vk: testing a robin_map key hash failure

Executable on device with:

./build.sh -p android -q arm64-v8a release test_compiler && adb push out/cmake-android-release-aarch64/filament/test/test_compiler /data/local/tmp/ && adb shell chmod 777 /data/local/tmp/test_compiler && adb shell /data/local/tmp/test_compiler

.github/actions/android-continuous/action.yml

@@ -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

.github/actions/ubuntu-apt-add-src/action.yml

@@ -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

.github/workflows/android-continuous.yml

@@ -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/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

.github/workflows/ios-continuous.yml

@@ -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
       - 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

.github/workflows/linux-continuous.yml

@@ -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
       - 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

.github/workflows/mac-continuous.yml

@@ -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
       - 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

.github/workflows/presubmit.yml

@@ -15,7 +15,7 @@ jobs:
 
     strategy:
       matrix:
-        os: [macos-14-xlarge, ubuntu-22.04-16core]
+        os: [macos-14, ubuntu-22.04]
 
     steps:
       - uses: actions/checkout@v4.1.6
@@ -29,7 +29,7 @@ jobs:
 
   build-windows:
     name: build-windows
-    runs-on: win-2019-16core
+    runs-on: windows-2019
 
     steps:
       - uses: actions/checkout@v4.1.6
@@ -40,7 +40,7 @@ jobs:
 
   build-android:
     name: build-android
-    runs-on: ubuntu-22.04-16core
+    runs-on: macos-14
 
     steps:
       - uses: actions/checkout@v4.1.6
@@ -49,14 +49,12 @@ jobs:
           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
@@ -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
       - 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

.github/workflows/release.yml

@@ -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
@@ -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:
@@ -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:
@@ -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:
@@ -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:
@@ -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

.github/workflows/web-continuous.yml

@@ -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
       - 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

.github/workflows/windows-continuous.yml

@@ -10,7 +10,7 @@ on:
 jobs:
   build-windows:
     name: build-windows
-    runs-on: windows-2019-32core
+    runs-on: windows-2019
 
     steps:
       - uses: actions/checkout@v4.1.6
@@ -18,7 +18,7 @@ jobs:
         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

CMakeLists.txt

@@ -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)

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.55.1'
+    implementation 'com.google.android.filament:filament-android:1.53.2'
 }
 ```
 
@@ -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', '~> 1.53.2'
 ```
 
+### 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

RELEASE_NOTES.md

@@ -7,42 +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.

android/filament-android/src/main/cpp/Engine.cpp

@@ -520,12 +520,13 @@ 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 +535,7 @@ 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 +544,6 @@ 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);
 }

android/filament-android/src/main/java/com/google/android/filament/Engine.java

@@ -224,12 +224,13 @@ 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;
         }
 
@@ -418,12 +419,12 @@ 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.
@@ -431,12 +432,12 @@ public class Engine {
          */
         public long resourceAllocatorCacheMaxAge = 1;
 
-        /**
+        /*
          * 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 +459,12 @@ 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) {
@@ -1412,11 +1406,12 @@ 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);

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.55.1
+VERSION_NAME=1.53.2
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

build.sh

@@ -64,9 +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."
@@ -183,8 +180,6 @@ BACKEND_DEBUG_FLAG_OPTION=""
 
 STEREOSCOPIC_OPTION=""
 
-OSMESA_OPTION=""
-
 IOS_BUILD_SIMULATOR=false
 BUILD_UNIVERSAL_LIBRARIES=false
 
@@ -245,7 +240,6 @@ function build_desktop_target {
             ${ASAN_UBSAN_OPTION} \
             ${BACKEND_DEBUG_FLAG_OPTION} \
             ${STEREOSCOPIC_OPTION} \
-            ${OSMESA_OPTION} \
             ${architectures} \
             ../..
         ln -sf "out/cmake-${lc_target}/compile_commands.json" \
@@ -802,7 +796,7 @@ 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:S:" opt; do
     case ${opt} in
         h)
             print_help
@@ -956,8 +950,6 @@ while getopts ":hacCfgijmp:q:uvslwedk:bx:S:X:" opt; do
                     exit 1
             esac
             ;;
-        X)  OSMESA_OPTION="-DFILAMENT_OSMESA_PATH=${OPTARG}"
-            ;;
         \?)
             echo "Invalid option: -${OPTARG}" >&2
             echo ""

build/android/build.sh

@@ -60,7 +60,13 @@ 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"
 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 +74,5 @@ 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

build/web/ci-common.sh

@@ -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

build/windows/build-github.bat

@@ -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%

docs/Materials.md.html

@@ -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,

filament/CMakeLists.txt

@@ -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

filament/backend/CMakeLists.txt

@@ -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)
@@ -368,11 +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)
-endif()
-
 if (MSVC)
     set(FILAMENT_WARNINGS /W3)
 else()
@@ -393,7 +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}>
 )
@@ -403,7 +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}>
 )
 
@@ -438,7 +424,6 @@ if (APPLE OR LINUX)
         test/test_StencilBuffer.cpp
         test/test_Scissor.cpp
         test/test_MipLevels.cpp
-        test/test_Handles.cpp
     )
     set(BACKEND_TEST_LIBS
         backend
@@ -510,40 +495,21 @@ 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()

filament/backend/include/backend/DescriptorSetOffsetArray.h

@@ -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

filament/backend/include/backend/DriverEnums.h

@@ -19,16 +19,13 @@
 #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 +97,6 @@ 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
@@ -196,70 +191,6 @@ 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 +270,15 @@ 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 +368,6 @@ 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 +693,9 @@ 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 +887,6 @@ 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 +895,6 @@ 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 +902,6 @@ 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 +1058,7 @@ 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 +1069,32 @@ 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.
@@ -1320,7 +1244,7 @@ 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
@@ -1335,8 +1259,6 @@ 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>

filament/backend/include/backend/Handle.h

@@ -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
 

filament/backend/include/backend/PipelineState.h

@@ -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

filament/backend/include/backend/Platform.h

@@ -68,6 +68,13 @@ public:
          */
         size_t handleArenaSize = 0;
 
+        /**
+         * 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;
+
         size_t metalUploadBufferSizeBytes = 512 * 1024;
 
         /**
@@ -91,13 +98,6 @@ public:
          * 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;

filament/backend/include/backend/Program.h

@@ -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.

filament/backend/include/backend/TargetBufferInfo.h

@@ -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 }} {
     }
 };
 

filament/backend/include/backend/platforms/PlatformEGLAndroid.h

@@ -89,11 +89,6 @@ protected:
      */
     AcquiredImage transformAcquiredImage(AcquiredImage source) noexcept override;
 
-protected:
-    bool makeCurrent(ContextType type,
-            SwapChain* drawSwapChain,
-            SwapChain* readSwapChain) noexcept override;
-
 private:
     struct InitializeJvmForPerformanceManagerIfNeeded {
         InitializeJvmForPerformanceManagerIfNeeded();
@@ -107,10 +102,6 @@ private:
 
     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

filament/backend/include/backend/platforms/PlatformOSMesa.h

@@ -1,64 +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;
-    void* mOsMesaApi = nullptr;
-};
-
-} // namespace filament::backend
-
-#endif // TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_OSMESA_H

filament/backend/include/private/backend/Driver.h

@@ -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>

filament/backend/include/private/backend/DriverAPI.inc

@@ -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
@@ -348,7 +305,6 @@ 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)
 

filament/backend/include/private/backend/DriverApi.h

@@ -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

filament/backend/include/private/backend/HandleAllocator.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 {
 
@@ -169,31 +168,13 @@ 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();
-                }
+                FILAMENT_CHECK_POSTCONDITION(expectedAge == age) <<
+                        "use-after-free of Handle with id=" << handle.getId();
             }
         }
 
@@ -202,18 +183,14 @@ public:
 
     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())) {
+        if (handle && isPoolHandle(handle.getId())) {
+            auto [p, tag] = handleToPointer(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;
+        return true;
     }
 
     template<typename Dp, typename B>
@@ -224,29 +201,6 @@ 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>
@@ -364,24 +318,12 @@ 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 +338,7 @@ 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 +353,7 @@ 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 +363,6 @@ 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;
 };

filament/backend/src/CircularBuffer.cpp

@@ -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

filament/backend/src/CommandBufferQueue.cpp

@@ -101,8 +101,9 @@ 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);
+    mCommandBuffersToExecute.push_back({ begin, end });
+    mCondition.notify_one();
 
     // circular buffer is too small, we corrupted the stream
     FILAMENT_CHECK_POSTCONDITION(used <= mFreeSpace) <<
@@ -111,13 +112,11 @@ void CommandBufferQueue::flush() noexcept {
             "Space used at this time: " << used <<
             " bytes, overflow: " << used - mFreeSpace << " bytes";
 
-    mFreeSpace -= used;
-    mCommandBuffersToExecute.push_back({ begin, end });
-    mCondition.notify_one();
-
     // 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): "
@@ -154,10 +153,8 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
 }
 
 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();
 }
 

filament/backend/src/CommandStream.cpp

@@ -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);

filament/backend/src/DriverBase.h

@@ -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;
 };

filament/backend/src/HandleAllocator.cpp

@@ -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>

filament/backend/src/PlatformFactory.cpp

@@ -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

filament/backend/src/Program.cpp

@@ -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;
 }

filament/backend/src/metal/MetalBuffer.h

@@ -73,11 +73,10 @@ public:
     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);
@@ -89,8 +88,6 @@ public:
                 return 1;
             case Type::STAGING:
                 return 2;
-            case Type::DESCRIPTOR_SET:
-                return 3;
         }
     }
 
@@ -163,8 +160,6 @@ 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 +169,8 @@ 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 +180,7 @@ public:
      * is no device allocation.
      *
      */
-    id<MTLBuffer> getGpuBufferForDraw() noexcept;
+    id<MTLBuffer> getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept;
 
     void* getCpuBuffer() const noexcept { return mCpuBuffer; }
 
@@ -216,10 +209,8 @@ private:
         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;
+    void uploadWithPoolBuffer(void* src, size_t size, size_t byteOffset) const;
+    void uploadWithBumpAllocator(void* src, size_t size, size_t byteOffset) const;
 
     UploadStrategy mUploadStrategy;
     TrackedMetalBuffer mBuffer;

filament/backend/src/metal/MetalBuffer.mm

@@ -40,15 +40,12 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
     // 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.
     {
@@ -56,8 +53,8 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
         mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
             TrackedMetalBuffer::Type::GENERIC };
     }
-    // mBuffer might fail to be allocated. Clients can check for this by calling
-    // wasAllocationSuccessful().
+    FILAMENT_CHECK_POSTCONDITION(mBuffer)
+            << "Could not allocate Metal buffer of size " << size << ".";
 }
 
 MetalBuffer::~MetalBuffer() {
@@ -66,20 +63,15 @@ 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;
     }
-
-    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();
+            << " at offset " << byteOffset;
     // 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();
+    FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3)) << "byteOffset must be a multiple of 4";
 
     // If we have a cpu buffer, we can directly copy into it.
     if (mCpuBuffer) {
@@ -89,21 +81,20 @@ void MetalBuffer::copyIntoBuffer(
 
     switch (mUploadStrategy) {
         case UploadStrategy::BUMP_ALLOCATOR:
-            uploadWithBumpAllocator(src, size, byteOffset, std::move(getHandleTag));
+            uploadWithBumpAllocator(src, size, byteOffset);
             break;
         case UploadStrategy::POOL:
-            uploadWithPoolBuffer(src, size, byteOffset, std::move(getHandleTag));
+            uploadWithPoolBuffer(src, size, byteOffset);
             break;
     }
 }
 
-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) {
@@ -146,7 +137,7 @@ 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;
         }
@@ -206,13 +197,9 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
     }
 }
 
-void MetalBuffer::uploadWithPoolBuffer(
-        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const {
+void MetalBuffer::uploadWithPoolBuffer(void* src, size_t size, size_t byteOffset) 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.
@@ -230,18 +217,10 @@ void MetalBuffer::uploadWithPoolBuffer(
     }];
 }
 
-void MetalBuffer::uploadWithBumpAllocator(
-        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) const {
+void MetalBuffer::uploadWithBumpAllocator(void* src, size_t size, size_t byteOffset) 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);
+    memcpy(static_cast<char*>(buffer.contents) + offset, src, size);
 
     // Encode a blit from the staging buffer into the private GPU buffer.
     id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);

filament/backend/src/metal/MetalContext.h

@@ -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>
@@ -48,13 +46,13 @@ 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
 
@@ -70,64 +68,16 @@ private:
     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 +93,7 @@ struct MetalContext {
     } highestSupportedGpuFamily;
 
     struct {
-        bool staticTextureTargetError;
+        bool a8xStaticTextureTargetError;
     } bugs;
 
     // sampleCountLookup[requestedSamples] gives a <= sample count supported by the device.
@@ -158,9 +108,10 @@ 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.
@@ -173,17 +124,23 @@ struct MetalContext {
 
     std::array<MetalPushConstantBuffer, Program::SHADER_TYPE_COUNT> currentPushConstants;
 
-    // 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;
+    MetalSamplerGroup* samplerBindings[Program::SAMPLER_BINDING_COUNT] = {};
 
-    // Keeps track of all alive textures.
+    // Keeps track of sampler groups we've finalized for the current render pass.
+    tsl::robin_set<MetalSamplerGroup*> finalizedSamplerGroups;
+
+    // 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;
 
