Simplify the output part of the kernel. There's no demonstrable advantage from more complexity.

phatk110722.cl

@@ -394,26 +394,26 @@ __kernel void search(	const uint state0, const uint state1, const uint state2, c
 #if defined(VECTORS4) || defined(VECTORS2)
 	if (Vals[7].x == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & (W[3].x >> 2)] =  W_3.x;
+		output[MAXBUFFERS] = output[NFLAG & W_3.x] =  W_3.x;
 	}
 	if (Vals[7].y == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & (W[3].y >> 2)] =  W_3.y;
+		output[MAXBUFFERS] = output[NFLAG & W_3.y] =  W_3.y;
 	}
 #ifdef VECTORS4
 	if (Vals[7].z == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & (W[3].z >> 2)] =  W_3.z;
+		output[MAXBUFFERS] = output[NFLAG & W_3.z] =  W_3.z;
 	}
 	if (Vals[7].w == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & (W[3].w >> 2)] =  W_3.w;
+		output[MAXBUFFERS] = output[NFLAG & W_3.w] =  W_3.w;
 	}
 #endif
 #else
 	if (Vals[7] == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & (W[3] >> 2)] =  W_3;
+		output[MAXBUFFERS] = output[NFLAG & W_3] =  W_3;
 	}
 #endif
 }