backend-test: centralize read pixel logic

2024-05-16 15:08:56 -07:00
643 changed files with 31310 additions and 46855 deletions
@@ -1,17 +0,0 @@
-name: 'Android Continuous'
-inputs:
-  build-abi:
-    description: 'The target platform ABI'
-    required: true
-    default: 'armeabi-v7a'
-runs:
-  using: "composite"
-  steps:
-    - uses: actions/setup-java@v3
-      with:
-        distribution: 'temurin'
-        java-version: '17'
-    - name: Run build script
-      run: |
-        cd build/android && printf "y" | ./build.sh continuous ${{ inputs.build-abi }}
-      shell: bash
@@ -1,9 +0,0 @@
-name: 'ubuntu apt add deb-src'
-runs:
-  using: "composite"
-  steps:
-    - name: "ubuntu apt add deb-src"
-      run: |
-          echo "deb-src http://archive.ubuntu.com/ubuntu jammy main restricted universe" | sudo tee /etc/apt/sources.list.d/my.list
-          sudo apt-get update
-      shell: bash
@@ -10,13 +10,30 @@ on:
 jobs:
  build-android:
    name: build-android
-    # We intentially use a larger runner here to enable larger disk space
-    # (standard linux runner will fail on disk space and faster build time).
-    runs-on: ubuntu-22.04-32core
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
-      - name: Run Android Continuous
-        uses: ./.github/actions/android-continuous
+      - uses: actions/checkout@v3.3.0
+      - uses: actions/setup-java@v3
        with:
-          build-abi: armeabi-v7a,arm64-v8a,x86_64
+          distribution: 'temurin'
+          java-version: '17'
+      - name: Run build script
+        run: |
+          cd build/android && printf "y" | ./build.sh continuous
+      - uses: actions/upload-artifact@v1.0.0
+        with:
+          name: filament-android
+          path: out/filament-android-release.aar
+      - uses: actions/upload-artifact@v1.0.0
+        with:
+          name: filamat-android-full
+          path: out/filamat-android-release.aar
+      - uses: actions/upload-artifact@v1.0.0
+        with:
+          name: gltfio-android-release
+          path: out/gltfio-android-release.aar
+      - uses: actions/upload-artifact@v1.0.0
+        with:
+          name: filament-utils-android-release
+          path: out/filament-utils-android-release.aar
@@ -10,14 +10,14 @@ on:
 jobs:
  build-ios:
    name: build-ios
-    runs-on: macos-14-xlarge
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          cd build/ios && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@v1.0.0
        with:
          name: filament-ios
          path: out/filament-release-ios.tgz
@@ -10,14 +10,14 @@ on:
 jobs:
  build-linux:
    name: build-linux
-    runs-on: ubuntu-22.04-16core
+    runs-on: ubuntu-22.04

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          cd build/linux && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@v1.0.0
        with:
          name: filament-linux
          path: out/filament-release-linux.tgz
@@ -10,14 +10,14 @@ on:
 jobs:
  build-mac:
    name: build-mac
-    runs-on: macos-14-xlarge
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          cd build/mac && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@v1.0.0
        with:
          name: filament-mac
          path: out/filament-release-darwin.tgz
@@ -13,7 +13,7 @@ jobs:
    name: npm-deploy
    runs-on: macos-14
    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
        with:
          ref: ${{ github.event.inputs.release_tag }}
      # Setup .npmrc file to publish to npm
@@ -15,10 +15,10 @@ jobs:

    strategy:
      matrix:
-        os: [macos-14-xlarge, ubuntu-22.04-16core]
+        os: [macos-14, ubuntu-22.04]

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          WORKFLOW_OS=`echo \`uname\` | sed "s/Darwin/mac/" | tr [:upper:] [:lower:]`
@@ -29,10 +29,10 @@ jobs:

  build-windows:
    name: build-windows
-    runs-on: win-2019-16core
+    runs-on: windows-2019

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          build\windows\build-github.bat presubmit
@@ -40,26 +40,24 @@ jobs:

  build-android:
    name: build-android
-    runs-on: ubuntu-22.04-16core
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - uses: actions/setup-java@v3
        with:
          distribution: 'temurin'
          java-version: '17'
      - name: Run build script
-        # Only build 1 64 bit target during presubmit to cut down build times during presubmit
-        # Continuous builds will build everything
        run: |
-          cd build/android && printf "y" | ./build.sh presubmit arm64-v8a
+          cd build/android && printf "y" | ./build.sh presubmit

  build-ios:
    name: build-iOS
-    runs-on: macos-14-xlarge
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          cd build/ios && printf "y" | ./build.sh presubmit
@@ -69,25 +67,10 @@ jobs:

  build-web:
    name: build-web
-    runs-on: ubuntu-22.04-16core
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          cd build/web && printf "y" | ./build.sh presubmit
-
-  test-renderdiff:
-    name: test-renderdiff
-    runs-on: ubuntu-22.04-32core
-
-    steps:
-      - uses: actions/checkout@v4.1.6
-      - uses: ./.github/actions/ubuntu-apt-add-src
-      - name: Run script
-        run: |
-          source ./build/linux/ci-common.sh && bash test/renderdiff_tests.sh
-      - uses: actions/upload-artifact@v4
-        with:
-          name: presubmit-renderdiff-result
-          path: ./out/renderdiff_tests
@@ -31,7 +31,7 @@ jobs:

    strategy:
      matrix:
-        os: [macos-14-xlarge, ubuntu-22.04-32core]
+        os: [macos-14, ubuntu-22.04]

    steps:
      - name: Decide Git ref
@@ -41,7 +41,7 @@ jobs:
          TAG=${REF##*/}
          echo "ref=${REF}" >> $GITHUB_OUTPUT
          echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
        with:
          ref: ${{ steps.git_ref.outputs.ref }}
      - name: Run build script
@@ -65,7 +65,7 @@ jobs:

  build-web:
    name: build-web
-    runs-on: ubuntu-22.04-16core
+    runs-on: macos-14
    if: github.event_name == 'release' || github.event.inputs.platform == 'web'

    steps:
@@ -76,7 +76,7 @@ jobs:
          TAG=${REF##*/}
          echo "ref=${REF}" >> $GITHUB_OUTPUT
          echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
        with:
          ref: ${{ steps.git_ref.outputs.ref }}
      - name: Run build script
@@ -98,7 +98,7 @@ jobs:

  build-android:
    name: build-android
-    runs-on: ubuntu-22.04-16core
+    runs-on: macos-14
    if: github.event_name == 'release' || github.event.inputs.platform == 'android'

    steps:
@@ -109,7 +109,7 @@ jobs:
          TAG=${REF##*/}
          echo "ref=${REF}" >> $GITHUB_OUTPUT
          echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
        with:
          ref: ${{ steps.git_ref.outputs.ref }}
      - uses: actions/setup-java@v3
@@ -120,7 +120,7 @@ jobs:
        env:
          TAG: ${{ steps.git_ref.outputs.tag }}
        run: |
-          cd build/android && printf "y" | ./build.sh release armeabi-v7a,arm64-v8a,x86,x86_64
+          cd build/android && printf "y" | ./build.sh release
          cd ../..
          mv out/filament-android-release.aar out/filament-${TAG}-android.aar
          mv out/filamat-android-release.aar out/filamat-${TAG}-android.aar
@@ -152,7 +152,7 @@ jobs:

  build-ios:
    name: build-ios
-    runs-on: macos-14-xlarge
+    runs-on: macos-14
    if: github.event_name == 'release' || github.event.inputs.platform == 'ios'

    steps:
@@ -163,7 +163,7 @@ jobs:
          TAG=${REF##*/}
          echo "ref=${REF}" >> $GITHUB_OUTPUT
          echo "tag=${TAG}" >> $GITHUB_OUTPUT
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
        with:
          ref: ${{ steps.git_ref.outputs.ref }}
      - name: Run build script
@@ -185,7 +185,7 @@ jobs:

  build-windows:
    name: build-windows
-    runs-on: windows-2019-32core
+    runs-on: windows-2019
    if: github.event_name == 'release' || github.event.inputs.platform == 'windows'

    steps:
@@ -197,7 +197,7 @@ jobs:
          echo "ref=${REF}" >> $GITHUB_OUTPUT
          echo "tag=${TAG}" >> $GITHUB_OUTPUT
        shell: bash
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
        with:
          ref: ${{ steps.git_ref.outputs.ref }}
      - name: Run build script
@@ -205,7 +205,6 @@ jobs:
          TAG: ${{ steps.git_ref.outputs.tag }}
        run: |
          build\windows\build-github.bat release
-          echo on
          move out\filament-windows.tgz out\filament-%TAG%-windows.tgz
        shell: cmd
      - uses: actions/github-script@v6
@@ -10,14 +10,14 @@ on:
 jobs:
  build-web:
    name: build-web
-    runs-on: ubuntu-22.04-16core
+    runs-on: macos-14

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          cd build/web && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@v1.0.0
        with:
          name: filament-web
          path: out/filament-release-web.tgz
@@ -10,15 +10,15 @@ on:
 jobs:
  build-windows:
    name: build-windows
-    runs-on: windows-2019-32core
+    runs-on: windows-2019

    steps:
-      - uses: actions/checkout@v4.1.6
+      - uses: actions/checkout@v3.3.0
      - name: Run build script
        run: |
          build\windows\build-github.bat continuous
        shell: cmd
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@v1.0.0
        with:
          name: filament-windows
          path: out/filament-windows.tgz
@@ -437,11 +437,10 @@ Finally simply open `docs/html/index.html` in your web browser.

 ## Software Rasterization

-We have tested swiftshader and Mesa for software rasterization on the Vulkan/GL backends.
-
-To use this for Vulkan, please first make sure that the [Vulkan SDK](https://www.lunarg.com/vulkan-sdk/) is
-installed on your machine. If you are doing a manual installation of the SDK on Linux, you will have
-to source `setup-env.sh` in the SDK's root folder to make sure the Vulkan loader is the first lib loaded.
+We have tested swiftshader for running software rasterization on the Vulkan backend. To use this,
+please first make sure that the [Vulkan SDK](https://www.lunarg.com/vulkan-sdk/) is installed on
+your machine. If you are doing a manual installation of the SDK on Linux, you will have to source
+`setup-env.sh` in the SDK's root folder to make sure the Vulkan loader is the first lib loaded.

 ### Swiftshader (Vulkan) [tested on macOS and Linux]

@@ -458,39 +457,5 @@ and then set `VK_ICD_FILENAMES` to the ICD json produced in the build. For examp
 export VK_ICD_FILENAMES=/Users/user/swiftshader/build/Darwin/vk_swiftshader_icd.json
 ```

-Build and run Filament as usual and specify the Vulkan backend when creating the Engine.
+Build Filament as normal and use the vulkan backend.

-### Mesa's LLVMPipe (GL) and Lavapipe (Vulkan) [tested on Linux]
-
-We will only cover steps that build Mesa from source. The official documentation of Mesa mentioned
-that in general precompiled libraries [are **not** made available](https://docs.mesa3d.org/precompiled.html).
-
-Download the repo and make sure you have the build depedencies. For example (assuming an Ubuntu/Debian distro),
-```shell
-git clone https://gitlab.freedesktop.org/mesa/mesa.git
-sudo apt-get build-dep mesa
-```
-
-To build both the GL and Vulkan rasterizers,
-
-```shell
-cd mesa
-mkdir -p out
-meson setup builddir/ -Dprefix=$(pwd)/out -Dglx=xlib -Dgallium-drivers=swrast -Dvulkan-drivers=swrast
-meson install -C builddir/
-```
-
-For GL, we need to ensure that we load the GL lib from the mesa output directory.  For example, to run
-the debug `gltf_viewer`, we would execute
-```shell
-LD_LIBRARY_PATH=/Users/user/mesa/out/lib/x86_64-linux-gnu \
-    ./out/cmake-debug/samples/gltf_viewer -a opengl
-```
-
-For Vulkan, we need to set the path to the ICD json, which tells the loader where to find the driver
-library. To run `gltf_viewer`, we would execute
-```shell
-VK_ICD_FILENAMES=/Users/user/mesa/out/share/vulkan/icd.d/lvp_icd.x86_64.json \
-    ./out/cmake-debug/samples/gltf_viewer -a vulkan
-
-```
@@ -45,8 +45,6 @@ option(FILAMENT_ENABLE_FEATURE_LEVEL_0 "Enable Feature Level 0" ON)

 option(FILAMENT_ENABLE_MULTIVIEW "Enable multiview for Filament" OFF)

-option(FILAMENT_SUPPORTS_OSMESA "Enable OSMesa (headless GL context) for Filament" OFF)
-
 set(FILAMENT_NDK_VERSION "" CACHE STRING
    "Android NDK version or version prefix to be used when building for Android."
 )
@@ -75,10 +73,6 @@ set(FILAMENT_BACKEND_DEBUG_FLAG "" CACHE STRING
    "A debug flag meant for enabling/disabling backend debugging paths"
 )

-set(FILAMENT_OSMESA_PATH "" CACHE STRING
-    "Path to the OSMesa header and lib"
-)
-
 # Enable exceptions by default in spirv-cross.
 set(SPIRV_CROSS_EXCEPTIONS_TO_ASSERTIONS OFF)

@@ -138,22 +132,12 @@ else()
 endif()

 if (LINUX)
-    if (NOT FILAMENT_OSMESA_PATH STREQUAL "")
-        if (NOT EXISTS ${FILAMENT_OSMESA_PATH}/)
-            message(FATAL_ERROR "Cannot find specified OSMesa build directory: ${FILAMENT_OSMESA_PATH}")
-        endif()
-        set(FILAMENT_SUPPORTS_OSMESA TRUE)
-    endif()
-
    if (FILAMENT_SUPPORTS_WAYLAND)
        add_definitions(-DFILAMENT_SUPPORTS_WAYLAND)
        set(FILAMENT_SUPPORTS_X11 FALSE)
    elseif (FILAMENT_SUPPORTS_EGL_ON_LINUX)
        add_definitions(-DFILAMENT_SUPPORTS_EGL_ON_LINUX)
        set(FILAMENT_SUPPORTS_X11 FALSE)
-    elseif (FILAMENT_SUPPORTS_OSMESA)
-        set(FILAMENT_SUPPORTS_X11 FALSE)
-        add_definitions(-DFILAMENT_SUPPORTS_OSMESA)
    else ()
        if (FILAMENT_SUPPORTS_XCB)
            add_definitions(-DFILAMENT_SUPPORTS_XCB)
@@ -540,15 +524,13 @@ else()
 endif()

 # This only affects the prebuilt shader files in gltfio and samples, not filament library.
-# The value can be either "instanced", "multiview", or "none"
-set(FILAMENT_SAMPLES_STEREO_TYPE "none" CACHE STRING
+# The value can be either "instanced" or "multiview".
+set(FILAMENT_SAMPLES_STEREO_TYPE "instanced" CACHE STRING
    "Stereoscopic type that shader files in gltfio and samples are built for."
 )
 string(TOLOWER "${FILAMENT_SAMPLES_STEREO_TYPE}" FILAMENT_SAMPLES_STEREO_TYPE)
-if (NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "instanced"
-    AND NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview"
-    AND NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "none")
-    message(FATAL_ERROR "Invalid stereo type: \"${FILAMENT_SAMPLES_STEREO_TYPE}\" choose either \"instanced\",  \"multiview\", or \"none\" ")
+if (NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "instanced" AND NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview")
+    message(FATAL_ERROR "Invalid stereo type: \"${FILAMENT_SAMPLES_STEREO_TYPE}\" choose either \"instanced\" or \"multiview\" ")
 endif ()

 # Compiling samples for multiview implies enabling multiview feature as well.
@@ -751,6 +733,7 @@ add_subdirectory(${FILAMENT}/filament)
 add_subdirectory(${FILAMENT}/shaders)
 add_subdirectory(${EXTERNAL}/basisu/tnt)
 add_subdirectory(${EXTERNAL}/civetweb/tnt)
+add_subdirectory(${EXTERNAL}/hat-trie/tnt)
 add_subdirectory(${EXTERNAL}/imgui/tnt)
 add_subdirectory(${EXTERNAL}/robin-map/tnt)
 add_subdirectory(${EXTERNAL}/smol-v/tnt)
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.55.1'
+    implementation 'com.google.android.filament:filament-android:1.51.8'
 }
 ```

@@ -51,9 +51,19 @@ Here are all the libraries available in the group `com.google.android.filament`:
 iOS projects can use CocoaPods to install the latest release:

 ```shell
-pod 'Filament', '~> 1.55.1'
+pod 'Filament', '~> 1.51.8'
 ```

+### Snapshots
+
+If you prefer to live on the edge, you can download a continuous build by following the following
+steps:
+
+1. Find the [commit](https://github.com/google/filament/commits/main) you're interested in.
+2. Click the green check mark under the commit message.
+3. Click on the _Details_ link for the platform you're interested in.
+4. On the top left click _Summary_, then in the _Artifacts_ section choose the desired artifact.
+
 ## Documentation

 - [Filament](https://google.github.io/filament/Filament.html), an in-depth explanation of
@@ -166,7 +176,6 @@ pod 'Filament', '~> 1.55.1'
  - [x] KHR_materials_unlit
  - [x] KHR_materials_variants
  - [x] KHR_materials_volume
-  - [x] KHR_materials_specular
  - [x] KHR_mesh_quantization
  - [x] KHR_texture_basisu
  - [x] KHR_texture_transform
@@ -7,66 +7,6 @@ A new header is inserted each time a *tag* is created.
 Instead, if you are authoring a PR for the main branch, add your release note to
 [NEW_RELEASE_NOTES.md](./NEW_RELEASE_NOTES.md).

-## v1.55.1
-
-
-## v1.55.0
- Add descriptor sets to describe shader resources. [⚠️ **New Material Version**]
-
-## v1.54.5
-
-
-## v1.54.4
-
- Add support for multi-layered render target with array textures.
-
-## v1.54.3
-
-
-## v1.54.2
-
- Add a `name` API to Filament objects for debugging handle use-after-free assertions
-
-## v1.54.1
-
-
-## v1.54.0
-
- materials: add a new `stereoscopicType` material parameter. [⚠️ **New Material Version**]
- Fix a crash when compiling shaders on IMG devices
-
-## v1.53.5
-
- engine: Fix bug causing certain sampler parameters to not be applied correctly in GLES 2.0 and on
-  certain GLES 3.0 drivers.
-
-## v1.53.4
-
-
-## v1.53.3
-
- Add drag and drop support for IBL files for desktop gltf_viewer.
-
-## v1.53.2
-
-
-## v1.53.1
-
-
-## v1.53.0
-
- engine: fix skinning normals with large transforms (b/342459864) [⚠️ **New Material Version**]
-
-## v1.52.3
-
-
-## v1.52.2
-
-
-## v1.52.1
-
- Add instructions for using Mesa for software rasterization
-
 ## v1.51.9


@@ -83,12 +83,12 @@ buildscript {
        'minSdk': 21,
        'targetSdk': 34,
        'compileSdk': 34,
-        'kotlin': '2.0.0',
-        'kotlin_coroutines': '1.9.0-RC',
+        'kotlin': '1.9.21',
+        'kotlin_coroutines': '1.7.3',
        'buildTools': '34.0.0',
-        'ndk': '27.0.11718014',
-        'androidx_core': '1.13.1',
-        'androidx_annotations': '1.8.0'
+        'ndk': '26.1.10909125',
+        'androidx_core': '1.12.0',
+        'androidx_annotations': '1.7.0'
    ]

    ext.deps = [
@@ -104,7 +104,7 @@ buildscript {
    ]

    dependencies {
-        classpath 'com.android.tools.build:gradle:8.4.1'
+        classpath 'com.android.tools.build:gradle:8.2.0'
        classpath "org.jetbrains.kotlin:kotlin-gradle-plugin:${versions.kotlin}"
    }

@@ -420,13 +420,6 @@ Java_com_google_android_filament_Engine_nSetPaused(JNIEnv*, jclass,
    engine->setPaused(paused);
 }

-extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_Engine_nUnprotected(JNIEnv*, jclass,
-        jlong nativeEngine, jboolean paused) {
-    Engine* engine = (Engine*) nativeEngine;
-    engine->unprotected();
-}
-
 // Managers...

 extern "C" JNIEXPORT jlong JNICALL
@@ -520,12 +513,13 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
 extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBuilderConfig(JNIEnv*,
        jclass, jlong nativeBuilder, jlong commandBufferSizeMB, jlong perRenderPassArenaSizeMB,
        jlong driverHandleArenaSizeMB, jlong minCommandBufferSizeMB, jlong perFrameCommandsSizeMB,
-        jlong jobSystemThreadCount, jboolean disableParallelShaderCompile,
+        jlong jobSystemThreadCount,
+        jlong textureUseAfterFreePoolSize, jboolean disableParallelShaderCompile,
        jint stereoscopicType, jlong stereoscopicEyeCount,
        jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge,
        jboolean disableHandleUseAfterFreeCheck,
        jint preferredShaderLanguage,
-        jboolean forceGLES2Context, jboolean assertNativeWindowIsValid) {
+        jboolean forceGLES2Context) {
    Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
    Engine::Config config = {
            .commandBufferSizeMB = (uint32_t) commandBufferSizeMB,
@@ -534,6 +528,7 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
            .minCommandBufferSizeMB = (uint32_t) minCommandBufferSizeMB,
            .perFrameCommandsSizeMB = (uint32_t) perFrameCommandsSizeMB,
            .jobSystemThreadCount = (uint32_t) jobSystemThreadCount,
+            .textureUseAfterFreePoolSize = (uint32_t) textureUseAfterFreePoolSize,
            .disableParallelShaderCompile = (bool) disableParallelShaderCompile,
            .stereoscopicType = (Engine::StereoscopicType) stereoscopicType,
            .stereoscopicEyeCount = (uint8_t) stereoscopicEyeCount,
@@ -542,7 +537,6 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
            .disableHandleUseAfterFreeCheck = (bool) disableHandleUseAfterFreeCheck,
            .preferredShaderLanguage = (Engine::Config::ShaderLanguage) preferredShaderLanguage,
            .forceGLES2Context = (bool) forceGLES2Context,
-            .assertNativeWindowIsValid = (bool) assertNativeWindowIsValid,
    };
    builder->config(&config);
 }
@@ -570,9 +564,3 @@ Java_com_google_android_filament_Engine_nBuilderBuild(JNIEnv*, jclass, jlong nat
    Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
    return (jlong) builder->build();
 }
-
-extern "C"
-JNIEXPORT jlong JNICALL
-Java_com_google_android_filament_Engine_getSteadyClockTimeNano(JNIEnv *env, jclass clazz) {
-    return (jlong)Engine::getSteadyClockTimeNano();
-}
@@ -25,17 +25,12 @@ using namespace filament;

 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Material_nBuilderBuild(JNIEnv *env, jclass,
-        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount) {
+        jlong nativeEngine, jobject buffer_, jint size) {
    Engine* engine = (Engine*) nativeEngine;
    AutoBuffer buffer(env, buffer_, size);
-    auto builder = Material::Builder();
-    if (shBandCount) {
-        builder.sphericalHarmonicsBandCount(shBandCount);
-    }
-    Material* material = builder
+    Material* material = Material::Builder()
            .package(buffer.getData(), buffer.getSize())
            .build(*engine);
-
    return (jlong) material;
 }

@@ -245,18 +245,12 @@ Java_com_google_android_filament_RenderableManager_nBuilderMorphing(JNIEnv*, jcl
 }

 extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nBuilderMorphingStandard(JNIEnv*, jclass,
-        jlong nativeBuilder, jlong nativeMorphTargetBuffer) {
+Java_com_google_android_filament_RenderableManager_nBuilderSetMorphTargetBufferAt(JNIEnv*, jclass,
+        jlong nativeBuilder, int level, int primitiveIndex, jlong nativeMorphTargetBuffer,
+        int offset, int count) {
    RenderableManager::Builder *builder = (RenderableManager::Builder *) nativeBuilder;
    MorphTargetBuffer *morphTargetBuffer = (MorphTargetBuffer *) nativeMorphTargetBuffer;
-    builder->morphing(morphTargetBuffer);
-}
-
-extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nBuilderSetMorphTargetBufferOffsetAt(JNIEnv*, jclass,
-        jlong nativeBuilder, int level, int primitiveIndex, int offset) {
-    RenderableManager::Builder *builder = (RenderableManager::Builder *) nativeBuilder;
-    builder->morphing(level, primitiveIndex, offset);
+    builder->morphing(level, primitiveIndex, morphTargetBuffer, offset, count);
 }

 extern "C" JNIEXPORT void JNICALL
@@ -328,12 +322,13 @@ Java_com_google_android_filament_RenderableManager_nSetMorphWeights(JNIEnv* env,
 }

 extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nSetMorphTargetBufferOffsetAt(JNIEnv*,
+Java_com_google_android_filament_RenderableManager_nSetMorphTargetBufferAt(JNIEnv*,
        jclass, jlong nativeRenderableManager, jint i, int level, jint primitiveIndex,
-        jlong, jint offset) {
+        jlong nativeMorphTargetBuffer, jint offset, jint count) {
    RenderableManager *rm = (RenderableManager *) nativeRenderableManager;
-    rm->setMorphTargetBufferOffsetAt((RenderableManager::Instance) i, (uint8_t) level,
-            (size_t) primitiveIndex, (size_t) offset);
+    MorphTargetBuffer *morphTargetBuffer = (MorphTargetBuffer *) nativeMorphTargetBuffer;
+    rm->setMorphTargetBufferAt((RenderableManager::Instance) i, (uint8_t) level,
+            (size_t) primitiveIndex, morphTargetBuffer, (size_t) offset, (size_t) count);
 }

 extern "C" JNIEXPORT jint JNICALL
@@ -28,14 +28,6 @@
 using namespace filament;
 using namespace backend;

-
-extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_Renderer_nSkipFrame(JNIEnv *, jclass, jlong nativeRenderer,
-        jlong vsyncSteadyClockTimeNano) {
-    Renderer *renderer = (Renderer *) nativeRenderer;
-    renderer->skipFrame(uint64_t(vsyncSteadyClockTimeNano));
-}
-
 extern "C" JNIEXPORT jboolean JNICALL
 Java_com_google_android_filament_Renderer_nBeginFrame(JNIEnv *, jclass, jlong nativeRenderer,
        jlong nativeSwapChain, jlong frameTimeNanos) {
@@ -195,10 +187,3 @@ Java_com_google_android_filament_Renderer_nSetPresentationTime(JNIEnv *, jclass
    Renderer *renderer = (Renderer *) nativeRenderer;
    renderer->setPresentationTime(monotonicClockNanos);
 }
-
-extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_Renderer_nSetVsyncTime(JNIEnv *, jclass,
-    jlong nativeRenderer, jlong steadyClockTimeNano) {
-    Renderer *renderer = (Renderer *) nativeRenderer;
-    renderer->setVsyncTime(steadyClockTimeNano);
-}
@@ -531,12 +531,3 @@ Java_com_google_android_filament_View_nGetFogEntity(JNIEnv *env, jclass clazz,
    View *view = (View *) nativeView;
    return (jint)view->getFogEntity().getId();
 }
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_google_android_filament_View_nClearFrameHistory(JNIEnv *env, jclass clazz,
-        jlong nativeView, jlong nativeEngine) {
-    View *view = (View *) nativeView;
-    Engine *engine = (Engine *) nativeEngine;
-    view->clearFrameHistory(*engine);
-}
@@ -159,12 +159,9 @@ public class Engine {
    };

    /**
-     * The type of technique for stereoscopic rendering. (Note that the materials used will need to be
-     * compatible with the chosen technique.)
+     * The type of technique for stereoscopic rendering
     */
    public enum StereoscopicType {
-        /** No stereoscopic rendering. */
-        NONE,
        /** Stereoscopic rendering is performed using instanced rendering technique. */
        INSTANCED,
        /** Stereoscopic rendering is performed using the multiview feature from the graphics backend. */
@@ -224,12 +221,13 @@ public class Engine {
            nSetBuilderConfig(mNativeBuilder, config.commandBufferSizeMB,
                    config.perRenderPassArenaSizeMB, config.driverHandleArenaSizeMB,
                    config.minCommandBufferSizeMB, config.perFrameCommandsSizeMB,
-                    config.jobSystemThreadCount, config.disableParallelShaderCompile,
+                    config.jobSystemThreadCount,
+                    config.textureUseAfterFreePoolSize, config.disableParallelShaderCompile,
                    config.stereoscopicType.ordinal(), config.stereoscopicEyeCount,
                    config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge,
                    config.disableHandleUseAfterFreeCheck,
                    config.preferredShaderLanguage.ordinal(),
-                    config.forceGLES2Context, config.assertNativeWindowIsValid);
+                    config.forceGLES2Context);
            return this;
        }

@@ -407,7 +405,7 @@ public class Engine {
         *
         * @see View#setStereoscopicOptions
         */
-        public StereoscopicType stereoscopicType = StereoscopicType.NONE;
+        public StereoscopicType stereoscopicType = StereoscopicType.INSTANCED;

        /**
         * The number of eyes to render when stereoscopic rendering is enabled. Supported values are
@@ -418,25 +416,22 @@ public class Engine {
         */
        public long stereoscopicEyeCount = 2;

-        /**
+        /*
         * @Deprecated This value is no longer used.
         */
        public long resourceAllocatorCacheSizeMB = 64;

-        /**
-         * This value determines how many frames texture entries are kept for in the cache. This
-         * is a soft limit, meaning some texture older than this are allowed to stay in the cache.
-         * Typically only one texture is evicted per frame.
-         * The default is 1.
+        /*
+         * This value determines for how many frames are texture entries kept in the cache.
         */
-        public long resourceAllocatorCacheMaxAge = 1;
+        public long resourceAllocatorCacheMaxAge = 2;

-        /**
+        /*
         * Disable backend handles use-after-free checks.
         */
        public boolean disableHandleUseAfterFreeCheck = false;

-        /**
+        /*
         * Sets a preferred shader language for Filament to use.
         *
         * The Metal backend supports two shader languages: MSL (Metal Shading Language) and
@@ -458,19 +453,12 @@ public class Engine {
        };
        public ShaderLanguage preferredShaderLanguage = ShaderLanguage.DEFAULT;

-        /**
+        /*
         * When the OpenGL ES backend is used, setting this value to true will force a GLES2.0
         * context if supported by the Platform, or if not, will have the backend pretend
         * it's a GLES2 context. Ignored on other backends.
         */
        public boolean forceGLES2Context = false;
-
-        /**
-         * Assert the native window associated to a SwapChain is valid when calling makeCurrent().
-         * This is only supported for:
-         *      - PlatformEGLAndroid
-         */
-        public boolean assertNativeWindowIsValid = false;
    }

    private Engine(long nativeEngine, Config config) {
@@ -1298,24 +1286,6 @@ public class Engine {
        nSetPaused(getNativeObject(), paused);
    }

-    /**
-     * Switch the command queue to unprotected mode. Protected mode can be activated via
-     * Renderer::beginFrame() using a protected SwapChain.
-     * @see Renderer
-     * @see SwapChain
-     */
-    public void unprotected() {
-        nUnprotected(getNativeObject());
-    }
-
-    /**
-     * Get the current time. This is a convenience function that simply returns the
-     * time in nanosecond since epoch of std::chrono::steady_clock.
-     * @return current time in nanosecond since epoch of std::chrono::steady_clock.
-     * @see Renderer#beginFrame
-     */
-    public static native long getSteadyClockTimeNano();
-
    @UsedByReflection("TextureHelper.java")
    public long getNativeObject() {
        if (mNativeObject == 0) {
@@ -1393,7 +1363,6 @@ public class Engine {
    private static native void nFlush(long nativeEngine);
    private static native boolean nIsPaused(long nativeEngine);
    private static native void nSetPaused(long nativeEngine, boolean paused);
-    private static native void nUnprotected(long nativeEngine);
    private static native long nGetTransformManager(long nativeEngine);
    private static native long nGetLightManager(long nativeEngine);
    private static native long nGetRenderableManager(long nativeEngine);
@@ -1412,11 +1381,12 @@ public class Engine {
    private static native void nSetBuilderConfig(long nativeBuilder, long commandBufferSizeMB,
            long perRenderPassArenaSizeMB, long driverHandleArenaSizeMB,
            long minCommandBufferSizeMB, long perFrameCommandsSizeMB, long jobSystemThreadCount,
-            boolean disableParallelShaderCompile, int stereoscopicType, long stereoscopicEyeCount,
+            long textureUseAfterFreePoolSize, boolean disableParallelShaderCompile,
+            int stereoscopicType, long stereoscopicEyeCount,
            long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge,
            boolean disableHandleUseAfterFreeCheck,
            int preferredShaderLanguage,
-            boolean forceGLES2Context, boolean assertNativeWindowIsValid);
+            boolean forceGLES2Context);
    private static native void nSetBuilderFeatureLevel(long nativeBuilder, int ordinal);
    private static native void nSetBuilderSharedContext(long nativeBuilder, long sharedContext);
    private static native void nSetBuilderPaused(long nativeBuilder, boolean paused);
@@ -346,7 +346,6 @@ public class Material {
    public static class Builder {
        private Buffer mBuffer;
        private int mSize;
-        private int mShBandCount = 0;

        /**
         * Specifies the material data. The material data is a binary blob produced by
@@ -362,22 +361,6 @@ public class Material {
            return this;
        }

-        /**
-         * Sets the quality of the indirect lights computations. This is only taken into account
-         * if this material is lit and in the surface domain. This setting will affect the
-         * IndirectLight computation if one is specified on the Scene and Spherical Harmonics
-         * are used for the irradiance.
-         *
-         * @param shBandCount Number of spherical harmonic bands. Must be 1, 2 or 3 (default).
-         * @return Reference to this Builder for chaining calls.
-         * @see IndirectLight
-         */
-        @NonNull
-        public Builder sphericalHarmonicsBandCount(@IntRange(from = 0) int shBandCount) {
-            mShBandCount = shBandCount;
-            return this;
-        }
-
        /**
         * Creates and returns the Material object.
         *
@@ -389,8 +372,7 @@ public class Material {
         */
        @NonNull
        public Material build(@NonNull Engine engine) {
-            long nativeMaterial = nBuilderBuild(engine.getNativeObject(),
-                mBuffer, mSize, mShBandCount);
+            long nativeMaterial = nBuilderBuild(engine.getNativeObject(), mBuffer, mSize);
            if (nativeMaterial == 0) throw new IllegalStateException("Couldn't create Material");
            return new Material(nativeMaterial);
        }
@@ -1041,7 +1023,7 @@ public class Material {
        mNativeObject = 0;
    }

-    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount);
+    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size);
    private static native long nCreateInstance(long nativeMaterial);
    private static native long nCreateInstanceWithName(long nativeMaterial, @NonNull String name);
    private static native long nGetDefaultInstance(long nativeMaterial);
@@ -74,7 +74,7 @@ public class MorphTargetBuffer {
         *
         * @exception IllegalStateException if the MorphTargetBuffer could not be created
         *
-         * @see #setMorphTargetBufferOffsetAt
+         * @see #setMorphTargetBufferAt
         */
        @NonNull
        public MorphTargetBuffer build(@NonNull Engine engine) {
@@ -524,7 +524,14 @@ public class RenderableManager {
        }

        /**
-         * Controls if the renderable has legacy vertex morphing targets, zero by default.
+         * Controls if the renderable has vertex morphing targets, zero by default. This is
+         * required to enable GPU morphing.
+         *
+         * <p>Filament supports two morphing modes: standard (default) and legacy.</p>
+         *
+         * <p>For standard morphing, A {@link MorphTargetBuffer} must be created and provided via
+         * {@link RenderableManager#setMorphTargetBufferAt}. Standard morphing supports up to
+         * <code>CONFIG_MAX_MORPH_TARGET_COUNT</code> morph targets.</p>
         *
         * For legacy morphing, the attached {@link VertexBuffer} must provide data in the
         * appropriate {@link VertexBuffer.VertexAttribute} slots (<code>MORPH_POSITION_0</code> etc).
@@ -542,22 +549,6 @@ public class RenderableManager {
            return this;
        }

-        /**
-         * Controls if the renderable has vertex morphing targets, zero by default.
-         *
-         * <p>For standard morphing, A {@link MorphTargetBuffer} must be provided.
-         * Standard morphing supports up to
-         * <code>CONFIG_MAX_MORPH_TARGET_COUNT</code> morph targets.</p>
-         *
-         * <p>See also {@link RenderableManager#setMorphWeights}, which can be called on a per-frame basis
-         * to advance the animation.</p>
-         */
-        @NonNull
-        public Builder morphing(@NonNull MorphTargetBuffer morphTargetBuffer) {
-            nBuilderMorphingStandard(mNativeBuilder, morphTargetBuffer.getNativeObject());
-            return this;
-        }
-
        /**
         * Specifies the morph target buffer for a primitive.
         *
@@ -569,13 +560,31 @@ public class RenderableManager {
         *
         * @param level the level of detail (lod), only 0 can be specified
         * @param primitiveIndex zero-based index of the primitive, must be less than the count passed to Builder constructor
+         * @param morphTargetBuffer specifies the morph target buffer
         * @param offset specifies where in the morph target buffer to start reading (expressed as a number of vertices)
+         * @param count number of vertices in the morph target buffer to read, must equal the geometry's count (for triangles, this should be a multiple of 3)
         */
        @NonNull
        public Builder morphing(@IntRange(from = 0) int level,
                                @IntRange(from = 0) int primitiveIndex,
-                                @IntRange(from = 0) int offset) {
-            nBuilderSetMorphTargetBufferOffsetAt(mNativeBuilder, level, primitiveIndex, offset);
+                                @NonNull MorphTargetBuffer morphTargetBuffer,
+                                @IntRange(from = 0) int offset,
+                                @IntRange(from = 0) int count) {
+            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex,
+                    morphTargetBuffer.getNativeObject(), offset, count);
+            return this;
+        }
+
+        /**
+         * Utility method to specify morph target buffer for a primitive.
+         * For details, see the {@link RenderableManager.Builder#morphing}.
+         */
+        @NonNull
+        public Builder morphing(@IntRange(from = 0) int level,
+                                @IntRange(from = 0) int primitiveIndex,
+                                @NonNull MorphTargetBuffer morphTargetBuffer) {
+            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex,
+                    morphTargetBuffer.getNativeObject(), 0, morphTargetBuffer.getVertexCount());
            return this;
        }

@@ -678,11 +687,26 @@ public class RenderableManager {
     *
     * @see Builder#morphing
     */
-    public void setMorphTargetBufferOffsetAt(@EntityInstance int i,
+    public void setMorphTargetBufferAt(@EntityInstance int i,
                                       @IntRange(from = 0) int level,
                                       @IntRange(from = 0) int primitiveIndex,
-                                       @IntRange(from = 0) int offset) {
-        nSetMorphTargetBufferOffsetAt(mNativeObject, i, level, primitiveIndex, 0, offset);
+                                       @NonNull MorphTargetBuffer morphTargetBuffer,
+                                       @IntRange(from = 0) int offset,
+                                       @IntRange(from = 0) int count) {
+        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex,
+                morphTargetBuffer.getNativeObject(), offset, count);
+    }
+
+    /**
+     * Utility method to change morph target buffer for the given primitive.
+     * For details, see the {@link RenderableManager#setMorphTargetBufferAt}.
+     */
+    public void setMorphTargetBufferAt(@EntityInstance int i,
+                                       @IntRange(from = 0) int level,
+                                       @IntRange(from = 0) int primitiveIndex,
+                                       @NonNull MorphTargetBuffer morphTargetBuffer) {
+        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex,
+                morphTargetBuffer.getNativeObject(), 0, morphTargetBuffer.getVertexCount());
    }

    /**
@@ -982,8 +1006,7 @@ public class RenderableManager {
    private static native int nBuilderSkinningBones(long nativeBuilder, int boneCount, Buffer bones, int remaining);
    private static native void nBuilderSkinningBuffer(long nativeBuilder, long nativeSkinningBuffer, int boneCount, int offset);
    private static native void nBuilderMorphing(long nativeBuilder, int targetCount);
-    private static native void nBuilderMorphingStandard(long nativeBuilder, long nativeMorphTargetBuffer);
-    private static native void nBuilderSetMorphTargetBufferOffsetAt(long nativeBuilder, int level, int primitiveIndex, int offset);
+    private static native void nBuilderSetMorphTargetBufferAt(long nativeBuilder, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset, int count);
    private static native void nBuilderEnableSkinningBuffers(long nativeBuilder, boolean enabled);
    private static native void nBuilderFog(long nativeBuilder, boolean enabled);
    private static native void nBuilderLightChannel(long nativeRenderableManager, int channel, boolean enable);
@@ -993,7 +1016,7 @@ public class RenderableManager {
    private static native int nSetBonesAsMatrices(long nativeObject, int i, Buffer matrices, int remaining, int boneCount, int offset);
    private static native int nSetBonesAsQuaternions(long nativeObject, int i, Buffer quaternions, int remaining, int boneCount, int offset);
    private static native void nSetMorphWeights(long nativeObject, int instance, float[] weights, int offset);
-    private static native void nSetMorphTargetBufferOffsetAt(long nativeObject, int i, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset);
+    private static native void nSetMorphTargetBufferAt(long nativeObject, int i, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset, int count);
    private static native int nGetMorphTargetCount(long nativeObject, int i);
    private static native void nSetAxisAlignedBoundingBox(long nativeRenderableManager, int i, float cx, float cy, float cz, float ex, float ey, float ez);
    private static native void nSetLayerMask(long nativeRenderableManager, int i, int select, int value);
@@ -284,33 +284,6 @@ public class Renderer {
        nSetPresentationTime(getNativeObject(), monotonicClockNanos);
    }

-    /**
-     * The use of this method is optional. It sets the VSYNC time expressed as the duration in
-     * nanosecond since epoch of std::chrono::steady_clock.
-     * If called, passing 0 to frameTimeNanos in Renderer.BeginFrame will use this
-     * time instead.
-     * @param steadyClockTimeNano duration in nanosecond since epoch of std::chrono::steady_clock
-     * @see Engine#getSteadyClockTimeNano
-     * @see Renderer#beginFrame
-     */
-    public void setVsyncTime(long steadyClockTimeNano) {
-        nSetVsyncTime(getNativeObject(), steadyClockTimeNano);
-    }
-
-    /**
-     * Call skipFrame when momentarily skipping frames, for instance if the content of the
-     * scene doesn't change.
-     *
-     * @param vsyncSteadyClockTimeNano The time in nanoseconds when the frame started being rendered,
-     *                       in the {@link System#nanoTime()} timebase. Divide this value by 1000000 to
-     *                       convert it to the {@link android.os.SystemClock#uptimeMillis()}
-     *                       time base. This typically comes from
-     *                       {@link android.view.Choreographer.FrameCallback}.
-     */
-    public void skipFrame(long vsyncSteadyClockTimeNano) {
-        nSkipFrame(getNativeObject(), vsyncSteadyClockTimeNano);
-    }
-
    /**
     * Sets up a frame for this <code>Renderer</code>.
     * <p><code>beginFrame</code> manages frame pacing, and returns whether or not a frame should be
@@ -729,8 +702,6 @@ public class Renderer {
    }

    private static native void nSetPresentationTime(long nativeObject, long monotonicClockNanos);
-    private static native void nSetVsyncTime(long nativeObject, long steadyClockTimeNano);
-    private static native void nSkipFrame(long nativeObject, long vsyncSteadyClockTimeNano);
    private static native boolean nBeginFrame(long nativeRenderer, long nativeSwapChain, long frameTimeNanos);
    private static native void nEndFrame(long nativeRenderer);
    private static native void nRender(long nativeRenderer, long nativeView);
@@ -1233,18 +1233,6 @@ public class View {
        return nGetFogEntity(getNativeObject());
    }

-    /**
-     * When certain temporal features are used (e.g.: TAA or Screen-space reflections), the view
-     * keeps a history of previous frame renders associated with the Renderer the view was last
-     * used with. When switching Renderer, it may be necessary to clear that history by calling
-     * this method. Similarly, if the whole content of the screen change, like when a cut-scene
-     * starts, clearing the history might be needed to avoid artifacts due to the previous frame
-     * being very different.
-     */
-    public void clearFrameHistory(Engine engine) {
-        nClearFrameHistory(getNativeObject(), engine.getNativeObject());
-    }
-
    public long getNativeObject() {
        if (mNativeObject == 0) {
            throw new IllegalStateException("Calling method on destroyed View");
@@ -1306,7 +1294,7 @@ public class View {
    private static native void nSetMaterialGlobal(long nativeView, int index, float x, float y, float z, float w);
    private static native void nGetMaterialGlobal(long nativeView, int index, float[] out);
    private static native int nGetFogEntity(long nativeView);
-    private static native void nClearFrameHistory(long nativeView, long nativeEngine);
+

    /**
     * List of available ambient occlusion techniques.
@@ -125,11 +125,6 @@ extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBu
    builder->groundPlane(a, b, c, d);
 }

-extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBuilderPanning(JNIEnv*, jclass, jlong nativeBuilder, jboolean enabled) {
-    Builder* builder = (Builder*) nativeBuilder;
-    builder->panning(enabled);
-}
-
 extern "C" JNIEXPORT long Java_com_google_android_filament_utils_Manipulator_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder, jint mode) {
    Builder* builder = (Builder*) nativeBuilder;
    return (jlong) builder->build((Mode) mode);
@@ -274,17 +274,6 @@ public class Manipulator {
            return this;
        }

-        /**
-         * Sets whether panning is enabled in the manipulator.
-         *
-         * @return this <code>Builder</code> object for chaining calls
-         */
-        @NonNull
-        public Builder panning(Boolean enabled) {
-            nBuilderPanning(mNativeBuilder, enabled);
-            return this;
-        }
-
        /**
         * Creates and returns the <code>Manipulator</code> object.
         *
@@ -494,7 +483,6 @@ public class Manipulator {
    private static native void nBuilderFlightPanSpeed(long nativeBuilder, float x, float y);
    private static native void nBuilderFlightMoveDamping(long nativeBuilder, float damping);
    private static native void nBuilderGroundPlane(long nativeBuilder, float a, float b, float c, float d);
-    private static native void nBuilderPanning(long nativeBuilder, Boolean enabled);
    private static native long nBuilderBuild(long nativeBuilder, int mode);

    private static native void nDestroyManipulator(long nativeManip);
@@ -120,6 +120,7 @@ set(GLTFIO_INCLUDE_DIRS
        ../../third_party/cgltf
        ../../third_party/meshoptimizer/src
        ../../third_party/robin-map
+        ../../third_party/hat-trie
        ../../third_party/stb
        ../../libs/utils/include
        ../../libs/ktxreader/include
@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.55.1
+VERSION_NAME=1.51.8

 POM_DESCRIPTION=Real-time physically based rendering engine for Android.