@@ -196,7 +153,6 @@ 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;
 

filament/backend/src/metal/MetalContext.mm

@@ -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) {
@@ -130,7 +125,6 @@ void submitPendingCommands(MetalContext* context) {
     assert_invariant(context->pendingCommandBuffer.status != MTLCommandBufferStatusCommitted);
     [context->pendingCommandBuffer commit];
     context->pendingCommandBuffer = nil;
-    context->pendingCommandBufferId++;
 }
 
 id<MTLTexture> getOrCreateEmptyTexture(MetalContext* context) {
@@ -173,6 +167,7 @@ void MetalPushConstantBuffer::setPushConstant(PushConstantVariant value, uint8_t
 
 void MetalPushConstantBuffer::setBytes(id<MTLCommandEncoder> encoder, ShaderStage stage) {
     constexpr size_t PUSH_CONSTANT_SIZE_BYTES = 4;
+    constexpr size_t PUSH_CONSTANT_BUFFER_INDEX = 26;
 
     static char buffer[MAX_PUSH_CONSTANT_COUNT * PUSH_CONSTANT_SIZE_BYTES];
     assert_invariant(mPushConstants.size() <= MAX_PUSH_CONSTANT_COUNT);

filament/backend/src/metal/MetalDriver.h

@@ -32,7 +32,6 @@
 #include <functional>
 #include <mutex>
 #include <vector>
-#include <deque>
 
 namespace filament {
 namespace backend {
@@ -58,11 +57,11 @@ 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 +73,10 @@ 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
@@ -151,6 +137,7 @@ 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);
 

filament/backend/src/metal/MetalExternalImage.h

@@ -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

filament/backend/src/metal/MetalExternalImage.mm

@@ -34,6 +34,10 @@
 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>
@@ -67,30 +71,18 @@ 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);
@@ -104,29 +96,30 @@ MetalExternalImage MetalExternalImage::createFromImage(
             << ".";
 
     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);
@@ -134,83 +127,70 @@ 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.";
 
-    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);
+    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault,
+            mContext.textureCache, image, nullptr, format, width, height, plane, &texture);
     FILAMENT_CHECK_POSTCONDITION(result == kCVReturnSuccess)
             << "Could not create a CVMetalTexture from CVPixelBuffer.";
 
@@ -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);
+    FILAMENT_CHECK_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

filament/backend/src/metal/MetalHandles.h

@@ -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();
@@ -124,7 +121,6 @@ private:
     struct {
         CallbackHandler* handler = nullptr;
         std::shared_ptr<FrameScheduledCallback> callback = nullptr;
-        uint64_t flags = 0;
     } frameScheduled;
 
     struct {
@@ -135,17 +131,19 @@ 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 +200,12 @@ 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 +227,43 @@ 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 +272,26 @@ 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 +302,95 @@ 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 +547,6 @@ 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
 

filament/backend/src/metal/MetalHandles.mm

@@ -74,6 +74,7 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, CAMetalLayer* nativeWindow
       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 +100,17 @@ 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 +121,7 @@ MTLPixelFormat MetalSwapChain::decideDepthStencilFormat(uint64_t flags) {
 }
 
 MetalSwapChain::~MetalSwapChain() {
+    externalImage.set(nullptr);
 }
 
 NSUInteger MetalSwapChain::getSurfaceWidth() const {
@@ -167,7 +171,7 @@ id<MTLTexture> MetalSwapChain::acquireDrawable() {
     }
 
     if (isPixelBuffer()) {
-        return externalImage.getMtlTexture();
+        return externalImage.getMetalTextureForDraw();
     }
 
     assert_invariant(isCaMetalLayer());
@@ -229,10 +233,9 @@ 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(
@@ -254,6 +257,10 @@ void MetalSwapChain::present() {
     }
 }
 
+#ifndef FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD
+#define FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD 1
+#endif
+
 class PresentDrawableData {
 public:
     PresentDrawableData() = delete;
@@ -261,10 +268,10 @@ public:
     PresentDrawableData& operator=(const PresentDrawableData&) = delete;
 
     static PresentDrawableData* create(id<CAMetalDrawable> drawable,
-            std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver, uint64_t flags) {
+            std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver) {
         assert_invariant(drawableMutex);
         assert_invariant(driver);
-        return new PresentDrawableData(drawable, drawableMutex, driver, flags);
+        return new PresentDrawableData(drawable, drawableMutex, driver);
     }
 
     static void maybePresentAndDestroyAsync(PresentDrawableData* that, bool shouldPresent) {
@@ -272,23 +279,20 @@ 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) {}
+            MetalDriver* driver)
+        : mDrawable(drawable), mDrawableMutex(drawableMutex), mDriver(driver) {}
 
     static void cleanupAndDestroy(PresentDrawableData *that) {
         if (that->mDrawable) {
@@ -303,7 +307,6 @@ private:
     id<CAMetalDrawable> mDrawable;
     std::shared_ptr<std::mutex> mDrawableMutex;
     MetalDriver* mDriver = nullptr;
-    uint64_t mFlags = 0;
 };
 
 void presentDrawable(bool presentFrame, void* user) {
@@ -320,8 +323,8 @@ 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)) {}
+                 std::shared_ptr<std::mutex> drawableMutex, MetalDriver* driver)
+            : f(callback), data(PresentDrawableData::create(drawable, drawableMutex, driver)) {}
         std::shared_ptr<FrameScheduledCallback> f;
         // PresentDrawableData* is destroyed by maybePresentAndDestroyAsync() later.
         std::unique_ptr<PresentDrawableData> data;
@@ -336,19 +339,14 @@ 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);
+        frameScheduled.callback, drawable, layerDrawableMutex, 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 +381,12 @@ 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 +482,11 @@ 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);
 }
@@ -505,9 +506,10 @@ 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)
@@ -593,28 +595,16 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             << ", 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 +618,44 @@ 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) {
@@ -770,12 +774,15 @@ 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,
@@ -899,6 +906,98 @@ 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) :
@@ -1244,193 +1343,5 @@ 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

filament/backend/src/metal/MetalState.h

@@ -33,28 +33,32 @@
 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
+//   17-25       Uniform buffers                          9   Program::UNIFORM_BINDING_COUNT
+//   26          Push constants                           1
+//   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
+//   17-25       Uniform buffers                          9   Program::UNIFORM_BINDING_COUNT
+//   26          Push constants                           1
+//   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 +71,10 @@ 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 +382,18 @@ 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 +415,6 @@ 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
 

filament/backend/src/metal/MetalState.mm

@@ -166,40 +166,28 @@ 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 +196,5 @@ 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

filament/backend/src/noop/NoopDriver.cpp

@@ -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 {};
 }
@@ -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

filament/backend/src/opengl/BindingMap.h

@@ -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

filament/backend/src/opengl/GLDescriptorSet.cpp

@@ -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

filament/backend/src/opengl/GLDescriptorSet.h

@@ -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

filament/backend/src/opengl/GLDescriptorSetLayout.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

filament/backend/src/opengl/GLTexture.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;
 };
 

filament/backend/src/opengl/OpenGLContext.cpp

@@ -552,14 +552,6 @@ 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 +676,6 @@ 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 +746,6 @@ 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 +927,15 @@ 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 +944,16 @@ 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;
+                }
             }
         }
     }

filament/backend/src/opengl/OpenGLContext.h

@@ -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;
@@ -509,6 +505,9 @@ 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;
 
@@ -559,9 +558,6 @@ 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 +627,6 @@ 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());

filament/backend/src/opengl/OpenGLDriver.h

@@ -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 {
+            Handle<HwTexture> th;
+            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 {
@@ -315,6 +317,10 @@ 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 +333,8 @@ 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 +346,9 @@ 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 +359,8 @@ 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 +384,6 @@ private:
     bool mRec709OutputColorspace = false;
 
     PushConstantBundle* mCurrentPushConstants = nullptr;
-    PipelineLayout::SetLayout mCurrentSetLayout;
 };
 
 // ------------------------------------------------------------------------------------------------

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -17,7 +17,6 @@
 #include "OpenGLProgram.h"
 
 #include "GLUtils.h"
-#include "GLTexture.h"
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"
 
@@ -25,7 +24,6 @@
 #include <backend/Program.h>
 #include <backend/Handle.h>
 
-#include <utils/BitmaskEnum.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/FixedCapacityVector.h>
@@ -34,10 +32,9 @@
 
 #include <algorithm>
 #include <array>
-#include <algorithm>
-#include <new>
 #include <string_view>
 #include <utility>
+#include <new>
 
 #include <stddef.h>
 #include <stdint.h>
@@ -49,8 +46,9 @@ 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 +57,16 @@ 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 +124,36 @@ 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 +165,51 @@ 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 +226,41 @@ 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
+            Handle<HwTexture> th = sb->textureUnitEntries[j].th;
+            if (th) { // program may not use all samplers of sampler group
+                GLTexture const* const t = gld->handle_cast<GLTexture const*>(th);
+                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);
 

filament/backend/src/opengl/OpenGLProgram.h

@@ -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>
@@ -72,25 +69,32 @@ public:
         }
 
         context.useProgram(gl.program);
+        if (UTILS_UNLIKELY(mUsedBindingsCount)) {
+            // We rely on GL state tracking to avoid unnecessary glBindTexture / glBindSampler
+            // calls.
+
+            // 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
+
+            // 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.
+
+            updateSamplers(gld);
+        }
         return true;
     }
 
-    GLuint getBufferBinding(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
-        return mBindingMap.get(set, binding);
-    }
-
-    GLuint getTextureUnit(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
-        return mBindingMap.get(set, binding);
-    }
-
-    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
-        return mBindingMap.getActiveDescriptors(set);
-    }
-
     // 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 +112,22 @@ 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 +135,15 @@ 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
 

filament/backend/src/opengl/OpenGLTimerQuery.cpp

@@ -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;
     });
 }

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -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;
@@ -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
@@ -597,31 +579,24 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
                 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*, 5> sources = {
+                    version.data(),
+                    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, 5> lengths = {
+                    (GLint)version.length(),
+                    (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

filament/backend/src/opengl/gl_headers.h

@@ -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

filament/backend/src/opengl/platforms/PlatformEGL.cpp

@@ -155,6 +155,10 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const Platform::DriverCon
     ext.egl.KHR_no_config_context = extensions.has("EGL_KHR_no_config_context");
     ext.egl.KHR_surfaceless_context = extensions.has("EGL_KHR_surfaceless_context");
     ext.egl.EXT_protected_content = extensions.has("EGL_EXT_protected_content");
+    if (ext.egl.KHR_create_context) {
+        // KHR_create_context implies KHR_surfaceless_context for ES3.x contexts
+        ext.egl.KHR_surfaceless_context = true;
+    }
 
     eglCreateSyncKHR = (PFNEGLCREATESYNCKHRPROC) eglGetProcAddress("eglCreateSyncKHR");
     eglDestroySyncKHR = (PFNEGLDESTROYSYNCKHRPROC) eglGetProcAddress("eglDestroySyncKHR");
@@ -553,8 +557,6 @@ void PlatformEGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     if (swapChain) {
         SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
         if (sc->sur != EGL_NO_SURFACE) {
-            // - if EGL_KHR_surfaceless_context is supported, mEGLDummySurface is EGL_NO_SURFACE.
-            // - this is actually a bit too aggressive, but it is a rare operation.
             egl.makeCurrent(mEGLDummySurface, mEGLDummySurface);
             eglDestroySurface(mEGLDisplay, sc->sur);
             delete sc;

filament/backend/src/opengl/platforms/PlatformEGLAndroid.cpp

@@ -25,14 +25,12 @@
 #include "ExternalStreamManagerAndroid.h"
 
 #include <android/api-level.h>
-#include <android/native_window.h>
 #include <android/hardware_buffer.h>
 
 #include <utils/android/PerformanceHintManager.h>
 
 #include <utils/compiler.h>
 #include <utils/ostream.h>
-#include <utils/Panic.h>
 #include <utils/Log.h>
 
 #include <EGL/egl.h>
@@ -44,9 +42,7 @@
 
 #include <chrono>
 #include <new>
-#include <string_view>
 
-#include <dlfcn.h>
 #include <unistd.h>
 
 #include <stddef.h>
@@ -84,6 +80,8 @@ UTILS_PRIVATE PFNEGLGETFRAMETIMESTAMPSANDROIDPROC eglGetFrameTimestampsANDROID =
 }
 using namespace glext;
 
+using EGLStream = Platform::Stream;
+
 // ---------------------------------------------------------------------------------------------
 
 PlatformEGLAndroid::InitializeJvmForPerformanceManagerIfNeeded::InitializeJvmForPerformanceManagerIfNeeded() {
@@ -103,73 +101,38 @@ PlatformEGLAndroid::PlatformEGLAndroid() noexcept
           mExternalStreamManager(ExternalStreamManagerAndroid::create()),
           mInitializeJvmForPerformanceManagerIfNeeded(),
           mPerformanceHintManager() {
-    mOSVersion = android_get_device_api_level();
-    if (mOSVersion < 0) {
-        mOSVersion = __ANDROID_API_FUTURE__;
+
+    char scratch[PROP_VALUE_MAX + 1];
+    int length = __system_property_get("ro.build.version.release", scratch);
+    int const androidVersion = length >= 0 ? atoi(scratch) : 1;
+    if (!androidVersion) {
+        mOSVersion = 1000; // if androidVersion is 0, it means "future"
+    } else {
+        length = __system_property_get("ro.build.version.sdk", scratch);
+        mOSVersion = length >= 0 ? atoi(scratch) : 1;
     }
 
-    mNativeWindowLib = dlopen("libnativewindow.so", RTLD_LOCAL | RTLD_NOW);
-    if (mNativeWindowLib) {
-        ANativeWindow_getBuffersDefaultDataSpace =
-                (int32_t(*)(ANativeWindow*))dlsym(mNativeWindowLib,
-                        "ANativeWindow_getBuffersDefaultDataSpace");
-    }
+    // This disables an ANGLE optimization on ARM, which turns out to be more costly for us
+    // see b/229017581
+    // We need to do this before we create the GL context.
+    // An alternative solution is use a system property:
+    //            __system_property_set(
+    //                    "debug.angle.feature_overrides_disabled",
+    //                    "preferSubmitAtFBOBoundary");
+    // but that would outlive this process, so the environment variable is better.
+    // We also make sure to not update the variable if it already exists.
+    // There is no harm setting this if we're not on ANGLE or ARM.
+    setenv("ANGLE_FEATURE_OVERRIDES_DISABLED", "preferSubmitAtFBOBoundary", false);
 }
 
