remove cleanup thread, add dither code from steve harris

git-svn-id: svn+ssh://jackaudio.org/trunk/jack@145 0c269be4-1314-0410-8aa9-9f06e86f4224

alsa_driver.c

@@ -54,6 +54,11 @@ alsa_driver_release_channel_dependent_memory (alsa_driver_t *driver)
 		free (driver->silent);
 		driver->silent = 0;
 	}
+
+	if (driver->dither_state) {
+		free (driver->dither_state);
+		driver->dither_state = 0;
+	}
 }
 
 static int
@@ -150,12 +155,27 @@ alsa_driver_setup_io_function_pointers (alsa_driver_t *driver)
 			driver->channel_copy = memcpy_fake;
 		}
 
-		if (driver->dither) {		
+		switch (driver->dither) {
+			case Rectangular:
 			printf("Rectangular dithering at 16 bits\n");
 			driver->write_via_copy = sample_move_dither_rect_d16_sS;
-		} else {
+			break;
+
+			case Triangular:
+			printf("Triangular dithering at 16 bits\n");
+			driver->write_via_copy = sample_move_dither_tri_d16_sS;
+			break;
+
+			case Shaped:
+			printf("Noise-shaped dithering at 16 bits\n");
+			driver->write_via_copy = sample_move_dither_shaped_d16_sS;
+			break;
+
+			default:
 			driver->write_via_copy = sample_move_d16_sS;
+			break;
 		}
+
 		driver->read_via_copy = sample_move_dS_s16;
 		break;
 
@@ -463,6 +483,8 @@ alsa_driver_set_parameters (alsa_driver_t *driver, nframes_t frames_per_cycle, n
 		for (chn = 0; chn < driver->playback_nchannels; chn++) {
 			driver->channel_done_bits |= (1<<chn);
 		}
+
+		driver->dither_state = (dither_state_t *) calloc ( driver->playback_nchannels, sizeof (dither_state_t));
 	}
 
 	if (driver->capture_handle) {
@@ -995,14 +1017,17 @@ alsa_driver_process (alsa_driver_t *driver, nframes_t nframes)
 				/* now move data from ports to channels */
 				
 				for (chn = 0, node = driver->playback_ports; node; node = g_slist_next (node), chn++) {
-					
+					sample_t *buf;
+
 					jack_port_t *port = (jack_port_t *) node->data;
 					
 					if (!jack_port_connected (port)) {
 						continue;
 					}
+
+					buf = jack_port_get_buffer (port, contiguous);
 					
-					alsa_driver_write_to_channel (driver, chn, jack_port_get_buffer (port, contiguous), contiguous);
+					alsa_driver_write_to_channel (driver, chn, buf, contiguous);
 				}
 			}
 
@@ -1256,7 +1281,7 @@ alsa_driver_new (char *name, char *alsa_device,
 		 int hw_monitoring,
 		 int capturing,
 		 int playing,
-		 int dither)
+		 DitherAlgorithm dither)
 {
 	int err;
 
@@ -1456,6 +1481,7 @@ alsa PCM driver args:
     -D (duplex, default: yes)
     -C (capture, default: duplex)
     -P (playback, default: duplex)
+    -z[r|t|s|-] (dither, rect|tri|shaped|off, default: off)
 ");
 }
 
@@ -1469,7 +1495,7 @@ driver_initialize (int argc, char **argv)
 	int hw_monitoring = FALSE;
 	int capture = FALSE;
 	int playback = FALSE;
-	int dither = FALSE;
+	DitherAlgorithm dither = None;
 	int i;
 
 	/* grrrr ... getopt() cannot be called in more than one "loop"
@@ -1518,7 +1544,24 @@ driver_initialize (int argc, char **argv)
 				break;
 
 			case 'z':
-				dither = TRUE;
+				switch (argv[i][2]) {
+					case '-':
+					dither = None;
+					break;
+
+					case 'r':
+					dither = Rectangular;
+					break;
+
+					case 's':
+					dither = Shaped;
+					break;
+
+					case 't':
+					default:
+					dither = Triangular;
+					break;
+				}
 				break;
 				
 			default:

client.c

@@ -1203,6 +1203,9 @@ jack_port_get_buffer (jack_port_t *port, nframes_t nframes)
 		   the buffer of the connected (output) port.
 		*/
 
