Merged lindenlab/viewer-release into default

14 files changed, 667 insertions, 68 deletions

diff --git a/indra/llrender/CMakeLists.txt b/indra/llrender/CMakeLists.txt
index 669b70aa43..dba12d048e 100755
--- a/indra/llrender/CMakeLists.txt
+++ b/indra/llrender/CMakeLists.txt
@@ -92,7 +92,7 @@ if (BUILD_HEADLESS)
 
   set_property(TARGET llrenderheadless
     PROPERTY COMPILE_DEFINITIONS LL_MESA=1 LL_MESA_HEADLESS=1
-    )
+      )
 
   target_link_libraries(llrenderheadless
     ${LLCOMMON_LIBRARIES}
diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 6c798aa4ed..088ba95b75 100755
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -86,7 +86,7 @@ void APIENTRY gl_debug_callback(GLenum source,
 	}
 	else
 	{
-		llwarns << "----- GL WARNING -------" << llendl;		
+		llwarns << "----- GL WARNING -------" << llendl;
 	}
 	llwarns << "Type: " << std::hex << type << llendl;
 	llwarns << "ID: " << std::hex << id << llendl;
@@ -216,6 +216,11 @@ PFNGLGETQUERYIVARBPROC glGetQueryivARB = NULL;
 PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB = NULL;
 PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB = NULL;
 
+// GL_ARB_timer_query
+PFNGLQUERYCOUNTERPROC glQueryCounter = NULL;
+PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v = NULL;
+PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v = NULL;
+
 // GL_ARB_point_parameters
 PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB = NULL;
 PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB = NULL;
@@ -421,6 +426,7 @@ LLGLManager::LLGLManager() :
 	mHasFragmentShader(FALSE),
 	mNumTextureImageUnits(0),
 	mHasOcclusionQuery(FALSE),
+	mHasTimerQuery(FALSE),
 	mHasOcclusionQuery2(FALSE),
 	mHasPointParameters(FALSE),
 	mHasDrawBuffers(FALSE),
@@ -445,7 +451,9 @@ LLGLManager::LLGLManager() :
 	mIsGFFX(FALSE),
 	mATIOffsetVerticalLines(FALSE),
 	mATIOldDriver(FALSE),
-
+#if LL_DARWIN
+	mIsMobileGF(FALSE),
+#endif
 	mHasRequirements(TRUE),
 
 	mHasSeparateSpecularColor(FALSE),
@@ -637,6 +645,13 @@ bool LLGLManager::initGL()
 		{
 			mIsGF3 = TRUE;
 		}
+#if LL_DARWIN
+		else if ((mGLRenderer.find("9400M") != std::string::npos)
+			  || (mGLRenderer.find("9600M") != std::string::npos))
+		{
+			mIsMobileGF = TRUE;
+		}
+#endif
 
 	}
 	else if (mGLVendor.find("INTEL") != std::string::npos
@@ -745,12 +760,13 @@ bool LLGLManager::initGL()
 	{ //using multisample textures on ATI results in black screen for some reason
 		mHasTextureMultisample = FALSE;
 	}
-#endif
+
 
 	if (mIsIntel && mGLVersion <= 3.f)
 	{ //never try to use framebuffer objects on older intel drivers (crashy)
 		mHasFramebufferObject = FALSE;
 	}
+#endif
 
 	if (mHasFramebufferObject)
 	{
@@ -947,13 +963,15 @@ void LLGLManager::initExtensions()
 	mHasARBEnvCombine = ExtensionExists("GL_ARB_texture_env_combine", gGLHExts.mSysExts);
 	mHasCompressedTextures = glh_init_extensions("GL_ARB_texture_compression");
 	mHasOcclusionQuery = ExtensionExists("GL_ARB_occlusion_query", gGLHExts.mSysExts);
+	mHasTimerQuery = ExtensionExists("GL_ARB_timer_query", gGLHExts.mSysExts);
 	mHasOcclusionQuery2 = ExtensionExists("GL_ARB_occlusion_query2", gGLHExts.mSysExts);
 	mHasVertexBufferObject = ExtensionExists("GL_ARB_vertex_buffer_object", gGLHExts.mSysExts);
 	mHasVertexArrayObject = ExtensionExists("GL_ARB_vertex_array_object", gGLHExts.mSysExts);
 	mHasSync = ExtensionExists("GL_ARB_sync", gGLHExts.mSysExts);
 	mHasMapBufferRange = ExtensionExists("GL_ARB_map_buffer_range", gGLHExts.mSysExts);
 	mHasFlushBufferRange = ExtensionExists("GL_APPLE_flush_buffer_range", gGLHExts.mSysExts);
-	mHasDepthClamp = ExtensionExists("GL_ARB_depth_clamp", gGLHExts.mSysExts) || ExtensionExists("GL_NV_depth_clamp", gGLHExts.mSysExts);
+	//mHasDepthClamp = ExtensionExists("GL_ARB_depth_clamp", gGLHExts.mSysExts) || ExtensionExists("GL_NV_depth_clamp", gGLHExts.mSysExts);
+	mHasDepthClamp = FALSE;
 	// mask out FBO support when packed_depth_stencil isn't there 'cause we need it for LLRenderTarget -Brad
 #ifdef GL_ARB_framebuffer_object
 	mHasFramebufferObject = ExtensionExists("GL_ARB_framebuffer_object", gGLHExts.mSysExts);
@@ -963,6 +981,15 @@ void LLGLManager::initExtensions()
 							ExtensionExists("GL_EXT_framebuffer_multisample", gGLHExts.mSysExts) &&
 							ExtensionExists("GL_EXT_packed_depth_stencil", gGLHExts.mSysExts);
 #endif
+#ifdef GL_EXT_texture_sRGB
+	mHassRGBTexture = ExtensionExists("GL_EXT_texture_sRGB", gGLHExts.mSysExts);
+#endif
+	
+#ifdef GL_ARB_framebuffer_sRGB
+	mHassRGBFramebuffer = ExtensionExists("GL_ARB_framebuffer_sRGB", gGLHExts.mSysExts);
+#else
+	mHassRGBFramebuffer = ExtensionExists("GL_EXT_framebuffer_sRGB", gGLHExts.mSysExts);
+#endif
 	
 	mHasMipMapGeneration = mHasFramebufferObject || mGLVersion >= 1.4f;
 
@@ -1253,6 +1280,13 @@ void LLGLManager::initExtensions()
 		glGetQueryObjectivARB = (PFNGLGETQUERYOBJECTIVARBPROC)GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjectivARB");
 		glGetQueryObjectuivARB = (PFNGLGETQUERYOBJECTUIVARBPROC)GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjectuivARB");
 	}
+	if (mHasTimerQuery)
+	{
+		llinfos << "initExtensions() TimerQuery-related procs..." << llendl;
+		glQueryCounter = (PFNGLQUERYCOUNTERPROC) GLH_EXT_GET_PROC_ADDRESS("glQueryCounter");
+		glGetQueryObjecti64v = (PFNGLGETQUERYOBJECTI64VPROC) GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjecti64v");
+		glGetQueryObjectui64v = (PFNGLGETQUERYOBJECTUI64VPROC) GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjectui64v");
+	}
 	if (mHasPointParameters)
 	{
 		llinfos << "initExtensions() PointParameters-related procs..." << llendl;
diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index d70e764769..60597fd090 100755
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -98,6 +98,7 @@ public:
 	BOOL mHasFragmentShader;
 	S32  mNumTextureImageUnits;
 	BOOL mHasOcclusionQuery;
+	BOOL mHasTimerQuery;
 	BOOL mHasOcclusionQuery2;
 	BOOL mHasPointParameters;
 	BOOL mHasDrawBuffers;
@@ -115,6 +116,8 @@ public:
 	BOOL mHasARBEnvCombine;
 	BOOL mHasCubeMap;
 	BOOL mHasDebugOutput;
+	BOOL mHassRGBTexture;
+	BOOL mHassRGBFramebuffer;
 
