preprocessing bugfix

examples_ptx/mergeSort/mergeSort.cu

@@ -190,6 +190,7 @@ void mergeSortGangKernel(
     s_val[programIndex + programCount] = srcVal[base + programIndex + programCount];
 
 #if 1
+#pragma unroll 1
     for (uniform int stride = 1; stride < 2*programCount; stride <<= 1)
     {
       const int lPos = programIndex & (stride - 1);
@@ -589,7 +590,7 @@ void mergeSort___export(
     // k20m:  2367 M/s
     mergeSortGang(iKey, iVal, srcKey, srcVal, N/(2*programCount));
 
-#if 1
+#if 0
     for (uniform int stride = 2*programCount; stride < N; stride <<= 1)
     {
       const uniform int lastSegmentElements = N % (2 * stride);
@@ -634,8 +635,8 @@ void mergeSort___export(
         oVal = tmpVal;
       }
     }
-  }
 #endif
+  }
 }
 extern "C"
 void mergeSort(

examples_ptx/mergeSort/mergeSort.ispc

@@ -738,6 +738,6 @@ void mergeSort(
         oVal = tmpVal;
       }
     }
-  }
 #endif
+  }
 }