SL-16099 Multi-threaded OpenGL usage on Windows, enable Core Profile and VAOs by default.

8 files changed, 252 insertions, 117 deletions

diff --git a/indra/newview/app_settings/settings.xml b/indra/newview/app_settings/settings.xml
index 05c3fc3bfe..d05d100ddf 100644
--- a/indra/newview/app_settings/settings.xml
+++ b/indra/newview/app_settings/settings.xml
@@ -9814,7 +9814,7 @@
       <key>Type</key>
       <string>Boolean</string>
       <key>Value</key>
-      <integer>0</integer>
+      <integer>1</integer>
     </map>
     <key>RenderGlow</key>
     <map>
@@ -10599,7 +10599,7 @@
       <key>Type</key>
       <string>Boolean</string>
       <key>Value</key>
-      <integer>0</integer>
+      <integer>1</integer>
     </map>
   <key>RenderUseTransformFeedback</key>
   <map>
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp
index 722a6caa65..e46e8feb14 100644
--- a/indra/newview/llappviewer.cpp
+++ b/indra/newview/llappviewer.cpp
@@ -613,7 +613,7 @@ static void settings_modify()
 	LLPipeline::sRenderDeferred		= LLPipeline::sRenderTransparentWater && LLPipeline::sRenderBump && gSavedSettings.getBOOL("RenderDeferred");
 	LLVOSurfacePatch::sLODFactor		= gSavedSettings.getF32("RenderTerrainLODFactor");
 	LLVOSurfacePatch::sLODFactor *= LLVOSurfacePatch::sLODFactor; //square lod factor to get exponential range of [1,4]
-	gDebugGL = gSavedSettings.getBOOL("RenderDebugGL") || gDebugSession;
+    gDebugGL = gSavedSettings.getBOOL("RenderDebugGL") || gDebugSession;
 	gDebugPipeline = gSavedSettings.getBOOL("RenderDebugPipeline");
 }
 
@@ -4831,6 +4831,7 @@ void LLAppViewer::idle()
 	LLNotificationsUI::LLToast::updateClass();
 	LLSmoothInterpolation::updateInterpolants();
 	LLMortician::updateClass();
+    LLImageGL::updateClass();
 	LLFilePickerThread::clearDead();  //calls LLFilePickerThread::notify()
 	LLDirPickerThread::clearDead();
 	F32 dt_raw = idle_timer.getElapsedTimeAndResetF32();
diff --git a/indra/newview/llnetmap.cpp b/indra/newview/llnetmap.cpp
index 112da55682..111b45612e 100644
--- a/indra/newview/llnetmap.cpp
+++ b/indra/newview/llnetmap.cpp
@@ -147,6 +147,7 @@ void LLNetMap::setScale( F32 scale )
 
 void LLNetMap::draw()
 {
+    LL_PROFILE_ZONE_SCOPED;
  	static LLFrameTimer map_timer;
 	static LLUIColor map_avatar_color = LLUIColorTable::instance().getColor("MapAvatarColor", LLColor4::white);
 	static LLUIColor map_avatar_friend_color = LLUIColorTable::instance().getColor("MapAvatarFriendColor", LLColor4::white);
diff --git a/indra/newview/lltexturefetch.cpp b/indra/newview/lltexturefetch.cpp
index f64db7beb5..63e561147d 100644
--- a/indra/newview/lltexturefetch.cpp
+++ b/indra/newview/lltexturefetch.cpp
@@ -314,6 +314,7 @@ private:
 		// Threads:  Ttc
 		virtual void completed(bool success)
 		{
+            LL_PROFILE_ZONE_SCOPED;
 			LLTextureFetchWorker* worker = mFetcher->getWorker(mID);
 			if (worker)
 			{
@@ -338,6 +339,7 @@ private:
 		// Threads:  Ttc
 		virtual void completed(bool success)
 		{
+            LL_PROFILE_ZONE_SCOPED;
 			LLTextureFetchWorker* worker = mFetcher->getWorker(mID);
 			if (worker)
 			{
@@ -362,6 +364,7 @@ private:
 		// Threads:  Tid
 		virtual void completed(bool success, LLImageRaw* raw, LLImageRaw* aux)
 		{
+            LL_PROFILE_ZONE_SCOPED;
 			LLTextureFetchWorker* worker = mFetcher->getWorker(mID);
 			if (worker)
 			{
@@ -1139,6 +1142,7 @@ void LLTextureFetchWorker::startWork(S32 param)
 // Threads:  Ttf
 bool LLTextureFetchWorker::doWork(S32 param)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	static const LLCore::HttpStatus http_not_found(HTTP_NOT_FOUND);						// 404
 	static const LLCore::HttpStatus http_service_unavail(HTTP_SERVICE_UNAVAILABLE);		// 503
 	static const LLCore::HttpStatus http_not_sat(HTTP_REQUESTED_RANGE_NOT_SATISFIABLE);	// 416;
diff --git a/indra/newview/llviewertexture.cpp b/indra/newview/llviewertexture.cpp
index af55f68cd2..274f53a160 100644
--- a/indra/newview/llviewertexture.cpp
+++ b/indra/newview/llviewertexture.cpp
@@ -184,6 +184,7 @@ void LLViewerTextureManager::findFetchedTextures(const LLUUID& id, std::vector<L
 
