#1870 Tune up for better experience on integrated intel with low memory (#1872)

* More deterministic vsize calculation.   Add control for choosing downscale method.
* Quick hack to make GLTF preview work again

28 files changed, 404 insertions, 165 deletions

diff --git a/indra/llfilesystem/llfilesystem.cpp b/indra/llfilesystem/llfilesystem.cpp
index 235aae0be3..912f48e216 100644
--- a/indra/llfilesystem/llfilesystem.cpp
+++ b/indra/llfilesystem/llfilesystem.cpp
@@ -79,6 +79,7 @@ LLFileSystem::~LLFileSystem()
 // static
 bool LLFileSystem::getExists(const LLUUID& file_id, const LLAssetType::EType file_type)
 {
+    LL_PROFILE_ZONE_SCOPED;
     std::string id_str;
     file_id.toString(id_str);
     const std::string extra_info = "";
diff --git a/indra/llmessage/llassetstorage.cpp b/indra/llmessage/llassetstorage.cpp
index 70a7a34a70..2de59c1b6a 100644
--- a/indra/llmessage/llassetstorage.cpp
+++ b/indra/llmessage/llassetstorage.cpp
@@ -484,6 +484,8 @@ void LLAssetStorage::getAssetData(const LLUUID uuid,
                                   void *user_data,
                                   bool is_priority)
 {
+    LL_PROFILE_ZONE_SCOPED;
+
     LL_DEBUGS("AssetStorage") << "LLAssetStorage::getAssetData() - " << uuid << "," << LLAssetType::lookup(type) << LL_ENDL;
 
     LL_DEBUGS("AssetStorage") << "ASSET_TRACE requesting " << uuid << " type " << LLAssetType::lookup(type) << LL_ENDL;
@@ -529,6 +531,7 @@ void LLAssetStorage::getAssetData(const LLUUID uuid,
 
     if (size > 0)
     {
+        LL_PROFILE_ZONE_NAMED("gad - file in cache");
         // we've already got the file
         // theoretically, partial files w/o a pending request shouldn't happen
         // unless there's a weird error
@@ -548,7 +551,7 @@ void LLAssetStorage::getAssetData(const LLUUID uuid,
         }
 
         bool duplicate = false;
-
+        LL_PROFILE_ZONE_NAMED("gad - check pending downloads");
         // check to see if there's a pending download of this uuid already
         for (request_list_t::iterator iter = mPendingDownloads.begin();
              iter != mPendingDownloads.end(); ++iter )
diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index 909dad2e85..be1c0a532a 100644
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -102,6 +102,9 @@ public:
     bool mIsNVIDIA;
     bool mIsIntel;
 
+    // hints to the render pipe
+    U32 mDownScaleMethod = 0; // see settings.xml RenderDownScaleMethod
+
 #if LL_DARWIN
     // Needed to distinguish problem cards on older Macs that break with Materials
     bool mIsMobileGF;
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 956bcef352..34200ef5cb 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -57,6 +57,9 @@ const F32 MIN_TEXTURE_LIFETIME = 10.f;
 U32 wpo2(U32 i);
 
 
+U32 LLImageGL::sFrameCount = 0;
+
+
 // texture memory accounting (for macOS)
 static LLMutex sTexMemMutex;
 static std::unordered_map<U32, U64> sTextureAllocs;
@@ -1003,7 +1006,7 @@ bool should_stagger_image_set(bool compressed)
 #else
     // glTexSubImage2D doesn't work with compressed textures on select tested Nvidia GPUs on Windows 10 -Cosmic,2023-03-08
     // Setting media textures off-thread seems faster when not using sub_image_lines (Nvidia/Windows 10) -Cosmic,2023-03-31
-    return !compressed && on_main_thread();
+    return !compressed && on_main_thread() && !gGLManager.mIsIntel;
 #endif
 }
 
@@ -1186,12 +1189,36 @@ void LLImageGL::generateTextures(S32 numTextures, U32 *textures)
 }
 
 // static
+void LLImageGL::updateClass()
+{
+    sFrameCount++;
+}
+
+// static
 void LLImageGL::deleteTextures(S32 numTextures, const U32 *textures)
 {
+    // wait a few frames before actually deleting the textures to avoid
+    // synchronization issues with the GPU
+    static std::vector<U32> sFreeList[4];
+
     if (gGLManager.mInited)
     {
-        free_tex_images(numTextures, textures);
-        glDeleteTextures(numTextures, textures);
+        LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+        U32 idx = sFrameCount % 4;
+
+        for (S32 i = 0; i < numTextures; ++i)
+        {
+            sFreeList[idx].push_back(textures[i]);
+        }
+
+        idx = (sFrameCount + 3) % 4;
+
+        if (!sFreeList[idx].empty())
+        {
+            glDeleteTextures((GLsizei) sFreeList[idx].size(), sFreeList[idx].data());
+            free_tex_images((GLsizei) sFreeList[idx].size(), sFreeList[idx].data());
+            sFreeList[idx].resize(0);
+        }
     }
 }
 
@@ -2335,44 +2362,85 @@ bool LLImageGL::scaleDown(S32 desired_discard)
     S32 desired_width = getWidth(desired_discard);
     S32 desired_height = getHeight(desired_discard);
 
-    U64 size = getBytes(desired_discard);
-    llassert(size <= 2048*2048*4); // we shouldn't be using this method to downscale huge textures, but it'll work
-    gGL.getTexUnit(0)->bind(this);
+    if (gGLManager.mDownScaleMethod == 0)
+    { // use an FBO to downscale the texture
+        // allocate new texture
+        U32 temp_texname = 0;
+        generateTextures(1, &temp_texname);
+        gGL.getTexUnit(0)->bindManual(LLTexUnit::TT_TEXTURE, temp_texname, true);
+        {
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glTexImage2D");
+            glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, NULL);
+        }
 
+        // account for new texture getting created
+        alloc_tex_image(desired_width, desired_height, mFormatPrimary);
 
-    if (sScratchPBO == 0)
-    {
-        glGenBuffers(1, &sScratchPBO);
-        sScratchPBOSize = 0;
-    }
+        // Use render-to-texture to scale down the texture
+        {
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glFramebufferTexture2D");
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, mTarget, temp_texname, 0);
+        }
 
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, sScratchPBO);
+        glViewport(0, 0, desired_width, desired_height);
 
