Parallel prefix sum added + minor amendements.

examples/sort/sort.cpp

@@ -31,6 +31,7 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
 */
 
+#include <stdio.h>
 #include <stdlib.h>
 #include <algorithm>
 #include <iostream>
@@ -55,10 +56,9 @@ static inline void progressbar (unsigned int x, unsigned int n, unsigned int w =
   if ((x != n) && (x % (n/100) != 0)) return;
 
   using namespace std;
-       
   float ratio  =  x/(float)n;
   int c =  ratio * w;
-	        
+
   cout << setw(3) << (int)(ratio*100) << "% [";
   for (int x=0; x<c; x++) cout << "=";
   for (int x=c; x<w; x++) cout << " ";
@@ -72,10 +72,10 @@ int main (int argc, char *argv[])
   unsigned int *code = new unsigned int [n];
   int *order = new int [n];
 
+  srand (0);
+
   for (i = 0; i < m; i ++)
   {
-    srand (0);
-
     for (j = 0; j < n; j ++) code [j] = random() % l;
 
     reset_and_start_timer();
@@ -86,14 +86,13 @@ int main (int argc, char *argv[])
 
     progressbar (i, m);
   }
-              
 
   printf("[sort ispc]:\t[%.3f] million cycles\n", tISPC1);
 
+  srand (0);
+
   for (i = 0; i < m; i ++)
   {
-    srand (0);
-
     for (j = 0; j < n; j ++) code [j] = random() % l;
 
     reset_and_start_timer();
@@ -107,10 +106,10 @@ int main (int argc, char *argv[])
               
   printf("[sort ispc+tasks]:\t[%.3f] million cycles\n", tISPC2);
 
+  srand (0);
+
   for (i = 0; i < m; i ++)
   {
-    srand (0);
-
     for (j = 0; j < n; j ++) code [j] = random() % l;
 
     reset_and_start_timer();

examples/sort/sort.ispc

@@ -142,12 +142,54 @@ task void unpack (uniform int span, uniform int n, uniform int64 pair[], uniform
   }
 }
 
+task void addup (uniform int h[], uniform int g[])
+{
+  uniform int * uniform u = &h[256*programCount*taskIndex];
+  uniform int i, x, y = 0;
+
+  for (i = 0; i < 256*programCount; i ++)
+  {
+    x = u[i];
+    u[i] = y;
+    y += x;
+  }
+
+  g[taskIndex] = y;
+}
+
+task void bumpup (uniform int h[], uniform int g[])
+{
+  uniform int * uniform u = &h[256*programCount*taskIndex];
+  uniform int z = g[taskIndex];
+
+  foreach (i = 0 ... 256*programCount)
+  {
+    u[i] += z;
+  }
+}
+
+static void prefix_sum (uniform int num, uniform int h[])
+{
+  uniform int * uniform g = uniform new int [num+1];
+  uniform int i;
+
+  launch[num] addup (h, g+1);
+  sync;
+
+  for (g[0] = 0, i = 1; i < num; i ++) g[i] += g[i-1];
+
+  launch[num] bumpup (h, g);
+  sync;
+
+  delete g;
+}
+
 export void sort_ispc (uniform int n, uniform unsigned int code[], uniform int order[], uniform int ntasks)
 {
   uniform int num = ntasks < 1 ? num_cores () : ntasks;
   uniform int span = n / num;
   uniform int hsize = 256*programCount*num;
-  uniform int * uniform hist = uniform new int [hsize+1];
+  uniform int * uniform hist = uniform new int [hsize];
   uniform int64 * uniform pair = uniform new int64 [n];
   uniform int64 * uniform temp = uniform new int64 [n];
   uniform int pass, i;
@@ -166,11 +208,10 @@ export void sort_ispc (uniform int n, uniform unsigned int code[], uniform int o
 
   for (pass = 0; pass < 4; pass ++)
   {
-    launch[num] histogram (span, n, pair, pass, hist+1);
+    launch[num] histogram (span, n, pair, pass, hist);
     sync;
 
-    /* TODO: parallelize and vectorize prefix sum */
-    for (hist[0] = 0, i = 1; i < hsize; i ++) hist[i] += hist[i-1];
+    prefix_sum (num, hist);
 
     launch[num] permutation (span, n, pair, pass, hist, temp);
     sync;