 	// Vendor-specific extensions
 	BOOL mIsATI;
@@ -126,6 +129,11 @@ public:
 	BOOL mATIOffsetVerticalLines;
 	BOOL mATIOldDriver;
 
+#if LL_DARWIN
+	// Needed to distinguish problem cards on older Macs that break with Materials
+	BOOL mIsMobileGF;
+#endif
+	
 	// Whether this version of GL is good enough for SL to use
 	BOOL mHasRequirements;
 
diff --git a/indra/llrender/llglheaders.h b/indra/llrender/llglheaders.h
index 509de51f4d..dace572953 100755
--- a/indra/llrender/llglheaders.h
+++ b/indra/llrender/llglheaders.h
@@ -116,6 +116,11 @@ extern PFNGLGETQUERYIVARBPROC glGetQueryivARB;
 extern PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB;
 extern PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB;
 
+// GL_ARB_timer_query
+extern PFNGLQUERYCOUNTERPROC glQueryCounter;
+extern PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v;
+extern PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v;
+
 // GL_ARB_point_parameters
 extern PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB;
 extern PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB;
@@ -378,6 +383,11 @@ extern PFNGLGETQUERYIVARBPROC glGetQueryivARB;
 extern PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB;
 extern PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB;
 
+// GL_ARB_timer_query
+extern PFNGLQUERYCOUNTERPROC glQueryCounter;
+extern PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v;
+extern PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v;
+
 // GL_ARB_point_parameters
 extern PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB;
 extern PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB;
@@ -619,6 +629,12 @@ extern PFNGLGETQUERYIVARBPROC glGetQueryivARB;
 extern PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB;
 extern PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB;
 
+// GL_ARB_timer_query
+extern PFNGLQUERYCOUNTERPROC glQueryCounter;
+extern PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v;
+extern PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v;
+
+
 // GL_ARB_point_parameters
 extern PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB;
 extern PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB;
diff --git a/indra/llrender/llglslshader.cpp b/indra/llrender/llglslshader.cpp
index 7cbf39096e..ac16e30796 100755
--- a/indra/llrender/llglslshader.cpp
+++ b/indra/llrender/llglslshader.cpp
@@ -53,6 +53,12 @@ GLhandleARB LLGLSLShader::sCurBoundShader = 0;
 LLGLSLShader* LLGLSLShader::sCurBoundShaderPtr = NULL;
 S32 LLGLSLShader::sIndexedTextureChannels = 0;
 bool LLGLSLShader::sNoFixedFunction = false;
+bool LLGLSLShader::sProfileEnabled = false;
+std::set<LLGLSLShader*> LLGLSLShader::sInstances;
+U64 LLGLSLShader::sTotalTimeElapsed = 0;
+U32 LLGLSLShader::sTotalTrianglesDrawn = 0;
+U64 LLGLSLShader::sTotalSamplesDrawn = 0;
+U32 LLGLSLShader::sTotalDrawCalls = 0;
 
 //UI shader -- declared here so llui_libtest will link properly
 LLGLSLShader	gUIProgram;
@@ -87,19 +93,240 @@ LLShaderFeatures::LLShaderFeatures()
 //===============================
 // LLGLSL Shader implementation
 //===============================
+
+//static
+void LLGLSLShader::initProfile()
+{
+	sProfileEnabled = true;
+	sTotalTimeElapsed = 0;
+	sTotalTrianglesDrawn = 0;
+	sTotalSamplesDrawn = 0;
+	sTotalDrawCalls = 0;
+
+	for (std::set<LLGLSLShader*>::iterator iter = sInstances.begin(); iter != sInstances.end(); ++iter)
+	{
+		(*iter)->clearStats();
+	}
+}
+
+
+struct LLGLSLShaderCompareTimeElapsed
+{
+		bool operator()(const LLGLSLShader* const& lhs, const LLGLSLShader* const& rhs)
+		{
+			return lhs->mTimeElapsed < rhs->mTimeElapsed;
+		}
+};
+
+//static
+void LLGLSLShader::finishProfile()
+{
+	sProfileEnabled = false;
+
+	std::vector<LLGLSLShader*> sorted;
+
+	for (std::set<LLGLSLShader*>::iterator iter = sInstances.begin(); iter != sInstances.end(); ++iter)
+	{
+		sorted.push_back(*iter);
+	}
+
+	std::sort(sorted.begin(), sorted.end(), LLGLSLShaderCompareTimeElapsed());
+
+	for (std::vector<LLGLSLShader*>::iterator iter = sorted.begin(); iter != sorted.end(); ++iter)
+	{
+		(*iter)->dumpStats();
+	}
+
+	llinfos << "-----------------------------------" << llendl;
+	llinfos << "Total rendering time: " << llformat("%.4f ms", sTotalTimeElapsed/1000000.f) << llendl;
+	llinfos << "Total samples drawn: " << llformat("%.4f million", sTotalSamplesDrawn/1000000.f) << llendl;
+	llinfos << "Total triangles drawn: " << llformat("%.3f million", sTotalTrianglesDrawn/1000000.f) << llendl;
+}
+
+void LLGLSLShader::clearStats()
+{
+	mTrianglesDrawn = 0;
+	mTimeElapsed = 0;
+	mSamplesDrawn = 0;
+	mDrawCalls = 0;
+	mTextureStateFetched = false;
+	mTextureMagFilter.clear();
+	mTextureMinFilter.clear();
+}
+
+void LLGLSLShader::dumpStats()
+{
+	if (mDrawCalls > 0)
+	{
+		llinfos << "=============================================" << llendl;
+		llinfos << mName << llendl;
+		for (U32 i = 0; i < mShaderFiles.size(); ++i)
+		{
+			llinfos << mShaderFiles[i].first << llendl;
+		}
+		for (U32 i = 0; i < mTexture.size(); ++i)
+		{
+			GLint idx = mTexture[i];
+			
+			if (idx >= 0)
+			{
+				GLint uniform_idx = getUniformLocation(i);
+				llinfos << mUniformNameMap[uniform_idx] << " - " << std::hex << mTextureMagFilter[i] << "/" << mTextureMinFilter[i] << std::dec << llendl;
+			}
+		}
+		llinfos << "=============================================" << llendl;
+
+		F32 ms = mTimeElapsed/1000000.f;
+		F32 seconds = ms/1000.f;
+
+		F32 pct_tris = (F32) mTrianglesDrawn/(F32)sTotalTrianglesDrawn*100.f;
+		F32 tris_sec = (F32) (mTrianglesDrawn/1000000.0);
+		tris_sec /= seconds;
+
+		F32 pct_samples = (F32) ((F64)mSamplesDrawn/(F64)sTotalSamplesDrawn)*100.f;
+		F32 samples_sec = (F32) mSamplesDrawn/1000000000.0;
+		samples_sec /= seconds;
+
+		F32 pct_calls = (F32) mDrawCalls/(F32)sTotalDrawCalls*100.f;
+		U32 avg_batch = mTrianglesDrawn/mDrawCalls;
+
+		llinfos << "Triangles Drawn: " << mTrianglesDrawn <<  " " << llformat("(%.2f pct of total, %.3f million/sec)", pct_tris, tris_sec ) << llendl;
+		llinfos << "Draw Calls: " << mDrawCalls << " " << llformat("(%.2f pct of total, avg %d tris/call)", pct_calls, avg_batch) << llendl;
+		llinfos << "SamplesDrawn: " << mSamplesDrawn << " " << llformat("(%.2f pct of total, %.3f billion/sec)", pct_samples, samples_sec) << llendl;
+		llinfos << "Time Elapsed: " << mTimeElapsed << " " << llformat("(%.2f pct of total, %.5f ms)\n", (F32) ((F64)mTimeElapsed/(F64)sTotalTimeElapsed)*100.f, ms) << llendl;
+	}
+}
+
+//static
+void LLGLSLShader::startProfile()
+{
+	if (sProfileEnabled && sCurBoundShaderPtr)
+	{
+		sCurBoundShaderPtr->placeProfileQuery();
+	}
+
+}
+
+//static
+void LLGLSLShader::stopProfile(U32 count, U32 mode)
+{
+	if (sProfileEnabled)
+	{
+		sCurBoundShaderPtr->readProfileQuery(count, mode);
+	}
+}
+
+void LLGLSLShader::placeProfileQuery()
+{
+#if !LL_DARWIN
+	if (mTimerQuery == 0)
+	{
+		glGenQueriesARB(1, &mTimerQuery);
+	}
+
+	if (!mTextureStateFetched)
+	{
+		mTextureStateFetched = true;
+		mTextureMagFilter.resize(mTexture.size());
+		mTextureMinFilter.resize(mTexture.size());
+
+		U32 cur_active = gGL.getCurrentTexUnitIndex();
+
+		for (U32 i = 0; i < mTexture.size(); ++i)
+		{
+			GLint idx = mTexture[i];
+
+			if (idx >= 0)
+			{
+				gGL.getTexUnit(idx)->activate();
+
+				U32 mag = 0xFFFFFFFF;
+				U32 min = 0xFFFFFFFF;
+
+				U32 type = LLTexUnit::getInternalType(gGL.getTexUnit(idx)->getCurrType());
+
+				glGetTexParameteriv(type, GL_TEXTURE_MAG_FILTER, (GLint*) &mag);
+				glGetTexParameteriv(type, GL_TEXTURE_MIN_FILTER, (GLint*) &min);
+
+				mTextureMagFilter[i] = mag;
+				mTextureMinFilter[i] = min;
+			}
+		}
+
+		gGL.getTexUnit(cur_active)->activate();
+	}
+
+
+	glBeginQueryARB(GL_SAMPLES_PASSED, 1);
+	glBeginQueryARB(GL_TIME_ELAPSED, mTimerQuery);
+#endif
+}
+
+void LLGLSLShader::readProfileQuery(U32 count, U32 mode)
+{
+#if !LL_DARWIN
+	glEndQueryARB(GL_TIME_ELAPSED);
+	glEndQueryARB(GL_SAMPLES_PASSED);
+	
+	U64 time_elapsed = 0;
+	glGetQueryObjectui64v(mTimerQuery, GL_QUERY_RESULT, &time_elapsed);
+
+	U64 samples_passed = 0;
+	glGetQueryObjectui64v(1, GL_QUERY_RESULT, &samples_passed);
+
+	sTotalTimeElapsed += time_elapsed;
+	mTimeElapsed += time_elapsed;
+
+	sTotalSamplesDrawn += samples_passed;
+	mSamplesDrawn += samples_passed;
+
+	U32 tri_count = 0;
+	switch (mode)
+	{
+		case LLRender::TRIANGLES: tri_count = count/3; break;
+		case LLRender::TRIANGLE_FAN: tri_count = count-2; break;
+		case LLRender::TRIANGLE_STRIP: tri_count = count-2; break;
+		default: tri_count = count; break; //points lines etc just use primitive count
+	}
+
+	mTrianglesDrawn += tri_count;
+	sTotalTrianglesDrawn += tri_count;
+
+	sTotalDrawCalls++;
+	mDrawCalls++;
+#endif
+}
+
+
+
 LLGLSLShader::LLGLSLShader()
-	: mProgramObject(0), mActiveTextureChannels(0), mShaderLevel(0), mShaderGroup(SG_DEFAULT), mUniformsDirty(FALSE)
+	: mProgramObject(0), 
+	  mAttributeMask(0),
+	  mTotalUniformSize(0),
+	  mActiveTextureChannels(0), 
+	  mShaderLevel(0), 
+	  mShaderGroup(SG_DEFAULT), 
+	  mUniformsDirty(FALSE),
+	  mTimerQuery(0)
 {
+	
+}
 
