7 files changed, 489 insertions, 163 deletions

diff --git a/indra/newview/gltf/animation.cpp b/indra/newview/gltf/animation.cpp
index 8b85eba3e5..3dff67d746 100644
--- a/indra/newview/gltf/animation.cpp
+++ b/indra/newview/gltf/animation.cpp
@@ -70,6 +70,14 @@ bool Animation::prep(Asset& asset)
         }
     }
 
+    for (auto& channel : mScaleChannels)
+    {
+        if (!channel.prep(asset, mSamplers[channel.mSampler]))
+        {
+            return false;
+        }
+    }
+
     return true;
 }
 
@@ -82,18 +90,37 @@ void Animation::update(Asset& asset, F32 dt)
 
 void Animation::apply(Asset& asset, float time)
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
+
     // convert time to animation loop time
     time = fmod(time, mMaxTime - mMinTime) + mMinTime;
 
     // apply each channel
-    for (auto& channel : mRotationChannels)
     {
-        channel.apply(asset, mSamplers[channel.mSampler], time);
+        LL_PROFILE_ZONE_NAMED_CATEGORY_GLTF("gltfanim - rotation");
+
+        for (auto& channel : mRotationChannels)
+        {
+            channel.apply(asset, mSamplers[channel.mSampler], time);
+        }
+    }
+
+    {
+        LL_PROFILE_ZONE_NAMED_CATEGORY_GLTF("gltfanim - translation");
+
+        for (auto& channel : mTranslationChannels)
+        {
+            channel.apply(asset, mSamplers[channel.mSampler], time);
+        }
     }
 
-    for (auto& channel : mTranslationChannels)
     {
-        channel.apply(asset, mSamplers[channel.mSampler], time);
+        LL_PROFILE_ZONE_NAMED_CATEGORY_GLTF("gltfanim - scale");
+
+        for (auto& channel : mScaleChannels)
+        {
+            channel.apply(asset, mSamplers[channel.mSampler], time);
+        }
     }
 };
 
@@ -178,7 +205,8 @@ const Animation::Channel& Animation::Channel::operator=(const Value& src)
 
 void Animation::Sampler::getFrameInfo(Asset& asset, F32 time, U32& frameIndex, F32& t)
 {
-    LL_PROFILE_ZONE_SCOPED;
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
+    llassert(mFrameTimes.size() > 1); // if there is only one frame, there is no need to interpolate
 
     if (time < mMinTime)
     {
@@ -187,32 +215,33 @@ void Animation::Sampler::getFrameInfo(Asset& asset, F32 time, U32& frameIndex, F
         return;
     }
 
-    if (mFrameTimes.size() > 1)
+    frameIndex = U32(mFrameTimes.size()) - 2;
+    t = 1.f;
+
+    if (time > mMaxTime)
     {
-        llassert(mFrameTimes.size() <= size_t(U32_MAX));
-        frameIndex = U32(mFrameTimes.size()) - 2;
-        t = 1.f;
+        return;
+    }
 
-        if (time > mMaxTime)
-        {
-            return;
-        }
+    if (time < mLastFrameTime)
+    {
+        mLastFrameIndex = 0;
+    }
 
-        for (U32 i = 0; i < (U32)mFrameTimes.size() - 1; i++)
+    mLastFrameTime = time;
+
+    U32 idx = mLastFrameIndex;
+
+    for (U32 i = idx; i < (U32)mFrameTimes.size() - 1; i++)
+    {
+        if (time >= mFrameTimes[i] && time < mFrameTimes[i + 1])
         {
-            if (time >= mFrameTimes[i] && time < mFrameTimes[i + 1])
-            {
-                frameIndex = i;
-                t = (time - mFrameTimes[i]) / (mFrameTimes[i + 1] - mFrameTimes[i]);
-                return;
-            }
+            frameIndex = i;
+            t = (time - mFrameTimes[i]) / (mFrameTimes[i + 1] - mFrameTimes[i]);
+            mLastFrameIndex = frameIndex;
+            return;
         }
     }
-    else
-    {
-        frameIndex = 0;
-        t = 0.0f;
-    }
 }
 
 bool Animation::RotationChannel::prep(Asset& asset, Animation::Sampler& sampler)
@@ -231,14 +260,14 @@ void Animation::RotationChannel::apply(Asset& asset, Sampler& sampler, F32 time)
 
     Node& node = asset.mNodes[mTarget.mNode];
 
-    sampler.getFrameInfo(asset, time, frameIndex, t);
-
-    if (sampler.mFrameTimes.size() == 1)
+    if (sampler.mFrameTimes.size() < 2)
     {
         node.setRotation(mRotations[0]);
     }
     else
     {
+        sampler.getFrameInfo(asset, time, frameIndex, t);
+
         // interpolate
         quat qf = glm::slerp(mRotations[frameIndex], mRotations[frameIndex + 1], t);
 
