Fix an off-by-one error in selecting the DP word to use to demux

dpstream.cu

@@ -88,9 +88,9 @@ extern bool dpcallback(void *data, unsigned short threadId,
     unsigned int b_0, unsigned int b_1, unsigned int b_2);
 
 #if WORDS > 1
-#define DEMUXWORD 1
+#define DEMUXWORD 2
 #else
-#define DEMUXWORD 0
+#define DEMUXWORD 1
 #endif
 
 // Return true if we should stop computing