+LLGLSLShader::~LLGLSLShader()
+{
+	
 }
 
 void LLGLSLShader::unload()
 {
+	sInstances.erase(this);
+
 	stop_glerror();
 	mAttribute.clear();
 	mTexture.clear();
 	mUniform.clear();
 	mShaderFiles.clear();
+	mDefines.clear();
 
 	if (mProgramObject)
 	{
@@ -133,6 +360,8 @@ BOOL LLGLSLShader::createShader(vector<string> * attributes,
 								U32 varying_count,
 								const char** varyings)
 {
+	sInstances.insert(this);
+
 	//reloading, reset matrix hash values
 	for (U32 i = 0; i < LLRender::NUM_MATRIX_MODES; ++i)
 	{
@@ -150,7 +379,7 @@ BOOL LLGLSLShader::createShader(vector<string> * attributes,
 	vector< pair<string,GLenum> >::iterator fileIter = mShaderFiles.begin();
 	for ( ; fileIter != mShaderFiles.end(); fileIter++ )
 	{
-		GLhandleARB shaderhandle = LLShaderMgr::instance()->loadShaderFile((*fileIter).first, mShaderLevel, (*fileIter).second, mFeatures.mIndexedTextureChannels);
+		GLhandleARB shaderhandle = LLShaderMgr::instance()->loadShaderFile((*fileIter).first, mShaderLevel, (*fileIter).second, &mDefines, mFeatures.mIndexedTextureChannels);
 		LL_DEBUGS("ShaderLoading") << "SHADER FILE: " << (*fileIter).first << " mShaderLevel=" << mShaderLevel << LL_ENDL;
 		if (shaderhandle > 0)
 		{
@@ -285,6 +514,8 @@ BOOL LLGLSLShader::mapAttributes(const vector<string> * attributes)
 	if (res)
 	{ //read back channel locations
 
+		mAttributeMask = 0;
+
 		//read back reserved channels first
 		for (U32 i = 0; i < LLShaderMgr::instance()->mReservedAttribs.size(); i++)
 		{
@@ -293,6 +524,7 @@ BOOL LLGLSLShader::mapAttributes(const vector<string> * attributes)
 			if (index != -1)
 			{
 				mAttribute[i] = index;
+				mAttributeMask |= 1 << i;
 				LL_DEBUGS("ShaderLoading") << "Attribute " << name << " assigned to channel " << index << LL_ENDL;
 			}
 		}
@@ -325,11 +557,56 @@ void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
 
 	GLenum type;
 	GLsizei length;
-	GLint size;
+	GLint size = -1;
 	char name[1024];		/* Flawfinder: ignore */
 	name[0] = 0;
 
