MAINT-2371 First set of profile guided optimizations.

Reviewed by Graham

3 files changed, 50 insertions, 9 deletions

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;