-PlatformEGLAndroid::~PlatformEGLAndroid() noexcept {
-    if (mNativeWindowLib) {
-        dlclose(mNativeWindowLib);
-    }
-}
+PlatformEGLAndroid::~PlatformEGLAndroid() noexcept = default;
+
 
 void PlatformEGLAndroid::terminate() noexcept {
     ExternalStreamManagerAndroid::destroy(&mExternalStreamManager);
     PlatformEGL::terminate();
 }
 
-static constexpr const std::string_view kNativeWindowInvalidMsg =
-        "ANativeWindow is invalid. It probably has been destroyed. EGL surface = ";
-
-bool PlatformEGLAndroid::makeCurrent(ContextType type,
-        SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
-
-    // fast & safe path
-    if (UTILS_LIKELY(!mAssertNativeWindowIsValid)) {
-        return PlatformEGL::makeCurrent(type, drawSwapChain, readSwapChain);
-    }
-
-    SwapChainEGL const* const dsc = static_cast<SwapChainEGL const*>(drawSwapChain);
-    if (ANativeWindow_getBuffersDefaultDataSpace) {
-        // anw can be nullptr if we're using a pbuffer surface
-        if (UTILS_LIKELY(dsc->nativeWindow)) {
-            // this a proxy of is_valid()
-            auto result = ANativeWindow_getBuffersDefaultDataSpace(dsc->nativeWindow);
-            FILAMENT_CHECK_POSTCONDITION(result >= 0) << kNativeWindowInvalidMsg << dsc->sur;
-        }
-    } else {
-        // If we don't have ANativeWindow_getBuffersDefaultDataSpace, we revert to using the
-        // private query() call.
-        // Shadow version if the real ANativeWindow, so we can access the query() hook. Query
-        // has existed since forever, probably Android 1.0.
-        struct NativeWindow {
-            // is valid query enum value
-            enum { IS_VALID = 17 };
-            uint64_t pad[18];
-            int (* query)(ANativeWindow const*, int, int*);
-        } const* pWindow = reinterpret_cast<NativeWindow const*>(dsc->nativeWindow);
-        int isValid = 0;
-        if (UTILS_LIKELY(pWindow->query)) { // just in case it's nullptr
-            int const err = pWindow->query(dsc->nativeWindow, NativeWindow::IS_VALID, &isValid);
-            if (UTILS_LIKELY(err >= 0)) { // in case the IS_VALID enum is not recognized
-                // query call succeeded
-                FILAMENT_CHECK_POSTCONDITION(isValid) << kNativeWindowInvalidMsg << dsc->sur;
-            }
-        }
-    }
-    return PlatformEGL::makeCurrent(type, drawSwapChain, readSwapChain);
-}
-
 void PlatformEGLAndroid::beginFrame(
         int64_t monotonic_clock_ns,
         int64_t refreshIntervalNs,
@@ -226,8 +189,6 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
                 "eglGetFrameTimestampsANDROID");
     }
 
-    mAssertNativeWindowIsValid = driverConfig.assertNativeWindowIsValid;
-
     return driver;
 }
 

filament/backend/src/opengl/platforms/PlatformOSMesa.cpp

@@ -1,149 +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 <backend/platforms/PlatformOSMesa.h>
-
-#include <utils/Log.h>
-#include <utils/Panic.h>
-
-#include <dlfcn.h>
-#include <memory>
-
-namespace filament::backend {
-
-using namespace backend;
-
-namespace {
-
-using BackingType = GLfloat;
-#define BACKING_GL_TYPE GL_FLOAT
-
-struct OSMesaSwapchain {
-    OSMesaSwapchain(uint32_t width, uint32_t height)
-        : width(width),
-          height(height),
-          buffer(new uint8_t[width * height * 4 * sizeof(BackingType)]) {}
-
-    uint32_t width = 0;
-    uint32_t height = 0;
-    std::unique_ptr<uint8_t[]> buffer;
-};
-
-struct OSMesaAPI {
-private:
-    using CreateContextFunc = OSMesaContext (*)(GLenum format, OSMesaContext);
-    using DestroyContextFunc = GLboolean (*)(OSMesaContext);
-    using MakeCurrentFunc = GLboolean (*)(OSMesaContext ctx, void* buffer, GLenum type,
-            GLsizei width, GLsizei height);
-    using GetProcAddressFunc = OSMESAproc (*)(const char* funcName);
-
-public:
-    CreateContextFunc OSMesaCreateContext;
-    DestroyContextFunc OSMesaDestroyContext;
-    MakeCurrentFunc OSMesaMakeCurrent;
-    GetProcAddressFunc OSMesaGetProcAddress;
-
-    OSMesaAPI() {
-        constexpr char const* libraryNames[] = {"libOSMesa.so", "libosmesa.so"};
-        for (char const* libName: libraryNames) {
-            mLib = dlopen(libName, RTLD_GLOBAL | RTLD_NOW);
-            if (mLib) {
-                break;
-            }
-        }
-        FILAMENT_CHECK_PRECONDITION(mLib)
-                << "Unable to dlopen libOSMesa to create a software GL context";
-
-        OSMesaGetProcAddress = (GetProcAddressFunc) dlsym(mLib, "OSMesaGetProcAddress");
-
-        OSMesaCreateContext = (CreateContextFunc) OSMesaGetProcAddress("OSMesaCreateContext");
-        OSMesaDestroyContext =
-                (DestroyContextFunc) OSMesaGetProcAddress("OSMesaDestroyContext");
-        OSMesaMakeCurrent = (MakeCurrentFunc) OSMesaGetProcAddress("OSMesaMakeCurrent");
-    }
-
-    ~OSMesaAPI() {
-        dlclose(mLib);
-    }
-private:
-    void* mLib = nullptr;
-};
-
-}// anonymous namespace
-
-Driver* PlatformOSMesa::createDriver(void* const sharedGLContext,
-        const DriverConfig& driverConfig) noexcept {
-    OSMesaAPI* api = new OSMesaAPI();
-    mOsMesaApi = api;
-
-    FILAMENT_CHECK_PRECONDITION(sharedGLContext == nullptr)
-            << "shared GL context is not supported with PlatformOSMesa";
-    mContext = api->OSMesaCreateContext(GL_RGBA, NULL);
-
-    // We need to do a no-op makecurrent here so that the context will be in a correct state before
-    // any GL calls.
-    auto chain = createSwapChain(1, 1, 0);
-    makeCurrent(ContextType::UNPROTECTED, chain, nullptr);
-    destroySwapChain(chain);
-
-    int result = bluegl::bind();
-    FILAMENT_CHECK_POSTCONDITION(!result) << "Unable to load OpenGL entry points.";
-
-    return OpenGLPlatform::createDefaultDriver(this, sharedGLContext, driverConfig);
-}
-
-void PlatformOSMesa::terminate() noexcept {
-    OSMesaAPI* api = (OSMesaAPI*) mOsMesaApi;
-    api->OSMesaDestroyContext(mContext);
-    delete api;
-    mOsMesaApi = nullptr;
-
-    bluegl::unbind();
-}
-
-Platform::SwapChain* PlatformOSMesa::createSwapChain(void* nativeWindow, uint64_t flags) noexcept {
-    FILAMENT_CHECK_POSTCONDITION(false) << "Cannot create non-headless swapchain";
-    return (SwapChain*) nativeWindow;
-}
-
-Platform::SwapChain* PlatformOSMesa::createSwapChain(uint32_t width, uint32_t height,
-        uint64_t flags) noexcept {
-    OSMesaSwapchain* swapchain = new OSMesaSwapchain(width, height);
-    return (SwapChain*) swapchain;
-}
-
-void PlatformOSMesa::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
-    OSMesaSwapchain* impl = (OSMesaSwapchain*) swapChain;
-    delete impl;
-}
-
-bool PlatformOSMesa::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
-    OSMesaAPI* api = (OSMesaAPI*) mOsMesaApi;
-    OSMesaSwapchain* impl = (OSMesaSwapchain*) drawSwapChain;
-
-    auto result = api->OSMesaMakeCurrent(mContext, (BackingType*) impl->buffer.get(),
-            BACKING_GL_TYPE, impl->width, impl->height);
-    FILAMENT_CHECK_POSTCONDITION(result == GL_TRUE) << "OSMesaMakeCurrent failed!";
-
-    return true;
-}
-
-void PlatformOSMesa::commit(Platform::SwapChain* swapChain) noexcept {
-    // No-op since we are not scanning out to a display.
-}
-
-} // namespace filament::backend

filament/backend/src/ostream.cpp

@@ -410,6 +410,7 @@ io::ostream& operator<<(io::ostream& out, const RasterState& rs) {
 io::ostream& operator<<(io::ostream& out, const TargetBufferInfo& tbi) {
     return out << "TargetBufferInfo{"
     << "handle=" << tbi.handle
+    << ", baseViewIndex=" << tbi.baseViewIndex
     << ", level=" << tbi.level
     << ", layer=" << tbi.layer << "}";
 }

filament/backend/src/vulkan/VulkanBlitter.cpp

@@ -15,7 +15,6 @@
  */
 
 #include "VulkanBlitter.h"
-#include "VulkanCommands.h"
 #include "VulkanContext.h"
 #include "VulkanFboCache.h"
 #include "VulkanHandles.h"
@@ -34,10 +33,9 @@ namespace filament::backend {
 
 namespace {
 
-inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, VkFilter filter,
+inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect, VkFilter filter,
         VulkanAttachment src, VulkanAttachment dst,
         const VkOffset3D srcRect[2], const VkOffset3D dstRect[2]) {
-    VkCommandBuffer const cmdbuf = commands->buffer();
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
         FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
@@ -51,8 +49,8 @@ inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, V
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
 
-    src.texture->transitionLayout(commands, srcRange, VulkanLayout::TRANSFER_SRC);
-    dst.texture->transitionLayout(commands, dstRange, VulkanLayout::TRANSFER_DST);
+    src.texture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
+    dst.texture->transitionLayout(cmdbuffer, dstRange, VulkanLayout::TRANSFER_DST);
 
     const VkImageBlit blitRegions[1] = {{
             .srcSubresource = { aspect, src.level, src.layer, 1 },
@@ -60,24 +58,23 @@ inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, V
             .dstSubresource = { aspect, dst.level, dst.layer, 1 },
             .dstOffsets = { dstRect[0], dstRect[1] },
     }};
-    vkCmdBlitImage(cmdbuf,
+    vkCmdBlitImage(cmdbuffer,
             src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
             dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
             1, blitRegions, filter);
 
     if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = src.texture->getDefaultLayout();
+        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
     }
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = dst.texture->getDefaultLayout();
+        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
     }
-    src.texture->transitionLayout(commands, srcRange, oldSrcLayout);
-    dst.texture->transitionLayout(commands, dstRange, oldDstLayout);
+    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
+    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
 }
 
-inline void resolveFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect,
+inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
         VulkanAttachment src, VulkanAttachment dst) {
-    VkCommandBuffer const cmdbuffer = commands->buffer();
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
         FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
@@ -91,8 +88,8 @@ inline void resolveFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
 
-    src.texture->transitionLayout(commands, srcRange, VulkanLayout::TRANSFER_SRC);
-    dst.texture->transitionLayout(commands, dstRange, VulkanLayout::TRANSFER_DST);
+    src.texture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
+    dst.texture->transitionLayout(cmdbuffer, dstRange, VulkanLayout::TRANSFER_DST);
 
     assert_invariant(
             aspect != VK_IMAGE_ASPECT_DEPTH_BIT && "Resolve with depth is not yet supported.");
@@ -109,13 +106,13 @@ inline void resolveFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect
             1, resolveRegions);
 
     if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = src.texture->getDefaultLayout();
+        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
     }
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = dst.texture->getDefaultLayout();
+        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
     }
-    src.texture->transitionLayout(commands, srcRange, oldSrcLayout);
-    dst.texture->transitionLayout(commands, dstRange, oldDstLayout);
+    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
+    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
 }
 
 struct BlitterUniforms {
@@ -152,9 +149,10 @@ void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {
 #endif
 
     VulkanCommandBuffer& commands = mCommands->get();
+    VkCommandBuffer const cmdbuffer = commands.buffer();
     commands.acquire(src.texture);
     commands.acquire(dst.texture);
-    resolveFast(&commands, aspect, src, dst);
+    resolveFast(cmdbuffer, aspect, src, dst);
 }
 
 void VulkanBlitter::blit(VkFilter filter,
@@ -177,9 +175,10 @@ void VulkanBlitter::blit(VkFilter filter,
     // src and dst should have the same aspect here
     VkImageAspectFlags const aspect = src.texture->getImageAspect();
     VulkanCommandBuffer& commands = mCommands->get();
+    VkCommandBuffer const cmdbuffer = commands.buffer();
     commands.acquire(src.texture);
     commands.acquire(dst.texture);
-    blitFast(&commands, aspect, filter, src, dst, srcRectPair, dstRectPair);
+    blitFast(cmdbuffer, aspect, filter, src, dst, srcRectPair, dstRectPair);
 }
 
 void VulkanBlitter::terminate() noexcept {

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -297,11 +297,11 @@ bool VulkanCommands::flush() {
 #endif
 
     auto& cmdfence = currentbuf->fence;
-    UTILS_UNUSED_IN_RELEASE VkResult result = VK_SUCCESS;
-    {
-        auto scope = cmdfence->setValue(VK_NOT_READY);
-        result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->getFence());
-    }
+    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
+    cmdfence->status.store(VK_NOT_READY);
+    UTILS_UNUSED_IN_RELEASE VkResult result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->fence);
+    cmdfence->condition.notify_all();
+    lock.unlock();
 
