MAINT-2371 First set of profile guided optimizations.

Reviewed by Graham

10 files changed, 203 insertions, 121 deletions

diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp
index 02c8d2b86f..77d89568df 100644
--- a/indra/llmath/llvolume.cpp
+++ b/indra/llmath/llvolume.cpp
@@ -5187,7 +5187,8 @@ LLVolumeFace::LLVolumeFace() :
 	mTexCoords(NULL),
 	mIndices(NULL),
 	mWeights(NULL),
-	mOctree(NULL)
+	mOctree(NULL),
+	mOptimized(FALSE)
 {
 	mExtents = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*3);
 	mExtents[0].splat(-0.5f);
@@ -5517,14 +5518,14 @@ class LLVCacheVertexData
 public:
 	S32 mIdx;
 	S32 mCacheTag;
-	F32 mScore;
+	F64 mScore;
 	U32 mActiveTriangles;
 	std::vector<LLVCacheTriangleData*> mTriangles;
 
 	LLVCacheVertexData()
 	{
 		mCacheTag = -1;
-		mScore = 0.f;
+		mScore = 0.0;
 		mActiveTriangles = 0;
 		mIdx = -1;
 	}
@@ -5534,13 +5535,13 @@ class LLVCacheTriangleData
 {
 public:
 	bool mActive;
-	F32 mScore;
+	F64 mScore;
 	LLVCacheVertexData* mVertex[3];
 
 	LLVCacheTriangleData()
 	{
 		mActive = true;
-		mScore = 0.f;
+		mScore = 0.0;
 		mVertex[0] = mVertex[1] = mVertex[2] = NULL;
 	}
 
@@ -5551,7 +5552,7 @@ public:
 		{
 			if (mVertex[i])
 			{
-				llassert_always(mVertex[i]->mActiveTriangles > 0);
+				llassert(mVertex[i]->mActiveTriangles > 0);
 				mVertex[i]->mActiveTriangles--;
 			}
 		}
@@ -5563,44 +5564,44 @@ public:
 	}
 };
 
-const F32 FindVertexScore_CacheDecayPower = 1.5f;
-const F32 FindVertexScore_LastTriScore = 0.75f;
-const F32 FindVertexScore_ValenceBoostScale = 2.0f;
-const F32 FindVertexScore_ValenceBoostPower = 0.5f;
+const F64 FindVertexScore_CacheDecayPower = 1.5;
+const F64 FindVertexScore_LastTriScore = 0.75;
+const F64 FindVertexScore_ValenceBoostScale = 2.0;
+const F64 FindVertexScore_ValenceBoostPower = 0.5;
 const U32 MaxSizeVertexCache = 32;
+const F64 FindVertexScore_Scaler = 1.0/(MaxSizeVertexCache-3);
 
-F32 find_vertex_score(LLVCacheVertexData& data)
+F64 find_vertex_score(LLVCacheVertexData& data)
 {
-	if (data.mActiveTriangles == 0)
-	{ //no triangle references this vertex
-		return -1.f;
-	}
-
-	F32 score = 0.f;
+	F64 score = -1.0;
 
-	S32 cache_idx = data.mCacheTag;
+	if (data.mActiveTriangles >= 0)
+	{ 
+		score = 0.0;
+		
+		S32 cache_idx = data.mCacheTag;
 
-	if (cache_idx < 0)
-	{
-		//not in cache
-	}
-	else
-	{
-		if (cache_idx < 3)
-		{ //vertex was in the last triangle
-			score = FindVertexScore_LastTriScore;
+		if (cache_idx < 0)
+		{
+			//not in cache
 		}
 		else
-		{ //more points for being higher in the cache
-			F32 scaler = 1.f/(MaxSizeVertexCache-3);
-			score = 1.f-((cache_idx-3)*scaler);
-			score = powf(score, FindVertexScore_CacheDecayPower);
+		{
+			if (cache_idx < 3)
+			{ //vertex was in the last triangle
+				score = FindVertexScore_LastTriScore;
+			}
+			else
+			{ //more points for being higher in the cache
+				score = 1.0-((cache_idx-3)*FindVertexScore_Scaler);
+				score = pow(score, FindVertexScore_CacheDecayPower);
+			}
 		}
-	}
 
