Merge pull request #59 from TomHarte/AYFixes

Reduces AY processor load

Components/AY38910/AY38910.cpp

@@ -12,8 +12,8 @@ using namespace GI;
 
 AY38910::AY38910() :
 	_selected_register(0),
-	_channel_output{0, 0, 0}, _channel_dividers{0, 0, 0}, _tone_generator_controls{0, 0, 0},
-	_noise_shift_register(0xffff), _noise_divider(0), _noise_output(0),
+	_tone_counters{0, 0, 0}, _tone_periods{0, 0, 0}, _tone_outputs{0, 0, 0},
+	_noise_shift_register(0xffff), _noise_period(0), _noise_counter(0), _noise_output(0),
 	_envelope_divider(0), _envelope_period(0), _envelope_position(0),
 	_output_registers{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
 {
@@ -82,60 +82,73 @@ void AY38910::set_clock_rate(double clock_rate)
 
 void AY38910::get_samples(unsigned int number_of_samples, int16_t *target)
 {
-	for(int c = 0; c < number_of_samples; c++)
+	int c = 0;
+	while((_master_divider&15) && c < number_of_samples)
 	{
-		// a master divider divides the clock by 16;
-		// resulting_steps will be 1 if a tick occurred, 0 otherwise
-		int former_master_divider = _master_divider;
+		target[c] = _output_volume;
 		_master_divider++;
-		int resulting_steps = ((_master_divider ^ former_master_divider) >> 4) & 1;
-
-		// Bluffer's guide to the stuff below: I wanted to avoid branches. If I avoid branches then
-		// I avoid stalls.
-		//
-		// Repeating patterns are:
-		//	(1) decrement, then shift a high-order bit right and mask to get 1 for did underflow, 0 otherwise;
-		//	(2) did_underflow * a + (did_underflow ^ 1) * b to pick between reloading and not reloading
-		int did_underflow;
-#define shift(x, r, steps) \
-	x -= steps;	\
-	did_underflow = (x >> 16)&1; \
-	x = did_underflow * r + (did_underflow^1) * x;
+		c++;
+	}
 
+	while(c < number_of_samples)
+	{
 #define step_channel(c) \
-	shift(_channel_dividers[c], _tone_generator_controls[c], resulting_steps);	\
-	_channel_output[c] ^= did_underflow;
+	if(_tone_counters[c]) _tone_counters[c]--;\
+	else\
+	{\
+		_tone_outputs[c] ^= 1;\
+		_tone_counters[c] = _tone_periods[c];\
+	}
 
 		// update the tone channels
 		step_channel(0);
 		step_channel(1);
 		step_channel(2);
 
+#undef step_channel
+
 		// ... the noise generator. This recomputes the new bit repeatedly but harmlessly, only shifting
 		// it into the official 17 upon divider underflow.
-		shift(_noise_divider, _output_registers[6]&0x1f, resulting_steps);
-		_noise_output ^= did_underflow&_noise_shift_register&1;
+		if(_noise_counter) _noise_counter--;
+		else
+		{
+			_noise_counter = _noise_period;
+			_noise_output ^= _noise_shift_register&1;
 			_noise_shift_register |= ((_noise_shift_register ^ (_noise_shift_register >> 3))&1) << 17;
-		_noise_shift_register >>= did_underflow;
+			_noise_shift_register >>= 1;
+		}
 
 		// ... and the envelope generator. Table based for pattern lookup, with a 'refill' step — a way of
 		// implementing non-repeating patterns by locking them to table position 0x1f.
-//		int envelope_divider = ((_master_divider ^ former_master_divider) >> 8) & 1;
-		shift(_envelope_divider, _envelope_period, resulting_steps);
-		_envelope_position += did_underflow;
-		int refill = _envelope_overflow_masks[_output_registers[13]] * (_envelope_position >> 5);
-		_envelope_position = (_envelope_position & 0x1f) | refill;
-		int envelope_volume = _envelope_shapes[_output_registers[13]][_envelope_position];
+		if(_envelope_divider) _envelope_divider--;
+		else
+		{
+			_envelope_divider = _envelope_period;
+			_envelope_position ++;
+			if(_envelope_position == 32) _envelope_position = _envelope_overflow_masks[_output_registers[13]];
+		}
 
-#undef step_channel
-#undef shift
+		evaluate_output_volume();
+
+		for(int ic = 0; ic < 16 && c < number_of_samples; ic++)
+		{
+			target[c] = _output_volume;
+			c++;
+			_master_divider++;
+		}
+	}
+}
+
+void AY38910::evaluate_output_volume()
+{
+	int envelope_volume = _envelope_shapes[_output_registers[13]][_envelope_position];
 