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
     if (result != VK_SUCCESS) {
@@ -340,7 +340,7 @@ void VulkanCommands::wait() {
         auto wrapper = mStorage[i].get();
         if (wrapper->buffer() != VK_NULL_HANDLE
                 && mCurrentCommandBufferIndex != static_cast<int8_t>(i)) {
-            fences[count++] = wrapper->fence->getFence();
+            fences[count++] = wrapper->fence->fence;
         }
     }
     if (count > 0) {
@@ -361,13 +361,12 @@ void VulkanCommands::gc() {
         if (wrapper->buffer() == VK_NULL_HANDLE) {
             continue;
         }
-        auto const vkfence = wrapper->fence->getFence();
-        VkResult const result = vkGetFenceStatus(mDevice, vkfence);
+        VkResult const result = vkGetFenceStatus(mDevice, wrapper->fence->fence);
         if (result != VK_SUCCESS) {
             continue;
         }
-        fences[count++] = vkfence;
-        wrapper->fence->setValue(VK_SUCCESS);
+        fences[count++] = wrapper->fence->fence;
+        wrapper->fence->status.store(VK_SUCCESS);
         wrapper->reset();
         mAvailableBufferCount++;
     }
@@ -384,9 +383,9 @@ void VulkanCommands::updateFences() {
         if (wrapper->buffer() != VK_NULL_HANDLE) {
             VulkanCmdFence* fence = wrapper->fence.get();
             if (fence) {
-                VkResult status = vkGetFenceStatus(mDevice, fence->getFence());
+                VkResult status = vkGetFenceStatus(mDevice, fence->fence);
                 // This is either VK_SUCCESS, VK_NOT_READY, or VK_ERROR_DEVICE_LOST.
-                fence->setValue(status);
+                fence->status.store(status, std::memory_order_relaxed);
             }
         }
     }

filament/backend/src/vulkan/VulkanCommands.h

@@ -61,40 +61,8 @@ private:
 
 // Wrapper to enable use of shared_ptr for implementing shared ownership of low-level Vulkan fences.
 struct VulkanCmdFence {
-    struct SetValueScope {
-    public:
-        ~SetValueScope() {
-            mHolder->mutex.unlock();
-            mHolder->condition.notify_all();
-        }
-
-    private:
-        SetValueScope(VulkanCmdFence* fenceHolder, VkResult result) :
-            mHolder(fenceHolder) {
-            mHolder->mutex.lock();
-            mHolder->status.store(result);
-        }
-        VulkanCmdFence* mHolder;
-        friend struct VulkanCmdFence;
-    };
-
     VulkanCmdFence(VkFence ifence);
     ~VulkanCmdFence() = default;
-
-    SetValueScope setValue(VkResult value) {
-        return {this, value};
-    }
-
-    VkFence& getFence() {
-        return fence;
-    }
-
-    VkResult getStatus() {
-        std::unique_lock<utils::Mutex> lock(mutex);
-        return status.load(std::memory_order_acquire);
-    }
-
-private:
     VkFence fence;
     utils::Condition condition;
     utils::Mutex mutex;

filament/backend/src/vulkan/VulkanContext.cpp

@@ -59,11 +59,7 @@ VkExtent2D VulkanAttachment::getExtent2D() const {
 
 VkImageView VulkanAttachment::getImageView() {
     assert_invariant(texture);
-    VkImageSubresourceRange range = getSubresourceRange();
-    if (range.layerCount > 1) {
-        return texture->getMultiviewAttachmentView(range);
-    }
-    return texture->getAttachmentView(range);
+    return texture->getAttachmentView(getSubresourceRange());
 }
 
 bool VulkanAttachment::isDepth() const {
@@ -77,7 +73,7 @@ VkImageSubresourceRange VulkanAttachment::getSubresourceRange() const {
             .baseMipLevel = uint32_t(level),
             .levelCount = 1,
             .baseArrayLayer = uint32_t(layer),
-            .layerCount = layerCount,
+            .layerCount = 1,
     };
 }
 

filament/backend/src/vulkan/VulkanContext.h

@@ -43,7 +43,6 @@ struct VulkanCommandBuffer;
 struct VulkanAttachment {
     VulkanTexture* texture = nullptr;
     uint8_t level = 0;
-    uint8_t layerCount = 1;
     uint16_t layer = 0;
 
     bool isDepth() const;
@@ -124,10 +123,6 @@ public:
         return mPhysicalDeviceFeatures.imageCubeArray == VK_TRUE;
     }
 
-    inline bool isDepthClampSupported() const noexcept {
-        return mPhysicalDeviceFeatures.depthClamp == VK_TRUE;
-    }
-
     inline bool isDebugMarkersSupported() const noexcept {
         return mDebugMarkersSupported;
     }
@@ -144,10 +139,6 @@ public:
         return mPhysicalDeviceFeatures.shaderClipDistance == VK_TRUE;
     }
 
-    inline bool isLazilyAllocatedMemorySupported() const noexcept {
-        return mLazilyAllocatedMemorySupported;
-    }
-
 private:
     VkPhysicalDeviceMemoryProperties mMemoryProperties = {};
     VkPhysicalDeviceProperties mPhysicalDeviceProperties = {};
@@ -155,7 +146,6 @@ private:
     bool mDebugMarkersSupported = false;
     bool mDebugUtilsSupported = false;
     bool mMultiviewEnabled = false;
-    bool mLazilyAllocatedMemorySupported = false;
 
     VkFormatList mDepthStencilFormats;
     VkFormatList mBlittableDepthStencilFormats;

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -88,6 +88,28 @@ VmaAllocator createAllocator(VkInstance instance, VkPhysicalDevice physicalDevic
     return allocator;
 }
 
+VulkanTexture* createEmptyTexture(VkDevice device, VkPhysicalDevice physicalDevice,
+        VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
+        VulkanStagePool& stagePool) {
+    VulkanTexture* emptyTexture = new VulkanTexture(device, physicalDevice, context, allocator,
+            commands, SamplerType::SAMPLER_2D, 1, TextureFormat::RGBA8, 1, 1, 1, 1,
+            TextureUsage::DEFAULT | TextureUsage::COLOR_ATTACHMENT | TextureUsage::SUBPASS_INPUT,
+            stagePool, true /* heap allocated */);
+    uint32_t black = 0;
+    PixelBufferDescriptor pbd(&black, 4, PixelDataFormat::RGBA, PixelDataType::UBYTE);
+    emptyTexture->updateImage(pbd, 1, 1, 1, 0, 0, 0, 0);
+    return emptyTexture;
+}
+
+VulkanBufferObject* createEmptyBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool,
+        VulkanCommands* commands) {
+    VulkanBufferObject* obj =
+            new VulkanBufferObject(allocator, stagePool, 1, BufferObjectBinding::UNIFORM);
+    uint8_t byte = 0;
+    obj->buffer.loadFromCpu(commands->get().buffer(), &byte, 0, 1);
+    return obj;
+}
+
 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
 VKAPI_ATTR VkBool32 VKAPI_CALL debugReportCallback(VkDebugReportFlagsEXT flags,
         VkDebugReportObjectTypeEXT objectType, uint64_t object, size_t location,
@@ -127,7 +149,6 @@ VKAPI_ATTR VkBool32 VKAPI_CALL debugUtilsCallback(VkDebugUtilsMessageSeverityFla
 }
 #endif // FVK_EANBLED(FVK_DEBUG_DEBUG_UTILS)
 
-
 }// anonymous namespace
 
 #if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
@@ -233,6 +254,17 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
 #endif
 
     mTimestamps = std::make_unique<VulkanTimestamps>(mPlatform->getDevice());
+
+    mEmptyTexture = createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
+            mContext, mAllocator, &mCommands, mStagePool);
+    mEmptyBufferObject = createEmptyBufferObject(mAllocator, mStagePool, &mCommands);
+
+    mDescriptorSetManager.setPlaceHolders(mSamplerCache.getSampler({}), mEmptyTexture,
+            mEmptyBufferObject);
+
+    mGetPipelineFunction = [this](VulkanDescriptorSetLayoutList const& layouts, VulkanProgram* program) {
+        return mPipelineLayoutCache.getLayout(layouts, program);
+    };
 }
 
 VulkanDriver::~VulkanDriver() noexcept = default;
@@ -297,6 +329,9 @@ void VulkanDriver::terminate() {
     // to those commands are no longer referenced.
     finish(0);
 
+    delete mEmptyBufferObject;
+    delete mEmptyTexture;
+
     // Command buffers should come first since it might have commands depending on resources that
     // are about to be destroyed.
     mCommands.terminate();
@@ -352,6 +387,7 @@ void VulkanDriver::collectGarbage() {
     mStagePool.gc();
     mFramebufferCache.gc();
     mPipelineCache.gc();
+    mDescriptorSetManager.gc();
 
 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
     mResourceAllocator.print();
@@ -367,8 +403,8 @@ void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch, CallbackHandler* handler,
-        FrameScheduledCallback&& callback, uint64_t flags) {
+void VulkanDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
 }
 
 void VulkanDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
@@ -386,28 +422,6 @@ void VulkanDriver::endFrame(uint32_t frameId) {
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::updateDescriptorSetBuffer(
-        backend::DescriptorSetHandle dsh,
-        backend::descriptor_binding_t binding,
-        backend::BufferObjectHandle boh,
-        uint32_t offset,
-        uint32_t size) {
-    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
-    VulkanBufferObject* obj = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    mDescriptorSetManager.updateBuffer(set, binding, obj, offset, size);
-}
-
-void VulkanDriver::updateDescriptorSetTexture(
-        backend::DescriptorSetHandle dsh,
-        backend::descriptor_binding_t binding,
-        backend::TextureHandle th,
-        SamplerParams params) {
-    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
-    VulkanTexture* texture = mResourceAllocator.handle_cast<VulkanTexture*>(th);
-    VkSampler const vksampler = mSamplerCache.getSampler(params);
-    mDescriptorSetManager.updateSampler(set, binding, texture, vksampler);
-}
-
 void VulkanDriver::flush(int) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("flush");
@@ -426,6 +440,12 @@ void VulkanDriver::finish(int dummy) {
     FVK_SYSTRACE_END();
 }
 
+void VulkanDriver::createSamplerGroupR(Handle<HwSamplerGroup> sbh, uint32_t count,
+        utils::FixedSizeString<32> debugName) {
+    auto sg = mResourceAllocator.construct<VulkanSamplerGroup>(sbh, count);
+    mResourceManager.acquire(sg);
+}
+
 void VulkanDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
         Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh,
         PrimitiveType pt) {
@@ -508,56 +528,23 @@ void VulkanDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
         TextureUsage usage) {
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
-            target, levels,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
             format, samples, w, h, depth, usage, mStagePool);
     mResourceManager.acquire(vktexture);
 }
 
-//void VulkanDriver::createTextureSwizzledR(Handle<HwTexture> th, SamplerType target, uint8_t levels,
-//        TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
-//        TextureUsage usage,
-//        TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) {
-//    TextureSwizzle swizzleArray[] = {r, g, b, a};
-//    const VkComponentMapping swizzleMap = getSwizzleMap(swizzleArray);
-//    auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-//            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
-//            target, levels, format, samples, w, h, depth, usage, mStagePool,
-//            false /*heap allocated */, swizzleMap);
-//    mResourceManager.acquire(vktexture);
-//}
-
-void VulkanDriver::createTextureViewR(Handle<HwTexture> th, Handle<HwTexture> srch,
-        uint8_t baseLevel, uint8_t levelCount) {
-    VulkanTexture const* src = mResourceAllocator.handle_cast<VulkanTexture const*>(srch);
+void VulkanDriver::createTextureSwizzledR(Handle<HwTexture> th, SamplerType target, uint8_t levels,
+        TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
+        TextureUsage usage,
+        TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) {
+    TextureSwizzle swizzleArray[] = {r, g, b, a};
+    const VkComponentMapping swizzleMap = getSwizzleMap(swizzleArray);
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
-            src, baseLevel, levelCount);
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
+            format, samples, w, h, depth, usage, mStagePool, false /*heap allocated */, swizzleMap);
     mResourceManager.acquire(vktexture);
 }
 
-void VulkanDriver::createTextureViewSwizzleR(Handle<HwTexture> th, Handle<HwTexture> srch,
-        backend::TextureSwizzle r, backend::TextureSwizzle g, backend::TextureSwizzle b,
-        backend::TextureSwizzle a) {
-    TextureSwizzle const swizzleArray[] = {r, g, b, a};
-    VkComponentMapping const swizzle = getSwizzleMap(swizzleArray);
-
-    VulkanTexture const* src = mResourceAllocator.handle_cast<VulkanTexture const*>(srch);
-    auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
-            src, swizzle);
-    mResourceManager.acquire(vktexture);
-}
-
-void VulkanDriver::createTextureExternalImageR(Handle<HwTexture> th, backend::TextureFormat format,
-        uint32_t width, uint32_t height, backend::TextureUsage usage, void* image) {
-}
-
-void VulkanDriver::createTextureExternalImagePlaneR(Handle<HwTexture> th,
-        backend::TextureFormat format, uint32_t width, uint32_t height, backend::TextureUsage usage,
-        void* image, uint32_t plane) {
-}
-
 void VulkanDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
         SamplerType target, uint8_t levels,
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
@@ -584,6 +571,7 @@ void VulkanDriver::destroyProgram(Handle<HwProgram> ph) {
         return;
     }
     auto vkprogram = mResourceAllocator.handle_cast<VulkanProgram*>(ph);
+    mDescriptorSetManager.clearProgram(vkprogram);
     mResourceManager.release(vkprogram);
 }
 