@@ -264,14 +293,14 @@ void Animation::TranslationChannel::apply(Asset& asset, Sampler& sampler, F32 ti
 
     Node& node = asset.mNodes[mTarget.mNode];
 
-    sampler.getFrameInfo(asset, time, frameIndex, t);
-
-    if (sampler.mFrameTimes.size() == 1)
+    if (sampler.mFrameTimes.size() < 2)
     {
         node.setTranslation(mTranslations[0]);
     }
     else
     {
+        sampler.getFrameInfo(asset, time, frameIndex, t);
+
         // interpolate
         const vec3& v0 = mTranslations[frameIndex];
         const vec3& v1 = mTranslations[frameIndex + 1];
@@ -298,14 +327,14 @@ void Animation::ScaleChannel::apply(Asset& asset, Sampler& sampler, F32 time)
 
     Node& node = asset.mNodes[mTarget.mNode];
 
-    sampler.getFrameInfo(asset, time, frameIndex, t);
-
-    if (sampler.mFrameTimes.size() == 1)
+    if (sampler.mFrameTimes.size() < 2)
     {
         node.setScale(mScales[0]);
     }
     else
     {
+        sampler.getFrameInfo(asset, time, frameIndex, t);
+
         // interpolate
         const vec3& v0 = mScales[frameIndex];
         const vec3& v1 = mScales[frameIndex + 1];
@@ -373,6 +402,7 @@ Skin::~Skin()
 void Skin::uploadMatrixPalette(Asset& asset)
 {
     // prepare matrix palette
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
 
     U32 max_joints = LLSkinningUtil::getMaxGLTFJointCount();
 
diff --git a/indra/newview/gltf/animation.h b/indra/newview/gltf/animation.h
index d5426fd4ce..ab8839470a 100644
--- a/indra/newview/gltf/animation.h
+++ b/indra/newview/gltf/animation.h
@@ -49,6 +49,9 @@ namespace LL
                 S32 mOutput = INVALID_INDEX;
                 std::string mInterpolation;
 
+                F32 mLastFrameTime = 0.f;
+                U32 mLastFrameIndex = 0;
+
                 bool prep(Asset& asset);
 
                 void serialize(boost::json::object& dst) const;
diff --git a/indra/newview/gltf/asset.cpp b/indra/newview/gltf/asset.cpp
index 21be69aae2..a454e68a92 100644
--- a/indra/newview/gltf/asset.cpp
+++ b/indra/newview/gltf/asset.cpp
@@ -35,6 +35,7 @@
 #include "buffer_util.h"
 #include <boost/url.hpp>
 #include "llimagejpeg.h"
+#include "../llskinningutil.h"
 
 using namespace LL::GLTF;
 using namespace boost::json;
@@ -86,7 +87,6 @@ namespace LL
     }
 }
 
-
 void Scene::updateTransforms(Asset& asset)
 {
     mat4 identity = glm::identity<mat4>();
@@ -98,26 +98,6 @@ void Scene::updateTransforms(Asset& asset)
     }
 }
 
-void Scene::updateRenderTransforms(Asset& asset, const mat4& modelview)
-{
-    for (auto& nodeIndex : mNodes)
-    {
-        Node& node = asset.mNodes[nodeIndex];
-        node.updateRenderTransforms(asset, modelview);
-    }
-}
-
-void Node::updateRenderTransforms(Asset& asset, const mat4& modelview)
-{
-    mRenderMatrix = modelview * mMatrix;
-
-    for (auto& childIndex : mChildren)
-    {
-        Node& child = asset.mNodes[childIndex];
-        child.updateRenderTransforms(asset, mRenderMatrix);
-    }
-}
-
 void Node::updateTransforms(Asset& asset, const mat4& parentMatrix)
 {
     makeMatrixValid();
@@ -137,19 +117,119 @@ void Node::updateTransforms(Asset& asset, const mat4& parentMatrix)
 
 void Asset::updateTransforms()
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
     for (auto& scene : mScenes)
     {
         scene.updateTransforms(*this);
     }
+
+    uploadTransforms();
 }
 