 void  LLViewerTextureManager::findTextures(const LLUUID& id, std::vector<LLViewerTexture*> &output)
 {
+    LL_PROFILE_ZONE_SCOPED;
     std::vector<LLViewerFetchedTexture*> fetched_output;
     gTextureList.findTexturesByID(id, fetched_output);
     std::vector<LLViewerFetchedTexture*>::iterator iter = fetched_output.begin();
@@ -485,6 +486,7 @@ static LLTrace::BlockTimerStatHandle FTM_TEXTURE_MEMORY_CHECK("Memory Check");
 //static 
 bool LLViewerTexture::isMemoryForTextureLow()
 {
+    LL_PROFILE_ZONE_SCOPED;
     // Note: we need to figure out a better source for 'min' values,
     // what is free for low end at minimal settings is 'nothing left'
     // for higher end gpus at high settings.
@@ -501,6 +503,7 @@ bool LLViewerTexture::isMemoryForTextureLow()
 //static
 bool LLViewerTexture::isMemoryForTextureSuficientlyFree()
 {
+    LL_PROFILE_ZONE_SCOPED;
     const S32Megabytes DESIRED_FREE_TEXTURE_MEMORY(50);
     const S32Megabytes DESIRED_FREE_MAIN_MEMORY(200);
 
@@ -514,6 +517,7 @@ bool LLViewerTexture::isMemoryForTextureSuficientlyFree()
 //static
 void LLViewerTexture::getGPUMemoryForTextures(S32Megabytes &gpu, S32Megabytes &physical)
 {
+    LL_PROFILE_ZONE_SCOPED;
     static LLFrameTimer timer;
     static S32Megabytes gpu_res = S32Megabytes(S32_MAX);
     static S32Megabytes physical_res = S32Megabytes(S32_MAX);
@@ -526,27 +530,29 @@ void LLViewerTexture::getGPUMemoryForTextures(S32Megabytes &gpu, S32Megabytes &p
     }
     timer.reset();
 
-    LL_RECORD_BLOCK_TIME(FTM_TEXTURE_MEMORY_CHECK);
-
-    if (gGLManager.mHasATIMemInfo)
     {
-        S32 meminfo[4];
-        glGetIntegerv(GL_TEXTURE_FREE_MEMORY_ATI, meminfo);
-        gpu_res = (S32Megabytes)meminfo[0];
+        LL_RECORD_BLOCK_TIME(FTM_TEXTURE_MEMORY_CHECK);
 
-        //check main memory, only works for windows.
-        LLMemory::updateMemoryInfo();
-        physical_res = LLMemory::getAvailableMemKB();
-    }
-    else if (gGLManager.mHasNVXMemInfo)
-    {
-        S32 free_memory;
-        glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &free_memory);
-        gpu_res = (S32Megabytes)(free_memory / 1024);
-    }
+        if (gGLManager.mHasATIMemInfo)
+        {
+            S32 meminfo[4];
+            glGetIntegerv(GL_TEXTURE_FREE_MEMORY_ATI, meminfo);
+            gpu_res = (S32Megabytes)meminfo[0];
 
-    gpu = gpu_res;
-    physical = physical_res;
+            //check main memory, only works for windows.
+            LLMemory::updateMemoryInfo();
+            physical_res = LLMemory::getAvailableMemKB();
+        }
+        else if (gGLManager.mHasNVXMemInfo)
+        {
+            S32 free_memory;
+            glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &free_memory);
+            gpu_res = (S32Megabytes)(free_memory / 1024);
+        }
+
+        gpu = gpu_res;
+        physical = physical_res;
+    }
 }
 
 static LLTrace::BlockTimerStatHandle FTM_TEXTURE_UPDATE_MEDIA("Media");
@@ -555,6 +561,7 @@ static LLTrace::BlockTimerStatHandle FTM_TEXTURE_UPDATE_TEST("Test");
 //static
 void LLViewerTexture::updateClass(const F32 velocity, const F32 angular_velocity)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	sCurrentTime = gFrameTimeSeconds;
 
 	LLTexturePipelineTester* tester = (LLTexturePipelineTester*)LLMetricPerformanceTesterBasic::getTester(sTesterName);
@@ -704,6 +711,7 @@ void LLViewerTexture::cleanup()
 
 void LLViewerTexture::notifyAboutCreatingTexture()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	for(U32 ch = 0; ch < LLRender::NUM_TEXTURE_CHANNELS; ++ch)
 	{
 		for(U32 f = 0; f < mNumFaces[ch]; f++)
@@ -715,6 +723,7 @@ void LLViewerTexture::notifyAboutCreatingTexture()
 
 void LLViewerTexture::notifyAboutMissingAsset()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	for(U32 ch = 0; ch < LLRender::NUM_TEXTURE_CHANNELS; ++ch)
 	{
 		for(U32 f = 0; f < mNumFaces[ch]; f++)
@@ -727,6 +736,7 @@ void LLViewerTexture::notifyAboutMissingAsset()
 // virtual
 void LLViewerTexture::dump()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	LLGLTexture::dump();
 
