Support 8 Dexed instances (TGs) with multi-core

* Core 1 kicks core 2 and 3 and processes two TGs then
* Cores 2 and 3 wait for a kick and process three TGs each then
* When all cores are ready, the output will be summed up
* All 8 TGs generate the same sound at the moment
* The maximum chunk size is limited to 4096 now

src/config.h

@@ -25,19 +25,28 @@
 
 #include <fatfs/ff.h>
 #include <Properties/propertiesfatfsfile.h>
+#include <circle/sysconfig.h>
 #include <string>
 
 class CConfig		// Configuration for MiniDexed
 {
 public:
-#if RASPPI == 1
+#ifndef ARM_ALLOW_MULTI_CORE
 	static const unsigned ToneGenerators = 1;
+#else
+	static const unsigned TGsCore1 = 2;		// process 2 TGs on core 1
+	static const unsigned TGsCore23 = 3;		// process 3 TGs on core 2 and 3 each
+	static const unsigned ToneGenerators = TGsCore1 + 2*TGsCore23;
+#endif
+
+#if RASPPI == 1
 	static const unsigned MaxNotes = 8;		// polyphony
 #else
-	static const unsigned ToneGenerators = 2;
-	static const unsigned MaxNotes = 16;		// polyphony
+	static const unsigned MaxNotes = 16;
 #endif
 
+	static const unsigned MaxChunkSize = 4096;
+
 #if RASPPI <= 3
 	static const unsigned MaxUSBMIDIDevices = 2;
 #else

src/minidexed.cpp

@@ -85,6 +85,13 @@ CMiniDexed::CMiniDexed (CConfig *pConfig, CInterruptSystem *pInterrupt,
 		m_pSoundDevice = new CPWMSoundBaseDevice (pInterrupt, pConfig->GetSampleRate (),
 							  pConfig->GetChunkSize ());
 	}
+
+#ifdef ARM_ALLOW_MULTI_CORE
+	for (unsigned nCore = 0; nCore < CORES; nCore++)
+	{
+		m_CoreStatus[nCore] = CoreStatusInit;
+	}
+#endif
 };
 
 bool CMiniDexed::Initialize (void)
@@ -175,13 +182,58 @@ void CMiniDexed::Process (bool bPlugAndPlayUpdated)
 
 void CMiniDexed::Run (unsigned nCore)
 {
+	assert (1 <= nCore && nCore < CORES);
+
 	if (nCore == 1)
 	{
-		while (1)
+		m_CoreStatus[nCore] = CoreStatusIdle;			// core 1 ready
+
+		// wait for cores 2 and 3 to be ready
+		for (unsigned nCore = 2; nCore < CORES; nCore++)
+		{
+			while (m_CoreStatus[nCore] != CoreStatusIdle)
+			{
+				// just wait
+			}
+		}
+
+		while (m_CoreStatus[nCore] != CoreStatusExit)
 		{
 			ProcessSound ();
 		}
 	}
+	else								// core 2 and 3
+	{
+		while (1)
+		{
+			m_CoreStatus[nCore] = CoreStatusIdle;		// ready to be kicked
+			while (m_CoreStatus[nCore] == CoreStatusIdle)
+			{
+				// just wait
+			}
+
+			// now kicked from core 1
+
+			if (m_CoreStatus[nCore] == CoreStatusExit)
+			{
+				m_CoreStatus[nCore] = CoreStatusUnknown;
+
+				break;
+			}
+
+			assert (m_CoreStatus[nCore] == CoreStatusBusy);
+
+			// process the TGs, assigned to this core (2 or 3)
+
+			assert (m_nFramesToProcess <= CConfig::MaxChunkSize);
+			unsigned nTG = CConfig::TGsCore1 + (nCore-2)*CConfig::TGsCore23;
+			for (unsigned i = 0; i < CConfig::TGsCore23; i++, nTG++)
+			{
+				assert (m_pTG[nTG]);
+				m_pTG[nTG]->getSamples (m_nFramesToProcess, m_OutputLevel[nTG]);
+			}
+		}
+	}
 }
 
 #endif
@@ -303,6 +355,8 @@ std::string CMiniDexed::GetVoiceName (unsigned nTG)
 	return Result;
 }
 
+#ifndef ARM_ALLOW_MULTI_CORE
+
 void CMiniDexed::ProcessSound (void)
 {
 	assert (m_pSoundDevice);
@@ -316,33 +370,77 @@ void CMiniDexed::ProcessSound (void)
 		}
 
 		int16_t SampleBuffer[nFrames];
+		m_pTG[0]->getSamples (nFrames, SampleBuffer);
+
+		if (   m_pSoundDevice->Write (SampleBuffer, sizeof SampleBuffer)
+		    != (int) sizeof SampleBuffer)
+		{
+			LOGERR ("Sound data dropped");
+		}
 