-    if (size > sScratchPBOSize)
-    {
-        glBufferData(GL_PIXEL_PACK_BUFFER, size, NULL, GL_STREAM_COPY);
-        sScratchPBOSize = size;
+        // draw a full screen triangle
+        gGL.getTexUnit(0)->bind(this);
+        glDrawArrays(GL_TRIANGLES, 0, 3);
+        gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+
+        // delete old texture and assign new texture name
+        deleteTextures(1, &mTexName);
+        mTexName = temp_texname;
+
+        if (mHasMipMaps)
+        { // generate mipmaps if needed
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
+            gGL.getTexUnit(0)->bind(this);
+            glGenerateMipmap(mTarget);
+            gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+        }
     }
+    else
+    { // use a PBO to downscale the texture
+        U64 size = getBytes(desired_discard);
+        llassert(size <= 2048 * 2048 * 4); // we shouldn't be using this method to downscale huge textures, but it'll work
+        gGL.getTexUnit(0)->bind(this);
+
+        if (sScratchPBO == 0)
+        {
+            glGenBuffers(1, &sScratchPBO);
+            sScratchPBOSize = 0;
+        }
 
-    glGetTexImage(mTarget, mip, mFormatPrimary, mFormatType, nullptr);
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, sScratchPBO);
 
-    free_tex_image(mTexName);
+        if (size > sScratchPBOSize)
+        {
+            glBufferData(GL_PIXEL_PACK_BUFFER, size, NULL, GL_STREAM_COPY);
+            sScratchPBOSize = size;
+        }
 
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+        glGetTexImage(mTarget, mip, mFormatPrimary, mFormatType, nullptr);
 
-    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, sScratchPBO);
-    glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, nullptr);
-    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+        free_tex_image(mTexName);
 
-    alloc_tex_image(desired_width, desired_height, mFormatPrimary);
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 
-    if (mHasMipMaps)
-    {
-        LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
-        glGenerateMipmap(mTarget);
-    }
+        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, sScratchPBO);
+        glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, nullptr);
+        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
-    gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+        alloc_tex_image(desired_width, desired_height, mFormatPrimary);
+
+        if (mHasMipMaps)
+        {
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
+            glGenerateMipmap(mTarget);
+        }
+
+        gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+    }
 
     mCurrentDiscardLevel = desired_discard;
 
diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h
index 3892e9c014..5073701c30 100644
--- a/indra/llrender/llimagegl.h
+++ b/indra/llrender/llimagegl.h
@@ -62,6 +62,9 @@ class LLImageGL : public LLRefCount
     friend class LLTexUnit;
 public:
 
+    // call once per frame
+    static void updateClass();
+
     // Get an estimate of how many bytes have been allocated in vram for textures.
     // Does not include mipmaps.
     // NOTE: multiplying this number by two gives a good estimate for total
@@ -276,7 +279,7 @@ protected:
 public:
     static std::unordered_set<LLImageGL*> sImageList;
     static S32 sCount;
-
+    static U32 sFrameCount;
     static F32 sLastFrameTime;
 
     // Global memory statistics
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index be9f3895e7..e6b4118beb 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -297,7 +297,11 @@ public:
         ALTERNATE_DIFFUSE_MAP = 1,
         NORMAL_MAP            = 1,
         SPECULAR_MAP          = 2,
-        NUM_TEXTURE_CHANNELS  = 3,
+        BASECOLOR_MAP         = 3,
+        METALLIC_ROUGHNESS_MAP = 4,
+        GLTF_NORMAL_MAP           = 5,
+        EMISSIVE_MAP          = 6,
+        NUM_TEXTURE_CHANNELS  = 7,
     };
 
     enum eVolumeTexIndex : U8
diff --git a/indra/llrender/llrendertarget.cpp b/indra/llrender/llrendertarget.cpp
index efdde77a32..60159a0497 100644
--- a/indra/llrender/llrendertarget.cpp
+++ b/indra/llrender/llrendertarget.cpp
@@ -426,14 +426,17 @@ void LLRenderTarget::bindTarget()
                             GL_COLOR_ATTACHMENT1,
                             GL_COLOR_ATTACHMENT2,
                             GL_COLOR_ATTACHMENT3};
-    glDrawBuffers(static_cast<GLsizei>(mTex.size()), drawbuffers);
 
     if (mTex.empty())
     { //no color buffer to draw to
         glDrawBuffer(GL_NONE);
         glReadBuffer(GL_NONE);
     }
-
+    else
+    {
+        glDrawBuffers(static_cast<GLsizei>(mTex.size()), drawbuffers);
+        glReadBuffer(GL_COLOR_ATTACHMENT0);
+    }
     check_framebuffer_status();
 
     glViewport(0, 0, mResX, mResY);
@@ -519,7 +522,8 @@ void LLRenderTarget::flush()
     llassert(sCurFBO == mFBO);
     llassert(sBoundTarget == this);
 
-    if (mGenerateMipMaps == LLTexUnit::TMG_AUTO) {
+    if (mGenerateMipMaps == LLTexUnit::TMG_AUTO)
+    {
         LL_PROFILE_GPU_ZONE("rt generate mipmaps");
         bindTexture(0, 0, LLTexUnit::TFO_TRILINEAR);
         glGenerateMipmap(GL_TEXTURE_2D);
@@ -540,6 +544,8 @@ void LLRenderTarget::flush()
         glViewport(gGLViewport[0], gGLViewport[1], gGLViewport[2], gGLViewport[3]);
         sCurResX = gGLViewport[2];
         sCurResY = gGLViewport[3];
+        glReadBuffer(GL_BACK);
+        glDrawBuffer(GL_BACK);
     }
 }
 
diff --git a/indra/llrender/llrendertarget.h b/indra/llrender/llrendertarget.h
index 340276a752..a1adf93fa1 100644
--- a/indra/llrender/llrendertarget.h
+++ b/indra/llrender/llrendertarget.h
@@ -172,6 +172,8 @@ public:
     // *HACK
     void swapFBORefs(LLRenderTarget& other);
 
+    static LLRenderTarget* sBoundTarget;
+
 protected:
     U32 mResX;
     U32 mResY;
@@ -186,8 +188,6 @@ protected:
     U32 mMipLevels;
 
     LLTexUnit::eTextureType mUsage;
-
-    static LLRenderTarget* sBoundTarget;
 };
 
 #endif
diff --git a/indra/newview/app_settings/settings.xml b/indra/newview/app_settings/settings.xml
index 09909aa9c4..7f645792a6 100644
--- a/indra/newview/app_settings/settings.xml
+++ b/indra/newview/app_settings/settings.xml
@@ -7732,7 +7732,7 @@
     <key>Comment</key>
     <string>Minimum of available physical memory in MB before textures get scaled down</string>
     <key>Persist</key>
-    <integer>1</integer>
+    <integer>0</integer>
     <key>Type</key>
     <string>U32</string>
     <key>Value</key>
@@ -7771,6 +7771,17 @@
     <key>Value</key>
     <integer>2048</integer>
   </map>
+    <key>RenderDownScaleMethod</key>
+    <map>
+        <key>Comment</key>
+        <string>Method to use to downscale images.  0 - FBO, 1 - PBO</string>
+        <key>Persist</key>
+        <integer>1</integer>
+        <key>Type</key>
+        <string>U32</string>
+        <key>Value</key>
+        <integer>1</integer>
+    </map>
     <key>RenderDebugTextureBind</key>
     <map>
       <key>Comment</key>
diff --git a/indra/newview/featuretable.txt b/indra/newview/featuretable.txt
index 03849a0326..af1587bd07 100644
--- a/indra/newview/featuretable.txt
+++ b/indra/newview/featuretable.txt
@@ -1,4 +1,4 @@
-version 60
+version 61
 // The version number above should be incremented IF AND ONLY IF some
 // change has been made that is sufficiently important to justify
 // resetting the graphics preferences of all users to the recommended
@@ -76,6 +76,7 @@ RenderGLMultiThreadedMedia         1   1
 RenderReflectionProbeResolution 1 128
 RenderScreenSpaceReflections 1  1
 RenderMirrors				1	1
+RenderDownScaleMethod       1   1
 