+		printf ("get buffer for port %s is at offset %d\n", port->shared->name,  
+			((jack_port_t *) node->data)->shared->offset);
+
 		return jack_port_buffer (((jack_port_t *) node->data));
 	}
 
@@ -1597,7 +1600,7 @@ jack_audio_port_mixdown (jack_port_t *port, nframes_t nframes)
 	node = port->connections;
 	input = (jack_port_t *) node->data;
 	buffer = jack_port_buffer (port);
-	
+
 	memcpy (buffer, jack_port_buffer (input), sizeof (sample_t) * nframes);
 
 	for (node = g_slist_next (node); node; node = g_slist_next (node)) {

configure.in

@@ -4,7 +4,7 @@ AC_INIT(client.c)
 AC_CONFIG_AUX_DIR(.)
 
 JACK_MAJOR_VERSION=0
-JACK_MINOR_VERSION=19
+JACK_MINOR_VERSION=20
 JACK_MICRO_VERSION=0
 
 BETA=
@@ -31,7 +31,7 @@ JACK_OPT_CFLAGS="-D_REENTRANT -O6 -Wall -fomit-frame-pointer -ffast-math -fstren
 
 AC_ARG_ENABLE(optimize,
      [  --enable-optimize       ask the compiler for its best optimizations.],
-     [ if test "x$enable_optimize" != "xno" ; then JACK_CFLAGS="$JACK_OPT_CFLAGS" ; fi ])
+     [ if test "x$enable_optimize" != "xno" ; then CFLAGS="$JACK_OPT_CFLAGS" ; fi ])
 
 AC_SUBST(JACK_CFLAGS)
 

engine.c

@@ -520,48 +520,6 @@ jack_process (jack_engine_t *engine, nframes_t nframes)
 	return engine->process_errors > 0;
 }
 
-static void *
-jack_cleanup_clients (void *arg)
-{
-	jack_engine_t *engine = (jack_engine_t *) arg;
-	jack_client_internal_t *client;
-	GSList *node, *tmp;
-	int need_sort = FALSE;
-
-	jack_lock_graph (engine);
-
-	/* remove all dead clients */
-
-	for (node = engine->clients; node; ) {
-
-		tmp = g_slist_next (node);
-
-		client = (jack_client_internal_t *) node->data;
-		
-		if (client->error) {
-			
-			if (engine->verbose) {
-				fprintf (stderr, "removing failed client %s (errors: %d)\n", 
-					 client->control->name, client->error);
-			}
-			
-			jack_remove_client (engine, (jack_client_internal_t *) node->data);
-			need_sort = TRUE;
-		}
-
-		node = tmp;
-	}
-
-	if (need_sort) {
-		jack_sort_graph (engine);
-	}
-
-	jack_unlock_graph (engine);
-	pthread_mutex_unlock (&engine->cleanup_lock);
-
-	return NULL;
-}
-
 static int
 jack_engine_post_process (jack_engine_t *engine)
 {
@@ -579,10 +537,6 @@ jack_engine_post_process (jack_engine_t *engine)
 		client = (jack_client_internal_t *) node->data;
 		ctl = client->control;
 		
-		if (engine->verbose) {
-			fprintf (stderr, "client %s state = %s\n", ctl->name, client_state_names[ctl->state]);
-		}
-
 		if (ctl->timed_out || (ctl->state > NotTriggered && ctl->state != Finished)) {
 			client->error++;
 		}
@@ -592,20 +546,40 @@ jack_engine_post_process (jack_engine_t *engine)
 		}
 	}
 
-	jack_unlock_graph (engine);
-
 	if (need_remove) {
 		
-		/* only one thread is allowed to run cleanup at a time. if
-		   one is already underway, don't bother starting another.
-		*/
+		GSList *tmp;
+		int need_sort = FALSE;
+		
+		/* remove all dead clients */
 		
-		if (pthread_mutex_trylock (&engine->cleanup_lock) == 0) {
-			pthread_t cleanup_thread;
-			pthread_create (&cleanup_thread, NULL, jack_cleanup_clients, engine);
+		for (node = engine->clients; node; ) {
+			
+			tmp = g_slist_next (node);
+			
+			client = (jack_client_internal_t *) node->data;
+			
+			if (client->error) {
+				
+				if (engine->verbose) {
+					fprintf (stderr, "removing failed client %s state = %s\n", 
+						 client->control->name, client_state_names[client->control->state]);
+				}
+				
+				jack_remove_client (engine, (jack_client_internal_t *) node->data);
+				need_sort = TRUE;
+			}
+			
+			node = tmp;
+		}
+			
+		if (need_sort) {
+			jack_sort_graph (engine);
 		}
 	}
 