 	LL_INFOS() << "LLViewerTexture"
@@ -762,6 +772,7 @@ bool LLViewerTexture::isActiveFetching()
 
 bool LLViewerTexture::bindDebugImage(const S32 stage)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (stage < 0) return false;
 
 	bool res = true;
@@ -780,6 +791,7 @@ bool LLViewerTexture::bindDebugImage(const S32 stage)
 
 bool LLViewerTexture::bindDefaultImage(S32 stage) 
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (stage < 0) return false;
 
 	bool res = true;
@@ -822,6 +834,7 @@ void LLViewerTexture::forceImmediateUpdate()
 
 void LLViewerTexture::addTextureStats(F32 virtual_size, BOOL needs_gltexture) const 
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(needs_gltexture)
 	{
 		mNeedsGLTexture = TRUE;
@@ -864,6 +877,7 @@ void LLViewerTexture::setKnownDrawSize(S32 width, S32 height)
 //virtual
 void LLViewerTexture::addFace(U32 ch, LLFace* facep) 
 {
+    LL_PROFILE_ZONE_SCOPED;
 	llassert(ch < LLRender::NUM_TEXTURE_CHANNELS);
 
 	if(mNumFaces[ch] >= mFaceList[ch].size())
@@ -879,6 +893,7 @@ void LLViewerTexture::addFace(U32 ch, LLFace* facep)
 //virtual
 void LLViewerTexture::removeFace(U32 ch, LLFace* facep) 
 {
+    LL_PROFILE_ZONE_SCOPED;
 	llassert(ch < LLRender::NUM_TEXTURE_CHANNELS);
 
 	if(mNumFaces[ch] > 1)
@@ -919,6 +934,7 @@ S32 LLViewerTexture::getNumFaces(U32 ch) const
 //virtual
 void LLViewerTexture::addVolume(U32 ch, LLVOVolume* volumep)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (mNumVolumes[ch] >= mVolumeList[ch].size())
 	{
 		mVolumeList[ch].resize(2 * mNumVolumes[ch] + 1);
@@ -932,6 +948,7 @@ void LLViewerTexture::addVolume(U32 ch, LLVOVolume* volumep)
 //virtual
 void LLViewerTexture::removeVolume(U32 ch, LLVOVolume* volumep)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (mNumVolumes[ch] > 1)
 	{
 		S32 index = volumep->getIndexInTex(ch); 
@@ -955,6 +972,7 @@ S32 LLViewerTexture::getNumVolumes(U32 ch) const
 
 void LLViewerTexture::reorganizeFaceList()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	static const F32 MAX_WAIT_TIME = 20.f; // seconds
 	static const U32 MAX_EXTRA_BUFFER_SIZE = 4;
 
@@ -978,6 +996,7 @@ void LLViewerTexture::reorganizeFaceList()
 
 void LLViewerTexture::reorganizeVolumeList()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	static const F32 MAX_WAIT_TIME = 20.f; // seconds
 	static const U32 MAX_EXTRA_BUFFER_SIZE = 4;
 
@@ -1180,6 +1199,7 @@ FTType LLViewerFetchedTexture::getFTType() const
 
 void LLViewerFetchedTexture::cleanup()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	for(callback_list_t::iterator iter = mLoadedCallbackList.begin();
 		iter != mLoadedCallbackList.end(); )
 	{
@@ -1205,6 +1225,7 @@ void LLViewerFetchedTexture::cleanup()
 //access the fast cache
 void LLViewerFetchedTexture::loadFromFastCache()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!mInFastCacheList)
 	{
 		return; //no need to access the fast cache.
@@ -1350,6 +1371,7 @@ void LLViewerFetchedTexture::dump()
 // ONLY called from LLViewerFetchedTextureList
 void LLViewerFetchedTexture::destroyTexture() 
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(LLImageGL::sGlobalTextureMemory < sMaxDesiredTextureMem * 0.95f)//not ready to release unused memory.
 	{
 		return ;
@@ -1366,6 +1388,7 @@ void LLViewerFetchedTexture::destroyTexture()
 
 void LLViewerFetchedTexture::addToCreateTexture()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	bool force_update = false;
 	if (getComponents() != mRawImage->getComponents())
 	{
@@ -1407,6 +1430,7 @@ void LLViewerFetchedTexture::addToCreateTexture()
 	}
 	else
 	{	
+        LL_PROFILE_ZONE_SCOPED;
 #if 1
 		//
 		//if mRequestedDiscardLevel > mDesiredDiscardLevel, we assume the required image res keep going up,
@@ -1451,99 +1475,100 @@ void LLViewerFetchedTexture::addToCreateTexture()
 			}
 		}
 #endif
-		mNeedsCreateTexture = TRUE;
-		gTextureList.mCreateTextureList.insert(this);
-	}	
+        scheduleCreateTexture();
+	}
 	return;
 }
 