-	//bonus points for having low valence
-	F32 valence_boost = powf((F32)data.mActiveTriangles, -FindVertexScore_ValenceBoostPower);
-	score += FindVertexScore_ValenceBoostScale * valence_boost;
+		//bonus points for having low valence
+		F64 valence_boost = pow((F64)data.mActiveTriangles, -FindVertexScore_ValenceBoostPower);
+		score += FindVertexScore_ValenceBoostScale * valence_boost;
+	}
 
 	return score;
 }
@@ -5720,7 +5721,7 @@ public:
 			if (mCache[i])
 			{
 				mCache[i]->mScore = find_vertex_score(*(mCache[i]));
-				llassert_always(mCache[i]->mCacheTag == i);
+				llassert(mCache[i]->mCacheTag == i);
 			}
 		}
 
@@ -5728,11 +5729,14 @@ public:
 		//update triangle scores
 		for (U32 i = 0; i < MaxSizeVertexCache+3; ++i)
 		{
-			if (mCache[i])
+			LLVCacheVertexData* data = mCache[i];
+			if (data)
 			{
-				for (U32 j = 0; j < mCache[i]->mTriangles.size(); ++j)
+				U32 count = data->mTriangles.size();
+
+				for (U32 j = 0; j < count; ++j)
 				{
-					LLVCacheTriangleData* tri = mCache[i]->mTriangles[j];
+					LLVCacheTriangleData* tri = data->mTriangles[j];
 					if (tri->mActive)
 					{
 						tri->mScore = tri->mVertex[0]->mScore;
@@ -5753,7 +5757,7 @@ public:
 		{
 			if (mCache[i])
 			{
-				llassert_always(mCache[i]->mCacheTag == -1);
+				llassert(mCache[i]->mCacheTag == -1);
 				mCache[i] = NULL;
 			}
 		}
@@ -5765,6 +5769,9 @@ void LLVolumeFace::cacheOptimize()
 { //optimize for vertex cache according to Forsyth method: 
   // http://home.comcast.net/~tom_forsyth/papers/fast_vert_cache_opt.html
 	
+	llassert(!mOptimized);
+	mOptimized = TRUE;
+
 	LLVCacheLRU cache;
 	
 	if (mNumVertices < 3)
@@ -5810,12 +5817,14 @@ void LLVolumeFace::cacheOptimize()
 
 	for (U32 i = 0; i < mNumVertices; i++)
 	{ //initialize score values (no cache -- might try a fifo cache here)
-		vertex_data[i].mScore = find_vertex_score(vertex_data[i]);
-		vertex_data[i].mActiveTriangles = vertex_data[i].mTriangles.size();
+		LLVCacheVertexData& data = vertex_data[i];
+
+		data.mScore = find_vertex_score(data);
+		data.mActiveTriangles = data.mTriangles.size();
 
-		for (U32 j = 0; j < vertex_data[i].mTriangles.size(); ++j)
+		for (U32 j = 0; j < data.mActiveTriangles; ++j)
 		{
-			vertex_data[i].mTriangles[j]->mScore += vertex_data[i].mScore;
+			data.mTriangles[j]->mScore += data.mScore;
 		}
 	}
 
diff --git a/indra/llmath/llvolume.h b/indra/llmath/llvolume.h
index c845556557..99158c1c44 100644
--- a/indra/llmath/llvolume.h
+++ b/indra/llmath/llvolume.h
@@ -933,6 +933,9 @@ public:
 
 	LLOctreeNode<LLVolumeTriangle>* mOctree;
 
