1541 Drive speed improvements

Specialized drive read loops for the most common cases, triggers a vast speed improvement in some scenarios. Improved saving speed. Removed some more code that's not really required right now.

src/Drive.cpp

@@ -358,6 +358,7 @@ void Drive::Reset()
 #if defined(EXPERIMENTALZERO)
 	LED = false;
 	cyclesForBit = 0;
+	UE7Counter = 16;
 #endif
 	headTrackPos = 18*2;		// Start with the head over track 19 (Very later Vorpal ie Cakifornia Games) need to have had the last head movement -ve
 	CLOCK_SEL_AB = 3;		// Track 18 will use speed zone 3 (encoder/decoder (ie UE7Counter) clocked at 1.2307Mhz)
@@ -568,20 +569,71 @@ bool Drive::Update()
 #define min(a,b) (((a) < (b)) ? (a) : (b))
 #define max(a,b) (((a) > (b)) ? (a) : (b))
 
-void Drive::DriveLoopRead()
+void Drive::DriveLoopReadNoFluxNoCycles()
+{
+	unsigned int cycles = 16;
+	fluxReversalCyclesLeft -= 16;
+	cyclesLeftForBit -= 16;
+	while (true)
+	{
+		if (cycles < UE7Counter)
+		{
+			UE7Counter -= cycles;
+			cycles = 0;
+			return;
+		}
+
+		cycles -= UE7Counter;
+
+		UE7Counter = 16 - CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
+									// The decoder consists of UF4 and UE5A. The ecoder has two outputs, Pin 1 of UE5A is the serial data output and pin 2 of UF4 (output B) is the serial clock output.
+		++UF4Counter &= 0xf; // Clock and clamp UF4.
+								// The UD2 read shift register is clocked by serial clock (the rising edge of encoder/decoder's UF4 B output (serial clock))
+								//	- ie on counts 2, 6, 10 and 14 (2 is the only count that outputs a 1 into readShiftRegister as the MSB bits of the count NORed together for other values are 0)
+		if ((UF4Counter & 0x3) == 2)
+		{
+			readShiftRegister <<= 1;
+			readShiftRegister |= (UF4Counter == 2); // Emulate UE5A and only shift in a 1 when pins 6 (output C) and 7 (output D) (bits 2 and 3 of UF4Counter are 0. ie the first count of the bit cell)
+
+			writeShiftRegister <<= 1;
+			// Note: SYNC can only trigger during reading as R/!W line is one of UC2's inputs.
+			if (((readShiftRegister & 0x3ff) == 0x3ff))	// if the last 10 bits are 1s then SYNC
+			{
+				UE3Counter = 0;	// Phase lock on to byte boundary
+				m_pVIA->GetPortB()->SetInput(0x80, false);			// PB7 active low SYNC
+			}
+			else
+			{
+				m_pVIA->GetPortB()->SetInput(0x80, true); // SYNC not asserted if not following the SYNC bits
+				UE3Counter++;
+			}
+		}
+		// UC5B (NOR used to invert UF4's output B serial clock) output high when UF4 counts 0,1,4,5,8,9,12 and 13
+		else if (((UF4Counter & 2) == 0) && (UE3Counter == 8))	// Phase locked on to byte boundary
+		{
+			UE3Counter = 0;
+			SO = (m_pVIA->GetFCR() & m6522::FCR_CA2_OUTPUT_MODE0) != 0;	// bit 2 of the FCR indicates "Byte Ready Active" turned on or not.
+			writeShiftRegister = (u8)(readShiftRegister & 0xff);
+			m_pVIA->GetPortA()->SetInput(writeShiftRegister);
+		}
+	};
+}
+
+void Drive::DriveLoopReadNoFlux()
 {
 	unsigned int minCycles;
-	unsigned int cycles = 0;
+	unsigned int cycles = 16;
+
+	fluxReversalCyclesLeft -= 16;
 
 	while (true)
 	{
-		minCycles = min(min(cyclesLeftForBit, fluxReversalCyclesLeft), 16 - max(UE7Counter, cycles));
+		minCycles = min(min(cyclesLeftForBit, cycles), UE7Counter);
 		cyclesLeftForBit -= minCycles;
-		fluxReversalCyclesLeft -= minCycles;
+		cycles -= minCycles;
-		cycles += minCycles;
+		UE7Counter -= minCycles;
-		UE7Counter += minCycles;
 