 	// The output level for a channel is:
 	//	1 if neither tone nor noise is enabled;
 	//	0 if either tone or noise is enabled and its value is low.
 	// (which is implemented here with reverse logic, assuming _channel_output and _noise_output are already inverted)
 #define level(c, tb, nb)	\
-	(((((_output_registers[7] >> tb)&1)^1) & _channel_output[c]) | ((((_output_registers[7] >> nb)&1)^1) & _noise_output)) ^ 1
+	(((((_output_registers[7] >> tb)&1)^1) & _tone_outputs[c]) | ((((_output_registers[7] >> nb)&1)^1) & _noise_output)) ^ 1
 
 	int channel_levels[3] = {
 		level(0, 0, 3),
@@ -155,20 +168,13 @@ void AY38910::get_samples(unsigned int number_of_samples, int16_t *target)
 	};
 #undef channel_volume
 
-		// Mix additively. TODO: non-linear volume.
-		target[c] = (int16_t)(
+	// Mix additively.
+	_output_volume = (int16_t)(
 		_volumes[volumes[0]] * channel_levels[0] +
 		_volumes[volumes[1]] * channel_levels[1] +
 		_volumes[volumes[2]] * channel_levels[2]
 	);
 }
-}
-
-void AY38910::skip_samples(unsigned int number_of_samples)
-{
-	// TODO
-//	printf("Skip %d\n", number_of_samples);
-}
 
 void AY38910::select_register(uint8_t r)
 {
@@ -186,23 +192,31 @@ void AY38910::set_register_value(uint8_t value)
 			switch(selected_register)
 			{
 				case 0: case 2: case 4:
-					_tone_generator_controls[selected_register >> 1] =
-						(_tone_generator_controls[selected_register >> 1] & ~0xff) | value;
+				case 1: case 3: case 5:
+				{
+					int channel = selected_register >> 1;
+
+					if(selected_register & 1)
+						_tone_periods[channel] = (_tone_periods[channel] & 0xff) | (uint16_t)((value&0xf) << 8);
+					else
+						_tone_periods[channel] = (_tone_periods[channel] & ~0xff) | value;
+					_tone_counters[channel] = _tone_periods[channel];
+				}
 				break;
 
-				case 1: case 3: case 5:
-					_tone_generator_controls[selected_register >> 1] =
-						(_tone_generator_controls[selected_register >> 1] & 0xff) | (uint16_t)((value&0xf) << 8);
+				case 6:
+					_noise_period = value & 0x1f;
+					_noise_counter = _noise_period;
 				break;
 
 				case 11:
 					_envelope_period = (_envelope_period & ~0xff) | value;
-//					printf("e: %d", _envelope_period);
+					_envelope_divider = _envelope_period;
 				break;
 
 				case 12:
 					_envelope_period = (_envelope_period & 0xff) | (int)(value << 8);
-//					printf("e: %d", _envelope_period);
+					_envelope_divider = _envelope_period;
 				break;
 
 				case 13:
@@ -211,6 +225,7 @@ void AY38910::set_register_value(uint8_t value)
 				break;
 			}
 			_output_registers[selected_register] = masked_value;
+			evaluate_output_volume();
 		});
 	}
 }

Components/AY38910/AY38910.hpp

@@ -49,31 +49,30 @@ class AY38910: public ::Outputs::Filter<AY38910> {
 
 		// to satisfy ::Outputs::Speaker (included via ::Outputs::Filter; not for public consumption
 		void get_samples(unsigned int number_of_samples, int16_t *target);
-		void skip_samples(unsigned int number_of_samples);
 
 	private:
 		int _selected_register;
 		uint8_t _registers[16], _output_registers[16];
 
-		int _tone_generator_controls[3];
-		int _channel_dividers[3];
-		int _channel_output[3];
-
-		int _volumes[16];
-
 		int _master_divider;
 
-		int _noise_divider;
+		int _tone_periods[3];
+		int _tone_counters[3];
+		int _tone_outputs[3];
+
+		int _noise_period;
+		int _noise_counter;
 		int _noise_shift_register;
 		int _noise_output;
 
 		int _envelope_period;
 		int _envelope_divider;
-
 		int _envelope_position;
 		int _envelope_shapes[16][32];
 		int _envelope_overflow_masks[16];
 
+		int _volumes[16];
+
 		enum ControlState {
 			Inactive,
 			LatchAddress,
@@ -86,6 +85,9 @@ class AY38910: public ::Outputs::Filter<AY38910> {
 		uint8_t get_register_value();
 
 		uint8_t _data_input, _data_output;
+
+		int16_t _output_volume;
+		inline void evaluate_output_volume();
 };
 