-void Asset::updateRenderTransforms(const mat4& modelview)
+void Asset::uploadTransforms()
 {
-    // use mAssetMatrix to update render transforms from node list
-    for (auto& node : mNodes)
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
+    // prepare matrix palette
+    U32 max_nodes = LLSkinningUtil::getMaxGLTFJointCount();
+
+    size_t node_count = llmin<size_t>(max_nodes, mNodes.size());
+
+    std::vector<mat4> t_mp;
+
+    t_mp.resize(node_count);
+
+    for (U32 i = 0; i < node_count; ++i)
     {
-        node.mRenderMatrix = modelview * node.mAssetMatrix;
+        Node& node = mNodes[i];
+        // build matrix palette in asset space
+        t_mp[i] = node.mAssetMatrix;
     }
+
+    std::vector<F32> glmp;
+
+    glmp.resize(node_count * 12);
+
+    F32* mp = glmp.data();
+
+    for (U32 i = 0; i < node_count; ++i)
+    {
+        F32* m = glm::value_ptr(t_mp[i]);
+
+        U32 idx = i * 12;
+
+        mp[idx + 0] = m[0];
+        mp[idx + 1] = m[1];
+        mp[idx + 2] = m[2];
+        mp[idx + 3] = m[12];
+
+        mp[idx + 4] = m[4];
+        mp[idx + 5] = m[5];
+        mp[idx + 6] = m[6];
+        mp[idx + 7] = m[13];
+
+        mp[idx + 8] = m[8];
+        mp[idx + 9] = m[9];
+        mp[idx + 10] = m[10];
+        mp[idx + 11] = m[14];
+    }
+
+    if (mNodesUBO == 0)
+    {
+        glGenBuffers(1, &mNodesUBO);
+    }
+
+    glBindBuffer(GL_UNIFORM_BUFFER, mNodesUBO);
+    glBufferData(GL_UNIFORM_BUFFER, glmp.size() * sizeof(F32), glmp.data(), GL_STREAM_DRAW);
+    glBindBuffer(GL_UNIFORM_BUFFER, 0);
+}
+
+void Asset::uploadMaterials()
+{
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
+    // see pbrmetallicroughnessV.glsl for the layout of the material UBO
+    std::vector<vec4> md;
+
+    U32 material_size = sizeof(vec4) * 12;
+    U32 max_materials = gGLManager.mMaxUniformBlockSize / material_size;
+
+    U32 mat_count = (U32)mMaterials.size();
+    mat_count = llmin(mat_count, max_materials);
+
+    md.resize(mat_count * 12);
+
+    for (U32 i = 0; i < mat_count*12; i += 12)
+    {
+        Material& material = mMaterials[i/12];
+
+        // add texture transforms and UV indices
+        material.mPbrMetallicRoughness.mBaseColorTexture.mTextureTransform.getPacked(&md[i+0]);
+        md[i + 1].g = (F32)material.mPbrMetallicRoughness.mBaseColorTexture.getTexCoord();
+        material.mNormalTexture.mTextureTransform.getPacked(&md[i + 2]);
+        md[i + 3].g = (F32)material.mNormalTexture.getTexCoord();
+        material.mPbrMetallicRoughness.mMetallicRoughnessTexture.mTextureTransform.getPacked(&md[i+4]);
+        md[i + 5].g = (F32)material.mPbrMetallicRoughness.mMetallicRoughnessTexture.getTexCoord();
+        material.mEmissiveTexture.mTextureTransform.getPacked(&md[i + 6]);
+        md[i + 7].g = (F32)material.mEmissiveTexture.getTexCoord();
+        material.mOcclusionTexture.mTextureTransform.getPacked(&md[i + 8]);
+        md[i + 9].g = (F32)material.mOcclusionTexture.getTexCoord();
+
+        // add material properties
+        F32 min_alpha = material.mAlphaMode == Material::AlphaMode::MASK ? material.mAlphaCutoff : -1.0f;
+        md[i + 10] = vec4(material.mEmissiveFactor, 1.f);
+        md[i + 11] = vec4(0.f,
+            material.mPbrMetallicRoughness.mRoughnessFactor,
+            material.mPbrMetallicRoughness.mMetallicFactor,
+            min_alpha);
+    }
+
+    if (mMaterialsUBO == 0)
+    {
+        glGenBuffers(1, &mMaterialsUBO);
+    }
+
+    glBindBuffer(GL_UNIFORM_BUFFER, mMaterialsUBO);
+    glBufferData(GL_UNIFORM_BUFFER, md.size() * sizeof(vec4), md.data(), GL_STREAM_DRAW);
+    glBindBuffer(GL_UNIFORM_BUFFER, 0);
 }
 
 S32 Asset::lineSegmentIntersect(const LLVector4a& start, const LLVector4a& end,
@@ -363,6 +443,7 @@ const Image& Image::operator=(const Value& src)
 
 void Asset::update()
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
     F32 dt = gFrameTimeSeconds - mLastUpdateTime;
 
     if (dt > 0.f)
@@ -383,11 +464,27 @@ void Asset::update()
         {
             skin.uploadMatrixPalette(*this);
         }
+
+        uploadMaterials();
+
+        {
+            LL_PROFILE_ZONE_NAMED_CATEGORY_GLTF("gltf - addTextureStats");
+
+            for (auto& image : mImages)
+            {
+                if (image.mTexture.notNull())
+                { // HACK - force texture to be loaded full rez
+                    // TODO: calculate actual vsize
+                    image.mTexture->addTextureStats(2048.f * 2048.f);
+                }
+            }
+        }
     }
 }
 
 bool Asset::prep()
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
     // check required extensions and fail if not supported
     bool unsupported = false;
     for (auto& extension : mExtensionsRequired)
@@ -445,6 +542,127 @@ bool Asset::prep()
         }
     }
 