@@ -607,7 +595,6 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
             colorTargets[i] = {
                 .texture = mResourceAllocator.handle_cast<VulkanTexture*>(color[i].handle),
                 .level = color[i].level,
-                .layerCount = layerCount,
                 .layer = color[i].layer,
             };
             UTILS_UNUSED_IN_RELEASE VkExtent2D extent = colorTargets[i].getExtent2D();
@@ -622,7 +609,6 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         depthStencil[0] = {
             .texture = mResourceAllocator.handle_cast<VulkanTexture*>(depth.handle),
             .level = depth.level,
-            .layerCount = layerCount,
             .layer = depth.layer,
         };
         UTILS_UNUSED_IN_RELEASE VkExtent2D extent = depthStencil[0].getExtent2D();
@@ -635,7 +621,6 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         depthStencil[1] = {
             .texture = mResourceAllocator.handle_cast<VulkanTexture*>(stencil.handle),
             .level = stencil.level,
-            .layerCount = layerCount,
             .layer = stencil.layer,
         };
         UTILS_UNUSED_IN_RELEASE VkExtent2D extent = depthStencil[1].getExtent2D();
@@ -650,10 +635,9 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
     assert_invariant(tmin == tmax);
     assert_invariant(tmin.x >= width && tmin.y >= height);
 
-    auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth,
-            mPlatform->getDevice(), mPlatform->getPhysicalDevice(), mContext, mAllocator,
-            &mCommands, &mResourceAllocator, width, height, samples, colorTargets, depthStencil,
-            mStagePool, layerCount);
+    auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth, mPlatform->getDevice(),
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, width, height,
+            samples, colorTargets, depthStencil, mStagePool);
     mResourceManager.acquire(renderTarget);
 }
 
@@ -681,7 +665,7 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
         flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
     }
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, &mCommands, &mResourceAllocator, mStagePool, nativeWindow, flags);
+            mAllocator, &mCommands, mStagePool, nativeWindow, flags);
     mResourceManager.acquire(swapChain);
 }
 
@@ -694,8 +678,7 @@ void VulkanDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t wi
     }
     assert_invariant(width > 0 && height > 0 && "Vulkan requires non-zero swap chain dimensions.");
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, &mCommands, &mResourceAllocator, mStagePool,
-            nullptr, flags, VkExtent2D{width, height});
+            mAllocator, &mCommands, mStagePool, nullptr, flags, VkExtent2D{width, height});
     mResourceManager.acquire(swapChain);
 }
 
@@ -703,25 +686,6 @@ void VulkanDriver::createTimerQueryR(Handle<HwTimerQuery> tqh, int) {
     // nothing to do, timer query was constructed in createTimerQueryS
 }
 
-void VulkanDriver::createDescriptorSetLayoutR(Handle<HwDescriptorSetLayout> dslh,
-        backend::DescriptorSetLayout&& info) {
-    VulkanDescriptorSetLayout* layout = mResourceAllocator.construct<VulkanDescriptorSetLayout>(
-            dslh, info);
-
-    // This will create a VkDescriptorSetLayout (which is cached) for this object.
-    mDescriptorSetManager.initVkLayout(layout);
-    mResourceManager.acquire(layout);
-}
-
-void VulkanDriver::createDescriptorSetR(Handle<HwDescriptorSet> dsh,
-        Handle<HwDescriptorSetLayout> dslh) {
-    auto layout = mResourceAllocator.handle_cast<VulkanDescriptorSetLayout*>(dslh);
-    mDescriptorSetManager.createSet(dsh, layout);
-
-    auto set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
-    mResourceManager.acquire(set);
-}
-
 Handle<HwVertexBufferInfo> VulkanDriver::createVertexBufferInfoS() noexcept {
     return mResourceAllocator.allocHandle<VulkanVertexBufferInfo>();
 }
@@ -742,19 +706,7 @@ Handle<HwTexture> VulkanDriver::createTextureS() noexcept {
     return mResourceAllocator.allocHandle<VulkanTexture>();
 }
 
-Handle<HwTexture> VulkanDriver::createTextureViewS() noexcept {
-    return mResourceAllocator.allocHandle<VulkanTexture>();
-}
-
-Handle<HwTexture> VulkanDriver::createTextureViewSwizzleS() noexcept {
-    return mResourceAllocator.allocHandle<VulkanTexture>();
-}
-
-Handle<HwTexture> VulkanDriver::createTextureExternalImageS() noexcept {
-    return mResourceAllocator.allocHandle<VulkanTexture>();
-}
-
-Handle<HwTexture> VulkanDriver::createTextureExternalImagePlaneS() noexcept {
+Handle<HwTexture> VulkanDriver::createTextureSwizzledS() noexcept {
     return mResourceAllocator.allocHandle<VulkanTexture>();
 }
 
@@ -762,6 +714,10 @@ Handle<HwTexture> VulkanDriver::importTextureS() noexcept {
     return mResourceAllocator.allocHandle<VulkanTexture>();
 }
 
+Handle<HwSamplerGroup> VulkanDriver::createSamplerGroupS() noexcept {
+    return mResourceAllocator.allocHandle<VulkanSamplerGroup>();
+}
+
 Handle<HwRenderPrimitive> VulkanDriver::createRenderPrimitiveS() noexcept {
     return mResourceAllocator.allocHandle<VulkanRenderPrimitive>();
 }
@@ -800,12 +756,21 @@ Handle<HwTimerQuery> VulkanDriver::createTimerQueryS() noexcept {
     return tqh;
 }
 
-Handle<HwDescriptorSetLayout> VulkanDriver::createDescriptorSetLayoutS() noexcept {
-    return mResourceAllocator.allocHandle<VulkanDescriptorSetLayout>();
-}
-
-Handle<HwDescriptorSet> VulkanDriver::createDescriptorSetS() noexcept {
-    return mResourceAllocator.allocHandle<VulkanDescriptorSet>();
+void VulkanDriver::destroySamplerGroup(Handle<HwSamplerGroup> sbh) {
+    if (!sbh) {
+        return;
+    }
+    // Unlike most of the other "Hw" handles, the sampler buffer is an abstract concept and does
+    // not map to any Vulkan objects. To handle destruction, the only thing we need to do is
+    // ensure that the next draw call doesn't try to access a zombie sampler buffer. Therefore,
+    // simply replace all weak references with null.
+    auto* hwsb = mResourceAllocator.handle_cast<VulkanSamplerGroup*>(sbh);
+    for (auto& binding : mSamplerBindings) {
+        if (binding == hwsb) {
+            binding = nullptr;
+        }
+    }
+    mResourceManager.release(hwsb);
 }
 
 void VulkanDriver::destroySwapChain(Handle<HwSwapChain> sch) {
@@ -830,17 +795,6 @@ void VulkanDriver::destroyTimerQuery(Handle<HwTimerQuery> tqh) {
     mThreadSafeResourceManager.release(vtq);
 }
 
-void VulkanDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> dslh) {
-    VulkanDescriptorSetLayout* layout = mResourceAllocator.handle_cast<VulkanDescriptorSetLayout*>(dslh);
-    mResourceManager.release(layout);
-}
-
-void VulkanDriver::destroyDescriptorSet(Handle<HwDescriptorSet> dsh) {
-    mDescriptorSetManager.destroySet(dsh);
-    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
-    mResourceManager.release(set);
-}
-
 Handle<HwStream> VulkanDriver::createStreamNative(void* nativeStream) {
     return {};
 }
@@ -878,7 +832,8 @@ FenceStatus VulkanDriver::getFenceStatus(Handle<HwFence> fh) {
 
     // Internally we use the VK_INCOMPLETE status to mean "not yet submitted".
     // When this fence gets submitted, its status changes to VK_NOT_READY.
-    if (cmdfence->getStatus() == VK_SUCCESS) {
+    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
+    if (cmdfence->status.load() == VK_SUCCESS) {
         return FenceStatus::CONDITION_SATISFIED;
     }
 
@@ -929,9 +884,7 @@ bool VulkanDriver::isRenderTargetFormatSupported(TextureFormat format) {
 }
 
 bool VulkanDriver::isFrameBufferFetchSupported() {
-    // TODO: we must fix this before landing descriptor set change.  Otherwise, the scuba tests will fail.
-    //return true;
-    return false;
+    return true;
 }
 
 bool VulkanDriver::isFrameBufferFetchMultiSampleSupported() {
@@ -985,10 +938,6 @@ bool VulkanDriver::isProtectedTexturesSupported() {
     return false;
 }
 
-bool VulkanDriver::isDepthClampSupported() {
-    return mContext.isDepthClampSupported();
-}
-
 bool VulkanDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU: {
@@ -1119,6 +1068,10 @@ void VulkanDriver::resetBufferObject(Handle<HwBufferObject> boh) {
     // This is only useful if updateBufferObjectUnsynchronized() is implemented unsynchronizedly.
 }
 
+void VulkanDriver::setMinMaxLevels(Handle<HwTexture> th, uint32_t minLevel, uint32_t maxLevel) {
+    mResourceAllocator.handle_cast<VulkanTexture*>(th)->setPrimaryRange(minLevel, maxLevel);
+}
+
 void VulkanDriver::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,
         PixelBufferDescriptor&& data) {
@@ -1204,6 +1157,23 @@ void VulkanDriver::generateMipmaps(Handle<HwTexture> th) {
         srcw = dstw;
         srch = dsth;
     } while ((srcw > 1 || srch > 1) && level < t->levels);
+    t->setPrimaryRange(0, t->levels - 1);
+}
+
+void VulkanDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
+        BufferDescriptor&& data) {
+    auto* sb = mResourceAllocator.handle_cast<VulkanSamplerGroup*>(sbh);
+
+    // FIXME: we shouldn't be using SamplerGroup here, instead the backend should create
+    //        a descriptor or any internal data-structure that represents the textures/samplers.
+    //        It's preferable to do as much work as possible here.
+    //        Here, we emulate the older backend API by re-creating a SamplerGroup from the
+    //        passed data.
+    SamplerGroup samplerGroup(data.size / sizeof(SamplerDescriptor));
+    memcpy(samplerGroup.data(), data.buffer, data.size);
+    *sb->sb = std::move(samplerGroup);
+
+    scheduleDestroy(std::move(data));
 }
 
 void VulkanDriver::compilePrograms(CompilerPriorityQueue priority,
@@ -1218,7 +1188,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     FVK_SYSTRACE_START("beginRenderPass");
 
     VulkanRenderTarget* const rt = mResourceAllocator.handle_cast<VulkanRenderTarget*>(rth);
-    VkExtent2D const extent = rt->getExtent();
+    const VkExtent2D extent = rt->getExtent();
     assert_invariant(rt == mDefaultRenderTarget || extent.width > 0 && extent.height > 0);
 
     // Filament has the expectation that the contents of the swap chain are not preserved on the
@@ -1249,10 +1219,34 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer const cmdbuffer = commands.buffer();
 
-    // Scissor is reset with each render pass
-    // This also takes care of VUID-vkCmdDrawIndexed-None-07832.
-    VkRect2D const scissor{ .offset = { 0, 0 }, .extent = extent };
-    vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
+    UTILS_NOUNROLL
+    for (uint8_t samplerGroupIdx = 0; samplerGroupIdx < Program::SAMPLER_BINDING_COUNT;
+            samplerGroupIdx++) {
+        VulkanSamplerGroup* vksb = mSamplerBindings[samplerGroupIdx];
+        if (!vksb) {
+            continue;
+        }
+        SamplerGroup* sb = vksb->sb.get();
+        for (size_t i = 0; i < sb->getSize(); i++) {
+            SamplerDescriptor const* boundSampler = sb->data() + i;
+            if (UTILS_LIKELY(boundSampler->t)) {
+                VulkanTexture* texture
+                        = mResourceAllocator.handle_cast<VulkanTexture*>(boundSampler->t);
+                if (!any(texture->usage & TextureUsage::DEPTH_ATTACHMENT)) {
+                    continue;
+                }
+                if (texture->getPrimaryImageLayout() == VulkanLayout::DEPTH_SAMPLER) {
+                    continue;
+                }
+                commands.acquire(texture);
+
+                // Transition the primary view, which is the sampler's view into the right layout.
+                texture->transitionLayout(cmdbuffer, texture->getPrimaryViewRange(),
+                        VulkanLayout::DEPTH_SAMPLER);
+                break;
+            }
+        }
+    }
 
     VulkanLayout currentDepthLayout = depth.getLayout();
 
@@ -1266,39 +1260,36 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         // If the depth attachment texture was previously sampled, then we need to manually
         // transition it to an attachment. This is necessary to also set up a barrier between the
         // previous read and the potentially coming write.
-        depth.texture->transitionLayout(&commands, depth.getSubresourceRange(),
-                VulkanLayout::DEPTH_ATTACHMENT);
-        currentDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
+        if (currentDepthLayout == VulkanLayout::DEPTH_SAMPLER) {
+            depth.texture->transitionLayout(cmdbuffer, depth.getSubresourceRange(),
+                    VulkanLayout::DEPTH_ATTACHMENT);
+            currentDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
+        }
     }
 
