa (selected by NRANDI, with 1 bit for sign) */
const randmtzig_uint64_t r = NRANDI;
const randmtzig_int64_t rabs=r>>1;
const int idx = (int)(rabs&0xFF);
const double x = ( r&1 ? -rabs : rabs) * wi[idx];
#else
double x;
int si,idx;
register randmtzig_uint32_t lo, hi;
randmtzig_int64_t rabs;
randmtzig_uint32_t *p = (randmtzig_uint32_t *)&rabs;
lo = dsfmt_gv_genrand_uint32();
idx = lo&0xFF;
hi = dsfmt_gv_genrand_uint32();
si = hi&UMASK;
p[0] = lo;
p[1] = hi&0x1FFFFF;
x = ( si ? -rabs : rabs ) * wi[idx];
# endif
if (rabs < (randmtzig_int64_t)ki[idx])
return x; /* 99.3% of the time we return here 1st try */
else if (idx == 0)
{
/* As stated in Marsaglia and Tsang
*
* For the normal tail, the method of Marsaglia[5] provides:
* generate x = -ln(U_1)/r, y = -ln(U_2), until y+y > x*x,
* then return r+x. Except that r+x is always in the positive
* tail!!!! Any thing random might be used to determine the
* sign, but as we already have r we might as well use it
*
* [PAK] but not the bottom 8 bits, since they are all 0 here!
*/
double xx, yy;
do
{
xx = - ZIGGURAT_NOR_INV_R * log (RANDU);
yy = - log (RANDU);
}
while ( yy+yy <= xx*xx);
return (rabs&0x100 ? -ZIGGURAT_NOR_R-xx : ZIGGURAT_NOR_R+xx);
}
else if ((fi[idx-1] - fi[idx]) * RANDU + fi[idx] < exp(-0.5*x*x))
return x;
}
}
inline static randmtzig_uint64_t randi54 (void)
{
const randmtzig_uint32_t lo = dsfmt_gv_genrand_uint32();
const randmtzig_uint32_t hi = dsfmt_gv_genrand_uint32()&0x3FFFFF;
//#ifndef __LP64__
#if 0
randmtzig_uint64_t u;
randmtzig_uint32_t *p = (randmtzig_uint32_t *)&u;
p[0] = lo;
p[1] = hi;
return u;
#else
return (((randmtzig_uint64_t)hi<<32)|lo);
#endif
}
int main(int ac, char *av[]) {
if (ac == 1) {
printf("Usage: randmtzig <n>\n");
return (-1);
}
int n = atoi(av[1]);
time_t t1;
dsfmt_gv_init_gen_rand(0);
double *p; posix_memalign((void **)&p, 16, n*sizeof(double));
uint32_t *u; posix_memalign((void **)&u, 16, 2*n*sizeof(uint32_t));
t1 = clock();
dsfmt_gv_fill_array_close_open(p, n);
printf("Uniform fill (n): %f\n", (clock() - t1) / (double) CLOCKS_PER_SEC);
t1 = clock();
for (int i = 0; i < n; i++) p[i] = dsfmt_gv_genrand_close_open();
printf("Uniform (n): %f\n", (clock() - t1) / (double) CLOCKS_PER_SEC);
t1 = clock();
for (int i = 0; i < 2*n; i++) u[i] = dsfmt_gv_genrand_uint32();
printf("Uniform 32-bit ints (2*n): %f\n", (clock() - t1) / (double) CLOCKS_PER_SEC);
memset((void *)p, 0, n*sizeof(double));
t1 = clock();
for (int i = 0; i < n; i++) p[i] = randmtzig_gv_randn();
printf("Normal (n): %f\n", (clock() - t1) / (double) CLOCKS_PER_SEC);
for (int i = 0; i < 10; i++) printf("%lf\n", p[i]);
return 0;
}
int main() {
// Initialize RNG
dsfmt_gv_init_gen_rand(0);
double t, tmin;
// fib(20)
assert(fib(20) == 6765);
int f = 0;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
f += fib(20);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("fib", tmin);
// parse_bin
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
char s[11];
for (int k=0; k<1000; ++k) {
uint32_t n = dsfmt_gv_genrand_uint32();
sprintf(s, "%x", n);
uint32_t m = (uint32_t)parse_int(s, 16);
assert(m == n);
}
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("parse_int", tmin);
// // array constructor
// tmin = INFINITY;
// for (int i=0; i<NITER; ++i) {
// t = clock_now();
// double *a = ones(200,200);
// free(a);
// t = clock_now()-t;
// if (t < tmin) tmin = t;
// }
// print_perf("ones", tmin);
//
// // A*A'
// //SUBROUTINE DGEMM(TRANSA,TRANSB,M,N,K,ALPHA,A,LDA,B,LDB,BETA,C,LDC)
// double *b = ones(200, 200);
// tmin = INFINITY;
// for (int i=0; i<NITER; ++i) {
// t = clock_now();
// double *c = matmul_aat(200, b);
// free(c);
// t = clock_now()-t;
// if (t < tmin) tmin = t;
// }
// free(b);
// print_perf("AtA", tmin);
// mandel
int mandel_sum;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
mandel_sum = mandelperf();
t = clock_now()-t;
if (t < tmin) tmin = t;
}
assert(mandel_sum == 14719);
print_perf("mandel", tmin);
// sort
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
double *d = myrand(5000);