@@ -1,6 +1,6 @@
 #Wed Nov 17 10:40:18 PST 2021
 distributionBase=GRADLE_USER_HOME
 distributionPath=wrapper/dists
-distributionUrl=https\://services.gradle.org/distributions/gradle-8.6-bin.zip
+distributionUrl=https\://services.gradle.org/distributions/gradle-8.2-bin.zip
 zipStoreBase=GRADLE_USER_HOME
 zipStorePath=wrapper/dists
@@ -64,12 +64,6 @@ function print_help {
    echo "        enabling debug paths in the backend from the build script. For example, make a"
    echo "        systrace-enabled build without directly changing #defines. Remember to add -f when"
    echo "        changing this option."
-    echo "    -X osmesa_path"
-    echo "        Indicates a path to a completed OSMesa build. OSMesa is used to create an offscreen GL"
-    echo "        context for software rasterization"
-    echo "    -S type"
-    echo "        Enable stereoscopic rendering where type is one of [instanced|multiview]. This is only"
-    echo "        meant for building the samples."
    echo ""
    echo "Build types:"
    echo "    release"
@@ -181,10 +175,6 @@ ASAN_UBSAN_OPTION=""

 BACKEND_DEBUG_FLAG_OPTION=""

-STEREOSCOPIC_OPTION=""
-
-OSMESA_OPTION=""
-
 IOS_BUILD_SIMULATOR=false
 BUILD_UNIVERSAL_LIBRARIES=false

@@ -244,8 +234,6 @@ function build_desktop_target {
            ${MATOPT_OPTION} \
            ${ASAN_UBSAN_OPTION} \
            ${BACKEND_DEBUG_FLAG_OPTION} \
-            ${STEREOSCOPIC_OPTION} \
-            ${OSMESA_OPTION} \
            ${architectures} \
            ../..
        ln -sf "out/cmake-${lc_target}/compile_commands.json" \
@@ -380,7 +368,6 @@ function build_android_target {
            ${MATOPT_OPTION} \
            ${VULKAN_ANDROID_OPTION} \
            ${BACKEND_DEBUG_FLAG_OPTION} \
-            ${STEREOSCOPIC_OPTION} \
            ../..
        ln -sf "out/cmake-android-${lc_target}-${arch}/compile_commands.json" \
           ../../compile_commands.json
@@ -615,7 +602,7 @@ function build_ios_target {
            -DCMAKE_TOOLCHAIN_FILE=../../third_party/clang/iOS.cmake \
            ${MATDBG_OPTION} \
            ${MATOPT_OPTION} \
-            ${STEREOSCOPIC_OPTION} \
+            ${BACKEND_DEBUG_FLAG_OPTION} \
            ../..
        ln -sf "out/cmake-ios-${lc_target}-${arch}/compile_commands.json" \
           ../../compile_commands.json
@@ -802,7 +789,7 @@ function check_debug_release_build {

 pushd "$(dirname "$0")" > /dev/null

-while getopts ":hacCfgijmp:q:uvslwedk:bx:S:X:" opt; do
+while getopts ":hacCfgijmp:q:uvslwedk:bx:" opt; do
    case ${opt} in
        h)
            print_help
@@ -942,22 +929,6 @@ while getopts ":hacCfgijmp:q:uvslwedk:bx:S:X:" opt; do
            ;;
        x)  BACKEND_DEBUG_FLAG_OPTION="-DFILAMENT_BACKEND_DEBUG_FLAG=${OPTARG}"
            ;;
-        S)  case $(echo "${OPTARG}" | tr '[:upper:]' '[:lower:]') in
-                instanced)
-                    STEREOSCOPIC_OPTION="-DFILAMENT_SAMPLES_STEREO_TYPE=instanced"
-                    ;;
-                multiview)
-                    STEREOSCOPIC_OPTION="-DFILAMENT_SAMPLES_STEREO_TYPE=multiview"
-                    ;;
-                *)
-                    echo "Unknown stereoscopic type ${OPTARG}"
-                    echo "Type must be one of [instanced|multiview]"
-                    echo ""
-                    exit 1
-            esac
-            ;;
-        X)  OSMESA_OPTION="-DFILAMENT_OSMESA_PATH=${OPTARG}"
-            ;;
        \?)
            echo "Invalid option: -${OPTARG}" >&2
            echo ""
@@ -57,10 +57,15 @@ FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION:-$(cat `dirname $0`/ndk.version)}

 # Install the required NDK version specifically (if not present)
 if [[ ! -d "${ANDROID_HOME}/ndk/$FILAMENT_NDK_VERSION" ]]; then
-    yes | ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager --licenses
-    ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager "ndk;$FILAMENT_NDK_VERSION"
+    ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager "ndk;$FILAMENT_NDK_VERSION" > /dev/null
 fi
+
+# Only build 1 64 bit target during presubmit to cut down build times during presubmit
+# Continuous builds will build everything
 ANDROID_ABIS=
+if [[ "$TARGET" == "presubmit" ]]; then
+  ANDROID_ABIS="-q arm64-v8a"
+fi

 # Build the Android sample-gltf-viewer APK during release.
 BUILD_SAMPLES=
@@ -68,19 +73,5 @@ if [[ "$TARGET" == "release" ]]; then
    BUILD_SAMPLES="-k sample-gltf-viewer"
 fi

-function build_android() {
-    local ABI=$1
-
-    # Do the following in two steps so that we do not run out of space
-    if [[ -n "${BUILD_DEBUG}" ]]; then
-        FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android -q ${ABI} -c ${BUILD_SAMPLES} ${GENERATE_ARCHIVES} ${BUILD_DEBUG}
-        rm -rf out/cmake-android-debug-*
-    fi
-    if [[ -n "${BUILD_RELEASE}" ]]; then
-        FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android -q ${ABI} -c ${BUILD_SAMPLES} ${GENERATE_ARCHIVES} ${BUILD_RELEASE}
-        rm -rf out/cmake-android-release-*
-    fi
-}
-
 pushd `dirname $0`/../.. > /dev/null
-build_android $2
+FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android $ANDROID_ABIS -c $BUILD_SAMPLES $GENERATE_ARCHIVES $BUILD_DEBUG $BUILD_RELEASE
@@ -1 +1 @@
-27.0.11718014
+26.1.10909125
@@ -1,6 +1,7 @@
 #!/bin/bash
 if [ `uname` == "Linux" ];then
-    source `dirname $0`/../linux/ci-common.sh
+    curl -OL https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-linux.zip
+    unzip -q ninja-linux.zip
 elif [ `uname` == "Darwin" ];then
    curl -OL https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-mac.zip
    unzip -q ninja-mac.zip
@@ -12,6 +13,9 @@ fi
 chmod +x ninja
 export PATH="$PWD:$PATH"

+# FIXME: kokoro machines have node and npm but currently they are symlinked to non-existent files
+# npm install -g typescript
+
 # Install emscripten.
 curl -L https://github.com/emscripten-core/emsdk/archive/refs/tags/3.1.15.zip > emsdk.zip
 unzip emsdk.zip ; mv emsdk-* emsdk ; cd emsdk
@@ -115,23 +115,17 @@ cmake ..\.. ^
    -DFILAMENT_SUPPORTS_VULKAN=ON ^
    || exit /b

-set build_flags=-j %NUMBER_OF_PROCESSORS%
-
-@echo on
-
-:: we've upgraded the windows machines, so the following are no longer accurate as of 09/19/24, but
-:: keeping around the comment for record.
-
 :: Attempt to fix "error C1060: compiler is out of heap space" seen on CI.
 :: Some resource libraries require significant heap space to compile, so first compile them serially.
-:: cmake --build . --target filagui --config %config% %build_flags% || exit /b
-:: cmake --build . --target uberarchive --config %config% %build_flags% || exit /b
-:: cmake --build . --target gltf-demo-resources --config %config% %build_flags% || exit /b
-:: cmake --build . --target filamentapp-resources --config %config% %build_flags% || exit /b
-:: cmake --build . --target sample-resources --config %config% %build_flags% || exit /b
-:: cmake --build . --target suzanne-resources --config %config% %build_flags% || exit /b
+@echo on
+cmake --build . --target filagui --config %config% || exit /b
+cmake --build . --target uberarchive --config %config% || exit /b
+cmake --build . --target gltf-demo-resources --config %config% || exit /b
+cmake --build . --target filamentapp-resources --config %config% || exit /b
+cmake --build . --target sample-resources --config %config% || exit /b
+cmake --build . --target suzanne-resources --config %config% || exit /b

-cmake --build . %INSTALL% --config %config% %build_flags% -- /m || exit /b
+cmake --build . %INSTALL% --config %config% -- /m || exit /b
@echo off

 echo Disk info after building variant: %variant%
@@ -1308,12 +1308,7 @@ Description
      declare a variable called `eyeDirection` you can access it in the fragment shader using
      `variable_eyeDirection`. In the vertex shader, the interpolant name is simply a member of
      the `MaterialVertexInputs` structure (`material.eyeDirection` in your example). Each
-      interpolant is of type `float4` (`vec4`) in the shaders. By default the precision of the
-      interpolant is `highp` in *both* the vertex and fragment shaders.
-      An alternate syntax can be used to specify both the name and precision of the interpolant.
-      In this case the specified precision is used as-is in both fragment and vertex stages, in
-      particular if `default` is specified the default precision is used is the fragment shader
-      (`mediump`) and in the vertex shader (`highp`).
+      interpolant is of type `float4` (`vec4`) in the shaders.

 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ JSON
 material {
@@ -1325,11 +1320,7 @@ material {
        }
    ],
    variables : [
-         eyeDirection,
-         {
-            name : eyeColor,
-            precision : medium
-         }
+         eyeDirection
    ],
    vertexDomain : device,
    depthWrite : false,
@@ -61,12 +61,10 @@ set(SRCS
        src/Engine.cpp
        src/Exposure.cpp
        src/Fence.cpp
-        src/FilamentBuilder.cpp
        src/FrameInfo.cpp
        src/FrameSkipper.cpp
        src/Froxelizer.cpp
        src/Frustum.cpp
-        src/HwDescriptorSetLayoutFactory.cpp
        src/HwRenderPrimitiveFactory.cpp
        src/HwVertexBufferInfoFactory.cpp
        src/IndexBuffer.cpp
@@ -77,6 +75,8 @@ set(SRCS
        src/MaterialInstance.cpp
        src/MaterialParser.cpp
        src/MorphTargetBuffer.cpp
+        src/PerViewUniforms.cpp
+        src/PerShadowMapUniforms.cpp
        src/PostProcessManager.cpp
        src/RenderPass.cpp
        src/RenderPrimitive.cpp
@@ -125,12 +125,6 @@ set(SRCS
        src/details/Texture.cpp
        src/details/VertexBuffer.cpp
        src/details/View.cpp
-        src/ds/ColorPassDescriptorSet.cpp
-        src/ds/DescriptorSet.cpp
-        src/ds/DescriptorSetLayout.cpp
-        src/ds/PostProcessDescriptorSet.cpp
-        src/ds/ShadowMapDescriptorSet.cpp
-        src/ds/SsrPassDescriptorSet.cpp
        src/fg/Blackboard.cpp
        src/fg/DependencyGraph.cpp
        src/fg/FrameGraph.cpp
@@ -154,21 +148,23 @@ set(PRIVATE_HDRS
        src/FrameInfo.h
        src/FrameSkipper.h
        src/Froxelizer.h
-        src/HwDescriptorSetLayoutFactory.h
        src/HwRenderPrimitiveFactory.h
        src/HwVertexBufferInfoFactory.h
        src/Intersections.h
        src/MaterialParser.h
+        src/PerViewUniforms.h
+        src/PerShadowMapUniforms.h
        src/PIDController.h
        src/PostProcessManager.h
+        src/RendererUtils.h
        src/RenderPass.h
        src/RenderPrimitive.h
-        src/RendererUtils.h
        src/ResourceAllocator.h
        src/ResourceList.h
        src/ShadowMap.h
        src/ShadowMapManager.h
        src/SharedHandle.h
+        src/TypedUniformBuffer.h
        src/UniformBuffer.h
        src/components/CameraManager.h
        src/components/LightManager.h
@@ -196,14 +192,6 @@ set(PRIVATE_HDRS
        src/details/Texture.h
        src/details/VertexBuffer.h
        src/details/View.h
-        src/downcast.h
-        src/ds/ColorPassDescriptorSet.h
-        src/ds/DescriptorSetLayout.h
-        src/ds/PostProcessDescriptorSet.h
-        src/ds/ShadowMapDescriptorSet.h
-        src/ds/SsrPassDescriptorSet.h
-        src/ds/TypedBuffer.h
-        src/ds/TypedUniformBuffer.h
        src/fg/Blackboard.h
        src/fg/FrameGraph.h
        src/fg/FrameGraphId.h
@@ -221,6 +209,7 @@ set(PRIVATE_HDRS
        src/materials/fsr/ffx_a.h
        src/materials/fsr/ffx_fsr1.h
        src/materials/fsr/ffx_fsr1_mobile.fs
+        src/downcast.h
 )

 set(MATERIAL_SRCS
@@ -12,7 +12,6 @@ set(PUBLIC_HDRS
        include/backend/AcquiredImage.h
        include/backend/BufferDescriptor.h
        include/backend/CallbackHandler.h
-        include/backend/DescriptorSetOffsetArray.h
        include/backend/DriverApiForward.h
        include/backend/DriverEnums.h
        include/backend/Handle.h
@@ -70,13 +69,9 @@ set(PRIVATE_HDRS
 if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3)
    list(APPEND SRCS
            include/backend/platforms/OpenGLPlatform.h
-            src/opengl/BindingMap.h
            src/opengl/gl_headers.cpp
            src/opengl/gl_headers.h
            src/opengl/GLBufferObject.h
-            src/opengl/GLDescriptorSet.cpp
-            src/opengl/GLDescriptorSet.h
-            src/opengl/GLDescriptorSetLayout.h
            src/opengl/GLTexture.h
            src/opengl/GLUtils.cpp
            src/opengl/GLUtils.h
@@ -119,8 +114,6 @@ if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3)
            list(APPEND SRCS src/opengl/platforms/PlatformGLX.cpp)
        elseif (FILAMENT_SUPPORTS_EGL_ON_LINUX)
            list(APPEND SRCS src/opengl/platforms/PlatformEGLHeadless.cpp)
-        elseif (FILAMENT_SUPPORTS_OSMESA)
-            list(APPEND SRCS src/opengl/platforms/PlatformOSMesa.cpp)
        endif()
    elseif (WIN32)
        list(APPEND SRCS src/opengl/platforms/PlatformWGL.cpp)
@@ -368,15 +361,6 @@ set(LINUX_LINKER_OPTIMIZATION_FLAGS
        -Wl,--exclude-libs,bluegl
 )

-if (LINUX AND FILAMENT_SUPPORTS_OSMESA)
-    set(OSMESA_COMPILE_FLAGS
-        -I${FILAMENT_OSMESA_PATH}/include/GL)
-    set(OSMESA_LINKER_FLAGS
-        -Wl,-L${FILAMENT_OSMESA_PATH}/lib/x86_64-linux-gnu/
-        -lOSMesa
-    )
-endif()
-
 if (MSVC)
    set(FILAMENT_WARNINGS /W3)
 else()
@@ -397,7 +381,6 @@ endif()

 target_compile_options(${TARGET} PRIVATE
        ${FILAMENT_WARNINGS}
-        ${OSMESA_COMPILE_FLAGS}
        $<$<CONFIG:Release>:${OPTIMIZATION_FLAGS}>
        $<$<AND:$<PLATFORM_ID:Darwin>,$<CONFIG:Release>>:${DARWIN_OPTIMIZATION_FLAGS}>
 )
@@ -407,7 +390,6 @@ if (FILAMENT_SUPPORTS_METAL)
 endif()

 target_link_libraries(${TARGET} PRIVATE
-        ${OSMESA_LINKER_FLAGS}
        $<$<AND:$<PLATFORM_ID:Linux>,$<CONFIG:Release>>:${LINUX_LINKER_OPTIMIZATION_FLAGS}>
 )

@@ -437,12 +419,10 @@ if (APPLE OR LINUX)
        test/test_BufferUpdates.cpp
        test/test_Callbacks.cpp
        test/test_MRT.cpp
-        test/test_PushConstants.cpp
        test/test_LoadImage.cpp
        test/test_StencilBuffer.cpp
        test/test_Scissor.cpp
        test/test_MipLevels.cpp
-        test/test_Handles.cpp
    )
    set(BACKEND_TEST_LIBS
        backend
@@ -514,40 +494,21 @@ endif()

 # ==================================================================================================
 # Compute tests
-#
-#if (NOT IOS AND NOT WEBGL)
-#
-#add_executable(compute_test
-#        test/ComputeTest.cpp
-#        test/Arguments.cpp
-#        test/test_ComputeBasic.cpp
-#        )
-#
-#target_link_libraries(compute_test PRIVATE
-#        backend
-#        getopt
-#        gtest
-#        )
-#
-#set_target_properties(compute_test PROPERTIES FOLDER Tests)
-#
-#endif()

-# ==================================================================================================
-# Metal utils tests
+if (NOT IOS AND NOT WEBGL)

-if (APPLE AND NOT IOS)
+add_executable(compute_test
+        test/ComputeTest.cpp
+        test/Arguments.cpp
+        test/test_ComputeBasic.cpp
+        )

-add_executable(metal_utils_test test/MetalTest.mm)
-
-target_compile_options(metal_utils_test PRIVATE "-fobjc-arc")
-
-target_link_libraries(metal_utils_test PRIVATE
+target_link_libraries(compute_test PRIVATE
        backend
        getopt
        gtest
        )

-set_target_properties(metal_utils_test PROPERTIES FOLDER Tests)
+set_target_properties(compute_test PROPERTIES FOLDER Tests)

 endif()
@@ -1,101 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H
-#define TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H
-
-#include <backend/DriverApiForward.h>
-
-#include <initializer_list>
-#include <memory>
-
-#include <stddef.h>
-#include <stdint.h>
-
-
-namespace filament::backend {
-
-void* allocateFromCommandStream(DriverApi& driver, size_t size, size_t alignment) noexcept;
-
-class DescriptorSetOffsetArray {
-public:
-    using value_type = uint32_t;
-    using reference = value_type&;
-    using const_reference = value_type const&;
-    using size_type = uint32_t;
-    using difference_type = int32_t;
-    using pointer = value_type*;
-    using const_pointer = value_type const*;
-    using iterator = pointer;
-    using const_iterator = const_pointer;
-
-    DescriptorSetOffsetArray() noexcept = default;
-
-    ~DescriptorSetOffsetArray() noexcept = default;
-
-    DescriptorSetOffsetArray(size_type size, DriverApi& driver) noexcept {
-        mOffsets = (value_type *)allocateFromCommandStream(driver,
-                size * sizeof(value_type), alignof(value_type));
-        std::uninitialized_fill_n(mOffsets, size, 0);
-    }
-
-    DescriptorSetOffsetArray(std::initializer_list<uint32_t> list, DriverApi& driver) noexcept {
-        mOffsets = (value_type *)allocateFromCommandStream(driver,
-                list.size() * sizeof(value_type), alignof(value_type));
-        std::uninitialized_copy(list.begin(), list.end(), mOffsets);
-    }
-
-    DescriptorSetOffsetArray(DescriptorSetOffsetArray const&) = delete;
-    DescriptorSetOffsetArray& operator=(DescriptorSetOffsetArray const&) = delete;
-
-    DescriptorSetOffsetArray(DescriptorSetOffsetArray&& rhs) noexcept
-            : mOffsets(rhs.mOffsets) {
-        rhs.mOffsets = nullptr;
-    }
-
-    DescriptorSetOffsetArray& operator=(DescriptorSetOffsetArray&& rhs) noexcept {
-        if (this != &rhs) {
-            mOffsets = rhs.mOffsets;
-            rhs.mOffsets = nullptr;
-        }
-        return *this;
-    }
-
-    bool empty() const noexcept { return mOffsets == nullptr; }
-
-    value_type* data() noexcept { return mOffsets; }
-    const value_type* data() const noexcept { return mOffsets; }
-
-
-    reference operator[](size_type n) noexcept {
-        return *(data() + n);
-    }
-
-    const_reference operator[](size_type n) const noexcept {
-        return *(data() + n);
-    }
-
-    void clear() noexcept {
-        mOffsets = nullptr;
-    }
-
-private:
-    value_type *mOffsets = nullptr;
-};
-
-} // namespace filament::backend
-
-#endif //TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H
@@ -19,16 +19,12 @@
 #ifndef TNT_FILAMENT_BACKEND_DRIVERENUMS_H
 #define TNT_FILAMENT_BACKEND_DRIVERENUMS_H

+#include <utils/BitmaskEnum.h>
 #include <utils/unwindows.h> // Because we define ERROR in the FenceStatus enum.

-#include <backend/Platform.h>
 #include <backend/PresentCallable.h>

-#include <utils/BitmaskEnum.h>
-#include <utils/FixedCapacityVector.h>
 #include <utils/Invocable.h>
-#include <utils/compiler.h>
-#include <utils/debug.h>
 #include <utils/ostream.h>

 #include <math/vec4.h>
@@ -100,8 +96,6 @@ static constexpr size_t MAX_VERTEX_ATTRIBUTE_COUNT  = 16;   // This is guarantee
 static constexpr size_t MAX_SAMPLER_COUNT           = 62;   // Maximum needed at feature level 3.
 static constexpr size_t MAX_VERTEX_BUFFER_COUNT     = 16;   // Max number of bound buffer objects.
 static constexpr size_t MAX_SSBO_COUNT              = 4;    // This is guaranteed by OpenGL ES.
-static constexpr size_t MAX_DESCRIPTOR_SET_COUNT    = 4;    // This is guaranteed by Vulkan.
-static constexpr size_t MAX_DESCRIPTOR_COUNT        = 64;   // per set

 static constexpr size_t MAX_PUSH_CONSTANT_COUNT     = 32;   // Vulkan 1.1 spec allows for 128-byte
                                                            // of push constant (we assume 4-byte
@@ -123,7 +117,7 @@ static_assert(MAX_VERTEX_BUFFER_COUNT <= MAX_VERTEX_ATTRIBUTE_COUNT,
        "The number of buffer objects that can be attached to a VertexBuffer must be "
        "less than or equal to the maximum number of vertex attributes.");

-static constexpr size_t CONFIG_UNIFORM_BINDING_COUNT = 9;   // This is guaranteed by OpenGL ES.
+static constexpr size_t CONFIG_UNIFORM_BINDING_COUNT = 10;  // This is guaranteed by OpenGL ES.
 static constexpr size_t CONFIG_SAMPLER_BINDING_COUNT = 4;   // This is guaranteed by OpenGL ES.

 /**
@@ -196,70 +190,6 @@ static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguag
    }
 }

-enum class ShaderStage : uint8_t {
-    VERTEX = 0,
-    FRAGMENT = 1,
-    COMPUTE = 2
-};
-
-static constexpr size_t PIPELINE_STAGE_COUNT = 2;
-enum class ShaderStageFlags : uint8_t {
-    NONE        =    0,
-    VERTEX      =    0x1,
-    FRAGMENT    =    0x2,
-    COMPUTE     =    0x4,
-    ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
-};
-
-static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
-    switch (type) {
-        case ShaderStage::VERTEX:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
-        case ShaderStage::FRAGMENT:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::FRAGMENT));
-        case ShaderStage::COMPUTE:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::COMPUTE));
-    }
-}
-
-enum class DescriptorType : uint8_t {
-    UNIFORM_BUFFER,
-    SHADER_STORAGE_BUFFER,
-    SAMPLER,
-    INPUT_ATTACHMENT,
-};
-
-enum class DescriptorFlags : uint8_t {
-    NONE = 0x00,
-    DYNAMIC_OFFSET = 0x01
-};
-
-using descriptor_set_t = uint8_t;
-
-using descriptor_binding_t = uint8_t;
-
-struct DescriptorSetLayoutBinding {
-    DescriptorType type;
-    ShaderStageFlags stageFlags;
-    descriptor_binding_t binding;
-    DescriptorFlags flags = DescriptorFlags::NONE;
-    uint16_t count = 0;
-
-    friend inline bool operator==(
-            DescriptorSetLayoutBinding const& lhs,
-            DescriptorSetLayoutBinding const& rhs) noexcept {
-        return lhs.type == rhs.type &&
-               lhs.flags == rhs.flags &&
-               lhs.count == rhs.count &&
-               lhs.stageFlags == rhs.stageFlags;
-    }
-};
-
-struct DescriptorSetLayout {
-    utils::FixedCapacityVector<DescriptorSetLayoutBinding> bindings;
-};
-
-
 /**
 * Bitmask for selecting render buffers
 */
@@ -339,6 +269,15 @@ enum class FenceStatus : int8_t {
    TIMEOUT_EXPIRED = 1,        //!< wait()'s timeout expired. The Fence condition is not satisfied.
 };

+/**
+ * Status codes for sync objects
+ */
+enum class SyncStatus : int8_t {
+    ERROR = -1,          //!< An error occurred. The Sync is not signaled.
+    SIGNALED = 0,        //!< The Sync is signaled.
+    NOT_SIGNALED = 1,    //!< The Sync is not signaled yet
+};
+
 static constexpr uint64_t FENCE_WAIT_FOR_EVER = uint64_t(-1);

 /**
@@ -428,18 +367,6 @@ enum class SamplerType : uint8_t {
    SAMPLER_CUBEMAP_ARRAY,  //!< Cube map array texture (feature level 2)
 };

-inline const char* stringify(SamplerType samplerType) {
-    switch (samplerType) {
-        case SamplerType::SAMPLER_2D: return "SAMPLER_2D";
-        case SamplerType::SAMPLER_2D_ARRAY: return "SAMPLER_2D_ARRAY";
-        case SamplerType::SAMPLER_CUBEMAP: return "SAMPLER_CUBEMAP";
-        case SamplerType::SAMPLER_EXTERNAL: return "SAMPLER_EXTERNAL";
-        case SamplerType::SAMPLER_3D: return "SAMPLER_3D";
-        case SamplerType::SAMPLER_CUBEMAP_ARRAY: return "SAMPLER_CUBEMAP_ARRAY";
-    }
-    return "UNKNOWN";
-}
-
 //! Subpass type
 enum class SubpassType : uint8_t {
    SUBPASS_INPUT
@@ -765,27 +692,9 @@ enum class TextureUsage : uint16_t {
    BLIT_SRC            = 0x0040,            //!< Texture can be used the source of a blit()
    BLIT_DST            = 0x0080,            //!< Texture can be used the destination of a blit()
    PROTECTED           = 0x0100,            //!< Texture can be used for protected content
-    DEFAULT             = UPLOADABLE | SAMPLEABLE,   //!< Default texture usage
-    ALL_ATTACHMENTS     = COLOR_ATTACHMENT | DEPTH_ATTACHMENT | STENCIL_ATTACHMENT | SUBPASS_INPUT,   //!< Mask of all attachments
+    DEFAULT             = UPLOADABLE | SAMPLEABLE   //!< Default texture usage
 };

-inline const char* stringify(TextureUsage usage) {
-    switch (usage) {
-        case TextureUsage::NONE: return "NONE";
-        case TextureUsage::COLOR_ATTACHMENT: return "COLOR_ATTACHMENT";
-        case TextureUsage::DEPTH_ATTACHMENT: return "DEPTH_ATTACHMENT";
-        case TextureUsage::STENCIL_ATTACHMENT: return "STENCIL_ATTACHMENT";
-        case TextureUsage::UPLOADABLE: return "UPLOADABLE";
-        case TextureUsage::SAMPLEABLE: return "SAMPLEABLE";
-        case TextureUsage::SUBPASS_INPUT: return "SUBPASS_INPUT";
-        case TextureUsage::BLIT_SRC: return "BLIT_SRC";
-        case TextureUsage::BLIT_DST: return "BLIT_DST";
-        case TextureUsage::PROTECTED: return "PROTECTED";
-        case TextureUsage::DEFAULT: return "DEFAULT";
-        default: return "UNKNOWN";
-    }
-}
-
 //! Texture swizzle
 enum class TextureSwizzle : uint8_t {
    SUBSTITUTE_ZERO,
@@ -977,9 +886,6 @@ struct SamplerParams { // NOLINT

    struct EqualTo {
        bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
-            assert_invariant(lhs.padding0 == 0);
-            assert_invariant(lhs.padding1 == 0);
-            assert_invariant(lhs.padding2 == 0);
            auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
            auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
            return *pLhs == *pRhs;
@@ -988,9 +894,6 @@ struct SamplerParams { // NOLINT

    struct LessThan {
        bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
-            assert_invariant(lhs.padding0 == 0);
-            assert_invariant(lhs.padding1 == 0);
-            assert_invariant(lhs.padding2 == 0);
            auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
            auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
            return *pLhs == *pRhs;
@@ -998,12 +901,6 @@ struct SamplerParams { // NOLINT
    };

 private:
-    friend inline bool operator == (SamplerParams lhs, SamplerParams rhs) noexcept {
-        return SamplerParams::EqualTo{}(lhs, rhs);
-    }
-    friend inline bool operator != (SamplerParams lhs, SamplerParams rhs) noexcept {
-        return  !SamplerParams::EqualTo{}(lhs, rhs);
-    }
    friend inline bool operator < (SamplerParams lhs, SamplerParams rhs) noexcept {
        return SamplerParams::LessThan{}(lhs, rhs);
    }
@@ -1160,7 +1057,7 @@ struct RasterState {
            bool inverseFrontFaces                      : 1;        // 31

            //! padding, must be 0
-            bool depthClamp                             : 1;        // 32
+            uint8_t padding                             : 1;        // 32
        };
        uint32_t u = 0;
    };
@@ -1171,6 +1068,32 @@ struct RasterState {
 * \privatesection
 */

+enum class ShaderStage : uint8_t {
+    VERTEX = 0,
+    FRAGMENT = 1,
+    COMPUTE = 2
+};
+
+static constexpr size_t PIPELINE_STAGE_COUNT = 2;
+enum class ShaderStageFlags : uint8_t {
+    NONE        =    0,
+    VERTEX      =    0x1,
+    FRAGMENT    =    0x2,
+    COMPUTE     =    0x4,
+    ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
+};
+
+static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
+    switch (type) {
+        case ShaderStage::VERTEX:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
+        case ShaderStage::FRAGMENT:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::FRAGMENT));
+        case ShaderStage::COMPUTE:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::COMPUTE));
+    }
+}
+
 /**
 * Selects which buffers to clear at the beginning of the render pass, as well as which buffers
 * can be discarded at the beginning and end of the render pass.
@@ -1320,14 +1243,20 @@ enum class Workaround : uint16_t {
    ADRENO_UNIFORM_ARRAY_CRASH,
    // Workaround a Metal pipeline compilation error with the message:
    // "Could not statically determine the target of a texture". See light_indirect.fs
-    METAL_STATIC_TEXTURE_TARGET_ERROR,
+    A8X_STATIC_TEXTURE_TARGET_ERROR,
    // Adreno drivers sometimes aren't able to blit into a layer of a texture array.
    DISABLE_BLIT_INTO_TEXTURE_ARRAY,
    // Multiple workarounds needed for PowerVR GPUs
    POWER_VR_SHADER_WORKAROUNDS,
 };

-using StereoscopicType = backend::Platform::StereoscopicType;
+//! The type of technique for stereoscopic rendering
+enum class StereoscopicType : uint8_t {
+    // Stereoscopic rendering is performed using instanced rendering technique.
+    INSTANCED,
+    // Stereoscopic rendering is performed using the multiview feature from the graphics backend.
+    MULTIVIEW,
+};

 } // namespace filament::backend

@@ -1335,8 +1264,6 @@ template<> struct utils::EnableBitMaskOperators<filament::backend::ShaderStageFl
        : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::TargetBufferFlags>
        : public std::true_type {};
-template<> struct utils::EnableBitMaskOperators<filament::backend::DescriptorFlags>
-        : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::TextureUsage>
        : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::StencilFace>
@@ -23,7 +23,6 @@
 #include <utils/debug.h>

 #include <type_traits> // FIXME: STL headers are not allowed in public headers
-#include <utility>

 #include <stdint.h>

@@ -42,8 +41,6 @@ struct HwTexture;
 struct HwTimerQuery;
 struct HwVertexBufferInfo;
 struct HwVertexBuffer;
-struct HwDescriptorSetLayout;
-struct HwDescriptorSet;

 /*
 * A handle to a backend resource. HandleBase is for internal use only.
@@ -107,18 +104,8 @@ struct Handle : public HandleBase {
    Handle(Handle const& rhs) noexcept = default;
    Handle(Handle&& rhs) noexcept = default;

-    // Explicitly redefine copy/move assignment operators rather than just using default here.
-    // Because it doesn't make a call to the parent's method automatically during the std::move
-    // function call(https://en.cppreference.com/w/cpp/algorithm/move) in certain compilers like
-    // NDK 25.1.8937393 and below (see b/371980551)
-    Handle& operator=(Handle const& rhs) noexcept {
-        HandleBase::operator=(rhs);
-        return *this;
-    }
-    Handle& operator=(Handle&& rhs) noexcept {
-        HandleBase::operator=(std::move(rhs));
-        return *this;
-    }
+    Handle& operator=(Handle const& rhs) noexcept = default;
+    Handle& operator=(Handle&& rhs) noexcept = default;

    explicit Handle(HandleId id) noexcept : HandleBase(id) { }

@@ -143,21 +130,19 @@ private:

 // Types used by the command stream
 // (we use this renaming because the macro-system doesn't deal well with "<" and ">")
-using BufferObjectHandle        = Handle<HwBufferObject>;
-using FenceHandle               = Handle<HwFence>;
-using IndexBufferHandle         = Handle<HwIndexBuffer>;
-using ProgramHandle             = Handle<HwProgram>;
-using RenderPrimitiveHandle     = Handle<HwRenderPrimitive>;
-using RenderTargetHandle        = Handle<HwRenderTarget>;
-using SamplerGroupHandle        = Handle<HwSamplerGroup>;
-using StreamHandle              = Handle<HwStream>;
-using SwapChainHandle           = Handle<HwSwapChain>;
-using TextureHandle             = Handle<HwTexture>;
-using TimerQueryHandle          = Handle<HwTimerQuery>;
-using VertexBufferHandle        = Handle<HwVertexBuffer>;
-using VertexBufferInfoHandle    = Handle<HwVertexBufferInfo>;
-using DescriptorSetLayoutHandle = Handle<HwDescriptorSetLayout>;
-using DescriptorSetHandle       = Handle<HwDescriptorSet>;
+using BufferObjectHandle     = Handle<HwBufferObject>;
+using FenceHandle            = Handle<HwFence>;
+using IndexBufferHandle      = Handle<HwIndexBuffer>;
+using ProgramHandle          = Handle<HwProgram>;
+using RenderPrimitiveHandle  = Handle<HwRenderPrimitive>;
+using RenderTargetHandle     = Handle<HwRenderTarget>;
+using SamplerGroupHandle     = Handle<HwSamplerGroup>;
+using StreamHandle           = Handle<HwStream>;
+using SwapChainHandle        = Handle<HwSwapChain>;
+using TextureHandle          = Handle<HwTexture>;
+using TimerQueryHandle       = Handle<HwTimerQuery>;
+using VertexBufferHandle     = Handle<HwVertexBuffer>;
+using VertexBufferInfoHandle = Handle<HwVertexBufferInfo>;