+    // prepare vertex buffers
+
+    // material count is number of materials + 1 for default material
+    U32 mat_count = (U32) mMaterials.size() + 1;
+
+    if (LLGLSLShader::sCurBoundShaderPtr == nullptr)
+    { // make sure a shader is bound to satisfy mVertexBuffer->setBuffer
+        gDebugProgram.bind();
+    }
+
+    for (S32 double_sided = 0; double_sided < 2; ++double_sided)
+    {
+        RenderData& rd = mRenderData[double_sided];
+        for (U32 i = 0; i < LLGLSLShader::NUM_GLTF_VARIANTS; ++i)
+        {
+            rd.mBatches[i].resize(mat_count);
+        }
+
+        // for each material
+        for (S32 mat_id = -1; mat_id < (S32)mMaterials.size(); ++mat_id)
+        {
+            // for each shader variant
+            U32 vertex_count[LLGLSLShader::NUM_GLTF_VARIANTS] = { 0 };
+            U32 index_count[LLGLSLShader::NUM_GLTF_VARIANTS] = { 0 };
+
+            S32 ds_mat = mat_id == -1 ? 0 : mMaterials[mat_id].mDoubleSided;
+            if (ds_mat != double_sided)
+            {
+                continue;
+            }
+
+            for (U32 variant = 0; variant < LLGLSLShader::NUM_GLTF_VARIANTS; ++variant)
+            {
+                U32 attribute_mask = 0;
+                // for each mesh
+                for (auto& mesh : mMeshes)
+                {
+                    // for each primitive
+                    for (auto& primitive : mesh.mPrimitives)
+                    {
+                        if (primitive.mMaterial == mat_id && primitive.mShaderVariant == variant)
+                        {
+                            // accumulate vertex and index counts
+                            primitive.mVertexOffset = vertex_count[variant];
+                            primitive.mIndexOffset = index_count[variant];
+
+                            vertex_count[variant] += primitive.getVertexCount();
+                            index_count[variant] += primitive.getIndexCount();
+
+                            // all primitives of a given variant and material should all have the same attribute mask
+                            llassert(attribute_mask == 0 || primitive.mAttributeMask == attribute_mask);
+                            attribute_mask |= primitive.mAttributeMask;
+                        }
+                    }
+                }
+
+                // allocate vertex buffer and pack it
+                if (vertex_count[variant] > 0)
+                {
+                    U32 mat_idx = mat_id + 1;
+                    LLVertexBuffer* vb = new LLVertexBuffer(attribute_mask);
+
+                    rd.mBatches[variant][mat_idx].mVertexBuffer = vb;
+                    vb->allocateBuffer(vertex_count[variant],
+                        index_count[variant] * 2); // hack double index count... TODO: find a better way to indicate 32-bit indices will be used
+                    vb->setBuffer();
+
+                    for (auto& mesh : mMeshes)
+                    {
+                        for (auto& primitive : mesh.mPrimitives)
+                        {
+                            if (primitive.mMaterial == mat_id && primitive.mShaderVariant == variant)
+                            {
+                                primitive.upload(vb);
+                            }
+                        }
+                    }
+
+                    vb->unmapBuffer();
+
+                    vb->unbind();
+                }
+            }
+        }
+    }
+
+    // sanity check that all primitives have a vertex buffer
+    for (auto& mesh : mMeshes)
+    {
+        for (auto& primitive : mesh.mPrimitives)
+        {
+            llassert(primitive.mVertexBuffer.notNull());
+        }
+    }
+
+    // build render batches
+    for (S32 node_id = 0; node_id < mNodes.size(); ++node_id)
+    {
+        Node& node = mNodes[node_id];
+
+        if (node.mMesh != INVALID_INDEX)
+        {
+            auto& mesh = mMeshes[node.mMesh];
+
+            S32 mat_idx = mesh.mPrimitives[0].mMaterial + 1;
+
+            S32 double_sided = mat_idx == 0 ? 0 : mMaterials[mat_idx - 1].mDoubleSided;
+
+            for (S32 j = 0; j < mesh.mPrimitives.size(); ++j)
+            {
+                auto& primitive = mesh.mPrimitives[j];
+
+                S32 variant = primitive.mShaderVariant;
+
+                RenderData& rd = mRenderData[double_sided];
+                RenderBatch& rb = rd.mBatches[variant][mat_idx];
+
+                rb.mPrimitives.push_back({ j, node_id });
+            }
+        }
+    }
     return true;
 }
 
@@ -455,6 +673,7 @@ Asset::Asset(const Value& src)
 
 bool Asset::load(std::string_view filename)
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_GLTF;
     mFilename = filename;
     std::string ext = gDirUtilp->getExtension(mFilename);
 
@@ -903,14 +1122,14 @@ bool Image::save(Asset& asset, const std::string& folder)
     return true;
 }
 