quicksort(d, 0, 5000-1);
free(d);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("quicksort", tmin);
// pi sum
double pi;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
pi = pisum();
t = clock_now()-t;
if (t < tmin) tmin = t;
}
assert(fabs(pi-1.644834071848065) < 1e-12);
print_perf("pi_sum", tmin);
// rand mat stat
struct double_pair r;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
r = randmatstat(1000);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
// assert(0.5 < r.s1 && r.s1 < 1.0 && 0.5 < r.s2 && r.s2 < 1.0);
print_perf("rand_mat_stat", tmin);
// rand mat mul
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
double *C = randmatmul(1000);
assert(0 <= C[0]);
free(C);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("rand_mat_mul", tmin);
// printfd
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
printfd(100000);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("printfd", tmin);
return 0;
}
int main() {
// Initialize RNG
dsfmt_gv_init_gen_rand(0);
double t, tmin;
// fib(20)
assert(fib(20) == 6765);
int f = 0;
tmin = INFINITY;
volatile int fibarg = 20; // prevent constant propagation
for (int i=0; i<NITER; ++i) {
t = clock_now();
for (int j = 0; j < 1000; j++)
f += fib(fibarg);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("fib", tmin / 1000);
// parse_bin
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
char s[11];
for (int k=0; k<1000 * 100; ++k) {
uint32_t n = dsfmt_gv_genrand_uint32();
sprintf(s, "%x", n);
uint32_t m = (uint32_t)parse_int(s, 16);
assert(m == n);
}
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("parse_int", tmin / 100);
// // array constructor
// tmin = INFINITY;
// for (int i=0; i<NITER; ++i) {
// t = clock_now();
// double *a = ones(200,200);
// free(a);
// t = clock_now()-t;
// if (t < tmin) tmin = t;
// }
// print_perf("ones", tmin);
//
// // A*A'
// //SUBROUTINE DGEMM(TRANSA,TRANSB,M,N,K,ALPHA,A,LDA,B,LDB,BETA,C,LDC)
// double *b = ones(200, 200);
// tmin = INFINITY;
// for (int i=0; i<NITER; ++i) {
// t = clock_now();
// double *c = matmul_aat(200, b);
// free(c);
// t = clock_now()-t;
// if (t < tmin) tmin = t;
// }
// free(b);
// print_perf("AtA", tmin);
// mandel
/* The initialization on the next line is deliberately volatile to
* prevent gcc from optimizing away the entire loop.
* (First observed in gcc 4.9.2)
*/
static volatile int mandel_sum_init = 0;
int mandel_sum2 = mandel_sum_init;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
int *M;
t = clock_now();
for (int j = 0; j < 100; j++) {
M = mandelperf();
if (j == 0) {
int mandel_sum = 0;
// for (int ii = 0; ii < 21; ii++) {
// for (int jj = 0; jj < 26; jj++) {
// printf("%4d", M[26*ii + jj]);
// }
// printf("\n");
// }
for (int k = 0; k < 21*26; k++) {
mandel_sum += M[k];
}
assert(mandel_sum == 14791);
mandel_sum2 += mandel_sum;
}
free(M);
}
t = clock_now()-t;
if (t < tmin) tmin = t;
}
assert(mandel_sum2 == 14791 * NITER);
print_perf("mandel", tmin / 100);
// sort
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
double *d = myrand(5000);
quicksort(d, 0, 5000-1);
free(d);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("quicksort", tmin);
// pi sum
double pi;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
pi = pisum();
t = clock_now()-t;
if (t < tmin) tmin = t;
}
assert(fabs(pi-1.644834071848065) < 1e-12);
print_perf("pi_sum", tmin);
// rand mat stat
struct double_pair r;
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
r = randmatstat(1000);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
// assert(0.5 < r.s1 && r.s1 < 1.0 && 0.5 < r.s2 && r.s2 < 1.0);
print_perf("rand_mat_stat", tmin);
// rand mat mul
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
double *C = randmatmul(1000);
assert(0 <= C[0]);
free(C);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("rand_mat_mul", tmin);
// printfd
tmin = INFINITY;
for (int i=0; i<NITER; ++i) {
t = clock_now();
printfd(100000);
t = clock_now()-t;
if (t < tmin) tmin = t;
}
print_perf("printfd", tmin);
return 0;
}