 } // namespace filament::backend

@@ -22,23 +22,15 @@

 #include <utils/ostream.h>

-#include <array>
-
 #include <stdint.h>

 namespace filament::backend {

 //! \privatesection

-struct PipelineLayout {
-    using SetLayout = std::array<Handle<HwDescriptorSetLayout>, MAX_DESCRIPTOR_SET_COUNT>;
-    SetLayout setLayout;      // 16
-};
-
 struct PipelineState {
    Handle<HwProgram> program;                                              //  4
    Handle<HwVertexBufferInfo> vertexBufferInfo;                            //  4
-    PipelineLayout pipelineLayout;                                          // 16
    RasterState rasterState;                                                //  4
    StencilState stencilState;                                              // 12
    PolygonOffset polygonOffset;                                            //  8
@@ -41,26 +41,6 @@ public:
    struct Fence {};
    struct Stream {};

-    /**
-     * The type of technique for stereoscopic rendering. (Note that the materials used will need to
-     * be compatible with the chosen technique.)
-     */
-    enum class StereoscopicType : uint8_t {
-        /**
-         * No stereoscopic rendering
-         */
-        NONE,
-        /**
-         * Stereoscopic rendering is performed using instanced rendering technique.
-         */
-        INSTANCED,
-        /**
-         * Stereoscopic rendering is performed using the multiview feature from the graphics
-         * backend.
-         */
-        MULTIVIEW,
-    };
-
    struct DriverConfig {
        /**
         * Size of handle arena in bytes. Setting to 0 indicates default value is to be used.
@@ -68,7 +48,12 @@ public:
         */
        size_t handleArenaSize = 0;

-        size_t metalUploadBufferSizeBytes = 512 * 1024;
+        /**
+         * This number of most-recently destroyed textures will be tracked for use-after-free.
+         * Throws an exception when a texture is freed but still bound to a SamplerGroup and used in
+         * a draw call. 0 disables completely. Currently only respected by the Metal backend.
+         */
+        size_t textureUseAfterFreePoolSize = 0;

        /**
         * Set to `true` to forcibly disable parallel shader compilation in the backend.
@@ -86,18 +71,6 @@ public:
         * GLES 3.x backends.
         */
        bool forceGLES2Context = false;
-
-        /**
-         * Sets the technique for stereoscopic rendering.
-         */
-        StereoscopicType stereoscopicType = StereoscopicType::NONE;
-
-        /**
-         * Assert the native window associated to a SwapChain is valid when calling makeCurrent().
-         * This is only supported for:
-         *      - PlatformEGLAndroid
-         */
-        bool assertNativeWindowIsValid = false;
    };

    Platform() noexcept;
@@ -24,11 +24,9 @@

 #include <backend/DriverEnums.h>

-#include <array>
-#include <unordered_map>
-#include <tuple>
-#include <utility>
-#include <variant>
+#include <array>    // FIXME: STL headers are not allowed in public headers
+#include <utility>  // FIXME: STL headers are not allowed in public headers
+#include <variant>  // FIXME: STL headers are not allowed in public headers

 #include <stddef.h>
 #include <stdint.h>
@@ -42,36 +40,29 @@ public:
    static constexpr size_t UNIFORM_BINDING_COUNT = CONFIG_UNIFORM_BINDING_COUNT;
    static constexpr size_t SAMPLER_BINDING_COUNT = CONFIG_SAMPLER_BINDING_COUNT;

-    struct Descriptor {
-        utils::CString name;
-        backend::DescriptorType type;
-        backend::descriptor_binding_t binding;
+    struct Sampler {
+        utils::CString name = {};   // name of the sampler in the shader
+        uint32_t binding = 0;       // binding point of the sampler in the shader
    };

-    struct SpecializationConstant {
-        using Type = std::variant<int32_t, float, bool>;
-        uint32_t id;    // id set in glsl
-        Type value;     // value and type
+    struct SamplerGroupData {
+        utils::FixedCapacityVector<Sampler> samplers;
+        ShaderStageFlags stageFlags = ShaderStageFlags::ALL_SHADER_STAGE_FLAGS;
    };

-    struct Uniform { // For ES2 support
+    struct Uniform {
        utils::CString name;    // full qualified name of the uniform field
        uint16_t offset;        // offset in 'uint32_t' into the uniform buffer
        uint8_t size;           // >1 for arrays
        UniformType type;       // uniform type
    };

-    using DescriptorBindingsInfo = utils::FixedCapacityVector<Descriptor>;
-    using DescriptorSetInfo = std::array<DescriptorBindingsInfo, MAX_DESCRIPTOR_SET_COUNT>;
-    using SpecializationConstantsInfo = utils::FixedCapacityVector<SpecializationConstant>;
+    using UniformBlockInfo = std::array<utils::CString, UNIFORM_BINDING_COUNT>;
+    using UniformInfo = utils::FixedCapacityVector<Uniform>;
+    using SamplerGroupInfo = std::array<SamplerGroupData, SAMPLER_BINDING_COUNT>;
    using ShaderBlob = utils::FixedCapacityVector<uint8_t>;
    using ShaderSource = std::array<ShaderBlob, SHADER_TYPE_COUNT>;

-    using AttributesInfo = utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>>;
-    using UniformInfo = utils::FixedCapacityVector<Uniform>;
-    using BindingUniformsInfo = utils::FixedCapacityVector<
-            std::tuple<uint8_t, utils::CString, Program::UniformInfo>>;
-
    Program() noexcept;

    Program(const Program& rhs) = delete;
@@ -88,19 +79,43 @@ public:
    Program& diagnostics(utils::CString const& name,
            utils::Invocable<utils::io::ostream&(utils::io::ostream& out)>&& logger);

-    // Sets one of the program's shader (e.g. vertex, fragment)
+    // sets one of the program's shader (e.g. vertex, fragment)
    // string-based shaders are null terminated, consequently the size parameter must include the
    // null terminating character.
    Program& shader(ShaderStage shader, void const* data, size_t size);

-    // Sets the language of the shader sources provided with shader() (defaults to ESSL3)
+    // sets the language of the shader sources provided with shader() (defaults to ESSL3)
    Program& shaderLanguage(ShaderLanguage shaderLanguage);

-    // Descriptor binding (set, binding, type -> shader name) info
-    Program& descriptorBindings(backend::descriptor_set_t set,
-            DescriptorBindingsInfo descriptorBindings) noexcept;
+    // Note: This is only needed for GLES3.0 backends, because the layout(binding=) syntax is
+    //       not permitted in glsl. The backend needs a way to associate a uniform block
+    //       to a binding point.
+    Program& uniformBlockBindings(
+            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept;

-    Program& specializationConstants(SpecializationConstantsInfo specConstants) noexcept;
+    // Note: This is only needed for GLES2.0, this is used to emulate UBO. This function tells
+    // the program everything it needs to know about the uniforms at a given binding
+    Program& uniforms(uint32_t index, UniformInfo const& uniforms) noexcept;
+
+    // Note: This is only needed for GLES2.0.
+    Program& attributes(
+            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes) noexcept;
+
+    // sets the 'bindingPoint' sampler group descriptor for this program.
+    // 'samplers' can be destroyed after this call.
+    // This effectively associates a set of (BindingPoints, index) to a texture unit in the shader.
+    // Or more precisely, what layout(binding=) is set to in GLSL.
+    Program& setSamplerGroup(size_t bindingPoint, ShaderStageFlags stageFlags,
+            Sampler const* samplers, size_t count) noexcept;
+
+    struct SpecializationConstant {
+        using Type = std::variant<int32_t, float, bool>;
+        uint32_t id;                                // id set in glsl
+        Type value;                                 // value and type
+    };
+
+    Program& specializationConstants(
+            utils::FixedCapacityVector<SpecializationConstant> specConstants) noexcept;

    struct PushConstant {
        utils::CString name;
@@ -114,40 +129,33 @@ public:

    Program& multiview(bool multiview) noexcept;

-    // For ES2 support only...
-    Program& uniforms(uint32_t index, utils::CString name, UniformInfo uniforms) noexcept;
-    Program& attributes(AttributesInfo attributes) noexcept;
-
-    //
-    // Getters for program construction...
-    //
-
    ShaderSource const& getShadersSource() const noexcept { return mShadersSource; }
    ShaderSource& getShadersSource() noexcept { return mShadersSource; }

+    UniformBlockInfo const& getUniformBlockBindings() const noexcept { return mUniformBlocks; }
+    UniformBlockInfo& getUniformBlockBindings() noexcept { return mUniformBlocks; }
+
+    SamplerGroupInfo const& getSamplerGroupInfo() const { return mSamplerGroups; }
+    SamplerGroupInfo& getSamplerGroupInfo() { return mSamplerGroups; }
+
+    auto const& getBindingUniformInfo() const { return mBindingUniformInfo; }
+    auto& getBindingUniformInfo() { return mBindingUniformInfo; }
+
+    auto const& getAttributes() const { return mAttributes; }
+    auto& getAttributes() { return mAttributes; }
+
    utils::CString const& getName() const noexcept { return mName; }
    utils::CString& getName() noexcept { return mName; }

    auto const& getShaderLanguage() const { return mShaderLanguage; }

-    uint64_t getCacheId() const noexcept { return mCacheId; }
-
-    bool isMultiview() const noexcept { return mMultiview; }
-
-    CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
-
-    SpecializationConstantsInfo const& getSpecializationConstants() const noexcept {
+    utils::FixedCapacityVector<SpecializationConstant> const& getSpecializationConstants() const noexcept {
        return mSpecializationConstants;
    }
-
-    SpecializationConstantsInfo& getSpecializationConstants() noexcept {
+    utils::FixedCapacityVector<SpecializationConstant>& getSpecializationConstants() noexcept {
        return mSpecializationConstants;
    }

-    DescriptorSetInfo& getDescriptorBindings() noexcept {
-        return mDescriptorBindings;
-    }
-
    utils::FixedCapacityVector<PushConstant> const& getPushConstants(
            ShaderStage stage) const noexcept {
        return mPushConstants[static_cast<uint8_t>(stage)];
@@ -157,29 +165,27 @@ public:
        return mPushConstants[static_cast<uint8_t>(stage)];
    }

-    auto const& getBindingUniformInfo() const { return mBindingUniformsInfo; }
-    auto& getBindingUniformInfo() { return mBindingUniformsInfo; }
+    uint64_t getCacheId() const noexcept { return mCacheId; }

-    auto const& getAttributes() const { return mAttributes; }
-    auto& getAttributes() { return mAttributes; }
+    bool isMultiview() const noexcept { return mMultiview; }
+
+    CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }

 private:
    friend utils::io::ostream& operator<<(utils::io::ostream& out, const Program& builder);

+    UniformBlockInfo mUniformBlocks = {};
+    SamplerGroupInfo mSamplerGroups = {};
    ShaderSource mShadersSource;
    ShaderLanguage mShaderLanguage = ShaderLanguage::ESSL3;
    utils::CString mName;
    uint64_t mCacheId{};
-    CompilerPriorityQueue mPriorityQueue = CompilerPriorityQueue::HIGH;
    utils::Invocable<utils::io::ostream&(utils::io::ostream& out)> mLogger;
-    SpecializationConstantsInfo mSpecializationConstants;
+    utils::FixedCapacityVector<SpecializationConstant> mSpecializationConstants;
    std::array<utils::FixedCapacityVector<PushConstant>, SHADER_TYPE_COUNT> mPushConstants;
-    DescriptorSetInfo mDescriptorBindings;
-
-    // For ES2 support only
-    AttributesInfo mAttributes;
-    BindingUniformsInfo mBindingUniformsInfo;
-
+    utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> mAttributes;
+    std::array<UniformInfo, Program::UNIFORM_BINDING_COUNT> mBindingUniformInfo;
+    CompilerPriorityQueue mPriorityQueue = CompilerPriorityQueue::HIGH;
    // Indicates the current engine was initialized with multiview stereo, and the variant for this
    // program contains STE flag. This will be referred later for the OpenGL shader compiler to
    // determine whether shader code replacement for the num_views should be performed.
@@ -30,6 +30,10 @@ namespace filament::backend {

 struct TargetBufferInfo {
    // note: the parameters of this constructor are not in the order of this structure's fields
+    TargetBufferInfo(Handle<HwTexture> handle, uint8_t level, uint16_t layer, uint8_t baseViewIndex) noexcept
+        : handle(handle), baseViewIndex(baseViewIndex), level(level), layer(layer) {
+    }
+
    TargetBufferInfo(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
            : handle(handle), level(level), layer(layer) {
    }
@@ -47,15 +51,14 @@ struct TargetBufferInfo {
    // texture to be used as render target
    Handle<HwTexture> handle;

+    // Starting layer index for multiview. This value is only used when the `layerCount` for the
+    // render target is greater than 1.
+    uint8_t baseViewIndex = 0;
+
    // level to be used
    uint8_t level = 0;

-    // - For cubemap textures, this indicates the face of the cubemap. See TextureCubemapFace for
-    //   the face->layer mapping)
-    // - For 2d array, cubemap array, and 3d textures, this indicates an index of a single layer of
-    //   them.
-    // - For multiview textures (i.e., layerCount for the RenderTarget is greater than 1), this
-    //   indicates a starting layer index of the current 2d array texture for multiview.
+    // For cubemaps and 3D textures. See TextureCubemapFace for the face->layer mapping
    uint16_t layer = 0;
 };

@@ -100,7 +103,7 @@ public:

    // this is here for backward compatibility
    MRT(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
-            : mInfos{{ handle, level, layer }} {
+            : mInfos{{ handle, level, layer, 0 }} {
    }
 };

@@ -89,28 +89,14 @@ protected:
     */
    AcquiredImage transformAcquiredImage(AcquiredImage source) noexcept override;

-protected:
-    bool makeCurrent(ContextType type,
-            SwapChain* drawSwapChain,
-            SwapChain* readSwapChain) noexcept override;
-
 private:
-    struct InitializeJvmForPerformanceManagerIfNeeded {
-        InitializeJvmForPerformanceManagerIfNeeded();
-    };
-
    int mOSVersion;
    ExternalStreamManagerAndroid& mExternalStreamManager;
-    InitializeJvmForPerformanceManagerIfNeeded const mInitializeJvmForPerformanceManagerIfNeeded;
    utils::PerformanceHintManager mPerformanceHintManager;
    utils::PerformanceHintManager::Session mPerformanceHintSession;

    using clock = std::chrono::high_resolution_clock;
    clock::time_point mStartTimeOfActualWork;
-
-    void* mNativeWindowLib = nullptr;
-    int32_t (*ANativeWindow_getBuffersDefaultDataSpace)(ANativeWindow* window) = nullptr;
-    bool mAssertNativeWindowIsValid = false;
 };

 } // namespace filament::backend
@@ -1,63 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H
-#define TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H
-
-#include <stdint.h>
-
-#include "bluegl/BlueGL.h"
-
-#include "osmesa.h"
-
-#include <backend/platforms/OpenGLPlatform.h>
-#include <backend/DriverEnums.h>
-
-namespace filament::backend {
-
-/**
- * A concrete implementation of OpenGLPlatform that uses OSMesa, which is an offscreen
- * context that can be used in conjunction with Mesa for software rasterization.
- * See https://docs.mesa3d.org/osmesa.html for more information.
- */
-class PlatformOSMesa : public OpenGLPlatform {
-protected:
-    // --------------------------------------------------------------------------------------------
-    // Platform Interface
-
-    Driver* createDriver(void* sharedGLContext, const DriverConfig& driverConfig) noexcept override;
-
-    int getOSVersion() const noexcept final override { return 0; }
-
-    // --------------------------------------------------------------------------------------------
-    // OpenGLPlatform Interface
-
-    void terminate() noexcept override;
-
-    SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
-    SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
-    void destroySwapChain(SwapChain* swapChain) noexcept override;
-    bool makeCurrent(ContextType type, SwapChain* drawSwapChain,
-            SwapChain* readSwapChain) noexcept override;
-    void commit(SwapChain* swapChain) noexcept override;
-
-private:
-    OSMesaContext mContext;
-};
-
-} // namespace filament::backend
-
-#endif // TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H
@@ -90,20 +90,6 @@ public:
        VkExtent2D extent = {0, 0};
    };

-    struct ImageSyncData {
-        static constexpr uint32_t INVALID_IMAGE_INDEX = UINT32_MAX;
-
-        // The index of the next image as returned by vkAcquireNextImage or equivalent.
-        uint32_t imageIndex = INVALID_IMAGE_INDEX;
-
-        // Semaphore to be signaled once the image is available.
-        VkSemaphore imageReadySemaphore = VK_NULL_HANDLE;
-
-        // A function called right before vkQueueSubmit. After this call, the image must be 
-        // available. This pointer can be null if imageReadySemaphore is not VK_NULL_HANDLE.
-        std::function<void(SwapChainPtr handle)> explicitImageReadyWait = nullptr;
-    };
-
    VulkanPlatform();

    ~VulkanPlatform() override;
@@ -141,12 +127,6 @@ public:
         * before recreating the swapchain. Default is true.
         */
        bool flushAndWaitOnWindowResize = true;
-
-        /**
-         * Whether the swapchain image should be transitioned to a layout suitable for
-         * presentation. Default is true.
-         */
-        bool transitionSwapChainImageLayoutForPresent = true;
    };

    /**
@@ -175,10 +155,13 @@ public:
     * corresponding VkImage will be used as the output color attachment. The client should signal
     * the `clientSignal` semaphore when the image is ready to be used by the backend.
     * @param handle         The handle returned by createSwapChain()
-     * @param outImageSyncData The synchronization data used for image readiness
+     * @param clientSignal   The semaphore that the client will signal to indicate that the backend
+     *                       may render into the image.
+     * @param index          Pointer to memory that will be filled with the index that corresponding
+     *                       to an image in the `SwapChainBundle.colors` array.
     * @return               Result of acquire
     */
-    virtual VkResult acquire(SwapChainPtr handle, ImageSyncData* outImageSyncData);
+    virtual VkResult acquire(SwapChainPtr handle, VkSemaphore clientSignal, uint32_t* index);

    /**
     * Present the image corresponding to `index` to the display. The client should wait on
@@ -18,7 +18,6 @@
 #define TNT_FILAMENT_BACKEND_PRIVATE_DRIVER_H

 #include <backend/CallbackHandler.h>
-#include <backend/DescriptorSetOffsetArray.h>
 #include <backend/DriverApiForward.h>
 #include <backend/DriverEnums.h>
 #include <backend/Handle.h>
@@ -139,8 +139,7 @@ DECL_DRIVER_API_N(beginFrame,
 DECL_DRIVER_API_N(setFrameScheduledCallback,
        backend::SwapChainHandle, sch,
        backend::CallbackHandler*, handler,
-        backend::FrameScheduledCallback&&, callback,
-        uint64_t, flags)
+        backend::FrameScheduledCallback&&, callback)

 DECL_DRIVER_API_N(setFrameCompletedCallback,
        backend::SwapChainHandle, sch,
@@ -163,10 +162,6 @@ DECL_DRIVER_API_0(finish)
 // reset state tracking, if the driver does any state tracking (e.g. GL)
 DECL_DRIVER_API_0(resetState)

-DECL_DRIVER_API_N(setDebugTag,
-        backend::HandleBase::HandleId, handleId,
-        utils::CString, tag)
-
 /*
 * Creating driver objects
 * -----------------------
@@ -201,33 +196,20 @@ DECL_DRIVER_API_R_N(backend::TextureHandle, createTexture,
        uint32_t, depth,
        backend::TextureUsage, usage)

-DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureView,
-        backend::TextureHandle, texture,
-        uint8_t, baseLevel,
-        uint8_t, levelCount)
-
-DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureViewSwizzle,
-        backend::TextureHandle, texture,
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureSwizzled,
+        backend::SamplerType, target,
+        uint8_t, levels,
+        backend::TextureFormat, format,
+        uint8_t, samples,
+        uint32_t, width,
+        uint32_t, height,
+        uint32_t, depth,
+        backend::TextureUsage, usage,
        backend::TextureSwizzle, r,
        backend::TextureSwizzle, g,
        backend::TextureSwizzle, b,
        backend::TextureSwizzle, a)

-DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureExternalImage,
-        backend::TextureFormat, format,
-        uint32_t, width,
-        uint32_t, height,
-        backend::TextureUsage, usage,
-        void*, image)
-
-DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureExternalImagePlane,
-        backend::TextureFormat, format,
-        uint32_t, width,
-        uint32_t, height,
-        backend::TextureUsage, usage,
-        void*, image,
-        uint32_t, plane)
-
 DECL_DRIVER_API_R_N(backend::TextureHandle, importTexture,
        intptr_t, id,
        backend::SamplerType, target,
@@ -239,6 +221,9 @@ DECL_DRIVER_API_R_N(backend::TextureHandle, importTexture,
        uint32_t, depth,
        backend::TextureUsage, usage)

+DECL_DRIVER_API_R_N(backend::SamplerGroupHandle, createSamplerGroup,
+        uint32_t, size, utils::FixedSizeString<32>, debugName)
+
 DECL_DRIVER_API_R_N(backend::RenderPrimitiveHandle, createRenderPrimitive,
        backend::VertexBufferHandle, vbh,
        backend::IndexBufferHandle, ibh,
@@ -272,53 +257,25 @@ DECL_DRIVER_API_R_N(backend::SwapChainHandle, createSwapChainHeadless,

 DECL_DRIVER_API_R_0(backend::TimerQueryHandle, createTimerQuery)

-DECL_DRIVER_API_R_N(backend::DescriptorSetLayoutHandle, createDescriptorSetLayout,
-        backend::DescriptorSetLayout&&, info)
-
-DECL_DRIVER_API_R_N(backend::DescriptorSetHandle, createDescriptorSet,
-        backend::DescriptorSetLayoutHandle, dslh)
-
-DECL_DRIVER_API_N(updateDescriptorSetBuffer,
-        backend::DescriptorSetHandle, dsh,
-        backend::descriptor_binding_t, binding,
-        backend::BufferObjectHandle, boh,
-        uint32_t, offset,
-        uint32_t, size
-)
-
-DECL_DRIVER_API_N(updateDescriptorSetTexture,
-        backend::DescriptorSetHandle, dsh,
-        backend::descriptor_binding_t, binding,
-        backend::TextureHandle, th,
-        SamplerParams, params
-)
-
-DECL_DRIVER_API_N(bindDescriptorSet,
-        backend::DescriptorSetHandle, dsh,
-        backend::descriptor_set_t, set,
-        backend::DescriptorSetOffsetArray&&, offsets
-)
-

 /*
 * Destroying driver objects
 * -------------------------
 */

-DECL_DRIVER_API_N(destroyVertexBuffer,          backend::VertexBufferHandle, vbh)
-DECL_DRIVER_API_N(destroyVertexBufferInfo,      backend::VertexBufferInfoHandle, vbih)
-DECL_DRIVER_API_N(destroyIndexBuffer,           backend::IndexBufferHandle, ibh)
-DECL_DRIVER_API_N(destroyBufferObject,          backend::BufferObjectHandle, ibh)
-DECL_DRIVER_API_N(destroyRenderPrimitive,       backend::RenderPrimitiveHandle, rph)
-DECL_DRIVER_API_N(destroyProgram,               backend::ProgramHandle, ph)
-DECL_DRIVER_API_N(destroyTexture,               backend::TextureHandle, th)
-DECL_DRIVER_API_N(destroyRenderTarget,          backend::RenderTargetHandle, rth)
-DECL_DRIVER_API_N(destroySwapChain,             backend::SwapChainHandle, sch)
-DECL_DRIVER_API_N(destroyStream,                backend::StreamHandle, sh)
-DECL_DRIVER_API_N(destroyTimerQuery,            backend::TimerQueryHandle, sh)
-DECL_DRIVER_API_N(destroyFence,                 backend::FenceHandle, fh)
-DECL_DRIVER_API_N(destroyDescriptorSetLayout,   backend::DescriptorSetLayoutHandle, dslh)
-DECL_DRIVER_API_N(destroyDescriptorSet,         backend::DescriptorSetHandle, dsh)
+DECL_DRIVER_API_N(destroyVertexBuffer,    backend::VertexBufferHandle, vbh)
+DECL_DRIVER_API_N(destroyVertexBufferInfo,backend::VertexBufferInfoHandle, vbih)
+DECL_DRIVER_API_N(destroyIndexBuffer,     backend::IndexBufferHandle, ibh)
+DECL_DRIVER_API_N(destroyBufferObject,    backend::BufferObjectHandle, ibh)
+DECL_DRIVER_API_N(destroyRenderPrimitive, backend::RenderPrimitiveHandle, rph)
+DECL_DRIVER_API_N(destroyProgram,         backend::ProgramHandle, ph)
+DECL_DRIVER_API_N(destroySamplerGroup,    backend::SamplerGroupHandle, sbh)
+DECL_DRIVER_API_N(destroyTexture,         backend::TextureHandle, th)
+DECL_DRIVER_API_N(destroyRenderTarget,    backend::RenderTargetHandle, rth)
+DECL_DRIVER_API_N(destroySwapChain,       backend::SwapChainHandle, sch)
+DECL_DRIVER_API_N(destroyStream,          backend::StreamHandle, sh)
+DECL_DRIVER_API_N(destroyTimerQuery,      backend::TimerQueryHandle, sh)
+DECL_DRIVER_API_N(destroyFence,           backend::FenceHandle, fh)

 /*
 * Synchronous APIs
@@ -343,12 +300,11 @@ DECL_DRIVER_API_SYNCHRONOUS_0(bool, isFrameTimeSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isAutoDepthResolveSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isSRGBSwapChainSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedContentSupported)
-DECL_DRIVER_API_SYNCHRONOUS_0(bool, isStereoSupported)
+DECL_DRIVER_API_SYNCHRONOUS_N(bool, isStereoSupported, backend::StereoscopicType, stereoscopicType)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isParallelShaderCompileSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthStencilResolveSupported)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isDepthStencilBlitSupported, backend::TextureFormat, format)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedTexturesSupported)
-DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthClampSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(uint8_t, getMaxDrawBuffers)
 DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getMaxUniformBufferSize)
 DECL_DRIVER_API_SYNCHRONOUS_0(math::float2, getClipSpaceParams)
@@ -385,6 +341,15 @@ DECL_DRIVER_API_N(updateBufferObjectUnsynchronized,
 DECL_DRIVER_API_N(resetBufferObject,
        backend::BufferObjectHandle, ibh)

+DECL_DRIVER_API_N(updateSamplerGroup,
+        backend::SamplerGroupHandle, ubh,
+        backend::BufferDescriptor&&, data)
+
+DECL_DRIVER_API_N(setMinMaxLevels,
+        backend::TextureHandle, th,
+        uint32_t, minLevel,
+        uint32_t, maxLevel)
+
 DECL_DRIVER_API_N(update3DImage,
        backend::TextureHandle, th,
        uint32_t, level,
@@ -399,12 +364,10 @@ DECL_DRIVER_API_N(update3DImage,
 DECL_DRIVER_API_N(generateMipmaps,
        backend::TextureHandle, th)

-// Deprecated
 DECL_DRIVER_API_N(setExternalImage,
        backend::TextureHandle, th,
        void*, image)

-// Deprecated
 DECL_DRIVER_API_N(setExternalImagePlane,
        backend::TextureHandle, th,
        void*, image,
@@ -451,16 +414,37 @@ DECL_DRIVER_API_N(commit,
 * -----------------------
 */

+DECL_DRIVER_API_N(bindUniformBuffer,
+        uint32_t, index,
+        backend::BufferObjectHandle, ubh)
+
+DECL_DRIVER_API_N(bindBufferRange,
+        BufferObjectBinding, bindingType,
+        uint32_t, index,
+        backend::BufferObjectHandle, ubh,
+        uint32_t, offset,
+        uint32_t, size)
+
+DECL_DRIVER_API_N(unbindBuffer,
+        BufferObjectBinding, bindingType,
+        uint32_t, index)
+
+DECL_DRIVER_API_N(bindSamplers,
+        uint32_t, index,
+        backend::SamplerGroupHandle, sbh)
+
 DECL_DRIVER_API_N(setPushConstant,
        backend::ShaderStage, stage,
        uint8_t, index,
        backend::PushConstantVariant, value)

 DECL_DRIVER_API_N(insertEventMarker,
-        const char*, string)
+        const char*, string,
+        uint32_t, len = 0)

 DECL_DRIVER_API_N(pushGroupMarker,
-        const char*, string)
+        const char*, string,
+        uint32_t, len = 0)

 DECL_DRIVER_API_0(popGroupMarker)

@@ -18,17 +18,6 @@
 #define TNT_FILAMENT_BACKEND_PRIVATE_DRIVERAPI_H

 #include "backend/DriverApiForward.h"
-
 #include "private/backend/CommandStream.h"

-#include <stddef.h>
-
-namespace filament::backend {
-
-inline void* allocateFromCommandStream(DriverApi& driver, size_t size, size_t alignment) noexcept {
-    return driver.allocate(size, alignment);
-}
-
-} // namespace filament::backend
-
 #endif // TNT_FILAMENT_BACKEND_PRIVATE_DRIVERAPI_H
@@ -20,12 +20,11 @@
 #include <backend/Handle.h>

 #include <utils/Allocator.h>
-#include <utils/CString.h>
 #include <utils/Log.h>
-#include <utils/Panic.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/ostream.h>
+#include <utils/Panic.h>

 #include <tsl/robin_map.h>

@@ -38,9 +37,9 @@
 #include <stddef.h>
 #include <stdint.h>

-#define HandleAllocatorGL  HandleAllocator<32,  96, 136>    // ~4520 / pool / MiB
+#define HandleAllocatorGL  HandleAllocator<32,  64, 136>    // ~4520 / pool / MiB
 #define HandleAllocatorVK  HandleAllocator<64, 160, 312>    // ~1820 / pool / MiB
-#define HandleAllocatorMTL HandleAllocator<32,  64, 552>    // ~1660 / pool / MiB
+#define HandleAllocatorMTL HandleAllocator<32,  48, 552>    // ~1660 / pool / MiB

 namespace filament::backend {

@@ -169,53 +168,19 @@ public:
        auto [p, tag] = handleToPointer(handle.getId());

        if (isPoolHandle(handle.getId())) {
-            // check for pool handle use-after-free
+            // check for use after free
            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
                uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
                auto const pNode = static_cast<typename Allocator::Node*>(p);
                uint8_t const expectedAge = pNode[-1].age;
-                // getHandleTag() is only called if the check fails.
-                FILAMENT_CHECK_POSTCONDITION(expectedAge == age)
-                        << "use-after-free of Handle with id=" << handle.getId()
-                        << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
-            }
-        } else {
-            // check for heap handle use-after-free
-            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
-                uint8_t const index = (handle.getId() & HANDLE_INDEX_MASK);
-                // if we've already handed out this handle index before, it's definitely a
-                // use-after-free, otherwise it's probably just a corrupted handle
-                if (index < mId) {
-                    FILAMENT_CHECK_POSTCONDITION(p != nullptr)
-                            << "use-after-free of heap Handle with id=" << handle.getId()
-                            << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
-                } else {
-                    FILAMENT_CHECK_POSTCONDITION(p != nullptr)
-                            << "corrupted heap Handle with id=" << handle.getId()
-                            << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
-                }
+                ASSERT_POSTCONDITION(expectedAge == age,
+                        "use-after-free of Handle with id=%d", handle.getId());
            }
        }

        return static_cast<Dp>(p);
    }

-    template<typename B>
-    bool is_valid(Handle<B>& handle) {
-        if (!handle) {
-            // null handles are invalid
-            return false;
-        }
-        auto [p, tag] = handleToPointer(handle.getId());
-        if (isPoolHandle(handle.getId())) {
-            uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
-            auto const pNode = static_cast<typename Allocator::Node*>(p);
-            uint8_t const expectedAge = pNode[-1].age;
-            return expectedAge == age;
-        }
-        return p != nullptr;
-    }
-
    template<typename Dp, typename B>
    inline typename std::enable_if_t<
            std::is_pointer_v<Dp> &&
@@ -224,29 +189,6 @@ public:
        return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
    }

-    void associateTagToHandle(HandleBase::HandleId id, utils::CString&& tag) noexcept {
-        // TODO: for now, only pool handles check for use-after-free, so we only keep tags for
-        // those
-        if (isPoolHandle(id)) {
-            // Truncate the age to get the debug tag
-            uint32_t const key = id & ~(HANDLE_DEBUG_TAG_MASK ^ HANDLE_AGE_MASK);
-            // This line is the costly part. In the future, we could potentially use a custom
-            // allocator.
-            mDebugTags[key] = std::move(tag);
-        }
-    }
-
-    utils::CString getHandleTag(HandleBase::HandleId id) const noexcept {
-        if (!isPoolHandle(id)) {
-            return "(no tag)";
-        }
-        uint32_t const key = id & ~(HANDLE_DEBUG_TAG_MASK ^ HANDLE_AGE_MASK);
-        if (auto pos = mDebugTags.find(key); pos != mDebugTags.end()) {
-            return pos->second;
-        }
-        return "(no tag)";
-    }
-
 private:

    template<typename D>
@@ -298,8 +240,8 @@ private:
            Node* const pNode = static_cast<Node*>(p);
            uint8_t& expectedAge = pNode[-1].age;
            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
-                FILAMENT_CHECK_POSTCONDITION(expectedAge == age) <<
-                        "double-free of Handle of size " << size << " at " << p;
+                ASSERT_POSTCONDITION(expectedAge == age,
+                        "double-free of Handle of size %d at %p", size, p);
            }
            expectedAge = (expectedAge + 1) & 0xF; // fixme

@@ -364,24 +306,12 @@ private:
        }
    }

-    // number if bits allotted to the handle's age (currently 4 max)
-    static constexpr uint32_t HANDLE_AGE_BIT_COUNT = 4;
-    // number if bits allotted to the handle's debug tag (HANDLE_AGE_BIT_COUNT max)
-    static constexpr uint32_t HANDLE_DEBUG_TAG_BIT_COUNT = 2;
-    // bit shift for both the age and debug tag
-    static constexpr uint32_t HANDLE_AGE_SHIFT = 27;
-    // mask for the heap (vs pool) flag
-    static constexpr uint32_t HANDLE_HEAP_FLAG = 0x80000000u;
-    // mask for the age
-    static constexpr uint32_t HANDLE_AGE_MASK =
-            ((1 << HANDLE_AGE_BIT_COUNT) - 1) << HANDLE_AGE_SHIFT;
-    // mask for the debug tag
-    static constexpr uint32_t HANDLE_DEBUG_TAG_MASK =
-            ((1 << HANDLE_DEBUG_TAG_BIT_COUNT) - 1) << HANDLE_AGE_SHIFT;
-    // mask for the index
-    static constexpr uint32_t HANDLE_INDEX_MASK = 0x07FFFFFFu;
-
-    static_assert(HANDLE_DEBUG_TAG_BIT_COUNT <= HANDLE_AGE_BIT_COUNT);
+    // we handle a 4 bits age per address
+    static constexpr uint32_t HANDLE_HEAP_FLAG      = 0x80000000u;      // pool vs heap handle
+    static constexpr uint32_t HANDLE_AGE_MASK       = 0x78000000u;      // handle's age
+    static constexpr uint32_t HANDLE_INDEX_MASK     = 0x07FFFFFFu;      // handle index
+    static constexpr uint32_t HANDLE_TAG_MASK       = HANDLE_AGE_MASK;
+    static constexpr uint32_t HANDLE_AGE_SHIFT      = 27;

    static bool isPoolHandle(HandleBase::HandleId id) noexcept {
        return (id & HANDLE_HEAP_FLAG) == 0u;
@@ -396,7 +326,7 @@ private:
        // a non-pool handle.
        if (UTILS_LIKELY(isPoolHandle(id))) {
            char* const base = (char*)mHandleArena.getArea().begin();
-            uint32_t const tag = id & HANDLE_AGE_MASK;
+            uint32_t const tag = id & HANDLE_TAG_MASK;
            size_t const offset = (id & HANDLE_INDEX_MASK) * Allocator::getAlignment();
            return { static_cast<void*>(base + offset), tag };
        }
@@ -411,7 +341,7 @@ private:
        size_t const offset = (char*)p - base;
        assert_invariant((offset % Allocator::getAlignment()) == 0);
        auto id = HandleBase::HandleId(offset / Allocator::getAlignment());
-        id |= tag & HANDLE_AGE_MASK;
+        id |= tag & HANDLE_TAG_MASK;
        assert_invariant((id & HANDLE_HEAP_FLAG) == 0);
        return id;
    }
@@ -421,7 +351,6 @@ private:
    // Below is only used when running out of space in the HandleArena
    mutable utils::Mutex mLock;
    tsl::robin_map<HandleBase::HandleId, void*> mOverflowMap;
-    tsl::robin_map<HandleBase::HandleId, utils::CString> mDebugTags;
    HandleBase::HandleId mId = 0;
    bool mUseAfterFreeCheckDisabled = false;
 };
@@ -27,8 +27,8 @@ PresentCallable::PresentCallable(PresentFn fn, void* user) noexcept
 }

 void PresentCallable::operator()(bool presentFrame) noexcept {
-    FILAMENT_CHECK_PRECONDITION(mPresentFn) << "This PresentCallable was already called. "
-                                               "PresentCallables should be called exactly once.";
+    ASSERT_PRECONDITION(mPresentFn, "This PresentCallable was already called. " \
+            "PresentCallables should be called exactly once.");
    mPresentFn(presentFrame, mUser);
    // Set mPresentFn to nullptr to denote that the callable has been called.
    mPresentFn = nullptr;
@@ -24,7 +24,7 @@
 #include <utils/debug.h>
 #include <utils/ostream.h>

-#if !defined(WIN32) && !defined(__EMSCRIPTEN__)
+#if !defined(WIN32) && !defined(__EMSCRIPTEN__) && !defined(IOS)
 #    include <sys/mman.h>
 #    include <unistd.h>
 #    define HAS_MMAP 1
@@ -32,11 +32,10 @@
 #    define HAS_MMAP 0
 #endif

-#include <stddef.h>
 #include <stdint.h>
-#include <stdio.h>
+#include <stddef.h>
 #include <stdlib.h>
-#include <string.h>
+#include <stdio.h>

 using namespace utils;

@@ -82,9 +81,6 @@ void* CircularBuffer::alloc(size_t size) noexcept {
            // map the circular buffer once...
            vaddr = mmap(reserve_vaddr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
            if (vaddr != MAP_FAILED) {
-                // populate the address space with pages (because this is a circular buffer,
-                // all the pages will be allocated eventually, might as well do it now)
-                memset(vaddr, 0, size);
                // and map the circular buffer again, behind the previous copy...
                vaddr_shadow = mmap((char*)vaddr + size, size,
                        PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
@@ -105,7 +101,7 @@ void* CircularBuffer::alloc(size_t size) noexcept {
    if (UTILS_UNLIKELY(mAshmemFd < 0)) {
        // ashmem failed
        if (vaddr_guard != MAP_FAILED) {
-            munmap(vaddr_guard, BLOCK_SIZE);
+            munmap(vaddr_guard, size);
        }

        if (vaddr_shadow != MAP_FAILED) {
@@ -123,11 +119,12 @@ void* CircularBuffer::alloc(size_t size) noexcept {
        data = mmap(nullptr, size * 2 + BLOCK_SIZE,
                PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);

-        FILAMENT_CHECK_POSTCONDITION(data != MAP_FAILED) <<
-                "couldn't allocate " << (size * 2 / 1024) <<
-                " KiB of virtual address space for the command buffer";
+        ASSERT_POSTCONDITION(data,
+                "couldn't allocate %u KiB of virtual address space for the command buffer",
+                (size * 2 / 1024));

-        slog.w << "Using 'soft' CircularBuffer (" << (size * 2 / 1024) << " KiB)" << io::endl;
+        slog.d << "WARNING: Using soft CircularBuffer (" << (size * 2 / 1024) << " KiB)"
+               << io::endl;

        // guard page at the end
        void* guard = (void*)(uintptr_t(data) + size * 2);
@@ -74,6 +74,8 @@ void CommandBufferQueue::setPaused(bool paused) {

 bool CommandBufferQueue::isExitRequested() const {
    std::lock_guard<utils::Mutex> const lock(mLock);
+    ASSERT_PRECONDITION( mExitRequested == 0 || mExitRequested == EXIT_REQUESTED,
+            "mExitRequested is corrupted (value = 0x%08x)!", mExitRequested);
    return (bool)mExitRequested;
 }