 // ONLY called from LLViewerTextureList
-BOOL LLViewerFetchedTexture::createTexture(S32 usename/*= 0*/)
+BOOL LLViewerFetchedTexture::preCreateTexture(S32 usename/*= 0*/)
 {
-	if (!mNeedsCreateTexture)
-	{
-		destroyRawImage();
-		return FALSE;
-	}
-	mNeedsCreateTexture = FALSE;
-	if (mRawImage.isNull())
-	{
-		LL_ERRS() << "LLViewerTexture trying to create texture with no Raw Image" << LL_ENDL;
-	}
-	if (mRawImage->isBufferInvalid())
-	{
-		LL_WARNS() << "Can't create a texture: invalid image data" << LL_ENDL;
-		destroyRawImage();
-		return FALSE;
-	}
-// 	LL_INFOS() << llformat("IMAGE Creating (%d) [%d x %d] Bytes: %d ",
-// 						mRawDiscardLevel, 
-// 						mRawImage->getWidth(), mRawImage->getHeight(),mRawImage->getDataSize())
-// 			<< mID.getString() << LL_ENDL;
-	BOOL res = TRUE;
+    LL_PROFILE_ZONE_SCOPED;
+    if (!mNeedsCreateTexture)
+    {
+        destroyRawImage();
+        return FALSE;
+    }
+    mNeedsCreateTexture = FALSE;
 
-	// store original size only for locally-sourced images
-	if (mUrl.compare(0, 7, "file://") == 0)
-	{
-		mOrigWidth = mRawImage->getWidth();
-		mOrigHeight = mRawImage->getHeight();
+    if (mRawImage.isNull())
+    {
+        LL_ERRS() << "LLViewerTexture trying to create texture with no Raw Image" << LL_ENDL;
+    }
+    if (mRawImage->isBufferInvalid())
+    {
+        LL_WARNS() << "Can't create a texture: invalid image data" << LL_ENDL;
+        destroyRawImage();
+        return FALSE;
+    }
+    // 	LL_INFOS() << llformat("IMAGE Creating (%d) [%d x %d] Bytes: %d ",
+    // 						mRawDiscardLevel, 
+    // 						mRawImage->getWidth(), mRawImage->getHeight(),mRawImage->getDataSize())
+    // 			<< mID.getString() << LL_ENDL;
+    BOOL res = TRUE;
+
+    // store original size only for locally-sourced images
+    if (mUrl.compare(0, 7, "file://") == 0)
+    {
+        mOrigWidth = mRawImage->getWidth();
+        mOrigHeight = mRawImage->getHeight();
 
         // This is only safe because it's a local image and fetcher doesn't use raw data
         // from local images, but this might become unsafe in case of changes to fetcher
-		if (mBoostLevel == BOOST_PREVIEW)
-		{ 
-			mRawImage->biasedScaleToPowerOfTwo(1024);
-		}
-		else
-		{ // leave black border, do not scale image content
-			mRawImage->expandToPowerOfTwo(MAX_IMAGE_SIZE, FALSE);
-		}
-		
-		mFullWidth = mRawImage->getWidth();
-		mFullHeight = mRawImage->getHeight();
-		setTexelsPerImage();
-	}
-	else
-	{
-		mOrigWidth = mFullWidth;
-		mOrigHeight = mFullHeight;
-	}
+        if (mBoostLevel == BOOST_PREVIEW)
+        {
+            mRawImage->biasedScaleToPowerOfTwo(1024);
+        }
+        else
+        { // leave black border, do not scale image content
+            mRawImage->expandToPowerOfTwo(MAX_IMAGE_SIZE, FALSE);
+        }
 
-	bool size_okay = true;
+        mFullWidth = mRawImage->getWidth();
+        mFullHeight = mRawImage->getHeight();
+        setTexelsPerImage();
+    }
+    else
+    {
+        mOrigWidth = mFullWidth;
+        mOrigHeight = mFullHeight;
+    }
 
-	S32 discard_level = mRawDiscardLevel;
-	if (mRawDiscardLevel < 0)
-	{
-		LL_DEBUGS() << "Negative raw discard level when creating image: " << mRawDiscardLevel << LL_ENDL;
-		discard_level = 0;
-	}
+    bool size_okay = true;
 
-	U32 raw_width = mRawImage->getWidth() << discard_level;
-	U32 raw_height = mRawImage->getHeight() << discard_level;
+    S32 discard_level = mRawDiscardLevel;
+    if (mRawDiscardLevel < 0)
+    {
+        LL_DEBUGS() << "Negative raw discard level when creating image: " << mRawDiscardLevel << LL_ENDL;
+        discard_level = 0;
+    }
 
