diff --git a/alsa_driver.c b/alsa_driver.c
index 67c3627..bbcc1ac 100644
--- a/alsa_driver.c
+++ b/alsa_driver.c
@@ -149,8 +149,13 @@ alsa_driver_setup_io_function_pointers (alsa_driver_t *driver)
 		} else {
 			driver->channel_copy = memcpy_fake;
 		}
-		
-		driver->write_via_copy = sample_move_d16_sS;
+
+		if (driver->dither) {		
+			printf("Rectangular dithering at 16 bits\n");
+			driver->write_via_copy = sample_move_dither_rect_d16_sS;
+		} else {
+			driver->write_via_copy = sample_move_d16_sS;
+		}
 		driver->read_via_copy = sample_move_dS_s16;
 		break;
 
@@ -1243,7 +1248,8 @@ alsa_driver_new (char *name, char *alsa_device,
 		 nframes_t rate,
 		 int hw_monitoring,
 		 int capturing,
-		 int playing)
+		 int playing,
+		 int dither)
 {
 	int err;
 
@@ -1289,6 +1295,8 @@ alsa_driver_new (char *name, char *alsa_device,
 	driver->playback_nfds = 0;
 	driver->capture_nfds = 0;
 
+	driver->dither = dither;
+
 	pthread_mutex_init (&driver->clock_sync_lock, 0);
 	driver->clock_sync_listeners = 0;
 
@@ -1454,6 +1462,7 @@ driver_initialize (int argc, char **argv)
 	int hw_monitoring = FALSE;
 	int capture = FALSE;
 	int playback = FALSE;
+	int dither = FALSE;
 	int i;
 
 	/* grrrr ... getopt() cannot be called in more than one "loop"
@@ -1500,6 +1509,10 @@ driver_initialize (int argc, char **argv)
 			case 'H':
 				hw_monitoring = 1;
 				break;
+
+			case 'z':
+				dither = TRUE;
+				break;
 				
 			default:
 				alsa_usage ();
@@ -1519,7 +1532,8 @@ driver_initialize (int argc, char **argv)
 	}
 
 	return alsa_driver_new ("alsa_pcm", pcm_name, frames_per_interrupt, 
-				user_nperiods, srate, hw_monitoring, capture, playback);
+				user_nperiods, srate, hw_monitoring, capture,
+			        playback, dither);
 }
 
 void
diff --git a/configure.in b/configure.in
index 24ca9db..64c081e 100644
--- a/configure.in
+++ b/configure.in
@@ -4,8 +4,8 @@ AC_INIT(client.c)
 AC_CONFIG_AUX_DIR(.)
 
 JACK_MAJOR_VERSION=0
-JACK_MINOR_VERSION=15
-JACK_MICRO_VERSION=6
+JACK_MINOR_VERSION=16
+JACK_MICRO_VERSION=0
 
 BETA=
 
diff --git a/engine.c b/engine.c
index f7b25f8..8714480 100644
--- a/engine.c
+++ b/engine.c
@@ -1675,9 +1675,9 @@ jack_rechain_graph (jack_engine_t *engine, int take_lock)
 				
 				if (subgraph_client == NULL) {
 					
-				/* start a new subgraph. the engine will start the chain
-				   by writing to the nth FIFO.
-				*/
+				        /* start a new subgraph. the engine will start the chain
+					   by writing to the nth FIFO.
+					*/
 					
 					subgraph_client = client;
 					subgraph_client->subgraph_start_fd = jack_get_fifo_fd (engine, n);
diff --git a/jack/alsa_driver.h b/jack/alsa_driver.h
index c5fddcc..80385b6 100644
--- a/jack/alsa_driver.h
+++ b/jack/alsa_driver.h
@@ -102,6 +102,8 @@ typedef struct {
     WriteCopyFunction write_via_copy;
     CopyCopyFunction channel_copy;
 
+    int dither;
+
     SampleClockMode clock_mode;
     GSList *clock_sync_listeners;
     pthread_mutex_t clock_sync_lock;
diff --git a/jack/memops.h b/jack/memops.h
index c388f50..3c5cd3e 100644
--- a/jack/memops.h
+++ b/jack/memops.h
@@ -26,9 +26,12 @@
 void sample_move_d32u24_sS           (char *dst, sample_t *src, unsigned long nsamples, unsigned long dst_skip);
 void sample_move_d16_sS              (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip);
 
+void sample_move_dither_rect_d16_sS  (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip);
+
 void sample_move_dS_s32u24           (sample_t *dst, char *src, unsigned long nsamples, unsigned long src_skip);
 void sample_move_dS_s16              (sample_t *dst, char *src, unsigned long nsamples, unsigned long src_skip);
 
+
 void sample_merge_d16_sS             (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip);
 void sample_merge_d32u24_sS          (char *dst, sample_t *src, unsigned long nsamples, unsigned long dst_skip);
 
@@ -63,11 +66,3 @@ void merge_memcpy_d16_s16            (char *dst, char *src, unsigned long src_by
 void merge_memcpy_d32_s32            (char *dst, char *src, unsigned long src_bytes, unsigned long foo, unsigned long bar);
 
 #endif /* __jack_memops_h__ */
-
-
-
-
-
-
-
-
diff --git a/memops.c b/memops.c
index e221533..a7fafd1 100644
--- a/memops.c
+++ b/memops.c
@@ -30,6 +30,13 @@
 #define SAMPLE_MAX_24BIT  8388608.0f
 #define SAMPLE_MAX_16BIT  32767.0f
 
+/* Round float to int (can't go in memops.h otherwise it gets multiply
+   defined) */
+inline int f_round(float f) {
+        f += (3<<22);
+        return *((int*)&f) - 0x4b400000;
+}
+
 /* XXX we could use rint(), but for now we'll accept whatever default
    floating-point => int conversion the compiler provides.  
 */
@@ -80,7 +87,30 @@ void sample_move_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsi
 		dst += dst_skip;
 		src++;
 	}
-}	
+}
+
+void sample_move_dither_rect_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip)
+	
+{
+	sample_t val;
+	int      tmp;
+
+	while (nsamples--) {
+		val = *src * (float)SAMPLE_MAX_16BIT;
+		val -= (float)rand() / (float)RAND_MAX;
+		/* swh: This could be some inline asm on x86 */
+		tmp = f_round(val);
+		if (tmp > SHRT_MAX) {
+			*((short *)dst) = SHRT_MAX;
+		} else if (tmp < SHRT_MIN) {
+			*((short *)dst) = SHRT_MIN;
+		} else {
+			*((short *) dst) = (short)tmp;
+		}
+		dst += dst_skip;
+		src++;
+	}
+}
 
 void sample_move_dS_s16 (sample_t *dst, char *src, unsigned long nsamples, unsigned long src_skip)