-		if (cycles == 16)
+		if (cycles == 0)
 			return;
 
 		if (cyclesLeftForBit == 0)
@@ -605,14 +657,9 @@ void Drive::DriveLoopRead()
 			}
 		}
 
-		if (fluxReversalCyclesLeft == 0)//Not entirely right, a flux reversal will be skipped if a bit read was going to happen
+		if (UE7Counter == 0x0) // The count carry (bit 4) clocks UF4.
 		{
-			ResetEncoderDecoder(2 * 16, /*25 * 16*/23 * 16); // Trigger a random noise generated zero crossing and start seeing more anywhere between 2us and 25us after this one.
+			UE7Counter = 16 - CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
-		}
-
-		if (UE7Counter == 0x10) // The count carry (bit 4) clocks UF4.
-		{
-			UE7Counter = CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
 										// The decoder consists of UF4 and UE5A. The ecoder has two outputs, Pin 1 of UE5A is the serial data output and pin 2 of UF4 (output B) is the serial clock output.
 			++UF4Counter &= 0xf; // Clock and clamp UF4.
 								 // The UD2 read shift register is clocked by serial clock (the rising edge of encoder/decoder's UF4 B output (serial clock))
@@ -644,21 +691,32 @@ void Drive::DriveLoopRead()
 				m_pVIA->GetPortA()->SetInput(writeShiftRegister);
 			}
 		}
-
-
 	};
 }
-
+void Drive::DriveLoopReadNoCycles()
-void Drive::DriveLoopWrite()
 {
-	const unsigned int cycleStep = 16 - CLOCK_SEL_AB;
+	unsigned int minCycles;
-	unsigned int cycles = 16 - UE7Counter;
+	unsigned int cycles = 16;
-	while(cycles < 16)
+
+	cyclesLeftForBit -= 16;
+	while (true)
 	{
-		//No check is required. This loops is all about the UE7Counter.
+		minCycles = min(cycles, min(UE7Counter, fluxReversalCyclesLeft));
-		//if (UE7Counter == 0x10) // The count carry (bit 4) clocks UF4.
+		fluxReversalCyclesLeft -= minCycles;
+		cycles -= minCycles;
+		UE7Counter -= minCycles;
+
+		if (cycles == 0)
+			return;
+
+		if (fluxReversalCyclesLeft == 0)//Not entirely right, a flux reversal will be skipped if a bit read was going to happen
 		{
-			UE7Counter = CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
+			ResetEncoderDecoder(2 * 16, /*25 * 16*/23 * 16); // Trigger a random noise generated zero crossing and start seeing more anywhere between 2us and 25us after this one.
+		}
+
+		if (UE7Counter == 0x0) // The count carry (bit 4) clocks UF4.
+		{
+			UE7Counter = 16 - CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
 										// The decoder consists of UF4 and UE5A. The ecoder has two outputs, Pin 1 of UE5A is the serial data output and pin 2 of UF4 (output B) is the serial clock output.
 			++UF4Counter &= 0xf; // Clock and clamp UF4.
 								 // The UD2 read shift register is clocked by serial clock (the rising edge of encoder/decoder's UF4 B output (serial clock))
@@ -668,28 +726,166 @@ void Drive::DriveLoopWrite()
 				readShiftRegister <<= 1;
 				readShiftRegister |= (UF4Counter == 2); // Emulate UE5A and only shift in a 1 when pins 6 (output C) and 7 (output D) (bits 2 and 3 of UF4Counter are 0. ie the first count of the bit cell)
 