+	//whether or not face has been cache optimized
+	BOOL mOptimized;
+
 private:
 	BOOL createUnCutCubeCap(LLVolume* volume, BOOL partial_build = FALSE);
 	BOOL createCap(LLVolume* volume, BOOL partial_build = FALSE);
diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 9d06dd6904..89f1f36297 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -59,6 +59,7 @@ BOOL gDebugGL = FALSE;
 BOOL gClothRipple = FALSE;
 BOOL gHeadlessClient = FALSE;
 BOOL gGLActive = FALSE;
+BOOL gGLDebugLoggingEnabled = TRUE;
 
 static const std::string HEADLESS_VENDOR_STRING("Linden Lab");
 static const std::string HEADLESS_RENDERER_STRING("Headless");
@@ -72,6 +73,7 @@ std::ofstream gFailLog;
 #define APIENTRY
 #endif
 
+
 void APIENTRY gl_debug_callback(GLenum source,
                                 GLenum type,
                                 GLuint id,
@@ -80,22 +82,25 @@ void APIENTRY gl_debug_callback(GLenum source,
                                 const GLchar* message,
                                 GLvoid* userParam)
 {
-	if (severity == GL_DEBUG_SEVERITY_HIGH_ARB)
-	{
-		llwarns << "----- GL ERROR --------" << llendl;
-	}
-	else
+	if (gGLDebugLoggingEnabled)
 	{
-		llwarns << "----- GL WARNING -------" << llendl;
-	}
-	llwarns << "Type: " << std::hex << type << llendl;
-	llwarns << "ID: " << std::hex << id << llendl;
-	llwarns << "Severity: " << std::hex << severity << llendl;
-	llwarns << "Message: " << message << llendl;
-	llwarns << "-----------------------" << llendl;
-	if (severity == GL_DEBUG_SEVERITY_HIGH_ARB)
-	{
-		llerrs << "Halting on GL Error" << llendl;
+		if (severity == GL_DEBUG_SEVERITY_HIGH_ARB)
+		{
+			llwarns << "----- GL ERROR --------" << llendl;
+		}
+		else
+		{
+			llwarns << "----- GL WARNING -------" << llendl;
+		}
+		llwarns << "Type: " << std::hex << type << llendl;
+		llwarns << "ID: " << std::hex << id << llendl;
+		llwarns << "Severity: " << std::hex << severity << llendl;
+		llwarns << "Message: " << message << llendl;
+		llwarns << "-----------------------" << llendl;
+		if (severity == GL_DEBUG_SEVERITY_HIGH_ARB)
+		{
+			llerrs << "Halting on GL Error" << llendl;
+		}
 	}
 }
 #endif
@@ -253,6 +258,7 @@ PFNGLBEGINTRANSFORMFEEDBACKPROC glBeginTransformFeedback = NULL;
 PFNGLENDTRANSFORMFEEDBACKPROC glEndTransformFeedback = NULL;
 PFNGLTRANSFORMFEEDBACKVARYINGSPROC glTransformFeedbackVaryings = NULL;
 PFNGLBINDBUFFERRANGEPROC glBindBufferRange = NULL;
+PFNGLBINDBUFFERBASEPROC glBindBufferBase = NULL;
 
