diff --git a/examples_ptx/mergeSort/mergeSort.ispc b/examples_ptx/mergeSort/mergeSort.ispc
index cdbd27b4..eb464e4e 100644
--- a/examples_ptx/mergeSort/mergeSort.ispc
+++ b/examples_ptx/mergeSort/mergeSort.ispc
@@ -505,6 +505,12 @@ void mergeElementaryIntervals(
       limitsB,
       stride,
       N);
+  if (lastSegmentElements <= stride)
+    foreach (i = 0 ... lastSegmentElements)
+    {
+      dstKey[N-lastSegmentElements+i] = srcKey[N-lastSegmentElements+i];
+      dstVal[N-lastSegmentElements+i] = srcVal[N-lastSegmentElements+i];
+    }
   sync;
 }
 
@@ -580,8 +586,6 @@ void mergeSort(
 #if 1
     for (uniform int stride = 2*programCount; stride < N; stride <<= 1)
     {
-      const uniform int lastSegmentElements = N % (2 * stride);
-
       // cpu: 30  gpu: 112 M/s
       {
 #if 1
@@ -602,15 +606,6 @@ void mergeSort(
         mergeElementaryIntervals(oKey, oVal, iKey, iVal, limitsA, limitsB, stride, N);
       }
 
-      if (lastSegmentElements <= stride)
-        foreach (i = 0 ... lastSegmentElements)
-        {
-          oKey[N-lastSegmentElements+i] = iKey[N-lastSegmentElements+i];
-          oVal[N-lastSegmentElements+i] = iVal[N-lastSegmentElements+i];
-        }
-      memory_barrier();
-
-
       {
         uniform Key_t * uniform tmpKey = iKey;
         iKey = oKey;
diff --git a/examples_ptx/tasksys.cpp b/examples_ptx/tasksys.cpp
index 1aae61f4..c60c1af2 100644
--- a/examples_ptx/tasksys.cpp
+++ b/examples_ptx/tasksys.cpp
@@ -962,7 +962,7 @@ InitTaskSystem() {
 
 inline void
 TaskGroup::Launch(int baseIndex, int count) {
-#pragma omp parallel for
+#pragma omp parallel for schedule(guided)
     for(int i = 0; i < count; i++) {
         TaskInfo *ti = GetTaskInfo(baseIndex + i);