diff --git a/examples_cuda/stencil/stencil.cu b/examples_cuda/stencil/stencil.cu
index 1533505b..9240c708 100644
--- a/examples_cuda/stencil/stencil.cu
+++ b/examples_cuda/stencil/stencil.cu
@@ -31,31 +31,31 @@ stencil_step( int x0,  int x1,
             const int x = xb + programIndex;
 
 #endif
-        int index = (z * Nxy) + (y * Nx) + x;
-#define A_cur(x, y, z) Ain[index + (x) + ((y) * Nx) + ((z) * Nxy)]
+            int index = (z * Nxy) + (y * Nx) + x;
+#define A_cur(x, y, z) __ldg(&Ain[index + (x) + ((y) * Nx) + ((z) * Nxy)])
 #define A_next(x, y, z) Aout[index + (x) + ((y) * Nx) + ((z) * Nxy)]
-        double div = 
-            coef0 *  A_cur(0, 0, 0) +
-            coef1 * (A_cur(+1, 0, 0) + A_cur(-1, 0, 0) +
-                      A_cur(0, +1, 0) + A_cur(0, -1, 0) +
-                       A_cur(0, 0, +1) + A_cur(0, 0, -1)) +
-            coef2 * (A_cur(+2, 0, 0) + A_cur(-2, 0, 0) +
-                       A_cur(0, +2, 0) + A_cur(0, -2, 0) +
-                       A_cur(0, 0, +2) + A_cur(0, 0, -2)) +
-            coef3 * (A_cur(+3, 0, 0) + A_cur(-3, 0, 0) +
-                       A_cur(0, +3, 0) + A_cur(0, -3, 0) +
-                       A_cur(0, 0, +3) + A_cur(0, 0, -3));
+            double div = 
+              coef0 *  A_cur(0, 0, 0) +
+              coef1 * (A_cur(+1, 0, 0) + A_cur(-1, 0, 0) +
+                  A_cur(0, +1, 0) + A_cur(0, -1, 0) +
+                  A_cur(0, 0, +1) + A_cur(0, 0, -1)) +
+              coef2 * (A_cur(+2, 0, 0) + A_cur(-2, 0, 0) +
+                  A_cur(0, +2, 0) + A_cur(0, -2, 0) +
+                  A_cur(0, 0, +2) + A_cur(0, 0, -2)) +
+              coef3 * (A_cur(+3, 0, 0) + A_cur(-3, 0, 0) +
+                  A_cur(0, +3, 0) + A_cur(0, -3, 0) +
+                  A_cur(0, 0, +3) + A_cur(0, 0, -3));
 
-        if (x < x1)
-          A_next(0, 0, 0) = 2.0 * A_cur(0, 0, 0) - A_next(0, 0, 0) + 
-            vsq[index] * div;
-    }
+            if (x < x1)
+              A_next(0, 0, 0) = 2.0 * A_cur(0, 0, 0) - A_next(0, 0, 0) + 
+                __ldg(&vsq[index]) * div;
+          }
 }
 
 
 #define SPANX 32
-#define SPANY 8
-#define SPANZ 8
+#define SPANY 2
+#define SPANZ 4
 
 __global__  void
 stencil_step_task( int x0,  int x1,
diff --git a/examples_cuda/stencil/stencil.ispc b/examples_cuda/stencil/stencil.ispc
index 72c28ef6..c4746868 100644
--- a/examples_cuda/stencil/stencil.ispc
+++ b/examples_cuda/stencil/stencil.ispc
@@ -40,8 +40,44 @@ stencil_step(uniform int x0, uniform int x1,
              uniform const double Ain[], uniform double Aout[]) {
     const uniform int Nxy = Nx * Ny;
 
+#if 0
+#define VER1
+#endif
+
+#ifdef VER1
+    const uniform int x1o = 1;
+    const uniform int x2o = 2;
+    const uniform int x3o = 3;
+    const uniform int y1o = Nx;
+    const uniform int y2o = Nx*2;
+    const uniform int y3o = Nx*3;
+    const uniform int z1o = Nxy;
+    const uniform int z2o = Nxy*2;
+    const uniform int z3o = Nxy*3;
+#endif
+
     foreach (z = z0 ... z1, y = y0 ... y1, x = x0 ... x1) 
     {
+#ifdef VER1
+
+      int index = (z * Nxy) + (y * Nx) + x;
+#define A_cur(x, y, z)  Ain [index + (x) + (y) + (z)]
+#define A_next(x, y, z) Aout[index + (x) + (y) + (z)]
+          double div = coef[0] * A_cur(0, 0, 0) +
+            coef[1] * (A_cur(+x1o, 0, 0) + A_cur(-x1o, 0, 0) +
+                A_cur(0, +y1o, 0) + A_cur(0, -y1o, 0) +
+                A_cur(0, 0, +z1o) + A_cur(0, 0, -z1o)) +
+            coef[2] * (A_cur(+x2o, 0, 0) + A_cur(-x2o, 0, 0) +
+                A_cur(0, +y2o, 0) + A_cur(0, -y2o, 0) +
+                A_cur(0, 0, +z2o) + A_cur(0, 0, -z2o)) +
+            coef[3] * (A_cur(+x3o, 0, 0) + A_cur(-x3o, 0, 0) + 
+                A_cur(0, +y3o, 0) + A_cur(0, -y3o, 0) +
+                A_cur(0, 0, +z3o) + A_cur(0, 0, -z3o));
+            A_next(0, 0, 0) = 2.0d0 * A_cur(0, 0, 0) - A_next(0, 0, 0) + 
+              vsq[index] * div;
+
+#else
+
       int index = (z * Nxy) + (y * Nx) + x;
 #define A_cur(x, y, z) Ain[index + (x) + ((y) * Nx) + ((z) * Nxy)]
 #define A_next(x, y, z) Aout[index + (x) + ((y) * Nx) + ((z) * Nxy)]
@@ -59,12 +95,14 @@ stencil_step(uniform int x0, uniform int x1,
       A_next(0, 0, 0) = 2.0 * A_cur(0, 0, 0) - A_next(0, 0, 0) + 
         vsq[index] * div;
 
+#endif
+
     }
 }
 
 #define SPANX 32
-#define SPANY 8
-#define SPANZ 8
+#define SPANY 2
+#define SPANZ 4
 
 static task void
 stencil_step_task(uniform int x0, uniform int x1,
diff --git a/examples_cuda/stencil/stencil_cu.cpp b/examples_cuda/stencil/stencil_cu.cpp
index a4674f59..de5f1854 100644
--- a/examples_cuda/stencil/stencil_cu.cpp
+++ b/examples_cuda/stencil/stencil_cu.cpp
@@ -129,7 +129,7 @@ int main() {
   // the minimum time of three runs.
   //
   double minTimeISPCTasks = 1e30;
-  const bool print_log = false;
+  bool print_log = true;
   const int nreg = 128;
   for (int i = 0; i < 3; ++i) {
     reset_and_start_timer();
@@ -152,6 +152,7 @@ int main() {
       &x0, &x1, &y0, &y1, &z0, &z1, &Nx, &Ny, &Nz,
       &d_coeff, &d_vsq, &d_Aispc0, &d_Aispc1};
     double dt = 1e3*CUDALaunch(NULL, func_name, func_args, print_log, nreg);
+    print_log = false;
     minTimeISPCTasks = std::min(minTimeISPCTasks, dt);
   }
   memcpyD2H(Aispc[1], d_Aispc1, bufsize);