-    uint8_t const renderTargetLayerCount = rt->getLayerCount();
-
     // Create the VkRenderPass or fetch it from cache.
     VulkanFboCache::RenderPassKey rpkey = {
+        .initialColorLayoutMask = 0,
+        .initialDepthLayout = currentDepthLayout,
         .depthFormat = depth.getFormat(),
         .clear = clearVal,
         .discardStart = discardStart,
         .discardEnd = discardEndVal,
-        .initialDepthLayout = currentDepthLayout,
         .samples = rt->getSamples(),
-        .needsResolveMask = 0,
-        .usesLazilyAllocatedMemory = 0,
         .subpassMask = uint8_t(params.subpassMask),
-        .viewCount = renderTargetLayerCount,
     };
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
-        VulkanAttachment const& info = rt->getColor(i);
+        const VulkanAttachment& info = rt->getColor(i);
         if (info.texture) {
-            assert_invariant(info.layerCount == renderTargetLayerCount);
+            rpkey.initialColorLayoutMask |= 1 << i;
             rpkey.colorFormat[i] = info.getFormat();
-            if (rpkey.samples > 1) {
-                const VulkanTexture* sidecar = info.texture->getSidecar();
-                if (sidecar && sidecar->isTransientAttachment()) {
-                    rpkey.usesLazilyAllocatedMemory |= (1 << i);
-                }
-                if (info.texture->samples == 1) {
-                    rpkey.needsResolveMask |= (1 << i);
-                }
+            if (rpkey.samples > 1 && info.texture->samples == 1) {
+                rpkey.needsResolveMask |= (1 << i);
+            }
+            if (info.texture->getPrimaryImageLayout() != VulkanLayout::COLOR_ATTACHMENT) {
+                ((VulkanTexture*) info.texture)
+                        ->transitionLayout(cmdbuffer, info.getSubresourceRange(),
+                                VulkanLayout::COLOR_ATTACHMENT);
             }
         } else {
             rpkey.colorFormat[i] = VK_FORMAT_UNDEFINED;
@@ -1318,43 +1309,28 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     };
     auto& renderPassAttachments = mRenderPassFboInfo.attachments;
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
-        VulkanAttachment& attachment = rt->getColor(i);
-        if (!attachment.texture) {
+        if (!rt->getColor(i).texture) {
             fbkey.color[i] = VK_NULL_HANDLE;
             fbkey.resolve[i] = VK_NULL_HANDLE;
-            continue;
-        }
-
-        if (fbkey.samples == 1) {
-            auto const& range = attachment.getSubresourceRange();
-            auto tex = attachment.texture;
-            if (tex->getLayout(range.baseMipLevel, range.baseArrayLayer) !=
-                            VulkanLayout::COLOR_ATTACHMENT &&
-                    !tex->transitionLayout(&commands, range, VulkanLayout::COLOR_ATTACHMENT)) {
-                // If the layout transition did not emit a barrier, we do it manually here.
-                tex->samplerToAttachmentBarrier(&commands, range);
-            }
-            renderPassAttachments.insert(attachment);
-
-            fbkey.color[i] = attachment.getImageView();
+        } else if (fbkey.samples == 1) {
+            auto& colorAttachment = rt->getColor(i);
+            renderPassAttachments.insert(colorAttachment);
+            fbkey.color[i] = colorAttachment.getImageView();
             fbkey.resolve[i] = VK_NULL_HANDLE;
             assert_invariant(fbkey.color[i]);
         } else {
             auto& msaaColorAttachment = rt->getMsaaColor(i);
-            auto const& msaaRange = attachment.getSubresourceRange();
-            msaaColorAttachment.texture->transitionLayout(&commands,
-                    msaaRange, VulkanLayout::COLOR_ATTACHMENT);
             renderPassAttachments.insert(msaaColorAttachment);
 
+            auto& colorAttachment = rt->getColor(i);
             fbkey.color[i] = msaaColorAttachment.getImageView();
 
-            VulkanTexture* texture = attachment.texture;
+            VulkanTexture* texture = colorAttachment.texture;
             if (texture->samples == 1) {
-                auto const& range = attachment.getSubresourceRange();
-                attachment.texture->transitionLayout(&commands,
-                        range, VulkanLayout::COLOR_ATTACHMENT);
-                renderPassAttachments.insert(attachment);
-                fbkey.resolve[i] = attachment.getImageView();
+                mRenderPassFboInfo.hasColorResolve = true;
+
+                renderPassAttachments.insert(colorAttachment);
+                fbkey.resolve[i] = colorAttachment.getImageView();
                 assert_invariant(fbkey.resolve[i]);
             }
             assert_invariant(fbkey.color[i]);
@@ -1472,27 +1448,57 @@ void VulkanDriver::endRenderPass(int) {
     assert_invariant(rt);
 
     // Since we might soon be sampling from the render target that we just wrote to, we need a
-    // pipeline barrier between framebuffer writes and shader reads.
+    // pipeline barrier between framebuffer writes and shader reads. This is a memory barrier rather
+    // than an image barrier. If we were to use image barriers here, we would potentially need to
+    // issue several of them when considering MRT. This would be very complex to set up and would
+    // require more state tracking, so we've chosen to use a memory barrier for simplicity and
+    // correctness.
     if (!rt->isSwapChain()) {
-        for (auto& attachment: mRenderPassFboInfo.attachments) {
-            auto const& range = attachment.getSubresourceRange();
+        for (auto const& attachment: mRenderPassFboInfo.attachments) {
             bool const isDepth = attachment.isDepth();
-            auto texture = attachment.texture;
-            if (isDepth) {
-                texture->setLayout(range, VulkanFboCache::FINAL_DEPTH_ATTACHMENT_LAYOUT);
-                if (!texture->transitionLayout(&commands, range, VulkanLayout::DEPTH_SAMPLER)) {
-                    texture->attachmentToSamplerBarrier(&commands, range);
-                }
-            } else {
-                texture->setLayout(range, VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT);
-                if (!texture->transitionLayout(&commands, range, VulkanLayout::READ_WRITE)) {
-                    texture->attachmentToSamplerBarrier(&commands, range);
-                }
+            VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+
+            // This is a workaround around a validation issue (might not be an actual driver issue).
+            if (mRenderPassFboInfo.hasColorResolve && !isDepth) {
+                srcStageMask = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
             }
+
+            VkPipelineStageFlags dstStageMask =
+                    VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
+            VkAccessFlags srcAccess = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+            VkAccessFlags dstAccess = VK_ACCESS_SHADER_READ_BIT;
+            VulkanLayout layout = VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT;
+            if (isDepth) {
+                srcAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+                dstAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
+                srcStageMask = VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
+                dstStageMask = VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
+                layout =  VulkanFboCache::FINAL_DEPTH_ATTACHMENT_LAYOUT;
+            }
+
+            auto const vkLayout = imgutil::getVkLayout(layout);
+            auto const& range = attachment.getSubresourceRange();
+            VkImageMemoryBarrier barrier = {
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = srcAccess,
+                .dstAccessMask = dstAccess,
+                .oldLayout = vkLayout,
+                .newLayout = vkLayout,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = attachment.getImage(),
+                .subresourceRange = range,
+            };
+
+            attachment.texture->setLayout(range, layout);
+            vkCmdPipelineBarrier(cmdbuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr,
+                    1, &barrier);
         }
     }
 
-    mRenderPassFboInfo = {};
+    mRenderPassFboInfo.clear();
+    mDescriptorSetManager.clearState();
     mCurrentRenderPass.renderTarget = nullptr;
     mCurrentRenderPass.renderPass = VK_NULL_HANDLE;
     FVK_SYSTRACE_END();
@@ -1551,6 +1557,33 @@ void VulkanDriver::commit(Handle<HwSwapChain> sch) {
     FVK_SYSTRACE_END();
 }
 
+void VulkanDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> boh) {
+    auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
+    VkDeviceSize const offset = 0;
+    VkDeviceSize const size = VK_WHOLE_SIZE;
+    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
+}
+
+void VulkanDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t index,
+        Handle<HwBufferObject> boh, uint32_t offset, uint32_t size) {
+
+    assert_invariant(bindingType == BufferObjectBinding::UNIFORM);
+
+    // TODO: implement BufferObjectBinding::SHADER_STORAGE case
+
+    auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
+    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
+}
+
+void VulkanDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
+    mDescriptorSetManager.clearBuffer((uint32_t) index);
+}
+
+void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
+    auto* hwsb = mResourceAllocator.handle_cast<VulkanSamplerGroup*>(sbh);
+    mSamplerBindings[index] = hwsb;
+}
+
 void VulkanDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
         backend::PushConstantVariant value) {
     assert_invariant(mBoundPipeline.program && "Expect a program when writing to push constants");
@@ -1559,13 +1592,13 @@ void VulkanDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
             value);
 }
 
-void VulkanDriver::insertEventMarker(char const* string) {
+void VulkanDriver::insertEventMarker(char const* string, uint32_t len) {
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands.insertEventMarker(string, strlen(string));
+    mCommands.insertEventMarker(string, len);
 #endif
 }
 
-void VulkanDriver::pushGroupMarker(char const* string) {
+void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
     // Turns out all the markers are 0-terminated, so we can just pass it without len.
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
     mCommands.pushGroupMarker(string);
@@ -1751,8 +1784,8 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
             *mResourceAllocator.handle_cast<VulkanVertexBufferInfo*>(pipelineState.vertexBufferInfo);
 
     Handle<HwProgram> programHandle = pipelineState.program;
-    RasterState const& rasterState = pipelineState.rasterState;
-    PolygonOffset const& depthOffset = pipelineState.polygonOffset;
+    RasterState rasterState = pipelineState.rasterState;
+    PolygonOffset depthOffset = pipelineState.polygonOffset;
 
     auto* program = mResourceAllocator.handle_cast<VulkanProgram*>(programHandle);
     commands->acquire(program);
@@ -1773,7 +1806,6 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
         .dstAlphaBlendFactor = getBlendFactor(rasterState.blendFunctionDstAlpha),
         .colorWriteMask = (VkColorComponentFlags) (rasterState.colorWrite ? 0xf : 0x0),
         .rasterizationSamples = rt->getSamples(),
-        .depthClamp = rasterState.depthClamp,
         .colorTargetCount = rt->getColorTargetCount(mCurrentRenderPass),
         .colorBlendOp = rasterState.blendEquationRGB,
         .alphaBlendOp =  rasterState.blendEquationAlpha,
@@ -1796,31 +1828,69 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
     mPipelineCache.bindPrimitiveTopology(topology);
     mPipelineCache.bindVertexArray(attribDesc, bufferDesc, vbi.getAttributeCount());
 
-    auto& setLayouts = pipelineState.pipelineLayout.setLayout;
-    VulkanDescriptorSetLayout::DescriptorSetLayoutArray layoutList;
-    uint8_t layoutCount = 0;
-    std::transform(setLayouts.begin(), setLayouts.end(), layoutList.begin(),
-            [&](Handle<HwDescriptorSetLayout> handle) -> VkDescriptorSetLayout {
-                if (!handle) {
-                    return VK_NULL_HANDLE;
-                }
-                auto layout = mResourceAllocator.handle_cast<VulkanDescriptorSetLayout*>(handle);
-                layoutCount++;
-                return layout->getVkLayout();
-            });
-    auto pipelineLayout = mPipelineLayoutCache.getLayout(layoutList, program);
+    // Query the program for the mapping from (SamplerGroupBinding,Offset) to (SamplerBinding),
+    // where "SamplerBinding" is the integer in the GLSL, and SamplerGroupBinding is the abstract
+    // Filament concept used to form groups of samplers.
 
-    constexpr uint8_t descriptorSetMaskTable[4] = {0x1, 0x3, 0x7, 0xF};
+    auto const& bindingToSamplerIndex = program->getBindingToSamplerIndex();
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+    auto const& bindingToName = program->getBindingToName();
+#endif
 
+    for (auto binding: program->getBindings()) {
+        uint16_t const indexPair = bindingToSamplerIndex[binding];
+        if (indexPair == 0xffff) {
+            continue;
+        }
+
+        uint16_t const samplerGroupInd = (indexPair >> 8) & 0xff;
+        uint16_t const samplerInd = (indexPair & 0xff);
+
+        VulkanSamplerGroup* vksb = mSamplerBindings[samplerGroupInd];
+        if (!vksb) {
+            continue;
+        }
+        SamplerDescriptor const* boundSampler = ((SamplerDescriptor*) vksb->sb->data()) + samplerInd;
+
+        if (UTILS_UNLIKELY(!boundSampler->t)) {
+            continue;
+        }
+        VulkanTexture* texture = mResourceAllocator.handle_cast<VulkanTexture*>(boundSampler->t);
+
+        // TODO: can this uninitialized check be checked in a higher layer?
+        // This fallback path is very flaky because the dummy texture might not have
+        // matching characteristics. (e.g. if the missing texture is a 3D texture)
+        if (UTILS_UNLIKELY(texture->getPrimaryImageLayout() == VulkanLayout::UNDEFINED)) {
+#if FVK_ENABLED(FVK_DEBUG_TEXTURE) && FVK_ENABLED_DEBUG_SAMPLER_NAME
+            FVK_LOGW << "Uninitialized texture bound to '" << bindingToName[binding] << "'";
+            FVK_LOGW << " in material '" << program->name.c_str() << "'";
+            FVK_LOGW << " at binding point " << +binding << utils::io::endl;
+#endif
+            texture = mEmptyTexture;
+        }
+
+        VkSampler const vksampler = mSamplerCache.getSampler(boundSampler->s);
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+        VulkanDriver::DebugUtils::setName(VK_OBJECT_TYPE_SAMPLER,
+                reinterpret_cast<uint64_t>(vksampler), bindingToName[binding].c_str());
+#endif
+        mDescriptorSetManager.updateSampler({}, binding, texture, vksampler);
+    }
+
+    auto const pipelineLayout = mDescriptorSetManager.bind(commands, program, mGetPipelineFunction);
     mBoundPipeline = {
         .program = program,
         .pipelineLayout = pipelineLayout,
-        .descriptorSetMask = DescriptorSetMask(descriptorSetMaskTable[layoutCount]),
     };
 
     mPipelineCache.bindLayout(pipelineLayout);
     mPipelineCache.bindPipeline(commands);
 
+    // Since we don't statically define scissor as part of the pipeline, we need to call scissor at
+    // least once. Context: VUID-vkCmdDrawIndexed-None-07832.
+    auto const& extent = rt->getExtent();
+    scissor({0, 0, extent.width, extent.height});
+
     FVK_SYSTRACE_END();
 }
 