-				SetNextBit((writeShiftRegister & 0x80));
-
 				writeShiftRegister <<= 1;
 				// Note: SYNC can only trigger during reading as R/!W line is one of UC2's inputs.
-				UE3Counter++;
+				if (((readShiftRegister & 0x3ff) == 0x3ff))	// if the last 10 bits are 1s then SYNC
+				{
+					UE3Counter = 0;	// Phase lock on to byte boundary
+					m_pVIA->GetPortB()->SetInput(0x80, false);			// PB7 active low SYNC
+				}
+				else
+				{
+					m_pVIA->GetPortB()->SetInput(0x80, true); // SYNC not asserted if not following the SYNC bits
+					UE3Counter++;
+				}
 			}
 			// UC5B (NOR used to invert UF4's output B serial clock) output high when UF4 counts 0,1,4,5,8,9,12 and 13
 			else if (((UF4Counter & 2) == 0) && (UE3Counter == 8))	// Phase locked on to byte boundary
 			{
 				UE3Counter = 0;
 				SO = (m_pVIA->GetFCR() & m6522::FCR_CA2_OUTPUT_MODE0) != 0;	// bit 2 of the FCR indicates "Byte Ready Active" turned on or not.
-				writeShiftRegister = m_pVIA->GetPortA()->GetOutput();
+				writeShiftRegister = (u8)(readShiftRegister & 0xff);
-			}
+				m_pVIA->GetPortA()->SetInput(writeShiftRegister);
-
-			if ((16-cycles) < cycleStep)
-			{
-				UE7Counter += 16 - cycles;
-				return;
 			}
 		}
-		UE7Counter += cycleStep;
+	};
-		cycles += cycleStep;
+}
+void Drive::DriveLoopRead()
+{
+	
+	if (fluxReversalCyclesLeft > 16 && cyclesLeftForBit > 16)
+	{
+		DriveLoopReadNoFluxNoCycles();
+		return;
+	}
+
+	if (fluxReversalCyclesLeft > 16)
+	{
+		DriveLoopReadNoFlux();
+		return;
+	}
+
+	if (cyclesLeftForBit > 16)
+	{
+		DriveLoopReadNoCycles();
+		return;
+	}
+
+	unsigned int minCycles;
+	unsigned int cycles = 16;
+
+	while (true)
+	{
+		minCycles = min(min(cyclesLeftForBit, cycles), min(UE7Counter, fluxReversalCyclesLeft));
+		cyclesLeftForBit -= minCycles;
+		fluxReversalCyclesLeft -= minCycles;
+		cycles -= minCycles;
+		UE7Counter -= minCycles;
+
+		if (cycles == 0)
+			return;
+
+		if (cyclesLeftForBit == 0)
+		{
+			//which is faster? single loop ceil check or the 3 lines below?
+			float fn = 2.0f * cyclesPerBit - cyclesForBit;
+			cyclesLeftForBit = (int)fn;
+			cyclesForBit = cyclesPerBit;
+			if (fn != (float)cyclesLeftForBit) {
+				++cyclesLeftForBit;
+				++cyclesForBit;
+			}
+
+			//cyclesForBit -= cyclesPerBit;
+			//cyclesLeftForBit = ceil(cyclesPerBit - cyclesForBit);
+			//cyclesForBit += cyclesLeftForBit;
+
+			if (GetNextBit())
+			{
+				ResetEncoderDecoder(18 * 16, /*20 * 16*/ 2 * 16);
+			}
+		}
+
+		if (fluxReversalCyclesLeft == 0)//Not entirely right, a flux reversal will be skipped if a bit read was going to happen
+		{
+			ResetEncoderDecoder(2 * 16, /*25 * 16*/23 * 16); // Trigger a random noise generated zero crossing and start seeing more anywhere between 2us and 25us after this one.
+		}
+
+		if (UE7Counter == 0x0) // The count carry (bit 4) clocks UF4.
+		{
+			UE7Counter = 16-CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
+										// The decoder consists of UF4 and UE5A. The ecoder has two outputs, Pin 1 of UE5A is the serial data output and pin 2 of UF4 (output B) is the serial clock output.
+			++UF4Counter &= 0xf; // Clock and clamp UF4.
+								 // The UD2 read shift register is clocked by serial clock (the rising edge of encoder/decoder's UF4 B output (serial clock))
