Merge branch 'master' into wii-network

Source/Core/Core/Core.vcxproj

@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup Label="ProjectConfigurations">
     <ProjectConfiguration Include="DebugFast|Win32">
@@ -34,12 +34,12 @@
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'" Label="Configuration">
     <UseDebugLibraries>true</UseDebugLibraries>
     <ConfigurationType>StaticLibrary</ConfigurationType>
-    <CharacterSet>MultiByte</CharacterSet>
+    <CharacterSet>Unicode</CharacterSet>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="Configuration">
     <UseDebugLibraries>true</UseDebugLibraries>
     <ConfigurationType>StaticLibrary</ConfigurationType>
-    <CharacterSet>MultiByte</CharacterSet>
+    <CharacterSet>Unicode</CharacterSet>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'" Label="Configuration">
     <UseDebugLibraries>false</UseDebugLibraries>
@@ -49,7 +49,7 @@
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='DebugFast|Win32'" Label="Configuration">
     <ConfigurationType>StaticLibrary</ConfigurationType>
     <UseDebugLibraries>false</UseDebugLibraries>
-    <CharacterSet>MultiByte</CharacterSet>
+    <CharacterSet>Unicode</CharacterSet>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="Configuration">
     <UseDebugLibraries>false</UseDebugLibraries>
@@ -59,7 +59,7 @@
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='DebugFast|x64'" Label="Configuration">
     <ConfigurationType>StaticLibrary</ConfigurationType>
     <UseDebugLibraries>false</UseDebugLibraries>
-    <CharacterSet>MultiByte</CharacterSet>
+    <CharacterSet>Unicode</CharacterSet>
   </PropertyGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
   <ImportGroup Label="ExtensionSettings">

Source/Core/Core/Src/DSPEmulator.h

@@ -44,6 +44,7 @@ public:
 	virtual void DSP_Update(int cycles) = 0;
 	virtual void DSP_StopSoundStream() = 0;
 	virtual void DSP_ClearAudioBuffer(bool mute) = 0;
+	virtual u32 DSP_UpdateRate() = 0;
 
 protected:
 	SoundStream *soundStream;

Source/Core/Core/Src/HW/DSPHLE/DSPHLE.cpp

@@ -28,6 +28,7 @@
 #include "../AudioInterface.h"
 #include "ConfigManager.h"
 #include "Core.h"
+#include "HW/SystemTimers.h"
 
 DSPHLE::DSPHLE() {
 	m_InitMixer = false;
@@ -87,6 +88,16 @@ void DSPHLE::DSP_Update(int cycles)
 		m_pUCode->Update(cycles / 6);
 }
 
+u32 DSPHLE::DSP_UpdateRate()
+{
+	// AX HLE uses 3ms (Wii) or 5ms (GC) timing period
+	int fields = SConfig::GetInstance().m_LocalCoreStartupParameter.bVBeam ? 2 : 1;
+	if (m_pUCode != NULL)
+		return (SystemTimers::GetTicksPerSecond() / 1000) * m_pUCode->GetUpdateMs() / fields;
+	else
+		return SystemTimers::GetTicksPerSecond() / 1000;
+}
+
 void DSPHLE::SendMailToDSP(u32 _uMail)
 {
 	if (m_pUCode != NULL) {

Source/Core/Core/Src/HW/DSPHLE/DSPHLE.h

@@ -46,6 +46,7 @@ public:
 	virtual void DSP_Update(int cycles);
 	virtual void DSP_StopSoundStream();
 	virtual void DSP_ClearAudioBuffer(bool mute);
+	virtual u32 DSP_UpdateRate();
 
 	CMailHandler& AccessMailHandler() { return m_MailHandler; }
 

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp

@@ -738,6 +738,11 @@ void CUCode_AX::Update(int cycles)
 	}
 }
 
+u32 CUCode_AX::GetUpdateMs()
+{
+	return 5;
+}
+
 void CUCode_AX::DoAXState(PointerWrap& p)
 {
 	p.Do(m_cmdlist);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h

@@ -73,6 +73,7 @@ public:
 	virtual void MixAdd(short* out_buffer, int nsamples);
 	virtual void Update(int cycles);
 	virtual void DoState(PointerWrap& p);
+	u32 GetUpdateMs();
 
 	// Needed because StdThread.h std::thread implem does not support member
 	// pointers. TODO(delroth): obsolete.

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp

@@ -672,6 +672,11 @@ void CUCode_AXWii::OutputWMSamples(u32* addresses)
 	}
 }
 
+u32 CUCode_AXWii::GetUpdateMs()
+{
+	return 3;
+}
+
 void CUCode_AXWii::DoState(PointerWrap &p)
 {
 	std::lock_guard<std::mutex> lk(m_processing);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h

@@ -25,6 +25,7 @@ class CUCode_AXWii : public CUCode_AX
 public:
 	CUCode_AXWii(DSPHLE *dsp_hle, u32 _CRC);
 	virtual ~CUCode_AXWii();
+	u32 GetUpdateMs();
 
 	virtual void DoState(PointerWrap &p);
 