@@ -1853,14 +1923,6 @@ void VulkanDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::bindDescriptorSet(
-        backend::DescriptorSetHandle dsh,
-        backend::descriptor_set_t setIndex,
-        backend::DescriptorSetOffsetArray&& offsets) {
-    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
-    mDescriptorSetManager.bind(setIndex, set, std::move(offsets));
-}
-
 void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("draw2");
@@ -1868,8 +1930,8 @@ void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
     VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer cmdbuffer = commands.buffer();
 
-    mDescriptorSetManager.commit(&commands, mBoundPipeline.pipelineLayout,
-            mBoundPipeline.descriptorSetMask);
+    // Bind "dynamic" UBOs if they need to change.
+    mDescriptorSetManager.dynamicBind(&commands, {});
 
     // Finally, make the actual draw call. TODO: support subranges
     const uint32_t firstIndex = indexOffset;
@@ -1899,9 +1961,6 @@ void VulkanDriver::scissor(Viewport scissorBox) {
     VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer cmdbuffer = commands.buffer();
 
-    // TODO: it's a common case that scissor() is called with (0, 0, maxint, maxint)
-    //       we should maybe have a fast path for this and avoid vkCmdSetScissor() if possible
-
     // Set scissoring.
     // clamp left-bottom to 0,0 and avoid overflows
     constexpr int32_t maxvali  = std::numeric_limits<int32_t>::max();
@@ -1920,7 +1979,7 @@ void VulkanDriver::scissor(Viewport scissorBox) {
             .extent = { uint32_t(r - l), uint32_t(t - b) }
     };
 
-    VulkanRenderTarget const* rt = mCurrentRenderPass.renderTarget;
+    const VulkanRenderTarget* rt = mCurrentRenderPass.renderTarget;
     rt->transformClientRectToPlatform(&scissor);
     vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
 }
@@ -1963,10 +2022,6 @@ void VulkanDriver::debugCommandBegin(CommandStream* cmds, bool synchronous, cons
 void VulkanDriver::resetState(int) {
 }
 
-void VulkanDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
-    mResourceAllocator.associateHandle(handleId, std::move(tag));
-}
-
 // explicit instantiation of the Dispatcher
 template class ConcreteDispatcher<VulkanDriver>;
 

filament/backend/src/vulkan/VulkanDriver.h

@@ -47,6 +47,21 @@ struct VulkanSamplerGroup;
 constexpr uint8_t MAX_RENDERTARGET_ATTACHMENT_TEXTURES =
         MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT * 2 + 1;
 
+// We need to store information about a render pass to enable better barriers at the end of a
+// renderpass.
+struct RenderPassFboBundle {
+    using AttachmentArray =
+            CappedArray<VulkanAttachment, MAX_RENDERTARGET_ATTACHMENT_TEXTURES>;
+
+    AttachmentArray attachments;
+    bool hasColorResolve = false;
+
+    void clear() {
+        attachments.clear();
+        hasColorResolve = false;
+    }
+};
+
 class VulkanDriver final : public DriverBase {
 public:
     static Driver* create(VulkanPlatform* platform, VulkanContext const& context,
@@ -115,6 +130,10 @@ private:
     VulkanPlatform* mPlatform = nullptr;
     std::unique_ptr<VulkanTimestamps> mTimestamps;
 
+    // Placeholder resources
+    VulkanTexture* mEmptyTexture;
+    VulkanBufferObject* mEmptyBufferObject;
+
     VulkanSwapChain* mCurrentSwapChain = nullptr;
     VulkanRenderTarget* mDefaultRenderTarget = nullptr;
     VulkanRenderPass mCurrentRenderPass = {};
@@ -140,19 +159,15 @@ private:
     VulkanReadPixels mReadPixels;
     VulkanDescriptorSetManager mDescriptorSetManager;
 
+    VulkanDescriptorSetManager::GetPipelineLayoutFunction mGetPipelineFunction;
+
     // This is necessary for us to write to push constants after binding a pipeline.
-    struct {
+    struct BoundPipeline {
         VulkanProgram* program;
         VkPipelineLayout pipelineLayout;
-        DescriptorSetMask descriptorSetMask;
-    } mBoundPipeline = {};
-
-    // We need to store information about a render pass to enable better barriers at the end of a
-    // renderpass.
-    struct {
-        using AttachmentArray = CappedArray<VulkanAttachment, MAX_RENDERTARGET_ATTACHMENT_TEXTURES>;
-        AttachmentArray attachments;
-    } mRenderPassFboInfo = {};
+    };
+    BoundPipeline mBoundPipeline = {};
+    RenderPassFboBundle mRenderPassFboInfo;
 
     bool const mIsSRGBSwapChainSupported;
     backend::StereoscopicType const mStereoscopicType;

filament/backend/src/vulkan/VulkanFboCache.cpp

@@ -31,6 +31,7 @@ namespace filament::backend {
 
 bool VulkanFboCache::RenderPassEq::operator()(const RenderPassKey& k1,
         const RenderPassKey& k2) const {
+    if (k1.initialColorLayoutMask != k2.initialColorLayoutMask) return false;
     if (k1.initialDepthLayout != k2.initialDepthLayout) return false;
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
         if (k1.colorFormat[i] != k2.colorFormat[i]) return false;
@@ -41,9 +42,7 @@ bool VulkanFboCache::RenderPassEq::operator()(const RenderPassKey& k1,
     if (k1.discardEnd != k2.discardEnd) return false;
     if (k1.samples != k2.samples) return false;
     if (k1.needsResolveMask != k2.needsResolveMask) return false;
-    if (k1.usesLazilyAllocatedMemory != k2.usesLazilyAllocatedMemory) return false;
     if (k1.subpassMask != k2.subpassMask) return false;
-    if (k1.viewCount != k2.viewCount) return false;
     return true;
 }
 
@@ -187,25 +186,6 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
         .pDependencies = dependencies
     };
 
-    VkRenderPassMultiviewCreateInfo multiviewCreateInfo = {};
-    uint32_t const subpassViewMask = (1 << config.viewCount) - 1;
-    // Prepare a view mask array for the maximum number of subpasses. All subpasses have all views
-    // activated.
-    uint32_t const viewMasks[2] = {subpassViewMask, subpassViewMask};
-    if (config.viewCount > 1) {
-        // Fill the multiview create info.
-        multiviewCreateInfo.sType = VK_STRUCTURE_TYPE_RENDER_PASS_MULTIVIEW_CREATE_INFO;
-        multiviewCreateInfo.pNext = nullptr;
-        multiviewCreateInfo.subpassCount = hasSubpasses ? 2u : 1u;
-        multiviewCreateInfo.pViewMasks = viewMasks;
-        multiviewCreateInfo.dependencyCount = 0;
-        multiviewCreateInfo.pViewOffsets = nullptr;
-        multiviewCreateInfo.correlationMaskCount = 1;
-        multiviewCreateInfo.pCorrelationMasks = &subpassViewMask;
-
-        renderPassInfo.pNext = &multiviewCreateInfo;
-    }
-
     int attachmentIndex = 0;
 
     // Populate the Color Attachments.
@@ -255,10 +235,12 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
             .format = config.colorFormat[i],
             .samples = (VkSampleCountFlagBits) config.samples,
             .loadOp = clear ? kClear : (discard ? kDontCare : kKeep),
-            .storeOp = (config.usesLazilyAllocatedMemory & (1 << i)) ? kDisableStore : kEnableStore,
+            .storeOp = kEnableStore,
             .stencilLoadOp = kDontCare,
             .stencilStoreOp = kDisableStore,
-            .initialLayout = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
+            .initialLayout = ((!discard && config.initialColorLayoutMask & (1 << i)) || clear)
+                                     ? imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT)
+                                     : imgutil::getVkLayout(VulkanLayout::UNDEFINED),
             .finalLayout = imgutil::getVkLayout(FINAL_COLOR_ATTACHMENT_LAYOUT),
         };
     }

filament/backend/src/vulkan/VulkanFboCache.h

@@ -42,25 +42,36 @@ public:
     // RenderPassKey is a small POD representing the immutable state that is used to construct
     // a VkRenderPass. It is hashed and used as a lookup key.
     struct alignas(8) RenderPassKey {
+        // For each target, we need to know three image layouts: the layout BEFORE the pass, the
+        // layout DURING the pass, and the layout AFTER the pass. Here are the rules:
+        // - For depth, we explicitly specify all three layouts.
+        // - Color targets have their initial image layout specified with a bitmask.
+        // - For each color target, the pre-existing layout is either UNDEFINED (0) or GENERAL (1).
+        // - The render pass and final images layout for color buffers is always
+        //   VulkanLayout::COLOR_ATTACHMENT.
+        uint8_t initialColorLayoutMask;
+
+        // Note that if VulkanLayout grows beyond 16, we'd need to up this.
+        VulkanLayout initialDepthLayout : 8;
+        uint8_t padding0;
+        uint8_t padding1;
+
         VkFormat colorFormat[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT]; // 32 bytes
         VkFormat depthFormat; // 4 bytes
         TargetBufferFlags clear; // 4 bytes
         TargetBufferFlags discardStart; // 4 bytes
         TargetBufferFlags discardEnd; // 4 bytes
-
-        VulkanLayout initialDepthLayout; // 1 byte
         uint8_t samples; // 1 byte
         uint8_t needsResolveMask; // 1 byte
-        uint8_t usesLazilyAllocatedMemory; // 1 byte
         uint8_t subpassMask; // 1 byte
-        uint8_t viewCount; // 1 byte
-		uint8_t padding[2];
+        uint8_t padding2; // 1 byte
     };
     struct RenderPassVal {
         VkRenderPass handle;
         uint32_t timestamp;
     };
     static_assert(0 == MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT % 8);
+    static_assert(sizeof(RenderPassKey::initialColorLayoutMask) == MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT / 8);
     static_assert(sizeof(TargetBufferFlags) == 4, "TargetBufferFlags has unexpected size.");
     static_assert(sizeof(VkFormat) == 4, "VkFormat has unexpected size.");
     static_assert(sizeof(RenderPassKey) == 56, "RenderPassKey has unexpected size.");

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -16,15 +16,13 @@
 
 #include "VulkanHandles.h"
 
-#include "VulkanConstants.h"
-
-// TODO: remove this by moving DebugUtils out of VulkanDriver
 #include "VulkanDriver.h"
-
+#include "VulkanConstants.h"
+#include "VulkanDriver.h"
 #include "VulkanMemory.h"
-#include "VulkanResourceAllocator.h"
 #include "VulkanUtility.h"
 #include "spirv/VulkanSpirvUtils.h"
+#include "utils/Log.h"
 
 #include <backend/platforms/VulkanPlatform.h>
 
@@ -56,14 +54,62 @@ void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
 }
 
 template<typename Bitmask>
-inline void fromStageFlags(backend::ShaderStageFlags stage, descriptor_binding_t binding,
-        Bitmask& mask) {
-    if ((bool) (stage & ShaderStageFlags::VERTEX)) {
-        mask.set(binding + getVertexStageShift<Bitmask>());
+static constexpr Bitmask fromStageFlags(ShaderStageFlags2 flags, uint8_t binding) {
+    Bitmask ret = 0;
+    if (flags & ShaderStageFlags2::VERTEX) {
+        ret |= (getVertexStage<Bitmask>() << binding);
     }
-    if ((bool) (stage & ShaderStageFlags::FRAGMENT)) {
-        mask.set(binding + getFragmentStageShift<Bitmask>());
+    if (flags & ShaderStageFlags2::FRAGMENT) {
+        ret |= (getFragmentStage<Bitmask>() << binding);
     }
+    return ret;
+}
+
+constexpr decltype(VulkanProgram::MAX_SHADER_MODULES) MAX_SHADER_MODULES =
+        VulkanProgram::MAX_SHADER_MODULES;
+
+using LayoutDescriptionList = VulkanProgram::LayoutDescriptionList;
+
+template<typename Bitmask>
+void addDescriptors(Bitmask mask,
+        utils::FixedCapacityVector<DescriptorSetLayoutBinding>& outputList) {
+    constexpr uint8_t MODULE_OFFSET = (sizeof(Bitmask) * 8) / MAX_SHADER_MODULES;
+    for (uint8_t i = 0; i < MODULE_OFFSET; ++i) {
+        bool const hasVertex = (mask & (1ULL << i)) != 0;
+        bool const hasFragment = (mask & (1ULL << (MODULE_OFFSET + i))) != 0;
+        if (!hasVertex && !hasFragment) {
+            continue;
+        }
+
+        DescriptorSetLayoutBinding binding{
+            .binding = i,
+            .flags = DescriptorFlags::NONE,
+            .count = 0,// This is always 0 for now as we pass the size of the UBOs in the Driver API
+                       // instead.
+        };
+        if (hasVertex) {
+            binding.stageFlags = ShaderStageFlags2::VERTEX;
+        }
+        if (hasFragment) {
+            binding.stageFlags = static_cast<ShaderStageFlags2>(
+                    binding.stageFlags | ShaderStageFlags2::FRAGMENT);
+        }
+        if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
+            binding.type = DescriptorType::UNIFORM_BUFFER;
+        } else if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
+            binding.type = DescriptorType::SAMPLER;
+        } else if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
+            binding.type = DescriptorType::INPUT_ATTACHMENT;
+        }
+        outputList.push_back(binding);
+    }
+}
+
+inline VkDescriptorSetLayout createDescriptorSetLayout(VkDevice device,
+        VkDescriptorSetLayoutCreateInfo const& info) {
+    VkDescriptorSetLayout layout;
+    vkCreateDescriptorSetLayout(device, &info, VKALLOC, &layout);
+    return layout;
 }
 
 inline VkShaderStageFlags getVkStage(backend::ShaderStage stage) {
@@ -77,48 +123,20 @@ inline VkShaderStageFlags getVkStage(backend::ShaderStage stage) {
     }
 }
 