-	if( raw_width > MAX_IMAGE_SIZE || raw_height > MAX_IMAGE_SIZE )
-	{
-		LL_INFOS() << "Width or height is greater than " << MAX_IMAGE_SIZE << ": (" << raw_width << "," << raw_height << ")" << LL_ENDL;
-		size_okay = false;
-	}
-	
-	if (!LLImageGL::checkSize(mRawImage->getWidth(), mRawImage->getHeight()))
-	{
-		// A non power-of-two image was uploaded (through a non standard client)
-		LL_INFOS() << "Non power of two width or height: (" << mRawImage->getWidth() << "," << mRawImage->getHeight() << ")" << LL_ENDL;
-		size_okay = false;
-	}
-	
-	if( !size_okay )
-	{
-		// An inappropriately-sized image was uploaded (through a non standard client)
-		// We treat these images as missing assets which causes them to
-		// be renderd as 'missing image' and to stop requesting data
-		LL_WARNS() << "!size_ok, setting as missing" << LL_ENDL;
-		setIsMissingAsset();
-		destroyRawImage();
-		return FALSE;
-	}
+    U32 raw_width = mRawImage->getWidth() << discard_level;
+    U32 raw_height = mRawImage->getHeight() << discard_level;
+
+    if (raw_width > MAX_IMAGE_SIZE || raw_height > MAX_IMAGE_SIZE)
+    {
+        LL_INFOS() << "Width or height is greater than " << MAX_IMAGE_SIZE << ": (" << raw_width << "," << raw_height << ")" << LL_ENDL;
+        size_okay = false;
+    }
+
+    if (!LLImageGL::checkSize(mRawImage->getWidth(), mRawImage->getHeight()))
+    {
+        // A non power-of-two image was uploaded (through a non standard client)
+        LL_INFOS() << "Non power of two width or height: (" << mRawImage->getWidth() << "," << mRawImage->getHeight() << ")" << LL_ENDL;
+        size_okay = false;
+    }
+
+    if (!size_okay)
+    {
+        // An inappropriately-sized image was uploaded (through a non standard client)
+        // We treat these images as missing assets which causes them to
+        // be renderd as 'missing image' and to stop requesting data
+        LL_WARNS() << "!size_ok, setting as missing" << LL_ENDL;
+        setIsMissingAsset();
+        destroyRawImage();
+        return FALSE;
+    }
 
     if (mGLTexturep->getHasExplicitFormat())
     {
@@ -1565,19 +1590,79 @@ BOOL LLViewerFetchedTexture::createTexture(S32 usename/*= 0*/)
         }
     }
 
-	res = mGLTexturep->createGLTexture(mRawDiscardLevel, mRawImage, usename, TRUE, mBoostLevel);
+    return res;
+}
 
-	notifyAboutCreatingTexture();
+BOOL LLViewerFetchedTexture::createTexture(S32 usename/*= 0*/)
+{
+    if (!mNeedsCreateTexture)
+    {
+        return FALSE;
+    }
 
-	setActive();
+	BOOL res = mGLTexturep->createGLTexture(mRawDiscardLevel, mRawImage, usename, TRUE, mBoostLevel);
+    
+	return res;
+}
 
-	if (!needsToSaveRawImage())
-	{
-		mNeedsAux = FALSE;
-		destroyRawImage();
-	}
+void LLViewerFetchedTexture::postCreateTexture()
+{
+    if (!mNeedsCreateTexture)
+    {
+        return;
+    }
 
-	return res;
+    notifyAboutCreatingTexture();
+
+    setActive();
+
+    if (!needsToSaveRawImage())
+    {
+        mNeedsAux = FALSE;
+        destroyRawImage();
+    }
+
+    mNeedsCreateTexture = FALSE;
+}
+
+void LLViewerFetchedTexture::scheduleCreateTexture()
+{
+    ref();
+    mNeedsCreateTexture = TRUE;
+    if (preCreateTexture())
+    {
+        mNeedsCreateTexture = TRUE;
+#if LL_WINDOWS //flip to 0 to revert to single-threaded OpenGL texture uploads
+        if (!LLImageGLThread::sInstance->post([this]()
+            {
+                //actually create the texture on a background thread
+                createTexture();
+                {
+                    LL_PROFILE_ZONE_NAMED("iglt - sync");
+                    if (gGLManager.mHasSync)
+                    {
+                        auto sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+                        glClientWaitSync(sync, 0, 0);
+                        glDeleteSync(sync);
+                    }
+                    else
+                    {
+                        glFinish();
+                    }
+                }
+                LLImageGLThread::sInstance->postCallback([this]()
+                    {
+                        //finalize on main thread
+                        postCreateTexture();
+                        unref();
+                    });
+            }))
+#endif
+        {
+            gTextureList.mCreateTextureList.insert(this);
+            unref();
+        }
+    }
 }
 
 // Call with 0,0 to turn this feature off.
@@ -1869,6 +1954,7 @@ void LLViewerFetchedTexture::setAdditionalDecodePriority(F32 priority)
 
 void LLViewerFetchedTexture::updateVirtualSize() 
 {	
+    LL_PROFILE_ZONE_SCOPED;
 	if(!mMaxVirtualSizeResetCounter)
 	{
 		addTextureStats(0.f, FALSE);//reset
@@ -1960,6 +2046,7 @@ bool LLViewerFetchedTexture::isActiveFetching()
 
 bool LLViewerFetchedTexture::updateFetch()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	static LLCachedControl<bool> textures_decode_disabled(gSavedSettings,"TextureDecodeDisabled", false);
 	static LLCachedControl<F32>  sCameraMotionThreshold(gSavedSettings,"TextureCameraMotionThreshold", 0.2);
 	static LLCachedControl<S32>  sCameraMotionBoost(gSavedSettings,"TextureCameraMotionBoost", 3);
@@ -2062,7 +2149,7 @@ bool LLViewerFetchedTexture::updateFetch()
 				}
 				else
 				{
-					mIsRawImageValid = TRUE;			
+					mIsRawImageValid = TRUE;
 					addToCreateTexture();
 				}
 