 
 //
@@ -366,6 +367,7 @@ RenderAnisotropic			1	0
 RenderFSAASamples			1	0
 RenderGLContextCoreProfile  1   0
 RenderGLMultiThreadedMedia  1   0
+RenderDownScaleMethod       1   0
 
 list AMD
 RenderGLMultiThreadedTextures       1   1
diff --git a/indra/newview/gltf/accessor.h b/indra/newview/gltf/accessor.h
index ec68c5f624..85ea0f2967 100644
--- a/indra/newview/gltf/accessor.h
+++ b/indra/newview/gltf/accessor.h
@@ -36,8 +36,6 @@ namespace LL
 {
     namespace GLTF
     {
-        constexpr S32 INVALID_INDEX = -1;
-
         class Buffer
         {
         public:
diff --git a/indra/newview/gltf/asset.cpp b/indra/newview/gltf/asset.cpp
index a454e68a92..e07befef3b 100644
--- a/indra/newview/gltf/asset.cpp
+++ b/indra/newview/gltf/asset.cpp
@@ -476,6 +476,7 @@ void Asset::update()
                 { // HACK - force texture to be loaded full rez
                     // TODO: calculate actual vsize
                     image.mTexture->addTextureStats(2048.f * 2048.f);
+                    image.mTexture->setBoostLevel(LLViewerTexture::BOOST_HIGH);
                 }
             }
         }
diff --git a/indra/newview/gltf/common.h b/indra/newview/gltf/common.h
index b9698d4017..742daff715 100644
--- a/indra/newview/gltf/common.h
+++ b/indra/newview/gltf/common.h
@@ -43,6 +43,8 @@ namespace LL
 {
     namespace GLTF
     {
+        constexpr S32 INVALID_INDEX = -1;
+
         using Value = boost::json::value;
 
         using mat4 = glm::mat4;
diff --git a/indra/newview/gltfscenemanager.cpp b/indra/newview/gltfscenemanager.cpp
index b161ec8492..58179f9dc3 100644
--- a/indra/newview/gltfscenemanager.cpp
+++ b/indra/newview/gltfscenemanager.cpp
@@ -345,6 +345,12 @@ void GLTFSceneManager::addGLTFObject(LLViewerObject* obj, LLUUID gltf_id)
     llassert(obj->getVolume()->getParams().getSculptID() == gltf_id);
     llassert(obj->getVolume()->getParams().getSculptType() == LL_SCULPT_TYPE_GLTF);
 
+    if (obj->mGLTFAsset)
+    { // object already has a GLTF asset, don't reload it
+        llassert(std::find(mObjects.begin(), mObjects.end(), obj) != mObjects.end());
+        return;
+    }
+
     obj->ref();
     gAssetStorage->getAssetData(gltf_id, LLAssetType::AT_GLTF, onGLTFLoadComplete, obj);
 }
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp
index 51099a480d..ac689f7319 100644
--- a/indra/newview/llappviewer.cpp
+++ b/indra/newview/llappviewer.cpp
@@ -2210,10 +2210,10 @@ bool LLAppViewer::initThreads()
         cores = llmin(cores, (S32) max_cores);
     }
 
-    // The only configurable thread count right now is ImageDecode
-    // The viewer typically starts around 8 threads not including image decode,
-    // so try to leave at least one core free
-    S32 image_decode_count = llclamp(cores - 9, 1, 8);
+    // always use at least 2 threads for image decoding to prevent
+    // a single texture blocking all other textures from decoding
+    S32 image_decode_count = llclamp(cores - 6, 2, 16);
+
     threadCounts["ImageDecode"] = image_decode_count;
     gSavedSettings.setLLSD("ThreadPoolSizes", threadCounts);
 
@@ -4609,6 +4609,10 @@ void LLAppViewer::idle()
 
     LLGLTFMaterialList::flushUpdates();
 
+    static LLCachedControl<U32> downscale_method(gSavedSettings, "RenderDownScaleMethod");
+    gGLManager.mDownScaleMethod = downscale_method;
+    LLImageGL::updateClass();
+
     // Service the WorkQueue we use for replies from worker threads.
     // Use function statics for the timeslice setting so we only have to fetch
     // and convert MainWorkTime once.
diff --git a/indra/newview/lldrawpoolalpha.cpp b/indra/newview/lldrawpoolalpha.cpp
index 34da5b29d4..2d702ecf32 100644
--- a/indra/newview/lldrawpoolalpha.cpp
+++ b/indra/newview/lldrawpoolalpha.cpp
@@ -81,10 +81,6 @@ LLDrawPoolAlpha::~LLDrawPoolAlpha()
 void LLDrawPoolAlpha::prerender()
 {
     mShaderLevel = LLViewerShaderMgr::instance()->getShaderLevel(LLViewerShaderMgr::SHADER_OBJECT);
-
-    // TODO: is this even necessay?  These are probably set to never discard
-    LLViewerFetchedTexture::sFlatNormalImagep->addTextureStats(1024.f*1024.f);
-    LLViewerFetchedTexture::sWhiteImagep->addTextureStats(1024.f * 1024.f);
 }
 
 S32 LLDrawPoolAlpha::getNumPostDeferredPasses()
diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp
index 0e8e64af69..777ea611b0 100644
--- a/indra/newview/llface.cpp
+++ b/indra/newview/llface.cpp
@@ -169,6 +169,8 @@ void LLFace::init(LLDrawable* drawablep, LLViewerObject* objp)
     mImportanceToCamera = 0.f ;
     mBoundingSphereRadius = 0.0f ;
 
+    mTexExtents[0].set(0, 0);
+    mTexExtents[1].set(1, 1);
     mHasMedia = false ;
     mIsMediaAllowed = true;
 }
@@ -2047,10 +2049,12 @@ void LLFace::resetVirtualSize()
 F32 LLFace::getTextureVirtualSize()
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+
     F32 radius;
     F32 cos_angle_to_view_dir;
     bool in_frustum = calcPixelArea(cos_angle_to_view_dir, radius);
 