+								 //	- ie on counts 2, 6, 10 and 14 (2 is the only count that outputs a 1 into readShiftRegister as the MSB bits of the count NORed together for other values are 0)
+			//if ((UF4Counter & 0x3) == 2)
+			if (UF4Counter == 2 || UF4Counter == 6) //You'd think the bit operation should be faster...
+			{
+				readShiftRegister <<= 1;
+				readShiftRegister |= (UF4Counter == 2); // Emulate UE5A and only shift in a 1 when pins 6 (output C) and 7 (output D) (bits 2 and 3 of UF4Counter are 0. ie the first count of the bit cell)
+
+				writeShiftRegister <<= 1;
+				// Note: SYNC can only trigger during reading as R/!W line is one of UC2's inputs.
+				if (((readShiftRegister & 0x3ff) == 0x3ff))	// if the last 10 bits are 1s then SYNC
+				{
+					UE3Counter = 0;	// Phase lock on to byte boundary
+					m_pVIA->GetPortB()->SetInput(0x80, false);			// PB7 active low SYNC
+				}
+				else
+				{
+					m_pVIA->GetPortB()->SetInput(0x80, true); // SYNC not asserted if not following the SYNC bits
+					UE3Counter++;
+				}
+			}
+			// UC5B (NOR used to invert UF4's output B serial clock) output high when UF4 counts 0,1,4,5,8,9,12 and 13
+			else if (((UF4Counter & 2) == 0) && (UE3Counter == 8))	// Phase locked on to byte boundary
+			{
+				UE3Counter = 0;
+				SO = (m_pVIA->GetFCR() & m6522::FCR_CA2_OUTPUT_MODE0) != 0;	// bit 2 of the FCR indicates "Byte Ready Active" turned on or not.
+				writeShiftRegister = (u8)(readShiftRegister & 0xff);
+				m_pVIA->GetPortA()->SetInput(writeShiftRegister);
+			}
+		}
+	};
+}
+
+void Drive::DriveLoopWrite()
+{
+	unsigned int cycles = 16;
+	while(true)
+	{
+		if (cycles < UE7Counter)
+		{
+			UE7Counter -= cycles;
+			cycles = 0;
+			return;
+		}
+		
+		cycles -= UE7Counter;
+
+		UE7Counter = 16 - CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6) ie preload the encoder/decoder clock for the current density settings.
+									// The decoder consists of UF4 and UE5A. The ecoder has two outputs, Pin 1 of UE5A is the serial data output and pin 2 of UF4 (output B) is the serial clock output.
+		++UF4Counter &= 0xf; // Clock and clamp UF4.
+								// The UD2 read shift register is clocked by serial clock (the rising edge of encoder/decoder's UF4 B output (serial clock))
+								//	- ie on counts 2, 6, 10 and 14 (2 is the only count that outputs a 1 into readShiftRegister as the MSB bits of the count NORed together for other values are 0)
+		if ((UF4Counter & 0x3) == 2)
+		{
+			readShiftRegister <<= 1;
+			readShiftRegister |= (UF4Counter == 2); // Emulate UE5A and only shift in a 1 when pins 6 (output C) and 7 (output D) (bits 2 and 3 of UF4Counter are 0. ie the first count of the bit cell)
+
+			SetNextBit((writeShiftRegister & 0x80));
+
+			writeShiftRegister <<= 1;
+			// Note: SYNC can only trigger during reading as R/!W line is one of UC2's inputs.
+			UE3Counter++;
+		}
+		// UC5B (NOR used to invert UF4's output B serial clock) output high when UF4 counts 0,1,4,5,8,9,12 and 13
+		else if (((UF4Counter & 2) == 0) && (UE3Counter == 8))	// Phase locked on to byte boundary
+		{
+			UE3Counter = 0;
+			SO = (m_pVIA->GetFCR() & m6522::FCR_CA2_OUTPUT_MODE0) != 0;	// bit 2 of the FCR indicates "Byte Ready Active" turned on or not.
+			writeShiftRegister = m_pVIA->GetPortA()->GetOutput();
+		}
 	}
 }
 #endif