+	jack_unlock_graph (engine);
+
 	return 0;
 }
 
@@ -1152,7 +1126,6 @@ jack_engine_new (int realtime, int rtpriority, int verbose)
 	engine->asio_mode = FALSE;
 
 	pthread_mutex_init (&engine->client_lock, 0);
-	pthread_mutex_init (&engine->cleanup_lock, 0);
 	pthread_mutex_init (&engine->buffer_lock, 0);
 	pthread_mutex_init (&engine->port_lock, 0);
 
@@ -2563,6 +2536,13 @@ jack_port_assign_buffer (jack_engine_t *engine, jack_port_internal_t *port)
 			break;
 		}
 	}
+
+	if (engine->verbose) {
+		fprintf (stderr, "port %s assigned buffer in shm key 0x%x at offset %d\n", 
+			 port->shared->name,
+			 port->shared->shm_key,
+			 port->shared->offset);
+	}
 	
 	if (port->shared->shm_key >= 0) {
 		engine->port_buffer_freelist = g_slist_remove (engine->port_buffer_freelist, bi);

jack/alsa_driver.h

@@ -34,7 +34,8 @@ typedef void (*ReadCopyFunction)  (sample_t *dst, char *src,
 				   unsigned long src_skip_bytes);
 typedef void (*WriteCopyFunction) (char *dst, sample_t *src,
 				   unsigned long src_bytes,
-				   unsigned long dst_skip_bytes);
+				   unsigned long dst_skip_bytes,
+				   dither_state_t *state);
 typedef void (*CopyCopyFunction)  (char *dst, char *src,
 				   unsigned long src_bytes,
 				   unsigned long dst_skip_bytes,
@@ -102,7 +103,8 @@ typedef struct {
     WriteCopyFunction write_via_copy;
     CopyCopyFunction channel_copy;
 
-    int dither;
+    int             dither;
+    dither_state_t *dither_state;
 
     SampleClockMode clock_mode;
     GSList *clock_sync_listeners;
@@ -161,7 +163,8 @@ static __inline__ void alsa_driver_write_to_channel (alsa_driver_t *driver,
 	driver->write_via_copy (driver->playback_addr[channel],
 				buf, 
 				nsamples, 
-				driver->playback_interleave_skip);
+				driver->playback_interleave_skip,
+				driver->dither_state+channel);
 	alsa_driver_mark_channel_done (driver, channel);
 }
 

jack/engine.h

@@ -38,7 +38,6 @@ struct _jack_engine {
     void (*process_unlock)(struct _jack_engine *);
     int  (*post_process)(struct _jack_engine *);
     pthread_mutex_t client_lock;
-    pthread_mutex_t cleanup_lock;
     pthread_mutex_t buffer_lock;
     pthread_mutex_t port_lock;
     int process_errors;

jack/memops.h

@@ -23,17 +23,18 @@
 
 #include <jack/types.h>
 
-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_d32u24_sS           (char *dst, sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state);
+void sample_move_d16_sS              (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state);
 
-void sample_move_dither_rect_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, dither_state_t *state);
+void sample_move_dither_tri_d16_sS  (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state);
+void sample_move_dither_shaped_d16_sS  (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state);
 
 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);
+void sample_merge_d16_sS             (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state);
+void sample_merge_d32u24_sS          (char *dst, sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state);
 
 static __inline__ void
 sample_merge (sample_t *dst, sample_t *src, unsigned long cnt)

jack/types.h

@@ -70,4 +70,21 @@ typedef	enum  {
 	NoSync = 0x8
 } ClockSyncStatus;
 
+typedef	enum  {
+	None,
+	Rectangular,
+	Triangular,
+	Shaped
+} DitherAlgorithm;
+
+#define DITHER_BUF_SIZE 8
+#define DITHER_BUF_MASK 7
+
+typedef struct {
+    unsigned int depth;
+    float rm1;
+    unsigned int idx;
+    float e[DITHER_BUF_SIZE];
+} dither_state_t;
+
 #endif /* __jack_engine_types_h__ */

memops.c

@@ -41,7 +41,7 @@ inline int f_round(float f) {
    floating-point => int conversion the compiler provides.  
 */
 