-void Material::TextureInfo::serialize(object& dst) const
+void TextureInfo::serialize(object& dst) const
 {
     write(mIndex, "index", dst, INVALID_INDEX);
     write(mTexCoord, "texCoord", dst, 0);
     write_extensions(dst, &mTextureTransform, "KHR_texture_transform");
 }
 
-S32 Material::TextureInfo::getTexCoord() const
+S32 TextureInfo::getTexCoord() const
 {
     if (mTextureTransform.mPresent && mTextureTransform.mTexCoord != INVALID_INDEX)
     {
@@ -928,7 +1147,7 @@ bool Material::isMultiUV() const
         mEmissiveTexture.getTexCoord() != 0;
 }
 
-const Material::TextureInfo& Material::TextureInfo::operator=(const Value& src)
+const TextureInfo& TextureInfo::operator=(const Value& src)
 {
     if (src.is_object())
     {
@@ -940,23 +1159,23 @@ const Material::TextureInfo& Material::TextureInfo::operator=(const Value& src)
     return *this;
 }
 
-bool Material::TextureInfo::operator==(const Material::TextureInfo& rhs) const
+bool TextureInfo::operator==(const TextureInfo& rhs) const
 {
     return mIndex == rhs.mIndex && mTexCoord == rhs.mTexCoord;
 }
 
-bool Material::TextureInfo::operator!=(const Material::TextureInfo& rhs) const
+bool TextureInfo::operator!=(const TextureInfo& rhs) const
 {
     return !(*this == rhs);
 }
 
-void Material::OcclusionTextureInfo::serialize(object& dst) const
+void OcclusionTextureInfo::serialize(object& dst) const
 {
     TextureInfo::serialize(dst);
     write(mStrength, "strength", dst, 1.f);
 }
 
-const Material::OcclusionTextureInfo& Material::OcclusionTextureInfo::operator=(const Value& src)
+const OcclusionTextureInfo& OcclusionTextureInfo::operator=(const Value& src)
 {
     TextureInfo::operator=(src);
 
@@ -968,13 +1187,13 @@ const Material::OcclusionTextureInfo& Material::OcclusionTextureInfo::operator=(
     return *this;
 }
 
-void Material::NormalTextureInfo::serialize(object& dst) const
+void NormalTextureInfo::serialize(object& dst) const
 {
     TextureInfo::serialize(dst);
     write(mScale, "scale", dst, 1.f);
 }
 
-const Material::NormalTextureInfo& Material::NormalTextureInfo::operator=(const Value& src)
+const NormalTextureInfo& NormalTextureInfo::operator=(const Value& src)
 {
     TextureInfo::operator=(src);
     if (src.is_object())
@@ -1035,18 +1254,12 @@ void Material::Unlit::serialize(object& dst) const
     // no members and object has already been created, nothing to do
 }
 
-void TextureTransform::getPacked(F32* packed) const
+void TextureTransform::getPacked(vec4* packed) const
 {
-    packed[0] = mScale.x;
-    packed[1] = mScale.y;
-    packed[2] = mRotation;
-    packed[3] = mOffset.x;
-    packed[4] = mOffset.y;
-
-    packed[5] = packed[6] = packed[7] = 0.f;
+    packed[0] = vec4(mScale.x, mScale.y, mRotation, mOffset.x);
+    packed[1] = vec4(mOffset.y, 0.f, 0.f, 0.f);
 }
 
-
 const TextureTransform& TextureTransform::operator=(const Value& src)
 {
     mPresent = true;
diff --git a/indra/newview/gltf/asset.h b/indra/newview/gltf/asset.h
index ea3f7d480a..27821659db 100644
--- a/indra/newview/gltf/asset.h
+++ b/indra/newview/gltf/asset.h
@@ -34,6 +34,7 @@
 #include "boost/json.hpp"
 #include "common.h"
 #include "../llviewertexture.h"
+#include "llglslshader.h"
 
 extern F32SecondsImplicit       gFrameTimeSeconds;
 
@@ -65,14 +66,51 @@ namespace LL
             vec2 mScale = vec2(1.f, 1.f);
             S32 mTexCoord = INVALID_INDEX;
 
-            // get the texture transform as a packed array of floats
-            // dst MUST point to at least 8 floats
-            void getPacked(F32* dst) const;
+            // get the texture transform as a packed array of vec4's
+            // dst MUST point to at least 2 vec4's
+            void getPacked(vec4* dst) const;
 
             const TextureTransform& operator=(const Value& src);
             void serialize(boost::json::object& dst) const;
         };
 