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_CARD.cpp

@@ -19,6 +19,7 @@
 #include "UCodes.h"
 #include "UCode_CARD.h"
 #include "../../DSP.h"
+#include "ConfigManager.h"
 
 
 CUCode_CARD::CUCode_CARD(DSPHLE *dsp_hle, u32 crc)
@@ -44,6 +45,11 @@ void CUCode_CARD::Update(int cycles)
 	}
 }
 
+u32 CUCode_CARD::GetUpdateMs()
+{
+	return SConfig::GetInstance().m_LocalCoreStartupParameter.bWii ? 3 : 5;
+}
+
 void CUCode_CARD::HandleMail(u32 _uMail)
 {
 	if (_uMail == 0xFF000000) // unlock card

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_CARD.h

@@ -25,6 +25,7 @@ class CUCode_CARD : public IUCode
 public:
 	CUCode_CARD(DSPHLE *dsp_hle, u32 crc);
 	virtual ~CUCode_CARD();
+	u32 GetUpdateMs();
 
 	void HandleMail(u32 _uMail);
 	void Update(int cycles);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_GBA.cpp

@@ -19,6 +19,7 @@
 #include "UCode_GBA.h"
 
 #include "../../DSP.h"
+#include "ConfigManager.h"
 
 CUCode_GBA::CUCode_GBA(DSPHLE *dsp_hle, u32 crc)
 : IUCode(dsp_hle, crc)
@@ -40,6 +41,11 @@ void CUCode_GBA::Update(int cycles)
 	}
 }
 
+u32 CUCode_GBA::GetUpdateMs()
+{
+	return SConfig::GetInstance().m_LocalCoreStartupParameter.bWii ? 3 : 5;
+}
+
 void CUCode_GBA::HandleMail(u32 _uMail)
 {
 	static bool nextmail_is_mramaddr = false;

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_GBA.h

@@ -23,6 +23,7 @@ struct CUCode_GBA : public IUCode
 {
 	CUCode_GBA(DSPHLE *dsp_hle, u32 crc);
 	virtual ~CUCode_GBA();
+	u32 GetUpdateMs();
 
 	void HandleMail(u32 _uMail);
 	void Update(int cycles);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_InitAudioSystem.cpp

@@ -17,6 +17,7 @@
 
 #include "UCodes.h"
 #include "UCode_InitAudioSystem.h"
+#include "ConfigManager.h"
 
 CUCode_InitAudioSystem::CUCode_InitAudioSystem(DSPHLE *dsp_hle, u32 crc)
 	: IUCode(dsp_hle, crc)
@@ -42,6 +43,11 @@ void CUCode_InitAudioSystem::Update(int cycles)
 	}
 }
 
+u32 CUCode_InitAudioSystem::GetUpdateMs()
+{
+	return SConfig::GetInstance().m_LocalCoreStartupParameter.bWii ? 3 : 5;
+}
+
 void CUCode_InitAudioSystem::HandleMail(u32 _uMail)
 {}
 

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_InitAudioSystem.h

@@ -25,6 +25,7 @@ class CUCode_InitAudioSystem : public IUCode
 public:
 	CUCode_InitAudioSystem(DSPHLE *dsp_hle, u32 crc);
 	virtual ~CUCode_InitAudioSystem();
+	u32 GetUpdateMs();
 
 	void HandleMail(u32 _uMail);
 	void Update(int cycles);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_ROM.cpp

@@ -19,6 +19,7 @@
 #include "UCode_ROM.h"
 #include "Hash.h"
 #include "../../Memmap.h"
+#include "ConfigManager.h"
 
 CUCode_Rom::CUCode_Rom(DSPHLE *dsp_hle, u32 crc)
 	: IUCode(dsp_hle, crc)
@@ -117,6 +118,11 @@ void CUCode_Rom::BootUCode()
 	m_DSPHLE->SetUCode(ector_crc);
 }
 
+u32 CUCode_Rom::GetUpdateMs()
+{
+	return SConfig::GetInstance().m_LocalCoreStartupParameter.bWii ? 3 : 5;
+}
+
 void CUCode_Rom::DoState(PointerWrap &p)
 {
 	p.Do(m_CurrentUCode);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_ROM.h

@@ -25,6 +25,7 @@ class CUCode_Rom : public IUCode
 public:
 	CUCode_Rom(DSPHLE *dsp_hle, u32 _crc);
 	virtual ~CUCode_Rom();
+	u32 GetUpdateMs();
 
 	void HandleMail(u32 _uMail);
 	void Update(int cycles);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_Zelda.cpp

@@ -27,6 +27,7 @@
 
 #include "WaveFile.h"
 #include "../../DSP.h"
+#include "ConfigManager.h"
 
 
 CUCode_Zelda::CUCode_Zelda(DSPHLE *dsp_hle, u32 _CRC)
@@ -565,6 +566,10 @@ void CUCode_Zelda::ExecuteList()
 	}
 }
 