+
 	glGetActiveUniformARB(mProgramObject, index, 1024, &length, &size, &type, (GLcharARB *)name);
+#if !LL_DARWIN
+	if (size > 0)
+	{
+		switch(type)
+		{
+			case GL_FLOAT_VEC2: size *= 2; break;
+			case GL_FLOAT_VEC3: size *= 3; break;
+			case GL_FLOAT_VEC4: size *= 4; break;
+			case GL_DOUBLE: size *= 2; break;
+			case GL_DOUBLE_VEC2: size *= 2; break;
+			case GL_DOUBLE_VEC3: size *= 6; break;
+			case GL_DOUBLE_VEC4: size *= 8; break;
+			case GL_INT_VEC2: size *= 2; break;
+			case GL_INT_VEC3: size *= 3; break;
+			case GL_INT_VEC4: size *= 4; break;
+			case GL_UNSIGNED_INT_VEC2: size *= 2; break;
+			case GL_UNSIGNED_INT_VEC3: size *= 3; break;
+			case GL_UNSIGNED_INT_VEC4: size *= 4; break;
+			case GL_BOOL_VEC2: size *= 2; break;
+			case GL_BOOL_VEC3: size *= 3; break;
+			case GL_BOOL_VEC4: size *= 4; break;
+			case GL_FLOAT_MAT2: size *= 4; break;
+			case GL_FLOAT_MAT3: size *= 9; break;
+			case GL_FLOAT_MAT4: size *= 16; break;
+			case GL_FLOAT_MAT2x3: size *= 6; break;
+			case GL_FLOAT_MAT2x4: size *= 8; break;
+			case GL_FLOAT_MAT3x2: size *= 6; break;
+			case GL_FLOAT_MAT3x4: size *= 12; break;
+			case GL_FLOAT_MAT4x2: size *= 8; break;
+			case GL_FLOAT_MAT4x3: size *= 12; break;
+			case GL_DOUBLE_MAT2: size *= 8; break;
+			case GL_DOUBLE_MAT3: size *= 18; break;
+			case GL_DOUBLE_MAT4: size *= 32; break;
+			case GL_DOUBLE_MAT2x3: size *= 12; break;
+			case GL_DOUBLE_MAT2x4: size *= 16; break;
+			case GL_DOUBLE_MAT3x2: size *= 12; break;
+			case GL_DOUBLE_MAT3x4: size *= 24; break;
+			case GL_DOUBLE_MAT4x2: size *= 16; break;
+			case GL_DOUBLE_MAT4x3: size *= 24; break;
+		}
+		mTotalUniformSize += size;
+	}
+#endif
+
 	S32 location = glGetUniformLocationARB(mProgramObject, name);
 	if (location != -1)
 	{
@@ -342,6 +619,7 @@ void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
 		}
 
 		mUniformMap[name] = location;
+		mUniformNameMap[location] = name;
 		LL_DEBUGS("ShaderLoading") << "Uniform " << name << " is at location " << location << LL_ENDL;
 	
 		//find the index of this uniform
@@ -372,11 +650,21 @@ void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
 			}
 		}
 	}
- }
+}
+
+void LLGLSLShader::addPermutation(std::string name, std::string value)
+{
+	mDefines[name] = value;
+}
+
+void LLGLSLShader::removePermutation(std::string name)
+{
+	mDefines[name].erase();
+}
 
 GLint LLGLSLShader::mapUniformTextureChannel(GLint location, GLenum type)
 {
-	if (type >= GL_SAMPLER_1D_ARB && type <= GL_SAMPLER_2D_RECT_SHADOW_ARB ||
+	if ((type >= GL_SAMPLER_1D_ARB && type <= GL_SAMPLER_2D_RECT_SHADOW_ARB) ||
 		type == GL_SAMPLER_2D_MULTISAMPLE)
 	{	//this here is a texture
 		glUniform1iARB(location, mActiveTextureChannels);
@@ -390,9 +678,11 @@ BOOL LLGLSLShader::mapUniforms(const vector<string> * uniforms)
 {
 	BOOL res = TRUE;
 	
+	mTotalUniformSize = 0;
 	mActiveTextureChannels = 0;
 	mUniform.clear();
 	mUniformMap.clear();
+	mUniformNameMap.clear();
 	mTexture.clear();
 	mValue.clear();
 	//initialize arrays
@@ -413,6 +703,7 @@ BOOL LLGLSLShader::mapUniforms(const vector<string> * uniforms)
 
 	unbind();
 
+	LL_DEBUGS("ShaderLoading") << "Total Uniform Size: " << mTotalUniformSize << llendl;
 	return res;
 }
 
@@ -471,6 +762,58 @@ void LLGLSLShader::bindNoShader(void)
 	}
 }
 
+S32 LLGLSLShader::bindTexture(const std::string &uniform, LLTexture *texture, LLTexUnit::eTextureType mode)
+{
+	S32 channel = 0;
+	channel = getUniformLocation(uniform);
+	
+	return bindTexture(channel, texture, mode);
+}
+
+S32 LLGLSLShader::bindTexture(S32 uniform, LLTexture *texture, LLTexUnit::eTextureType mode)
+{
+	if (uniform < 0 || uniform >= (S32)mTexture.size())
+	{
+		UNIFORM_ERRS << "Uniform out of range: " << uniform << LL_ENDL;
+		return -1;
+	}
+	
+	uniform = mTexture[uniform];
+	
+	if (uniform > -1)
+	{
+		gGL.getTexUnit(uniform)->bind(texture, mode);
+	}
+	
+	return uniform;
+}
+
+S32 LLGLSLShader::unbindTexture(const std::string &uniform, LLTexUnit::eTextureType mode)
+{
+	S32 channel = 0;
+	channel = getUniformLocation(uniform);
+	
+	return unbindTexture(channel);
+}
+
+S32 LLGLSLShader::unbindTexture(S32 uniform, LLTexUnit::eTextureType mode)
+{
+	if (uniform < 0 || uniform >= (S32)mTexture.size())
+	{
+		UNIFORM_ERRS << "Uniform out of range: " << uniform << LL_ENDL;
+		return -1;
+	}
+	
+	uniform = mTexture[uniform];
+	
+	if (uniform > -1)
+	{
+		gGL.getTexUnit(uniform)->unbind(mode);
+	}
+	
+	return uniform;
+}
+
 S32 LLGLSLShader::enableTexture(S32 uniform, LLTexUnit::eTextureType mode)
 {
 	if (uniform < 0 || uniform >= (S32)mTexture.size())
@@ -857,6 +1200,23 @@ void LLGLSLShader::uniform1i(const string& uniform, GLint v)
 	}
 }
 
+void LLGLSLShader::uniform2i(const string& uniform, GLint i, GLint j)
+{
+	GLint location = getUniformLocation(uniform);
+				
+	if (location >= 0)
+	{
+		std::map<GLint, LLVector4>::iterator iter = mValue.find(location);
+		LLVector4 vec(i,j,0.f,0.f);
+		if (iter == mValue.end() || shouldChange(iter->second,vec))
+		{
+			glUniform2iARB(location, i, j);
+			mValue[location] = vec;
+		}
+	}
+}
+
+
 void LLGLSLShader::uniform1f(const string& uniform, GLfloat v)
 {
 	GLint location = getUniformLocation(uniform);
diff --git a/indra/llrender/llglslshader.h b/indra/llrender/llglslshader.h
index cf21101e35..eabdb9fc92 100755
--- a/indra/llrender/llglslshader.h
+++ b/indra/llrender/llglslshader.h
@@ -67,14 +67,29 @@ public:
 		SG_WATER
 	};
 	
+	static std::set<LLGLSLShader*> sInstances;
+	static bool sProfileEnabled;
+
 	LLGLSLShader();
+	~LLGLSLShader();
 
 	static GLhandleARB sCurBoundShader;
 	static LLGLSLShader* sCurBoundShaderPtr;
 	static S32 sIndexedTextureChannels;
 	static bool sNoFixedFunction;
 
+	static void initProfile();
+	static void finishProfile();
+
+	static void startProfile();
+	static void stopProfile(U32 count, U32 mode);
+
 	void unload();
+	void clearStats();
+	void dumpStats();
+	void placeProfileQuery();
+	void readProfileQuery(U32 count, U32 mode);
+
 	BOOL createShader(std::vector<std::string> * attributes,
 						std::vector<std::string> * uniforms,
 						U32 varying_count = 0,
@@ -96,6 +111,7 @@ public:
 	void uniform3fv(U32 index, U32 count, const GLfloat* v);
 	void uniform4fv(U32 index, U32 count, const GLfloat* v);
 	void uniform1i(const std::string& uniform, GLint i);
+	void uniform2i(const std::string& uniform, GLint i, GLint j);
 	void uniform1f(const std::string& uniform, GLfloat v);
 	void uniform2f(const std::string& uniform, GLfloat x, GLfloat y);
 	void uniform3f(const std::string& uniform, GLfloat x, GLfloat y, GLfloat z);
@@ -123,12 +139,22 @@ public:
 	GLint getAttribLocation(U32 attrib);
 	GLint mapUniformTextureChannel(GLint location, GLenum type);
 	