src/Drive.h

@@ -51,6 +51,9 @@ public:
 #if defined(EXPERIMENTALZERO)
 	void DriveLoopWrite();
 	void DriveLoopRead();
+	void DriveLoopReadNoFluxNoCycles();
+	void DriveLoopReadNoFlux();
+	void DriveLoopReadNoCycles();
 #endif
 
 	void Insert(DiskImage* diskImage);
@@ -69,7 +72,7 @@ private:
 	int32_t localSeed;
 	inline void ResetEncoderDecoder(unsigned int min, unsigned int /*max*/span)
 	{
-		UE7Counter = CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6)
+		UE7Counter = 16 - CLOCK_SEL_AB;	// A and B inputs of UE7 come from the VIA's CLOCK SEL A/B outputs (ie PB5/6)
 		UF4Counter = 0;
 		localSeed = ((localSeed * 1103515245) + 12345) & 0x7fffffff;
 		fluxReversalCyclesLeft = (span) * (localSeed >> 11) + min;

src/Pi1581.cpp

@@ -93,6 +93,7 @@ extern u16 pc;
 u8 read6502_1581(u16 address)
 {
 	u8 value = 0;
+#if not defined(EXPERIMENTALZERO)
 	if (address & 0x8000)
 	{
 		value = roms.Read1581(address);
@@ -115,6 +116,7 @@ u8 read6502_1581(u16 address)
 	{
 		value = address >> 8;	// Empty address bus
 	}
+#endif
 	return value;
 }
 
@@ -127,6 +129,7 @@ u8 peek6502_1581(u16 address)
 
 void write6502_1581(u16 address, const u8 value)
 {
+#if not defined(EXPERIMENTALZERO)
 	if (address & 0x8000)
 	{
 		return;
@@ -145,6 +148,7 @@ void write6502_1581(u16 address, const u8 value)
 	{
 		s_u8Memory[address & 0x1fff] = value;
 	}
+#endif
 }
 
 static void CIAPortA_OnPortOut(void* pUserData, unsigned char status)

src/iec_bus.h

@@ -277,7 +277,7 @@ public:
 			RPI_SetGpioPinFunction((rpi_gpio_pin_t)PIGPIO_OUT_SRQ, FS_OUTPUT);
 		}
 	
-
+#if not defined(EXPERIMENTALZERO)
 		// Set up audio.
 		write32(CM_PWMDIV, CM_PASSWORD + 0x2000);
 		write32(CM_PWMCTL, CM_PASSWORD + CM_ENAB + CM_SRC_OSCILLATOR);	// Use Default 100MHz Clock
@@ -285,7 +285,7 @@ public:
 		write32(PWM_RNG1, 0x1B4);	// 8bit 44100Hz Mono
 		write32(PWM_RNG2, 0x1B4);
 		write32(PWM_CTL, PWM_USEF2 + PWM_PWEN2 + PWM_USEF1 + PWM_PWEN1 + PWM_CLRF1);
-
+#endif
 
 		int buttonCount = sizeof(ButtonPinFlags) / sizeof(unsigned);
 		for (index = 0; index < buttonCount; ++index)

src/main.cpp

@@ -101,7 +101,9 @@ u8 s_u8Memory[0xc000];
 int numberOfUSBMassStorageDevices = 0;
 DiskCaddy diskCaddy;
 Pi1541 pi1541;
+#if not defined(EXPERIMENTALZERO)
 Pi1581 pi1581;
+#endif
 CEMMCDevice	m_EMMC;
 Screen screen;
 ScreenLCD* screenLCD = 0;
@@ -118,8 +120,9 @@ bool USBKeyboardDetected = false;
 bool selectedViaIECCommands = false;
 u16 pc;
 