+
     if (mPixelArea < F_ALMOST_ZERO || !in_frustum)
     {
         setVirtualSize(0.f) ;
diff --git a/indra/newview/lltextureview.cpp b/indra/newview/lltextureview.cpp
index 9d7319d5b8..9e86548c2d 100644
--- a/indra/newview/lltextureview.cpp
+++ b/indra/newview/lltextureview.cpp
@@ -530,6 +530,9 @@ void LLGLTexMemBar::draw()
    F64 raw_image_bytes_MB = raw_image_bytes / (1024.0 * 1024.0);
    F64 saved_raw_image_bytes_MB = saved_raw_image_bytes / (1024.0 * 1024.0);
    F64 aux_raw_image_bytes_MB = aux_raw_image_bytes / (1024.0 * 1024.0);
+   F64 texture_bytes_alloc = LLImageGL::getTextureBytesAllocated() / 1024.0 / 1024.0 * 1.3333f; // add 33% for mipmaps
+   F64 vertex_bytes_alloc = LLVertexBuffer::getBytesAllocated() / 1024.0 / 1024.0;
+   F64 render_bytes_alloc = LLRenderTarget::sBytesAllocated / 1024.0 / 1024.0;
 
     //----------------------------------------------------------------------------
     LLGLSUIDefault gls_ui;
@@ -559,7 +562,7 @@ void LLGLTexMemBar::draw()
 
     // draw a background above first line.... no idea where the rest of the background comes from for the below text
     gGL.color4f(0.f, 0.f, 0.f, 0.25f);
-    gl_rect_2d(-10, getRect().getHeight() + line_height + 1, getRect().getWidth()+2, getRect().getHeight()+2);
+    gl_rect_2d(-10, getRect().getHeight() + line_height*2 + 1, getRect().getWidth()+2, getRect().getHeight()+2);
 
     text = llformat("Est. Free: %d MB Sys Free: %d MB FBO: %d MB Bias: %.2f Cache: %.1f/%.1f MB",
                     (S32)LLViewerTexture::sFreeVRAMMegabytes,
@@ -568,12 +571,20 @@ void LLGLTexMemBar::draw()
                     discard_bias,
                     cache_usage,
                     cache_max_usage);
-    LLFontGL::getFontMonospace()->renderUTF8(text, 0, 0, v_offset + line_height*7,
+    LLFontGL::getFontMonospace()->renderUTF8(text, 0, 0, v_offset + line_height*8,
                                              text_color, LLFontGL::LEFT, LLFontGL::TOP);
 
     text = llformat("Images: %d   Raw: %d (%.2f MB)  Saved: %d (%.2f MB) Aux: %d (%.2f MB)", image_count, raw_image_count, raw_image_bytes_MB,
         saved_raw_image_count, saved_raw_image_bytes_MB,
         aux_raw_image_count, aux_raw_image_bytes_MB);
+    LLFontGL::getFontMonospace()->renderUTF8(text, 0, 0, v_offset + line_height * 7,
+        text_color, LLFontGL::LEFT, LLFontGL::TOP);
+
+    text = llformat("Textures: %.2f MB  Vertex: %.2f MB  Render: %.2f MB  Total: %.2f MB",
+                    texture_bytes_alloc,
+                    vertex_bytes_alloc,
+                    render_bytes_alloc,
+        texture_bytes_alloc+vertex_bytes_alloc+render_bytes_alloc);
     LLFontGL::getFontMonospace()->renderUTF8(text, 0, 0, v_offset + line_height * 6,
         text_color, LLFontGL::LEFT, LLFontGL::TOP);
 
diff --git a/indra/newview/llviewerassetstorage.cpp b/indra/newview/llviewerassetstorage.cpp
index 26707f5d9a..5ab9f76e47 100644
--- a/indra/newview/llviewerassetstorage.cpp
+++ b/indra/newview/llviewerassetstorage.cpp
@@ -376,6 +376,7 @@ void LLViewerAssetStorage::queueRequestHttp(
     bool duplicate,
     bool is_priority)
 {
+    LL_PROFILE_ZONE_SCOPED;
     LL_DEBUGS("ViewerAsset") << "Request asset via HTTP " << uuid << " type " << LLAssetType::lookup(atype) << LL_ENDL;
 
     bool with_http = true;
diff --git a/indra/newview/llviewertexture.cpp b/indra/newview/llviewertexture.cpp
index b2ed86707c..99c917ca6f 100644
--- a/indra/newview/llviewertexture.cpp
+++ b/indra/newview/llviewertexture.cpp
@@ -89,7 +89,7 @@ S32 LLViewerTexture::sRawCount = 0;
 S32 LLViewerTexture::sAuxCount = 0;
 LLFrameTimer LLViewerTexture::sEvaluationTimer;
 F32 LLViewerTexture::sDesiredDiscardBias = 0.f;
-F32 LLViewerTexture::sDesiredDiscardScale = 1.1f;
+
 S32 LLViewerTexture::sMaxSculptRez = 128; //max sculpt image size
 const S32 MAX_CACHED_RAW_IMAGE_AREA = 64 * 64;
 const S32 MAX_CACHED_RAW_SCULPT_IMAGE_AREA = LLViewerTexture::sMaxSculptRez * LLViewerTexture::sMaxSculptRez;
@@ -513,15 +513,18 @@ void LLViewerTexture::updateClass()
     F32 target = llmax(budget - 512.f, MIN_VRAM_BUDGET);
     sFreeVRAMMegabytes = target - used;
 
-    F32 over_pct = llmax((used-target) / target, 0.f);
+    F32 over_pct = (used - target) / target;
+
+    bool is_low = over_pct > 0.f;
 
     if (isSystemMemoryLow())
     {
+        is_low = true;
         // System RAM is low -> ramp up discard bias over time to free memory
         if (sEvaluationTimer.getElapsedTimeF32() > MEMORY_CHECK_WAIT_TIME)
         {
             static LLCachedControl<F32> low_mem_min_discard_increment(gSavedSettings, "RenderLowMemMinDiscardIncrement", .1f);
-            sDesiredDiscardBias += llmax(low_mem_min_discard_increment, over_pct);
+            sDesiredDiscardBias += (F32) low_mem_min_discard_increment * (F32) gFrameIntervalSeconds;
             sEvaluationTimer.reset();
         }
     }
@@ -529,12 +532,28 @@ void LLViewerTexture::updateClass()
     {
         sDesiredDiscardBias = llmax(sDesiredDiscardBias, 1.f + over_pct);
 
-        if (sDesiredDiscardBias > 1.f)
+        if (sDesiredDiscardBias > 1.f && over_pct < 0.f)
         {
             sDesiredDiscardBias -= gFrameIntervalSeconds * 0.01;
         }
     }
 
+    static bool was_low = false;
+    if (is_low && !was_low)
+    {
+        LL_WARNS() << "Low system memory detected, emergency downrezzing off screen textures" << LL_ENDL;
+        sDesiredDiscardBias = llmax(sDesiredDiscardBias, 1.5f);
+
+        for (auto image : gTextureList)
+        {
+            gTextureList.updateImageDecodePriority(image);
+        }
+    }
+
+    was_low = is_low;
+
+    sDesiredDiscardBias = llclamp(sDesiredDiscardBias, 1.f, 3.f);
+
     LLViewerTexture::sFreezeImageUpdates = false;
 }
 
