Fix in computeMaxGradient kernel of pipeline

llvm/test/VISC/parboil/benchmarks/pipeline/src/visc/main_demo.cc

@@ -541,24 +541,21 @@ void computeMaxGradientLeaf(float *G, size_t bytesG,
   int dimx = __visc__getNumNodeInstances_x(thisNode);
 
   gid = lx + 2*px*dimx;
-
-  if (gid == 0)
-    *maxG = 0;
   
-  for (unsigned stride = blockDim.x; stride > 32; stride >>= 1) {
-    if ((gid + stride < m*n) && (threadIdx.x < stride))
+  for (unsigned stride = dimx; stride > 32; stride >>= 1) {
+    if ((gid + stride < m*n) && (lx < stride))
       if (G[gid + stride] > G[gid])
         G[gid] = G[gid + stride];
-    __syncthreads();
+    __visc__barrier();
   }
 
   for (unsigned stride = 32; stride >= 1; stride >>= 1) {
-    if ((gid + stride < m*n) && (threadIdx.x < stride))
+    if ((gid + stride < m*n) && (lx < stride))
       if (G[gid + stride] > G[gid])
         G[gid] = G[gid + stride];
   }
 
-  if (threadIdx.x == 0)
+  if (lx == 0)
     __visc__atomic_max(maxG,G[gid]);
 
   __visc__return(m);