@@ -101,23 +103,22 @@ void CommandBufferQueue::flush() noexcept {
    size_t const used = std::distance(
            static_cast<char const*>(begin), static_cast<char const*>(end));

-
    std::unique_lock<utils::Mutex> lock(mLock);
-
-    // circular buffer is too small, we corrupted the stream
-    FILAMENT_CHECK_POSTCONDITION(used <= mFreeSpace) <<
-            "Backend CommandStream overflow. Commands are corrupted and unrecoverable.\n"
-            "Please increase minCommandBufferSizeMB inside the Config passed to Engine::create.\n"
-            "Space used at this time: " << used <<
-            " bytes, overflow: " << used - mFreeSpace << " bytes";
-
-    mFreeSpace -= used;
    mCommandBuffersToExecute.push_back({ begin, end });
    mCondition.notify_one();

+    // circular buffer is too small, we corrupted the stream
+    ASSERT_POSTCONDITION(used <= mFreeSpace,
+            "Backend CommandStream overflow. Commands are corrupted and unrecoverable.\n"
+            "Please increase minCommandBufferSizeMB inside the Config passed to Engine::create.\n"
+            "Space used at this time: %u bytes, overflow: %u bytes",
+            (unsigned)used, unsigned(used - mFreeSpace));
+
    // wait until there is enough space in the buffer
+    mFreeSpace -= used;
    if (UTILS_UNLIKELY(mFreeSpace < requiredSize)) {

+
 #ifndef NDEBUG
        size_t const totalUsed = circularBuffer.size() - mFreeSpace;
        slog.d << "CommandStream used too much space (will block): "
@@ -130,11 +131,9 @@ void CommandBufferQueue::flush() noexcept {
 #endif

        SYSTRACE_NAME("waiting: CircularBuffer::flush()");
-
-        FILAMENT_CHECK_POSTCONDITION(!mPaused) <<
+        ASSERT_POSTCONDITION(!mPaused,
                "CommandStream is full, but since the rendering thread is paused, "
-                "the buffer cannot flush and we will deadlock. Instead, abort.";
-
+                "the buffer cannot flush and we will deadlock. Instead, abort.");
        mCondition.wait(lock, [this, requiredSize]() -> bool {
            // TODO: on macOS, we need to call pumpEvents from time to time
            return mFreeSpace >= requiredSize;
@@ -150,14 +149,16 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
    while ((mCommandBuffersToExecute.empty() || mPaused) && !mExitRequested) {
        mCondition.wait(lock);
    }
+
+    ASSERT_PRECONDITION( mExitRequested == 0 || mExitRequested == EXIT_REQUESTED,
+            "mExitRequested is corrupted (value = 0x%08x)!", mExitRequested);
+
    return std::move(mCommandBuffersToExecute);
 }

 void CommandBufferQueue::releaseBuffer(CommandBufferQueue::Range const& buffer) {
-    size_t const used = std::distance(
-            static_cast<char const*>(buffer.begin), static_cast<char const*>(buffer.end));
    std::lock_guard<utils::Mutex> const lock(mLock);
-    mFreeSpace += used;
+    mFreeSpace += uintptr_t(buffer.end) - uintptr_t(buffer.begin);
    mCondition.notify_one();
 }

@@ -20,16 +20,11 @@
 #include <utils/CallStack.h>
 #endif

-#include <utils/compiler.h>
 #include <utils/Log.h>
-#include <utils/ostream.h>
 #include <utils/Profiler.h>
 #include <utils/Systrace.h>

-#include <cstddef>
 #include <functional>
-#include <string>
-#include <utility>

 #ifdef __ANDROID__
 #include <sys/system_properties.h>
@@ -79,8 +74,8 @@ CommandStream::CommandStream(Driver& driver, CircularBuffer& buffer) noexcept
 }

 void CommandStream::execute(void* buffer) {
-    // NOTE: we can't use SYSTRACE_CALL() or similar here because, execute() below, also
-    // uses systrace BEGIN/END and the END is not guaranteed to be happening in this scope.
+    SYSTRACE_CALL();
+    SYSTRACE_CONTEXT();

    Profiler profiler;

@@ -105,7 +100,6 @@ void CommandStream::execute(void* buffer) {
            // we want to remove all this when tracing is completely disabled
            profiler.stop();
            UTILS_UNUSED Profiler::Counters const counters = profiler.readCounters();
-            SYSTRACE_CONTEXT();
            SYSTRACE_VALUE32("GLThread (I)", counters.getInstructions());
            SYSTRACE_VALUE32("GLThread (C)", counters.getCpuCycles());
            SYSTRACE_VALUE32("GLThread (CPI x10)", counters.getCPI() * 10);
@@ -101,14 +101,6 @@ struct HwProgram : public HwBase {
    HwProgram() noexcept = default;
 };

-struct HwDescriptorSetLayout : public HwBase {
-    HwDescriptorSetLayout() noexcept = default;
-};
-
-struct HwDescriptorSet : public HwBase {
-    HwDescriptorSet() noexcept = default;
-};
-
 struct HwSamplerGroup : public HwBase {
    HwSamplerGroup() noexcept = default;
 };
@@ -80,9 +80,6 @@ HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size,
        bool disableUseAfterFreeCheck) noexcept
    : mHandleArena(name, size, disableUseAfterFreeCheck),
      mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
-    // Reserve initial space for debug tags. This prevents excessive calls to malloc when the first
-    // few tags are set.
-    mDebugTags.reserve(512);
 }

 template <size_t P0, size_t P1, size_t P2>
@@ -116,9 +113,9 @@ HandleBase::HandleId HandleAllocator<P0, P1, P2>::allocateHandleSlow(size_t size

    HandleBase::HandleId id = (++mId) | HANDLE_HEAP_FLAG;

-    FILAMENT_CHECK_POSTCONDITION(mId < HANDLE_HEAP_FLAG) <<
+    ASSERT_POSTCONDITION(mId < HANDLE_HEAP_FLAG,
            "No more Handle ids available! This can happen if HandleAllocator arena has been full"
-            " for a while. Please increase FILAMENT_OPENGL_HANDLE_ARENA_SIZE_IN_MB";
+            " for a while. Please increase FILAMENT_OPENGL_HANDLE_ARENA_SIZE_IN_MB");

    mOverflowMap.emplace(id, p);
    lock.unlock();
@@ -41,10 +41,6 @@
        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
            #include "backend/platforms/PlatformEGLHeadless.h"
        #endif
-    #elif defined(FILAMENT_SUPPORTS_OSMESA)
-        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
-            #include "backend/platforms/PlatformOSMesa.h"
-        #endif
    #endif
 #elif defined(WIN32)
    #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
@@ -128,8 +124,6 @@ Platform* PlatformFactory::create(Backend* backend) noexcept {
                return new PlatformGLX();
            #elif defined(FILAMENT_SUPPORTS_EGL_ON_LINUX)
                return new PlatformEGLHeadless();
-            #elif defined(FILAMENT_SUPPORTS_OSMESA)
-                return new PlatformOSMesa();
            #else
                return nullptr;
            #endif
@@ -14,18 +14,7 @@
 * limitations under the License.
 */

-#include <backend/Program.h>
-#include <backend/DriverEnums.h>
-
-#include <utils/debug.h>
-#include <utils/CString.h>
-#include <utils/ostream.h>
-#include <utils/Invocable.h>
-
-#include <utility>
-
-#include <stddef.h>
-#include <stdint.h>
+#include "backend/Program.h"

 namespace filament::backend {

@@ -63,24 +52,41 @@ Program& Program::shaderLanguage(ShaderLanguage shaderLanguage) {
    return *this;
 }

-Program& Program::descriptorBindings(backend::descriptor_set_t set,
-        DescriptorBindingsInfo descriptorBindings) noexcept {
-    mDescriptorBindings[set] = std::move(descriptorBindings);
+Program& Program::uniformBlockBindings(
+        FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept {
+    for (auto const& item : uniformBlockBindings) {
+        assert_invariant(item.second < UNIFORM_BINDING_COUNT);
+        mUniformBlocks[item.second] = item.first;
+    }
    return *this;
 }

-Program& Program::uniforms(uint32_t index, utils::CString name, UniformInfo uniforms) noexcept {
-    mBindingUniformsInfo.reserve(mBindingUniformsInfo.capacity() + 1);
-    mBindingUniformsInfo.emplace_back(index, std::move(name), std::move(uniforms));
+Program& Program::uniforms(uint32_t index, UniformInfo const& uniforms) noexcept {
+    assert_invariant(index < UNIFORM_BINDING_COUNT);
+    mBindingUniformInfo[index] = uniforms;
    return *this;
 }

-Program& Program::attributes(AttributesInfo attributes) noexcept {
+
+Program& Program::attributes(
+        utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes) noexcept {
    mAttributes = std::move(attributes);
    return *this;
 }

-Program& Program::specializationConstants(SpecializationConstantsInfo specConstants) noexcept {
+Program& Program::setSamplerGroup(size_t bindingPoint, ShaderStageFlags stageFlags,
+        const Program::Sampler* samplers, size_t count) noexcept {
+    auto& groupData = mSamplerGroups[bindingPoint];
+    groupData.stageFlags = stageFlags;
+    auto& samplerList = groupData.samplers;
+    samplerList.reserve(count);
+    samplerList.resize(count);
+    std::copy_n(samplers, count, samplerList.data());
+    return *this;
+}
+
+Program& Program::specializationConstants(
+        FixedCapacityVector<SpecializationConstant> specConstants) noexcept {
    mSpecializationConstants = std::move(specConstants);
    return *this;
 }
@@ -16,23 +16,13 @@

 #include "private/backend/VirtualMachineEnv.h"

-#include <utils/compiler.h>
 #include <utils/debug.h>

-#include <jni.h>
-
 namespace filament {

 JavaVM* VirtualMachineEnv::sVirtualMachine = nullptr;

-/*
- * This is typically called by filament_jni.so when it is loaded. If filament_jni.so is not used,
- * then this must be called manually -- however, this is a problem because VirtualMachineEnv.h
- * is currently private and part of backend.
- * For now, we authorize this usage, but we will need to fix it; by making a proper public
- * API for this.
- */
-UTILS_PUBLIC
+// This is called when the library is loaded. We need this to get a reference to the global VM
 UTILS_NOINLINE
 jint VirtualMachineEnv::JNI_OnLoad(JavaVM* vm) noexcept {
    JNIEnv* env = nullptr;
@@ -109,8 +109,9 @@ inline bool MTLSizeEqual(T a, T b) noexcept {
 MetalBlitter::MetalBlitter(MetalContext& context) noexcept : mContext(context) { }

 void MetalBlitter::blit(id<MTLCommandBuffer> cmdBuffer, const BlitArgs& args, const char* label) {
-    FILAMENT_CHECK_PRECONDITION(args.source.region.size.depth == args.destination.region.size.depth)
-            << "Blitting requires the source and destination regions to have the same depth.";
+
+    ASSERT_PRECONDITION(args.source.region.size.depth == args.destination.region.size.depth,
+            "Blitting requires the source and destination regions to have the same depth.");

    // Determine if the blit for color or depth are eligible to use a MTLBlitCommandEncoder.
    // blitFastPath returns true upon success.
@@ -326,8 +327,7 @@ id<MTLFunction> MetalBlitter::compileFragmentFunction(BlitFunctionKey key) const
            utils::slog.e << description << utils::io::endl;
        }
    }
-    FILAMENT_CHECK_POSTCONDITION(library && function)
-            << "Unable to compile fragment shader for MetalBlitter.";
+    ASSERT_POSTCONDITION(library && function, "Unable to compile fragment shader for MetalBlitter.");

    return function;
 }
@@ -352,8 +352,7 @@ id<MTLFunction> MetalBlitter::getBlitVertexFunction() {
            utils::slog.e << description << utils::io::endl;
        }
    }
-    FILAMENT_CHECK_POSTCONDITION(library && function)
-            << "Unable to compile vertex shader for MetalBlitter.";
+    ASSERT_POSTCONDITION(library && function, "Unable to compile vertex shader for MetalBlitter.");

    mVertexFunction = function;

@@ -65,19 +65,21 @@ private:
    const char* mName;
 };

-class TrackedMetalBuffer {
-public:
+#ifndef FILAMENT_METAL_BUFFER_TRACKING
+#define FILAMENT_METAL_BUFFER_TRACKING 0
+#endif

+class MetalBufferTracking {
+public:
    static constexpr size_t EXCESS_BUFFER_COUNT = 30000;

    enum class Type {
        NONE = 0,
        GENERIC = 1,
-        RING = 2,       // deprecated
+        RING = 2,
        STAGING = 3,
-        DESCRIPTOR_SET = 4,
    };
-    static constexpr size_t TypeCount = 4;
+    static constexpr size_t TypeCount = 3;

    static constexpr auto toIndex(Type t) {
        assert_invariant(t != Type::NONE);
@@ -89,71 +91,60 @@ public:
                return 1;
            case Type::STAGING:
                return 2;
-            case Type::DESCRIPTOR_SET:
-                return 3;
        }
    }

-    TrackedMetalBuffer() noexcept : mBuffer(nil) {}
-    TrackedMetalBuffer(nullptr_t) noexcept : mBuffer(nil) {}
-    TrackedMetalBuffer(id<MTLBuffer> buffer, Type type) : mBuffer(buffer), mType(type) {
+#if FILAMENT_METAL_BUFFER_TRACKING
+    static void initialize() {
+        static dispatch_once_t onceToken;
+        dispatch_once(&onceToken, ^{
+            for (size_t i = 0; i < TypeCount; i++) {
+                aliveBuffers[i] = [NSHashTable weakObjectsHashTable];
+            }
+        });
+    }
+
+    static void setPlatform(MetalPlatform* p) { platform = p; }
+
+    static void track(id<MTLBuffer> buffer, Type type) {
        assert_invariant(type != Type::NONE);
-        if (buffer) {
-            aliveBuffers[toIndex(type)]++;
-            mType = type;
-            if (getAliveBuffers() >= EXCESS_BUFFER_COUNT) {
-                if (platform && platform->hasDebugUpdateStatFunc()) {
-                    platform->debugUpdateStat("filament.metal.excess_buffers_allocated",
-                            TrackedMetalBuffer::getAliveBuffers());
-                }
+        if (UTILS_UNLIKELY(getAliveBuffers() >= EXCESS_BUFFER_COUNT)) {
+            if (platform && platform->hasDebugUpdateStatFunc()) {
+                platform->debugUpdateStat("filament.metal.excess_buffers_allocated",
+                        MetalBufferTracking::getAliveBuffers());
            }
        }
+        [aliveBuffers[toIndex(type)] addObject:buffer];
    }

-    ~TrackedMetalBuffer() {
-        if (mBuffer) {
-            assert_invariant(mType != Type::NONE);
-            aliveBuffers[toIndex(mType)]--;
-        }
-    }
-
-    TrackedMetalBuffer(TrackedMetalBuffer&&) = delete;
-    TrackedMetalBuffer(TrackedMetalBuffer const&) = delete;
-    TrackedMetalBuffer& operator=(TrackedMetalBuffer const&) = delete;
-
-    TrackedMetalBuffer& operator=(TrackedMetalBuffer&& rhs) noexcept {
-        swap(rhs);
-        return *this;
-    }
-
-    id<MTLBuffer> get() const noexcept { return mBuffer; }
-    operator bool() const noexcept { return bool(mBuffer); }
-
    static uint64_t getAliveBuffers() {
        uint64_t sum = 0;
-        for (const auto& v : aliveBuffers) {
-            sum += v;
+        for (size_t i = 1; i < TypeCount; i++) {
+            sum += getAliveBuffers(static_cast<Type>(i));
        }
        return sum;
    }

    static uint64_t getAliveBuffers(Type type) {
        assert_invariant(type != Type::NONE);
-        return aliveBuffers[toIndex(type)];
+        NSHashTable* hashTable = aliveBuffers[toIndex(type)];
+        // Caution! We can't simply use hashTable.count here, which is inaccurate.
+        // See http://cocoamine.net/blog/2013/12/13/nsmaptable-and-zeroing-weak-references/
+        return hashTable.objectEnumerator.allObjects.count;
    }
-    static void setPlatform(MetalPlatform* p) { platform = p; }
+#else
+    static void initialize() {}
+    static void setPlatform(MetalPlatform* p) {}
+    static id<MTLBuffer> track(id<MTLBuffer> buffer, Type type) { return buffer; }
+    static uint64_t getAliveBuffers() { return 0; }
+    static uint64_t getAliveBuffers(Type type) { return 0; }
+#endif

 private:
-    void swap(TrackedMetalBuffer& other) noexcept {
-        std::swap(mBuffer, other.mBuffer);
-        std::swap(mType, other.mType);
-    }
-
-    id<MTLBuffer> mBuffer;
-    Type mType = Type::NONE;
-
+#if FILAMENT_METAL_BUFFER_TRACKING
+    static std::array<NSHashTable<id<MTLBuffer>>*, TypeCount> aliveBuffers;
    static MetalPlatform* platform;
-    static std::array<uint64_t, TypeCount> aliveBuffers;
+#endif
 };

 class MetalBuffer {
@@ -163,8 +154,6 @@ public:
         size_t size, bool forceGpuBuffer = false);
    ~MetalBuffer();

-    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer || mCpuBuffer; }
-
    MetalBuffer(const MetalBuffer& rhs) = delete;
    MetalBuffer& operator=(const MetalBuffer& rhs) = delete;

@@ -174,10 +163,8 @@ public:
     * Update the buffer with data inside src. Potentially allocates a new buffer allocation to hold
     * the bytes which will be released when the current frame is finished.
     */
-    using TagResolver = utils::Invocable<const char*(void)>;
-    void copyIntoBuffer(void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag);
-    void copyIntoBufferUnsynchronized(
-            void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag);
+    void copyIntoBuffer(void* src, size_t size, size_t byteOffset);
+    void copyIntoBufferUnsynchronized(void* src, size_t size, size_t byteOffset);

    /**
     * Denotes that this buffer is used for a draw call ensuring that its allocation remains valid
@@ -187,7 +174,7 @@ public:
     * is no device allocation.
     *
     */
-    id<MTLBuffer> getGpuBufferForDraw() noexcept;
+    id<MTLBuffer> getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept;

    void* getCpuBuffer() const noexcept { return mCpuBuffer; }

@@ -211,18 +198,7 @@ public:

 private:

-    enum class UploadStrategy {
-        POOL,
-        BUMP_ALLOCATOR,
-    };
-
-    void uploadWithPoolBuffer(
-            void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const;
-    void uploadWithBumpAllocator(
-            void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const;
-
-    UploadStrategy mUploadStrategy;
-    TrackedMetalBuffer mBuffer;
+    id<MTLBuffer> mBuffer;
    size_t mBufferSize = 0;
    void* mCpuBuffer = nullptr;
    MetalContext& mContext;
@@ -271,9 +247,11 @@ public:
          mBufferOptions(options),
          mSlotSizeBytes(computeSlotSize(layout)),
          mSlotCount(slotCount) {
-        ScopedAllocationTimer timer("ring");
-        mBuffer = { [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions],
-            TrackedMetalBuffer::Type::RING };
+        {
+            ScopedAllocationTimer timer("ring");
+            mBuffer = [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions];
+        }
+        MetalBufferTracking::track(mBuffer, MetalBufferTracking::Type::RING);
        assert_invariant(mBuffer);
    }

@@ -293,11 +271,11 @@ public:
            // finishes executing.
            {
                ScopedAllocationTimer timer("ring");
-                mAuxBuffer = { [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions],
-                    TrackedMetalBuffer::Type::RING };
+                mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
            }
+            MetalBufferTracking::track(mAuxBuffer, MetalBufferTracking::Type::RING);
            assert_invariant(mAuxBuffer);
-            return { mAuxBuffer.get(), 0 };
+            return { mAuxBuffer, 0 };
        }
        mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
        mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);
@@ -326,9 +304,9 @@ public:
     */
    std::pair<id<MTLBuffer>, NSUInteger> getCurrentAllocation() const {
        if (UTILS_UNLIKELY(mAuxBuffer)) {
-            return { mAuxBuffer.get(), 0 };
+            return { mAuxBuffer, 0 };
        }
-        return { mBuffer.get(), mCurrentSlot * mSlotSizeBytes };
+        return { mBuffer, mCurrentSlot * mSlotSizeBytes };
    }

    bool canAccomodateLayout(MTLSizeAndAlign layout) const {
@@ -337,8 +315,8 @@ public:

 private:
    id<MTLDevice> mDevice;
-    TrackedMetalBuffer mBuffer;
-    TrackedMetalBuffer mAuxBuffer;
+    id<MTLBuffer> mBuffer;
+    id<MTLBuffer> mAuxBuffer;

    MTLResourceOptions mBufferOptions;

@@ -22,42 +22,33 @@
 namespace filament {
 namespace backend {

-std::array<uint64_t, TrackedMetalBuffer::TypeCount> TrackedMetalBuffer::aliveBuffers = { 0 };
-MetalPlatform* TrackedMetalBuffer::platform = nullptr;
 MetalPlatform* ScopedAllocationTimer::platform = nullptr;

-MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
-        size_t size, bool forceGpuBuffer)
-    : mBufferSize(size), mContext(context) {
-    const MetalBumpAllocator& allocator = *mContext.bumpAllocator;
-    // VERTEX is also used for index buffers
-    if (allocator.getCapacity() > 0 && bindingType == BufferObjectBinding::VERTEX) {
-        mUploadStrategy = UploadStrategy::BUMP_ALLOCATOR;
-    } else {
-        mUploadStrategy = UploadStrategy::POOL;
-    }
+#if FILAMENT_METAL_BUFFER_TRACKING
+std::array<NSHashTable<id<MTLBuffer>>*, MetalBufferTracking::TypeCount>
+        MetalBufferTracking::aliveBuffers;
+MetalPlatform* MetalBufferTracking::platform = nullptr;
+#endif

+MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
+        size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
    // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
    // buffer. Instead, we use immediate command encoder methods like setVertexBytes:length:atIndex:.
    // This won't work for SSBOs, since they are read/write.
-
-    /*
    if (size <= 4 * 1024 && bindingType != BufferObjectBinding::SHADER_STORAGE &&
            usage == BufferUsage::DYNAMIC && !forceGpuBuffer) {
        mBuffer = nil;
        mCpuBuffer = malloc(size);
        return;
    }
-    */

    // Otherwise, we allocate a private GPU buffer.
    {
        ScopedAllocationTimer timer("generic");
-        mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
-            TrackedMetalBuffer::Type::GENERIC };
+        mBuffer = [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate];
    }
-    // mBuffer might fail to be allocated. Clients can check for this by calling
-    // wasAllocationSuccessful().
+    MetalBufferTracking::track(mBuffer, MetalBufferTracking::Type::GENERIC);
+    ASSERT_POSTCONDITION(mBuffer, "Could not allocate Metal buffer of size %zu.", size);
 }

 MetalBuffer::~MetalBuffer() {
@@ -66,51 +57,56 @@ MetalBuffer::~MetalBuffer() {
    }
 }

-void MetalBuffer::copyIntoBuffer(
-        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
+void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
    if (size <= 0) {
        return;
    }
+    ASSERT_PRECONDITION(size + byteOffset <= mBufferSize,
+            "Attempting to copy %zu bytes into a buffer of size %zu at offset %zu",
+            size, mBufferSize, byteOffset);

-    FILAMENT_CHECK_PRECONDITION(src)
-            << "copyIntoBuffer called with a null src, tag=" << getHandleTag();
-    FILAMENT_CHECK_PRECONDITION(size + byteOffset <= mBufferSize)
-            << "Attempting to copy " << size << " bytes into a buffer of size " << mBufferSize
-            << " at offset " << byteOffset << ", tag=" << getHandleTag();
-    // The copy blit requires that byteOffset be a multiple of 4.
-    FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3))
-            << "byteOffset must be a multiple of 4, tag=" << getHandleTag();
-
-    // If we have a cpu buffer, we can directly copy into it.
+    // Either copy into the Metal buffer or into our cpu buffer.
    if (mCpuBuffer) {
        memcpy(static_cast<uint8_t*>(mCpuBuffer) + byteOffset, src, size);
        return;
    }

-    switch (mUploadStrategy) {
-        case UploadStrategy::BUMP_ALLOCATOR:
-            uploadWithBumpAllocator(src, size, byteOffset, std::move(getHandleTag));
-            break;
-        case UploadStrategy::POOL:
-            uploadWithPoolBuffer(src, size, byteOffset, std::move(getHandleTag));
-            break;
-    }
+    // Acquire a staging buffer to hold the contents of this update.
+    MetalBufferPool* bufferPool = mContext.bufferPool;
+    const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
+    memcpy(staging->buffer.contents, src, size);
+
+    // The blit below requires that byteOffset be a multiple of 4.
+    ASSERT_PRECONDITION(!(byteOffset & 0x3u), "byteOffset must be a multiple of 4");
+
+    // Encode a blit from the staging buffer into the private GPU buffer.
+    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
+    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
+    blitEncoder.label = @"Buffer upload blit";
+    [blitEncoder copyFromBuffer:staging->buffer
+                   sourceOffset:0
+                       toBuffer:mBuffer
+              destinationOffset:byteOffset
+                           size:size];
+    [blitEncoder endEncoding];
+    [cmdBuffer addCompletedHandler:^(id<MTLCommandBuffer> cb) {
+        bufferPool->releaseBuffer(staging);
+    }];
 }

-void MetalBuffer::copyIntoBufferUnsynchronized(
-        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
+void MetalBuffer::copyIntoBufferUnsynchronized(void* src, size_t size, size_t byteOffset) {
    // TODO: implement the unsynchronized version
-    copyIntoBuffer(src, size, byteOffset, std::move(getHandleTag));
+    copyIntoBuffer(src, size, byteOffset);
 }

-id<MTLBuffer> MetalBuffer::getGpuBufferForDraw() noexcept {
+id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept {
    // If there's a CPU buffer, then we return nil here, as the CPU-side buffer will be bound
    // separately.
    if (mCpuBuffer) {
        return nil;
    }
    assert_invariant(mBuffer);
-    return mBuffer.get();
+    return mBuffer;
 }

 void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncoder> encoder,
@@ -146,7 +142,7 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
        }
        // getGpuBufferForDraw() might return nil, which means there isn't a device allocation for
        // this buffer. In this case, we'll bind the buffer below with the CPU-side memory.
-        id<MTLBuffer> gpuBuffer = buffer->getGpuBufferForDraw();
+        id<MTLBuffer> gpuBuffer = buffer->getGpuBufferForDraw(cmdBuffer);
        if (!gpuBuffer) {
            continue;
        }
@@ -206,54 +202,5 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
    }
 }

-void MetalBuffer::uploadWithPoolBuffer(
-        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const {
-    MetalBufferPool* bufferPool = mContext.bufferPool;
-    const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
-    FILAMENT_CHECK_POSTCONDITION(staging)
-            << "uploadWithPoolbuffer unable to acquire staging buffer of size " << size
-            << ", tag=" << getHandleTag();
-    memcpy(staging->buffer.get().contents, src, size);
-
-    // Encode a blit from the staging buffer into the private GPU buffer.
-    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
-    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
-    blitEncoder.label = @"Buffer upload blit - pool buffer";
-    [blitEncoder copyFromBuffer:staging->buffer.get()
-                   sourceOffset:0
-                       toBuffer:mBuffer.get()
-              destinationOffset:byteOffset
-                           size:size];
-    [blitEncoder endEncoding];
-    [cmdBuffer addCompletedHandler:^(id<MTLCommandBuffer> cb) {
-        bufferPool->releaseBuffer(staging);
-    }];
-}
-
-void MetalBuffer::uploadWithBumpAllocator(
-        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const {
-    MetalBumpAllocator& allocator = *mContext.bumpAllocator;
-    auto [buffer, offset] = allocator.allocateStagingArea(size);
-    FILAMENT_CHECK_POSTCONDITION(buffer)
-            << "uploadWithBumpAllocator unable to acquire staging area of size " << size
-            << ", tag=" << getHandleTag();
-    void* const contents = buffer.contents;
-    FILAMENT_CHECK_POSTCONDITION(contents)
-            << "uploadWithBumpAllocator unable to acquire pointer to staging area, size " << size
-            << ", tag=" << getHandleTag();
-    memcpy(static_cast<char*>(contents) + offset, src, size);
-
-    // Encode a blit from the staging buffer into the private GPU buffer.
-    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
-    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
-    blitEncoder.label = @"Buffer upload blit - bump allocator";
-    [blitEncoder copyFromBuffer:buffer
-                   sourceOffset:offset
-                       toBuffer:mBuffer.get()
-              destinationOffset:byteOffset
-                           size:size];
-    [blitEncoder endEncoding];
-}
-
 } // namespace backend
 } // namespace filament
@@ -32,34 +32,12 @@ struct MetalContext;

 // Immutable POD representing a shared CPU-GPU buffer.
 struct MetalBufferPoolEntry {
-    TrackedMetalBuffer buffer;
+    id<MTLBuffer> buffer;
    size_t capacity;
    mutable uint64_t lastAccessed;
    mutable uint32_t referenceCount;
 };

-class MetalBumpAllocator {
-public:
-    MetalBumpAllocator(id<MTLDevice> device, size_t capacity);
-
-    /**
-     * Allocates a staging area of the given size. Returns a pair of the buffer and the offset
-     * within the buffer. The buffer is guaranteed to be at least the given size, but may be larger.
-     * Clients must not write to the buffer beyond the returned offset + size.
-     * Clients are responsible for holding a reference to the returned buffer.
-     * Allocations are guaranteed to be aligned to 4 bytes.
-     */
-    std::pair<id<MTLBuffer>, size_t> allocateStagingArea(size_t size);
-
-    size_t getCapacity() const noexcept { return mCapacity; }
-
-private:
-    id<MTLDevice> mDevice;
-    TrackedMetalBuffer mCurrentUploadBuffer = nil;
-    size_t mHead = 0;
-    size_t mCapacity;
-};
-
 // Manages a pool of Metal buffers, periodically releasing ones that have been unused for awhile.
 class MetalBufferPool {
 public:
@@ -48,10 +48,10 @@ MetalBufferPoolEntry const* MetalBufferPool::acquireBuffer(size_t numBytes) {
        buffer = [mContext.device newBufferWithLength:numBytes
                                              options:MTLResourceStorageModeShared];
    }
-    FILAMENT_CHECK_POSTCONDITION(buffer)
-            << "Could not allocate Metal staging buffer of size " << numBytes << ".";
+    MetalBufferTracking::track(buffer, MetalBufferTracking::Type::STAGING);
+    ASSERT_POSTCONDITION(buffer, "Could not allocate Metal staging buffer of size %zu.", numBytes);
    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry {
-        .buffer = { buffer, TrackedMetalBuffer::Type::STAGING },
+        .buffer = buffer,
        .capacity = numBytes,
        .lastAccessed = mCurrentFrame,
        .referenceCount = 1
@@ -116,39 +116,5 @@ void MetalBufferPool::reset() noexcept {
    mFreeStages.clear();
 }

-MetalBumpAllocator::MetalBumpAllocator(id<MTLDevice> device, size_t capacity)
-    : mDevice(device), mCapacity(capacity) {
-    if (mCapacity > 0) {
-        mCurrentUploadBuffer = { [device newBufferWithLength:capacity options:MTLStorageModeShared],
-            TrackedMetalBuffer::Type::STAGING };
-    }
-}
-
-std::pair<id<MTLBuffer>, size_t> MetalBumpAllocator::allocateStagingArea(size_t size) {
-    if (size == 0) {
-        return { nil, 0 };
-    }
-    if (size > mCapacity) {
-        return { [mDevice newBufferWithLength:size options:MTLStorageModeShared], 0 };
-    }
-    assert_invariant(mCurrentUploadBuffer);
-
-    // Align the head to a 4-byte boundary.
-    mHead = (mHead + 3) & ~3;
-
-    if (UTILS_LIKELY(mHead + size <= mCapacity)) {
-        const size_t oldHead = mHead;
-        mHead += size;
-        return { mCurrentUploadBuffer.get(), oldHead };
-    }
-
-    // We're finished with the current allocation.
-    mCurrentUploadBuffer = { [mDevice newBufferWithLength:mCapacity options:MTLStorageModeShared],
-        TrackedMetalBuffer::Type::STAGING };
-    mHead = size;
-
-    return { mCurrentUploadBuffer.get(), 0 };
-}
-
 } // namespace backend
 } // namespace filament
@@ -21,8 +21,6 @@
 #include "MetalShaderCompiler.h"
 #include "MetalState.h"

-#include <backend/DriverEnums.h>
-
 #include <CoreVideo/CVMetalTextureCache.h>
 #include <Metal/Metal.h>
 #include <QuartzCore/QuartzCore.h>
@@ -46,88 +44,27 @@ namespace backend {
 class MetalDriver;
 class MetalBlitter;
 class MetalBufferPool;
-class MetalBumpAllocator;
 class MetalRenderTarget;
+class MetalSamplerGroup;
 class MetalSwapChain;
 class MetalTexture;
 class MetalTimerQueryInterface;
 struct MetalUniformBuffer;
 struct MetalIndexBuffer;
 struct MetalVertexBuffer;
-struct MetalDescriptorSet;

 constexpr static uint8_t MAX_SAMPLE_COUNT = 8;  // Metal devices support at most 8 MSAA samples

-class MetalPushConstantBuffer {
-public:
-    void setPushConstant(PushConstantVariant value, uint8_t index);
-    bool isDirty() const { return mDirty; }
-    void setBytes(id<MTLCommandEncoder> encoder, ShaderStage stage);
-    void clear();
-
-private:
-    std::vector<PushConstantVariant> mPushConstants;
-    bool mDirty = false;
-};
-
-class MetalDynamicOffsets {
-public:
-    void setOffsets(uint32_t set, const uint32_t* offsets, uint32_t count) {
-        assert(set < MAX_DESCRIPTOR_SET_COUNT);
-
-        auto getStartIndexForSet = [&](uint32_t s) {
-            uint32_t startIndex = 0;
-            for (uint32_t i = 0; i < s; i++) {
-                startIndex += mCounts[i];
-            }
-            return startIndex;
-        };
-
-        const bool resizeNecessary = mCounts[set] != count;
-        if (UTILS_UNLIKELY(resizeNecessary)) {
-            int delta = count - mCounts[set];
-
-            auto thisSetStart = mOffsets.begin() + getStartIndexForSet(set);
-            if (delta > 0) {
-                mOffsets.insert(thisSetStart, delta, 0);
-            } else {
-                mOffsets.erase(thisSetStart, thisSetStart - delta);
-            }
-
-            mCounts[set] = count;
-        }
-
-        if (resizeNecessary ||
-                !std::equal(
-                        offsets, offsets + count, mOffsets.begin() + getStartIndexForSet(set))) {
-            std::copy(offsets, offsets + count, mOffsets.begin() + getStartIndexForSet(set));
-            mDirty = true;
-        }
-    }
-    bool isDirty() const { return mDirty; }
-    void setDirty(bool dirty) { mDirty = dirty; }
-
-    std::pair<uint32_t, const uint32_t*> getOffsets() const {
-        return { mOffsets.size(), mOffsets.data() };
-    }
-
-private:
-    std::array<uint32_t, MAX_DESCRIPTOR_SET_COUNT> mCounts = { 0 };
-    std::vector<uint32_t> mOffsets;
-    bool mDirty = false;
-};
-
 struct MetalContext {
+    explicit MetalContext(size_t metalFreedTextureListSize)
+        : texturesToDestroy(metalFreedTextureListSize) {}
+
    MetalDriver* driver;
    id<MTLDevice> device = nullptr;
    id<MTLCommandQueue> commandQueue = nullptr;

-    // The ID of pendingCommandBuffer (or the next command buffer, if pendingCommandBuffer is nil).
-    uint64_t pendingCommandBufferId = 1;
-    // read from driver thread, set from completion handlers
-    std::atomic<uint64_t> latestCompletedCommandBufferId = 0;
-    id<MTLCommandBuffer> pendingCommandBuffer = nil;
-    id<MTLRenderCommandEncoder> currentRenderPassEncoder = nil;
+    id<MTLCommandBuffer> pendingCommandBuffer = nullptr;
+    id<MTLRenderCommandEncoder> currentRenderPassEncoder = nullptr;

    std::atomic<bool> memorylessLimitsReached = false;

@@ -143,7 +80,7 @@ struct MetalContext {
    } highestSupportedGpuFamily;

    struct {
-        bool staticTextureTargetError;
+        bool a8xStaticTextureTargetError;
    } bugs;

    // sampleCountLookup[requestedSamples] gives a <= sample count supported by the device.
@@ -158,9 +95,10 @@ struct MetalContext {
    // State trackers.
    PipelineStateTracker pipelineState;
    DepthStencilStateTracker depthStencilState;
+    std::array<BufferState, Program::UNIFORM_BINDING_COUNT> uniformState;
+    std::array<BufferState, MAX_SSBO_COUNT> ssboState;
    CullModeStateTracker cullModeState;
    WindingStateTracker windingState;
-    DepthClampStateTracker depthClampState;
    Handle<HwRenderPrimitive> currentRenderPrimitive;

    // State caches.
@@ -171,21 +109,24 @@ struct MetalContext {

    PolygonOffset currentPolygonOffset = {0.0f, 0.0f};

-    std::array<MetalPushConstantBuffer, Program::SHADER_TYPE_COUNT> currentPushConstants;
+    MetalSamplerGroup* samplerBindings[Program::SAMPLER_BINDING_COUNT] = {};

-    // Keeps track of descriptor sets we've finalized for the current render pass.
-    tsl::robin_set<MetalDescriptorSet*> finalizedDescriptorSets;
-    std::array<MetalDescriptorSet*, MAX_DESCRIPTOR_SET_COUNT> currentDescriptorSets = {};
-    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::VERTEX> vertexDescriptorBindings;
-    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::FRAGMENT> fragmentDescriptorBindings;
-    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::COMPUTE> computeDescriptorBindings;
-    MetalDynamicOffsets dynamicOffsets;
+    // Keeps track of sampler groups we've finalized for the current render pass.
+    tsl::robin_set<MetalSamplerGroup*> finalizedSamplerGroups;

-    // Keeps track of all alive textures.
+    // Keeps track of all alive sampler groups, textures.
+    tsl::robin_set<MetalSamplerGroup*> samplerGroups;
    tsl::robin_set<MetalTexture*> textures;

+    // This circular buffer implements delayed destruction for Metal texture handles. It keeps a
+    // handle to a fixed number of the most recently destroyed texture handles. When we're asked to
+    // destroy a texture handle, we free its texture memory, but keep the MetalTexture object alive,
+    // marking it as "terminated". If we later are asked to use that texture, we can check its
+    // terminated status and throw an Objective-C error instead of crashing, which is helpful for
+    // debugging use-after-free issues in release builds.
+    utils::FixedCircularBuffer<Handle<HwTexture>> texturesToDestroy;
+
    MetalBufferPool* bufferPool;
-    MetalBumpAllocator* bumpAllocator;

    MetalSwapChain* currentDrawSwapChain = nil;
    MetalSwapChain* currentReadSwapChain = nil;
@@ -196,7 +137,6 @@ struct MetalContext {

    // Empty texture used to prevent GPU errors when a sampler has been bound without a texture.
    id<MTLTexture> emptyTexture = nil;
-    id<MTLBuffer> emptyBuffer = nil;

    MetalBlitter* blitter = nullptr;

@@ -101,14 +101,9 @@ id<MTLCommandBuffer> getPendingCommandBuffer(MetalContext* context) {
    context->pendingCommandBuffer = [context->commandQueue commandBuffer];
    // It's safe for this block to capture the context variable. MetalDriver::terminate will ensure
    // all frames and their completion handlers finish before context is deallocated.
-    uint64_t thisCommandBufferId = context->pendingCommandBufferId;
    [context->pendingCommandBuffer addCompletedHandler:^(id <MTLCommandBuffer> buffer) {
        context->resourceTracker.clearResources((__bridge void*) buffer);
-
-        // Command buffers should complete in order, so latestCompletedCommandBufferId will only
-        // ever increase.
-        context->latestCompletedCommandBufferId = thisCommandBufferId;
-
+        
        auto errorCode = (MTLCommandBufferError)buffer.error.code;
        if (@available(macOS 11.0, *)) {
            if (errorCode == MTLCommandBufferErrorMemoryless) {
@@ -118,8 +113,7 @@ id<MTLCommandBuffer> getPendingCommandBuffer(MetalContext* context) {
            }
        }
    }];
-    FILAMENT_CHECK_POSTCONDITION(context->pendingCommandBuffer)
-            << "Could not obtain command buffer.";
+    ASSERT_POSTCONDITION(context->pendingCommandBuffer, "Could not obtain command buffer.");
    return context->pendingCommandBuffer;
 }