@@ -615,16 +634,15 @@ void LLViewerTexture::init(bool firstinit)
     mParcelMedia = NULL;
 
     memset(&mNumVolumes, 0, sizeof(U32)* LLRender::NUM_VOLUME_TEXTURE_CHANNELS);
-    mFaceList[LLRender::DIFFUSE_MAP].clear();
-    mFaceList[LLRender::NORMAL_MAP].clear();
-    mFaceList[LLRender::SPECULAR_MAP].clear();
-    mNumFaces[LLRender::DIFFUSE_MAP] =
-    mNumFaces[LLRender::NORMAL_MAP] =
-    mNumFaces[LLRender::SPECULAR_MAP] = 0;
-
     mVolumeList[LLRender::LIGHT_TEX].clear();
     mVolumeList[LLRender::SCULPT_TEX].clear();
 
+    for (U32 i = 0; i < LLRender::NUM_TEXTURE_CHANNELS; i++)
+    {
+        mNumFaces[i] = 0;
+        mFaceList[i].clear();
+    }
+
     mMainQueue  = LL::WorkQueue::getInstance("mainloop");
     mImageQueue = LL::WorkQueue::getInstance("LLImageGL");
 }
@@ -1608,7 +1626,11 @@ void LLViewerFetchedTexture::scheduleCreateTexture()
             }
             else
             {
-                gTextureList.mCreateTextureList.insert(this);
+                if (!mCreatePending)
+                {
+                    mCreatePending = true;
+                    gTextureList.mCreateTextureList.push(this);
+                }
             }
         }
     }
@@ -1632,13 +1654,12 @@ void LLViewerFetchedTexture::setKnownDrawSize(S32 width, S32 height)
 
 void LLViewerFetchedTexture::setDebugText(const std::string& text)
 {
-    for (U32 ch = 0; ch < LLRender::NUM_TEXTURE_CHANNELS; ++ch)
+    for (U32 i = 0; i < LLRender::NUM_TEXTURE_CHANNELS; ++i)
     {
-        llassert(mNumFaces[ch] <= mFaceList[ch].size());
-
-        for (U32 i = 0; i < mNumFaces[ch]; i++)
+        for (S32 fi = 0; fi < getNumFaces(i); ++fi)
         {
-            LLFace* facep = mFaceList[ch][i];
+            LLFace* facep = (*(getFaceList(i)))[fi];
+
             if (facep)
             {
                 LLDrawable* drawable = facep->getDrawable();
@@ -1651,10 +1672,15 @@ void LLViewerFetchedTexture::setDebugText(const std::string& text)
     }
 }
 
+extern bool gCubeSnapshot;
+
 //virtual
 void LLViewerFetchedTexture::processTextureStats()
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+    llassert(!gCubeSnapshot);  // should only be called when the main camera is active
+    llassert(!LLPipeline::sShadowRender);
+
     if(mFullyLoaded)
     {
         if(mDesiredDiscardLevel > mMinDesiredDiscardLevel)//need to load more
@@ -2859,6 +2885,8 @@ void LLViewerLODTexture::processTextureStats()
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
     updateVirtualSize();
 
+    bool did_downscale = false;
+
     static LLCachedControl<bool> textures_fullres(gSavedSettings,"TextureLoadFullRes", false);
 
     { // restrict texture resolution to download based on RenderMaxTextureResolution
@@ -2916,10 +2944,7 @@ void LLViewerLODTexture::processTextureStats()
             mDiscardVirtualSize = mMaxVirtualSize;
             mCalculatedDiscardLevel = discard_level;
         }
-        if (mBoostLevel < LLGLTexture::BOOST_SCULPTED)
-        {
-            discard_level *= sDesiredDiscardScale; // scale (default 1.1f)
-        }
+
         discard_level = floorf(discard_level);
 
         F32 min_discard = 0.f;
@@ -2945,10 +2970,9 @@ void LLViewerLODTexture::processTextureStats()
         //
 
         S32 current_discard = getDiscardLevel();
-        if (mBoostLevel < LLGLTexture::BOOST_AVATAR_BAKED &&
-            current_discard >= 0)
+        if (mBoostLevel < LLGLTexture::BOOST_AVATAR_BAKED)
         {
-            if (current_discard < (mDesiredDiscardLevel-1) && !mForceToSaveRawImage)
+            if (current_discard < mDesiredDiscardLevel && !mForceToSaveRawImage)
             { // should scale down
                 scaleDown();
             }
@@ -2968,9 +2992,6 @@ void LLViewerLODTexture::processTextureStats()
         mDesiredDiscardLevel = llmin(mDesiredDiscardLevel, (S8)mDesiredSavedRawDiscardLevel);
     }
 
-    // decay max virtual size over time
-    mMaxVirtualSize *= 0.8f;
-
     // selection manager will immediately reset BOOST_SELECTED but never unsets it
     // unset it immediately after we consume it
     if (getBoostLevel() == BOOST_SELECTED)
@@ -2979,14 +3000,22 @@ void LLViewerLODTexture::processTextureStats()
     }
 }
 
+extern LLGLSLShader gCopyProgram;
+
 bool LLViewerLODTexture::scaleDown()
 {
-    if (mGLTexturep.isNull())
+    if (mGLTexturep.isNull() || !mGLTexturep->getHasGLTexture())
     {
         return false;
     }
 
-    return mGLTexturep->scaleDown(mDesiredDiscardLevel);
+    if (!mDownScalePending)
+    {
+        mDownScalePending = true;
+        gTextureList.mDownScaleQueue.push(this);
+    }
+
+    return true;
 }
 
 //----------------------------------------------------------------------------------------------
diff --git a/indra/newview/llviewertexture.h b/indra/newview/llviewertexture.h
index b1e0494a30..5b2e3aaab7 100644
--- a/indra/newview/llviewertexture.h
+++ b/indra/newview/llviewertexture.h
@@ -37,6 +37,7 @@
 #include "llmetricperformancetester.h"
 #include "httpcommon.h"
 #include "workqueue.h"
+#include "gltf/common.h"
 
 #include <map>
 #include <list>
@@ -102,7 +103,6 @@ public:
         DYNAMIC_TEXTURE,
         FETCHED_TEXTURE,
         LOD_TEXTURE,
-        ATLAS_TEXTURE,
         INVALID_TEXTURE_TYPE
     };
 
@@ -173,6 +173,15 @@ public:
     LLViewerMediaTexture* getParcelMedia() const { return mParcelMedia;}
 
     /*virtual*/ void updateBindStatsForTester() ;
+
+    struct MaterialEntry
+    {
+        S32 mIndex = LL::GLTF::INVALID_INDEX;
+        std::shared_ptr<LL::GLTF::Asset> mAsset;
+    };
+    typedef std::vector<MaterialEntry> material_list_t;
+    material_list_t   mMaterialList;  // reverse pointer pointing to LL::GLTF::Materials using this image as texture
+
 protected:
     void cleanup() ;
     void init(bool firstinit) ;
@@ -214,7 +223,6 @@ public:
     static S32 sAuxCount;
     static LLFrameTimer sEvaluationTimer;
     static F32 sDesiredDiscardBias;
-    static F32 sDesiredDiscardScale;
     static S32 sMaxSculptRez ;
     static U32 sMinLargeImageSize ;
     static U32 sMaxSmallImageSize ;
@@ -414,6 +422,9 @@ public:
 
     /*virtual*/bool  isActiveFetching() override; //is actively in fetching by the fetching pipeline.
 