+	void addPermutation(std::string name, std::string value);
+	void removePermutation(std::string name);
+	
 	//enable/disable texture channel for specified uniform
 	//if given texture uniform is active in the shader, 
 	//the corresponding channel will be active upon return
 	//returns channel texture is enabled in from [0-MAX)
 	S32 enableTexture(S32 uniform, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE);
-	S32 disableTexture(S32 uniform, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE); 
+	S32 disableTexture(S32 uniform, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE);
+	
+	// bindTexture returns the texture unit we've bound the texture to.
+	// You can reuse the return value to unbind a texture when required.
+	S32 bindTexture(const std::string& uniform, LLTexture *texture, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE);
+	S32 bindTexture(S32 uniform, LLTexture *texture, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE);
+	S32 unbindTexture(const std::string& uniform, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE);
+	S32 unbindTexture(S32 uniform, LLTexUnit::eTextureType mode = LLTexUnit::TT_TEXTURE);
 	
     BOOL link(BOOL suppress_errors = FALSE);
 	void bind();
@@ -142,10 +168,13 @@ public:
 
 	GLhandleARB mProgramObject;
 	std::vector<GLint> mAttribute; //lookup table of attribute enum to attribute channel
+	U32 mAttributeMask;  //mask of which reserved attributes are set (lines up with LLVertexBuffer::getTypeMask())
 	std::vector<GLint> mUniform;   //lookup table of uniform enum to uniform location
 	std::map<std::string, GLint> mUniformMap;  //lookup map of uniform name to uniform location
+	std::map<GLint, std::string> mUniformNameMap; //lookup map of uniform location to uniform name
 	std::map<GLint, LLVector4> mValue; //lookup map of uniform location to last known value
 	std::vector<GLint> mTexture;
+	S32 mTotalUniformSize;
 	S32 mActiveTextureChannels;
 	S32 mShaderLevel;
 	S32 mShaderGroup;
@@ -153,6 +182,23 @@ public:
 	LLShaderFeatures mFeatures;
 	std::vector< std::pair< std::string, GLenum > > mShaderFiles;
 	std::string mName;
+	boost::unordered_map<std::string, std::string> mDefines;
+
+	//statistcis for profiling shader performance
+	U32 mTimerQuery;
+	U64 mTimeElapsed;
+	static U64 sTotalTimeElapsed;
+	U32 mTrianglesDrawn;
+	static U32 sTotalTrianglesDrawn;
+	U64 mSamplesDrawn;
+	static U64 sTotalSamplesDrawn;
+	U32 mDrawCalls;
+	static U32 sTotalDrawCalls;
+
+	bool mTextureStateFetched;
+	std::vector<U32> mTextureMagFilter;
+	std::vector<U32> mTextureMinFilter;
+	
 };
 
 //UI shader (declared here so llui_libtest will link properly)
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 5c171d372c..38764eba23 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -748,12 +748,16 @@ void LLImageGL::setImage(const U8* data_in, BOOL data_hasmips)
 				S32 height = getHeight(mCurrentDiscardLevel);
 				S32 nummips = mMaxDiscardLevel - mCurrentDiscardLevel + 1;
 				S32 w = width, h = height;
+
+
+				const U8* new_data = 0;
+				(void)new_data;
+
 				const U8* prev_mip_data = 0;
 				const U8* cur_mip_data = 0;
 #ifdef SHOW_ASSERT
 				S32 cur_mip_size = 0;
 #endif
-				
 				mMipLevels = nummips;
 
 				for (int m=0; m<nummips; m++)
@@ -773,14 +777,22 @@ void LLImageGL::setImage(const U8* data_in, BOOL data_hasmips)
 						llassert(cur_mip_size == bytes*4);
 #endif
 						U8* new_data = new U8[bytes];
+
+#ifdef SHOW_ASSERT
+						llassert(prev_mip_data);
+						llassert(cur_mip_size == bytes*4);
 						llassert_always(new_data);
+#endif
+
 						LLImageBase::generateMip(prev_mip_data, new_data, w, h, mComponents);
 						cur_mip_data = new_data;
 #ifdef SHOW_ASSERT
 						cur_mip_size = bytes; 
 #endif
+
 					}
 					llassert(w > 0 && h > 0 && cur_mip_data);
+					(void)cur_mip_data;
 					{
 // 						LLFastTimer t1(FTM_TEMP4);
 						if(mFormatSwapBytes)
@@ -1119,30 +1131,30 @@ void LLImageGL::deleteTextures(LLTexUnit::eTextureType type, U32 format, S32 mip
 			default:
 			{
 				if (type == LLTexUnit::TT_CUBE_MAP || mip_levels == -1)
-				{ //unknown internal format or unknown number of mip levels, not safe to reuse
-					glDeleteTextures(numTextures, textures);
-				}
-				else
-				{
-					for (S32 i = 0; i < numTextures; ++i)
-					{ //remove texture from VRAM by setting its size to zero
+		{ //unknown internal format or unknown number of mip levels, not safe to reuse
+			glDeleteTextures(numTextures, textures);
+		}
+		else
+		{
+			for (S32 i = 0; i < numTextures; ++i)
+			{ //remove texture from VRAM by setting its size to zero
 
-						for (S32 j = 0; j <= mip_levels; j++)
-						{
-							gGL.getTexUnit(0)->bindManual(type, textures[i]);
+				for (S32 j = 0; j <= mip_levels; j++)
+				{
+					gGL.getTexUnit(0)->bindManual(type, textures[i]);
 							U32 internal_type = LLTexUnit::getInternalType(type);
 							glTexImage2D(internal_type, j, format, 0, 0, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
 							stop_glerror();
-						}
+				}
 
-						llassert(std::find(sDeadTextureList[type][format].begin(),
-							sDeadTextureList[type][format].end(), textures[i]) == 
-							sDeadTextureList[type][format].end());
+				llassert(std::find(sDeadTextureList[type][format].begin(),
+								   sDeadTextureList[type][format].end(), textures[i]) == 
+								   sDeadTextureList[type][format].end());
 
-						sDeadTextureList[type][format].push_back(textures[i]);
-					}	
-				}				
-			}
+				sDeadTextureList[type][format].push_back(textures[i]);
+			}	
+		}
+	}
 			break;
 		}
 	}
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 78a310e525..98222939e7 100755
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -262,6 +262,14 @@ class LLRender
 	friend class LLTexUnit;
 public:
 
+	enum eTexIndex
+	{
+		DIFFUSE_MAP = 0,
+		NORMAL_MAP,
+		SPECULAR_MAP,
+		NUM_TEXTURE_CHANNELS,
+	};
+	
 	typedef enum {
 		TRIANGLES = 0,
 		TRIANGLE_STRIP,
diff --git a/indra/llrender/llrendertarget.cpp b/indra/llrender/llrendertarget.cpp
index 5fb4fc8e52..6e22712b94 100755
--- a/indra/llrender/llrendertarget.cpp
+++ b/indra/llrender/llrendertarget.cpp
@@ -53,11 +53,19 @@ void check_framebuffer_status()
 bool LLRenderTarget::sUseFBO = false;
 U32 LLRenderTarget::sCurFBO = 0;
 
+
+extern S32 gGLViewport[4];
+
+U32 LLRenderTarget::sCurResX = 0;
+U32 LLRenderTarget::sCurResY = 0;
+
 LLRenderTarget::LLRenderTarget() :
 	mResX(0),
 	mResY(0),
 	mFBO(0),
 	mPreviousFBO(0),
+	mPreviousResX(0),
+	mPreviousResY(0),
 	mDepth(0),
 	mStencil(0),
 	mUseDepth(false),
@@ -390,13 +398,12 @@ void LLRenderTarget::bindTarget()
 {
 	if (mFBO)
 	{
-		mPreviousFBO = sCurFBO;
-
 		stop_glerror();
 		