+u32 CUCode_Zelda::GetUpdateMs()
+{
+	return SConfig::GetInstance().m_LocalCoreStartupParameter.bWii ? 3 : 5;
+}
 
 void CUCode_Zelda::DoState(PointerWrap &p)
 {

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_Zelda.h

@@ -133,6 +133,7 @@ class CUCode_Zelda : public IUCode
 public:
 	CUCode_Zelda(DSPHLE *dsp_hle, u32 _CRC);
 	virtual ~CUCode_Zelda();
+	u32 GetUpdateMs();
 
 	void HandleMail(u32 _uMail);
 	void HandleMail_LightVersion(u32 _uMail);

Source/Core/Core/Src/HW/DSPHLE/UCodes/UCodes.h

@@ -92,6 +92,7 @@ public:
 	// Cycles are out of the 81/121mhz the DSP runs at.
 	virtual void Update(int cycles) = 0;
 	virtual void MixAdd(short* buffer, int size) {}
+	virtual u32 GetUpdateMs() = 0;
 
 	virtual void DoState(PointerWrap &p) { DoStateShared(p); }
 

Source/Core/Core/Src/HW/DSPLLE/DSPLLE.cpp

@@ -326,6 +326,11 @@ void DSPLLE::DSP_Update(int cycles)
 	}
 }
 
+u32 DSPLLE::DSP_UpdateRate()
+{
+	return 12600; // TO BE TWEAKED
+}
+
 void DSPLLE::DSP_SendAIBuffer(unsigned int address, unsigned int num_samples)
 {
 	if (!soundStream)

Source/Core/Core/Src/HW/DSPLLE/DSPLLE.h

@@ -44,6 +44,7 @@ public:
 	virtual void DSP_Update(int cycles);
 	virtual void DSP_StopSoundStream();
 	virtual void DSP_ClearAudioBuffer(bool mute);
+	virtual u32 DSP_UpdateRate();
 
 private:
 	static void dsp_thread(DSPLLE* lpParameter);

Source/Core/Core/Src/HW/SystemTimers.cpp

@@ -119,9 +119,6 @@ int et_PatchEngine;	// PatchEngine updates every 1/60th of a second by default
 // These are badly educated guesses
 // Feel free to experiment. Set these in Init below.
 int
-	// These shouldn't be period controlled either, most likely.
-	DSP_PERIOD,
-
 	// This is a fixed value, don't change it 
 	AUDIO_DMA_PERIOD,
 
@@ -149,8 +146,8 @@ void DSPCallback(u64 userdata, int cyclesLate)
 {
 	//splits up the cycle budget in case lle is used
 	//for hle, just gives all of the slice to hle
-	DSP::UpdateDSPSlice(DSP_PERIOD - cyclesLate);
+	DSP::UpdateDSPSlice(DSP::GetDSPEmulator()->DSP_UpdateRate() - cyclesLate);
-	CoreTiming::ScheduleEvent(DSP_PERIOD - cyclesLate, et_DSP);
+	CoreTiming::ScheduleEvent(DSP::GetDSPEmulator()->DSP_UpdateRate() - cyclesLate, et_DSP);
 }
 
 void AudioDMACallback(u64 userdata, int cyclesLate)
@@ -256,17 +253,6 @@ void Init()
 		IPC_HLE_PERIOD = GetTicksPerSecond() / (freq * fields);
 	}
 
-	if (DSP::GetDSPEmulator()->IsLLE())
-	{
-		DSP_PERIOD = 12600; // TO BE TWEAKED
-	}
-	else
-	{
-		// AX HLE uses 3ms (Wii) or 5ms (GC) timing period
-		int ms_to_process = SConfig::GetInstance().m_LocalCoreStartupParameter.bWii ? 3 : 5;
-		DSP_PERIOD = (int)(GetTicksPerSecond() / 1000) * ms_to_process / fields;
-	}
-
 	// System internal sample rate is fixed at 32KHz * 4 (16bit Stereo) / 32 bytes DMA
 	AUDIO_DMA_PERIOD = CPU_CORE_CLOCK / (AudioInterface::GetAIDSampleRate() * 4 / 32);
 
@@ -292,7 +278,7 @@ void Init()
 	et_PatchEngine = CoreTiming::RegisterEvent("PatchEngine", PatchEngineCallback);
 
 	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerLine(), et_VI);
-	CoreTiming::ScheduleEvent(DSP_PERIOD, et_DSP);
+	CoreTiming::ScheduleEvent(0, et_DSP);
 	CoreTiming::ScheduleEvent(VideoInterface::GetTicksPerFrame(), et_SI);
 	CoreTiming::ScheduleEvent(AUDIO_DMA_PERIOD, et_AudioDMA);
 	if (SConfig::GetInstance().m_LocalCoreStartupParameter.bSyncGPU)

Source/Core/VideoCommon/Src/IndexGenerator.cpp

@@ -18,17 +18,9 @@
 #include <cstddef>
 
 #include "Common.h"
+#include "VideoConfig.h"
 #include "IndexGenerator.h"
 
-/*
-*
-QUAD simulator
-
-0   1   4   5
-3   2   7   6
-012023 147172 ...
-*/
-
 //Init
 u16 *IndexGenerator::Tptr;
 u16 *IndexGenerator::BASETptr;
@@ -41,6 +33,32 @@ u32 IndexGenerator::numL;
 u32 IndexGenerator::numP;
 u32 IndexGenerator::index;
 