+    bool mCreatePending = false;    // if true, this is in gTextureList.mCreateTextureList
+    mutable bool mDownScalePending = false; // if true, this is in gTextureList.mDownScaleQueue
+
 protected:
     S32 getCurrentDiscardLevelForFetching() ;
     void forceToRefetchTexture(S32 desired_discard = 0, F32 kept_time = 60.f);
@@ -424,11 +435,6 @@ private:
 
     void saveRawImage() ;
 
-    //for atlas
-    void resetFaceAtlas() ;
-    void invalidateAtlas(bool rebuild_geom) ;
-    bool insertToAtlas() ;
-
 private:
     bool  mFullyLoaded;
     bool  mInDebug;
@@ -539,9 +545,10 @@ public:
     /*virtual*/ void processTextureStats();
     bool isUpdateFrozen() ;
 
+    bool scaleDown();
+
 private:
     void init(bool firstinit) ;
-    bool scaleDown() ;
 
 private:
     F32 mDiscardVirtualSize;        // Virtual size used to calculate desired discard
diff --git a/indra/newview/llviewertexturelist.cpp b/indra/newview/llviewertexturelist.cpp
index 2f685474e1..c436566297 100644
--- a/indra/newview/llviewertexturelist.cpp
+++ b/indra/newview/llviewertexturelist.cpp
@@ -70,6 +70,8 @@ S32 LLViewerTextureList::sNumImages = 0;
 
 LLViewerTextureList gTextureList;
 
+extern LLGLSLShader gCopyProgram;
+
 ETexListType get_element_type(S32 priority)
 {
     return (priority == LLViewerFetchedTexture::BOOST_ICON || priority == LLViewerFetchedTexture::BOOST_THUMBNAIL) ? TEX_LIST_SCALE : TEX_LIST_STANDARD;
@@ -352,8 +354,11 @@ void LLViewerTextureList::shutdown()
     mCallbackList.clear();
 
     // Flush all of the references
-    mLoadingStreamList.clear();
-    mCreateTextureList.clear();
+    while (!mCreateTextureList.empty())
+    {
+        mCreateTextureList.front()->mCreatePending = false;
+        mCreateTextureList.pop();
+    }
     mFastCacheList.clear();
 
     mUUIDMap.clear();
@@ -897,14 +902,6 @@ void LLViewerTextureList::clearFetchingRequests()
     }
 }
 
-static void touch_texture(LLViewerFetchedTexture* tex, F32 vsize)
-{
-    if (tex)
-    {
-        tex->addTextureStats(vsize);
-    }
-}
-
 extern bool gCubeSnapshot;
 
 void LLViewerTextureList::updateImageDecodePriority(LLViewerFetchedTexture* imagep)
@@ -921,58 +918,67 @@ void LLViewerTextureList::updateImageDecodePriority(LLViewerFetchedTexture* imag
     static LLCachedControl<F32> texture_scale_min(gSavedSettings, "TextureScaleMinAreaFactor", 0.04f);
     static LLCachedControl<F32> texture_scale_max(gSavedSettings, "TextureScaleMaxAreaFactor", 25.f);
 
+    if (imagep->getType() == LLViewerTexture::LOD_TEXTURE && imagep->getBoostLevel() == LLViewerTexture::BOOST_NONE)
+    { // reset max virtual size for unboosted LOD_TEXTURES
+        // this is an alternative to decaying mMaxVirtualSize over time
+        // that keeps textures from continously downrezzing and uprezzing in the background
+        imagep->mMaxVirtualSize = 0.f;
+    }
+
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE
+    for (U32 i = 0; i < LLRender::NUM_TEXTURE_CHANNELS; ++i)
     {
-        for (U32 i = 0; i < LLRender::NUM_TEXTURE_CHANNELS; ++i)
+        for (S32 fi = 0; fi < imagep->getNumFaces(i); ++fi)
         {
-            for (S32 fi = 0; fi < imagep->getNumFaces(i); ++fi)
+            LLFace* face = (*(imagep->getFaceList(i)))[fi];
+
+            if (face && face->getViewerObject())
             {
-                LLFace* face = (*(imagep->getFaceList(i)))[fi];
+                F32 radius;
+                F32 cos_angle_to_view_dir;
+                bool in_frustum = face->calcPixelArea(cos_angle_to_view_dir, radius);
+                static LLCachedControl<F32> bias_unimportant_threshold(gSavedSettings, "TextureBiasUnimportantFactor", 0.25f);
+                F32 vsize = face->getPixelArea();
+
+                // Scale desired texture resolution higher or lower depending on texture scale
+                //
+                // Minimum usage examples: a 1024x1024 texture with aplhabet, runing string
+                // shows one letter at a time
+                //
+                // Maximum usage examples: huge chunk of terrain repeats texture
+                const LLTextureEntry* te = face->getTextureEntry();
+                F32 min_scale = te ? llmin(fabsf(te->getScaleS()), fabsf(te->getScaleT())) : 1.f;
+                min_scale = llclamp(min_scale * min_scale, texture_scale_min(), texture_scale_max());
+                vsize /= min_scale;
+
+                // if bias is > 2, apply to on-screen textures as well
+                bool apply_bias = LLViewerTexture::sDesiredDiscardBias > 2.f;
+
+                // apply bias to off screen objects or objects that are small on screen all the time
+                if (!in_frustum || !face->getDrawable()->isVisible() || face->getImportanceToCamera() < bias_unimportant_threshold)
+                { // further reduce by discard bias when off screen or occluded
+                    apply_bias = true;
+                }
 