 //GL_ARB_debug_output
 PFNGLDEBUGMESSAGECONTROLARBPROC glDebugMessageControlARB = NULL;
@@ -1224,6 +1230,7 @@ void LLGLManager::initExtensions()
 		glEndTransformFeedback = (PFNGLENDTRANSFORMFEEDBACKPROC) GLH_EXT_GET_PROC_ADDRESS("glEndTransformFeedback");
 		glTransformFeedbackVaryings = (PFNGLTRANSFORMFEEDBACKVARYINGSPROC) GLH_EXT_GET_PROC_ADDRESS("glTransformFeedbackVaryings");
 		glBindBufferRange = (PFNGLBINDBUFFERRANGEPROC) GLH_EXT_GET_PROC_ADDRESS("glBindBufferRange");
+		glBindBufferBase = (PFNGLBINDBUFFERBASEPROC) GLH_EXT_GET_PROC_ADDRESS("glBindBufferBase");
 	}
 	if (mHasDebugOutput)
 	{
diff --git a/indra/llrender/llglheaders.h b/indra/llrender/llglheaders.h
index 509de51f4d..a92ed428da 100644
--- a/indra/llrender/llglheaders.h
+++ b/indra/llrender/llglheaders.h
@@ -533,6 +533,7 @@ extern PFNGLBEGINTRANSFORMFEEDBACKPROC glBeginTransformFeedback;
 extern PFNGLENDTRANSFORMFEEDBACKPROC glEndTransformFeedback;
 extern PFNGLTRANSFORMFEEDBACKVARYINGSPROC glTransformFeedbackVaryings;
 extern PFNGLBINDBUFFERRANGEPROC glBindBufferRange;
+extern PFNGLBINDBUFFERBASEPROC glBindBufferBase;
 
 
 #elif LL_WINDOWS
@@ -771,6 +772,7 @@ extern PFNGLBEGINTRANSFORMFEEDBACKPROC glBeginTransformFeedback;
 extern PFNGLENDTRANSFORMFEEDBACKPROC glEndTransformFeedback;
 extern PFNGLTRANSFORMFEEDBACKVARYINGSPROC glTransformFeedbackVaryings;
 extern PFNGLBINDBUFFERRANGEPROC glBindBufferRange;
+extern PFNGLBINDBUFFERBASEPROC glBindBufferBase;
 
 //GL_ARB_debug_output
 extern PFNGLDEBUGMESSAGECONTROLARBPROC glDebugMessageControlARB;
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index a12e9caf4c..cb82cbfb74 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -225,46 +225,56 @@ void LLTexUnit::disable(void)
 bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
 {
 	stop_glerror();
-	if (mIndex < 0) return false;
-
-	gGL.flush();
-
-	LLImageGL* gl_tex = NULL ;
-	if (texture == NULL || !(gl_tex = texture->getGLTexture()))
-	{
-		llwarns << "NULL LLTexUnit::bind texture" << llendl;
-		return false;
-	}
-
-	if (!gl_tex->getTexName()) //if texture does not exist
+	if (mIndex >= 0)
 	{
-		//if deleted, will re-generate it immediately
-		texture->forceImmediateUpdate() ;
-
-		gl_tex->forceUpdateBindStats() ;
-		return texture->bindDefaultImage(mIndex);
-	}
+		gGL.flush();
 
-	//in audit, replace the selected texture by the default one.
-	if ((mCurrTexture != gl_tex->getTexName()) || forceBind)
-	{
-		activate();
-		enable(gl_tex->getTarget());
-		mCurrTexture = gl_tex->getTexName();
-		glBindTexture(sGLTextureType[gl_tex->getTarget()], mCurrTexture);
-		if(gl_tex->updateBindStats(gl_tex->mTextureMemory))
+		LLImageGL* gl_tex = NULL ;
+		if (texture != NULL && (gl_tex = texture->getGLTexture()))
 		{
-			texture->setActive() ;
-			texture->updateBindStatsForTester() ;
+			if (gl_tex->getTexName()) //if texture exists
+			{
+				//in audit, replace the selected texture by the default one.
+				if ((mCurrTexture != gl_tex->getTexName()) || forceBind)
+				{
+					activate();
+					enable(gl_tex->getTarget());
+					mCurrTexture = gl_tex->getTexName();
+					glBindTexture(sGLTextureType[gl_tex->getTarget()], mCurrTexture);
+					if(gl_tex->updateBindStats(gl_tex->mTextureMemory))
+					{
+						texture->setActive() ;
+						texture->updateBindStatsForTester() ;
+					}
+					mHasMipMaps = gl_tex->mHasMipMaps;
+					if (gl_tex->mTexOptionsDirty)
+					{
+						gl_tex->mTexOptionsDirty = false;
+						setTextureAddressMode(gl_tex->mAddressMode);
+						setTextureFilteringOption(gl_tex->mFilterOption);
+					}
+				}
+			}
+			else
+			{
+				//if deleted, will re-generate it immediately
+				texture->forceImmediateUpdate() ;
+
+				gl_tex->forceUpdateBindStats() ;
+				return texture->bindDefaultImage(mIndex);
+			}
 		}
-		mHasMipMaps = gl_tex->mHasMipMaps;
-		if (gl_tex->mTexOptionsDirty)
+		else
 		{
-			gl_tex->mTexOptionsDirty = false;
-			setTextureAddressMode(gl_tex->mAddressMode);
-			setTextureFilteringOption(gl_tex->mFilterOption);
+			llwarns << "NULL LLTexUnit::bind texture" << llendl;
+			return false;
 		}
 	}
+	else
+	{ // mIndex < 0
+		return false;
+	}
+
 	return true;
 }
 
