Significantly reduce stack memory usage

Oops, there were a lot of large local buffers in use here.

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

Audio/Chain/ChunkList.h

@@ -45,6 +45,9 @@ NS_ASSUME_NONNULL_BEGIN
 
 	uint32_t inputChannelConfig;
 	BOOL inputLossless;
+
+	uint8_t *tempData;
+	size_t tempDataSize;
 }
 
 @property(readonly) double listDuration;

Audio/Chain/ChunkList.m

@@ -408,6 +408,9 @@ static void convert_be_to_le(uint8_t *buffer, size_t bitsPerSample, size_t bytes
 		dsd2pcm = NULL;
 	}
 #endif
+	if(tempData) {
+		free(tempData);
+	}
 }
 
 - (void)reset {
@@ -572,22 +575,24 @@ static void convert_be_to_le(uint8_t *buffer, size_t bitsPerSample, size_t bytes
 	NSData *inputData = [inChunk removeSamples:samplesRead];
 
 #if DSD_DECIMATE
-	const size_t sizeFactor = 2;
+	const size_t sizeFactor = 3;
 #else
-	const size_t sizeFactor = (bitsPerSample == 1) ? 9 : 2;
+	const size_t sizeFactor = (bitsPerSample == 1) ? 9 : 3;
 #endif
-	uint8_t tempData[samplesRead * floatFormat.mBytesPerPacket * sizeFactor + 32]; // Either two buffers plus padding, and/or double precision in case of endian flip
-	
+	size_t newSize = samplesRead * floatFormat.mBytesPerPacket * sizeFactor + 64;
+	if(!tempData || tempDataSize < newSize)
+		tempData = realloc(tempData, tempDataSize = newSize); // Either two buffers plus padding, and/or double precision in case of endian flip
+
 	// double buffer system, with alignment
 	const size_t buffer_adder_base = (samplesRead * floatFormat.mBytesPerPacket + 31) & ~31;
-	
+
 	NSUInteger bytesReadFromInput = samplesRead * inputFormat.mBytesPerPacket;
-	
+
 	uint8_t *inputBuffer = (uint8_t *)[inputData bytes];
 	BOOL inputChanged = NO;
-	
+
 	BOOL hdcdSustained = NO;
-	
+
 	if(bytesReadFromInput && isBigEndian) {
 		// Time for endian swap!
 		memcpy(&tempData[0], [inputData bytes], bytesReadFromInput);
@@ -598,10 +603,11 @@ static void convert_be_to_le(uint8_t *buffer, size_t bitsPerSample, size_t bytes
 
 	if(bytesReadFromInput && isFloat && bitsPerSample == 64) {
 		// Time for precision loss from weird inputs
+		const size_t buffer_adder = (inputBuffer == &tempData[0]) ? buffer_adder_base * 2 : 0;
 		samplesRead = bytesReadFromInput / sizeof(double);
-		convert_f64_to_f32((float *)(&tempData[0]), (const double *)inputBuffer, samplesRead);
+		convert_f64_to_f32((float *)(&tempData[buffer_adder]), (const double *)inputBuffer, samplesRead);
 		bytesReadFromInput = samplesRead * sizeof(float);
-		inputBuffer = (uint8_t *)(&tempData[0]);
+		inputBuffer = &tempData[buffer_adder];
 		inputChanged = YES;
 		bitsPerSample = 32;
 	}

Audio/Chain/Node.m

@@ -211,9 +211,7 @@
 }
 
 - (void)launchThread {
-	NSThread *thread = [[NSThread alloc] initWithTarget:self selector:@selector(threadEntry:) object:nil];
-	[thread setStackSize:1024 * 1024]; // Dammit, this new code makes the nodes overflow the stack size, so let's double the stack
-	[thread start];
+	[NSThread detachNewThreadSelector:@selector(threadEntry:) toTarget:self withObject:nil];
 }
 
 - (void)setPreviousNode:(id)p {

Audio/Output/OutputAVFoundation.h

@@ -117,11 +117,15 @@ using std::atomic_long;
 
 	float *samplePtr;
 	float tempBuffer[512 * 32];
+	float r8bTempBuffer[4096 * 32];
 	float inputBuffer[4096 * 32]; // 4096 samples times maximum supported channel count
 	float fsurroundBuffer[4096 * 6];
 	float hrtfBuffer[4096 * 2];
 	float eqBuffer[4096 * 32];
 
+	float visAudio[4096];
+	float visTemp[8192];
+
 #ifdef OUTPUT_LOG
 	FILE *_logFile;
 #endif

Audio/Output/OutputAVFoundation.m

@@ -63,8 +63,6 @@ static OSStatus eqRenderCallback(void *inRefCon, AudioUnitRenderActionFlags *ioA
 - (int)renderInput:(int)amountToRead toBuffer:(float *)buffer {
 	int amountRead = 0;
 
-	float visAudio[amountToRead]; // Chunk size
-
 	if(stopping == YES || [outputController shouldContinue] == NO) {
 		// Chain is dead, fill out the serial number pointer forever with silence
 		stopping = YES;
@@ -151,7 +149,6 @@ static OSStatus eqRenderCallback(void *inRefCon, AudioUnitRenderActionFlags *ioA
 
 		[visController postSampleRate:44100.0];
 
-		float visTemp[8192];
 		if(newFormat.mSampleRate != 44100.0) {
 			if(newFormat.mSampleRate != lastVisRate) {
 				if(r8bvis) {
@@ -768,8 +765,6 @@ current_device_listener(AudioObjectID inObjectID, UInt32 inNumberAddresses, cons
 		if([self processEndOfStream]) break;
 	} while(inputRendered < 4096);
 
-	float tempBuffer[4096 * 32];
-
 	int samplesRenderedTotal = 0;
 
 	for(size_t i = 0; i < 2;) {
@@ -781,7 +776,7 @@ current_device_listener(AudioObjectID inObjectID, UInt32 inNumberAddresses, cons
 				continue;
 			}
 			[currentPtsLock lock];
-			samplesRendered = r8bstate_flush(r8bold, &tempBuffer[0], 4096);
+			samplesRendered = r8bstate_flush(r8bold, &r8bTempBuffer[0], 4096);
 			[currentPtsLock unlock];
 			if(!samplesRendered) {
 				r8bstate_delete(r8bold);
@@ -790,7 +785,7 @@ current_device_listener(AudioObjectID inObjectID, UInt32 inNumberAddresses, cons
 				++i;
 				continue;
 			}
-			samplePtr = &tempBuffer[0];
+			samplePtr = &r8bTempBuffer[0];
 		} else {
 			samplesRendered = inputRendered;
 			samplePtr = &inputBuffer[0];