 };

Concurrency/AsyncTaskQueue.cpp

@@ -10,8 +10,14 @@
 
 using namespace Concurrency;
 
-AsyncTaskQueue::AsyncTaskQueue() : should_destruct_(false)
+AsyncTaskQueue::AsyncTaskQueue()
+#ifndef __APPLE__
+	: should_destruct_(false)
+#endif
 {
+#ifdef __APPLE__
+	serial_dispatch_queue_ = dispatch_queue_create("com.thomasharte.clocksignal.asyntaskqueue", DISPATCH_QUEUE_SERIAL);
+#else
 	thread_.reset(new std::thread([this]() {
 		while(!should_destruct_)
 		{
@@ -39,25 +45,37 @@ AsyncTaskQueue::AsyncTaskQueue() : should_destruct_(false)
 			}
 		}
 	}));
+#endif
 }
 
 AsyncTaskQueue::~AsyncTaskQueue()
 {
+#ifdef __APPLE__
+	dispatch_release(serial_dispatch_queue_);
+#else
 	should_destruct_ = true;
 	enqueue([](){});
 	thread_->join();
 	thread_.reset();
+#endif
 }
 
 void AsyncTaskQueue::enqueue(std::function<void(void)> function)
 {
+#ifdef __APPLE__
+	dispatch_async(serial_dispatch_queue_, ^{function();});
+#else
 	std::lock_guard<std::mutex> lock(queue_mutex_);
 	pending_tasks_.push_back(function);
 	processing_condition_.notify_all();
+#endif
 }
 
 void AsyncTaskQueue::flush()
 {
+#ifdef __APPLE__
+	dispatch_sync(serial_dispatch_queue_, ^{});
+#else
 	std::shared_ptr<std::mutex> flush_mutex(new std::mutex);
 	std::shared_ptr<std::condition_variable> flush_condition(new std::condition_variable);
 	std::unique_lock<std::mutex> lock(*flush_mutex);
@@ -66,4 +84,5 @@ void AsyncTaskQueue::flush()
 		flush_condition->notify_all();
 	});
 	flush_condition->wait(lock);
+#endif
 }

Concurrency/AsyncTaskQueue.hpp

@@ -14,6 +14,10 @@
 #include <list>
 #include <condition_variable>
 
+#ifdef __APPLE__
+#include <dispatch/dispatch.h>
+#endif
+
 namespace Concurrency {
 
 /*!
@@ -42,12 +46,16 @@ class AsyncTaskQueue {
 		void flush();
 
 	private:
+#ifdef __APPLE__
+		dispatch_queue_t serial_dispatch_queue_;
+#else
 		std::unique_ptr<std::thread> thread_;
 
 		std::mutex queue_mutex_;
 		std::list<std::function<void(void)>> pending_tasks_;
 		std::condition_variable processing_condition_;
 		std::atomic_bool should_destruct_;
+#endif
 };
 
 }

OSBindings/Mac/Clock Signal/Audio/CSAudioQueue.m

@@ -14,22 +14,13 @@
 @implementation CSAudioQueue
 {
 	AudioQueueRef _audioQueue;
-	size_t _queuedSamples;
-	BOOL _hasHad256;
 }
 
 #pragma mark - AudioQueue callbacks
 
 - (void)audioQueue:(AudioQueueRef)theAudioQueue didCallbackWithBuffer:(AudioQueueBufferRef)buffer
 {
-	size_t samplesInBuffer = (size_t)(buffer->mAudioDataByteSize / sizeof(int16_t));
-	if(_queuedSamples >= 128 && _queuedSamples  - samplesInBuffer < 128 && _hasHad256)
-	{
-		_hasHad256 = NO;
 	[self.delegate audioQueueIsRunningDry:self];
-	}
-	_queuedSamples -= samplesInBuffer;
-
 	AudioQueueFreeBuffer(_audioQueue, buffer);
 }
 
@@ -106,8 +97,6 @@ static void audioOutputCallback(
 {
 	AudioQueueBufferRef newBuffer;
 	size_t bufferBytes = lengthInSamples * sizeof(int16_t);
-	_queuedSamples += lengthInSamples;
-	_hasHad256 |= (_queuedSamples >= 256);
 
 	AudioQueueAllocateBuffer(_audioQueue, (UInt32)bufferBytes, &newBuffer);
 	memcpy(newBuffer->mAudioData, buffer, bufferBytes);