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index dfbd8cd4ee..f152911b24 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -85,6 +85,7 @@ const U32 LL_VBO_POOL_SEED_COUNT = vbo_block_index(LL_VBO_POOL_MAX_SEED_SIZE);
 //static
 LLVBOPool LLVertexBuffer::sStreamVBOPool(GL_STREAM_DRAW_ARB, GL_ARRAY_BUFFER_ARB);
 LLVBOPool LLVertexBuffer::sDynamicVBOPool(GL_DYNAMIC_DRAW_ARB, GL_ARRAY_BUFFER_ARB);
+LLVBOPool LLVertexBuffer::sDynamicCopyVBOPool(GL_DYNAMIC_COPY_ARB, GL_ARRAY_BUFFER_ARB);
 LLVBOPool LLVertexBuffer::sStreamIBOPool(GL_STREAM_DRAW_ARB, GL_ELEMENT_ARRAY_BUFFER_ARB);
 LLVBOPool LLVertexBuffer::sDynamicIBOPool(GL_DYNAMIC_DRAW_ARB, GL_ELEMENT_ARRAY_BUFFER_ARB);
 
@@ -199,7 +200,10 @@ volatile U8* LLVBOPool::allocate(U32& name, U32 size, bool for_seed)
 		if (LLVertexBuffer::sDisableVBOMapping || mUsage != GL_DYNAMIC_DRAW_ARB)
 		{
 			glBufferDataARB(mType, size, 0, mUsage);
-			ret = (U8*) ll_aligned_malloc_16(size);
+			if (mUsage != GL_DYNAMIC_COPY_ARB)
+			{ //data will be provided by application
+				ret = (U8*) ll_aligned_malloc_16(size);
+			}
 		}
 		else
 		{ //always use a true hint of static draw when allocating non-client-backed buffers
@@ -393,6 +397,7 @@ void LLVertexBuffer::seedPools()
 {
 	sStreamVBOPool.seedPool();
 	sDynamicVBOPool.seedPool();
+	sDynamicCopyVBOPool.seedPool();
 	sStreamIBOPool.seedPool();
 	sDynamicIBOPool.seedPool();
 }
@@ -875,6 +880,7 @@ void LLVertexBuffer::cleanupClass()
 	sDynamicIBOPool.cleanup();
 	sStreamVBOPool.cleanup();
 	sDynamicVBOPool.cleanup();
+	sDynamicCopyVBOPool.cleanup();
 
 	if(sPrivatePoolp)
 	{
@@ -911,13 +917,16 @@ S32 LLVertexBuffer::determineUsage(S32 usage)
 	
 	if (ret_usage && ret_usage != GL_STREAM_DRAW_ARB)
 	{ //only stream_draw and dynamic_draw are supported when using VBOs, dynamic draw is the default
-		if (sDisableVBOMapping)
-		{ //always use stream draw if VBO mapping is disabled
-			ret_usage = GL_STREAM_DRAW_ARB;
-		}
-		else
+		if (ret_usage != GL_DYNAMIC_COPY_ARB)
 		{
-			ret_usage = GL_DYNAMIC_DRAW_ARB;
+			if (sDisableVBOMapping)
+			{ //always use stream draw if VBO mapping is disabled
+				ret_usage = GL_STREAM_DRAW_ARB;
+			}
+			else
+			{
+				ret_usage = GL_DYNAMIC_DRAW_ARB;
+			}
 		}
 	}
 	