-                if (face && face->getViewerObject() && face->getTextureEntry())
+                if (apply_bias)
                 {
-                    F32 radius;
-                    F32 cos_angle_to_view_dir;
-                    BOOL in_frustum = face->calcPixelArea(cos_angle_to_view_dir, radius);
-                    static LLCachedControl<F32> bias_unimportant_threshold(gSavedSettings, "TextureBiasUnimportantFactor", 0.25f);
-                    F32 vsize = face->getPixelArea();
-
-                    // Scale desired texture resolution higher or lower depending on texture scale
-                    //
-                    // Minimum usage examples: a 1024x1024 texture with aplhabet, runing string
-                    // shows one letter at a time
-                    //
-                    // Maximum usage examples: huge chunk of terrain repeats texture
-                    const LLTextureEntry* te = face->getTextureEntry();
-                    F32 min_scale = te ? llmin(fabsf(te->getScaleS()), fabsf(te->getScaleT())) : 1.f;
-                    min_scale = llclamp(min_scale*min_scale, texture_scale_min(), texture_scale_max());
-
-                    vsize /= min_scale;
-                    if (!in_frustum || !face->getDrawable()->isVisible() || face->getImportanceToCamera() < bias_unimportant_threshold)
-                    { // further reduce by discard bias when off screen or occluded
-                        vsize /= LLViewerTexture::sDesiredDiscardBias;
-                    }
-                    // if a GLTF material is present, ignore that face
-                    // as far as this texture stats go, but update the GLTF material
-                    // stats
-                    LLFetchedGLTFMaterial* mat = te ? (LLFetchedGLTFMaterial*)te->getGLTFRenderMaterial() : nullptr;
-                    llassert(mat == nullptr || dynamic_cast<LLFetchedGLTFMaterial*>(te->getGLTFRenderMaterial()) != nullptr);
-                    if (mat)
-                    {
-                        touch_texture(mat->mBaseColorTexture, vsize);
-                        touch_texture(mat->mNormalTexture, vsize);
-                        touch_texture(mat->mMetallicRoughnessTexture, vsize);
-                        touch_texture(mat->mEmissiveTexture, vsize);
-                    }
-                    else
-                    {
-                        imagep->addTextureStats(vsize);
-                    }
+                    F32 bias = powf(4, LLViewerTexture::sDesiredDiscardBias - 1.f);
+                    bias = llround(bias);
+                    vsize /= bias;
                 }
+
+                imagep->addTextureStats(vsize);
             }
         }
     }
 
+    // make sure to addTextureStats for any spotlights that are using this texture
+    for (S32 vi = 0; vi < imagep->getNumVolumes(LLRender::LIGHT_TEX); ++vi)
+    {
+        LLVOVolume* volume = (*imagep->getVolumeList(LLRender::LIGHT_TEX))[vi];
+        volume->updateSpotLightPriority();
+    }
+
     //imagep->setDebugText(llformat("%.3f - %d", sqrtf(imagep->getMaxVirtualSize()), imagep->getBoostLevel()));
 
     F32 lazy_flush_timeout = 30.f; // stop decoding
@@ -1066,22 +1072,65 @@ F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time)
     //
 
     LLTimer create_timer;
-    image_list_t::iterator enditer = mCreateTextureList.begin();
-    for (image_list_t::iterator iter = mCreateTextureList.begin();
-         iter != mCreateTextureList.end();)
+
+    if (!mDownScaleQueue.empty() && gPipeline.mDownResMap.isComplete())
     {
-        image_list_t::iterator curiter = iter++;
-        enditer = iter;
-        LLViewerFetchedTexture *imagep = *curiter;
+        // just in case we downres textures, bind downresmap and copy program
+        gPipeline.mDownResMap.bindTarget();
+        gCopyProgram.bind();
+        gPipeline.mScreenTriangleVB->setBuffer();
+
+        // give time to downscaling first -- if mDownScaleQueue is not empty, we're running out of memory and need
+        // to free up memory by discarding off screen textures quickly
+
+        // do at least 5 and make sure we don't get too far behind even if it violates
+        // the time limit.  If we don't downscale quickly the viewer will hit swap and may
+        // freeze.
+        S32 min_count = (S32)mCreateTextureList.size() / 20 + 5;
+
+        while (!mDownScaleQueue.empty())
+        {
+            LLViewerFetchedTexture* image = mDownScaleQueue.front();
+            llassert(image->mDownScalePending);
+
+            LLImageGL* img = image->getGLTexture();
+            if (img && img->getHasGLTexture())
+            {
+                img->scaleDown(image->getDesiredDiscardLevel());
+            }
+
+            image->mDownScalePending = false;
+            mDownScaleQueue.pop();
+
+            if (create_timer.getElapsedTimeF32() > max_time && --min_count <= 0)
+            {
+                break;
+            }
+        }
+
+        gCopyProgram.unbind();
+        gPipeline.mDownResMap.flush();
+    }
+
+    // do at least 5 and make sure we don't get too far behind even if it violates
+    // the time limit.  Textures pending creation have a copy of their texture data
+    // in system memory, so we don't want to let them pile up.
+    S32 min_count = (S32) mCreateTextureList.size() / 20 + 5;
+
+    while (!mCreateTextureList.empty())
+    {
+        LLViewerFetchedTexture *imagep =  mCreateTextureList.front();
+        llassert(imagep->mCreatePending);
         imagep->createTexture();
         imagep->postCreateTexture();
+        imagep->mCreatePending = false;
+        mCreateTextureList.pop();
 
-        if (create_timer.getElapsedTimeF32() > max_time)
+        if (create_timer.getElapsedTimeF32() > max_time && --min_count <= 0)
         {
             break;
         }
     }
-    mCreateTextureList.erase(mCreateTextureList.begin(), enditer);
     return create_timer.getElapsedTimeF32();
 }
 
@@ -1124,7 +1173,10 @@ void LLViewerTextureList::forceImmediateUpdate(LLViewerFetchedTexture* imagep)
         removeImageFromList(imagep);
     }
 
-    imagep->processTextureStats();
+    if (!gCubeSnapshot)
+    { // never call processTextureStats in a cube snapshot
+        imagep->processTextureStats();
+    }
     imagep->sMaxVirtualSize = LLViewerFetchedTexture::sMaxVirtualSize;
     addImageToList(imagep);
 
@@ -1134,6 +1186,7 @@ void LLViewerTextureList::forceImmediateUpdate(LLViewerFetchedTexture* imagep)
 F32 LLViewerTextureList::updateImagesFetchTextures(F32 max_time)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+
     typedef std::vector<LLPointer<LLViewerFetchedTexture> > entries_list_t;
     entries_list_t entries;
 
diff --git a/indra/newview/llviewertexturelist.h b/indra/newview/llviewertexturelist.h
index 2779ad9f91..813fda1647 100644
--- a/indra/newview/llviewertexturelist.h
+++ b/indra/newview/llviewertexturelist.h
@@ -144,12 +144,13 @@ public:
     void clearFetchingRequests();
     void setDebugFetching(LLViewerFetchedTexture* tex, S32 debug_level);
 