@@ -130,7 +124,6 @@ void submitPendingCommands(MetalContext* context) {
    assert_invariant(context->pendingCommandBuffer.status != MTLCommandBufferStatusCommitted);
    [context->pendingCommandBuffer commit];
    context->pendingCommandBuffer = nil;
-    context->pendingCommandBufferId++;
 }

 id<MTLTexture> getOrCreateEmptyTexture(MetalContext* context) {
@@ -160,67 +153,5 @@ bool isInRenderPass(MetalContext* context) {
    return context->currentRenderPassEncoder != nil;
 }

-void MetalPushConstantBuffer::setPushConstant(PushConstantVariant value, uint8_t index) {
-    if (mPushConstants.size() <= index) {
-        mPushConstants.resize(index + 1);
-        mDirty = true;
-    }
-    if (UTILS_LIKELY(mPushConstants[index] != value)) {
-        mDirty = true;
-        mPushConstants[index] = value;
-    }
-}
-
-void MetalPushConstantBuffer::setBytes(id<MTLCommandEncoder> encoder, ShaderStage stage) {
-    constexpr size_t PUSH_CONSTANT_SIZE_BYTES = 4;
-
-    static char buffer[MAX_PUSH_CONSTANT_COUNT * PUSH_CONSTANT_SIZE_BYTES];
-    assert_invariant(mPushConstants.size() <= MAX_PUSH_CONSTANT_COUNT);
-
-    size_t bufferSize = PUSH_CONSTANT_SIZE_BYTES * mPushConstants.size();
-    for (size_t i = 0; i < mPushConstants.size(); i++) {
-        const auto& constant = mPushConstants[i];
-        std::visit(
-                [i](auto arg) {
-                    if constexpr (std::is_same_v<decltype(arg), bool>) {
-                        // bool push constants are converted to uints in MSL.
-                        // We must ensure we write all the bytes for boolean values to work
-                        // correctly.
-                        uint32_t boolAsUint = arg ? 0x00000001 : 0x00000000;
-                        *(reinterpret_cast<uint32_t*>(buffer + PUSH_CONSTANT_SIZE_BYTES * i)) =
-                                boolAsUint;
-                    } else {
-                        *(decltype(arg)*)(buffer + PUSH_CONSTANT_SIZE_BYTES * i) = arg;
-                    }
-                },
-                constant);
-    }
-
-    switch (stage) {
-        case ShaderStage::VERTEX:
-            [(id<MTLRenderCommandEncoder>)encoder setVertexBytes:buffer
-                                                          length:bufferSize
-                                                         atIndex:PUSH_CONSTANT_BUFFER_INDEX];
-            break;
-        case ShaderStage::FRAGMENT:
-            [(id<MTLRenderCommandEncoder>)encoder setFragmentBytes:buffer
-                                                            length:bufferSize
-                                                           atIndex:PUSH_CONSTANT_BUFFER_INDEX];
-            break;
-        case ShaderStage::COMPUTE:
-            [(id<MTLComputeCommandEncoder>)encoder setBytes:buffer
-                                                     length:bufferSize
-                                                    atIndex:PUSH_CONSTANT_BUFFER_INDEX];
-            break;
-    }
-
-    mDirty = false;
-}
-
-void MetalPushConstantBuffer::clear() {
-    mPushConstants.clear();
-    mDirty = false;
-}
-
 } // namespace backend
 } // namespace filament
@@ -17,7 +17,6 @@
 #ifndef TNT_FILAMENT_DRIVER_METALDRIVER_H
 #define TNT_FILAMENT_DRIVER_METALDRIVER_H

-#include <backend/DriverEnums.h>
 #include "private/backend/Driver.h"
 #include "DriverBase.h"

@@ -32,7 +31,6 @@
 #include <functional>
 #include <mutex>
 #include <vector>
-#include <deque>

 namespace filament {
 namespace backend {
@@ -58,11 +56,11 @@ class MetalDriver final : public DriverBase {

 public:
    static Driver* create(MetalPlatform* platform, const Platform::DriverConfig& driverConfig);
+    void runAtNextTick(const std::function<void()>& fn) noexcept;

 private:

    friend class MetalSwapChain;
-    friend struct MetalDescriptorSet;

    MetalPlatform& mPlatform;
    MetalContext* mContext;
@@ -74,23 +72,10 @@ private:

    /*
     * Tasks run regularly on the driver thread.
-     * Not thread-safe; tasks are run from the driver thead and must be enqueued from the driver
-     * thread.
     */
-    void runAtNextTick(const std::function<void()>& fn) noexcept;
    void executeTickOps() noexcept;
    std::vector<std::function<void()>> mTickOps;
-
-    // Tasks regularly executed on the driver thread after a command buffer has completed
-    struct DeferredTask {
-        DeferredTask(uint64_t commandBufferId, utils::Invocable<void()>&& fn) noexcept
-            : commandBufferId(commandBufferId), fn(std::move(fn)) {}
-        uint64_t commandBufferId;     // after this command buffer completes
-        utils::Invocable<void()> fn;  // execute this task
-    };
-    void executeAfterCurrentCommandBufferCompletes(utils::Invocable<void()>&& fn) noexcept;
-    void executeDeferredOps() noexcept;
-    std::deque<DeferredTask> mDeferredTasks;
+    std::mutex mTickOpsLock;

    /*
     * Driver interface
@@ -151,10 +136,10 @@ private:
    inline void setRenderPrimitiveBuffer(Handle<HwRenderPrimitive> rph, PrimitiveType pt,
            Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh);

+    void finalizeSamplerGroup(MetalSamplerGroup* sg);
    void enumerateBoundBuffers(BufferObjectBinding bindingType,
            const std::function<void(const BufferState&, MetalBuffer*, uint32_t)>& f);

-    backend::StereoscopicType const mStereoscopicType;
 };

 } // namespace backend
@@ -71,7 +71,7 @@ constexpr inline MTLIndexType getIndexType(size_t elementSize) noexcept {
    } else if (elementSize == 4) {
        return MTLIndexTypeUInt32;
    }
-    FILAMENT_CHECK_POSTCONDITION(false) << "Index element size not supported.";
+    ASSERT_POSTCONDITION(false, "Index element size not supported.");
 }

 constexpr inline MTLVertexFormat getMetalFormat(ElementType type, bool normalized) noexcept {
@@ -100,7 +100,7 @@ constexpr inline MTLVertexFormat getMetalFormat(ElementType type, bool normalize
            case ElementType::SHORT4: return MTLVertexFormatShort4Normalized;
            case ElementType::USHORT4: return MTLVertexFormatUShort4Normalized;
            default:
-                FILAMENT_CHECK_POSTCONDITION(false) << "Normalized format does not exist.";
+                ASSERT_POSTCONDITION(false, "Normalized format does not exist.");
                return MTLVertexFormatInvalid;
        }
    }
@@ -326,8 +326,7 @@ constexpr inline MTLCullMode getMetalCullMode(CullingMode cullMode) noexcept {
        case CullingMode::FRONT: return MTLCullModeFront;
        case CullingMode::BACK: return MTLCullModeBack;
        case CullingMode::FRONT_AND_BACK:
-            FILAMENT_CHECK_POSTCONDITION(false)
-                    << "FRONT_AND_BACK culling is not supported in Metal.";
+            ASSERT_POSTCONDITION(false, "FRONT_AND_BACK culling is not supported in Metal.");
    }
 }

@@ -32,75 +32,100 @@ struct MetalContext;
 * texture.
 */
 class MetalExternalImage {
+
 public:
-    MetalExternalImage() = default;

-    MetalExternalImage(MetalExternalImage&&);
-    MetalExternalImage& operator=(MetalExternalImage&&);
-    ~MetalExternalImage() noexcept;
-
-    MetalExternalImage(const MetalExternalImage&) = delete;
-    MetalExternalImage& operator=(const MetalExternalImage&) = delete;
+    MetalExternalImage(MetalContext& context,
+            TextureSwizzle r = TextureSwizzle::CHANNEL_0,
+            TextureSwizzle g = TextureSwizzle::CHANNEL_1,
+            TextureSwizzle b = TextureSwizzle::CHANNEL_2,
+            TextureSwizzle a = TextureSwizzle::CHANNEL_3) noexcept;

    /**
-     * While the texture is used for rendering, this MetalExternalImage must be kept alive.
+     * @return true, if this MetalExternalImage is holding a live external image. Returns false
+     * until set has been called with a valid CVPixelBuffer. The image can be cleared via
+     * set(nullptr), and isValid will return false again.
     */
-    id<MTLTexture> getMtlTexture() const noexcept;
-
-    bool isValid() const noexcept {
-        return mImage != nil || mRgbTexture != nullptr;
-    }
-
-    NSUInteger getWidth() const noexcept;
-    NSUInteger getHeight() const noexcept;
+    bool isValid() const noexcept;

    /**
-     * Create an external image with the passed-in CVPixelBuffer.
+     * Set this external image to the passed-in CVPixelBuffer. Future calls to
+     * getMetalTextureForDraw will return a texture backed by this CVPixelBuffer. Previous
+     * CVPixelBuffers and related resources will be released when all GPU work using them has
+     * finished.
     *
-     * Ownership is taken of the CVPixelBuffer, which will be released when the returned
-     * MetalExternalImage is destroyed (or, in the case of a YCbCr image, after the conversion has
-     * completed).
-     *
-     * Calling set with a YCbCr image will encode a compute pass to convert the image from
-     * YCbCr to RGB.
+     * Calling set with a YCbCr image will encode a compute pass to convert the image from YCbCr to
+     * RGB.
     */
-    static MetalExternalImage createFromImage(MetalContext& context, CVPixelBufferRef image);
+    void set(CVPixelBufferRef image) noexcept;

    /**
-     * Create an external image with a specific plane of the passed-in CVPixelBuffer.
-     *
-     * Ownership is taken of the CVPixelBuffer, which will be released when the returned
-     * MetalExternalImage is destroyed.
+     * Set this external image to a specific plane of the passed-in CVPixelBuffer. Future calls to
+     * getMetalTextureForDraw will return a texture backed by a single plane of this CVPixelBuffer.
+     * Previous CVPixelBuffers and related resources will be released when all GPU work using them
+     * has finished.
     */
-    static MetalExternalImage createFromImagePlane(
-            MetalContext& context, CVPixelBufferRef image, uint32_t plane);
+    void set(CVPixelBufferRef image, size_t plane) noexcept;

-    static void assertWritableImage(CVPixelBufferRef image);
+    /**
+     * Returns the width of the external image, or 0 if one is not set. For YCbCr images, returns
+     * the width of the luminance plane.
+     */
+    size_t getWidth() const noexcept { return mWidth; }
+
+    /**
+     * Returns the height of the external image, or 0 if one is not set. For YCbCr images, returns
+     * the height of the luminance plane.
+     */
+    size_t getHeight() const noexcept { return mHeight; }
+
+    /**
+     * Get a Metal texture used to draw this image and denote that it is used for the current frame.
+     * For future frames that use this external image, getMetalTextureForDraw must be called again.
+     */
+    id<MTLTexture> getMetalTextureForDraw() const noexcept;

    /**
     * Free resources. Should be called at least once when no further calls to set will occur.
     */
    static void shutdown(MetalContext& context) noexcept;

-private:
-    MetalExternalImage(CVPixelBufferRef image, CVMetalTextureRef texture) noexcept
-        : mImage(image), mTexture(texture) {}
-    explicit MetalExternalImage(id<MTLTexture> texture) noexcept : mRgbTexture(texture) {}
+    static void assertWritableImage(CVPixelBufferRef image);

-    static id<MTLTexture> createRgbTexture(id<MTLDevice> device, size_t width, size_t height);
-    static CVMetalTextureRef createTextureFromImage(CVMetalTextureCacheRef textureCache,
-            CVPixelBufferRef image, MTLPixelFormat format, size_t plane);
-    static void ensureComputePipelineState(MetalContext& context);
-    static id<MTLCommandBuffer> encodeColorConversionPass(MetalContext& context,
-            id<MTLTexture> inYPlane, id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture);
+private:
+
+    void unset();
+
+    CVMetalTextureRef createTextureFromImage(CVPixelBufferRef image, MTLPixelFormat format,
+            size_t plane);
+    id<MTLTexture> createRgbTexture(size_t width, size_t height);
+    id<MTLTexture> createSwizzledTextureView(id<MTLTexture> texture) const;
+    id<MTLTexture> createSwizzledTextureView(CVMetalTextureRef texture) const;
+    void ensureComputePipelineState();
+    id<MTLCommandBuffer> encodeColorConversionPass(id<MTLTexture> inYPlane, id<MTLTexture>
+            inCbCrTexture, id<MTLTexture> outTexture);

    static constexpr size_t Y_PLANE = 0;
    static constexpr size_t CBCR_PLANE = 1;

-    // TODO: this could probably be a union.
+    MetalContext& mContext;
+
+    // If the external image has a single plane, mImage and mTexture hold references to the image
+    // and created Metal texture, respectively.
+    // mTextureView is a view of mTexture with any swizzling applied.
    CVPixelBufferRef mImage = nullptr;
    CVMetalTextureRef mTexture = nullptr;
+    id<MTLTexture> mTextureView = nullptr;
+    size_t mWidth = 0;
+    size_t mHeight = 0;
+
+    // If the external image is in the YCbCr format, this holds the result of the converted RGB
+    // texture.
    id<MTLTexture> mRgbTexture = nil;
+
+    struct {
+        TextureSwizzle r, g, b, a;
+    } mSwizzle;
 };

 } // namespace backend
@@ -29,11 +29,15 @@
        auto description = [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding]; \
        utils::slog.e << description << utils::io::endl;                                           \
    }                                                                                              \
-    FILAMENT_CHECK_POSTCONDITION(error == nil) << message;
+    ASSERT_POSTCONDITION(error == nil, message);

 namespace filament {
 namespace backend {

+static const auto cvBufferDeleter = [](const void* buffer) {
+    CVBufferRelease((CVMetalTextureRef) buffer);
+};
+
 static const char* kernel = R"(
 #include <metal_stdlib>
 #include <simd/simd.h>
@@ -67,66 +71,55 @@ ycbcrToRgb(texture2d<half, access::read>  inYTexture    [[texture(0)]],
 }
 )";

-NSUInteger MetalExternalImage::getWidth() const noexcept {
-    if (mImage) {
-        return CVPixelBufferGetWidth(mImage);
-    }
-    if (mRgbTexture) {
-        return mRgbTexture.width;
-    }
-    return 0;
+MetalExternalImage::MetalExternalImage(MetalContext& context, TextureSwizzle r, TextureSwizzle g,
+        TextureSwizzle b, TextureSwizzle a) noexcept : mContext(context), mSwizzle{r, g, b, a} { }
+
+bool MetalExternalImage::isValid() const noexcept {
+    return mRgbTexture != nil || mImage != nullptr;
 }

-NSUInteger MetalExternalImage::getHeight() const noexcept {
-    if (mImage) {
-        return CVPixelBufferGetHeight(mImage);
-    }
-    if (mRgbTexture) {
-        return mRgbTexture.height;
-    }
-    return 0;
-}
+void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
+    unset();

-MetalExternalImage MetalExternalImage::createFromImage(
-        MetalContext& context, CVPixelBufferRef image) {
    if (!image) {
-        return {};
+        return;
    }

    OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_POSTCONDITION(formatType == kCVPixelFormatType_32BGRA ||
-            formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange)
-            << "Metal external images must be in either 32BGRA or 420f format.";
+    ASSERT_POSTCONDITION(formatType == kCVPixelFormatType_32BGRA ||
+                         formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange,
+            "Metal external images must be in either 32BGRA or 420f format.");

    size_t planeCount = CVPixelBufferGetPlaneCount(image);
-    FILAMENT_CHECK_POSTCONDITION(planeCount == 0 || planeCount == 2)
-            << "The Metal backend does not support images with plane counts of " << planeCount
-            << ".";
+    ASSERT_POSTCONDITION(planeCount == 0 || planeCount == 2,
+            "The Metal backend does not support images with plane counts of %d.", planeCount);
+

    if (planeCount == 0) {
-        CVMetalTextureRef texture =
-                createTextureFromImage(context.textureCache, image, MTLPixelFormatBGRA8Unorm, 0);
-        return { CVPixelBufferRetain(image), texture };
+        mImage = image;
+        mTexture = createTextureFromImage(image, MTLPixelFormatBGRA8Unorm, 0);
+        mTextureView = createSwizzledTextureView(mTexture);
+        mWidth = CVPixelBufferGetWidth(image);
+        mHeight = CVPixelBufferGetHeight(image);
    }

    if (planeCount == 2) {
-        CVPixelBufferRetain(image);
-
-        CVMetalTextureRef yPlane =
-                createTextureFromImage(context.textureCache, image, MTLPixelFormatR8Unorm, Y_PLANE);
-        CVMetalTextureRef cbcrPlane =
-                createTextureFromImage(context.textureCache, image, MTLPixelFormatRG8Unorm, CBCR_PLANE);
+        CVMetalTextureRef yPlane = createTextureFromImage(image, MTLPixelFormatR8Unorm, Y_PLANE);
+        CVMetalTextureRef cbcrPlane = createTextureFromImage(image, MTLPixelFormatRG8Unorm,
+                CBCR_PLANE);

        // Get the size of luminance plane.
-        NSUInteger width = CVPixelBufferGetWidthOfPlane(image, Y_PLANE);
-        NSUInteger height = CVPixelBufferGetHeightOfPlane(image, Y_PLANE);
+        mWidth = CVPixelBufferGetWidthOfPlane(image, Y_PLANE);
+        mHeight = CVPixelBufferGetHeightOfPlane(image, Y_PLANE);

-        id<MTLTexture> rgbTexture = createRgbTexture(context.device, width, height);
-        id<MTLCommandBuffer> commandBuffer = encodeColorConversionPass(context,
+        id<MTLTexture> rgbTexture = createRgbTexture(mWidth, mHeight);
+        id<MTLCommandBuffer> commandBuffer = encodeColorConversionPass(
                CVMetalTextureGetTexture(yPlane),
                CVMetalTextureGetTexture(cbcrPlane),
                rgbTexture);

+        mRgbTexture = createSwizzledTextureView(rgbTexture);
+
        [commandBuffer addCompletedHandler:^(id <MTLCommandBuffer> o) {
            CVBufferRelease(yPlane);
            CVBufferRelease(cbcrPlane);
@@ -134,85 +127,72 @@ MetalExternalImage MetalExternalImage::createFromImage(
        }];

        [commandBuffer commit];
-        return MetalExternalImage { rgbTexture };
    }
-
-    return {};
 }

-MetalExternalImage MetalExternalImage::createFromImagePlane(
-        MetalContext& context, CVPixelBufferRef image, uint32_t plane) {
+void MetalExternalImage::set(CVPixelBufferRef image, size_t plane) noexcept {
+    unset();
+
    if (!image) {
-        return {};
+        return;
    }

    const OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_POSTCONDITION(formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange)
-            << "Metal planar external images must be in the 420f format.";
-    FILAMENT_CHECK_POSTCONDITION(plane == 0 || plane == 1)
-            << "Metal planar external images must be created from planes 0 or 1.";
+    ASSERT_POSTCONDITION(formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange,
+            "Metal planar external images must be in the 420f format.");

-    auto getPlaneFormat = [](size_t plane) {
-        // Right now Metal only supports kCVPixelFormatType_420YpCbCr8BiPlanarFullRange planar
-        // external images, so we can make the following assumptions about the format of each plane.
-        if (plane == 0) {
-            return MTLPixelFormatR8Unorm;  // luminance
-        }
-        if (plane == 1) {
-            return MTLPixelFormatRG8Unorm;  // CbCr
-        }
-        return MTLPixelFormatInvalid;
+    mImage = image;
+
+    auto getPlaneFormat = [] (size_t plane) {
+      // Right now Metal only supports kCVPixelFormatType_420YpCbCr8BiPlanarFullRange planar
+      // external images, so we can make the following assumptions about the format of each plane.
+      if (plane == 0) {
+        return MTLPixelFormatR8Unorm; // luminance
+      }
+      if (plane == 1) {
+        // CbCr
+        return MTLPixelFormatRG8Unorm; // CbCr
+      }
+      return MTLPixelFormatInvalid;
    };

    const MTLPixelFormat format = getPlaneFormat(plane);
    assert_invariant(format != MTLPixelFormatInvalid);
-    CVMetalTextureRef mTexture = createTextureFromImage(context.textureCache, image, format, plane);
-    return { CVPixelBufferRetain(image), mTexture };
+    mTexture = createTextureFromImage(image, format, plane);
+    mTextureView = createSwizzledTextureView(mTexture);
 }

-MetalExternalImage::MetalExternalImage(MetalExternalImage&& rhs) {
-    std::swap(mImage, rhs.mImage);
-    std::swap(mTexture, rhs.mTexture);
-    std::swap(mRgbTexture, rhs.mRgbTexture);
-}
-
-MetalExternalImage& MetalExternalImage::operator=(MetalExternalImage&& rhs) {
-    CVPixelBufferRelease(mImage);
-    CVBufferRelease(mTexture);
-    mImage = nullptr;
-    mTexture = nullptr;
-    mRgbTexture = nullptr;
-    std::swap(mImage, rhs.mImage);
-    std::swap(mTexture, rhs.mTexture);
-    std::swap(mRgbTexture, rhs.mRgbTexture);
-    return *this;
-}
-
-MetalExternalImage::~MetalExternalImage() noexcept {
-    CVPixelBufferRelease(mImage);
-    CVBufferRelease(mTexture);
-}
-
-id<MTLTexture> MetalExternalImage::getMtlTexture() const noexcept {
+id<MTLTexture> MetalExternalImage::getMetalTextureForDraw() const noexcept {
    if (mRgbTexture) {
        return mRgbTexture;
    }
-    if (mTexture) {
-        return CVMetalTextureGetTexture(mTexture);
+
+    // Retain the image and Metal texture until the GPU has finished with this frame. This does
+    // not need to be done for the RGB texture, because it is an Objective-C object whose
+    // lifetime is automatically managed by Metal.
+    auto& tracker = mContext.resourceTracker;
+    auto commandBuffer = getPendingCommandBuffer(&mContext);
+    if (tracker.trackResource((__bridge void*) commandBuffer, mImage, cvBufferDeleter)) {
+        CVPixelBufferRetain(mImage);
    }
-    return nil;
+    if (tracker.trackResource((__bridge void*) commandBuffer, mTexture, cvBufferDeleter)) {
+        CVBufferRetain(mTexture);
+    }
+
+    assert_invariant(mTextureView);
+    return mTextureView;
 }

-CVMetalTextureRef MetalExternalImage::createTextureFromImage(CVMetalTextureCacheRef textureCache,
-        CVPixelBufferRef image, MTLPixelFormat format, size_t plane) {
+CVMetalTextureRef MetalExternalImage::createTextureFromImage(CVPixelBufferRef image,
+        MTLPixelFormat format, size_t plane) {
    const size_t width = CVPixelBufferGetWidthOfPlane(image, plane);
    const size_t height = CVPixelBufferGetHeightOfPlane(image, plane);

    CVMetalTextureRef texture;
-    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault, textureCache,
-            image, nullptr, format, width, height, plane, &texture);
-    FILAMENT_CHECK_POSTCONDITION(result == kCVReturnSuccess)
-            << "Could not create a CVMetalTexture from CVPixelBuffer.";
+    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault,
+            mContext.textureCache, image, nullptr, format, width, height, plane, &texture);
+    ASSERT_POSTCONDITION(result == kCVReturnSuccess,
+            "Could not create a CVMetalTexture from CVPixelBuffer.");

    return texture;
 }
@@ -221,19 +201,58 @@ void MetalExternalImage::shutdown(MetalContext& context) noexcept {
    context.externalImageComputePipelineState = nil;
 }

-id<MTLTexture> MetalExternalImage::createRgbTexture(
-        id<MTLDevice> device, size_t width, size_t height) {
+void MetalExternalImage::assertWritableImage(CVPixelBufferRef image) {
+    OSType formatType = CVPixelBufferGetPixelFormatType(image);
+    ASSERT_PRECONDITION(formatType == kCVPixelFormatType_32BGRA,
+            "Metal SwapChain images must be in the 32BGRA format.");
+}
+
+void MetalExternalImage::unset() {
+    CVPixelBufferRelease(mImage);
+    CVBufferRelease(mTexture);
+
+    mImage = nullptr;
+    mTexture = nullptr;
+    mTextureView = nil;
+    mRgbTexture = nil;
+    mWidth = 0;
+    mHeight = 0;
+}
+
+id<MTLTexture> MetalExternalImage::createRgbTexture(size_t width, size_t height) {
    MTLTextureDescriptor *descriptor =
            [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
                                                               width:width
                                                              height:height
                                                           mipmapped:NO];
    descriptor.usage = MTLTextureUsageShaderWrite | MTLTextureUsageShaderRead;
-    return [device newTextureWithDescriptor:descriptor];
+    return [mContext.device newTextureWithDescriptor:descriptor];
 }

-void MetalExternalImage::ensureComputePipelineState(MetalContext& context) {
-    if (context.externalImageComputePipelineState != nil) {
+id<MTLTexture> MetalExternalImage::createSwizzledTextureView(id<MTLTexture> texture) const {
+    const bool isDefaultSwizzle =
+            mSwizzle.r == TextureSwizzle::CHANNEL_0 &&
+            mSwizzle.g == TextureSwizzle::CHANNEL_1 &&
+            mSwizzle.b == TextureSwizzle::CHANNEL_2 &&
+            mSwizzle.a == TextureSwizzle::CHANNEL_3;
+    if (!isDefaultSwizzle && mContext.supportsTextureSwizzling) {
+        // Even though we've already checked supportsTextureSwizzling, we still need to guard these
+        // calls with @availability, otherwise the API usage will generate compiler warnings.
+        if (@available(iOS 13, *)) {
+            texture = createTextureViewWithSwizzle(texture,
+                    getSwizzleChannels(mSwizzle.r, mSwizzle.g, mSwizzle.b, mSwizzle.a));
+        }
+    }
+    return texture;
+}
+
+id<MTLTexture> MetalExternalImage::createSwizzledTextureView(CVMetalTextureRef ref) const {
+    id<MTLTexture> texture = CVMetalTextureGetTexture(ref);
+    return createSwizzledTextureView(texture);
+}
+
+void MetalExternalImage::ensureComputePipelineState() {
+    if (mContext.externalImageComputePipelineState != nil) {
        return;
    }

@@ -241,28 +260,29 @@ void MetalExternalImage::ensureComputePipelineState(MetalContext& context) {

    NSString* objcSource = [NSString stringWithCString:kernel
                                              encoding:NSUTF8StringEncoding];
-    id<MTLLibrary> library = [context.device newLibraryWithSource:objcSource
-                                                          options:nil
-                                                            error:&error];
+    id<MTLLibrary> library = [mContext.device newLibraryWithSource:objcSource
+                                                            options:nil
+                                                              error:&error];
    NSERROR_CHECK("Unable to compile Metal shading library.");

    id<MTLFunction> kernelFunction = [library newFunctionWithName:@"ycbcrToRgb"];

-    context.externalImageComputePipelineState =
-            [context.device newComputePipelineStateWithFunction:kernelFunction error:&error];
+    mContext.externalImageComputePipelineState =
+            [mContext.device newComputePipelineStateWithFunction:kernelFunction
+                                                           error:&error];
    NSERROR_CHECK("Unable to create Metal compute pipeline state.");
 }

-id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(MetalContext& context,
-        id<MTLTexture> inYPlane, id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture) {
-    ensureComputePipelineState(context);
+id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(id<MTLTexture> inYPlane,
+        id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture) {
+    ensureComputePipelineState();

-    id<MTLCommandBuffer> commandBuffer = [context.commandQueue commandBuffer];
+    id<MTLCommandBuffer> commandBuffer = [mContext.commandQueue commandBuffer];
    commandBuffer.label = @"YCbCr to RGB conversion";

    id<MTLComputeCommandEncoder> computeEncoder = [commandBuffer computeCommandEncoder];

-    [computeEncoder setComputePipelineState:context.externalImageComputePipelineState];
+    [computeEncoder setComputePipelineState:mContext.externalImageComputePipelineState];
    [computeEncoder setTexture:inYPlane atIndex:0];
    [computeEncoder setTexture:inCbCrTexture atIndex:1];
    [computeEncoder setTexture:outTexture atIndex:2];
@@ -280,11 +300,5 @@ id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(MetalContext&
    return commandBuffer;
 }

-void MetalExternalImage::assertWritableImage(CVPixelBufferRef image) {
-    OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_PRECONDITION(formatType == kCVPixelFormatType_32BGRA)
-            << "Metal SwapChain images must be in the 32BGRA format.";
-}
-
 } // namespace backend
 } // namespace filament
@@ -44,7 +44,6 @@
 #include <condition_variable>
 #include <memory>
 #include <type_traits>
-#include <vector>

 namespace filament {
 namespace backend {
@@ -74,8 +73,7 @@ public:

    void releaseDrawable();

-    void setFrameScheduledCallback(
-            CallbackHandler* handler, FrameScheduledCallback&& callback, uint64_t flags);
+    void setFrameScheduledCallback(CallbackHandler* handler, FrameScheduledCallback&& callback);
    void setFrameCompletedCallback(
            CallbackHandler* handler, utils::Invocable<void(void)>&& callback);

@@ -86,8 +84,6 @@ public:
    NSUInteger getSurfaceWidth() const;
    NSUInteger getSurfaceHeight() const;

-    bool isPixelBuffer() const { return type == SwapChainType::CVPIXELBUFFERREF; }
-
 private:

    enum class SwapChainType {
@@ -97,6 +93,7 @@ private:
    };
    bool isCaMetalLayer() const { return type == SwapChainType::CAMETALLAYER; }
    bool isHeadless() const { return type == SwapChainType::HEADLESS; }
+    bool isPixelBuffer() const { return type == SwapChainType::CVPIXELBUFFERREF; }

    void scheduleFrameScheduledCallback();
    void scheduleFrameCompletedCallback();
@@ -112,7 +109,6 @@ private:
    NSUInteger headlessWidth = 0;
    NSUInteger headlessHeight = 0;
    CAMetalLayer* layer = nullptr;
-    std::shared_ptr<std::mutex> layerDrawableMutex;
    MetalExternalImage externalImage;
    SwapChainType type;

@@ -124,7 +120,6 @@ private:
    struct {
        CallbackHandler* handler = nullptr;
        std::shared_ptr<FrameScheduledCallback> callback = nullptr;
-        uint64_t flags = 0;
    } frameScheduled;

    struct {
@@ -135,17 +130,19 @@ private:

 class MetalBufferObject : public HwBufferObject {
 public:
-
-    using TagResolver = MetalBuffer::TagResolver;
-
    MetalBufferObject(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
         uint32_t byteCount);

-    void updateBuffer(void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag);
-    void updateBufferUnsynchronized(
-            void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag);
+    void updateBuffer(void* data, size_t size, uint32_t byteOffset);
+    void updateBufferUnsynchronized(void* data, size_t size, uint32_t byteOffset);
    MetalBuffer* getBuffer() { return &buffer; }

+    // Tracks which uniform/ssbo buffers this buffer object is bound into.
+    static_assert(Program::UNIFORM_BINDING_COUNT <= 32);
+    static_assert(MAX_SSBO_COUNT <= 32);
+    utils::bitset32 boundUniformBuffers;
+    utils::bitset32 boundSsbos;
+
 private:
    MetalBuffer buffer;
 };
@@ -202,10 +199,12 @@ public:
    MetalProgram(MetalContext& context, Program&& program) noexcept;

    const MetalShaderCompiler::MetalFunctionBundle& getFunctions();
+    const Program::SamplerGroupInfo& getSamplerGroupInfo() { return samplerGroupInfo; }

 private:
    void initialize();

+    Program::SamplerGroupInfo samplerGroupInfo;
    MetalContext& mContext;
    MetalShaderCompiler::MetalFunctionBundle mFunctionBundle;
    MetalShaderCompiler::program_token_t mToken;
@@ -227,42 +226,43 @@ struct PixelBufferShape {
 class MetalTexture : public HwTexture {
 public:
    MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
-            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth,
-            TextureUsage usage) noexcept;
-
-    // constructors for creating texture views
-    MetalTexture(MetalContext& context, MetalTexture const* src, uint8_t baseLevel,
-            uint8_t levelCount) noexcept;
-    MetalTexture(MetalContext& context, MetalTexture const* src, TextureSwizzle r, TextureSwizzle g,
-            TextureSwizzle b, TextureSwizzle a) noexcept;
+            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
+            TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a)
+            noexcept;

    // Constructor for importing an id<MTLTexture> outside of Filament.
    MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
            id<MTLTexture> texture) noexcept;

-    // Constructors for importing external images.
-    MetalTexture(MetalContext& context, TextureFormat format, uint32_t width, uint32_t height,
-            TextureUsage usage, CVPixelBufferRef image) noexcept;
-    MetalTexture(MetalContext& context, TextureFormat format, uint32_t width, uint32_t height,
-            TextureUsage usage, CVPixelBufferRef image, uint32_t plane) noexcept;
+    ~MetalTexture();

-    // Returns an id<MTLTexture> suitable for reading in a shader, taking into account swizzle.
-    id<MTLTexture> getMtlTextureForRead() const noexcept;
+    // Returns an id<MTLTexture> suitable for reading in a shader, taking into account swizzle and
+    // LOD clamping.
+    id<MTLTexture> getMtlTextureForRead() noexcept;

    // Returns the id<MTLTexture> for attaching to a render pass.
-    id<MTLTexture> getMtlTextureForWrite() const noexcept {
+    id<MTLTexture> getMtlTextureForWrite() noexcept {
        return texture;
    }

-    std::shared_ptr<MetalExternalImage> getExternalImage() const noexcept { return externalImage; }
-
    void loadImage(uint32_t level, MTLRegion region, PixelBufferDescriptor& p) noexcept;
    void generateMipmaps() noexcept;

+    // A texture starts out with none of its mip levels (also referred to as LODs) available for
+    // reading. 4 actions update the range of LODs available:
+    // - calling loadImage
+    // - calling generateMipmaps
+    // - using the texture as a render target attachment
+    // - calling setMinMaxLevels
+    // A texture's available mips are consistent throughout a render pass.
+    void setLodRange(uint16_t minLevel, uint16_t maxLevel);
+    void extendLodRangeTo(uint16_t level);
+
    static MTLPixelFormat decidePixelFormat(MetalContext* context, TextureFormat format);

    MetalContext& context;
+    MetalExternalImage externalImage;

    // A "sidecar" texture used to implement automatic MSAA resolve.
    // This is created by MetalRenderTarget and stored here so it can be used with multiple
@@ -271,6 +271,26 @@ public:

    MTLPixelFormat devicePixelFormat;

+    // Frees memory associated with this texture and marks it as "terminated".
+    // Used to track "use after free" scenario.
+    void terminate() noexcept;
+    bool isTerminated() const noexcept { return terminated; }
+    inline void checkUseAfterFree(const char* samplerGroupDebugName, size_t textureIndex) const {
+        if (UTILS_LIKELY(!isTerminated())) {
+            return;
+        }
+        NSString* reason =
+                [NSString stringWithFormat:
+                                  @"Filament Metal texture use after free, sampler group = "
+                                  @"%s, texture index = %zu",
+                          samplerGroupDebugName, textureIndex];
+        NSException* useAfterFreeException =
+                [NSException exceptionWithName:@"MetalTextureUseAfterFree"
+                                        reason:reason
+                                      userInfo:nil];
+        [useAfterFreeException raise];
+    }
+
 private:
    void loadSlice(uint32_t level, MTLRegion region, uint32_t byteOffset, uint32_t slice,
            PixelBufferDescriptor const& data) noexcept;
@@ -281,12 +301,95 @@ private:

    id<MTLTexture> texture = nil;

-    std::shared_ptr<MetalExternalImage> externalImage;
-
    // If non-nil, a swizzled texture view to use instead of "texture".
    // Filament swizzling only affects texture reads, so this should not be used when the texture is
    // bound as a render target attachment.
    id<MTLTexture> swizzledTextureView = nil;
+    id<MTLTexture> lodTextureView = nil;
+
+    uint16_t minLod = std::numeric_limits<uint16_t>::max();
+    uint16_t maxLod = 0;
+
+    bool terminated = false;
+};
+
+class MetalSamplerGroup : public HwSamplerGroup {
+public:
+    explicit MetalSamplerGroup(size_t size, utils::FixedSizeString<32> name) noexcept
+        : size(size),
+          debugName(name),
+          textureHandles(size, Handle<HwTexture>()),
+          textures(size, nil),
+          samplers(size, nil) {}
+
+    inline void setTextureHandle(size_t index, Handle<HwTexture> th) {
+        assert_invariant(!finalized);
+        textureHandles[index] = th;
+    }
+
+    // This method is only used for debugging, to ensure all texture handles are alive.
+    const auto& getTextureHandles() const {
+        return textureHandles;
+    }
+
+    // Encode a MTLTexture into this SamplerGroup at the given index.
+    inline void setFinalizedTexture(size_t index, id<MTLTexture> t) {
+        assert_invariant(!finalized);
+        textures[index] = t;
+    }
+
+    // Encode a MTLSamplerState into this SamplerGroup at the given index.
+    inline void setFinalizedSampler(size_t index, id<MTLSamplerState> s) {
+        assert_invariant(!finalized);
+        samplers[index] = s;
+    }
+
+    // A SamplerGroup is "finalized" when all of its textures have been set and is ready for use in
+    // a draw call.
+    // Once a SamplerGroup is finalized, it must be reset or mutated to be written into again.
+    void finalize();
+    bool isFinalized() const noexcept { return finalized; }
+
+    // Both of these methods "unfinalize" a SamplerGroup, allowing it to be updated via calls to
+    // setFinalizedTexture or setFinalizedSampler. The difference is that when reset is called, all
+    // the samplers/textures must be rebound. The MTLArgumentEncoder must be specified, in case
+    // the texture types have changed.
+    // Mutate re-encodes the current set of samplers/textures into the new argument
+    // buffer.
+    void reset(id<MTLCommandBuffer> cmdBuffer, id<MTLArgumentEncoder> e, id<MTLDevice> device);
+    void mutate(id<MTLCommandBuffer> cmdBuffer);
+
+    id<MTLBuffer> getArgumentBuffer() const {
+        assert_invariant(finalized);
+        return argBuffer->getCurrentAllocation().first;
+    }
+
+    NSUInteger getArgumentBufferOffset() const {
+        return argBuffer->getCurrentAllocation().second;
+    }
+
+    inline std::pair<Handle<HwTexture>, id<MTLTexture>> getFinalizedTexture(size_t index) {
+        return {textureHandles[index], textures[index]};
+    }
+
+    // Calls the Metal useResource:usage:stages: method for all the textures in this SamplerGroup.
+    void useResources(id<MTLRenderCommandEncoder> renderPassEncoder);
+
+    size_t size;
+    utils::FixedSizeString<32> debugName;
+
+public:
+
+    // These vectors are kept in sync with one another.
+    utils::FixedCapacityVector<Handle<HwTexture>> textureHandles;
+    utils::FixedCapacityVector<id<MTLTexture>> textures;
+    utils::FixedCapacityVector<id<MTLSamplerState>> samplers;
+
+    id<MTLArgumentEncoder> encoder;
+
+    std::unique_ptr<MetalRingBuffer> argBuffer = nullptr;
+
+    bool finalized = false;
 };

 class MetalRenderTarget : public HwRenderTarget {
@@ -443,61 +546,6 @@ struct MetalTimerQuery : public HwTimerQuery {
    std::shared_ptr<Status> status;
 };

