From 3e4840df0c874753672288ffbc9046e134cf2ddf Mon Sep 17 00:00:00 2001
From: Henrik Rydgard <hrydgard@gmail.com>
Date: Thu, 10 Jan 2013 23:41:26 +0100
Subject: [PATCH 1/3] Don't lock the audio mutex if sound is off

---
 Core/HLE/__sceAudio.cpp | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

diff --git a/Core/HLE/__sceAudio.cpp b/Core/HLE/__sceAudio.cpp
index 48261e457b..89338fb447 100644
--- a/Core/HLE/__sceAudio.cpp
+++ b/Core/HLE/__sceAudio.cpp
@@ -198,9 +198,8 @@ void __AudioUpdate()
 		}
 	}
 
-	section.lock();
-
 	if (g_Config.bEnableSound) {
+		section.lock();
 		if (outAudioQueue.room() >= hwBlockSize * 2) {
 			// Push the mixed samples onto the output audio queue.
 			for (int i = 0; i < hwBlockSize; i++) {
@@ -210,14 +209,14 @@ void __AudioUpdate()
 				outAudioQueue.push((s16)sampleL);
 				outAudioQueue.push((s16)sampleR);
 			}
-		} else {
-			// This happens quite a lot. There's still something slightly off
-			// about the amount of audio we produce.
-			DEBUG_LOG(HLE, "Audio outbuffer overrun! room = %i / %i", outAudioQueue.room(), (u32)outAudioQueue.capacity());
 		}
+		section.unlock();
+	} else {
+		// This happens quite a lot. There's still something slightly off
+		// about the amount of audio we produce.
+		DEBUG_LOG(HLE, "Audio outbuffer overrun! room = %i / %i", outAudioQueue.room(), (u32)outAudioQueue.capacity());
 	}
 	
-	section.unlock();
 }
 
 void __AudioSetOutputFrequency(int freq)

From 97f3638212432756fe709b232c29a9e5c0621103 Mon Sep 17 00:00:00 2001
From: Henrik Rydgard <hrydgard@gmail.com>
Date: Thu, 10 Jan 2013 23:42:46 +0100
Subject: [PATCH 2/3] Add kernelstats Useful for finding out if some syscall
 takes unexpectedly much CPU.

---
 Core/HLE/HLE.cpp       | 13 +++++++++++++
 Core/HLE/sceKernel.cpp |  1 +
 Core/HLE/sceKernel.h   | 22 +++++++++++++++++++++-
 3 files changed, 35 insertions(+), 1 deletion(-)

diff --git a/Core/HLE/HLE.cpp b/Core/HLE/HLE.cpp
index 4fcb6b93c1..c12e7b1c63 100644
--- a/Core/HLE/HLE.cpp
+++ b/Core/HLE/HLE.cpp
@@ -15,6 +15,7 @@
 // Official git repository and contact information can be found at
 // https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
 
+#include "base/timeutil.h"
 #include "HLE.h"
 #include <map>
 #include <vector>
@@ -344,6 +345,8 @@ inline void hleFinishSyscall(int modulenum, int funcnum)
 
 void CallSyscall(u32 op)
 {
+	time_update();
+	double start = time_now_d();
 	u32 callno = (op >> 6) & 0xFFFFF; //20 bits
 	int funcnum = callno & 0xFFF;
 	int modulenum = (callno & 0xFF000) >> 12;
@@ -365,4 +368,14 @@ void CallSyscall(u32 op)
 	{
 		ERROR_LOG(HLE,"Unimplemented HLE function %s", moduleDB[modulenum].funcTable[funcnum].name);
 	}
+	time_update();
+	double total = time_now_d() - start;
+	if (total > kernelStats.slowestSyscallTime) {
+		const char *name = moduleDB[modulenum].funcTable[funcnum].name;
+		if (0 != strcmp(name, "_sceKernelIdle")) {
+			kernelStats.slowestSyscallTime = total;
+			kernelStats.slowestSyscallName = name;
+		}
+	}
+	kernelStats.msInSyscalls += total;
 }
diff --git a/Core/HLE/sceKernel.cpp b/Core/HLE/sceKernel.cpp
index 26909f5574..1dae0dd5ab 100644
--- a/Core/HLE/sceKernel.cpp
+++ b/Core/HLE/sceKernel.cpp
@@ -71,6 +71,7 @@
 
 static bool kernelRunning = false;
 KernelObjectPool kernelObjects;
+KernelStats kernelStats;
 