@@ -1067,10 +1076,15 @@ void LLVertexBuffer::genBuffer(U32 size)
 	{
 		mMappedData = sStreamVBOPool.allocate(mGLBuffer, mSize);
 	}
-	else
+	else if (mUsage == GL_DYNAMIC_DRAW_ARB)
 	{
 		mMappedData = sDynamicVBOPool.allocate(mGLBuffer, mSize);
 	}
+	else
+	{
+		mMappedData = sDynamicCopyVBOPool.allocate(mGLBuffer, mSize);
+	}
+
 	
 	sGLCount++;
 }
@@ -1820,6 +1834,8 @@ void LLVertexBuffer::unmapBuffer()
 
 	if (mMappedData && mVertexLocked)
 	{
+		llassert(mUsage != GL_DYNAMIC_COPY_ARB);
+		
 		LLFastTimer t(FTM_VBO_UNMAP);
 		bindGLBuffer(true);
 		updated_all = mIndexLocked; //both vertex and index buffers done updating
@@ -2094,22 +2110,15 @@ bool LLVertexBuffer::bindGLBuffer(bool force_bind)
 
 	if (useVBOs() && (force_bind || (mGLBuffer && (mGLBuffer != sGLRenderBuffer || !sVBOActive))))
 	{
-		LLFastTimer t(FTM_BIND_GL_BUFFER);
-		/*if (sMapped)
-		{
-			llerrs << "VBO bound while another VBO mapped!" << llendl;
-		}*/
+		//LLFastTimer t(FTM_BIND_GL_BUFFER); <-- this timer is showing up as a hotspot (irony)
+		
 		glBindBufferARB(GL_ARRAY_BUFFER_ARB, mGLBuffer);
 		sGLRenderBuffer = mGLBuffer;
 		sBindCount++;
 		sVBOActive = true;
 
-		if (mGLArray)
-		{
-			llassert(sGLRenderArray == mGLArray);
-			//mCachedRenderBuffer = mGLBuffer;
-		}
-
+		llassert(!mGLArray || sGLRenderArray == mGLArray);
+		
 		ret = true;
 	}
 
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index 11fa4ab6a0..a3400ae80c 100644
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -125,9 +125,10 @@ public:
 
 	static LLVBOPool sStreamVBOPool;
 	static LLVBOPool sDynamicVBOPool;
+	static LLVBOPool sDynamicCopyVBOPool;
 	static LLVBOPool sStreamIBOPool;
 	static LLVBOPool sDynamicIBOPool;
-
+	
 	static std::list<U32> sAvailableVAOName;
 	static U32 sCurVAOName;
 
diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp
index a7e225843c..6b3127decf 100755
--- a/indra/newview/llface.cpp
+++ b/indra/newview/llface.cpp
@@ -53,6 +53,7 @@
 #include "llviewershadermgr.h"
 #include "llvoavatar.h"
 
+extern BOOL gGLDebugLoggingEnabled;
 