+        class TextureInfo
+        {
+        public:
+            S32 mIndex = INVALID_INDEX;
+            S32 mTexCoord = 0;
+
+            TextureTransform mTextureTransform;
+
+            bool operator==(const TextureInfo& rhs) const;
+            bool operator!=(const TextureInfo& rhs) const;
+
+            // get the UV channel that should be used for sampling this texture
+            // returns mTextureTransform.mTexCoord if present and valid, otherwise mTexCoord
+            S32 getTexCoord() const;
+
+            const TextureInfo& operator=(const Value& src);
+            void serialize(boost::json::object& dst) const;
+        };
+
+        class NormalTextureInfo : public TextureInfo
+        {
+        public:
+            F32 mScale = 1.0f;
+
+            const NormalTextureInfo& operator=(const Value& src);
+            void serialize(boost::json::object& dst) const;
+        };
+
+        class OcclusionTextureInfo : public TextureInfo
+        {
+        public:
+            F32 mStrength = 1.0f;
+
+            const OcclusionTextureInfo& operator=(const Value& src);
+            void serialize(boost::json::object& dst) const;
+        };
+
         class Material
         {
         public:
@@ -91,42 +129,6 @@ namespace LL
                 BLEND
             };
 
-            class TextureInfo
-            {
-            public:
-                S32 mIndex = INVALID_INDEX;
-                S32 mTexCoord = 0;
-
-                TextureTransform mTextureTransform;
-
-                bool operator==(const TextureInfo& rhs) const;
-                bool operator!=(const TextureInfo& rhs) const;
-
-                // get the UV channel that should be used for sampling this texture
-                // returns mTextureTransform.mTexCoord if present and valid, otherwise mTexCoord
-                S32 getTexCoord() const;
-
-                const TextureInfo& operator=(const Value& src);
-                void serialize(boost::json::object& dst) const;
-            };
-
-            class NormalTextureInfo : public TextureInfo
-            {
-            public:
-                F32 mScale = 1.0f;
-
-                const NormalTextureInfo& operator=(const Value& src);
-                void serialize(boost::json::object& dst) const;
-            };
-
-            class OcclusionTextureInfo : public TextureInfo
-            {
-            public:
-                F32 mStrength = 1.0f;
-
-                const OcclusionTextureInfo& operator=(const Value& src);
-                void serialize(boost::json::object& dst) const;
-            };
 
             class PbrMetallicRoughness
             {
@@ -179,7 +181,6 @@ namespace LL
         {
         public:
             mat4 mMatrix = glm::identity<mat4>(); //local transform
-            mat4 mRenderMatrix; //transform for rendering
             mat4 mAssetMatrix; //transform from local to asset space
             mat4 mAssetMatrixInv; //transform from asset to local space
 
@@ -206,10 +207,6 @@ namespace LL
             const Node& operator=(const Value& src);
             void serialize(boost::json::object& dst) const;
 
-            // Set mRenderMatrix to a transform that can be used for the current render pass
-            // modelview -- parent's render matrix
-            void updateRenderTransforms(Asset& asset, const mat4& modelview);
-
             // update mAssetMatrix and mAssetMatrixInv
             void updateTransforms(Asset& asset, const mat4& parentMatrix);
 
@@ -322,6 +319,31 @@ namespace LL
             bool prep(Asset& asset);
         };
 
+        // Render Batch -- vertex buffer and list of primitives to render using
+        // said vertex buffer
+        class RenderBatch
+        {
+        public:
+            struct PrimitiveData
+            {
+                S32 mPrimitiveIndex = INVALID_INDEX;
+                S32 mNodeIndex = INVALID_INDEX;
+            };
+
+            LLPointer<LLVertexBuffer> mVertexBuffer;
+            std::vector<PrimitiveData> mPrimitives;
+        };
+
+        class RenderData
+        {
+        public:
+            // list of render batches
+            // indexed by [material index + 1](0 is reserved for default material)
+            // there should be exactly one render batch per material per variant
+            std::vector<RenderBatch> mBatches[LLGLSLShader::NUM_GLTF_VARIANTS];
+        };
+
+
         // C++ representation of a GLTF Asset
         class Asset
         {
@@ -359,6 +381,16 @@ namespace LL
             // the last time update() was called according to gFrameTimeSeconds
             F32 mLastUpdateTime = gFrameTimeSeconds;
 
+            // data used for rendering
+            // 0 - single sided
+            // 1 - double sided
+            RenderData mRenderData[2];
+
+            // UBO for storing node transforms
+            U32 mNodesUBO = 0;
+
+            // UBO for storing material data
+            U32 mMaterialsUBO = 0;
 
             // prepare for first time use
             bool prep();
@@ -373,8 +405,11 @@ namespace LL
             // update asset-to-node and node-to-asset transforms
             void updateTransforms();
 
-            // update node render transforms
-            void updateRenderTransforms(const mat4& modelview);
+            // upload matrices to UBO
+            void uploadTransforms();
+
+            // upload materils to UBO
+            void uploadMaterials();
 
