diff --git a/builtins/target-nvptx64.ll b/builtins/target-nvptx64.ll
index 297b9779..85b5ead2 100644
--- a/builtins/target-nvptx64.ll
+++ b/builtins/target-nvptx64.ll
@@ -397,19 +397,19 @@ define  i64 @__movmsk(<1 x i1>) nounwind readnone alwaysinline {
 
 define  i1 @__any(<1 x i1>) nounwind readnone alwaysinline {
   %v = extractelement <1 x i1> %0, i32 0
-  %cmp = icmp ne i1 %v, 0
-  ret i1 %cmp
+;   %cmp = icmp ne i1 %v, 0
+  ret i1 %v
 }
 
 define  i1 @__all(<1 x i1>) nounwind readnone alwaysinline {
   %v = extractelement <1 x i1> %0, i32 0
-  %cmp = icmp eq i1 %v, 1
-  ret i1 %cmp
+;  %cmp = icmp eq i1 %v, 1
+  ret i1 %v
 }
 
 define  i1 @__none(<1 x i1>) nounwind readnone alwaysinline {
   %v = extractelement <1 x i1> %0, i32 0
-  %cmp = icmp eq i1 %v, 0
+  %cmp = xor i1 %v,  1; ;icmp eq i1 %v, 0
   ret i1 %cmp
 }
 
diff --git a/examples_cuda/rt/rt.cu b/examples_cuda/rt/rt.cu
index 5857a873..2575662d 100644
--- a/examples_cuda/rt/rt.cu
+++ b/examples_cuda/rt/rt.cu
@@ -249,8 +249,6 @@ static inline bool BVHIntersect(const  LinearBVHNode nodes[],
                     if (TriIntersect(tris[primitivesOffset+i], ray))
                         hit = true;
                 }
-                if (todoOffset == 0) 
-                    break;
                 nodeNum = todo[--todoOffset];
             }
             else {
@@ -275,10 +273,10 @@ static inline bool BVHIntersect(const  LinearBVHNode nodes[],
             }
         }
         else {
-            if (todoOffset == 0)
-                break;
             nodeNum = todo[--todoOffset];
         }
+            if (todoOffset == 0)
+                break;
     }
     r.maxt = ray.maxt;
     r.hitId = ray.hitId;
diff --git a/examples_cuda/rt/rt.ispc b/examples_cuda/rt/rt.ispc
index 602c92fa..230a5288 100644
--- a/examples_cuda/rt/rt.ispc
+++ b/examples_cuda/rt/rt.ispc
@@ -219,6 +219,8 @@ inline inline bool BVHIntersect(const uniform LinearBVHNode nodes[],
     uniform int todoOffset = 0, nodeNum = 0;
     uniform int todo[64];
 
+#define FAST
+
     while (true) {
         // Check ray against BVH node
         uniform LinearBVHNode node = nodes[nodeNum];
@@ -233,8 +235,10 @@ inline inline bool BVHIntersect(const uniform LinearBVHNode nodes[],
                     if (TriIntersect(tris[primitivesOffset+i], ray))
                         hit = true;
                 }
+#ifndef FAST
                 if (todoOffset == 0) 
                     break;
+#endif
                 nodeNum = todo[--todoOffset];
             }
             else 
@@ -261,10 +265,16 @@ inline inline bool BVHIntersect(const uniform LinearBVHNode nodes[],
             }
         }
         else {
+#ifndef FAST
             if (todoOffset == 0)
                 break;
+#endif
             nodeNum = todo[--todoOffset];
         }
+#ifdef FAST
+        if (todoOffset == 0)
+          break;
+#endif
     }
     r.maxt = ray.maxt;
     r.hitId = ray.hitId;