+#if not defined(EXPERIMENTALZERO)
 SpinLock core0RefreshingScreen;
-
+#endif
 unsigned int screenWidth = 1024;
 unsigned int screenHeight = 768;
 
@@ -609,6 +612,7 @@ EmulatingMode BeginEmulating(FileBrowser* fileBrowser, const char* filenameForIc
 	DiskImage* diskImage = diskCaddy.SelectFirstImage();
 	if (diskImage)
 	{
+#if not defined(EXPERIMENTALZERO)
 		if (diskImage->IsD81())
 		{
 			pi1581.Insert(diskImage);
@@ -617,6 +621,7 @@ EmulatingMode BeginEmulating(FileBrowser* fileBrowser, const char* filenameForIc
 			return EMULATING_1581;
 		}
 		else
+#endif
 		{
 			pi1541.drive.Insert(diskImage);
 			fileBrowser->DisplayDiskInfo(diskImage, filenameForIcon);
@@ -668,7 +673,9 @@ void GlobalSetDeviceID(u8 id)
 	deviceID = id;
 	m_IEC_Commands.SetDeviceId(id);
 	pi1541.SetDeviceID(id);
+#if not defined(EXPERIMENTALZERO)
 	pi1581.SetDeviceID(id);
+#endif
 }
 
 void CheckAutoMountImage(EXIT_TYPE reset_reason , FileBrowser* fileBrowser)
@@ -712,10 +719,6 @@ EXIT_TYPE Emulate1541(FileBrowser* fileBrowser)
 	if (numberOfImagesMax > 10)
 		numberOfImagesMax = 10;
 
-	core0RefreshingScreen.Acquire();
-	diskCaddy.Display();
-	core0RefreshingScreen.Release();
-
 	inputMappings->directDiskSwapRequest = 0;
 	// Force an update on all the buttons now before we start emulation mode. 
 	IEC_Bus::ReadBrowseMode();
@@ -731,8 +734,6 @@ EXIT_TYPE Emulate1541(FileBrowser* fileBrowser)
 	pi1541.Reset();	// will call IEC_Bus::Reset();
 	IEC_Bus::OutputLED = false;
 	IEC_Bus::LetSRQBePulledHigh();
-	float avgTimer = 0.0f;
-
 	ctBefore = read32(ARM_SYSTIMER_CLO);
 
 	//resetWhileEmulating = false;
@@ -756,7 +757,6 @@ EXIT_TYPE Emulate1541(FileBrowser* fileBrowser)
 	}
 	bool buttonState = false;
 	bool prevButtonState = false;
-
 	while (true)
 	{
 
@@ -1274,8 +1274,9 @@ void emulator()
 			IEC_Bus::Reset();
 
 			IEC_Bus::LetSRQBePulledHigh();
-
+#if not defined(EXPERIMENTALZERO)
 			core0RefreshingScreen.Acquire();
+#endif
 			IEC_Bus::WaitMicroSeconds(100);
 
 			roms.ResetCurrentROMIndex();
@@ -1285,9 +1286,9 @@ void emulator()
 			fileBrowser->ClearSelections();
 
 			fileBrowser->RefeshDisplay(); // Just redisplay the current folder.
-
+#if not defined(EXPERIMENTALZERO)
 			core0RefreshingScreen.Release();
-
+#endif
 			selectedViaIECCommands = false;
 
 			inputMappings->Reset();
@@ -1590,7 +1591,7 @@ static void CheckOptions()
 
 	deviceID = (u8)options.GetDeviceID();
 	DEBUG_LOG("DeviceID = %d\r\n", deviceID);
-
+#if not defined(EXPERIMENTALZERO)
 	const char* FontROMName = options.GetRomFontName();
 	if (FontROMName)
 	{
@@ -1625,7 +1626,7 @@ static void CheckOptions()
 			//DEBUG_LOG("Read ROM %s from options\r\n", ROMName);
 		}
 	}
-
+#endif
 	const char* ROMName1581 = options.GetRomName1581();
 	if (ROMName1581)
 	{