+static const u16 s_primitive_restart = -1;
+
+static void (*primitive_table[8])(u32);
+
+void IndexGenerator::Init()
+{
+	if(g_Config.backend_info.bSupportsPrimitiveRestart)
+	{
+		primitive_table[0] = IndexGenerator::AddQuads<true>;
+		primitive_table[2] = IndexGenerator::AddList<true>;
+		primitive_table[3] = IndexGenerator::AddStrip<true>;
+		primitive_table[4] = IndexGenerator::AddFan<true>;
+	}
+	else
+	{
+		primitive_table[0] = IndexGenerator::AddQuads<false>;
+		primitive_table[2] = IndexGenerator::AddList<false>;
+		primitive_table[3] = IndexGenerator::AddStrip<false>;
+		primitive_table[4] = IndexGenerator::AddFan<false>;
+	}
+	primitive_table[1] = NULL;
+	primitive_table[5] = &IndexGenerator::AddLineList;
+	primitive_table[6] = &IndexGenerator::AddLineStrip;
+	primitive_table[7] = &IndexGenerator::AddPoints;
+}
+
 void IndexGenerator::Start(u16* Triangleptr, u16* Lineptr, u16* Pointptr)
 {
 	Tptr = Triangleptr;
@@ -57,82 +75,150 @@ void IndexGenerator::Start(u16* Triangleptr, u16* Lineptr, u16* Pointptr)
 
 void IndexGenerator::AddIndices(int primitive, u32 numVerts)
 {
-	//switch (primitive)
-	//{
-	//case GX_DRAW_QUADS:          IndexGenerator::AddQuads(numVerts);		break;
-	//case GX_DRAW_TRIANGLES:      IndexGenerator::AddList(numVerts);		break;
-	//case GX_DRAW_TRIANGLE_STRIP: IndexGenerator::AddStrip(numVerts);		break;
-	//case GX_DRAW_TRIANGLE_FAN:   IndexGenerator::AddFan(numVerts);		break;
-	//case GX_DRAW_LINES:          IndexGenerator::AddLineList(numVerts);	break;
-	//case GX_DRAW_LINE_STRIP:     IndexGenerator::AddLineStrip(numVerts);	break;
-	//case GX_DRAW_POINTS:         IndexGenerator::AddPoints(numVerts);		break;
-	//}
-
-	static void (*const primitive_table[])(u32) =
-	{
-		IndexGenerator::AddQuads,
-		NULL,
-		IndexGenerator::AddList,
-		IndexGenerator::AddStrip,
-		IndexGenerator::AddFan,
-		IndexGenerator::AddLineList,
-		IndexGenerator::AddLineStrip,
-		IndexGenerator::AddPoints,
-	};
-
 	primitive_table[primitive](numVerts);
 	index += numVerts;
 }
 
 // Triangles
-__forceinline void IndexGenerator::WriteTriangle(u32 index1, u32 index2, u32 index3)
+template <bool pr> __forceinline void IndexGenerator::WriteTriangle(u32 index1, u32 index2, u32 index3)
 {
 	*Tptr++ = index1;
 	*Tptr++ = index2;
 	*Tptr++ = index3;
-
+	if(pr)
+		*Tptr++ = s_primitive_restart;
+	
 	++numT;
 }
 
-void IndexGenerator::AddList(u32 const numVerts)
+template <bool pr> void IndexGenerator::AddList(u32 const numVerts)
 {
 	auto const numTris = numVerts / 3;
 	for (u32 i = 0; i != numTris; ++i)
 	{
-		WriteTriangle(index + i * 3, index + i * 3 + 1, index + i * 3 + 2);
+		WriteTriangle<pr>(index + i * 3, index + i * 3 + 1, index + i * 3 + 2);
 	}
+	u32 remainingVerts = numVerts - numTris*3;
+	if(remainingVerts)
+		ERROR_LOG(VIDEO, "AddList: unknown count of vertices found");
 }
 
-void IndexGenerator::AddStrip(u32 const numVerts)
+template <bool pr> void IndexGenerator::AddStrip(u32 const numVerts)
 {
-	bool wind = false;
+	if(pr) {
-	for (u32 i = 2; i < numVerts; ++i)
+		for (u32 i = 0; i < numVerts; ++i)
-	{
+		{
-		WriteTriangle(
+			*Tptr++ = index + i;
-			index + i - 2,
+		}
-			index + i - !wind,
+		*Tptr++ = s_primitive_restart;
-			index + i - wind);
+		numT += numVerts - 2;
+		
+	} else {
+		bool wind = false;
+		for (u32 i = 2; i < numVerts; ++i)
+		{
+			WriteTriangle<pr>(
+				index + i - 2,
+				index + i - !wind,
+				index + i - wind);
 