-class MetalDescriptorSetLayout : public HwDescriptorSetLayout {
-public:
-    MetalDescriptorSetLayout(DescriptorSetLayout&& layout) noexcept;
-
-    const auto& getBindings() const noexcept { return mLayout.bindings; }
-
-    size_t getDynamicOffsetCount() const noexcept { return mDynamicOffsetCount; }
-
-    /**
-     * Get an argument encoder for this descriptor set and shader stage.
-     * textureTypes should only include the textures present in the corresponding shader stage.
-     */
-    id<MTLArgumentEncoder> getArgumentEncoder(id<MTLDevice> device, ShaderStage stage,
-            utils::FixedCapacityVector<MTLTextureType> const& textureTypes);
-
-private:
-    id<MTLArgumentEncoder> getArgumentEncoderSlow(id<MTLDevice> device, ShaderStage stage,
-            utils::FixedCapacityVector<MTLTextureType> const& textureTypes);
-
-    DescriptorSetLayout mLayout;
-    size_t mDynamicOffsetCount = 0;
-    std::array<id<MTLArgumentEncoder>, Program::SHADER_TYPE_COUNT> mCachedArgumentEncoder = { nil };
-    std::array<utils::FixedCapacityVector<MTLTextureType>, Program::SHADER_TYPE_COUNT>
-            mCachedTextureTypes;
-};
-
-struct MetalDescriptorSet : public HwDescriptorSet {
-    MetalDescriptorSet(MetalDescriptorSetLayout* layout) noexcept;
-
-    void finalize(MetalDriver* driver);
-
-    id<MTLBuffer> finalizeAndGetBuffer(MetalDriver* driver, ShaderStage stage);
-
-    MetalDescriptorSetLayout* layout;
-
-    struct BufferBinding {
-        id<MTLBuffer> buffer;
-        uint32_t offset;
-        uint32_t size;
-    };
-    struct TextureBinding {
-        id<MTLTexture> texture;
-        SamplerParams sampler;
-    };
-    tsl::robin_map<descriptor_binding_t, BufferBinding> buffers;
-    tsl::robin_map<descriptor_binding_t, TextureBinding> textures;
-
-    std::vector<id<MTLResource>> vertexResources;
-    std::vector<id<MTLResource>> fragmentResources;
-
-    std::vector<std::shared_ptr<MetalExternalImage>> externalImages;
-
-    std::array<TrackedMetalBuffer, Program::SHADER_TYPE_COUNT> cachedBuffer = { nil };
-};
-
 } // namespace backend
 } // namespace filament

@@ -73,7 +73,7 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, CAMetalLayer* nativeWindow
    : context(context),
      depthStencilFormat(decideDepthStencilFormat(flags)),
      layer(nativeWindow),
-      layerDrawableMutex(std::make_shared<std::mutex>()),
+      externalImage(context),
      type(SwapChainType::CAMETALLAYER) {

    if (!(flags & SwapChain::CONFIG_TRANSPARENT) && !nativeWindow.opaque) {
@@ -99,15 +99,17 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, int32_t width, int32_t hei
      depthStencilFormat(decideDepthStencilFormat(flags)),
      headlessWidth(width),
      headlessHeight(height),
+      externalImage(context),
      type(SwapChainType::HEADLESS) {}

 MetalSwapChain::MetalSwapChain(MetalContext& context, CVPixelBufferRef pixelBuffer, uint64_t flags)
    : context(context),
      depthStencilFormat(decideDepthStencilFormat(flags)),
-      externalImage(MetalExternalImage::createFromImage(context, pixelBuffer)),
+      externalImage(context),
      type(SwapChainType::CVPIXELBUFFERREF) {
    assert_invariant(flags & SWAP_CHAIN_CONFIG_APPLE_CVPIXELBUFFER);
    MetalExternalImage::assertWritableImage(pixelBuffer);
+    externalImage.set(pixelBuffer);
    assert_invariant(externalImage.isValid());
 }

@@ -118,6 +120,7 @@ MTLPixelFormat MetalSwapChain::decideDepthStencilFormat(uint64_t flags) {
 }

 MetalSwapChain::~MetalSwapChain() {
+    externalImage.set(nullptr);
 }

 NSUInteger MetalSwapChain::getSurfaceWidth() const {
@@ -167,28 +170,18 @@ id<MTLTexture> MetalSwapChain::acquireDrawable() {
    }

    if (isPixelBuffer()) {
-        return externalImage.getMtlTexture();
+        return externalImage.getMetalTextureForDraw();
    }

    assert_invariant(isCaMetalLayer());
+    drawable = [layer nextDrawable];

-    // CAMetalLayer's drawable pool is not thread safe. Use a mutex when
-    // calling -nextDrawable, or when releasing the last known reference
-    // to any CAMetalDrawable returned from a previous -nextDrawable.
-    {
-        std::lock_guard<std::mutex> lock(*layerDrawableMutex);
-        drawable = [layer nextDrawable];
-    }
-
-    FILAMENT_CHECK_POSTCONDITION(drawable != nil) << "Could not obtain drawable.";
+    ASSERT_POSTCONDITION(drawable != nil, "Could not obtain drawable.");
    return drawable.texture;
 }

 void MetalSwapChain::releaseDrawable() {
-    if (drawable) {
-        std::lock_guard<std::mutex> lock(*layerDrawableMutex);
-        drawable = nil;
-    }
+    drawable = nil;
 }

 id<MTLTexture> MetalSwapChain::acquireDepthTexture() {
@@ -229,10 +222,9 @@ void MetalSwapChain::ensureDepthStencilTexture() {
 }

 void MetalSwapChain::setFrameScheduledCallback(
-        CallbackHandler* handler, FrameScheduledCallback&& callback, uint64_t flags) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
    frameScheduled.handler = handler;
    frameScheduled.callback = std::make_shared<FrameScheduledCallback>(std::move(callback));
-    frameScheduled.flags = flags;
 }

 void MetalSwapChain::setFrameCompletedCallback(
@@ -254,17 +246,19 @@ void MetalSwapChain::present() {
    }
 }

+#ifndef FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD
+#define FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD 1
+#endif
+
 class PresentDrawableData {
 public:
    PresentDrawableData() = delete;
    PresentDrawableData(const PresentDrawableData&) = delete;
    PresentDrawableData& operator=(const PresentDrawableData&) = delete;

-    static PresentDrawableData* create(id<CAMetalDrawable> drawable,
-            std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver, uint64_t flags) {
-        assert_invariant(drawableMutex);
+    static PresentDrawableData* create(id<CAMetalDrawable> drawable, MetalDriver* driver) {
        assert_invariant(driver);
-        return new PresentDrawableData(drawable, drawableMutex, driver, flags);
+        return new PresentDrawableData(drawable, driver);
    }

    static void maybePresentAndDestroyAsync(PresentDrawableData* that, bool shouldPresent) {
@@ -272,38 +266,28 @@ public:
           [that->mDrawable present];
        }

-        if (that->mFlags & SwapChain::CALLBACK_DEFAULT_USE_METAL_COMPLETION_HANDLER) {
-            cleanupAndDestroy(that);
-        } else {
-            // mDrawable is acquired on the driver thread. Typically, we would release this object
-            // on the same thread, but after receiving consistent crash reports from within
-            // [CAMetalDrawable dealloc], we suspect this object requires releasing on the main
-            // thread.
-            dispatch_async(dispatch_get_main_queue(), ^{
-                cleanupAndDestroy(that);
-            });
-        }
+#if FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD == 1
+        // mDrawable is acquired on the driver thread. Typically, we would release this object on
+        // the same thread, but after receiving consistent crash reports from within
+        // [CAMetalDrawable dealloc], we suspect this object requires releasing on the main thread.
+        dispatch_async(dispatch_get_main_queue(), ^{ cleanupAndDestroy(that); });
+#else
+        that->mDriver->runAtNextTick([that]() { cleanupAndDestroy(that); });
+#endif
    }

 private:
-    PresentDrawableData(id<CAMetalDrawable> drawable, std::shared_ptr<std::mutex> drawableMutex,
-            MetalDriver* driver, uint64_t flags)
-        : mDrawable(drawable), mDrawableMutex(drawableMutex), mDriver(driver), mFlags(flags) {}
+    PresentDrawableData(id<CAMetalDrawable> drawable, MetalDriver* driver)
+        : mDrawable(drawable), mDriver(driver) {}

    static void cleanupAndDestroy(PresentDrawableData *that) {
-        if (that->mDrawable) {
-            std::lock_guard<std::mutex> lock(*(that->mDrawableMutex));
-            that->mDrawable = nil;
-        }
-        that->mDrawableMutex.reset();
+        that->mDrawable = nil;
        that->mDriver = nullptr;
        delete that;
    }

    id<CAMetalDrawable> mDrawable;
-    std::shared_ptr<std::mutex> mDrawableMutex;
    MetalDriver* mDriver = nullptr;
-    uint64_t mFlags = 0;
 };

 void presentDrawable(bool presentFrame, void* user) {
@@ -320,8 +304,8 @@ void MetalSwapChain::scheduleFrameScheduledCallback() {

    struct Callback {
        Callback(std::shared_ptr<FrameScheduledCallback> callback, id<CAMetalDrawable> drawable,
-                 std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver, uint64_t flags)
-            : f(callback), data(PresentDrawableData::create(drawable, drawableMutex, driver, flags)) {}
+                MetalDriver* driver)
+            : f(callback), data(PresentDrawableData::create(drawable, driver)) {}
        std::shared_ptr<FrameScheduledCallback> f;
        // PresentDrawableData* is destroyed by maybePresentAndDestroyAsync() later.
        std::unique_ptr<PresentDrawableData> data;
@@ -336,19 +320,14 @@ void MetalSwapChain::scheduleFrameScheduledCallback() {

    // This callback pointer will be captured by the block. Even if the scheduled handler is never
    // called, the unique_ptr will still ensure we don't leak memory.
-    uint64_t const flags = frameScheduled.flags;
-    __block auto callback = std::make_unique<Callback>(
-            frameScheduled.callback, drawable, layerDrawableMutex, context.driver, flags);
+    __block auto callback =
+            std::make_unique<Callback>(frameScheduled.callback, drawable, context.driver);

    backend::CallbackHandler* handler = frameScheduled.handler;
    MetalDriver* driver = context.driver;
    [getPendingCommandBuffer(&context) addScheduledHandler:^(id<MTLCommandBuffer> cb) {
        Callback* user = callback.release();
-        if (flags & SwapChain::CALLBACK_DEFAULT_USE_METAL_COMPLETION_HANDLER) {
-            Callback::func(user);
-        } else {
-            driver->scheduleCallback(handler, user, &Callback::func);
-        }
+        driver->scheduleCallback(handler, user, &Callback::func);
    }];
 }

@@ -383,14 +362,12 @@ MetalBufferObject::MetalBufferObject(MetalContext& context, BufferObjectBinding
        BufferUsage usage, uint32_t byteCount)
        : HwBufferObject(byteCount), buffer(context, bindingType, usage, byteCount) {}

-void MetalBufferObject::updateBuffer(
-        void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag) {
-    buffer.copyIntoBuffer(data, size, byteOffset, std::move(getHandleTag));
+void MetalBufferObject::updateBuffer(void* data, size_t size, uint32_t byteOffset) {
+    buffer.copyIntoBuffer(data, size, byteOffset);
 }

-void MetalBufferObject::updateBufferUnsynchronized(
-        void* data, size_t size, uint32_t byteOffset, TagResolver&& getHandleTag) {
-    buffer.copyIntoBufferUnsynchronized(data, size, byteOffset, std::move(getHandleTag));
+void MetalBufferObject::updateBufferUnsynchronized(void* data, size_t size, uint32_t byteOffset) {
+    buffer.copyIntoBufferUnsynchronized(data, size, byteOffset);
 }

 MetalVertexBufferInfo::MetalVertexBufferInfo(MetalContext& context, uint8_t bufferCount,
@@ -486,6 +463,11 @@ void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,

 MetalProgram::MetalProgram(MetalContext& context, Program&& program) noexcept
    : HwProgram(program.getName()), mContext(context) {
+
+    // Save this program's SamplerGroupInfo, it's used during draw calls to bind sampler groups to
+    // the appropriate stage(s).
+    samplerGroupInfo = program.getSamplerGroupInfo();
+
    mToken = context.shaderCompiler->createProgram(program.getName(), std::move(program));
    assert_invariant(mToken);
 }
@@ -505,21 +487,21 @@ void MetalProgram::initialize() {

 MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t levels,
        TextureFormat format, uint8_t samples, uint32_t width, uint32_t height, uint32_t depth,
-        TextureUsage usage) noexcept
-    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context) {
-    assert_invariant(target != SamplerType::SAMPLER_EXTERNAL);
+        TextureUsage usage, TextureSwizzle r, TextureSwizzle g, TextureSwizzle b,
+        TextureSwizzle a) noexcept
+    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
+        externalImage(context, r, g, b, a) {

    devicePixelFormat = decidePixelFormat(&context, format);
-    FILAMENT_CHECK_POSTCONDITION(devicePixelFormat != MTLPixelFormatInvalid)
-            << "Texture format not supported.";
+    ASSERT_POSTCONDITION(devicePixelFormat != MTLPixelFormatInvalid, "Texture format not supported.");

    const BOOL mipmapped = levels > 1;
    const BOOL multisampled = samples > 1;

 #if defined(IOS)
    const BOOL textureArray = target == SamplerType::SAMPLER_2D_ARRAY;
-    FILAMENT_CHECK_PRECONDITION(!textureArray || !multisampled)
-            << "iOS does not support multisampled texture arrays.";
+    ASSERT_PRECONDITION(!textureArray || !multisampled,
+            "iOS does not support multisampled texture arrays.");
 #endif

    const auto get2DTextureType = [](SamplerType target, bool isMultisampled) {
@@ -554,12 +536,12 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
            descriptor.usage = getMetalTextureUsage(usage);
            descriptor.storageMode = MTLStorageModePrivate;
            texture = [context.device newTextureWithDescriptor:descriptor];
+            ASSERT_POSTCONDITION(texture != nil, "Could not create Metal texture. Out of memory?");
            break;
        case SamplerType::SAMPLER_CUBEMAP:
        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
-            FILAMENT_CHECK_POSTCONDITION(!multisampled)
-                    << "Multisampled cubemap faces not supported.";
-            FILAMENT_CHECK_POSTCONDITION(width == height) << "Cubemap faces must be square.";
+            ASSERT_POSTCONDITION(!multisampled, "Multisampled cubemap faces not supported.");
+            ASSERT_POSTCONDITION(width == height, "Cubemap faces must be square.");
            descriptor = [MTLTextureDescriptor textureCubeDescriptorWithPixelFormat:devicePixelFormat
                                                                               size:width
                                                                          mipmapped:mipmapped];
@@ -568,6 +550,7 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
            descriptor.usage = getMetalTextureUsage(usage);
            descriptor.storageMode = MTLStorageModePrivate;
            texture = [context.device newTextureWithDescriptor:descriptor];
+            ASSERT_POSTCONDITION(texture != nil, "Could not create Metal texture. Out of memory?");
            break;
        case SamplerType::SAMPLER_3D:
            descriptor = [MTLTextureDescriptor new];
@@ -580,6 +563,7 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
            descriptor.usage = getMetalTextureUsage(usage);
            descriptor.storageMode = MTLStorageModePrivate;
            texture = [context.device newTextureWithDescriptor:descriptor];
+            ASSERT_POSTCONDITION(texture != nil, "Could not create Metal texture. Out of memory?");
            break;
        case SamplerType::SAMPLER_EXTERNAL:
            // If we're using external textures (CVPixelBufferRefs), we don't need to make any
@@ -588,33 +572,15 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
            break;
    }

-    FILAMENT_CHECK_POSTCONDITION(target == SamplerType::SAMPLER_EXTERNAL || texture != nil)
-            << "Could not create Metal texture (SamplerType = " << int(target)
-            << ", levels = " << int(levels) << ", MTLPixelFormat = " << int(devicePixelFormat)
-            << ", width = " << width << ", height = " << height << ", depth = " << depth
-            << "). Out of memory?";
-}
-
-MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, uint8_t baseLevel,
-        uint8_t levelCount) noexcept
-    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
-              src->format, src->usage),
-      context(context),
-      devicePixelFormat(src->devicePixelFormat),
-      externalImage(src->externalImage) {
-    texture = createTextureViewWithLodRange(
-            src->getMtlTextureForRead(), baseLevel, baseLevel + levelCount - 1);
-}
-
-MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, TextureSwizzle r,
-        TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) noexcept
-    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
-              src->format, src->usage),
-      context(context),
-      devicePixelFormat(src->devicePixelFormat),
-      externalImage(src->externalImage) {
-    texture = src->getMtlTextureForRead();
-    if (context.supportsTextureSwizzling) {
+    // If swizzling is set, set up a swizzled texture view that we'll use when sampling this texture.
+    const bool isDefaultSwizzle =
+            r == TextureSwizzle::CHANNEL_0 &&
+            g == TextureSwizzle::CHANNEL_1 &&
+            b == TextureSwizzle::CHANNEL_2 &&
+            a == TextureSwizzle::CHANNEL_3;
+    // If texture is nil, then it must be a SAMPLER_EXTERNAL texture.
+    // Swizzling for external textures is handled inside MetalExternalImage.
+    if (!isDefaultSwizzle && texture && context.supportsTextureSwizzling) {
        // Even though we've already checked context.supportsTextureSwizzling, we still need to
        // guard these calls with @availability, otherwise the API usage will generate compiler
        // warnings.
@@ -628,30 +594,44 @@ MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, Textu
 MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
        uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
        id<MTLTexture> metalTexture) noexcept
-    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context) {
+    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
+        externalImage(context) {
    texture = metalTexture;
+    setLodRange(0, levels - 1);
 }

-MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t width,
-        uint32_t height, TextureUsage usage, CVPixelBufferRef image) noexcept
-    : HwTexture(SamplerType::SAMPLER_EXTERNAL, 1, 1, width, height, 1, format, usage),
-      context(context),
-      externalImage(std::make_shared<MetalExternalImage>(
-              MetalExternalImage::createFromImage(context, image))) {
-    texture = externalImage->getMtlTexture();
+void MetalTexture::terminate() noexcept {
+    texture = nil;
+    swizzledTextureView = nil;
+    lodTextureView = nil;
+    msaaSidecar = nil;
+    externalImage.set(nullptr);
+    terminated = true;
 }

-MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t width,
-        uint32_t height, TextureUsage usage, CVPixelBufferRef image, uint32_t plane) noexcept
-    : HwTexture(SamplerType::SAMPLER_EXTERNAL, 1, 1, width, height, 1, format, usage),
-      context(context),
-      externalImage(std::make_shared<MetalExternalImage>(
-              MetalExternalImage::createFromImagePlane(context, image, plane))) {
-    texture = externalImage->getMtlTexture();
+MetalTexture::~MetalTexture() {
+    externalImage.set(nullptr);
 }

-id<MTLTexture> MetalTexture::getMtlTextureForRead() const noexcept {
-    return swizzledTextureView ? swizzledTextureView : texture;
+id<MTLTexture> MetalTexture::getMtlTextureForRead() noexcept {
+    if (lodTextureView) {
+        return lodTextureView;
+    }
+    // The texture's swizzle remains constant throughout its lifetime, however its LOD range can
+    // change. We'll cache the LOD view, and set lodTextureView to nil if minLod or maxLod is
+    // updated.
+    id<MTLTexture> t = swizzledTextureView ? swizzledTextureView : texture;
+    if (!t) {
+        return nil;
+    }
+    if (UTILS_UNLIKELY(minLod > maxLod)) {
+        // If the texture does not have any available LODs, provide a view of only level 0.
+        // Filament should prevent this from ever occurring.
+        lodTextureView = createTextureViewWithLodRange(t, 0, 0);
+        return lodTextureView;
+    }
+    lodTextureView = createTextureViewWithLodRange(t, minLod, maxLod);
+    return lodTextureView;
 }

 MTLPixelFormat MetalTexture::decidePixelFormat(MetalContext* context, TextureFormat format) {
@@ -770,21 +750,24 @@ void MetalTexture::loadImage(uint32_t level, MTLRegion region, PixelBufferDescri
            assert_invariant(false);
        }
    }
+
+    extendLodRangeTo(level);
 }

 void MetalTexture::generateMipmaps() noexcept {
    id <MTLBlitCommandEncoder> blitEncoder = [getPendingCommandBuffer(&context) blitCommandEncoder];
    [blitEncoder generateMipmapsForTexture:texture];
    [blitEncoder endEncoding];
+    setLodRange(0, texture.mipmapLevelCount - 1);
 }

 void MetalTexture::loadSlice(uint32_t level, MTLRegion region, uint32_t byteOffset, uint32_t slice,
        PixelBufferDescriptor const& data) noexcept {
    const PixelBufferShape shape = PixelBufferShape::compute(data, format, region.size, byteOffset);

-    FILAMENT_CHECK_PRECONDITION(data.size >= shape.totalBytes)
-            << "Expected buffer size of at least " << shape.totalBytes
-            << " but received PixelBufferDescriptor with size " << data.size << ".";
+    ASSERT_PRECONDITION(data.size >= shape.totalBytes,
+            "Expected buffer size of at least %d but "
+            "received PixelBufferDescriptor with size %d.", shape.totalBytes, data.size);

    // Earlier versions of iOS don't have the maxBufferLength query, but 256 MB is a safe bet.
    NSUInteger deviceMaxBufferLength = 256 * 1024 * 1024;   // 256 MB
@@ -817,13 +800,13 @@ void MetalTexture::loadWithCopyBuffer(uint32_t level, uint32_t slice, MTLRegion
        PixelBufferDescriptor const& data, const PixelBufferShape& shape) {
    const size_t stagingBufferSize = shape.totalBytes;
    auto entry = context.bufferPool->acquireBuffer(stagingBufferSize);
-    memcpy(entry->buffer.get().contents,
+    memcpy(entry->buffer.contents,
            static_cast<uint8_t*>(data.buffer) + shape.sourceOffset,
            stagingBufferSize);
    id<MTLCommandBuffer> blitCommandBuffer = getPendingCommandBuffer(&context);
    id<MTLBlitCommandEncoder> blitCommandEncoder = [blitCommandBuffer blitCommandEncoder];
    blitCommandEncoder.label = @"Texture upload buffer blit";
-    [blitCommandEncoder copyFromBuffer:entry->buffer.get()
+    [blitCommandEncoder copyFromBuffer:entry->buffer
                          sourceOffset:0
                     sourceBytesPerRow:shape.bytesPerRow
                   sourceBytesPerImage:shape.bytesPerSlice
@@ -899,6 +882,98 @@ void MetalTexture::loadWithBlit(uint32_t level, uint32_t slice, MTLRegion region
    context.blitter->blit(getPendingCommandBuffer(&context), args, "Texture upload blit");
 }

+void MetalTexture::extendLodRangeTo(uint16_t level) {
+    assert_invariant(!isInRenderPass(&context));
+    minLod = std::min(minLod, level);
+    maxLod = std::max(maxLod, level);
+    lodTextureView = nil;
+}
+
+void MetalTexture::setLodRange(uint16_t min, uint16_t max) {
+    assert_invariant(!isInRenderPass(&context));
+    assert_invariant(min <= max);
+    minLod = min;
+    maxLod = max;
+    lodTextureView = nil;
+}
+
+void MetalSamplerGroup::finalize() {
+    assert_invariant(encoder);
+    // TODO: we should be able to encode textures and samplers inside setFinalizedTexture and
+    // setFinalizedSampler as they become available, but Metal doesn't seem to like this; the arg
+    // buffer gets encoded incorrectly. This warrants more investigation.
+
+    auto [buffer, offset] = argBuffer->getCurrentAllocation();
+    [encoder setArgumentBuffer:buffer offset:offset];
+
+    // Encode all textures and samplers.
+    for (size_t s = 0; s < size; s++) {
+        [encoder setTexture:textures[s] atIndex:(s * 2 + 0)];
+        [encoder setSamplerState:samplers[s] atIndex:(s * 2 + 1)];
+    }
+
+    finalized = true;
+}
+
+void MetalSamplerGroup::reset(id<MTLCommandBuffer> cmdBuffer, id<MTLArgumentEncoder> e,
+        id<MTLDevice> device) {
+    encoder = e;
+
+    // The number of slots in the ring buffer we use to manage argument buffer allocations.
+    // This number was chosen to avoid running out of slots and having to allocate a "fallback"
+    // buffer when SamplerGroups are updated multiple times a frame. This value can reduced after
+    // auditing Filament's calls to updateSamplerGroup, which should be as few times as possible.
+    // For example, the bloom downsample pass should be refactored to maintain two separate
+    // MaterialInstances instead of "ping ponging" between two texture bindings, which causes a
+    // single SamplerGroup to be updated many times a frame.
+    static constexpr auto METAL_ARGUMENT_BUFFER_SLOTS = 32;
+
+    MTLSizeAndAlign argBufferLayout;
+    argBufferLayout.size = encoder.encodedLength;
+    argBufferLayout.align = encoder.alignment;
+    // Chances are, even though the MTLArgumentEncoder might change, the required size and alignment
+    // probably won't. So we can re-use the previous ring buffer.
+    if (UTILS_UNLIKELY(!argBuffer || !argBuffer->canAccomodateLayout(argBufferLayout))) {
+        argBuffer = std::make_unique<MetalRingBuffer>(device, MTLResourceStorageModeShared,
+                argBufferLayout, METAL_ARGUMENT_BUFFER_SLOTS);
+    } else {
+        argBuffer->createNewAllocation(cmdBuffer);
+    }
+
+    // Clear all textures and samplers.
+    assert_invariant(textureHandles.size() == textures.size());
+    assert_invariant(textures.size() == samplers.size());
+    for (size_t s = 0; s < textureHandles.size(); s++) {
+        textureHandles[s] = {};
+        textures[s] = nil;
+        samplers[s] = nil;
+    }
+
+    finalized = false;
+}
+
+void MetalSamplerGroup::mutate(id<MTLCommandBuffer> cmdBuffer) {
+    assert_invariant(finalized);    // only makes sense to mutate if this sampler group is finalized
+    assert_invariant(argBuffer);
+    argBuffer->createNewAllocation(cmdBuffer);
+    finalized = false;
+}
+
+void MetalSamplerGroup::useResources(id<MTLRenderCommandEncoder> renderPassEncoder) {
+    assert_invariant(finalized);
+    if (@available(iOS 13, *)) {
+        // TODO: pass only the appropriate stages to useResources.
+        [renderPassEncoder useResources:textures.data()
+                                  count:textures.size()
+                                  usage:MTLResourceUsageRead | MTLResourceUsageSample
+                                 stages:MTLRenderStageFragment | MTLRenderStageVertex];
+    } else {
+        [renderPassEncoder useResources:textures.data()
+                                  count:textures.size()
+                                  usage:MTLResourceUsageRead | MTLResourceUsageSample];
+    }
+}
+
 MetalRenderTarget::MetalRenderTarget(MetalContext* context, uint32_t width, uint32_t height,
        uint8_t samples, Attachment colorAttachments[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
        Attachment depthAttachment, Attachment stencilAttachment) :
@@ -913,9 +988,9 @@ MetalRenderTarget::MetalRenderTarget(MetalContext* context, uint32_t width, uint
        }
        color[i] = colorAttachments[i];

-        FILAMENT_CHECK_PRECONDITION(color[i].getSampleCount() <= samples)
-                << "MetalRenderTarget was initialized with a MSAA COLOR" << i
-                << " texture, but sample count is " << samples << ".";
+        ASSERT_PRECONDITION(color[i].getSampleCount() <= samples,
+                "MetalRenderTarget was initialized with a MSAA COLOR%d texture, but sample count is %d.",
+                i, samples);

        auto t = color[i].metalTexture;
        const auto twidth = std::max(1u, t->width >> color[i].level);
@@ -938,10 +1013,9 @@ MetalRenderTarget::MetalRenderTarget(MetalContext* context, uint32_t width, uint
    if (depthAttachment) {
        depth = depthAttachment;

-        FILAMENT_CHECK_PRECONDITION(depth.getSampleCount() <= samples)
-                << "MetalRenderTarget was initialized with a MSAA DEPTH texture, but sample count "
-                   "is "
-                << samples << ".";
+        ASSERT_PRECONDITION(depth.getSampleCount() <= samples,
+                "MetalRenderTarget was initialized with a MSAA DEPTH texture, but sample count is %d.",
+                samples);

        auto t = depth.metalTexture;
        const auto twidth = std::max(1u, t->width >> depth.level);
@@ -964,10 +1038,9 @@ MetalRenderTarget::MetalRenderTarget(MetalContext* context, uint32_t width, uint
    if (stencilAttachment) {
        stencil = stencilAttachment;

-        FILAMENT_CHECK_PRECONDITION(stencil.getSampleCount() <= samples)
-                << "MetalRenderTarget was initialized with a MSAA STENCIL texture, but sample "
-                   "count is "
-                << samples << ".";
+        ASSERT_PRECONDITION(stencil.getSampleCount() <= samples,
+                "MetalRenderTarget was initialized with a MSAA STENCIL texture, but sample count is %d.",
+                samples);

        auto t = stencil.metalTexture;
        const auto twidth = std::max(1u, t->width >> stencil.level);
@@ -1244,193 +1317,5 @@ FenceStatus MetalFence::wait(uint64_t timeoutNs) {
    return FenceStatus::ERROR;
 }

-MetalDescriptorSetLayout::MetalDescriptorSetLayout(DescriptorSetLayout&& l) noexcept
-    : mLayout(std::move(l)) {
-    size_t dynamicBindings = 0;
-    for (const auto& binding : mLayout.bindings) {
-        if (any(binding.flags & DescriptorFlags::DYNAMIC_OFFSET)) {
-            dynamicBindings++;
-        }
-    }
-    mDynamicOffsetCount = dynamicBindings;
-}
-
-id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoder(id<MTLDevice> device, ShaderStage stage,
-        utils::FixedCapacityVector<MTLTextureType> const& textureTypes) {
-    auto const index = static_cast<size_t>(stage);
-    assert_invariant(index < mCachedArgumentEncoder.size());
-    if (mCachedArgumentEncoder[index] &&
-            std::equal(
-                    textureTypes.begin(), textureTypes.end(), mCachedTextureTypes[index].begin())) {
-        return mCachedArgumentEncoder[index];
-    }
-    mCachedArgumentEncoder[index] = getArgumentEncoderSlow(device, stage, textureTypes);
-    mCachedTextureTypes[index] = textureTypes;
-    return mCachedArgumentEncoder[index];
-}
-
-id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoderSlow(id<MTLDevice> device,
-        ShaderStage stage, utils::FixedCapacityVector<MTLTextureType> const& textureTypes) {
-    auto const& bindings = getBindings();
-    NSMutableArray<MTLArgumentDescriptor*>* arguments = [NSMutableArray new];
-    // Important! The bindings must be sorted by binding number. This has already been done inside
-    // createDescriptorSetLayout.
-    size_t textureIndex = 0;
-    for (auto const& binding : bindings) {
-        if (!hasShaderType(binding.stageFlags, stage)) {
-            continue;
-        }
-        switch (binding.type) {
-            case DescriptorType::UNIFORM_BUFFER:
-            case DescriptorType::SHADER_STORAGE_BUFFER: {
-                MTLArgumentDescriptor* bufferArgument = [MTLArgumentDescriptor argumentDescriptor];
-                bufferArgument.index = binding.binding * 2;
-                bufferArgument.dataType = MTLDataTypePointer;
-                bufferArgument.access = MTLArgumentAccessReadOnly;
-                [arguments addObject:bufferArgument];
-                break;
-            }
-            case DescriptorType::SAMPLER: {
-                MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
-                textureArgument.index = binding.binding * 2;
-                textureArgument.dataType = MTLDataTypeTexture;
-                MTLTextureType textureType = MTLTextureType2D;
-                if (textureIndex < textureTypes.size()) {
-                    textureType = textureTypes[textureIndex++];
-                }
-                textureArgument.textureType = textureType;
-                textureArgument.access = MTLArgumentAccessReadOnly;
-                [arguments addObject:textureArgument];
-
-                MTLArgumentDescriptor* samplerArgument = [MTLArgumentDescriptor argumentDescriptor];
-                samplerArgument.index = binding.binding * 2 + 1;
-                samplerArgument.dataType = MTLDataTypeSampler;
-                textureArgument.access = MTLArgumentAccessReadOnly;
-                [arguments addObject:samplerArgument];
-                break;
-            }
-            case DescriptorType::INPUT_ATTACHMENT:
-                // TODO: support INPUT_ATTACHMENT
-                assert_invariant(false);
-                break;
-        }
-    }
-    return [device newArgumentEncoderWithArguments:arguments];
-}
-
-MetalDescriptorSet::MetalDescriptorSet(MetalDescriptorSetLayout* layout) noexcept
-    : layout(layout) {}
-
-void MetalDescriptorSet::finalize(MetalDriver* driver) {
-    [driver->mContext->currentRenderPassEncoder useResource:driver->mContext->emptyBuffer
-                                                      usage:MTLResourceUsageRead];
-    [driver->mContext->currentRenderPassEncoder
-            useResource:getOrCreateEmptyTexture(driver->mContext)
-                  usage:MTLResourceUsageRead];
-
-    if (@available(iOS 13.0, *)) {
-        [driver->mContext->currentRenderPassEncoder useResources:vertexResources.data()
-                                                           count:vertexResources.size()
-                                                           usage:MTLResourceUsageRead
-                                                          stages:MTLRenderStageVertex];
-        [driver->mContext->currentRenderPassEncoder useResources:fragmentResources.data()
-                                                           count:fragmentResources.size()
-                                                           usage:MTLResourceUsageRead
-                                                          stages:MTLRenderStageFragment];
-    } else {
-        [driver->mContext->currentRenderPassEncoder useResources:vertexResources.data()
-                                                           count:vertexResources.size()
-                                                           usage:MTLResourceUsageRead];
-        [driver->mContext->currentRenderPassEncoder useResources:fragmentResources.data()
-                                                           count:fragmentResources.size()
-                                                           usage:MTLResourceUsageRead];
-    }
-}
-
-id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, ShaderStage stage) {
-    auto const index = static_cast<size_t>(stage);
-    assert_invariant(index < cachedBuffer.size());
-    auto& buffer = cachedBuffer[index];
-
-    if (buffer) {
-        return buffer.get();
-    }
-
-    // Map all the texture bindings to their respective texture types.
-    auto const& bindings = layout->getBindings();
-    auto textureTypes = utils::FixedCapacityVector<MTLTextureType>::with_capacity(bindings.size());
-    for (auto const& binding : bindings) {
-        if (!hasShaderType(binding.stageFlags, stage)) {
-            continue;
-        }
-        MTLTextureType textureType = MTLTextureType2D;
-        if (auto found = textures.find(binding.binding); found != textures.end()) {
-            auto const& textureBinding = textures[binding.binding];
-            textureType = textureBinding.texture.textureType;
-        }
-        textureTypes.push_back(textureType);
-    }
-
-    MetalContext const& context = *driver->mContext;
-
-    id<MTLArgumentEncoder> encoder =
-            layout->getArgumentEncoder(context.device, stage, textureTypes);
-
-    {
-        ScopedAllocationTimer timer("descriptor_set");
-        buffer = { [context.device newBufferWithLength:encoder.encodedLength
-                                               options:MTLResourceStorageModeShared],
-            TrackedMetalBuffer::Type::DESCRIPTOR_SET };
-    }
-    [encoder setArgumentBuffer:buffer.get() offset:0];
-
-    for (auto const& binding : bindings) {
-        if (!hasShaderType(binding.stageFlags, stage)) {
-            continue;
-        }
-        switch (binding.type) {
-            case DescriptorType::UNIFORM_BUFFER:
-            case DescriptorType::SHADER_STORAGE_BUFFER: {
-                auto found = buffers.find(binding.binding);
-                if (found == buffers.end()) {
-                    [encoder setBuffer:driver->mContext->emptyBuffer
-                                offset:0
-                               atIndex:binding.binding * 2];
-                    continue;
-                }
-
-                auto const& bufferBinding = buffers[binding.binding];
-                [encoder setBuffer:bufferBinding.buffer
-                            offset:bufferBinding.offset
-                           atIndex:binding.binding * 2];
-                break;
-            }
-            case DescriptorType::SAMPLER: {
-                auto found = textures.find(binding.binding);
-                if (found == textures.end()) {
-                    [encoder setTexture:driver->mContext->emptyTexture atIndex:binding.binding * 2];
-                    id<MTLSamplerState> sampler =
-                            driver->mContext->samplerStateCache.getOrCreateState({});
-                    [encoder setSamplerState:sampler atIndex:binding.binding * 2 + 1];
-                    continue;
-                }
-
-                auto const& textureBinding = textures[binding.binding];
-                [encoder setTexture:textureBinding.texture atIndex:binding.binding * 2];
-                SamplerState samplerState { .samplerParams = textureBinding.sampler };
-                id<MTLSamplerState> sampler =
-                        driver->mContext->samplerStateCache.getOrCreateState(samplerState);
-                [encoder setSamplerState:sampler atIndex:binding.binding * 2 + 1];
-                break;
-            }
-            case DescriptorType::INPUT_ATTACHMENT:
-                assert_invariant(false);
-                break;
-        }
-    }
-
-    return buffer.get();
-}
-
 } // namespace backend
 } // namespace filament
@@ -33,28 +33,30 @@
 namespace filament {
 namespace backend {

+inline bool operator==(const SamplerParams& lhs, const SamplerParams& rhs) {
+    return SamplerParams::EqualTo{}(lhs, rhs);
+}
+
 //   Rasterization Bindings
 //   ----------------------
 //   Bindings    Buffer name                          Count
 //   ------------------------------------------------------
 //   0           Zero buffer (placeholder vertex buffer)  1
 //   1-16        Filament vertex buffers                 16   limited by MAX_VERTEX_BUFFER_COUNT
-//   20          Push constants                           1
-//   21-24       Descriptor sets (argument buffers)       4   limited by MAX_DESCRIPTOR_SET_COUNT
-//   25          Dynamic offset buffer                    1
+//   17-26       Uniform buffers                         10   Program::UNIFORM_BINDING_COUNT
+//   27-30       Sampler groups (argument buffers)        4   Program::SAMPLER_BINDING_COUNT
 //
-//   Total                                               23
+//   Total                                               31

 //   Compute Bindings
 //   ----------------------
 //   Bindings    Buffer name                          Count
 //   ------------------------------------------------------
 //   0-3         SSBO buffers                             4   MAX_SSBO_COUNT
-//   20          Push constants                           1
-//   21-24       Descriptor sets (argument buffers)       4   limited by MAX_DESCRIPTOR_SET_COUNT
-//   25          Dynamic offset buffer                    1
+//   17-26       Uniform buffers                         10   Program::UNIFORM_BINDING_COUNT
+//   27-30       Sampler groups (argument buffers)        4   Program::SAMPLER_BINDING_COUNT
 //
-//   Total                                               10
+//   Total                                               18

 // The total number of vertex buffer "slots" that the Metal backend can bind.
 // + 1 to account for the zero buffer, a placeholder buffer used internally by the Metal backend.
@@ -67,11 +69,10 @@ static constexpr uint32_t ZERO_VERTEX_BUFFER_BINDING = 0u;

 static constexpr uint32_t USER_VERTEX_BUFFER_BINDING_START = 1u;

-
 // These constants must match the equivalent in CodeGenerator.h.
-static constexpr uint32_t PUSH_CONSTANT_BUFFER_INDEX = 20u;
-static constexpr uint32_t DESCRIPTOR_SET_BINDING_START = 21u;
-static constexpr uint32_t DYNAMIC_OFFSET_BINDING = 25u;
+static constexpr uint32_t UNIFORM_BUFFER_BINDING_START = 17u;
+static constexpr uint32_t SSBO_BINDING_START = 0u;
+static constexpr uint32_t SAMPLER_GROUP_BINDING_START = 27u;

 // Forward declarations necessary here, definitions at end of file.
 inline bool operator==(const MTLViewport& lhs, const MTLViewport& rhs);
@@ -379,22 +380,18 @@ using SamplerStateCache = StateCache<SamplerState, id<MTLSamplerState>, SamplerS

 using CullModeStateTracker = StateTracker<MTLCullMode>;
 using WindingStateTracker = StateTracker<MTLWinding>;
-using DepthClampStateTracker = StateTracker<MTLDepthClipMode>;

 // Argument encoder

 struct ArgumentEncoderState {
-    NSUInteger bufferCount;
    utils::FixedCapacityVector<MTLTextureType> textureTypes;

-    explicit ArgumentEncoderState(
-            NSUInteger bufferCount, utils::FixedCapacityVector<MTLTextureType>&& types)
-        : bufferCount(bufferCount), textureTypes(std::move(types)) {}
+    explicit ArgumentEncoderState(utils::FixedCapacityVector<MTLTextureType>&& types)
+        : textureTypes(std::move(types)) {}

    bool operator==(const ArgumentEncoderState& rhs) const noexcept {
        return std::equal(textureTypes.begin(), textureTypes.end(), rhs.textureTypes.begin(),
-                       rhs.textureTypes.end()) &&
-                bufferCount == rhs.bufferCount;
+                rhs.textureTypes.end());
    }

    bool operator!=(const ArgumentEncoderState& rhs) const noexcept {
@@ -416,30 +413,6 @@ struct ArgumentEncoderCreator {
 using ArgumentEncoderCache = StateCache<ArgumentEncoderState, id<MTLArgumentEncoder>,
        ArgumentEncoderCreator, ArgumentEncoderHasher>;