 #define LL_MAX_INDICES_COUNT 1000000
 
@@ -1166,6 +1167,15 @@ static LLFastTimer::DeclareTimer FTM_FACE_GEOM_COLOR("Color");
 static LLFastTimer::DeclareTimer FTM_FACE_GEOM_EMISSIVE("Emissive");
 static LLFastTimer::DeclareTimer FTM_FACE_GEOM_WEIGHTS("Weights");
 static LLFastTimer::DeclareTimer FTM_FACE_GEOM_BINORMAL("Binormal");
+
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK("Face Feedback");
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_POSITION("Feedback Position");
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_NORMAL("Feedback  Normal");
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_TEXTURE("Feedback  Texture");
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_COLOR("Feedback  Color");
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_EMISSIVE("Feedback  Emissive");
+static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_BINORMAL("Feedback Binormal");
+
 static LLFastTimer::DeclareTimer FTM_FACE_GEOM_INDEX("Index");
 static LLFastTimer::DeclareTimer FTM_FACE_GEOM_INDEX_TAIL("Tail");
 static LLFastTimer::DeclareTimer FTM_FACE_POSITION_STORE("Pos");
@@ -1389,12 +1399,14 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 #ifdef GL_TRANSFORM_FEEDBACK_BUFFER
 	if (use_transform_feedback &&
+		mVertexBuffer->getUsage() == GL_DYNAMIC_COPY_ARB &&
 		gTransformPositionProgram.mProgramObject && //transform shaders are loaded
 		mVertexBuffer->useVBOs() && //target buffer is in VRAM
 		!rebuild_weights && //TODO: add support for weights
 		!volume.isUnique()) //source volume is NOT flexi
 	{ //use transform feedback to pack vertex buffer
-
+		//gGLDebugLoggingEnabled = TRUE;
+		LLFastTimer t(FTM_FACE_GEOM_FEEDBACK);
 		LLVertexBuffer* buff = (LLVertexBuffer*) vf.mVertexBuffer.get();
 
 		if (vf.mVertexBuffer.isNull() || buff->getNumVerts() != vf.mNumVertices)
@@ -1411,7 +1423,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (rebuild_pos)
 		{
-			LLFastTimer t(FTM_FACE_GEOM_POSITION);
+			LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_POSITION);
 			gTransformPositionProgram.bind();
 
 			mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_VERTEX, mGeomIndex, mGeomCount);
@@ -1436,7 +1448,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (rebuild_color)
 		{
-			LLFastTimer t(FTM_FACE_GEOM_COLOR);
+			LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_COLOR);
 			gTransformColorProgram.bind();
 			
 			mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_COLOR, mGeomIndex, mGeomCount);
@@ -1452,7 +1464,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (rebuild_emissive)
 		{
-			LLFastTimer t(FTM_FACE_GEOM_EMISSIVE);
+			LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_EMISSIVE);
 			gTransformColorProgram.bind();
 			
 			mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_EMISSIVE, mGeomIndex, mGeomCount);
@@ -1473,7 +1485,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (rebuild_normal)
 		{
-			LLFastTimer t(FTM_FACE_GEOM_NORMAL);
+			LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_NORMAL);
 			gTransformNormalProgram.bind();
 			
 			mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_NORMAL, mGeomIndex, mGeomCount);
@@ -1486,7 +1498,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (rebuild_binormal)
 		{
-			LLFastTimer t(FTM_FACE_GEOM_BINORMAL);
+			LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_BINORMAL);
 			gTransformBinormalProgram.bind();
 			
 			mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_BINORMAL, mGeomIndex, mGeomCount);
@@ -1499,7 +1511,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (rebuild_tcoord)
 		{
-			LLFastTimer t(FTM_FACE_GEOM_TEXTURE);
+			LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_TEXTURE);
 			gTransformTexCoordProgram.bind();
 			
 			mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_TEXCOORD0, mGeomIndex, mGeomCount);
@@ -1522,13 +1534,13 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 		}
 
 		glBindBufferARB(GL_TRANSFORM_FEEDBACK_BUFFER, 0);
-
 		gGL.popMatrix();
 