@@ -2890,6 +2977,7 @@ void LLViewerFetchedTexture::destroyRawImage()
 //virtual
 void LLViewerFetchedTexture::switchToCachedImage()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(mCachedRawImage.notNull())
 	{
 		mRawImage = mCachedRawImage;
@@ -2901,12 +2989,12 @@ void LLViewerFetchedTexture::switchToCachedImage()
 			mComponents = mRawImage->getComponents();
 			mGLTexturep->setComponents(mComponents);
 			gTextureList.dirtyImage(this);
-		}			
+		}
 
 		mIsRawImageValid = TRUE;
 		mRawDiscardLevel = mCachedRawDiscardLevel;
-		gTextureList.mCreateTextureList.insert(this);
-		mNeedsCreateTexture = TRUE;		
+
+        scheduleCreateTexture();
 	}
 }
 
@@ -3180,6 +3268,7 @@ bool LLViewerLODTexture::isUpdateFrozen()
 //virtual
 void LLViewerLODTexture::processTextureStats()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	updateVirtualSize();
 	
 	static LLCachedControl<bool> textures_fullres(gSavedSettings,"TextureLoadFullRes", false);
diff --git a/indra/newview/llviewertexture.h b/indra/newview/llviewertexture.h
index 69568cc825..a5a1fb2c16 100644
--- a/indra/newview/llviewertexture.h
+++ b/indra/newview/llviewertexture.h
@@ -321,9 +321,13 @@ public:
 
 	void addToCreateTexture();
 
-
-	 // ONLY call from LLViewerTextureList
+    //call to determine if createTexture is necessary
+    BOOL preCreateTexture(S32 usename = 0);
+	 // ONLY call from LLViewerTextureList or ImageGL background thread
 	BOOL createTexture(S32 usename = 0);
+    void postCreateTexture();
+    void scheduleCreateTexture();
+
 	void destroyTexture() ;
 
 	virtual void processTextureStats() ;
diff --git a/indra/newview/llviewertexturelist.cpp b/indra/newview/llviewertexturelist.cpp
index 12495078e9..db740b69e9 100644
--- a/indra/newview/llviewertexturelist.cpp
+++ b/indra/newview/llviewertexturelist.cpp
@@ -114,6 +114,7 @@ void LLViewerTextureList::init()
 
 void LLViewerTextureList::doPreloadImages()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	LL_DEBUGS("ViewerImages") << "Preloading images..." << LL_ENDL;
 	
 	llassert_always(mInitialized) ;
@@ -205,6 +206,7 @@ static std::string get_texture_list_name()
 
 void LLViewerTextureList::doPrefetchImages()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (LLAppViewer::instance()->getPurgeCache())
 	{
 		// cache was purged, no point
@@ -258,6 +260,7 @@ LLViewerTextureList::~LLViewerTextureList()
 
 void LLViewerTextureList::shutdown()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// clear out preloads
 	mImagePreloads.clear();
 
@@ -333,6 +336,7 @@ void LLViewerTextureList::shutdown()
 
 void LLViewerTextureList::dump()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	LL_INFOS() << "LLViewerTextureList::dump()" << LL_ENDL;
 	for (image_priority_list_t::iterator it = mImageList.begin(); it != mImageList.end(); ++it)
 	{
@@ -377,6 +381,7 @@ LLViewerFetchedTexture* LLViewerTextureList::getImageFromFile(const std::string&
 												   LLGLenum primary_format, 
 												   const LLUUID& force_id)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!mInitialized)
 	{
 		return NULL ;
@@ -404,6 +409,7 @@ LLViewerFetchedTexture* LLViewerTextureList::getImageFromUrl(const std::string&
 												   LLGLenum primary_format, 
 												   const LLUUID& force_id)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!mInitialized)
 	{
 		return NULL ;
@@ -492,6 +498,7 @@ LLViewerFetchedTexture* LLViewerTextureList::getImage(const LLUUID &image_id,
 												   LLGLenum primary_format,
 												   LLHost request_from_host)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!mInitialized)
 	{
 		return NULL ;
@@ -554,6 +561,7 @@ LLViewerFetchedTexture* LLViewerTextureList::createImage(const LLUUID &image_id,
 												   LLGLenum primary_format,
 												   LLHost request_from_host)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	static LLCachedControl<bool> fast_cache_fetching_enabled(gSavedSettings, "FastCacheFetchEnabled", true);
 
 	LLPointer<LLViewerFetchedTexture> imagep ;
@@ -609,6 +617,7 @@ LLViewerFetchedTexture* LLViewerTextureList::createImage(const LLUUID &image_id,
 
 void LLViewerTextureList::findTexturesByID(const LLUUID &image_id, std::vector<LLViewerFetchedTexture*> &output)
 {
+    LL_PROFILE_ZONE_SCOPED;
     LLTextureKey search_key(image_id, TEX_LIST_STANDARD);
     uuid_map_t::iterator iter = mUUIDMap.lower_bound(search_key);
     while (iter != mUUIDMap.end() && iter->first.textureId == image_id)
@@ -634,6 +643,7 @@ LLViewerFetchedTexture *LLViewerTextureList::findImage(const LLUUID &image_id, E
 