-#if RASPPI > 1
-		for (unsigned i = 0; i < CConfig::ToneGenerators; i++)
+		if (m_bProfileEnabled)
 		{
-			int16_t TempBuffer[nFrames];
+			m_GetChunkTimer.Stop ();
+		}
+	}
+}
+
+#else	// #ifdef ARM_ALLOW_MULTI_CORE
 
+void CMiniDexed::ProcessSound (void)
+{
+	assert (m_pSoundDevice);
+
+	unsigned nFrames = m_nQueueSizeFrames - m_pSoundDevice->GetQueueFramesAvail ();
+	if (nFrames >= m_nQueueSizeFrames/2)
+	{
+		if (m_bProfileEnabled)
+		{
+			m_GetChunkTimer.Start ();
+		}
+
+		m_nFramesToProcess = nFrames;
+
+		// kick secondary cores
+		for (unsigned nCore = 2; nCore < CORES; nCore++)
+		{
+			assert (m_CoreStatus[nCore] == CoreStatusIdle);
+			m_CoreStatus[nCore] = CoreStatusBusy;
+		}
+
+		// process the TGs assigned to core 1
+		assert (nFrames <= CConfig::MaxChunkSize);
+		for (unsigned i = 0; i < CConfig::TGsCore1; i++)
+		{
 			assert (m_pTG[i]);
-			m_pTG[i]->getSamples (nFrames, TempBuffer);
+			m_pTG[i]->getSamples (nFrames, m_OutputLevel[i]);
+		}
 
-			if (i == 0)
-			{
-				for (unsigned j = 0; j < nFrames; j++)
-				{
-					SampleBuffer[j] = TempBuffer[j] / CConfig::ToneGenerators;
-				}
-			}
-			else
+		// wait for cores 2 and 3 to complete their work
+		for (unsigned nCore = 2; nCore < CORES; nCore++)
+		{
+			while (m_CoreStatus[nCore] != CoreStatusIdle)
 			{
-				for (unsigned j = 0; j < nFrames; j++)
-				{
-					SampleBuffer[j] += TempBuffer[j] / CConfig::ToneGenerators;
-				}
+				// just wait
 			}
 		}
-#else
-		m_pTG[0]->getSamples (nFrames, SampleBuffer);
-#endif
+
+		// now mix the output of all TGs
+		int16_t SampleBuffer[nFrames];
+		assert (CConfig::ToneGenerators == 8);
+		for (unsigned i = 0; i < nFrames; i++)
+		{
+			int32_t nSample =   m_OutputLevel[0][i]
+					  + m_OutputLevel[1][i]
+					  + m_OutputLevel[2][i]
+					  + m_OutputLevel[3][i]
+					  + m_OutputLevel[4][i]
+					  + m_OutputLevel[5][i]
+					  + m_OutputLevel[6][i]
+					  + m_OutputLevel[7][i];
+
+			SampleBuffer[i] = (int16_t) (nSample / CConfig::ToneGenerators);
+		}
 
 		if (   m_pSoundDevice->Write (SampleBuffer, sizeof SampleBuffer)
 		    != (int) sizeof SampleBuffer)
@@ -356,3 +454,5 @@ void CMiniDexed::ProcessSound (void)
 		}
 	}
 }
+
+#endif

src/minidexed.h

@@ -73,6 +73,17 @@ public:
 private:
 	void ProcessSound (void);
 
+#ifdef ARM_ALLOW_MULTI_CORE
+	enum TCoreStatus
+	{
+		CoreStatusInit,
+		CoreStatusIdle,
+		CoreStatusBusy,
+		CoreStatusExit,
+		CoreStatusUnknown
+	};
+#endif
+
 private:
 	CConfig *m_pConfig;
 
@@ -89,6 +100,12 @@ private:
 	CSoundBaseDevice *m_pSoundDevice;
 	unsigned m_nQueueSizeFrames;
 
+#ifdef ARM_ALLOW_MULTI_CORE
+	volatile TCoreStatus m_CoreStatus[CORES];
+	volatile unsigned m_nFramesToProcess;
+	int16_t m_OutputLevel[CConfig::ToneGenerators][CConfig::MaxChunkSize];
+#endif
+
 	CPerformanceTimer m_GetChunkTimer;
 	bool m_bProfileEnabled;
 };