-void sample_move_d32u24_sS (char *dst, sample_t *src, unsigned long nsamples, unsigned long dst_skip)
+void sample_move_d32u24_sS (char *dst, sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state)
 
 {
 	/* ALERT: signed sign-extension portability !!! */
@@ -64,7 +64,7 @@ void sample_move_dS_s32u24 (sample_t *dst, char *src, unsigned long nsamples, un
 	}
 }	
 
-void sample_move_d16_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, dither_state_t *state)
 	
 {
 	sample_t val;
@@ -89,7 +89,7 @@ void sample_move_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsi
 	}
 }
 
-void sample_move_dither_rect_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, dither_state_t *state)
 	
 {
 	sample_t val;
@@ -112,6 +112,82 @@ void sample_move_dither_rect_d16_sS (char *dst,  sample_t *src, unsigned long ns
 	}
 }
 
+void sample_move_dither_tri_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state)
+	
+{
+	sample_t x;
+	float    r;
+	float    rm1 = state->rm1;
+	int      y;
+
+	while (nsamples--) {
+		x = *src * (float)SAMPLE_MAX_16BIT;
+		r = 2.0f * (float)rand() / (float)RAND_MAX - 1.0f;
+		x += r - rm1;
+		rm1 = r;
+		/* swh: This could be some inline asm on x86 */
+		y = f_round(x);
+
+		if (y > SHRT_MAX) {
+			*((short *)dst) = SHRT_MAX;
+		} else if (y < SHRT_MIN) {
+			*((short *)dst) = SHRT_MIN;
+		} else {
+			*((short *) dst) = (short)y;
+		}
+
+		dst += dst_skip;
+		src++;
+	}
+	state->rm1 = rm1;
+}
+
+void sample_move_dither_shaped_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state)
+	
+{
+	sample_t     x;
+	sample_t     xe; /* the innput sample - filtered error */
+	sample_t     xp; /* x' */
+	float        r;
+	float        rm1 = state->rm1;
+	unsigned int idx = state->idx;
+	int          y;
+
+	while (nsamples--) {
+		x = *src * (float)SAMPLE_MAX_16BIT;
+		r = 2.0f * (float)rand() / (float)RAND_MAX - 1.0f;
+		/* Filter the error with Lipshitz's minimally audible FIR:
+		   [2.033 -2.165 1.959 -1.590 0.6149] */
+		xe = x
+		     - state->e[idx] * 2.033f
+		     + state->e[(idx - 1) & DITHER_BUF_MASK] * 2.165f
+		     - state->e[(idx - 2) & DITHER_BUF_MASK] * 1.959f
+		     + state->e[(idx - 3) & DITHER_BUF_MASK] * 1.590f
+		     - state->e[(idx - 4) & DITHER_BUF_MASK] * 0.6149f;
+		xp = xe + r - rm1;
+		rm1 = r;
+
+		/* This could be some inline asm on x86 */
+		y = f_round(xp);
+
+		/* Intrinsic z^-1 delay */
+		idx = (idx + 1) & DITHER_BUF_MASK;
+		state->e[idx] = y - xe;
+
+		if (y > SHRT_MAX) {
+			*((short *)dst) = SHRT_MAX;
+		} else if (y < SHRT_MIN) {
+			*((short *)dst) = SHRT_MIN;
+		} else {
+			*((short *) dst) = (short)y;
+		}
+		dst += dst_skip;
+		src++;
+	}
+	state->rm1 = rm1;
+	state->idx = idx;
+}
+
 void sample_move_dS_s16 (sample_t *dst, char *src, unsigned long nsamples, unsigned long src_skip) 
 	
 {
@@ -123,7 +199,7 @@ void sample_move_dS_s16 (sample_t *dst, char *src, unsigned long nsamples, unsig
 	}
 }	
 
-void sample_merge_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip)
+void sample_merge_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, unsigned long dst_skip, dither_state_t *state)
 {
 	short val;
 
@@ -144,7 +220,7 @@ void sample_merge_d16_sS (char *dst,  sample_t *src, unsigned long nsamples, uns
 	}
 }	
 
-void sample_merge_d32u24_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, dither_state_t *state)
 
 {
 	/* ALERT: signed sign-extension portability !!! */