 void LLViewerTextureList::addImageToList(LLViewerFetchedTexture *image)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	assert_main_thread();
 	llassert_always(mInitialized) ;
 	llassert(image);
@@ -653,6 +663,7 @@ void LLViewerTextureList::addImageToList(LLViewerFetchedTexture *image)
 
 void LLViewerTextureList::removeImageFromList(LLViewerFetchedTexture *image)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	assert_main_thread();
 	llassert_always(mInitialized) ;
 	llassert(image);
@@ -701,6 +712,7 @@ void LLViewerTextureList::removeImageFromList(LLViewerFetchedTexture *image)
 
 void LLViewerTextureList::addImage(LLViewerFetchedTexture *new_image, ETexListType tex_type)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (!new_image)
 	{
 		return;
@@ -724,6 +736,7 @@ void LLViewerTextureList::addImage(LLViewerFetchedTexture *new_image, ETexListTy
 
 void LLViewerTextureList::deleteImage(LLViewerFetchedTexture *image)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if( image)
 	{
 		if (image->hasCallbacks())
@@ -845,6 +858,7 @@ void LLViewerTextureList::updateImages(F32 max_time)
 
 void LLViewerTextureList::clearFetchingRequests()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (LLAppViewer::getTextureFetch()->getNumRequests() == 0)
 	{
 		return;
@@ -862,6 +876,7 @@ void LLViewerTextureList::clearFetchingRequests()
 
 void LLViewerTextureList::updateImagesDecodePriorities()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// Update the decode priority for N images each frame
 	{
 		F32 lazy_flush_timeout = 30.f; // stop decoding
@@ -977,6 +992,7 @@ void LLViewerTextureList::updateImagesDecodePriorities()
 
 void LLViewerTextureList::setDebugFetching(LLViewerFetchedTexture* tex, S32 debug_level)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!tex->setDebugFetching(debug_level))
 	{
 		return;
@@ -1025,6 +1041,7 @@ void LLViewerTextureList::setDebugFetching(LLViewerFetchedTexture* tex, S32 debu
 
 F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (gGLManager.mIsDisabled) return 0.0f;
 	
 	//
@@ -1041,6 +1058,7 @@ F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time)
 		enditer = iter;
 		LLViewerFetchedTexture *imagep = *curiter;
 		imagep->createTexture();
+        imagep->postCreateTexture();
 		if (create_timer.getElapsedTimeF32() > max_time)
 		{
 			break;
@@ -1052,6 +1070,7 @@ F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time)
 
 F32 LLViewerTextureList::updateImagesLoadingFastCache(F32 max_time)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (gGLManager.mIsDisabled) return 0.0f;
 	if(mFastCacheList.empty())
 	{
@@ -1082,6 +1101,7 @@ F32 LLViewerTextureList::updateImagesLoadingFastCache(F32 max_time)
 
 void LLViewerTextureList::forceImmediateUpdate(LLViewerFetchedTexture* imagep)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!imagep)
 	{
 		return ;
@@ -1101,6 +1121,7 @@ void LLViewerTextureList::forceImmediateUpdate(LLViewerFetchedTexture* imagep)
 
 F32 LLViewerTextureList::updateImagesFetchTextures(F32 max_time)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	LLTimer image_op_timer;
 	
 	// Update fetch for N images each frame
@@ -1176,6 +1197,7 @@ F32 LLViewerTextureList::updateImagesFetchTextures(F32 max_time)
 
 void LLViewerTextureList::updateImagesUpdateStats()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (mForceResetTextureStats)
 	{
 		for (image_priority_list_t::iterator iter = mImageList.begin();
@@ -1190,6 +1212,7 @@ void LLViewerTextureList::updateImagesUpdateStats()
 
 void LLViewerTextureList::decodeAllImages(F32 max_time)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	LLTimer timer;
 
 	//loading from fast cache 
@@ -1259,6 +1282,7 @@ BOOL LLViewerTextureList::createUploadFile(const std::string& filename,
 										 const std::string& out_filename,
 										 const U8 codec)
 {	
+    LL_PROFILE_ZONE_SCOPED;
 	// Load the image
 	LLPointer<LLImageFormatted> image = LLImageFormatted::createFromType(codec);
 	if (image.isNull())
@@ -1312,6 +1336,7 @@ BOOL LLViewerTextureList::createUploadFile(const std::string& filename,
 // note: modifies the argument raw_image!!!!
 LLPointer<LLImageJ2C> LLViewerTextureList::convertToUploadFile(LLPointer<LLImageRaw> raw_image)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	raw_image->biasedScaleToPowerOfTwo(LLViewerFetchedTexture::MAX_IMAGE_SIZE_DEFAULT);
 	LLPointer<LLImageJ2C> compressedImage = new LLImageJ2C();
 	