-private:
     // do some book keeping on the specified texture
     // - updates decode priority
     // - updates desired discard level
     // - cleans up textures that haven't been referenced in awhile
     void updateImageDecodePriority(LLViewerFetchedTexture* imagep);
+
+private:
     F32  updateImagesCreateTextures(F32 max_time);
     F32  updateImagesFetchTextures(F32 max_time);
     void updateImagesUpdateStats();
@@ -211,8 +212,14 @@ private:
 
 public:
     typedef std::unordered_set<LLPointer<LLViewerFetchedTexture> > image_list_t;
-    image_list_t mLoadingStreamList;
-    image_list_t mCreateTextureList;
+    typedef std::queue<LLPointer<LLViewerFetchedTexture> > image_queue_t;
+
+    // images that have been loaded but are waiting to be uploaded to GL
+    image_queue_t mCreateTextureList;
+
+    // images that must be downscaled quickly so we don't run out of memory
+    image_queue_t mDownScaleQueue;
+
     image_list_t mCallbackList;
     image_list_t mFastCacheList;
 
diff --git a/indra/newview/llvograss.cpp b/indra/newview/llvograss.cpp
index e04bdbf1d9..6903af2619 100644
--- a/indra/newview/llvograss.cpp
+++ b/indra/newview/llvograss.cpp
@@ -451,6 +451,9 @@ void LLVOGrass::plantBlades()
         face->setVertexBuffer(NULL);
         face->setTEOffset(0);
         face->mCenterLocal = mPosition + mRegionp->getOriginAgent();
+        const LLVector4a* ext = mDrawable->getSpatialExtents();
+        face->mExtents[0] = ext[0];
+        face->mExtents[1] = ext[1];
     }
 
     mDepth = (face->mCenterLocal - LLViewerCamera::getInstance()->getOrigin())*LLViewerCamera::getInstance()->getAtAxis();
diff --git a/indra/newview/llvotree.cpp b/indra/newview/llvotree.cpp
index a981c60ef2..d982592ee7 100644
--- a/indra/newview/llvotree.cpp
+++ b/indra/newview/llvotree.cpp
@@ -461,9 +461,7 @@ void LLVOTree::updateTextures()
         {
             setDebugText(llformat("%4.0f", (F32) sqrt(mPixelArea)));
         }
-        mTreeImagep->addTextureStats(mPixelArea);
     }
-
 }
 
 
@@ -479,7 +477,7 @@ LLDrawable* LLVOTree::createDrawable(LLPipeline *pipeline)
     // Just a placeholder for an actual object...
     LLFace *facep = mDrawable->addFace(poolp, mTreeImagep);
     facep->setSize(1, 3);
-
+    facep->setTexture(LLRender::DIFFUSE_MAP, mTreeImagep);
     updateRadius();
 
     return mDrawable;
@@ -1169,6 +1167,10 @@ void LLVOTree::updateSpatialExtents(LLVector4a& newMin, LLVector4a& newMax)
     LLVector4a pos;
     pos.load3(center.mV);
     mDrawable->setPositionGroup(pos);
+
+    LLFace* facep = mDrawable->getFace(0);
+    facep->mExtents[0] = newMin;
+    facep->mExtents[1] = newMax;
 }
 
 bool LLVOTree::lineSegmentIntersect(const LLVector4a& start, const LLVector4a& end, S32 face, bool pick_transparent, bool pick_rigged, bool pick_unselectable, S32 *face_hitp,
diff --git a/indra/newview/llvovolume.cpp b/indra/newview/llvovolume.cpp
index 7ca3fbd1ad..45713071bb 100644
--- a/indra/newview/llvovolume.cpp
+++ b/indra/newview/llvovolume.cpp
@@ -5745,18 +5745,23 @@ void LLVolumeGeometryManager::rebuildGeom(LLSpatialGroup* group)
                 {
                     continue;
                 }
-#if 0
-#if LL_RELEASE_WITH_DEBUG_INFO
-                const LLUUID pbr_id( "49c88210-7238-2a6b-70ac-92d4f35963cf" );
-                const LLUUID obj_id( vobj->getID() );
-                bool is_pbr = (obj_id == pbr_id);
-#else
-                bool is_pbr = false;
-#endif
-#else
-                LLGLTFMaterial *gltf_mat = facep->getTextureEntry()->getGLTFRenderMaterial();
+
+                LLFetchedGLTFMaterial *gltf_mat = (LLFetchedGLTFMaterial*) facep->getTextureEntry()->getGLTFRenderMaterial();
                 bool is_pbr = gltf_mat != nullptr;
-#endif
+
+                if (is_pbr)
+                {
+                    // tell texture streaming system to ignore blinn-phong textures
+                    facep->setTexture(LLRender::DIFFUSE_MAP, nullptr);
+                    facep->setTexture(LLRender::NORMAL_MAP, nullptr);
+                    facep->setTexture(LLRender::SPECULAR_MAP, nullptr);
+
+                    // let texture streaming system know about PBR textures
+                    facep->setTexture(LLRender::BASECOLOR_MAP, gltf_mat->mBaseColorTexture);
+                    facep->setTexture(LLRender::GLTF_NORMAL_MAP, gltf_mat->mNormalTexture);
+                    facep->setTexture(LLRender::METALLIC_ROUGHNESS_MAP, gltf_mat->mMetallicRoughnessTexture);
+                    facep->setTexture(LLRender::EMISSIVE_MAP, gltf_mat->mEmissiveTexture);
+                }
 
                 //ALWAYS null out vertex buffer on rebuild -- if the face lands in a render
                 // batch, it will recover its vertex buffer reference from the spatial group
diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp
index b9b1264b16..f83774c39e 100644
--- a/indra/newview/pipeline.cpp
+++ b/indra/newview/pipeline.cpp
@@ -865,6 +865,10 @@ bool LLPipeline::allocateScreenBuffer(U32 resX, U32 resY, U32 samples)
     const U32 post_color_fmt = post_hdr ? GL_RGBA16F : GL_RGBA;
     mPostMap.allocate(resX, resY, post_color_fmt);
 
+    // used to scale down textures
+    // See LLViwerTextureList::updateImagesCreateTextures and LLImageGL::scaleDown
+    mDownResMap.allocate(4, 4, GL_RGBA);
+
     //HACK make screenbuffer allocations start failing after 30 seconds
     if (gSavedSettings.getBOOL("SimulateFBOFailure"))
     {
@@ -1111,6 +1115,8 @@ void LLPipeline::releaseGLBuffers()
 
     mPostMap.release();
 
+    mDownResMap.release();
+
     for (U32 i = 0; i < 3; i++)
     {
         mGlow[i].release();
diff --git a/indra/newview/pipeline.h b/indra/newview/pipeline.h
index 1d2033b4f0..6c79d20149 100644
--- a/indra/newview/pipeline.h
+++ b/indra/newview/pipeline.h
@@ -725,6 +725,9 @@ public:
     // tonemapped and gamma corrected render ready for post
     LLRenderTarget          mPostMap;
 
+    // downres scratch space for GPU downscaling of textures
+    LLRenderTarget          mDownResMap;
+
     LLCullResult            mSky;
     LLCullResult            mReflectedObjects;
     LLCullResult            mRefractedObjects;