-template <NSUInteger N, ShaderStage stage>
-class MetalBufferBindings {
-public:
-    MetalBufferBindings() { invalidate(); }
-
-    void invalidate() {
-        mDirtyBuffers.reset();
-        mDirtyOffsets.reset();
-        for (int i = 0; i < int(N); i++) {
-            mDirtyBuffers.set(i, true);
-            mDirtyOffsets.set(i, true);
-        }
-    }
-    void setBuffer(const id<MTLBuffer> buffer, NSUInteger offset, NSUInteger index);
-    void bindBuffers(id<MTLCommandEncoder> encoder, NSUInteger startIndex);
-
-private:
-    static_assert(N <= 8);
-    std::array<__weak id<MTLBuffer>, N> mBuffers = { nil };
-    std::array<NSUInteger, N> mOffsets = { 0 };
-    utils::bitset8 mDirtyBuffers;
-    utils::bitset8 mDirtyOffsets;
-};
-
 } // namespace backend
 } // namespace filament

@@ -90,17 +90,11 @@ id<MTLRenderPipelineState> PipelineStateCreator::operator()(id<MTLDevice> device
    NSError* error = nullptr;
    id<MTLRenderPipelineState> pipeline = [device newRenderPipelineStateWithDescriptor:descriptor
                                                                                 error:&error];
-    if (UTILS_UNLIKELY(pipeline == nil)) {
-        NSString *errorMessage =
-            [NSString stringWithFormat:@"Could not create Metal pipeline state: %@",
-                error ? error.localizedDescription : @"unknown error"];
-        auto description = [errorMessage cStringUsingEncoding:NSUTF8StringEncoding];
+    if (error) {
+        auto description = [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
        utils::slog.e << description << utils::io::endl;
-        [[NSException exceptionWithName:@"MetalRenderPipelineFailure"
-                                 reason:errorMessage
-                               userInfo:nil] raise];
    }
-    FILAMENT_CHECK_POSTCONDITION(error == nil) << "Could not create Metal pipeline state.";
+    ASSERT_POSTCONDITION(error == nil, "Could not create Metal pipeline state.");

    return pipeline;
 }
@@ -166,40 +160,28 @@ id<MTLSamplerState> SamplerStateCreator::operator()(id<MTLDevice> device,
 id<MTLArgumentEncoder> ArgumentEncoderCreator::operator()(id<MTLDevice> device,
        const ArgumentEncoderState &state) noexcept {
    const auto& textureTypes = state.textureTypes;
-    const auto& textureCount = textureTypes.size();
-    const auto& bufferCount = state.bufferCount;
-    assert_invariant(textureCount > 0);
+    const auto& count = textureTypes.size();
+    assert_invariant(count > 0);

    // Metal has separate data types for textures versus samplers, so the argument buffer layout
    // alternates between texture and sampler, i.e.:
-    // buffer0
-    // buffer1
    // textureA
    // samplerA
    // textureB
    // samplerB
    // etc
    NSMutableArray<MTLArgumentDescriptor*>* arguments =
-            [NSMutableArray arrayWithCapacity:(bufferCount + textureCount * 2)];
-    size_t i = 0;
-    for (size_t j = 0; j < bufferCount; j++) {
-        MTLArgumentDescriptor* bufferArgument = [MTLArgumentDescriptor argumentDescriptor];
-        bufferArgument.index = i++;
-        bufferArgument.dataType = MTLDataTypePointer;
-        bufferArgument.access = MTLArgumentAccessReadOnly;
-        [arguments addObject:bufferArgument];
-    }
-
-    for (size_t j = 0; j < textureCount; j++) {
+            [NSMutableArray arrayWithCapacity:(count * 2)];
+    for (size_t i = 0; i < count; i++) {
        MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
-        textureArgument.index = i++;
+        textureArgument.index = i * 2 + 0;
        textureArgument.dataType = MTLDataTypeTexture;
        textureArgument.textureType = textureTypes[i];
        textureArgument.access = MTLArgumentAccessReadOnly;
        [arguments addObject:textureArgument];

        MTLArgumentDescriptor* samplerArgument = [MTLArgumentDescriptor argumentDescriptor];
-        samplerArgument.index = i++;
+        samplerArgument.index = i * 2 + 1;
        samplerArgument.dataType = MTLDataTypeSampler;
        textureArgument.access = MTLArgumentAccessReadOnly;
        [arguments addObject:samplerArgument];
@@ -208,64 +190,5 @@ id<MTLArgumentEncoder> ArgumentEncoderCreator::operator()(id<MTLDevice> device,
    return [device newArgumentEncoderWithArguments:arguments];
 }

-template <NSUInteger N, ShaderStage stage>
-void MetalBufferBindings<N, stage>::setBuffer(const id<MTLBuffer> buffer, NSUInteger offset, NSUInteger index) {
-    assert_invariant(offset + 1 <= N);
-
-    if (mBuffers[index] != buffer) {
-        mBuffers[index] = buffer;
-        mDirtyBuffers.set(index);
-    }
-
-    if (mOffsets[index] != offset) {
-        mOffsets[index] = offset;
-        mDirtyOffsets.set(index);
-    }
-}
-
-template <NSUInteger N, ShaderStage stage>
-void MetalBufferBindings<N, stage>::bindBuffers(
-        id<MTLCommandEncoder> encoder, NSUInteger startIndex) {
-    if (mDirtyBuffers.none() && mDirtyOffsets.none()) {
-        return;
-    }
-
-    utils::bitset8 onlyOffsetDirty = mDirtyOffsets & ~mDirtyBuffers;
-    onlyOffsetDirty.forEachSetBit([&](size_t i) {
-        if constexpr (stage == ShaderStage::VERTEX) {
-            [(id<MTLRenderCommandEncoder>)encoder setVertexBufferOffset:mOffsets[i]
-                                                                atIndex:i + startIndex];
-        } else if constexpr (stage == ShaderStage::FRAGMENT) {
-            [(id<MTLRenderCommandEncoder>)encoder setFragmentBufferOffset:mOffsets[i]
-                                                                  atIndex:i + startIndex];
-        } else if constexpr (stage == ShaderStage::COMPUTE) {
-            [(id<MTLComputeCommandEncoder>)encoder setBufferOffset:mOffsets[i]
-                                                           atIndex:i + startIndex];
-        }
-    });
-    mDirtyOffsets.reset();
-
-    mDirtyBuffers.forEachSetBit([&](size_t i) {
-        if constexpr (stage == ShaderStage::VERTEX) {
-            [(id<MTLRenderCommandEncoder>)encoder setVertexBuffer:mBuffers[i]
-                                                           offset:mOffsets[i]
-                                                          atIndex:i + startIndex];
-        } else if constexpr (stage == ShaderStage::FRAGMENT) {
-            [(id<MTLRenderCommandEncoder>)encoder setFragmentBuffer:mBuffers[i]
-                                                             offset:mOffsets[i]
-                                                            atIndex:i + startIndex];
-        } else if constexpr (stage == ShaderStage::COMPUTE) {
-            [(id<MTLComputeCommandEncoder>)encoder setBuffer:mBuffers[i]
-                                                      offset:mOffsets[i]
-                                                     atIndex:i + startIndex];
-        }
-    });
-    mDirtyBuffers.reset();
-}
-
-template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::VERTEX>;
-template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::FRAGMENT>;
-template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::COMPUTE>;
-
 } // namespace backend
 } // namespace filament
@@ -54,7 +54,7 @@ void NoopDriver::beginFrame(int64_t monotonic_clock_ns,
 }

 void NoopDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, FrameScheduledCallback&& callback, uint64_t flags) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {

 }

@@ -99,6 +99,9 @@ void NoopDriver::destroyProgram(Handle<HwProgram> ph) {
 void NoopDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
 }

+void NoopDriver::destroySamplerGroup(Handle<HwSamplerGroup> sbh) {
+}
+
 void NoopDriver::destroySwapChain(Handle<HwSwapChain> sch) {
 }

@@ -108,12 +111,6 @@ void NoopDriver::destroyStream(Handle<HwStream> sh) {
 void NoopDriver::destroyTimerQuery(Handle<HwTimerQuery> tqh) {
 }

-void NoopDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> tqh) {
-}
-
-void NoopDriver::destroyDescriptorSet(Handle<HwDescriptorSet> tqh) {
-}
-
 Handle<HwStream> NoopDriver::createStreamNative(void* nativeStream) {
    return {};
 }
@@ -185,7 +182,7 @@ bool NoopDriver::isProtectedContentSupported() {
    return false;
 }

-bool NoopDriver::isStereoSupported() {
+bool NoopDriver::isStereoSupported(backend::StereoscopicType) {
    return false;
 }

@@ -205,10 +202,6 @@ bool NoopDriver::isProtectedTexturesSupported() {
    return true;
 }

-bool NoopDriver::isDepthClampSupported() {
-    return false;
-}
-
 bool NoopDriver::isWorkaroundNeeded(Workaround) {
    return false;
 }
@@ -251,6 +244,9 @@ void NoopDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t inde
        Handle<HwBufferObject> boh) {
 }

+void NoopDriver::setMinMaxLevels(Handle<HwTexture> th, uint32_t minLevel, uint32_t maxLevel) {
+}
+
 void NoopDriver::update3DImage(Handle<HwTexture> th,
        uint32_t level, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset,
        uint32_t width, uint32_t height, uint32_t depth,
@@ -276,6 +272,11 @@ void NoopDriver::setExternalStream(Handle<HwTexture> th, Handle<HwStream> sh) {

 void NoopDriver::generateMipmaps(Handle<HwTexture> th) { }

+void NoopDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
+        BufferDescriptor&& data) {
+    scheduleDestroy(std::move(data));
+}
+
 void NoopDriver::compilePrograms(CompilerPriorityQueue priority,
        CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
    if (callback) {
@@ -298,14 +299,27 @@ void NoopDriver::makeCurrent(Handle<HwSwapChain> drawSch, Handle<HwSwapChain> re
 void NoopDriver::commit(Handle<HwSwapChain> sch) {
 }

+void NoopDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> ubh) {
+}
+
+void NoopDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t index,
+        Handle<HwBufferObject> ubh, uint32_t offset, uint32_t size) {
+}
+
+void NoopDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
+}
+
+void NoopDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
+}
+
 void NoopDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
        backend::PushConstantVariant value) {
 }

-void NoopDriver::insertEventMarker(char const* string) {
+void NoopDriver::insertEventMarker(char const* string, uint32_t len) {
 }

-void NoopDriver::pushGroupMarker(char const* string) {
+void NoopDriver::pushGroupMarker(char const* string,  uint32_t len) {
 }

 void NoopDriver::popGroupMarker(int) {
@@ -374,28 +388,4 @@ void NoopDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
 void NoopDriver::resetState(int) {
 }

-void NoopDriver::updateDescriptorSetBuffer(
-        backend::DescriptorSetHandle dsh,
-        backend::descriptor_binding_t binding,
-        backend::BufferObjectHandle boh,
-        uint32_t offset,
-        uint32_t size) {
-}
-
-void NoopDriver::updateDescriptorSetTexture(
-        backend::DescriptorSetHandle dsh,
-        backend::descriptor_binding_t binding,
-        backend::TextureHandle th,
-        SamplerParams params) {
-}
-
-void NoopDriver::bindDescriptorSet(
-        backend::DescriptorSetHandle dsh,
-        backend::descriptor_set_t set,
-        backend::DescriptorSetOffsetArray&& offsets) {
-}
-
-void NoopDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
-}
-
 } // namespace filament
@@ -1,89 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H
-#define TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H
-
-#include <backend/DriverEnums.h>
-
-#include "gl_headers.h"
-
-#include <utils/bitset.h>
-#include <utils/debug.h>
-
-#include <new>
-
-#include <stddef.h>
-#include <stdint.h>
-#include <string.h>
-
-namespace filament::backend {
-
-class BindingMap {
-    struct CompressedBinding {
-        // this is in fact a GLuint, but we only want 8-bits
-        uint8_t binding : 7;
-        uint8_t sampler : 1;
-    };
-
-    CompressedBinding (*mStorage)[MAX_DESCRIPTOR_COUNT];
-
-    utils::bitset64 mActiveDescriptors[MAX_DESCRIPTOR_SET_COUNT];
-
-public:
-    BindingMap() noexcept
-            : mStorage(new (std::nothrow) CompressedBinding[MAX_DESCRIPTOR_SET_COUNT][MAX_DESCRIPTOR_COUNT]) {
-#ifndef NDEBUG
-        memset(mStorage, 0xFF, sizeof(CompressedBinding[MAX_DESCRIPTOR_SET_COUNT][MAX_DESCRIPTOR_COUNT]));
-#endif
-    }
-
-    ~BindingMap() noexcept {
-        delete [] mStorage;
-    }
-
-    BindingMap(BindingMap const&) noexcept = delete;
-    BindingMap(BindingMap&&) noexcept = delete;
-    BindingMap& operator=(BindingMap const&) noexcept = delete;
-    BindingMap& operator=(BindingMap&&) noexcept = delete;
-
-    struct Binding {
-        GLuint binding;
-        DescriptorType type;
-    };
-
-    void insert(descriptor_set_t set, descriptor_binding_t binding, Binding entry) noexcept {
-        assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
-        assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
-        assert_invariant(entry.binding < 128); // we reserve 1 bit for the type right now
-        mStorage[set][binding] = { (uint8_t)entry.binding, entry.type == DescriptorType::SAMPLER };
-        mActiveDescriptors[set].set(binding);
-    }
-
-    GLuint get(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
-        assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
-        assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
-        return mStorage[set][binding].binding;
-    }
-
-    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
-        return mActiveDescriptors[set];
-    }
-};
-
-} // namespace filament::backend
-
-#endif //TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H
@@ -1,363 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "GLDescriptorSet.h"
-
-#include "GLBufferObject.h"
-#include "GLDescriptorSetLayout.h"
-#include "GLTexture.h"
-#include "GLUtils.h"
-#include "OpenGLDriver.h"
-#include "OpenGLContext.h"
-#include "OpenGLProgram.h"
-
-#include "gl_headers.h"
-
-#include <private/backend/HandleAllocator.h>
-
-#include <backend/DriverEnums.h>
-#include <backend/Handle.h>
-
-#include <utils/BitmaskEnum.h>
-#include <utils/Log.h>
-#include <utils/Panic.h>
-#include <utils/bitset.h>
-#include <utils/compiler.h>
-#include <utils/debug.h>
-#include <algorithm>
-
-#include <type_traits>
-#include <utility>
-#include <variant>
-
-#include <stddef.h>
-#include <stdint.h>
-
-namespace filament::backend {
-
-GLDescriptorSet::GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle dslh,
-        GLDescriptorSetLayout const* layout) noexcept
-        : descriptors(layout->maxDescriptorBinding + 1),
-          dslh(std::move(dslh)) {
-
-    // We have allocated enough storage for all descriptors. Now allocate the empty descriptor
-    // themselves.
-    for (auto const& entry : layout->bindings) {
-        size_t const index = entry.binding;
-
-        // now we'll initialize the alternative for each way we can handle this descriptor.
-        auto& desc = descriptors[index].desc;
-        switch (entry.type) {
-            case DescriptorType::UNIFORM_BUFFER: {
-                // A uniform buffer can have dynamic offsets or not and have special handling for
-                // ES2 (where we need to emulate it). That's four alternatives.
-                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
-                dynamicBuffers.set(index, dynamicOffset);
-                if (UTILS_UNLIKELY(gl.isES2())) {
-                    if (dynamicOffset) {
-                        dynamicBufferCount++;
-                    }
-                    desc.emplace<BufferGLES2>(dynamicOffset);
-                } else {
-                    auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::UNIFORM);
-                    if (dynamicOffset) {
-                        dynamicBufferCount++;
-                        desc.emplace<DynamicBuffer>(type);
-                    } else {
-                        desc.emplace<Buffer>(type);
-                    }
-                }
-                break;
-            }
-            case DescriptorType::SHADER_STORAGE_BUFFER: {
-                // shader storage buffers are not supported on ES2, So that's two alternatives.
-                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
-                dynamicBuffers.set(index, dynamicOffset);
-                auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::SHADER_STORAGE);
-                if (dynamicOffset) {
-                    dynamicBufferCount++;
-                    desc.emplace<DynamicBuffer>(type);
-                } else {
-                    desc.emplace<Buffer>(type);
-                }
-                break;
-            }
-            case DescriptorType::SAMPLER:
-                if (UTILS_UNLIKELY(gl.isES2())) {
-                    desc.emplace<SamplerGLES2>();
-                } else {
-                    const bool anisotropyWorkaround =
-                            gl.ext.EXT_texture_filter_anisotropic &&
-                            gl.bugs.texture_filter_anisotropic_broken_on_sampler;
-                    if (anisotropyWorkaround) {
-                        desc.emplace<SamplerWithAnisotropyWorkaround>();
-                    } else {
-                        desc.emplace<Sampler>();
-                    }
-                }
-                break;
-            case DescriptorType::INPUT_ATTACHMENT:
-                break;
-        }
-    }
-}
-
-void GLDescriptorSet::update(OpenGLContext&,
-        descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept {
-    assert_invariant(binding < descriptors.size());
-    std::visit([=](auto&& arg) {
-        using T = std::decay_t<decltype(arg)>;
-        if constexpr (std::is_same_v<T, Buffer> || std::is_same_v<T, DynamicBuffer>) {
-            assert_invariant(arg.target != 0);
-            arg.id = bo ? bo->gl.id : 0;
-            arg.offset = uint32_t(offset);
-            arg.size = uint32_t(size);
-            assert_invariant(arg.id || (!arg.size && !offset));
-        } else if constexpr (std::is_same_v<T, BufferGLES2>) {
-            arg.bo = bo;
-            arg.offset = uint32_t(offset);
-        } else {
-            // API usage error. User asked to update the wrong type of descriptor.
-            PANIC_PRECONDITION("descriptor %d is not a buffer", +binding);
-        }
-    }, descriptors[binding].desc);
-}
-
-void GLDescriptorSet::update(OpenGLContext& gl,
-        descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept {
-    assert_invariant(binding < descriptors.size());
-    std::visit([=, &gl](auto&& arg) mutable {
-        using T = std::decay_t<decltype(arg)>;
-        if constexpr (std::is_same_v<T, Sampler> ||
-                      std::is_same_v<T, SamplerWithAnisotropyWorkaround> ||
-                      std::is_same_v<T, SamplerGLES2>) {
-            if (UTILS_UNLIKELY(t && t->target == SamplerType::SAMPLER_EXTERNAL)) {
-                // From OES_EGL_image_external spec:
-                // "The default s and t wrap modes are CLAMP_TO_EDGE, and it is an INVALID_ENUM
-                //  error to set the wrap mode to any other value."
-                params.wrapS = SamplerWrapMode::CLAMP_TO_EDGE;
-                params.wrapT = SamplerWrapMode::CLAMP_TO_EDGE;
-                params.wrapR = SamplerWrapMode::CLAMP_TO_EDGE;
-            }
-            // GLES3.x specification forbids depth textures to be filtered.
-            if (t && isDepthFormat(t->format)
-                    && params.compareMode == SamplerCompareMode::NONE) {
-                params.filterMag = SamplerMagFilter::NEAREST;
-                switch (params.filterMin) {
-                    case SamplerMinFilter::LINEAR:
-                        params.filterMin = SamplerMinFilter::NEAREST;
-                        break;
-                    case SamplerMinFilter::LINEAR_MIPMAP_NEAREST:
-                    case SamplerMinFilter::NEAREST_MIPMAP_LINEAR:
-                    case SamplerMinFilter::LINEAR_MIPMAP_LINEAR:
-                        params.filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST;
-                        break;
-                    default:
-                        break;
-                }
-            }
-
-            arg.target = t ? t->gl.target : 0;
-            arg.id = t ? t->gl.id : 0;
-            if constexpr (std::is_same_v<T, Sampler> ||
-                          std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
-                if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
-                    arg.anisotropy = float(1u << params.anisotropyLog2);
-                }
-                if (t) {
-                    arg.ref = t->ref;
-                    arg.baseLevel = t->gl.baseLevel;
-                    arg.maxLevel = t->gl.maxLevel;
-                    arg.swizzle = t->gl.swizzle;
-                }
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-                arg.sampler = gl.getSampler(params);
-#else
-                (void)gl;
-#endif
-            } else {
-                arg.params = params;
-            }
-        } else {
-            // API usage error. User asked to update the wrong type of descriptor.
-            PANIC_PRECONDITION("descriptor %d is not a texture", +binding);
-        }
-    }, descriptors[binding].desc);
-}
-
-template<typename T>
-void GLDescriptorSet::updateTextureView(OpenGLContext& gl,
-        HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept {
-    // The common case is that we don't have a ref handle (we only have one if
-    // the texture ever had a View on it).
-    assert_invariant(desc.ref);
-    GLTextureRef* const ref = handleAllocator.handle_cast<GLTextureRef*>(desc.ref);
-    if (UTILS_UNLIKELY((desc.baseLevel != ref->baseLevel || desc.maxLevel != ref->maxLevel))) {
-        // If we have views, then it's still uncommon that we'll switch often
-        // handle the case where we reset to the original texture
-        GLint baseLevel = GLint(desc.baseLevel); // NOLINT(*-signed-char-misuse)
-        GLint maxLevel = GLint(desc.maxLevel); // NOLINT(*-signed-char-misuse)
-        if (baseLevel > maxLevel) {
-            baseLevel = 0;
-            maxLevel = 1000; // per OpenGL spec
-        }
-        // that is very unfortunate that we have to call activeTexture here
-        gl.activeTexture(unit);
-        glTexParameteri(desc.target, GL_TEXTURE_BASE_LEVEL, baseLevel);
-        glTexParameteri(desc.target, GL_TEXTURE_MAX_LEVEL,  maxLevel);
-        ref->baseLevel = desc.baseLevel;
-        ref->maxLevel = desc.maxLevel;
-    }
-    if (UTILS_UNLIKELY(desc.swizzle != ref->swizzle)) {
-        using namespace GLUtils;
-        gl.activeTexture(unit);
-#if !defined(__EMSCRIPTEN__)  && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_R, (GLint)getSwizzleChannel(desc.swizzle[0]));
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_G, (GLint)getSwizzleChannel(desc.swizzle[1]));
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_B, (GLint)getSwizzleChannel(desc.swizzle[2]));
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_A, (GLint)getSwizzleChannel(desc.swizzle[3]));
-#endif
-        ref->swizzle = desc.swizzle;
-    }
-}
-
-void GLDescriptorSet::bind(
-        OpenGLContext& gl,
-        HandleAllocatorGL& handleAllocator,
-        OpenGLProgram const& p,
-        descriptor_set_t set, uint32_t const* offsets, bool offsetsOnly) const noexcept {
-    // TODO: check that offsets is sized correctly
-    size_t dynamicOffsetIndex = 0;
-
-    utils::bitset64 activeDescriptorBindings = p.getActiveDescriptors(set);
-    if (offsetsOnly) {
-        activeDescriptorBindings &= dynamicBuffers;
-    }
-
-    // loop only over the active indices for this program
-    activeDescriptorBindings.forEachSetBit(
-            [this,&gl, &handleAllocator, &p, set, offsets, &dynamicOffsetIndex]
-            (size_t binding) {
-
-        // This would fail here if we're trying to set a descriptor that doesn't exist in the
-        // program. In other words, a mismatch between the program's layout and this descriptor-set.
-        assert_invariant(binding < descriptors.size());
-
-        auto const& entry = descriptors[binding];
-        std::visit(
-                [&gl, &handleAllocator, &p, &dynamicOffsetIndex, set, binding, offsets]
-                (auto&& arg) {
-            using T = std::decay_t<decltype(arg)>;
-            if constexpr (std::is_same_v<T, Buffer>) {
-                GLuint const bindingPoint = p.getBufferBinding(set, binding);
-                GLintptr const offset = arg.offset;
-                assert_invariant(arg.id || (!arg.size && !offset));
-                gl.bindBufferRange(arg.target, bindingPoint, arg.id, offset, arg.size);
-            } else if constexpr (std::is_same_v<T, DynamicBuffer>) {
-                GLuint const bindingPoint = p.getBufferBinding(set, binding);
-                GLintptr const offset = arg.offset + offsets[dynamicOffsetIndex++];
-                assert_invariant(arg.id || (!arg.size && !offset));
-                gl.bindBufferRange(arg.target, bindingPoint, arg.id, offset, arg.size);
-            } else if constexpr (std::is_same_v<T, BufferGLES2>) {
-                GLuint const bindingPoint = p.getBufferBinding(set, binding);
-                GLintptr offset = arg.offset;
-                if (arg.dynamicOffset) {
-                    offset += offsets[dynamicOffsetIndex++];
-                }
-                if (arg.bo) {
-                    auto buffer = static_cast<char const*>(arg.bo->gl.buffer) + offset;
-                    p.updateUniforms(bindingPoint, arg.bo->gl.id, buffer, arg.bo->age);
-                }
-            } else if constexpr (std::is_same_v<T, Sampler>) {
-                GLuint const unit = p.getTextureUnit(set, binding);
-                if (arg.target) {
-                    gl.bindTexture(unit, arg.target, arg.id);
-                    gl.bindSampler(unit, arg.sampler);
-                    if (UTILS_UNLIKELY(arg.ref)) {
-                        updateTextureView(gl, handleAllocator, unit, arg);
-                    }
-                } else {
-                    gl.unbindTextureUnit(unit);
-                }
-            } else if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
-                GLuint const unit = p.getTextureUnit(set, binding);
-                if (arg.target) {
-                    gl.bindTexture(unit, arg.target, arg.id);
-                    gl.bindSampler(unit, arg.sampler);
-                    if (UTILS_UNLIKELY(arg.ref)) {
-                        updateTextureView(gl, handleAllocator, unit, arg);
-                    }
-#if defined(GL_EXT_texture_filter_anisotropic)
-                    // Driver claims to support anisotropic filtering, but it fails when set on
-                    // the sampler, we have to set it on the texture instead.
-                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
-                            std::min(gl.gets.max_anisotropy, float(arg.anisotropy)));
-#endif
-                } else {
-                    gl.unbindTextureUnit(unit);
-                }
-            } else if constexpr (std::is_same_v<T, SamplerGLES2>) {
-                // in ES2 the sampler parameters need to be set on the texture itself
-                GLuint const unit = p.getTextureUnit(set, binding);
-                if (arg.target) {
-                    gl.bindTexture(unit, arg.target, arg.id);
-                    SamplerParams const params = arg.params;
-                    glTexParameteri(arg.target, GL_TEXTURE_MIN_FILTER,
-                            (GLint)GLUtils::getTextureFilter(params.filterMin));
-                    glTexParameteri(arg.target, GL_TEXTURE_MAG_FILTER,
-                            (GLint)GLUtils::getTextureFilter(params.filterMag));
-                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_S,
-                            (GLint)GLUtils::getWrapMode(params.wrapS));
-                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_T,
-                            (GLint)GLUtils::getWrapMode(params.wrapT));
-#if defined(GL_EXT_texture_filter_anisotropic)
-                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
-                            std::min(gl.gets.max_anisotropy, arg.anisotropy));
-#endif
-                } else {
-                    gl.unbindTextureUnit(unit);
-                }
-            }
-        }, entry.desc);
-    });
-    CHECK_GL_ERROR(utils::slog.e)
-}
-
-void GLDescriptorSet::validate(HandleAllocatorGL& allocator,
-        DescriptorSetLayoutHandle pipelineLayout) const {
-
-    if (UTILS_UNLIKELY(dslh != pipelineLayout)) {
-        auto* const dsl = allocator.handle_cast < GLDescriptorSetLayout const * > (dslh);
-        auto* const cur = allocator.handle_cast < GLDescriptorSetLayout const * > (pipelineLayout);
-
-        UTILS_UNUSED_IN_RELEASE
-        bool const pipelineLayoutMatchesDescriptorSetLayout = std::equal(
-                dsl->bindings.begin(), dsl->bindings.end(),
-                cur->bindings.begin(),
-                [](DescriptorSetLayoutBinding const& lhs,
-                        DescriptorSetLayoutBinding const& rhs) {
-                    return lhs.type == rhs.type &&
-                           lhs.stageFlags == rhs.stageFlags &&
-                           lhs.binding == rhs.binding &&
-                           lhs.flags == rhs.flags &&
-                           lhs.count == rhs.count;
-                });
-
-        assert_invariant(pipelineLayoutMatchesDescriptorSetLayout);
-    }
-}
-
-} // namespace filament::backend
@@ -1,175 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H
-#define TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H
-
-#include "DriverBase.h"
-
-#include "gl_headers.h"
-
-#include <private/backend/HandleAllocator.h>
-
-#include <backend/DriverEnums.h>
-#include <backend/Handle.h>
-
-#include <utils/bitset.h>
-#include <utils/FixedCapacityVector.h>
-
-#include <math/half.h>
-
-#include <array>
-#include <variant>
-
-#include <stddef.h>
-#include <stdint.h>
-
-namespace filament::backend {
-
-struct GLBufferObject;
-struct GLTexture;
-struct GLTextureRef;
-struct GLDescriptorSetLayout;
-class OpenGLProgram;
-class OpenGLContext;
-class OpenGLDriver;
-
-struct GLDescriptorSet : public HwDescriptorSet {
-
-    using HwDescriptorSet::HwDescriptorSet;
-
-    GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle dslh,
-            GLDescriptorSetLayout const* layout) noexcept;
-
-    // update a buffer descriptor in the set
-    void update(OpenGLContext& gl,
-            descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept;
-
-    // update a sampler descriptor in the set
-    void update(OpenGLContext& gl,
-            descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept;
-
-    // conceptually bind the set to the command buffer
-    void bind(
-            OpenGLContext& gl,
-            HandleAllocatorGL& handleAllocator,
-            OpenGLProgram const& p,
-            descriptor_set_t set, uint32_t const* offsets, bool offsetsOnly) const noexcept;
-
-    uint32_t getDynamicBufferCount() const noexcept {
-        return dynamicBufferCount;
-    }
-
-    void validate(HandleAllocatorGL& allocator, DescriptorSetLayoutHandle pipelineLayout) const;
-
-private:
-    // a Buffer Descriptor such as SSBO or UBO with static offset
-    struct Buffer {
-        // Workaround: we cannot define the following as Buffer() = default because one of our
-        // clients has their compiler set up where such declaration (possibly coupled with explicit)
-        // will be considered a deleted constructor.
-        Buffer() {}
-
-        explicit Buffer(GLenum target) noexcept : target(target) {}
-        GLenum target;                          // 4
-        GLuint id = 0;                          // 4
-        uint32_t offset = 0;                    // 4
-        uint32_t size = 0;                      // 4
-    };
-
-    // a Buffer Descriptor such as SSBO or UBO with dynamic offset
-    struct DynamicBuffer {
-        DynamicBuffer() = default;
-        explicit DynamicBuffer(GLenum target) noexcept : target(target) { }
-        GLenum target;                          // 4
-        GLuint id = 0;                          // 4
-        uint32_t offset = 0;                    // 4
-        uint32_t size = 0;                      // 4
-    };
-
-    // a UBO descriptor for ES2
-    struct BufferGLES2 {
-        BufferGLES2() = default;
-        explicit BufferGLES2(bool dynamicOffset) noexcept : dynamicOffset(dynamicOffset) { }
-        GLBufferObject const* bo = nullptr;     // 8
-        uint32_t offset = 0;                    // 4
-        bool dynamicOffset = false;             // 4
-    };
-
-    // A sampler descriptor
-    struct Sampler {
-        GLenum target = 0;                      // 4
-        GLuint id = 0;                          // 4
-        GLuint sampler = 0;                     // 4
-        Handle<GLTextureRef> ref;               // 4
-        int8_t baseLevel = 0x7f;                // 1
-        int8_t maxLevel = -1;                   // 1
-        std::array<TextureSwizzle, 4> swizzle{  // 4
-                TextureSwizzle::CHANNEL_0,
-                TextureSwizzle::CHANNEL_1,
-                TextureSwizzle::CHANNEL_2,
-                TextureSwizzle::CHANNEL_3
-        };
-    };
-
-    struct SamplerWithAnisotropyWorkaround {
-        GLenum target = 0;                      // 4
-        GLuint id = 0;                          // 4
-        GLuint sampler = 0;                     // 4
-        Handle<GLTextureRef> ref;               // 4
-        math::half anisotropy = 1.0f;           // 2
-        int8_t baseLevel = 0x7f;                // 1
-        int8_t maxLevel = -1;                   // 1
-        std::array<TextureSwizzle, 4> swizzle{  // 4
-                TextureSwizzle::CHANNEL_0,
-                TextureSwizzle::CHANNEL_1,
-                TextureSwizzle::CHANNEL_2,
-                TextureSwizzle::CHANNEL_3
-        };
-    };
-
-    // A sampler descriptor for ES2
-    struct SamplerGLES2 {
-        GLenum target = 0;                      // 4
-        GLuint id = 0;                          // 4
-        SamplerParams params{};                 // 4
-        float anisotropy = 1.0f;                // 4
-    };
-    struct Descriptor {
-        std::variant<
-                Buffer,
-                DynamicBuffer,
-                BufferGLES2,
-                Sampler,
-                SamplerWithAnisotropyWorkaround,
-                SamplerGLES2> desc;
-    };
-    static_assert(sizeof(Descriptor) <= 32);
-
-    template<typename T>
-    static void updateTextureView(OpenGLContext& gl,
-            HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept;
-
-    utils::FixedCapacityVector<Descriptor> descriptors;     // 16
-    utils::bitset64 dynamicBuffers;                         // 8
-    DescriptorSetLayoutHandle dslh;                         // 4
-    uint8_t dynamicBufferCount = 0;                         // 1
-};
-static_assert(sizeof(GLDescriptorSet) <= 32);
-
-} // namespace filament::backend
-
-#endif //TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H
@@ -1,52 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H
-#define TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H
-
-#include "DriverBase.h"
-
-#include <backend/DriverEnums.h>
-
-#include <algorithm>
-#include <utility>
-
-#include <stdint.h>
-
-namespace filament::backend {
-
-struct GLDescriptorSetLayout : public HwDescriptorSetLayout, public DescriptorSetLayout {
-    using HwDescriptorSetLayout::HwDescriptorSetLayout;
-    explicit GLDescriptorSetLayout(DescriptorSetLayout&& layout) noexcept
-            : DescriptorSetLayout(std::move(layout)) {
-
-        std::sort(bindings.begin(), bindings.end(),
-                [](auto&& lhs, auto&& rhs){
-            return lhs.binding < rhs.binding;
-        });
-
-        auto p = std::max_element(bindings.cbegin(), bindings.cend(),
-                [](auto const& lhs, auto const& rhs) {
-            return lhs.binding < rhs.binding;
-        });
-        maxDescriptorBinding = p->binding;
-    }
-    uint8_t maxDescriptorBinding = 0;
-};
-
-} // namespace filament::backend
-
-#endif //TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H
@@ -21,32 +21,12 @@

 #include "gl_headers.h"

-#include <backend/Handle.h>
-#include <backend/DriverEnums.h>
 #include <backend/platforms/OpenGLPlatform.h>

-#include <array>
-
 #include <stdint.h>

 namespace filament::backend {

-struct GLTextureRef {
-    GLTextureRef() = default;
-    // view reference counter
-    uint16_t count = 1;
-    // current per-view values of the texture (in GL we can only have a single View active at
-    // a time, and this tracks that state). It's used to avoid unnecessarily change state.
-    int8_t baseLevel = 127;
-    int8_t maxLevel = -1;
-    std::array<TextureSwizzle, 4> swizzle{
-            TextureSwizzle::CHANNEL_0,
-            TextureSwizzle::CHANNEL_1,
-            TextureSwizzle::CHANNEL_2,
-            TextureSwizzle::CHANNEL_3
-    };
-};
-
 struct GLTexture : public HwTexture {
    using HwTexture::HwTexture;
    struct GL {
@@ -64,14 +44,8 @@ struct GLTexture : public HwTexture {
        bool imported           : 1;
        uint8_t sidecarSamples  : 4;
        uint8_t reserved1       : 3;
-        std::array<TextureSwizzle, 4> swizzle{
-                TextureSwizzle::CHANNEL_0,
-                TextureSwizzle::CHANNEL_1,
-                TextureSwizzle::CHANNEL_2,
-                TextureSwizzle::CHANNEL_3
-        };
    } gl;
-    mutable Handle<GLTextureRef> ref;
+
    OpenGLPlatform::ExternalTexture* externalTexture = nullptr;
 };

@@ -66,8 +66,7 @@ bool OpenGLContext::queryOpenGLVersion(GLint* major, GLint* minor) noexcept {
 OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
        Platform::DriverConfig const& driverConfig) noexcept
        : mPlatform(platform),
-          mSamplerMap(32),
-          mDriverConfig(driverConfig) {
+          mSamplerMap(32) {

    state.vao.p = &mDefaultVAO;

@@ -367,8 +366,7 @@ void OpenGLContext::setDefaultState() noexcept {
    }
 #endif

-    if (ext.EXT_clip_cull_distance
-            && mDriverConfig.stereoscopicType == StereoscopicType::INSTANCED) {
+    if (ext.EXT_clip_cull_distance) {
        glEnable(GL_CLIP_DISTANCE0);
        glEnable(GL_CLIP_DISTANCE1);
    }
@@ -552,14 +550,6 @@ void OpenGLContext::initBugs(Bugs* bugs, Extensions const& exts,
        } else if (strstr(renderer, "AMD") ||
                   strstr(renderer, "ATI")) {
            // AMD/ATI GPU
-        } else if (strstr(vendor, "Mesa")) {
-            // Seen on
-            //  [Mesa],
-            //  [llvmpipe (LLVM 17.0.6, 256 bits)],
-            //  [4.5 (Core Profile) Mesa 24.0.6-1],
-            //  [4.50]
-            // not known which version are affected
-            bugs->rebind_buffer_after_deletion = true;
        } else if (strstr(renderer, "Mozilla")) {
            bugs->disable_invalidate_framebuffer = true;
        }
@@ -684,7 +674,6 @@ void OpenGLContext::initExtensionsGLES(Extensions* ext, GLint major, GLint minor
 #ifndef __EMSCRIPTEN__
    ext->EXT_debug_marker = exts.has("GL_EXT_debug_marker"sv);
 #endif
-    ext->EXT_depth_clamp = exts.has("GL_EXT_depth_clamp"sv);
    ext->EXT_discard_framebuffer = exts.has("GL_EXT_discard_framebuffer"sv);
 #ifndef __EMSCRIPTEN__
    ext->EXT_disjoint_timer_query = exts.has("GL_EXT_disjoint_timer_query"sv);
@@ -755,7 +744,6 @@ void OpenGLContext::initExtensionsGL(Extensions* ext, GLint major, GLint minor)
    ext->EXT_color_buffer_half_float = true;  // Assumes core profile.
    ext->EXT_clip_cull_distance = true;
    ext->EXT_debug_marker = exts.has("GL_EXT_debug_marker"sv);
-    ext->EXT_depth_clamp = true;
    ext->EXT_discard_framebuffer = false;
    ext->EXT_disjoint_timer_query = true;
    ext->EXT_multisampled_render_to_texture = false;
@@ -937,19 +925,15 @@ void OpenGLContext::unbindSampler(GLuint sampler) noexcept {
    }
 }

-void OpenGLContext::deleteBuffer(GLuint buffer, GLenum target) noexcept {
-    glDeleteBuffers(1, &buffer);
-
+void OpenGLContext::deleteBuffers(GLsizei n, const GLuint* buffers, GLenum target) noexcept {
+    glDeleteBuffers(n, buffers);
    // bindings of bound buffers are reset to 0
-    size_t const targetIndex = getIndexForBufferTarget(target);
-    auto& genericBinding = state.buffers.genericBinding[targetIndex];
-    if (genericBinding == buffer) {
-        genericBinding = 0;
-    }
-
-    if (UTILS_UNLIKELY(bugs.rebind_buffer_after_deletion)) {
-        if (genericBinding) {
-            glBindBuffer(target, genericBinding);
+    const size_t targetIndex = getIndexForBufferTarget(target);
+    auto& genericBuffer = state.buffers.genericBinding[targetIndex];
+    UTILS_NOUNROLL
+    for (GLsizei i = 0; i < n; ++i) {
+        if (genericBuffer == buffers[i]) {
+            genericBuffer = 0;
        }
    }

@@ -958,13 +942,16 @@ void OpenGLContext::deleteBuffer(GLuint buffer, GLenum target) noexcept {
            (target != GL_UNIFORM_BUFFER && target != GL_TRANSFORM_FEEDBACK_BUFFER));

    if (target == GL_UNIFORM_BUFFER || target == GL_TRANSFORM_FEEDBACK_BUFFER) {
-        auto& indexedBinding = state.buffers.targets[targetIndex];
-        UTILS_NOUNROLL
-        for (auto& entry: indexedBinding.buffers) {
-            if (entry.name == buffer) {
-                entry.name = 0;
-                entry.offset = 0;
-                entry.size = 0;
+        auto& indexedBuffer = state.buffers.targets[targetIndex];
+        UTILS_NOUNROLL // clang generates >1 KiB of code!!
+        for (GLsizei i = 0; i < n; ++i) {
+            UTILS_NOUNROLL
+            for (auto& buffer : indexedBuffer.buffers) {
+                if (buffer.name == buffers[i]) {
+                    buffer.name = 0;
+                    buffer.offset = 0;
+                    buffer.size = 0;
+                }
            }
        }
    }
@@ -60,19 +60,10 @@ public:
    struct RenderPrimitive {
        static_assert(MAX_VERTEX_ATTRIBUTE_COUNT <= 16);