 void __KernelInit()
 {
diff --git a/Core/HLE/sceKernel.h b/Core/HLE/sceKernel.h
index 0b77682136..52ad5f8ca2 100644
--- a/Core/HLE/sceKernel.h
+++ b/Core/HLE/sceKernel.h
@@ -377,6 +377,7 @@ public:
 			return t;
 		}
 	}
+
 	template <class T>
 	T* GetByModuleByEntryAddr(u32 entryAddr)
 	{
@@ -410,13 +411,32 @@ public:
 	int GetCount();
 
 private:
-	enum {maxCount=4096, handleOffset=0x100};
+	enum {
+		maxCount=4096,
+		handleOffset=0x100
+	};
 	KernelObject *pool[maxCount];
 	bool occupied[maxCount];
 };
 
 extern KernelObjectPool kernelObjects;
 
+struct KernelStats {
+	void Reset() {
+		memset(this, 0, sizeof(this));
+	}
+	void ResetFrame() {
+		msInSyscalls = 0;
+		slowestSyscallTime = 0;
+		slowestSyscallName = 0;
+	}
+
+	double msInSyscalls;
+	double slowestSyscallTime;
+	const char *slowestSyscallName;
+};
+
+extern KernelStats kernelStats;
 
 void Register_ThreadManForUser();
 void Register_LoadExecForUser();

From ca0d74762dbcf26ed79722c2bc489bd50c13f32e Mon Sep 17 00:00:00 2001
From: Henrik Rydgard <hrydgard@gmail.com>
Date: Thu, 10 Jan 2013 23:49:33 +0100
Subject: [PATCH 3/3] Some realtime profiling stuff. Enable FZ (flush-to-zero)
 mode.

---
 CMakeLists.txt                      |  1 +
 Core/HLE/sceDisplay.cpp             |  8 ++++++++
 GPU/GLES/DisplayListInterpreter.cpp |  2 +-
 GPU/GLES/Framebuffer.cpp            |  3 ++-
 GPU/GLES/TextureCache.cpp           | 12 +++++++++++-
 GPU/GLES/TextureCache.h             |  1 +
 GPU/GPUCommon.cpp                   |  8 ++++++--
 GPU/GPUState.h                      |  2 ++
 android/jni/EmuScreen.cpp           |  1 +
 android/jni/NativeApp.cpp           |  7 +++++++
 native                              |  2 +-
 11 files changed, 41 insertions(+), 6 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9c85ba7a16..fcdda619e8 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -112,6 +112,7 @@ if(NOT MSVC)
 	add_definitions(-O2)
 	add_definitions(-Wno-multichar)
 	add_definitions(-fno-strict-aliasing)
+	add_definitions(-ffast-math)
 	if(NOT APPLE)
 		add_definitions(-Wno-psabi)
 		add_definitions(-D_XOPEN_SOURCE=600 -D_XOPEN_SOURCE_EXTENDED -D__BSD_VISIBLE=1)
diff --git a/Core/HLE/sceDisplay.cpp b/Core/HLE/sceDisplay.cpp
index e85be05e2c..d047027cdc 100644
--- a/Core/HLE/sceDisplay.cpp
+++ b/Core/HLE/sceDisplay.cpp
@@ -211,6 +211,9 @@ void hleEnterVblank(u64 userdata, int cyclesLate) {
 		char stats[512];
 		sprintf(stats,
 			"Frames: %i\n"
+			"DL processing time: %0.2f ms\n"
+			"Kernel processing time: %0.2f ms\n"
+			"Slowest syscall: %s : %0.2f ms\n"
 			"Draw calls: %i\n"
 			"Draw flushes: %i\n"
 			"Vertices Transformed: %i\n"
@@ -221,6 +224,10 @@ void hleEnterVblank(u64 userdata, int cyclesLate) {
 			"Fragment shaders loaded: %i\n"
 			"Combined shaders loaded: %i\n",
 			gpuStats.numFrames,
+			gpuStats.msProcessingDisplayLists * 1000.0f,
+			kernelStats.msInSyscalls * 1000.0f,
+			kernelStats.slowestSyscallName ? kernelStats.slowestSyscallName : "(none)",
+			kernelStats.slowestSyscallTime * 1000.0f,
 			gpuStats.numDrawCalls,
 			gpuStats.numFlushes,
 			gpuStats.numVertsTransformed,
@@ -238,6 +245,7 @@ void hleEnterVblank(u64 userdata, int cyclesLate) {
 		PPGeEnd();
 