-		wind ^= true;
+			wind ^= true;
+		}
 	}
 }
 
-void IndexGenerator::AddFan(u32 numVerts)
+/**
+ * FAN simulator:
+ * 
+ *   2---3
+ *  / \ / \
+ * 1---0---4
+ * 
+ * would generate this triangles:
+ * 012, 023, 034
+ * 
+ * rotated (for better striping):
+ * 120, 302, 034
+ * 
+ * as odd ones have to winded, following strip is fine:
+ * 12034
+ * 
+ * so we use 6 indices for 3 triangles
+ */
+
+template <bool pr> void IndexGenerator::AddFan(u32 numVerts)
 {
-	for (u32 i = 2; i < numVerts; ++i)
+	u32 i = 2;
+	
+	if(pr) {
+		for(; i<=numVerts-3; i+=3) {
+			*Tptr++ = index + i - 1;
+			*Tptr++ = index + i + 0;
+			*Tptr++ = index;
+			*Tptr++ = index + i + 1;
+			*Tptr++ = index + i + 2;
+			*Tptr++ = s_primitive_restart;
+			numT += 3;
+		}
+		
+		for(; i<=numVerts-2; i+=2) {
+			*Tptr++ = index + i - 1;
+			*Tptr++ = index + i + 0;
+			*Tptr++ = index;
+			*Tptr++ = index + i + 1;
+			*Tptr++ = s_primitive_restart;
+			numT += 2;
+		}
+	}
+	
+	for (; i < numVerts; ++i)
 	{
-		WriteTriangle(index, index + i - 1, index + i);
+		WriteTriangle<pr>(index, index + i - 1, index + i);
 	}
 }
 
-void IndexGenerator::AddQuads(u32 numVerts)
+/*
+ * QUAD simulator
+ * 
+ * 0---1   4---5
+ * |\  |   |\  |
+ * | \ |   | \ |
+ * |  \|   |  \|
+ * 3---2   7---6
+ * 
+ * 012,023, 456,467 ...
+ * or 120,302, 564,746
+ * or as strip: 1203, 5647
+ * 
+ * Warning: 
+ * A simple triangle has to be rendered for three vertices.
+ * ZWW do this for sun rays
+ */
+template <bool pr> void IndexGenerator::AddQuads(u32 numVerts)
 {
 	auto const numQuads = numVerts / 4;
 	for (u32 i = 0; i != numQuads; ++i)
 	{
-		WriteTriangle(index + i * 4, index + i * 4 + 1, index + i * 4 + 2);
+		if(pr) {
-		WriteTriangle(index + i * 4, index + i * 4 + 2, index + i * 4 + 3);
+			*Tptr++ = index + i * 4 + 1;
+			*Tptr++ = index + i * 4 + 2;
+			*Tptr++ = index + i * 4 + 0;
+			*Tptr++ = index + i * 4 + 3;
+			*Tptr++ = s_primitive_restart;
+			numT += 2;
+		} else {
+			WriteTriangle<pr>(index + i * 4, index + i * 4 + 1, index + i * 4 + 2);
+			WriteTriangle<pr>(index + i * 4, index + i * 4 + 2, index + i * 4 + 3);
+		}
 	}
+	// three vertices remaining, so render a triangle
+	u32 remainingVerts = numVerts - numQuads*4;
+	if(remainingVerts == 3)
+	{
+		WriteTriangle<pr>(index+numVerts-3, index+numVerts-2, index+numVerts-1);
+	}
+	else if(remainingVerts)
+		ERROR_LOG(VIDEO, "AddQuads: unknown count of vertices found");
 }
 
 // Lines
@@ -145,8 +231,14 @@ void IndexGenerator::AddLineList(u32 numVerts)
 		*Lptr++ = index + i * 2 + 1;
 		++numL;
 	}
+	u32 remainingVerts = numVerts - numLines*2;
+	if(remainingVerts)
+		ERROR_LOG(VIDEO, "AddLineList: unknown count of vertices found");
+	
 }
 
+// shouldn't be used as strips as LineLists are much more common
+// so converting them to lists
 void IndexGenerator::AddLineStrip(u32 numVerts)
 {
 	for (u32 i = 1; i < numVerts; ++i)
@@ -170,6 +262,6 @@ void IndexGenerator::AddPoints(u32 numVerts)
 
 u32 IndexGenerator::GetRemainingIndices()
 {
-	u32 max_index = 65535;
+	u32 max_index = 65534; // -1 is reserved for primitive restart (ogl + dx11)
 	return max_index - index;
 }

Source/Core/VideoCommon/Src/IndexGenerator.h

@@ -26,6 +26,7 @@ class IndexGenerator
 {
 public:
 	// Init
+	static void Init();
 	static void Start(u16 *Triangleptr,u16 *Lineptr,u16 *Pointptr);
 
 	static void AddIndices(int primitive, u32 numVertices);
@@ -54,10 +55,10 @@ public:
 */
 private:
 	// Triangles
-	static void AddList(u32 numVerts);
+	template <bool pr> static void AddList(u32 numVerts);
-	static void AddStrip(u32 numVerts);
+	template <bool pr> static void AddStrip(u32 numVerts);
-	static void AddFan(u32 numVerts);
+	template <bool pr> static void AddFan(u32 numVerts);
-	static void AddQuads(u32 numVerts);
+	template <bool pr> static void AddQuads(u32 numVerts);
 
 	// Lines
 	static void AddLineList(u32 numVerts);
@@ -66,7 +67,7 @@ private:
 	// Points
 	static void AddPoints(u32 numVerts);
 
