small improvement

examples_ptx/radixSort/radixSort.cu

@@ -103,11 +103,11 @@ void sortPass(
 #endif
 
 
-  int scatter;
   for (int i = programIndex; i < nloc; i += programCount)
     if (i < nloc)
     {
       const int key = mask & ((unsigned int)keys[i] >> bit);
+      int scatter;
       /* not a vector friendly loop */
 #pragma unroll 1  /* needed, otherwise compiler unroll and optimizes the result :S */
       for (int iv = 0; iv < programCount; iv++)

examples_ptx/radixSort/radixSort.ispc

@@ -196,6 +196,9 @@ export void radixSort_alloc(const uniform int n)
 {
   assert(memoryPool == NULL);
   numBlocks     = num_cores()*4;
+#ifdef __NVPTX__
+  numBlocks     = 13*32*4; //num_cores()*4;
+#endif
   nSharedCounts = NUMDIGITS*numBlocks;
   nCountsGlobal = NUMDIGITS;
   nExcScan      = NUMDIGITS*numBlocks;