 		gpuStats.resetFrame();
+		kernelStats.ResetFrame();
 	}
 
 	host->EndFrame();
diff --git a/GPU/GLES/DisplayListInterpreter.cpp b/GPU/GLES/DisplayListInterpreter.cpp
index f1a9d1030d..2d7aee60e2 100644
--- a/GPU/GLES/DisplayListInterpreter.cpp
+++ b/GPU/GLES/DisplayListInterpreter.cpp
@@ -199,7 +199,7 @@ void GLES_GPU::DumpNextFrame() {
 }
 
 void GLES_GPU::BeginFrame() {
-	TextureCache_Decimate();
+	TextureCache_StartFrame();
 
 	if (dumpNextFrame_) {
 		NOTICE_LOG(G3D, "DUMPING THIS FRAME");
diff --git a/GPU/GLES/Framebuffer.cpp b/GPU/GLES/Framebuffer.cpp
index f3a6519fa0..963560db1b 100644
--- a/GPU/GLES/Framebuffer.cpp
+++ b/GPU/GLES/Framebuffer.cpp
@@ -166,7 +166,8 @@ void FramebufferManager::DrawActiveTexture(float w, float h, bool flip) {
 	Matrix4x4 ortho;
 	ortho.setOrtho(0, 480, 272, 0, -1, 1);
 	glUniformMatrix4fv(draw2dprogram->u_viewproj, 1, GL_FALSE, ortho.getReadPtr());
-
+	glBindBuffer(GL_ARRAY_BUFFER, 0);
+	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0);
 	glEnableVertexAttribArray(draw2dprogram->a_position);
 	glEnableVertexAttribArray(draw2dprogram->a_texcoord0);
 	glVertexAttribPointer(draw2dprogram->a_position, 3, GL_FLOAT, GL_FALSE, 12, pos);
diff --git a/GPU/GLES/TextureCache.cpp b/GPU/GLES/TextureCache.cpp
index 534564187b..14aec11244 100644
--- a/GPU/GLES/TextureCache.cpp
+++ b/GPU/GLES/TextureCache.cpp
@@ -598,6 +598,13 @@ void convertColors(u8 *finalBuf, GLuint dstFmt, int numPixels) {
 	}
 }
 
+int lastBoundTexture = -1;
+
+void TextureCache_StartFrame() {
+	lastBoundTexture = -1;
+	TextureCache_Decimate();
+}
+
 void PSPSetTexture() {
 	u32 texaddr = (gstate.texaddr[0] & 0xFFFFF0) | ((gstate.texbufwidth[0]<<8) & 0xFF000000);
 	texaddr &= 0xFFFFFFF;
@@ -656,7 +663,10 @@ void PSPSetTexture() {
 		if (match) {
 			//got one!
 			entry.frameCounter = gpuStats.numFrames;
-			glBindTexture(GL_TEXTURE_2D, entry.texture);
+			if (entry.texture != lastBoundTexture) {
+				glBindTexture(GL_TEXTURE_2D, entry.texture);
+				lastBoundTexture = entry.texture;
+			}
 			UpdateSamplingParams(entry, false);
 			DEBUG_LOG(G3D, "Texture at %08x Found in Cache, applying", texaddr);
 			return; //Done!
diff --git a/GPU/GLES/TextureCache.h b/GPU/GLES/TextureCache.h
index 7be01ad2e3..a9e062170f 100644
--- a/GPU/GLES/TextureCache.h
+++ b/GPU/GLES/TextureCache.h
@@ -24,6 +24,7 @@ void PSPSetTexture();
 void TextureCache_Init();
 void TextureCache_Shutdown();
 void TextureCache_Clear(bool delete_them);
+void TextureCache_StartFrame();
 void TextureCache_Decimate();  // Run this once per frame to get rid of old textures.
 void TextureCache_Invalidate(u32 addr, int size, bool force);
 void TextureCache_InvalidateAll(bool force);
diff --git a/GPU/GPUCommon.cpp b/GPU/GPUCommon.cpp
index bb877840cf..b55554aa9e 100644
--- a/GPU/GPUCommon.cpp
+++ b/GPU/GPUCommon.cpp
@@ -1,3 +1,4 @@
+#include "base/timeutil.h"
 #include "../Core/MemMap.h"
 #include "GeDisasm.h"
 #include "GPUCommon.h"
@@ -41,7 +42,6 @@ u32 GPUCommon::EnqueueList(u32 listpc, u32 stall, int subIntrBase, bool head)
 