+		mPreviousFBO = sCurFBO;
 		glBindFramebuffer(GL_FRAMEBUFFER, mFBO);
 		sCurFBO = mFBO;
-
+		
 		stop_glerror();
 		if (gGLManager.mHasDrawBuffers)
 		{ //setup multiple render targets
@@ -418,7 +425,12 @@ void LLRenderTarget::bindTarget()
 		stop_glerror();
 	}
 
+	mPreviousResX = sCurResX;
+	mPreviousResY = sCurResY;
 	glViewport(0, 0, mResX, mResY);
+	sCurResX = mResX;
+	sCurResY = mResY;
+
 	sBoundTarget = this;
 }
 
@@ -489,6 +501,20 @@ void LLRenderTarget::flush(bool fetch_depth)
 		stop_glerror();
 		glBindFramebuffer(GL_FRAMEBUFFER, mPreviousFBO);
 		sCurFBO = mPreviousFBO;
+
+		if (mPreviousFBO)
+		{
+			glViewport(0, 0, mPreviousResX, mPreviousResY);
+			sCurResX = mPreviousResX;
+			sCurResY = mPreviousResY;
+		}
+		else
+		{
+			glViewport(gGLViewport[0],gGLViewport[1],gGLViewport[2],gGLViewport[3]);
+			sCurResX = gGLViewport[2];
+			sCurResY = gGLViewport[3];
+		}
+						
 		stop_glerror();
 	}
 }
diff --git a/indra/llrender/llrendertarget.h b/indra/llrender/llrendertarget.h
index 765a727b5b..66a9874a6b 100755
--- a/indra/llrender/llrendertarget.h
+++ b/indra/llrender/llrendertarget.h
@@ -63,6 +63,9 @@ public:
 	static bool sUseFBO; 
 	static U32 sBytesAllocated;
 	static U32 sCurFBO;
+	static U32 sCurResX;
+	static U32 sCurResY;
+
 
 	LLRenderTarget();
 	~LLRenderTarget();
@@ -146,6 +149,9 @@ protected:
 	std::vector<U32> mInternalFormat;
 	U32 mFBO;
 	U32 mPreviousFBO;
+	U32 mPreviousResX;
+	U32 mPreviousResY;
+
 	U32 mDepth;
 	bool mStencil;
 	bool mUseDepth;
diff --git a/indra/llrender/llshadermgr.cpp b/indra/llrender/llshadermgr.cpp
index b6a9a6b653..fea4ee2819 100755
--- a/indra/llrender/llshadermgr.cpp
+++ b/indra/llrender/llshadermgr.cpp
@@ -521,7 +521,7 @@ void LLShaderMgr::dumpObjectLog(GLhandleARB ret, BOOL warns)
 	}
  }
 
-GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shader_level, GLenum type, S32 texture_index_channels)
+GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shader_level, GLenum type, boost::unordered_map<std::string, std::string>* defines, S32 texture_index_channels)
 {
 	GLenum error = GL_NO_ERROR;
 	if (gDebugGL)
@@ -650,13 +650,15 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 			text[count++] = strdup("#define shadow2DRect(a,b) vec2(texture(a,b))\n");
 		}
 	}
-
-	//copy preprocessor definitions into buffer
-	for (std::map<std::string,std::string>::iterator iter = mDefinitions.begin(); iter != mDefinitions.end(); ++iter)
+	
+	if (defines)
+	{
+		for (boost::unordered_map<std::string,std::string>::iterator iter = defines->begin(); iter != defines->end(); ++iter)
 	{
 		std::string define = "#define " + iter->first + " " + iter->second + "\n";
 		text[count++] = (GLcharARB *) strdup(define.c_str());
 	}
+	}
 
 	if (texture_index_channels > 0 && type == GL_FRAGMENT_SHADER_ARB)
 	{
@@ -693,6 +695,8 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 		}
 		*/
 
+		text[count++] = strdup("#define HAS_DIFFUSE_LOOKUP 1\n");
+
 		//uniform declartion
 		for (S32 i = 0; i < texture_index_channels; ++i)
 		{
@@ -750,6 +754,10 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 			llerrs << "Indexed texture rendering requires GLSL 1.30 or later." << llendl;
 		}
 	}
+	else
+	{
+		text[count++] = strdup("#define HAS_DIFFUSE_LOOKUP 0\n");
+	}
 
 	//copy file into memory
 	while( fgets((char *)buff, 1024, file) != NULL && count < LL_ARRAY_SIZE(text) ) 
@@ -806,7 +814,6 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 				//an error occured, print log
 				LL_WARNS("ShaderLoading") << "GLSL Compilation Error: (" << error << ") in " << filename << LL_ENDL;
 				dumpObjectLog(ret);
-
 #if LL_WINDOWS
 				std::stringstream ostr;
 				//dump shader source for debugging
@@ -824,8 +831,20 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 				}
 
 				LL_WARNS("ShaderLoading") << "\n" << ostr.str() << llendl;
-#endif // LL_WINDOWS
-
+#else
+				std::string str;
+				
+				for (GLuint i = 0; i < count; i++) {
+					str.append(text[i]);
+					
+					if (i % 128 == 0)
+					{
+						LL_WARNS("ShaderLoading") << str << llendl;
+						str = "";
+					}
+				}
+#endif
+				
 				ret = 0;
 			}
 		}
@@ -854,7 +873,7 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 		if (shader_level > 1)
 		{
 			shader_level--;
-			return loadShaderFile(filename,shader_level,type,texture_index_channels);
+			return loadShaderFile(filename,shader_level,type, defines, texture_index_channels);
 		}
 		LL_WARNS("ShaderLoading") << "Failed to load " << filename << LL_ENDL;	
 	}
@@ -958,7 +977,7 @@ void LLShaderMgr::initAttribsAndUniforms()
 	mReservedAttribs.push_back("texcoord3");
 	mReservedAttribs.push_back("diffuse_color");
 	mReservedAttribs.push_back("emissive");
-	mReservedAttribs.push_back("binormal");
+	mReservedAttribs.push_back("tangent");
 	mReservedAttribs.push_back("weight");
 	mReservedAttribs.push_back("weight4");
 	mReservedAttribs.push_back("clothing");
@@ -1055,6 +1074,7 @@ void LLShaderMgr::initAttribsAndUniforms()
 
 
 	mReservedUniforms.push_back("minimum_alpha");
+	mReservedUniforms.push_back("emissive_brightness");
 
 	mReservedUniforms.push_back("shadow_matrix");
 	mReservedUniforms.push_back("env_mat");
@@ -1115,6 +1135,12 @@ void LLShaderMgr::initAttribsAndUniforms()
 	mReservedUniforms.push_back("lightMap");
 	mReservedUniforms.push_back("bloomMap");
 	mReservedUniforms.push_back("projectionMap");
+	
+	mReservedUniforms.push_back("global_gamma");
+	mReservedUniforms.push_back("texture_gamma");
+	
+	mReservedUniforms.push_back("specular_color");
+	mReservedUniforms.push_back("env_intensity");
 
 	llassert(mReservedUniforms.size() == END_RESERVED_UNIFORMS);
 
diff --git a/indra/llrender/llshadermgr.h b/indra/llrender/llshadermgr.h
index 7a16b7c20f..c049e935b8 100755
--- a/indra/llrender/llshadermgr.h
+++ b/indra/llrender/llshadermgr.h
@@ -109,6 +109,7 @@ public:
 		GLOW_DELTA,
 
 		MINIMUM_ALPHA,
+		EMISSIVE_BRIGHTNESS,
 
 		DEFERRED_SHADOW_MATRIX,
 		DEFERRED_ENV_MAT,
@@ -164,6 +165,13 @@ public:
 		DEFERRED_LIGHT,
 		DEFERRED_BLOOM,
 		DEFERRED_PROJECTION,
+		
+		GLOBAL_GAMMA,
+		TEXTURE_GAMMA,
+		
+		SPECULAR_COLOR,
+		ENVIRONMENT_INTENSITY,
+		
 		END_RESERVED_UNIFORMS
 	} eGLSLReservedUniforms;
 