             // return the index of the node that the line segment intersects with, or -1 if no hit
             // input and output values must be in this asset's local coordinate frame
diff --git a/indra/newview/gltf/common.h b/indra/newview/gltf/common.h
index 4f660d7cfc..b9698d4017 100644
--- a/indra/newview/gltf/common.h
+++ b/indra/newview/gltf/common.h
@@ -64,6 +64,9 @@ namespace LL
 
         class Asset;
         class Material;
+        class TextureInfo;
+        class NormalTextureInfo;
+        class OcclusionTextureInfo;
         class Mesh;
         class Node;
         class Scene;
@@ -78,6 +81,17 @@ namespace LL
         class Accessor;
         class BufferView;
         class Buffer;
+
+        enum class TextureType : U8
+        {
+            BASE_COLOR = 0,
+            NORMAL,
+            METALLIC_ROUGHNESS,
+            OCCLUSION,
+            EMISSIVE
+        };
+
+        constexpr U32 TEXTURE_TYPE_COUNT = 5;
     }
 }
 
diff --git a/indra/newview/gltf/primitive.cpp b/indra/newview/gltf/primitive.cpp
index 2280c7004e..e1579374d4 100644
--- a/indra/newview/gltf/primitive.cpp
+++ b/indra/newview/gltf/primitive.cpp
@@ -380,11 +380,22 @@ bool Primitive::prep(Asset& asset)
             }
         }
     }
+    else
+    { //everything must be indexed at runtime
+        mIndexArray.resize(mPositions.size());
+        for (U32 i = 0; i < mPositions.size(); ++i)
+        {
+            mIndexArray[i] = i;
+        }
+    }
 
     U32 mask = LLVertexBuffer::MAP_VERTEX;
 
+    mShaderVariant = 0;
+
     if (!mWeights.empty())
     {
+        mShaderVariant |= LLGLSLShader::GLTFVariant::RIGGED;
         mask |= LLVertexBuffer::MAP_WEIGHT4;
         mask |= LLVertexBuffer::MAP_JOINT;
     }
@@ -406,9 +417,6 @@ bool Primitive::prep(Asset& asset)
         mColors.resize(mPositions.size(), LLColor4U::white);
     }
 
-    mShaderVariant = 0;
-
-    // TODO: support colorless vertex buffers
     mask |= LLVertexBuffer::MAP_COLOR;
 
     bool unlit = false;
@@ -506,69 +514,79 @@ bool Primitive::prep(Asset& asset)
         mask |= LLVertexBuffer::MAP_TANGENT;
     }
 
-    if (LLGLSLShader::sCurBoundShaderPtr == nullptr)
-    { // make sure a shader is bound to satisfy mVertexBuffer->setBuffer
-        gDebugProgram.bind();
+    mAttributeMask = mask;
+
+    if (mMaterial != INVALID_INDEX)
+    {
+        Material& material = asset.mMaterials[mMaterial];
+        if (material.mAlphaMode == Material::AlphaMode::BLEND)
+        {
+            mShaderVariant |= LLGLSLShader::GLTFVariant::ALPHA_BLEND;
+        }
     }
 
-    mVertexBuffer = new LLVertexBuffer(mask);
+    createOctree();
+
+    return true;
+}
+
+void Primitive::upload(LLVertexBuffer* buffer)
+{
+    mVertexBuffer = buffer;
     // we store these buffer sizes as S32 elsewhere
     llassert(mPositions.size() <= size_t(S32_MAX));
     llassert(mIndexArray.size() <= size_t(S32_MAX / 2));
-    mVertexBuffer->allocateBuffer(U32(mPositions.size()), U32(mIndexArray.size() * 2)); // double the size of the index buffer for 32-bit indices
 
-    mVertexBuffer->setBuffer();
-    mVertexBuffer->setPositionData(mPositions.data());
-    mVertexBuffer->setColorData(mColors.data());
+    llassert(mVertexBuffer != nullptr);
+
+    // assert that buffer can hold this primitive
+    llassert(mVertexBuffer->getNumVerts() >= mPositions.size() + mVertexOffset);
+    llassert(mVertexBuffer->getNumIndices() >= mIndexArray.size() + mIndexOffset);
+    llassert(mVertexBuffer->getTypeMask() == mAttributeMask);
+
+    U32 offset = mVertexOffset;
+    U32 count = getVertexCount();
+
+    mVertexBuffer->setPositionData(mPositions.data(), offset, count);
+    mVertexBuffer->setColorData(mColors.data(), offset, count);
 
     if (!mNormals.empty())
     {
-        mVertexBuffer->setNormalData(mNormals.data());
+        mVertexBuffer->setNormalData(mNormals.data(), offset, count);
     }
     if (!mTangents.empty())
     {
-        mVertexBuffer->setTangentData(mTangents.data());
+        mVertexBuffer->setTangentData(mTangents.data(), offset, count);
     }
 
     if (!mWeights.empty())
     {
-        mShaderVariant |= LLGLSLShader::GLTFVariant::RIGGED;
-        mVertexBuffer->setWeight4Data(mWeights.data());
-        mVertexBuffer->setJointData(mJoints.data());
+        mVertexBuffer->setWeight4Data(mWeights.data(), offset, count);
+        mVertexBuffer->setJointData(mJoints.data(), offset, count);
     }
 