 void GPUCommon::UpdateStall(int listid, u32 newstall)
 {
-	
 	for (auto iter = dlQueue.begin(); iter != dlQueue.end(); ++iter)
 	{
 		DisplayList &cur = *iter;
@@ -56,6 +56,8 @@ void GPUCommon::UpdateStall(int listid, u32 newstall)
 
 bool GPUCommon::InterpretList(DisplayList &list)
 {
+	time_update();
+	double start = time_now_d();
 	currentList = &list;
 	// Reset stackptr for safety
 	stackptr = 0;
@@ -91,13 +93,15 @@ bool GPUCommon::InterpretList(DisplayList &list)
 		list.pc += 4;
 		prev = op;
 	}
+	time_update();
+	gpuStats.msProcessingDisplayLists += time_now_d() - start;
 	return true;
 }
 
 bool GPUCommon::ProcessDLQueue()
 {
 	DisplayListQueue::iterator iter = dlQueue.begin();
-	while (!(iter == dlQueue.end()))
+	while (iter != dlQueue.end())
 	{
 		DisplayList &l = *iter;
 		DEBUG_LOG(G3D,"Okay, starting DL execution at %08x - stall = %08x", l.pc, l.stall);
diff --git a/GPU/GPUState.h b/GPU/GPUState.h
index 4b3a58a6a9..dfbd97dec8 100644
--- a/GPU/GPUState.h
+++ b/GPU/GPUState.h
@@ -263,6 +263,7 @@ struct GPUStatistics
 		numShaderSwitches = 0;
 		numFlushes = 0;
 		numTexturesDecoded = 0;
+		msProcessingDisplayLists = 0;
 	}
 
 	// Per frame statistics
@@ -274,6 +275,7 @@ struct GPUStatistics
 	int numTextureSwitches;
 	int numShaderSwitches;
 	int numTexturesDecoded;
+	double msProcessingDisplayLists;
 
 	// Total statistics, updated by the GPU core in UpdateStats
 	int numFrames;
diff --git a/android/jni/EmuScreen.cpp b/android/jni/EmuScreen.cpp
index 0dffb2548f..623bec0621 100644
--- a/android/jni/EmuScreen.cpp
+++ b/android/jni/EmuScreen.cpp
@@ -41,6 +41,7 @@
 
 EmuScreen::EmuScreen(const std::string &filename) : invalid_(true)
 {
+	CheckGLExtensions();
 	std::string fileToStart = filename;
 	// This is probably where we should start up the emulated PSP.
 	INFO_LOG(BOOT, "Starting up hardware.");
diff --git a/android/jni/NativeApp.cpp b/android/jni/NativeApp.cpp
index 9faf23f009..4877028375 100644
--- a/android/jni/NativeApp.cpp
+++ b/android/jni/NativeApp.cpp
@@ -31,6 +31,7 @@
 #include "gfx/gl_lost_manager.h"
 #include "gfx/texture.h"
 #include "input/input_state.h"
+#include "math/math_util.h"
 #include "math/lin/matrix4x4.h"
 #include "ui/screen.h"
 #include "ui/ui.h"
@@ -155,6 +156,7 @@ void NativeGetAppInfo(std::string *app_dir_name, std::string *app_nice_name, boo
 
 void NativeInit(int argc, const char *argv[], const char *savegame_directory, const char *external_directory, const char *installID)
 {
+	EnableFZ();
 	std::string user_data_path = savegame_directory;
 
 	// We want this to be FIRST.
@@ -308,6 +310,11 @@ void NativeInitGraphics()
 
 void NativeRender()
 {
+	EnableFZ();
+	// Clearing the screen at the start of the frame is an optimization for tiled mobile GPUs, as it then doesn't need to keep it around between frames.
+	glClearColor(0,0,0,1);
+	glClear(GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+
 	glstate.Restore();
 	glViewport(0, 0, pixel_xres, pixel_yres);
 	Matrix4x4 ortho;
diff --git a/native b/native
index c63061ddfd..7ceecd22b4 160000
--- a/native
+++ b/native
@@ -1 +1 @@
-Subproject commit c63061ddfd5aec3b9dd51aa4c71150de905d8d1e
+Subproject commit 7ceecd22b421c697d4d8c8ad6bda7654226be8ca