[Audio Threads] Join output device workgroup

On Big Sur or newer, it is possible to join the audio threads to the
same OS workgroup as the audio output device, improving response.

Signed-off-by: Christopher Snowhill <kode54@gmail.com>

Audio/Chain/ConverterNode.mm

@@ -424,6 +424,8 @@ static void convert_be_to_le(uint8_t *buffer, size_t bitsPerSample, size_t bytes
 	// when the end of stream is reached. Convert function instead processes what it can,
 	// and returns 0 samples when it has nothing more to process at the end of stream.
 	while([self shouldContinue] == YES) {
+		[self followWorkgroup];
+
 		int amountConverted;
 		while(paused) {
 			usleep(500);
@@ -435,6 +437,7 @@ static void convert_be_to_le(uint8_t *buffer, size_t bitsPerSample, size_t bytes
 			if(paused) {
 				continue;
 			} else if(streamFormatChanged) {
+				[self leaveWorkgroup];
 				[self cleanUp];
 				[self setupWithInputFormat:newInputFormat withInputConfig:newInputChannelConfig outputFormat:outputFormat outputConfig:outputChannelConfig isLossless:rememberedLossless];
 				continue;

Audio/Chain/InputNode.m

@@ -148,6 +148,7 @@
 
 	while([self shouldContinue] == YES && [self endOfStream] == NO) {
 		if(shouldSeek == YES) {
+			[self leaveWorkgroup];
 			BufferChain *bufferChain = [[controller controller] bufferChain];
 			ConverterNode *converter = [bufferChain converter];
 			DLog(@"SEEKING! Resetting Buffer");
@@ -173,6 +174,8 @@
 		}
 
 		if(amountInBuffer < CHUNK_SIZE) {
+			[self followWorkgroup];
+
 			int framesToRead = CHUNK_SIZE - amountInBuffer;
 			int framesRead;
 			@autoreleasepool {

Audio/Chain/Node.h

@@ -10,6 +10,8 @@
 #import "Semaphore.h"
 #import <Cocoa/Cocoa.h>
 
+#import <os/workgroup.h>
+
 #define BUFFER_SIZE 1024 * 1024
 #define CHUNK_SIZE 16 * 1024
 
@@ -31,6 +33,9 @@
 	AudioStreamBasicDescription nodeFormat;
 	uint32_t nodeChannelConfig;
 	BOOL nodeLossless;
+
+	os_workgroup_t wg;
+	os_workgroup_join_token_s wgToken;
 }
 - (id _Nullable)initWithController:(id _Nonnull)c previous:(id _Nullable)p;
 
@@ -42,6 +47,9 @@
 - (void)process; // Should be overwriten by subclass
 - (void)threadEntry:(id _Nullable)arg;
 
+- (void)followWorkgroup;
+- (void)leaveWorkgroup;
+
 - (void)launchThread;
 
 - (void)setShouldReset:(BOOL)s;

Audio/Chain/Node.m

@@ -11,6 +11,8 @@
 #import "BufferChain.h"
 #import "Logging.h"
 
+#import "OutputCoreAudio.h"
+
 @implementation Node
 
 - (id)initWithController:(id)c previous:(id)p {
@@ -91,7 +93,38 @@
 
 - (void)threadEntry:(id)arg {
 	@autoreleasepool {
+		[self followWorkgroup];
 		[self process];
+		[self leaveWorkgroup];
+	}
+}
+
+- (void)followWorkgroup {
+	if(@available(macOS 11, *)) {
+		if(currentWorkgroup != wg) {
+			if(wg) {
+				os_workgroup_leave(wg, &wgToken);
+			}
+			wg = currentWorkgroup;
+			if(wg) {
+				int result = os_workgroup_join(wg, &wgToken);
+				if(result == 0) return;
+				if(result == EALREADY) {
+					DLog(@"Thread already in workgroup");
+				} else {
+					DLog(@"Cannot join workgroup, error %d", result);
+				}
+			}
+		}
+	}
+}
+
+- (void)leaveWorkgroup {
+	if(@available(macOS 11, *)) {
+		if(wg) {
+			os_workgroup_leave(wg, &wgToken);
+			wg = nil;
+		}
 	}
 }
 