-	static void WriteTriangle(u32 index1, u32 index2, u32 index3);
+	template <bool pr> static void WriteTriangle(u32 index1, u32 index2, u32 index3);
 
 	static u16 *Tptr;
 	static u16 *BASETptr;

Source/Core/VideoCommon/Src/VertexLoader.cpp

@@ -659,6 +659,10 @@ void VertexLoader::SetVAT(u32 _group0, u32 _group1, u32 _group2)
 	m_VtxAttr.texCoord[7].Elements	= vat.g2.Tex7CoordElements;
 	m_VtxAttr.texCoord[7].Format	= vat.g2.Tex7CoordFormat;
 	m_VtxAttr.texCoord[7].Frac		= vat.g2.Tex7Frac;
+	
+	if(!m_VtxAttr.ByteDequant) {
+		ERROR_LOG(VIDEO, "ByteDequant is set to zero");
+	}
 };
 
 void VertexLoader::AppendToString(std::string *dest) const

Source/Core/VideoCommon/Src/VertexManagerBase.cpp

@@ -74,28 +74,54 @@ bool VertexManager::IsFlushed() const
 
 u32 VertexManager::GetRemainingIndices(int primitive)
 {
-	switch (primitive)
+	
-	{
+	if(g_Config.backend_info.bSupportsPrimitiveRestart) {
-	case GX_DRAW_QUADS:
+		switch (primitive)
-		return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 6 * 4;
+		{
-	case GX_DRAW_TRIANGLES:
+		case GX_DRAW_QUADS:
-		return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen());
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 5 * 4;
-	case GX_DRAW_TRIANGLE_STRIP:
+		case GX_DRAW_TRIANGLES:
-		return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 3 + 2;
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 4 * 3;
-	case GX_DRAW_TRIANGLE_FAN:
+		case GX_DRAW_TRIANGLE_STRIP:
-		return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 3 + 2;
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 1 - 1;
+		case GX_DRAW_TRIANGLE_FAN:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 6 * 4 + 1;
 
-	case GX_DRAW_LINES:
+		case GX_DRAW_LINES:
-		return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen());
+			return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen());
-	case GX_DRAW_LINE_STRIP:
+		case GX_DRAW_LINE_STRIP:
-		return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen()) / 2 + 1;
+			return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen()) / 2 + 1;
 
-	case GX_DRAW_POINTS:
+		case GX_DRAW_POINTS:
-		return (MAXIBUFFERSIZE - IndexGenerator::GetPointindexLen());
+			return (MAXIBUFFERSIZE - IndexGenerator::GetPointindexLen());
 
-	default:
+		default:
-		return 0;
+			return 0;
-	}
+		}
+	} else {
+		switch (primitive)
+		{
+		case GX_DRAW_QUADS:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 6 * 4;
+		case GX_DRAW_TRIANGLES:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen());
+		case GX_DRAW_TRIANGLE_STRIP:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 3 + 2;
+		case GX_DRAW_TRIANGLE_FAN:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 3 + 2;
+
+		case GX_DRAW_LINES:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen());
+		case GX_DRAW_LINE_STRIP:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen()) / 2 + 1;
+
+		case GX_DRAW_POINTS:
+			return (MAXIBUFFERSIZE - IndexGenerator::GetPointindexLen());
+
+		default:
+			return 0;
+		}
+	} 
 }
 
 void VertexManager::AddVertices(int primitive, u32 numVertices)

Source/Core/VideoCommon/Src/VideoConfig.h

@@ -165,6 +165,7 @@ struct VideoConfig
 		bool bSupportsDualSourceBlend; // only supported by D3D11 and OpenGL
 		bool bSupportsFormatReinterpretation;
 		bool bSupportsPixelLighting;
+		bool bSupportsPrimitiveRestart;
 		bool bSupportsSeparateAlphaFunction;
 		bool bSupportsGLSLUBO; // needed by pixelShaderGen, so must stay in videoCommon
 	} backend_info;

Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp

@@ -154,7 +154,7 @@ void VertexManager::Draw(UINT stride)
 	
 	if (IndexGenerator::GetNumTriangles() > 0)
 	{
-		D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
+		D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
 		D3D::context->DrawIndexed(IndexGenerator::GetTriangleindexLen(), m_triangle_draw_index, 0);
 		INCSTAT(stats.thisFrame.numIndexedDrawCalls);
 	}

Source/Plugins/Plugin_VideoDX11/Src/main.cpp

@@ -35,6 +35,7 @@
 #include "Debugger/DebuggerPanel.h"
 #include "DLCache.h"
 #include "EmuWindow.h"
+#include "IndexGenerator.h"
 #include "FileUtil.h"
 #include "Globals.h"
 #include "IniFile.h"
@@ -96,6 +97,7 @@ void InitBackendInfo()
 	g_Config.backend_info.bSupportsDualSourceBlend = true;
 	g_Config.backend_info.bSupportsFormatReinterpretation = true;
 	g_Config.backend_info.bSupportsPixelLighting = true;
+	g_Config.backend_info.bSupportsPrimitiveRestart = true;
 