 		if (cur_shader)
 		{
 			cur_shader->bind();
 		}
+		//gGLDebugLoggingEnabled = FALSE;
 	}
 	else
 #endif
diff --git a/indra/newview/llmeshrepository.cpp b/indra/newview/llmeshrepository.cpp
index 5b65687090..ae48898e82 100755
--- a/indra/newview/llmeshrepository.cpp
+++ b/indra/newview/llmeshrepository.cpp
@@ -588,7 +588,7 @@ void LLMeshRepoThread::run()
 					if (!fetchMeshLOD(req.mMeshParams, req.mLOD, count))//failed, resubmit
 					{
 						mMutex->lock();
-						mLODReqQ.push(req) ; 
+						mLODReqQ.push(req); 
 						mMutex->unlock();
 					}
 				}
diff --git a/indra/newview/llvovolume.cpp b/indra/newview/llvovolume.cpp
index b853112f74..6a18534484 100644
--- a/indra/newview/llvovolume.cpp
+++ b/indra/newview/llvovolume.cpp
@@ -99,6 +99,8 @@ static LLFastTimer::DeclareTimer FTM_GEN_TRIANGLES("Generate Triangles");
 static LLFastTimer::DeclareTimer FTM_GEN_VOLUME("Generate Volumes");
 static LLFastTimer::DeclareTimer FTM_VOLUME_TEXTURES("Volume Textures");
 
+extern BOOL gGLDebugLoggingEnabled;
+
 // Implementation class of LLMediaDataClientObject.  See llmediadataclient.h
 class LLMediaDataClientObjectImpl : public LLMediaDataClientObject
 {
@@ -1067,7 +1069,9 @@ BOOL LLVOVolume::setVolume(const LLVolumeParams &params_in, const S32 detail, bo
 					break;
 				}
 				volume->genBinormals(i);
+				//gGLDebugLoggingEnabled = TRUE;
 				LLFace::cacheFaceInVRAM(face);
+				//gGLDebugLoggingEnabled = FALSE;
 			}
 		}
 		
@@ -4836,6 +4840,16 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 
 	U32 buffer_usage = group->mBufferUsage;
 	
+	static LLCachedControl<bool> use_transform_feedback(gSavedSettings, "RenderUseTransformFeedback");
+
+	if (use_transform_feedback &&
+		gTransformPositionProgram.mProgramObject && //transform shaders are loaded
+		buffer_usage == GL_DYNAMIC_DRAW_ARB && //target buffer is in VRAM
+		!(mask & LLVertexBuffer::MAP_WEIGHT4)) //TODO: add support for weights
+	{
+		buffer_usage = GL_DYNAMIC_COPY_ARB;
+	}
+
 #if LL_DARWIN
 	// HACK from Leslie:
 	// Disable VBO usage for alpha on Mac OS X because it kills the framerate
@@ -4895,6 +4909,8 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 	//NEVER use more than 16 texture index channels (workaround for prevalent driver bug)
 	texture_index_channels = llmin(texture_index_channels, 16);
 
+	bool flexi = false;
+
 	while (face_iter != faces.end())
 	{
 		//pull off next face
@@ -4921,6 +4937,8 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 		U32 index_count = facep->getIndicesCount();
 		U32 geom_count = facep->getGeomCount();
 
+		flexi = flexi || facep->getViewerObject()->getVolume()->isUnique();
+
 		//sum up vertices needed for this render batch
 		std::vector<LLFace*>::iterator i = face_iter;
 		++i;
@@ -4989,6 +5007,9 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 						}
 
 						++i;
+
+						flexi = flexi || facep->getViewerObject()->getVolume()->isUnique();
+
 						index_count += facep->getIndicesCount();
 						geom_count += facep->getGeomCount();
 
@@ -5018,10 +5039,18 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 					++i;
 					index_count += facep->getIndicesCount();
 					geom_count += facep->getGeomCount();
+
+					flexi = flexi || facep->getViewerObject()->getVolume()->isUnique();
 				}
 			}
 		}
 
+
+		if (flexi && buffer_usage && buffer_usage != GL_STREAM_DRAW_ARB)
+		{
+			buffer_usage = GL_STREAM_DRAW_ARB;
+		}
+
 		//create vertex buffer
 		LLVertexBuffer* buffer = NULL;