@@ -176,7 +184,7 @@ public:
 	void dumpObjectLog(GLhandleARB ret, BOOL warns = TRUE);
 	BOOL	linkProgramObject(GLhandleARB obj, BOOL suppress_errors = FALSE);
 	BOOL	validateProgramObject(GLhandleARB obj);
-	GLhandleARB loadShaderFile(const std::string& filename, S32 & shader_level, GLenum type, S32 texture_index_channels = -1);
+	GLhandleARB loadShaderFile(const std::string& filename, S32 & shader_level, GLenum type, boost::unordered_map<std::string, std::string>* defines = NULL, S32 texture_index_channels = -1);
 
 	// Implemented in the application to actually point to the shader directory.
 	virtual std::string getShaderDirPrefix(void) = 0; // Pure Virtual
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 4909b43e8a..01541026b1 100755
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -342,13 +342,32 @@ S32 LLVertexBuffer::sTypeSize[LLVertexBuffer::TYPE_MAX] =
 	sizeof(LLVector2), // TYPE_TEXCOORD3,
 	sizeof(LLColor4U), // TYPE_COLOR,
 	sizeof(LLColor4U), // TYPE_EMISSIVE, only alpha is used currently
-	sizeof(LLVector4), // TYPE_BINORMAL,
+	sizeof(LLVector4), // TYPE_TANGENT,
 	sizeof(F32),	   // TYPE_WEIGHT,
 	sizeof(LLVector4), // TYPE_WEIGHT4,
 	sizeof(LLVector4), // TYPE_CLOTHWEIGHT,
 	sizeof(LLVector4), // TYPE_TEXTURE_INDEX (actually exists as position.w), no extra data, but stride is 16 bytes
 };
 
+static std::string vb_type_name[] =
+{
+	"TYPE_VERTEX",
+	"TYPE_NORMAL",
+	"TYPE_TEXCOORD0",
+	"TYPE_TEXCOORD1",
+	"TYPE_TEXCOORD2",
+	"TYPE_TEXCOORD3",
+	"TYPE_COLOR",
+	"TYPE_EMISSIVE",
+	"TYPE_TANGENT",
+	"TYPE_WEIGHT",
+	"TYPE_WEIGHT4",
+	"TYPE_CLOTHWEIGHT",
+	"TYPE_TEXTURE_INDEX",
+	"TYPE_MAX",
+	"TYPE_INDEX",	
+};
+
 U32 LLVertexBuffer::sGLMode[LLRender::NUM_MODES] = 
 {
 	GL_TRIANGLES,
@@ -523,16 +542,16 @@ void LLVertexBuffer::setupClientArrays(U32 data_mask)
 				}
 			}
 
-			if (sLastMask & MAP_BINORMAL)
+			if (sLastMask & MAP_TANGENT)
 			{
-				if (!(data_mask & MAP_BINORMAL))
+				if (!(data_mask & MAP_TANGENT))
 				{
 					glClientActiveTextureARB(GL_TEXTURE2_ARB);
 					glDisableClientState(GL_TEXTURE_COORD_ARRAY);
 					glClientActiveTextureARB(GL_TEXTURE0_ARB);
 				}
 			}
-			else if (data_mask & MAP_BINORMAL)
+			else if (data_mask & MAP_TANGENT)
 			{
 				glClientActiveTextureARB(GL_TEXTURE2_ARB);
 				glEnableClientState(GL_TEXTURE_COORD_ARRAY);
@@ -593,8 +612,9 @@ void LLVertexBuffer::drawArrays(U32 mode, const std::vector<LLVector3>& pos, con
 		glVertexPointer(3, GL_FLOAT, 0, pos[0].mV);
 		glNormalPointer(GL_FLOAT, 0, norm[0].mV);
 	}
-
+	LLGLSLShader::startProfile();
 	glDrawArrays(sGLMode[mode], 0, count);
+	LLGLSLShader::stopProfile(count, mode);
 }
 
 //static
@@ -631,7 +651,9 @@ void LLVertexBuffer::drawElements(U32 mode, const LLVector4a* pos, const LLVecto
 		glVertexPointer(3, GL_FLOAT, 16, pos);
 	}
 
+	LLGLSLShader::startProfile();
 	glDrawElements(sGLMode[mode], num_indices, GL_UNSIGNED_SHORT, indicesp);
+	LLGLSLShader::stopProfile(num_indices, mode);
 }
 
 void LLVertexBuffer::validateRange(U32 start, U32 end, U32 count, U32 indices_offset) const
@@ -731,9 +753,14 @@ void LLVertexBuffer::drawRange(U32 mode, U32 start, U32 end, U32 count, U32 indi
 	U16* idx = ((U16*) getIndicesPointer())+indices_offset;
 
 	stop_glerror();
+	LLGLSLShader::startProfile();
 	glDrawRangeElements(sGLMode[mode], start, end, count, GL_UNSIGNED_SHORT, 
 		idx);
+	LLGLSLShader::stopProfile(count, mode);
 	stop_glerror();
+
+	
+
 	placeFence();
 }
 
@@ -777,8 +804,10 @@ void LLVertexBuffer::draw(U32 mode, U32 count, U32 indices_offset) const
 	}
 
 	stop_glerror();
+	LLGLSLShader::startProfile();
 	glDrawElements(sGLMode[mode], count, GL_UNSIGNED_SHORT,
 		((U16*) getIndicesPointer()) + indices_offset);
+	LLGLSLShader::stopProfile(count, mode);
 	stop_glerror();
 	placeFence();
 }
@@ -820,9 +849,12 @@ void LLVertexBuffer::drawArrays(U32 mode, U32 first, U32 count) const
 
