Performance improvements. Faster threadsaferefcount, faster instancetracker, use hashtable and pre-hashed strings for GLSL uniform lookup, and use dup_handles curl interface to avoid re-initing ares resolvers expensively.

5 files changed, 79 insertions, 50 deletions

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..5801946a74 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -225,26 +225,15 @@ void LLTexUnit::disable(void)
 bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
 {
 	stop_glerror();
-	if (mIndex < 0) return false;
-
+	if (mIndex >= 0)
+	{
 	gGL.flush();
 
 	LLImageGL* gl_tex = NULL ;
-	if (texture == NULL || !(gl_tex = texture->getGLTexture()))
+		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 (gl_tex->getTexName()) //if texture exists
 	{
-		//if deleted, will re-generate it immediately
-		texture->forceImmediateUpdate() ;
-
-		gl_tex->forceUpdateBindStats() ;
-		return texture->bindDefaultImage(mIndex);
-	}
-
 	//in audit, replace the selected texture by the default one.
 	if ((mCurrTexture != gl_tex->getTexName()) || forceBind)
 	{
@@ -265,6 +254,27 @@ bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
 			setTextureFilteringOption(gl_tex->mFilterOption);
 		}
 	}
+			}
+			else
+			{
+				//if deleted, will re-generate it immediately
+				texture->forceImmediateUpdate() ;
+
+				gl_tex->forceUpdateBindStats() ;
+				return texture->bindDefaultImage(mIndex);
+			}
+		}
+		else
+		{
+			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;