From 69f2911e04ffb1b19eef1fafb8c040af271f656e Mon Sep 17 00:00:00 2001
From: Tor Aamodt <aamodt@ece.ubc.ca>
Date: Thu, 15 Jul 2010 18:09:46 -0800
Subject: creating branch for adding support for CUDA 3.x and Fermi

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 6829]
---
 src/intersim/rng_double.cpp | 115 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 115 insertions(+)
 create mode 100644 src/intersim/rng_double.cpp

(limited to 'src/intersim/rng_double.cpp')

diff --git a/src/intersim/rng_double.cpp b/src/intersim/rng_double.cpp
new file mode 100644
index 0000000..58a3c2c
--- /dev/null
+++ b/src/intersim/rng_double.cpp
@@ -0,0 +1,115 @@
+/*    This program by D E Knuth is in the public domain and freely copyable
+*    AS LONG AS YOU MAKE ABSOLUTELY NO CHANGES!
+*    It is explained in Seminumerical Algorithms, 3rd edition, Section 3.6
+*    (or in the errata to the 2nd edition --- see
+*        http://www-cs-faculty.stanford.edu/~knuth/taocp.html
+*    in the changes to Volume 2 on pages 171 and following).              */
+
+/*    N.B. The MODIFICATIONS introduced in the 9th printing (2002) are
+      included here; there's no backwards compatibility with the original. */
+
+/*    This version also adopts Brendan McKay's suggestion to
+      accommodate naive users who forget to call ranf_start(seed).         */
+
+/*    If you find any bugs, please report them immediately to
+ *                 taocp@cs.stanford.edu
+ *    (and you will be rewarded if the bug is genuine). Thanks!            */
+
+/************ see the book for explanations and caveats! *******************/
+/************ in particular, you need two's complement arithmetic **********/
+
+#define KK 100                     /* the long lag */
+#define LL  37                     /* the short lag */
+#define mod_sum(x,y) (((x)+(y))-(int)((x)+(y)))   /* (x+y) mod 1.0 */
+
+double ran_u[KK];           /* the generator state */
+
+#ifdef __STDC__
+void ranf_array(double aa[], int n)
+#else
+void ranf_array(aa,n)    /* put n new random fractions in aa */
+double *aa;   /* destination */
+int n;      /* array length (must be at least KK) */
+#endif
+{
+   register int i,j;
+   for (j=0;j<KK;j++) aa[j]=ran_u[j];
+   for (;j<n;j++) aa[j]=mod_sum(aa[j-KK],aa[j-LL]);
+   for (i=0;i<LL;i++,j++) ran_u[i]=mod_sum(aa[j-KK],aa[j-LL]);
+   for (;i<KK;i++,j++) ran_u[i]=mod_sum(aa[j-KK],ran_u[i-LL]);
+}
+
+/* the following routines are adapted from exercise 3.6--15 */
+/* after calling ranf_start, get new randoms by, e.g., "x=ranf_arr_next()" */
+
+#define QUALITY 1009 /* recommended quality level for high-res use */
+double ranf_arr_buf[QUALITY];
+double ranf_arr_dummy=-1.0, ranf_arr_started=-1.0;
+double *ranf_arr_ptr=&ranf_arr_dummy; /* the next random fraction, or -1 */
+
+#define TT  70   /* guaranteed separation between streams */
+#define is_odd(s) ((s)&1)
+
+#ifdef __STDC__
+void ranf_start(long seed)
+#else
+void ranf_start(seed)    /* do this before using ranf_array */
+long seed;            /* selector for different streams */
+#endif
+{
+   register int t,s,j;
+   double u[KK+KK-1];
+   double ulp=(1.0/(1L<<30))/(1L<<22);               /* 2 to the -52 */
+   double ss=2.0*ulp*((seed&0x3fffffff)+2);
+
+   for (j=0;j<KK;j++) {
+      u[j]=ss;                                /* bootstrap the buffer */
+      ss+=ss; if (ss>=1.0) ss-=1.0-2*ulp;  /* cyclic shift of 51 bits */
+   }
+   u[1]+=ulp;                     /* make u[1] (and only u[1]) "odd" */
+   for (s=seed&0x3fffffff,t=TT-1; t; ) {
+      for (j=KK-1;j>0;j--)
+         u[j+j]=u[j],u[j+j-1]=0.0;                         /* "square" */
+      for (j=KK+KK-2;j>=KK;j--) {
+         u[j-(KK-LL)]=mod_sum(u[j-(KK-LL)],u[j]);
+         u[j-KK]=mod_sum(u[j-KK],u[j]);
+      }
+      if (is_odd(s)) {                             /* "multiply by z" */
+         for (j=KK;j>0;j--) u[j]=u[j-1];
+         u[0]=u[KK];                    /* shift the buffer cyclically */
+         u[LL]=mod_sum(u[LL],u[KK]);
+      }
+      if (s) s>>=1;
+      else t--;
+   }
+   for (j=0;j<LL;j++) ran_u[j+KK-LL]=u[j];
+   for (;j<KK;j++) ran_u[j-LL]=u[j];
+   for (j=0;j<10;j++) ranf_array(u,KK+KK-1);  /* warm things up */
+   ranf_arr_ptr=&ranf_arr_started;
+}
+
+#define ranf_arr_next() (*ranf_arr_ptr>=0? *ranf_arr_ptr++: ranf_arr_cycle())
+double ranf_arr_cycle()
+{
+   if (ranf_arr_ptr==&ranf_arr_dummy)
+      ranf_start(314159L); /* the user forgot to initialize */
+   ranf_array(ranf_arr_buf,QUALITY);
+   ranf_arr_buf[100]=-1;
+   ranf_arr_ptr=ranf_arr_buf+1;
+   return ranf_arr_buf[0];
+}
+
+#include <stdio.h>
+int main()
+{
+   register int m; double a[2009]; /* a rudimentary test */
+   ranf_start(310952);
+   for (m=0;m<2009;m++) ranf_array(a,1009);
+   printf("%.20f\n", ran_u[0]);            /* 0.36410514377569680455 */
+   /* beware of buggy printf routines that do not give full accuracy here! */
+   ranf_start(310952);
+   for (m=0;m<1009;m++) ranf_array(a,2009);
+   printf("%.20f\n", ran_u[0]);            /* 0.36410514377569680455 */
+   return 0;
+}
+
-- 
cgit v1.3