 	{
 		LLFastTimer t2(FTM_GL_DRAW_ARRAYS);
-		stop_glerror();
-		glDrawArrays(sGLMode[mode], first, count);
-	}
+	stop_glerror();
+	LLGLSLShader::startProfile();
+	glDrawArrays(sGLMode[mode], first, count);
+	LLGLSLShader::stopProfile(count, mode);
+        }
+
 	stop_glerror();
 	placeFence();
 }
@@ -1322,7 +1354,7 @@ void LLVertexBuffer::setupVertexArray()
 		2, //TYPE_TEXCOORD3,
 		4, //TYPE_COLOR,
 		4, //TYPE_EMISSIVE,
-		3, //TYPE_BINORMAL,
+		4, //TYPE_TANGENT,
 		1, //TYPE_WEIGHT,
 		4, //TYPE_WEIGHT4,
 		4, //TYPE_CLOTHWEIGHT,
@@ -1339,7 +1371,7 @@ void LLVertexBuffer::setupVertexArray()
 		GL_FLOAT, //TYPE_TEXCOORD3,
 		GL_UNSIGNED_BYTE, //TYPE_COLOR,
 		GL_UNSIGNED_BYTE, //TYPE_EMISSIVE,
-		GL_FLOAT,   //TYPE_BINORMAL,
+		GL_FLOAT,   //TYPE_TANGENT,
 		GL_FLOAT, //TYPE_WEIGHT,
 		GL_FLOAT, //TYPE_WEIGHT4,
 		GL_FLOAT, //TYPE_CLOTHWEIGHT,
@@ -1356,7 +1388,7 @@ void LLVertexBuffer::setupVertexArray()
 		false, //TYPE_TEXCOORD3,
 		false, //TYPE_COLOR,
 		false, //TYPE_EMISSIVE,
-		false, //TYPE_BINORMAL,
+		false, //TYPE_TANGENT,
 		false, //TYPE_WEIGHT,
 		false, //TYPE_WEIGHT4,
 		false, //TYPE_CLOTHWEIGHT,
@@ -1373,7 +1405,7 @@ void LLVertexBuffer::setupVertexArray()
 		GL_FALSE, //TYPE_TEXCOORD3,
 		GL_TRUE, //TYPE_COLOR,
 		GL_TRUE, //TYPE_EMISSIVE,
-		GL_FALSE,   //TYPE_BINORMAL,
+		GL_FALSE,   //TYPE_TANGENT,
 		GL_FALSE, //TYPE_WEIGHT,
 		GL_FALSE, //TYPE_WEIGHT4,
 		GL_FALSE, //TYPE_CLOTHWEIGHT,
@@ -2038,14 +2070,21 @@ bool LLVertexBuffer::getTexCoord1Strider(LLStrider<LLVector2>& strider, S32 inde
 {
 	return VertexBufferStrider<LLVector2,TYPE_TEXCOORD1>::get(*this, strider, index, count, map_range);
 }
-
+bool LLVertexBuffer::getTexCoord2Strider(LLStrider<LLVector2>& strider, S32 index, S32 count, bool map_range)
+{
+	return VertexBufferStrider<LLVector2,TYPE_TEXCOORD2>::get(*this, strider, index, count, map_range);
+}
 bool LLVertexBuffer::getNormalStrider(LLStrider<LLVector3>& strider, S32 index, S32 count, bool map_range)
 {
 	return VertexBufferStrider<LLVector3,TYPE_NORMAL>::get(*this, strider, index, count, map_range);
 }
-bool LLVertexBuffer::getBinormalStrider(LLStrider<LLVector3>& strider, S32 index, S32 count, bool map_range)
+bool LLVertexBuffer::getTangentStrider(LLStrider<LLVector3>& strider, S32 index, S32 count, bool map_range)
 {
-	return VertexBufferStrider<LLVector3,TYPE_BINORMAL>::get(*this, strider, index, count, map_range);
+	return VertexBufferStrider<LLVector3,TYPE_TANGENT>::get(*this, strider, index, count, map_range);
+}
+bool LLVertexBuffer::getTangentStrider(LLStrider<LLVector4a>& strider, S32 index, S32 count, bool map_range)
+{
+	return VertexBufferStrider<LLVector4a,TYPE_TANGENT>::get(*this, strider, index, count, map_range);
 }
 bool LLVertexBuffer::getColorStrider(LLStrider<LLColor4U>& strider, S32 index, S32 count, bool map_range)
 {
@@ -2200,7 +2239,7 @@ void LLVertexBuffer::setBuffer(U32 data_mask)
 
 			if ((data_mask & required_mask) != required_mask)
 			{
-				llerrs << "Shader consumption mismatches data provision." << llendl;
+				llwarns << "Shader consumption mismatches data provision." << llendl;
 			}
 		}
 	}
@@ -2318,6 +2357,14 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 
 	if (gDebugGL && ((data_mask & mTypeMask) != data_mask))
 	{
+		for (U32 i = 0; i < LLVertexBuffer::TYPE_MAX; ++i)
+		{
+			U32 mask = 1 << i;
+			if (mask & data_mask && !(mask & mTypeMask))
+			{ //bit set in data_mask, but not set in mTypeMask
+				llwarns << "Missing required component " << vb_type_name[i] << llendl;
+			}
+		}
 		llerrs << "LLVertexBuffer::setupVertexBuffer missing required components for supplied data mask." << llendl;
 	}
 
@@ -2347,11 +2394,11 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 			void* ptr = (void*)(base + mOffsets[TYPE_TEXCOORD1]);
 			glVertexAttribPointerARB(loc,2,GL_FLOAT, GL_FALSE, LLVertexBuffer::sTypeSize[TYPE_TEXCOORD1], ptr);
 		}
-		if (data_mask & MAP_BINORMAL)
+		if (data_mask & MAP_TANGENT)
 		{
-			S32 loc = TYPE_BINORMAL;
-			void* ptr = (void*)(base + mOffsets[TYPE_BINORMAL]);
-			glVertexAttribPointerARB(loc, 3,GL_FLOAT, GL_FALSE, LLVertexBuffer::sTypeSize[TYPE_BINORMAL], ptr);
+			S32 loc = TYPE_TANGENT;
+			void* ptr = (void*)(base + mOffsets[TYPE_TANGENT]);
+			glVertexAttribPointerARB(loc, 4,GL_FLOAT, GL_FALSE, LLVertexBuffer::sTypeSize[TYPE_TANGENT], ptr);
 		}
 		if (data_mask & MAP_TEXCOORD0)
 		{
@@ -2429,10 +2476,10 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 			glTexCoordPointer(2,GL_FLOAT, LLVertexBuffer::sTypeSize[TYPE_TEXCOORD1], (void*)(base + mOffsets[TYPE_TEXCOORD1]));
 			glClientActiveTextureARB(GL_TEXTURE0_ARB);
 		}
-		if (data_mask & MAP_BINORMAL)
+		if (data_mask & MAP_TANGENT)
 		{
 			glClientActiveTextureARB(GL_TEXTURE2_ARB);
-			glTexCoordPointer(3,GL_FLOAT, LLVertexBuffer::sTypeSize[TYPE_BINORMAL], (void*)(base + mOffsets[TYPE_BINORMAL]));
+			glTexCoordPointer(4,GL_FLOAT, LLVertexBuffer::sTypeSize[TYPE_TANGENT], (void*)(base + mOffsets[TYPE_TANGENT]));
 			glClientActiveTextureARB(GL_TEXTURE0_ARB);
 		}
 		if (data_mask & MAP_TEXCOORD0)
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index 11fa4ab6a0..04806c1d8c 100755
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -174,7 +174,7 @@ public:
 		TYPE_TEXCOORD3,
 		TYPE_COLOR,
 		TYPE_EMISSIVE,
-		TYPE_BINORMAL,
+		TYPE_TANGENT,
 		TYPE_WEIGHT,
 		TYPE_WEIGHT4,
 		TYPE_CLOTHWEIGHT,
@@ -192,7 +192,7 @@ public:
 		MAP_COLOR = (1<<TYPE_COLOR),
 		MAP_EMISSIVE = (1<<TYPE_EMISSIVE),
 		// These use VertexAttribPointer and should possibly be made generic
-		MAP_BINORMAL = (1<<TYPE_BINORMAL),
+		MAP_TANGENT = (1<<TYPE_TANGENT),
 		MAP_WEIGHT = (1<<TYPE_WEIGHT),
 		MAP_WEIGHT4 = (1<<TYPE_WEIGHT4),
 		MAP_CLOTHWEIGHT = (1<<TYPE_CLOTHWEIGHT),
@@ -250,8 +250,10 @@ public:
 	bool getIndexStrider(LLStrider<U16>& strider, S32 index=0, S32 count = -1, bool map_range = false);
 	bool getTexCoord0Strider(LLStrider<LLVector2>& strider, S32 index=0, S32 count = -1, bool map_range = false);
 	bool getTexCoord1Strider(LLStrider<LLVector2>& strider, S32 index=0, S32 count = -1, bool map_range = false);
+	bool getTexCoord2Strider(LLStrider<LLVector2>& strider, S32 index=0, S32 count = -1, bool map_range = false);
 	bool getNormalStrider(LLStrider<LLVector3>& strider, S32 index=0, S32 count = -1, bool map_range = false);
-	bool getBinormalStrider(LLStrider<LLVector3>& strider, S32 index=0, S32 count = -1, bool map_range = false);
+	bool getTangentStrider(LLStrider<LLVector3>& strider, S32 index=0, S32 count = -1, bool map_range = false);
+	bool getTangentStrider(LLStrider<LLVector4a>& strider, S32 index=0, S32 count = -1, bool map_range = false);
 	bool getColorStrider(LLStrider<LLColor4U>& strider, S32 index=0, S32 count = -1, bool map_range = false);
 	bool getTextureIndexStrider(LLStrider<LLColor4U>& strider, S32 index=0, S32 count = -1, bool map_range = false);
 	bool getEmissiveStrider(LLStrider<LLColor4U>& strider, S32 index=0, S32 count = -1, bool map_range = false);