@@ -1345,6 +1370,7 @@ LLPointer<LLImageJ2C> LLViewerTextureList::convertToUploadFile(LLPointer<LLImage
 // Returns min setting for TextureMemory (in MB)
 S32Megabytes LLViewerTextureList::getMinVideoRamSetting()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	U32Megabytes system_ram = gSysMemory.getPhysicalMemoryKB();
 	//min texture mem sets to 64M if total physical mem is more than 1.5GB
 	return (system_ram > U32Megabytes(1500)) ? S32Megabytes(64) : gMinVideoRam ;
@@ -1354,6 +1380,7 @@ S32Megabytes LLViewerTextureList::getMinVideoRamSetting()
 // Returns max setting for TextureMemory (in MB)
 S32Megabytes LLViewerTextureList::getMaxVideoRamSetting(bool get_recommended, float mem_multiplier)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	S32Megabytes max_texmem;
 	if (gGLManager.mVRAM != 0)
 	{
@@ -1407,6 +1434,7 @@ const S32Megabytes VIDEO_CARD_FRAMEBUFFER_MEM(12);
 const S32Megabytes MIN_MEM_FOR_NON_TEXTURE(512);
 void LLViewerTextureList::updateMaxResidentTexMem(S32Megabytes mem)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// Initialize the image pipeline VRAM settings
 	S32Megabytes cur_mem(gSavedSettings.getS32("TextureMemory"));
 	F32 mem_multiplier = gSavedSettings.getF32("RenderTextureMemoryMultiple");
@@ -1647,6 +1675,7 @@ void LLUIImageList::cleanUp()
 
 LLUIImagePtr LLUIImageList::getUIImageByID(const LLUUID& image_id, S32 priority)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// use id as image name
 	std::string image_name = image_id.asString();
 
@@ -1665,6 +1694,7 @@ LLUIImagePtr LLUIImageList::getUIImageByID(const LLUUID& image_id, S32 priority)
 
 LLUIImagePtr LLUIImageList::getUIImage(const std::string& image_name, S32 priority)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// look for existing image
 	uuid_ui_image_map_t::iterator found_it = mUIImages.find(image_name);
 	if (found_it != mUIImages.end())
@@ -1682,6 +1712,7 @@ LLUIImagePtr LLUIImageList::loadUIImageByName(const std::string& name, const std
 											  BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLViewerTexture::EBoostLevel boost_priority,
 											  LLUIImage::EScaleStyle scale_style)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (boost_priority == LLGLTexture::BOOST_NONE)
 	{
 		boost_priority = LLGLTexture::BOOST_UI;
@@ -1694,6 +1725,7 @@ LLUIImagePtr LLUIImageList::loadUIImageByID(const LLUUID& id,
 											BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLViewerTexture::EBoostLevel boost_priority,
 											LLUIImage::EScaleStyle scale_style)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (boost_priority == LLGLTexture::BOOST_NONE)
 	{
 		boost_priority = LLGLTexture::BOOST_UI;
@@ -1705,6 +1737,7 @@ LLUIImagePtr LLUIImageList::loadUIImageByID(const LLUUID& id,
 LLUIImagePtr LLUIImageList::loadUIImage(LLViewerFetchedTexture* imagep, const std::string& name, BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect,
 										LLUIImage::EScaleStyle scale_style)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if (!imagep) return NULL;
 
 	imagep->setAddressMode(LLTexUnit::TAM_CLAMP);
@@ -1742,6 +1775,7 @@ LLUIImagePtr LLUIImageList::loadUIImage(LLViewerFetchedTexture* imagep, const st
 
 LLUIImagePtr LLUIImageList::preloadUIImage(const std::string& name, const std::string& filename, BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLUIImage::EScaleStyle scale_style)
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// look for existing image
 	uuid_ui_image_map_t::iterator found_it = mUIImages.find(name);
 	if (found_it != mUIImages.end())
@@ -1756,6 +1790,7 @@ LLUIImagePtr LLUIImageList::preloadUIImage(const std::string& name, const std::s
 //static 
 void LLUIImageList::onUIImageLoaded( BOOL success, LLViewerFetchedTexture *src_vi, LLImageRaw* src, LLImageRaw* src_aux, S32 discard_level, BOOL final, void* user_data )
 {
+    LL_PROFILE_ZONE_SCOPED;
 	if(!success || !user_data) 
 	{
 		return;
@@ -1857,6 +1892,7 @@ struct UIImageDeclarations : public LLInitParam::Block<UIImageDeclarations>
 
 bool LLUIImageList::initFromFile()
 {
+    LL_PROFILE_ZONE_SCOPED;
 	// Look for textures.xml in all the right places. Pass
 	// constraint=LLDir::ALL_SKINS because we want to overlay textures.xml
 	// from all the skins directories.
diff --git a/indra/newview/llviewerwindow.cpp b/indra/newview/llviewerwindow.cpp
index c52a4e98d3..82ece85c1b 100644
--- a/indra/newview/llviewerwindow.cpp
+++ b/indra/newview/llviewerwindow.cpp
@@ -2011,7 +2011,7 @@ LLViewerWindow::LLViewerWindow(const Params& p)
 		
 	// Init the image list.  Must happen after GL is initialized and before the images that
 	// LLViewerWindow needs are requested.
-	LLImageGL::initClass(LLViewerTexture::MAX_GL_IMAGE_CATEGORY) ;
+	LLImageGL::initClass(mWindow, LLViewerTexture::MAX_GL_IMAGE_CATEGORY) ;
 	gTextureList.init();
 	LLViewerTextureManager::init() ;
 	gBumpImageList.init();