 	IDXGIFactory* factory;
 	IDXGIAdapter* ad;
@@ -192,6 +194,7 @@ void VideoBackend::Video_Prepare()
 	// VideoCommon
 	BPInit();
 	Fifo_Init();
+	IndexGenerator::Init();
 	VertexLoaderManager::Init();
 	OpcodeDecoder_Init();
 	VertexShaderManager::Init();

Source/Plugins/Plugin_VideoDX9/Src/main.cpp

@@ -51,6 +51,7 @@
 #include "VideoState.h"
 #include "Render.h"
 #include "DLCache.h"
+#include "IndexGenerator.h"
 #include "IniFile.h"
 #include "Core.h"
 #include "Host.h"
@@ -102,6 +103,7 @@ void InitBackendInfo()
 	g_Config.backend_info.bUseRGBATextures = false;
 	g_Config.backend_info.bUseMinimalMipCount = true;
 	g_Config.backend_info.bSupports3DVision = true;
+	g_Config.backend_info.bSupportsPrimitiveRestart = false; // TODO: figure out if it does
 	g_Config.backend_info.bSupportsSeparateAlphaFunction = device_caps.PrimitiveMiscCaps & D3DPMISCCAPS_SEPARATEALPHABLEND;
 	// Dual source blend disabled by default until a proper method to test for support is found	
 	g_Config.backend_info.bSupportsDualSourceBlend = false;
@@ -183,6 +185,7 @@ void VideoBackend::Video_Prepare()
 	// VideoCommon
 	BPInit();
 	Fifo_Init();
+	IndexGenerator::Init();
 	VertexLoaderManager::Init();
 	OpcodeDecoder_Init();
 	VertexShaderManager::Init();

Source/Plugins/Plugin_VideoOGL/Src/Render.cpp

@@ -338,6 +338,7 @@ Renderer::Renderer()
 	
 	g_Config.backend_info.bSupportsDualSourceBlend = GLEW_ARB_blend_func_extended;
 	g_Config.backend_info.bSupportsGLSLUBO = GLEW_ARB_uniform_buffer_object;
+	g_Config.backend_info.bSupportsPrimitiveRestart = GLEW_VERSION_3_1 || GLEW_NV_primitive_restart;
 	
 	g_ogl_config.bSupportsGLSLCache = GLEW_ARB_get_program_binary;
 	g_ogl_config.bSupportsGLPinnedMemory = GLEW_AMD_pinned_memory;
@@ -345,6 +346,7 @@ Renderer::Renderer()
 	g_ogl_config.bSupportsGLBaseVertex = GLEW_ARB_draw_elements_base_vertex;
 	g_ogl_config.bSupportCoverageMSAA = GLEW_NV_framebuffer_multisample_coverage;
 	g_ogl_config.bSupportSampleShading = GLEW_ARB_sample_shading;
+	g_ogl_config.bSupportOGL31 = GLEW_VERSION_3_1;
 	
 	g_ogl_config.gl_vendor = (const char*)glGetString(GL_VENDOR);
 	g_ogl_config.gl_renderer = (const char*)glGetString(GL_RENDERER);
@@ -397,9 +399,10 @@ Renderer::Renderer()
 				g_ogl_config.gl_renderer,
 				g_ogl_config.gl_version).c_str(), 5000);
 	
-	WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s",
+	WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s%s",
 			g_ActiveConfig.backend_info.bSupportsDualSourceBlend ? "" : "DualSourceBlend ",
 			g_ActiveConfig.backend_info.bSupportsGLSLUBO ? "" : "UniformBuffer ",
+			g_ActiveConfig.backend_info.bSupportsPrimitiveRestart ? "" : "PrimitiveRestart ",
 			g_ogl_config.bSupportsGLPinnedMemory ? "" : "PinnedMemory ",
 			g_ogl_config.bSupportsGLSLCache ? "" : "ShaderCache ",
 			g_ogl_config.bSupportsGLBaseVertex ? "" : "BaseVertex ",
@@ -475,6 +478,20 @@ Renderer::Renderer()
 	glScissor(0, 0, GetTargetWidth(), GetTargetHeight());
 	glBlendColor(0, 0, 0, 0.5f);
 	glClearDepth(1.0f);
+	
+	if(g_ActiveConfig.backend_info.bSupportsPrimitiveRestart)
+	{
+		if(g_ogl_config.bSupportOGL31)
+		{
+			glEnable(GL_PRIMITIVE_RESTART);
+			glPrimitiveRestartIndex(65535);
+		}
+		else
+		{
+			glEnableClientState(GL_PRIMITIVE_RESTART_NV);
+			glPrimitiveRestartIndexNV(65535);
+		}
+	}
 
 	UpdateActiveConfig();
 }

Source/Plugins/Plugin_VideoOGL/Src/Render.h

@@ -24,6 +24,7 @@ extern struct VideoConfig {
 	bool bSupportCoverageMSAA;
 	bool bSupportSampleShading;
 	GLSL_VERSION eSupportedGLSLVersion;
+	bool bSupportOGL31;
 	
 	const char *gl_vendor;
 	const char *gl_renderer;

Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp

@@ -358,10 +358,10 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 				(GLfloat)targetSource.left, (GLfloat)targetSource.bottom,
 				-1.f, -1.f,
 				(GLfloat)targetSource.left, (GLfloat)targetSource.top,
-				1.f, -1.f,
-				(GLfloat)targetSource.right, (GLfloat)targetSource.top,
 				1.f, 1.f,
-				(GLfloat)targetSource.right, (GLfloat)targetSource.bottom
+				(GLfloat)targetSource.right, (GLfloat)targetSource.bottom,
+				1.f, -1.f,
+				(GLfloat)targetSource.right, (GLfloat)targetSource.top
 			};
 			
 			glBindBuffer(GL_ARRAY_BUFFER, vbo_it->second.vbo);
@@ -371,7 +371,7 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 		} 
 
 		glBindVertexArray(vbo_it->second.vao);
-		glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
+		glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 
 		glBindTexture(GL_TEXTURE_RECTANGLE_ARB, 0);
 

Source/Plugins/Plugin_VideoOGL/Src/TextureConverter.cpp

@@ -250,10 +250,10 @@ void EncodeToRamUsingShader(GLuint srcTexture, const TargetRectangle& sourceRc,
 			(float)sourceRc.left, (float)sourceRc.top,
 			-1.f, 1.f,
 			(float)sourceRc.left, (float)sourceRc.bottom,
-			1.f, 1.f,
-			(float)sourceRc.right, (float)sourceRc.bottom,
 			1.f, -1.f,
-			(float)sourceRc.right, (float)sourceRc.top
+			(float)sourceRc.right, (float)sourceRc.top,
+			1.f, 1.f,
+			(float)sourceRc.right, (float)sourceRc.bottom
 		};
 		glBindBuffer(GL_ARRAY_BUFFER, s_encode_VBO );
 		glBufferData(GL_ARRAY_BUFFER, 4*4*sizeof(GLfloat), vertices, GL_STREAM_DRAW);
@@ -262,7 +262,7 @@ void EncodeToRamUsingShader(GLuint srcTexture, const TargetRectangle& sourceRc,
 	} 
 
 	glBindVertexArray( s_encode_VAO );
-	glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
+	glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 	
 	glBindTexture(GL_TEXTURE_RECTANGLE_ARB, 0);
 	
@@ -426,10 +426,10 @@ void DecodeToTexture(u32 xfbAddr, int srcWidth, int srcHeight, GLuint destRender
 			(float)srcFmtWidth, (float)srcHeight,
 			1.f, 1.f,
 			(float)srcFmtWidth, 0.f,
-			-1.f, 1.f,
-			0.f, 0.f,
 			-1.f, -1.f,
-			0.f, (float)srcHeight
+			0.f, (float)srcHeight,
+			-1.f, 1.f,
+			0.f, 0.f
 		};
 		
 		glBindBuffer(GL_ARRAY_BUFFER, s_decode_VBO );
@@ -440,7 +440,7 @@ void DecodeToTexture(u32 xfbAddr, int srcWidth, int srcHeight, GLuint destRender
 	}
 	
 	glBindVertexArray( s_decode_VAO );
-	glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
+	glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 	
 	GL_REPORT_ERRORD();
 

Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp

@@ -124,10 +124,12 @@ void VertexManager::Draw(u32 stride)
 	u32 triangle_index_size = IndexGenerator::GetTriangleindexLen();
 	u32 line_index_size = IndexGenerator::GetLineindexLen();
 	u32 point_index_size = IndexGenerator::GetPointindexLen();
+	GLenum triangle_mode = g_ActiveConfig.backend_info.bSupportsPrimitiveRestart?GL_TRIANGLE_STRIP:GL_TRIANGLES;
+	
 	if(g_ogl_config.bSupportsGLBaseVertex) {
 		if (triangle_index_size > 0)
 		{
-			glDrawElementsBaseVertex(GL_TRIANGLES, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0], s_baseVertex);
+			glDrawElementsBaseVertex(triangle_mode, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0], s_baseVertex);
 			INCSTAT(stats.thisFrame.numIndexedDrawCalls);
 		}
 		if (line_index_size > 0)
@@ -143,7 +145,7 @@ void VertexManager::Draw(u32 stride)
 	} else {
 		if (triangle_index_size > 0)
 		{
-			glDrawElements(GL_TRIANGLES, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0]);
+			glDrawElements(triangle_mode, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0]);
 			INCSTAT(stats.thisFrame.numIndexedDrawCalls);
 		}
 		if (line_index_size > 0)

Source/Plugins/Plugin_VideoOGL/Src/main.cpp

@@ -96,6 +96,7 @@ Make AA apply instantly during gameplay if possible
 #include "PerfQuery.h"
 
 #include "VideoState.h"
+#include "IndexGenerator.h"
 #include "VideoBackend.h"
 #include "ConfigManager.h"
 
@@ -199,6 +200,7 @@ void VideoBackend::Video_Prepare()
 	g_perf_query = new PerfQuery;
 	Fifo_Init(); // must be done before OpcodeDecoder_Init()
 	OpcodeDecoder_Init();
+	IndexGenerator::Init();
 	VertexShaderManager::Init();
 	PixelShaderManager::Init();
 	ProgramShaderCache::Init();