Audio/Output/OutputCoreAudio.h

@@ -33,6 +33,10 @@ using std::atomic_long;
 #import <stdio.h>
 #endif
 
+#import <os/workgroup.h>
+
+extern volatile os_workgroup_t currentWorkgroup;
+
 @class OutputNode;
 
 @interface OutputCoreAudio : NSObject {
@@ -84,6 +88,9 @@ using std::atomic_long;
 
 	VisualizationController *visController;
 
+	os_workgroup_t wg;
+	os_workgroup_join_token_s wgToken;
+
 #ifdef OUTPUT_LOG
 	FILE *_logFile;
 #endif

Audio/Output/OutputCoreAudio.m

@@ -17,6 +17,8 @@
 
 #import <Accelerate/Accelerate.h>
 
+volatile os_workgroup_t currentWorkgroup;
+
 extern void scale_by_volume(float *buffer, size_t count, float volume);
 
 static NSString *CogPlaybackDidBeginNotficiation = @"CogPlaybackDidBeginNotficiation";
@@ -260,6 +262,22 @@ default_device_changed(AudioObjectID inObjectID, UInt32 inNumberAddresses, const
 	atomic_store(&bytesRendered, 0);
 	NSMutableArray *delayedEvents = [[NSMutableArray alloc] init];
 	BOOL delayedEventsPopped = YES;
+
+	if(@available(macOS 11, *)) {
+		if(currentWorkgroup) {
+			wg = currentWorkgroup;
+			int result = os_workgroup_join(wg, &wgToken);
+			if(result != 0) {
+				if(result == EALREADY) {
+					DLog(@"Output thread already in workgroup");
+				} else {
+					DLog(@"Output thread could not be added to workgroup, error = %d", result);
+				}
+				wg = nil;
+			}
+		}
+	}
+
 	while(!stopping) {
 		if(++eventCount == 48) {
 			[self resetIfOutputChanged];
@@ -344,6 +362,14 @@ default_device_changed(AudioObjectID inObjectID, UInt32 inNumberAddresses, const
 		[readSemaphore signal];
 		[writeSemaphore timedWait:5000];
 	}
+
+	if(@available(macOS 11, *)) {
+		if(wg) {
+			os_workgroup_leave(wg, &wgToken);
+			wg = nil;
+		}
+	}
+
 	stopped = YES;
 	if(!stopInvoked)
 		[self stop];
@@ -780,6 +806,10 @@ default_device_changed(AudioObjectID inObjectID, UInt32 inNumberAddresses, const
 
 	visController = [VisualizationController sharedController];
 
+	if(@available(macOS 11, *)) {
+		currentWorkgroup = _au.osWorkgroup;
+	}
+
 	[[NSUserDefaultsController sharedUserDefaultsController] addObserver:self forKeyPath:@"values.outputDevice" options:0 context:NULL];
 	[[NSUserDefaultsController sharedUserDefaultsController] addObserver:self forKeyPath:@"values.GraphicEQenable" options:0 context:NULL];
 	[[NSUserDefaultsController sharedUserDefaultsController] addObserver:self forKeyPath:@"values.eqPreamp" options:(NSKeyValueObservingOptionInitial | NSKeyValueObservingOptionNew) context:NULL];
@@ -809,6 +839,10 @@ default_device_changed(AudioObjectID inObjectID, UInt32 inNumberAddresses, const
 }
 
 - (void)stop {
+	if(@available(macOS 11, *)) {
+		currentWorkgroup = nil;
+	}
+
 	stopInvoked = YES;
 	if(observersapplied) {
 		observersapplied = NO;