Fix out-of-bounds atomic writes in spmv fixup kernel.

This fix was suggested in PR NVIDIA#160.

cub/agent/agent_spmv_orig.cuh

@@ -647,9 +647,24 @@ struct AgentSpmv
         if (threadIdx.x == 0)
         {
             if (HAS_ALPHA)
+            {
                 tile_carry.value *= spmv_params.alpha;
+            }
 
             tile_carry.key += tile_start_coord.x;
+            if (tile_carry.key >= spmv_params.num_rows)
+            {
+                // Prevent the AgentSegmentFixup kernel from updating
+                // out-of-bounds values.
+                // FIXME This creates several no-op atomicAdd(..., 0) calls.
+                // Ideally the fixup kernel would be aware of the keys. This
+                // should be possible as the fixup kernel is only used by spmv.
+                // This should give a perf boost by eliminating these
+                // unneccessary global atomic updates.
+                tile_carry.key = spmv_params.num_rows - 1;
+                tile_carry.value = ValueT{};
+            };
+
             d_tile_carry_pairs[tile_idx]    = tile_carry;
         }
     }