     // flip texcoord y, upload, then flip back (keep the off-spec data in vram only)
     vertical_flip(mTexCoords0);
-    mVertexBuffer->setTexCoord0Data(mTexCoords0.data());
+    mVertexBuffer->setTexCoord0Data(mTexCoords0.data(), offset, count);
     vertical_flip(mTexCoords0);
 
     if (!mTexCoords1.empty())
     {
         vertical_flip(mTexCoords1);
-        mVertexBuffer->setTexCoord1Data(mTexCoords1.data());
+        mVertexBuffer->setTexCoord1Data(mTexCoords1.data(), offset, count);
         vertical_flip(mTexCoords1);
     }
 
-
     if (!mIndexArray.empty())
     {
-        mVertexBuffer->setIndexData(mIndexArray.data());
-    }
-
-    createOctree();
-
-    mVertexBuffer->unbind();
-
-    if (mMaterial != INVALID_INDEX)
-    {
-        Material& material = asset.mMaterials[mMaterial];
-        if (material.mAlphaMode == Material::AlphaMode::BLEND)
+        std::vector<U32> index_array;
+        index_array.resize(mIndexArray.size());
+        for (U32 i = 0; i < mIndexArray.size(); ++i)
         {
-            mShaderVariant |= LLGLSLShader::GLTFVariant::ALPHA_BLEND;
+            index_array[i] = mIndexArray[i] + mVertexOffset;
         }
+        mVertexBuffer->setIndexData(index_array.data(), mIndexOffset, getIndexCount());
     }
-
-    return true;
 }
 
 void initOctreeTriangle(LLVolumeTriangle* tri, F32 scaler, S32 i0, S32 i1, S32 i2, const LLVector4a& v0, const LLVector4a& v1, const LLVector4a& v2)
@@ -616,7 +634,7 @@ void Primitive::createOctree()
 
     if (mMode == Mode::TRIANGLES)
     {
-        const U32 num_triangles = mVertexBuffer->getNumIndices() / 3;
+        const U32 num_triangles = getIndexCount() / 3;
         // Initialize all the triangles we need
         mOctreeTriangles.resize(num_triangles);
 
@@ -640,7 +658,7 @@ void Primitive::createOctree()
     }
     else if (mMode == Mode::TRIANGLE_STRIP)
     {
-        const U32 num_triangles = mVertexBuffer->getNumIndices() - 2;
+        const U32 num_triangles = getIndexCount() - 2;
         // Initialize all the triangles we need
         mOctreeTriangles.resize(num_triangles);
 
@@ -664,7 +682,7 @@ void Primitive::createOctree()
     }
     else if (mMode == Mode::TRIANGLE_FAN)
     {
-        const U32 num_triangles = mVertexBuffer->getNumIndices() - 2;
+        const U32 num_triangles = getIndexCount() - 2;
         // Initialize all the triangles we need
         mOctreeTriangles.resize(num_triangles);
 
diff --git a/indra/newview/gltf/primitive.h b/indra/newview/gltf/primitive.h
index 7cc05cf831..304eb26432 100644
--- a/indra/newview/gltf/primitive.h
+++ b/indra/newview/gltf/primitive.h
@@ -54,10 +54,7 @@ namespace LL
 
             ~Primitive();
 
-            // GPU copy of mesh data
-            LLPointer<LLVertexBuffer> mVertexBuffer;
-
-            // CPU copy of mesh data, keep these as LLVector types for compatibility with raycasting code
+            // CPU copy of mesh data
             std::vector<LLVector2> mTexCoords0;
             std::vector<LLVector2> mTexCoords1;
             std::vector<LLVector4a> mNormals;
@@ -80,6 +77,17 @@ namespace LL
             // shader variant according to LLGLSLShader::GLTFVariant flags
             U8 mShaderVariant = 0;
 
+            // vertex attribute mask
+            U32 mAttributeMask = 0;
+
+            // backpointer to vertex buffer (owned by Asset)
+            LLPointer<LLVertexBuffer> mVertexBuffer;
+            U32 mVertexOffset = 0;
+            U32 mIndexOffset = 0;
+
+            U32 getVertexCount() const { return (U32) mPositions.size(); }
+            U32 getIndexCount() const { return (U32) mIndexArray.size(); }
+
             std::unordered_map<std::string, S32> mAttributes;
 
             // create octree based on vertex buffer
@@ -100,6 +108,11 @@ namespace LL
             const Primitive& operator=(const Value& src);
 
             bool prep(Asset& asset);
+
+            // upload geometry to given vertex buffer
+            // asserts that buffer is bound
+            // asserts that buffer is valid for this primitive
+            void upload(LLVertexBuffer* buffer);
         };
     }
 }