-        GLuint vao[2] = {};                                     // 8
+        GLuint vao[2] = {};                                     // 4
        GLuint elementArray = 0;                                // 4
-        GLenum indicesType = 0;                                 // 4
-
-        // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
-        // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
-        // immutable.
-        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4
-
        mutable utils::bitset<uint16_t> vertexAttribArray;      // 2

-        uint8_t reserved[2] = {};                               // 2
-
        // if this differs from vertexBufferWithObjects->bufferObjectsVersion, this VAO needs to
        // be updated (see OpenGLDriver::updateVertexArrayObject())
        uint8_t vertexBufferVersion = 0;                        // 1
@@ -85,11 +76,16 @@ public:
        // See OpenGLContext::bindVertexArray()
        uint8_t nameVersion = 0;                                // 1

-        // Size in bytes of indices in the index buffer (1 or 2)
-        uint8_t indicesShift = 0;                                // 1
+        // Size in bytes of indices in the index buffer
+        uint8_t indicesSize = 0;                                // 1
+
+        // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
+        // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
+        // immutable.
+        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4

        GLenum getIndicesType() const noexcept {
-            return indicesType;
+            return indicesSize == 4 ? GL_UNSIGNED_INT : GL_UNSIGNED_SHORT;
        }
    } gl;

@@ -190,7 +186,7 @@ public:
    inline void viewport(GLint left, GLint bottom, GLsizei width, GLsizei height) noexcept;
    inline void depthRange(GLclampf near, GLclampf far) noexcept;

-    void deleteBuffer(GLuint buffer, GLenum target) noexcept;
+    void deleteBuffers(GLsizei n, const GLuint* buffers, GLenum target) noexcept;
    void deleteVertexArray(GLuint vao) noexcept;

    void destroyWithContext(size_t index, std::function<void(OpenGLContext&)> const& closure) noexcept;
@@ -224,9 +220,8 @@ public:
        bool EXT_color_buffer_float;
        bool EXT_color_buffer_half_float;
        bool EXT_debug_marker;
-        bool EXT_depth_clamp;
-        bool EXT_discard_framebuffer;
        bool EXT_disjoint_timer_query;
+        bool EXT_discard_framebuffer;
        bool EXT_multisampled_render_to_texture2;
        bool EXT_multisampled_render_to_texture;
        bool EXT_protected_textures;
@@ -244,10 +239,10 @@ public:
        bool KHR_parallel_shader_compile;
        bool KHR_texture_compression_astc_hdr;
        bool KHR_texture_compression_astc_ldr;
-        bool OES_EGL_image_external_essl3;
-        bool OES_depth24;
        bool OES_depth_texture;
+        bool OES_depth24;
        bool OES_packed_depth_stencil;
+        bool OES_EGL_image_external_essl3;
        bool OES_rgb8_rgba8;
        bool OES_standard_derivatives;
        bool OES_texture_npot;
@@ -316,15 +311,10 @@ public:
        // a glFinish. So we must delay the destruction until we know the GPU is finished.
        bool delay_fbo_destruction;

-        // Mesa sometimes clears the generic buffer binding when *another* buffer is destroyed,
-        // if that other buffer is bound on an *indexed* buffer binding.
-        bool rebind_buffer_after_deletion;
-
        // Force feature level 0. Typically used for low end ES3 devices with significant driver
        // bugs or performance issues.
        bool force_feature_level0;

-
    } bugs = {};

    // state getters -- as needed.
@@ -483,6 +473,12 @@ public:

    void unbindEverything() noexcept;
    void synchronizeStateAndCache(size_t index) noexcept;
+    void setEs2UniformBinding(size_t index, GLuint id, void const* data, uint16_t age) noexcept {
+        mUniformBindings[index] = { id, data, age };
+    }
+    auto getEs2UniformBinding(size_t index) const noexcept {
+        return mUniformBindings[index];
+    }

 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
    GLuint getSamplerSlow(SamplerParams sp) const noexcept;
@@ -509,11 +505,12 @@ private:
    std::vector<std::function<void(OpenGLContext&)>> mDestroyWithNormalContext;
    RenderPrimitive mDefaultVAO;
    std::optional<GLuint> mDefaultFbo[2];
+    std::array<
+            std::tuple<GLuint, void const*, uint16_t>,
+            CONFIG_UNIFORM_BINDING_COUNT> mUniformBindings = {};
    mutable tsl::robin_map<SamplerParams, GLuint,
            SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;

-    Platform::DriverConfig const mDriverConfig;
-
    void bindFramebufferResolved(GLenum target, GLuint buffer) noexcept;

    const std::array<std::tuple<bool const&, char const*, char const*>, sizeof(bugs)> mBugDatabase{{
@@ -559,9 +556,6 @@ private:
            {   bugs.delay_fbo_destruction,
                    "delay_fbo_destruction",
                    ""},
-            {   bugs.rebind_buffer_after_deletion,
-                    "rebind_buffer_after_deletion",
-                    ""},
            {   bugs.force_feature_level0,
                    "force_feature_level0",
                    ""},
@@ -631,7 +625,6 @@ constexpr size_t OpenGLContext::getIndexForCap(GLenum cap) noexcept { //NOLINT
 #ifdef BACKEND_OPENGL_VERSION_GL
        case GL_PROGRAM_POINT_SIZE:             index = 10; break;
 #endif
-        case GL_DEPTH_CLAMP:                    index = 11; break;
        default: break;
    }
    assert_invariant(index < state.enables.caps.size());
@@ -21,8 +21,6 @@
 #include "OpenGLContext.h"
 #include "OpenGLTimerQuery.h"
 #include "GLBufferObject.h"
-#include "GLDescriptorSet.h"
-#include "GLDescriptorSetLayout.h"
 #include "GLTexture.h"
 #include "ShaderCompilerService.h"

@@ -38,7 +36,6 @@
 #include "private/backend/Driver.h"
 #include "private/backend/HandleAllocator.h"

-#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -55,7 +52,6 @@
 #include <tuple>
 #include <type_traits>
 #include <utility>
-#include <variant>
 #include <vector>

 #include <stddef.h>
@@ -85,7 +81,7 @@ public:
            const Platform::DriverConfig& driverConfig) noexcept;

    class DebugMarker {
-        UTILS_UNUSED OpenGLDriver& driver;
+        OpenGLDriver& driver;
    public:
        DebugMarker(OpenGLDriver& driver, const char* string) noexcept;
        ~DebugMarker() noexcept;
@@ -127,6 +123,16 @@ public:
        } gl;
    };

+    struct GLSamplerGroup : public HwSamplerGroup {
+        using HwSamplerGroup::HwSamplerGroup;
+        struct Entry {
+            GLTexture const* texture = nullptr;
+            GLuint sampler = 0u;
+        };
+        utils::FixedCapacityVector<Entry> textureUnitEntries;
+        explicit GLSamplerGroup(size_t size) noexcept : textureUnitEntries(size) { }
+    };
+
    struct GLRenderPrimitive : public HwRenderPrimitive {
        using HwRenderPrimitive::HwRenderPrimitive;
        OpenGLContext::RenderPrimitive gl;
@@ -139,10 +145,6 @@ public:

    using GLTimerQuery = filament::backend::GLTimerQuery;

-    using GLDescriptorSetLayout = filament::backend::GLDescriptorSetLayout;
-
-    using GLDescriptorSet = filament::backend::GLDescriptorSet;
-
    struct GLStream : public HwStream {
        using HwStream::HwStream;
        struct Info {
@@ -254,11 +256,6 @@ private:
        return mHandleAllocator.handle_cast<Dp, B>(handle);
    }

-    template<typename B>
-    bool is_valid(Handle<B>& handle) {
-        return mHandleAllocator.is_valid(handle);
-    }
-
    template<typename Dp, typename B>
    inline typename std::enable_if_t<
            std::is_pointer_v<Dp> &&
@@ -315,6 +312,10 @@ private:
    void resolvePass(ResolveAction action, GLRenderTarget const* rt,
            TargetBufferFlags discardFlags) noexcept;

+    const std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT>& getSamplerBindings() const {
+        return mSamplerBindings;
+    }
+
    using AttachmentArray = std::array<GLenum, MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT + 2>;
    static GLsizei getAttachments(AttachmentArray& attachments, TargetBufferFlags buffers,
            bool isDefaultFramebuffer) noexcept;
@@ -327,16 +328,8 @@ private:
    GLboolean mRenderPassStencilWrite{};

    GLRenderPrimitive const* mBoundRenderPrimitive = nullptr;
-    OpenGLProgram* mBoundProgram = nullptr;
    bool mValidProgram = false;
-    utils::bitset8 mInvalidDescriptorSetBindings;
-    utils::bitset8 mInvalidDescriptorSetBindingOffsets;
-    void updateDescriptors(utils::bitset8 invalidDescriptorSets) noexcept;

-    struct {
-        backend::DescriptorSetHandle dsh;
-        std::array<uint32_t, CONFIG_UNIFORM_BINDING_COUNT> offsets;
-    } mBoundDescriptorSets[MAX_DESCRIPTOR_SET_COUNT];

    void clearWithRasterPipe(TargetBufferFlags clearFlags,
            math::float4 const& linearColor, GLfloat depth, GLint stencil) noexcept;
@@ -348,6 +341,9 @@ private:
    // ES2 only. Uniform buffer emulation binding points
    GLuint mLastAssignedEmulatedUboId = 0;

+    // sampler buffer binding points (nullptr if not used)
+    std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT> mSamplerBindings = {};   // 4 pointers
+
    // this must be accessed from the driver thread only
    std::vector<GLTexture*> mTexturesWithStreamsAttached;

@@ -358,6 +354,8 @@ private:
    void detachStream(GLTexture* t) noexcept;
    void replaceStream(GLTexture* t, GLStream* stream) noexcept;

+    void updateTextureLodRange(GLTexture* texture, int8_t targetLevel) noexcept;
+
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
    // tasks executed on the main thread after the fence signaled
    void whenGpuCommandsComplete(const std::function<void()>& fn) noexcept;
@@ -381,7 +379,6 @@ private:
    bool mRec709OutputColorspace = false;

    PushConstantBundle* mCurrentPushConstants = nullptr;
-    PipelineLayout::SetLayout mCurrentSetLayout;
 };

 // ------------------------------------------------------------------------------------------------
@@ -17,30 +17,24 @@
 #include "OpenGLProgram.h"

 #include "GLUtils.h"
-#include "GLTexture.h"
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"

-#include <backend/DriverEnums.h>
 #include <backend/Program.h>
-#include <backend/Handle.h>

-#include <utils/BitmaskEnum.h>
-#include <utils/compiler.h>
+#include <private/backend/BackendUtils.h>
+
 #include <utils/debug.h>
-#include <utils/FixedCapacityVector.h>
+#include <utils/compiler.h>
 #include <utils/Log.h>
 #include <utils/Systrace.h>

-#include <algorithm>
 #include <array>
-#include <algorithm>
-#include <new>
 #include <string_view>
 #include <utility>
+#include <new>

 #include <stddef.h>
-#include <stdint.h>

 namespace filament::backend {

@@ -49,8 +43,9 @@ using namespace utils;
 using namespace backend;

 struct OpenGLProgram::LazyInitializationData {
-    Program::DescriptorSetInfo descriptorBindings;
-    Program::BindingUniformsInfo bindingUniformInfo;
+    Program::UniformBlockInfo uniformBlockInfo;
+    Program::SamplerGroupInfo samplerGroupInfo;
+    std::array<Program::UniformInfo, Program::UNIFORM_BINDING_COUNT> bindingUniformInfo;
    utils::FixedCapacityVector<Program::PushConstant> vertexPushConstants;
    utils::FixedCapacityVector<Program::PushConstant> fragmentPushConstants;
 };
@@ -59,14 +54,16 @@ struct OpenGLProgram::LazyInitializationData {
 OpenGLProgram::OpenGLProgram() noexcept = default;

 OpenGLProgram::OpenGLProgram(OpenGLDriver& gld, Program&& program) noexcept
-        : HwProgram(std::move(program.getName())), mRec709Location(-1) {
+        : HwProgram(std::move(program.getName())) {
    auto* const lazyInitializationData = new(std::nothrow) LazyInitializationData();
+    lazyInitializationData->samplerGroupInfo = std::move(program.getSamplerGroupInfo());
    if (UTILS_UNLIKELY(gld.getContext().isES2())) {
        lazyInitializationData->bindingUniformInfo = std::move(program.getBindingUniformInfo());
+    } else {
+        lazyInitializationData->uniformBlockInfo = std::move(program.getUniformBlockBindings());
    }
    lazyInitializationData->vertexPushConstants = std::move(program.getPushConstants(ShaderStage::VERTEX));
    lazyInitializationData->fragmentPushConstants = std::move(program.getPushConstants(ShaderStage::FRAGMENT));
-    lazyInitializationData->descriptorBindings = std::move(program.getDescriptorBindings());

    ShaderCompilerService& compiler = gld.getShaderCompilerService();
    mToken = compiler.createProgram(name, std::move(program));
@@ -124,86 +121,36 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra

    SYSTRACE_CALL();

-    // from the pipeline layout we compute a mapping from {set, binding} to {binding}
-    // for both buffers and textures
-
-    for (auto&& entry: lazyInitializationData.descriptorBindings) {
-        std::sort(entry.begin(), entry.end(),
-                [](Program::Descriptor const& lhs, Program::Descriptor const& rhs) {
-                    return lhs.binding < rhs.binding;
-                });
-    }
-
-    GLuint tmu = 0;
-    GLuint binding = 0;
-
-    // needed for samplers
-    context.useProgram(program);
-
-    UTILS_NOUNROLL
-    for (backend::descriptor_set_t set = 0; set < MAX_DESCRIPTOR_SET_COUNT; set++) {
-        for (Program::Descriptor const& entry: lazyInitializationData.descriptorBindings[set]) {
-            switch (entry.type) {
-                case DescriptorType::UNIFORM_BUFFER:
-                case DescriptorType::SHADER_STORAGE_BUFFER: {
-                    if (!entry.name.empty()) {
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-                        if (UTILS_LIKELY(!context.isES2())) {
-                            GLuint const index = glGetUniformBlockIndex(program,
-                                    entry.name.c_str());
-                            if (index != GL_INVALID_INDEX) {
-                                // this can fail if the program doesn't use this descriptor
-                                glUniformBlockBinding(program, index, binding);
-                                mBindingMap.insert(set, entry.binding,
-                                        { binding, entry.type });
-                                ++binding;
-                            }
-                        } else
-#endif
-                        {
-                            auto pos = std::find_if(lazyInitializationData.bindingUniformInfo.begin(),
-                                    lazyInitializationData.bindingUniformInfo.end(),
-                                    [&name = entry.name](const auto& item) {
-                                return std::get<1>(item) == name;
-                            });
-                            if (pos != lazyInitializationData.bindingUniformInfo.end()) {
-                                binding = std::get<0>(*pos);
-                                mBindingMap.insert(set, entry.binding, { binding, entry.type });
-                            }
-                        }
-                    }
-                    break;
+    if (!context.isES2()) {
+        // Note: This is only needed, because the layout(binding=) syntax is not permitted in glsl
+        // (ES3.0 and GL4.1). The backend needs a way to associate a uniform block to a binding point.
+        UTILS_NOUNROLL
+        for (GLuint binding = 0, n = lazyInitializationData.uniformBlockInfo.size();
+                binding < n; binding++) {
+            auto const& name = lazyInitializationData.uniformBlockInfo[binding];
+            if (!name.empty()) {
+                GLuint const index = glGetUniformBlockIndex(program, name.c_str());
+                if (index != GL_INVALID_INDEX) {
+                    glUniformBlockBinding(program, index, binding);
                }
-                case DescriptorType::SAMPLER: {
-                    if (!entry.name.empty()) {
-                        GLint const loc = glGetUniformLocation(program, entry.name.c_str());
-                        if (loc >= 0) {
-                            // this can fail if the program doesn't use this descriptor
-                            mBindingMap.insert(set, entry.binding, { tmu, entry.type });
-                            glUniform1i(loc, GLint(tmu));
-                            ++tmu;
-                        }
-                    }
-                    break;
-                }
-                case DescriptorType::INPUT_ATTACHMENT:
-                    break;
+                CHECK_GL_ERROR(utils::slog.e)
            }
-            CHECK_GL_ERROR(utils::slog.e)
        }
-    }
-
-    if (context.isES2()) {
+    } else
+#endif
+    {
        // ES2 initialization of (fake) UBOs
        UniformsRecord* const uniformsRecords = new(std::nothrow) UniformsRecord[Program::UNIFORM_BINDING_COUNT];
        UTILS_NOUNROLL
-        for (auto&& [index, name, uniforms] : lazyInitializationData.bindingUniformInfo) {
-            uniformsRecords[index].locations.reserve(uniforms.size());
-            uniformsRecords[index].locations.resize(uniforms.size());
+        for (GLuint binding = 0, n = Program::UNIFORM_BINDING_COUNT; binding < n; binding++) {
+            Program::UniformInfo& uniforms = lazyInitializationData.bindingUniformInfo[binding];
+            uniformsRecords[binding].locations.reserve(uniforms.size());
+            uniformsRecords[binding].locations.resize(uniforms.size());
            for (size_t j = 0, c = uniforms.size(); j < c; j++) {
                GLint const loc = glGetUniformLocation(program, uniforms[j].name.c_str());
-                uniformsRecords[index].locations[j] = loc;
-                if (UTILS_UNLIKELY(index == 0)) {
+                uniformsRecords[binding].locations[j] = loc;
+                if (UTILS_UNLIKELY(binding == 0)) {
                    // This is a bit of a gross hack here, we stash the location of
                    // "frameUniforms.rec709", which obviously the backend shouldn't know about,
                    // which is used for emulating the "rec709" colorspace in the shader.
@@ -215,11 +162,51 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
                    }
                }
            }
-            uniformsRecords[index].uniforms = std::move(uniforms);
+            uniformsRecords[binding].uniforms = std::move(uniforms);
        }
        mUniformsRecords = uniformsRecords;
    }

+    uint8_t usedBindingCount = 0;
+    uint8_t tmu = 0;
+
+    UTILS_NOUNROLL
+    for (size_t i = 0, c = lazyInitializationData.samplerGroupInfo.size(); i < c; i++) {
+        auto const& samplers = lazyInitializationData.samplerGroupInfo[i].samplers;
+        if (samplers.empty()) {
+            // this binding point doesn't have any samplers, skip it.
+            continue;
+        }
+
+        // keep this in the loop, so we skip it in the rare case a program doesn't have
+        // sampler. The context cache will prevent repeated calls to GL.
+        context.useProgram(program);
+
+        bool atLeastOneSamplerUsed = false;
+        UTILS_NOUNROLL
+        for (const Program::Sampler& sampler: samplers) {
+            // find its location and associate a TMU to it
+            GLint const loc = glGetUniformLocation(program, sampler.name.c_str());
+            if (loc >= 0) {
+                // this can fail if the program doesn't use this sampler
+                glUniform1i(loc, tmu);
+                atLeastOneSamplerUsed = true;
+            }
+            tmu++;
+        }
+
+        // if this program doesn't use any sampler from this HwSamplerGroup, just cancel the
+        // whole group.
+        if (atLeastOneSamplerUsed) {
+            // Cache the sampler uniform locations for each interface block
+            mUsedSamplerBindingPoints[usedBindingCount] = i;
+            usedBindingCount++;
+        } else {
+            tmu -= samplers.size();
+        }
+    }
+    mUsedBindingsCount = usedBindingCount;
+
    auto& vertexConstants = lazyInitializationData.vertexPushConstants;
    auto& fragmentConstants = lazyInitializationData.fragmentPushConstants;

@@ -236,8 +223,40 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
    }
 }

-void OpenGLProgram::updateUniforms(
-        uint32_t index, GLuint id, void const* buffer, uint16_t age) const noexcept {
+void OpenGLProgram::updateSamplers(OpenGLDriver* const gld) const noexcept {
+    using GLTexture = OpenGLDriver::GLTexture;
+
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+    bool const es2 = gld->getContext().isES2();
+#endif
+
+    // cache a few member variable locally, outside the loop
+    auto const& UTILS_RESTRICT samplerBindings = gld->getSamplerBindings();
+    auto const& UTILS_RESTRICT usedBindingPoints = mUsedSamplerBindingPoints;
+
+    for (uint8_t i = 0, tmu = 0, n = mUsedBindingsCount; i < n; i++) {
+        size_t const binding = usedBindingPoints[i];
+        assert_invariant(binding < Program::SAMPLER_BINDING_COUNT);
+        auto const * const sb = samplerBindings[binding];
+        assert_invariant(sb);
+        if (!sb) continue; // should never happen, this would be a user error.
+        for (uint8_t j = 0, m = sb->textureUnitEntries.size(); j < m; ++j, ++tmu) { // "<=" on purpose here
+            const GLTexture* const t = sb->textureUnitEntries[j].texture;
+            if (t) { // program may not use all samplers of sampler group
+                gld->bindTexture(tmu, t);
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+                if (UTILS_LIKELY(!es2)) {
+                    GLuint const s = sb->textureUnitEntries[j].sampler;
+                    gld->bindSampler(tmu, s);
+                }
+#endif
+            }
+        }
+    }
+    CHECK_GL_ERROR(utils::slog.e)
+}
+
+void OpenGLProgram::updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept {
    assert_invariant(mUniformsRecords);
    assert_invariant(buffer);

@@ -19,20 +19,17 @@

 #include "DriverBase.h"

-#include "BindingMap.h"
 #include "OpenGLContext.h"
 #include "ShaderCompilerService.h"

 #include <private/backend/Driver.h>
-
-#include <backend/DriverEnums.h>
 #include <backend/Program.h>

-#include <utils/bitset.h>
 #include <utils/compiler.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Slice.h>

+#include <array>
 #include <limits>

 #include <stddef.h>
@@ -56,41 +53,37 @@ public:

    bool isValid() const noexcept { return mToken || gl.program != 0; }

-    bool use(OpenGLDriver* const gld, OpenGLContext& context) noexcept {
-        // both non-null is impossible by construction
-        assert_invariant(!mToken || !gl.program);
-
-        if (UTILS_UNLIKELY(mToken && !gl.program)) {
-            // first time a program is used
+    void use(OpenGLDriver* const gld, OpenGLContext& context) noexcept {
+        if (UTILS_UNLIKELY(!gl.program)) {
            initialize(*gld);
        }

-        if (UTILS_UNLIKELY(!gl.program)) {
-            // compilation failed (token should be null)
-            assert_invariant(!mToken);
-            return false;
-        }
-
        context.useProgram(gl.program);
-        return true;
-    }
+        if (UTILS_UNLIKELY(mUsedBindingsCount)) {
+            // We rely on GL state tracking to avoid unnecessary glBindTexture / glBindSampler
+            // calls.

-    GLuint getBufferBinding(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
-        return mBindingMap.get(set, binding);
-    }
+            // we need to do this if:
+            // - the content of mSamplerBindings has changed
+            // - the content of any bound sampler buffer has changed
+            // ... since last time we used this program

-    GLuint getTextureUnit(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
-        return mBindingMap.get(set, binding);
-    }
+            // Turns out the former might be relatively cheap to check, the latter requires
+            // a bit less. Compared to what updateSamplers() actually does, which is
+            // pretty little, I'm not sure if we'll get ahead.

-    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
-        return mBindingMap.getActiveDescriptors(set);
+            updateSamplers(gld);
+        }
    }

    // For ES2 only
-    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) const noexcept;
+    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept;
    void setRec709ColorSpace(bool rec709) const noexcept;

+    struct {
+        GLuint program = 0;
+    } gl;                                               // 4 bytes
+
    PushConstantBundle getPushConstants() {
        auto fragBegin = mPushConstants.begin() + mPushConstantFragmentStageOffset;
        return {
@@ -108,15 +101,22 @@ private:
    void initializeProgramState(OpenGLContext& context, GLuint program,
            LazyInitializationData& lazyInitializationData) noexcept;

-    BindingMap mBindingMap;     // 8 bytes + out-of-line 256 bytes
+    void updateSamplers(OpenGLDriver* gld) const noexcept;
+
+    // number of bindings actually used by this program
+    std::array<uint8_t, Program::SAMPLER_BINDING_COUNT> mUsedSamplerBindingPoints;   // 4 bytes

    ShaderCompilerService::program_token_t mToken{};    // 16 bytes

-    // Note that this can be replaced with a raw pointer and an uint8_t (for size) to reduce the
-    // size of the container to 9 bytes if there is a need in the future.
-    utils::FixedCapacityVector<std::pair<GLint, ConstantType>> mPushConstants;// 16 bytes
+    uint8_t mUsedBindingsCount = 0u;                    // 1 byte
+    UTILS_UNUSED uint8_t padding[2] = {};               // 2 byte
+
+    // Push constant array offset for fragment stage constants.
+    uint8_t mPushConstantFragmentStageOffset = 0u;      // 1 byte

    // only needed for ES2
+    GLint mRec709Location = -1;                         // 4 bytes
+
    using LocationInfo = utils::FixedCapacityVector<GLint>;
    struct UniformsRecord {
        Program::UniformInfo uniforms;
@@ -124,20 +124,15 @@ private:
        mutable GLuint id = 0;
        mutable uint16_t age = std::numeric_limits<uint16_t>::max();
    };
-    UniformsRecord const* mUniformsRecords = nullptr;
-    GLint mRec709Location : 24;     // 4 bytes
+    UniformsRecord const* mUniformsRecords = nullptr;               // 8 bytes

-    // Push constant array offset for fragment stage constants.
-    GLint mPushConstantFragmentStageOffset : 8;      // 1 byte
-
-public:
-    struct {
-        GLuint program = 0;
-    } gl;                                               // 4 bytes
+    // Note that this can be replaced with a raw pointer and an uint8_t (for size) to reduce the
+    // size of the container to 9 bytes if there is a need in the future.
+    utils::FixedCapacityVector<std::pair<GLint, ConstantType>> mPushConstants;// 16 bytes
 };

-// if OpenGLProgram is larger than 96 bytes, it'll fall in a larger Handle bucket.
-static_assert(sizeof(OpenGLProgram) <= 96); // currently 96 bytes
+// if OpenGLProgram is larger tha 64 bytes, it'll fall in a larger Handle bucket.
+static_assert(sizeof(OpenGLProgram) <= 64); // currently 64 bytes

 } // namespace filament::backend

@@ -143,25 +143,21 @@ void TimerQueryNativeFactory::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQ

    driver.runEveryNowAndThen([&context = mContext, weak]() -> bool {
        auto state = weak.lock();
-        if (!state) {
-            // The timer query state has been destroyed on the way, very likely due to the IBL
-            // prefilter context destruction. We still return true to get this element removed from
-            // the query list.
-            return true;
+        if (state) {
+            GLuint available = 0;
+            context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
+            CHECK_GL_ERROR(utils::slog.e)
+            if (!available) {
+                // we need to try this one again later
+                return false;
+            }
+            GLuint64 elapsedTime = 0;
+            // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
+            context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
+            state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
+        } else {
+            state->elapsed.store(int64_t(TimerQueryResult::ERROR), std::memory_order_relaxed);
        }
-
-        GLuint available = 0;
-        context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
-        CHECK_GL_ERROR(utils::slog.e)
-        if (!available) {
-            // we need to try this one again later
-            return false;
-        }
-        GLuint64 elapsedTime = 0;
-        // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
-        context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
-        state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
-
        return true;
    });
 }
@@ -26,28 +26,16 @@

 #include <utils/compiler.h>
 #include <utils/CString.h>
-#include <utils/debug.h>
-#include <utils/FixedCapacityVector.h>
 #include <utils/JobSystem.h>
 #include <utils/Log.h>
-#include <utils/ostream.h>
-#include <utils/Panic.h>
 #include <utils/Systrace.h>

-#include <array>
 #include <cctype>
 #include <chrono>
-#include <mutex>
-#include <memory>
 #include <string>
 #include <string_view>
-#include <thread>
-#include <utility>
 #include <variant>

-#include <stddef.h>
-#include <stdint.h>
-
 namespace filament::backend {

 using namespace utils;
@@ -371,7 +359,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
 GLuint ShaderCompilerService::getProgram(ShaderCompilerService::program_token_t& token) {
    GLuint const program = initialize(token);
    assert_invariant(token == nullptr);
-#if !FILAMENT_ENABLE_MATDBG
+#ifndef FILAMENT_ENABLE_MATDBG
    assert_invariant(program);
 #endif
    return program;
@@ -488,12 +476,6 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
    // check status of program linking and shader compilation, logs error and free all resources
    // in case of error.
    bool const success = checkProgramStatus(token);
-
-    // Unless we have matdbg, we panic if a program is invalid. Otherwise, we'd get a UB.
-    // The compilation error has been logged to log.e by this point.
-    FILAMENT_CHECK_POSTCONDITION(FILAMENT_ENABLE_MATDBG || success)
-            << "OpenGL program " << token->name.c_str_safe() << " failed to link or compile";
-
    if (UTILS_LIKELY(success)) {
        program = token->gl.program;
        // no need to keep the shaders around
@@ -590,38 +572,24 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,

            // split shader source, so we can insert the specialization constants and the packing
            // functions
-            auto [version, prolog, body] = splitShaderSource({ shader_src, shader_len });
+            auto const [prolog, body] = splitShaderSource({ shader_src, shader_len });

-            // enable ESSL 3.10 if available
-            if (context.isAtLeastGLES<3, 1>()) {
-                version = "#version 310 es\n";
-            }
-
-            std::array<std::string_view, 5> sources = {
-                version,
-                prolog,
-                specializationConstantString,
-                packingFunctions,
-                { body.data(), body.size() - 1 }  // null-terminated
+            const std::array<const char*, 4> sources = {
+                    prolog.data(),
+                    specializationConstantString.c_str(),
+                    packingFunctions.data(),
+                    body.data()
            };

-            // Some of the sources may be zero-length. Remove them as to avoid passing lengths of
-            // zero to glShaderSource(). glShaderSource should work with lengths of zero, but some
-            // drivers instead interpret zero as a sentinel for a null-terminated string.
-            auto partitionPoint = std::stable_partition(
-                    sources.begin(), sources.end(), [](std::string_view s) { return !s.empty(); });
-            size_t count = std::distance(sources.begin(), partitionPoint);
-
-            std::array<const char*, 5> shaderStrings;
-            std::array<GLint, 5> lengths;
-            for (size_t i = 0; i < count; i++) {
-                shaderStrings[i] = sources[i].data();
-                lengths[i] = sources[i].size();
-            }
+            const std::array<GLint, 4> lengths = {
+                    (GLint)prolog.length(),
+                    (GLint)specializationConstantString.length(),
+                    (GLint)packingFunctions.length(),
+                    (GLint)body.length() - 1 // null terminated
+            };

            GLuint const shaderId = glCreateShader(glShaderType);
-            glShaderSource(shaderId, count, shaderStrings.data(), lengths.data());
-
+            glShaderSource(shaderId, sources.size(), sources.data(), lengths.data());
            glCompileShader(shaderId);

 #ifndef NDEBUG
@@ -693,7 +661,6 @@ void ShaderCompilerService::process_OVR_multiview2(OpenGLContext& context,

 // Tragically, OpenGL 4.1 doesn't support unpackHalf2x16 (appeared in 4.2) and
 // macOS doesn't support GL_ARB_shading_language_packing
-// Also GLES3.0 didn't have the full set of packing/unpacking functions
 std::string_view ShaderCompilerService::process_ARB_shading_language_packing(OpenGLContext& context) noexcept {
    using namespace std::literals;
 #ifdef BACKEND_OPENGL_VERSION_GL
@@ -733,102 +700,31 @@ highp uint packHalf2x16(vec2 v) {
    highp uint y = fp32tou16(v.y);
    return (y << 16u) | x;
 }
-highp uint packUnorm4x8(mediump vec4 v) {
-    v = round(clamp(v, 0.0, 1.0) * 255.0);
-    highp uint a = uint(v.x);
-    highp uint b = uint(v.y) <<  8;
-    highp uint c = uint(v.z) << 16;
-    highp uint d = uint(v.w) << 24;
-    return (a|b|c|d);
-}
-highp uint packSnorm4x8(mediump vec4 v) {
-    v = round(clamp(v, -1.0, 1.0) * 127.0);
-    highp uint a = uint((int(v.x) & 0xff));
-    highp uint b = uint((int(v.y) & 0xff)) <<  8;
-    highp uint c = uint((int(v.z) & 0xff)) << 16;
-    highp uint d = uint((int(v.w) & 0xff)) << 24;
-    return (a|b|c|d);
-}
-mediump vec4 unpackUnorm4x8(highp uint v) {
-    return vec4(float((v & 0x000000ffu)      ),
-                float((v & 0x0000ff00u) >>  8),
-                float((v & 0x00ff0000u) >> 16),
-                float((v & 0xff000000u) >> 24)) / 255.0;
-}
-mediump vec4 unpackSnorm4x8(highp uint v) {
-    int a = int(((v       ) & 0xffu) << 24u) >> 24 ;
-    int b = int(((v >>  8u) & 0xffu) << 24u) >> 24 ;
-    int c = int(((v >> 16u) & 0xffu) << 24u) >> 24 ;
-    int d = int(((v >> 24u) & 0xffu) << 24u) >> 24 ;
-    return clamp(vec4(float(a), float(b), float(c), float(d)) / 127.0, -1.0, 1.0);
-}
 )"sv;
    }
 #endif // BACKEND_OPENGL_VERSION_GL
-
-#ifdef BACKEND_OPENGL_VERSION_GLES
-    if (!context.isES2() && !context.isAtLeastGLES<3, 1>()) {
-        return R"(
-
-highp uint packUnorm4x8(mediump vec4 v) {
-    v = round(clamp(v, 0.0, 1.0) * 255.0);
-    highp uint a = uint(v.x);
-    highp uint b = uint(v.y) <<  8;
-    highp uint c = uint(v.z) << 16;
-    highp uint d = uint(v.w) << 24;
-    return (a|b|c|d);
-}
-highp uint packSnorm4x8(mediump vec4 v) {
-    v = round(clamp(v, -1.0, 1.0) * 127.0);
-    highp uint a = uint((int(v.x) & 0xff));
-    highp uint b = uint((int(v.y) & 0xff)) <<  8;
-    highp uint c = uint((int(v.z) & 0xff)) << 16;
-    highp uint d = uint((int(v.w) & 0xff)) << 24;
-    return (a|b|c|d);
-}
-mediump vec4 unpackUnorm4x8(highp uint v) {
-    return vec4(float((v & 0x000000ffu)      ),
-                float((v & 0x0000ff00u) >>  8),
-                float((v & 0x00ff0000u) >> 16),
-                float((v & 0xff000000u) >> 24)) / 255.0;
-}
-mediump vec4 unpackSnorm4x8(highp uint v) {
-    int a = int(((v       ) & 0xffu) << 24u) >> 24 ;
-    int b = int(((v >>  8u) & 0xffu) << 24u) >> 24 ;
-    int c = int(((v >> 16u) & 0xffu) << 24u) >> 24 ;
-    int d = int(((v >> 24u) & 0xffu) << 24u) >> 24 ;
-    return clamp(vec4(float(a), float(b), float(c), float(d)) / 127.0, -1.0, 1.0);
-}
-)"sv;
-    }
-#endif // BACKEND_OPENGL_VERSION_GLES
    return ""sv;
 }

-// split shader source code in three:
-// - the version line
-// - extensions
-// - everything else
-std::array<std::string_view, 3> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
-    auto version_start = source.find("#version");
-    assert_invariant(version_start != std::string_view::npos);
+// split shader source code in two, the first section goes from the start to the line after the
+// last #extension, and the 2nd part goes from there to the end.
+std::array<std::string_view, 2> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
+    auto start = source.find("#version");
+    assert_invariant(start != std::string_view::npos);

-    auto version_eol = source.find('\n', version_start) + 1;
-    assert_invariant(version_eol != std::string_view::npos);
-
-    auto prolog_start = version_eol;
-    auto prolog_eol = source.rfind("\n#extension"); // last #extension line
-    if (prolog_eol == std::string_view::npos) {
-        prolog_eol = prolog_start;
+    auto pos = source.rfind("\n#extension");
+    if (pos == std::string_view::npos) {
+        pos = start;
    } else {
-        prolog_eol = source.find('\n', prolog_eol + 1) + 1;
+        ++pos;
    }
-    auto body_start = prolog_eol;

-    std::string_view const version = source.substr(version_start, version_eol - version_start);
-    std::string_view const prolog = source.substr(prolog_start, prolog_eol - prolog_start);
-    std::string_view const body = source.substr(body_start, source.length() - body_start);
-    return { version, prolog, body };
+    auto eol = source.find('\n', pos) + 1;
+    assert_invariant(eol != std::string_view::npos);
+
+    std::string_view const version = source.substr(start, eol - start);
+    std::string_view const body = source.substr(version.length(), source.length() - version.length());
+    return { version, body };
 }

 /*
@@ -146,7 +146,7 @@ private:

    static std::string_view process_ARB_shading_language_packing(OpenGLContext& context) noexcept;

-    static std::array<std::string_view, 3> splitShaderSource(std::string_view source) noexcept;
+    static std::array<std::string_view, 2> splitShaderSource(std::string_view source) noexcept;

    static GLuint linkProgram(OpenGLContext& context,
            std::array<GLuint, Program::SHADER_TYPE_COUNT> shaders,
@@ -201,12 +201,6 @@ using namespace glext;
 #   define GL_CLIP_DISTANCE1                        0x3001
 #endif

-#if defined(GL_EXT_depth_clamp)
-#   define GL_DEPTH_CLAMP                           GL_DEPTH_CLAMP_EXT
-#else
-#   define GL_DEPTH_CLAMP                           0x864F
-#endif
-
 #if defined(GL_KHR_debug)
 #   define GL_DEBUG_OUTPUT                          GL_DEBUG_OUTPUT_KHR
 #   define GL_DEBUG_OUTPUT_SYNCHRONOUS              GL_DEBUG_OUTPUT_SYNCHRONOUS_KHR
@@ -111,8 +111,8 @@ bool CocoaExternalImage::set(CVPixelBufferRef image) noexcept {
    }

    OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_POSTCONDITION(formatType == kCVPixelFormatType_32BGRA)
-            << "macOS external images must be 32BGRA format.";
+    ASSERT_POSTCONDITION(formatType == kCVPixelFormatType_32BGRA,
+            "macOS external images must be 32BGRA format.");

    // The pixel buffer must be locked whenever we do rendering with it. We'll unlock it before
    // releasing.
@@ -135,14 +135,13 @@ bool CocoaTouchExternalImage::set(CVPixelBufferRef image) noexcept {
    }

    OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_POSTCONDITION(formatType == kCVPixelFormatType_32BGRA ||
-            formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange)
-            << "iOS external images must be in either 32BGRA or 420f format.";
+    ASSERT_POSTCONDITION(formatType == kCVPixelFormatType_32BGRA ||
+                         formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange,
+            "iOS external images must be in either 32BGRA or 420f format.");

    size_t planeCount = CVPixelBufferGetPlaneCount(image);
-    FILAMENT_CHECK_POSTCONDITION(planeCount == 0 || planeCount == 2)
-            << "The OpenGL backend does not support images with plane counts of " << planeCount
-            << ".";
+    ASSERT_POSTCONDITION(planeCount == 0 || planeCount == 2,
+            "The OpenGL backend does not support images with plane counts of %d.", planeCount);

    // The pixel buffer must be locked whenever we do rendering with it. We'll unlock it before
    // releasing.
@@ -162,7 +162,7 @@ Driver* PlatformCocoaGL::createDriver(void* sharedContext, const Platform::Drive
    pImpl->mGLContext = nsOpenGLContext;

    int result = bluegl::bind();
-    FILAMENT_CHECK_POSTCONDITION(!result) << "Unable to load OpenGL entry points.";
+    ASSERT_POSTCONDITION(!result, "Unable to load OpenGL entry points.");

    UTILS_UNUSED_IN_RELEASE CVReturn success = CVOpenGLTextureCacheCreate(kCFAllocatorDefault, nullptr,
            [pImpl->mGLContext CGLContextObj], [pImpl->mGLContext.pixelFormat CGLPixelFormatObj], nullptr,
--- a/Show More
+++ b/Show More