-using BitmaskGroup = VulkanDescriptorSetLayout::Bitmask;
-BitmaskGroup fromBackendLayout(DescriptorSetLayout const& layout) {
-    BitmaskGroup mask;
-    for (auto const& binding: layout.bindings) {
-        switch (binding.type) {
-            case DescriptorType::UNIFORM_BUFFER: {
-                if ((binding.flags & DescriptorFlags::DYNAMIC_OFFSET) != DescriptorFlags::NONE) {
-                    fromStageFlags(binding.stageFlags, binding.binding, mask.dynamicUbo);
-                } else {
-                    fromStageFlags(binding.stageFlags, binding.binding, mask.ubo);
-                }
-                break;
-            }
-            case DescriptorType::SAMPLER: {
-                fromStageFlags(binding.stageFlags, binding.binding, mask.sampler);
-                break;
-            }
-            case DescriptorType::INPUT_ATTACHMENT: {
-                fromStageFlags(binding.stageFlags, binding.binding, mask.inputAttachment);
-                break;
-            }
-            case DescriptorType::SHADER_STORAGE_BUFFER:
-                PANIC_POSTCONDITION("Shader storage is not supported");
-                break;
-        }
-    }
-    return mask;
-}
-
 } // anonymous namespace
 
-VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(DescriptorSetLayout const& layout)
+
+VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(VkDevice device,
+        VkDescriptorSetLayoutCreateInfo const& info, Bitmask const& bitmask)
     : VulkanResource(VulkanResourceType::DESCRIPTOR_SET_LAYOUT),
-      bitmask(fromBackendLayout(layout)),
+      mDevice(device),
+      vklayout(createDescriptorSetLayout(device, info)),
+      bitmask(bitmask),
+      bindings(getBindings(bitmask)),
       count(Count::fromLayoutBitmask(bitmask)) {}
 
-void VulkanDescriptorSet::acquire(VulkanTexture* texture) {
-    mResources.acquire(texture);
-}
-
-void VulkanDescriptorSet::acquire(VulkanBufferObject* bufferObject) {
-    mResources.acquire(bufferObject);
+VulkanDescriptorSetLayout::~VulkanDescriptorSetLayout() {
+    vkDestroyDescriptorSetLayout(mDevice, vklayout, VKALLOC);
 }
 
 PushConstantDescription::PushConstantDescription(backend::Program const& program) noexcept {
@@ -172,11 +190,24 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
       mInfo(new(std::nothrow) PipelineInfo(builder)),
       mDevice(device) {
 
+    constexpr uint8_t UBO_MODULE_OFFSET = (sizeof(UniformBufferBitmask) * 8) / MAX_SHADER_MODULES;
+    constexpr uint8_t SAMPLER_MODULE_OFFSET = (sizeof(SamplerBitmask) * 8) / MAX_SHADER_MODULES;
+    constexpr uint8_t INPUT_ATTACHMENT_MODULE_OFFSET =
+            (sizeof(InputAttachmentBitmask) * 8) / MAX_SHADER_MODULES;
+
     Program::ShaderSource const& blobs = builder.getShadersSource();
     auto& modules = mInfo->shaders;
+
     auto const& specializationConstants = builder.getSpecializationConstants();
+
     std::vector<uint32_t> shader;
 
+    // TODO: this will be moved out of the shader as the descriptor set layout will be provided by
+    // Filament instead of parsed from the shaders. See [GDSR] in VulkanDescriptorSetManager.h
+    UniformBufferBitmask uboMask = 0;
+    SamplerBitmask samplerMask = 0;
+    InputAttachmentBitmask inputAttachmentMask = 0;
+
     static_assert(static_cast<ShaderStage>(0) == ShaderStage::VERTEX &&
             static_cast<ShaderStage>(1) == ShaderStage::FRAGMENT &&
             MAX_SHADER_MODULES == 2);
@@ -193,6 +224,12 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
             dataSize = shader.size() * 4;
         }
 
+        auto const [ubo, sampler, inputAttachment] = getProgramBindings(blob);
+        uboMask |= (static_cast<UniformBufferBitmask>(ubo) << (UBO_MODULE_OFFSET * i));
+        samplerMask |= (static_cast<SamplerBitmask>(sampler) << (SAMPLER_MODULE_OFFSET * i));
+        inputAttachmentMask |= (static_cast<InputAttachmentBitmask>(inputAttachment)
+                                << (INPUT_ATTACHMENT_MODULE_OFFSET * i));
+
         VkShaderModule& module = modules[i];
         VkShaderModuleCreateInfo moduleInfo = {
             .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
@@ -220,6 +257,40 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
     }
 
+    LayoutDescriptionList& layouts = mInfo->layouts;
+    layouts[0].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
+            countBits(collapseStages(uboMask)));
+    layouts[1].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
+            countBits(collapseStages(samplerMask)));
+    layouts[2].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
+            countBits(collapseStages(inputAttachmentMask)));
+
+    addDescriptors(uboMask, layouts[0].bindings);
+    addDescriptors(samplerMask, layouts[1].bindings);
+    addDescriptors(inputAttachmentMask, layouts[2].bindings);
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+    auto& bindingToName = mInfo->bindingToName;
+#endif
+
+    auto& groupInfo = builder.getSamplerGroupInfo();
+    auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
+    auto& bindings = mInfo->bindings;
+    for (uint8_t groupInd = 0; groupInd < Program::SAMPLER_BINDING_COUNT; groupInd++) {
+        auto const& group = groupInfo[groupInd];
+        auto const& samplers = group.samplers;
+        for (size_t i = 0; i < samplers.size(); ++i) {
+            uint32_t const binding = samplers[i].binding;
+            bindingToSamplerIndex[binding] = (groupInd << 8) | (0xff & i);
+            assert_invariant(bindings.find(binding) == bindings.end());
+            bindings.insert(binding);
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+            bindingToName[binding] = samplers[i].name.c_str();
+#endif
+        }
+    }
+
 #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
     FVK_LOGD << "Created VulkanProgram " << builder << ", shaders = (" << modules[0]
              << ", " << modules[1] << ")" << utils::io::endl;
@@ -250,15 +321,13 @@ void VulkanRenderTarget::bindToSwapChain(VulkanSwapChain& swapChain) {
 
 VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physicalDevice,
         VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
-        VulkanResourceAllocator* handleAllocator,
         uint32_t width, uint32_t height, uint8_t samples,
         VulkanAttachment color[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
-        VulkanAttachment depthStencil[2], VulkanStagePool& stagePool, uint8_t layerCount)
+        VulkanAttachment depthStencil[2], VulkanStagePool& stagePool)
     : HwRenderTarget(width, height),
       VulkanResource(VulkanResourceType::RENDER_TARGET),
       mOffscreen(true),
-      mSamples(samples),
-      mLayerCount(layerCount) {
+      mSamples(samples) {
     for (int index = 0; index < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; index++) {
         mColor[index] = color[index];
     }
@@ -282,16 +351,10 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica
         if (texture && texture->samples == 1) {
             auto msTexture = texture->getSidecar();
             if (UTILS_UNLIKELY(!msTexture)) {
-                // Clear all usage flags that are not related to attachments, so that we can
-                // use the transient usage flag.
-                const TextureUsage usage = texture->usage & TextureUsage::ALL_ATTACHMENTS;
-                assert_invariant(static_cast<uint16_t>(usage) != 0U);
-
                 // TODO: This should be allocated with the ResourceAllocator.
                 msTexture = new VulkanTexture(device, physicalDevice, context, allocator, commands,
-                        handleAllocator,
                         texture->target, ((VulkanTexture const*) texture)->levels, texture->format,
-                        samples, texture->width, texture->height, texture->depth, usage,
+                        samples, texture->width, texture->height, texture->depth, texture->usage,
                         stagePool, true /* heap allocated */);
                 texture->setSidecar(msTexture);
             }
@@ -319,8 +382,7 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica
     VulkanTexture* msTexture = depthTexture->getSidecar();
     if (UTILS_UNLIKELY(!msTexture)) {
         msTexture = new VulkanTexture(device, physicalDevice, context, allocator,
-                commands, handleAllocator,
-                depthTexture->target, msLevel, depthTexture->format, samples,
+                commands, depthTexture->target, msLevel, depthTexture->format, samples,
                 depthTexture->width, depthTexture->height, depthTexture->depth, depthTexture->usage,
                 stagePool, true /* heap allocated */);
         depthTexture->setSidecar(msTexture);
@@ -472,7 +534,15 @@ bool VulkanTimerQuery::isCompleted() noexcept {
     // timestamp has at least been written into a processed command buffer.
 
     // This fence indicates that the corresponding buffer has been completed.
-    return mFence && mFence->getStatus() == VK_SUCCESS;
+    if (!mFence) {
+        return false;
+    }
+    VkResult status = mFence->status.load(std::memory_order_relaxed);
+    if (status != VK_SUCCESS) {
+        return false;
+    }
+
+    return true;
 }
 
 VulkanTimerQuery::~VulkanTimerQuery() = default;
@@ -488,4 +558,35 @@ VulkanRenderPrimitive::VulkanRenderPrimitive(VulkanResourceAllocator* resourceAl
     mResources.acquire(indexBuffer);
 }
 
+using Bitmask = VulkanDescriptorSetLayout::Bitmask;
+
+Bitmask Bitmask::fromBackendLayout(descset::DescriptorSetLayout const& layout) {
+    Bitmask mask;
+    for (auto const& binding: layout.bindings) {
+        switch (binding.type) {
+            case descset::DescriptorType::UNIFORM_BUFFER: {
+                if (binding.flags == descset::DescriptorFlags::DYNAMIC_OFFSET) {
+                    mask.dynamicUbo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
+                            binding.binding);
+                } else {
+                    mask.ubo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
+                            binding.binding);
+                }
+                break;
+            }
+            case descset::DescriptorType::SAMPLER: {
+                mask.sampler |= fromStageFlags<SamplerBitmask>(binding.stageFlags, binding.binding);
+                break;
+            }
+            case descset::DescriptorType::INPUT_ATTACHMENT: {
+                mask.inputAttachment |=
+                        fromStageFlags<InputAttachmentBitmask>(binding.stageFlags, binding.binding);
+                break;
+            }
+        }
+    }
+    return mask;
+}
+
+
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanHandles.h

@@ -29,52 +29,39 @@
 #include <private/backend/SamplerGroup.h>
 #include <backend/Program.h>
 
-#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Mutex.h>
 #include <utils/StructureOfArrays.h>
 
 namespace filament::backend {
 
-namespace {
-// Counts the total number of descriptors for both vertex and fragment stages.
-template<typename Bitmask>
-inline uint8_t collapsedCount(Bitmask const& mask) {
-    static_assert(sizeof(mask) <= 64);
-    constexpr uint64_t VERTEX_MASK = (1ULL << getFragmentStageShift<Bitmask>()) - 1ULL;
-    constexpr uint64_t FRAGMENT_MASK = (VERTEX_MASK << getFragmentStageShift<Bitmask>());
-    uint64_t val = mask.getValue();
-    val = ((val & VERTEX_MASK) >> getVertexStageShift<Bitmask>()) |
-        ((val & FRAGMENT_MASK) >> getFragmentStageShift<Bitmask>());
-    return (uint8_t) Bitmask(val).count();
-}
-
-} // anonymous namespace
+using namespace descset;
 
 class VulkanTimestamps;
-struct VulkanBufferObject;
 
-struct VulkanDescriptorSetLayout : public VulkanResource, HwDescriptorSetLayout {
-    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 4;
-    static constexpr uint8_t MAX_BINDINGS = 25;
-
-    using DescriptorSetLayoutArray = std::array<VkDescriptorSetLayout,
-            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+struct VulkanDescriptorSetLayout : public VulkanResource {
+    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 3;
 
     // The bitmask representation of a set layout.
     struct Bitmask {
-        // TODO: better utiltize the space below and use bitset instead.
-        UniformBufferBitmask ubo;         // 8 bytes
-        UniformBufferBitmask dynamicUbo;  // 8 bytes
-        SamplerBitmask sampler;           // 8 bytes
-        InputAttachmentBitmask inputAttachment; // 8 bytes
+        UniformBufferBitmask ubo = 0;         // 4 bytes
+        UniformBufferBitmask dynamicUbo = 0;  // 4 bytes
+        SamplerBitmask sampler = 0;           // 8 bytes
+        InputAttachmentBitmask inputAttachment = 0; // 1 bytes
+
+        // Because we're using this struct as hash key, must make it's 8-bytes aligned, with no
+        // unaccounted bytes.
+        uint8_t padding0 = 0; // 1 bytes
+        uint16_t padding1 = 0;// 2 bytes
+        uint32_t padding2 = 0;// 4 bytes
 
         bool operator==(Bitmask const& right) const {
             return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
                    inputAttachment == right.inputAttachment;
         }
+
+        static Bitmask fromBackendLayout(descset::DescriptorSetLayout const& layout);
     };
-    static_assert(sizeof(Bitmask) == 32);
 
     // This is a convenience struct to quickly check layout compatibility in terms of descriptor set
     // pools.
@@ -84,10 +71,6 @@ struct VulkanDescriptorSetLayout : public VulkanResource, HwDescriptorSetLayout
         uint32_t sampler = 0;
         uint32_t inputAttachment = 0;
 
-        inline uint32_t total() const {
-            return ubo + dynamicUbo + sampler + inputAttachment;
-        }
-
         bool operator==(Count const& right) const noexcept {
             return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
                    inputAttachment == right.inputAttachment;
@@ -95,10 +78,10 @@ struct VulkanDescriptorSetLayout : public VulkanResource, HwDescriptorSetLayout
 
         static inline Count fromLayoutBitmask(Bitmask const& mask) {
             return {
-                .ubo = collapsedCount(mask.ubo),
-                .dynamicUbo = collapsedCount(mask.dynamicUbo),
-                .sampler = collapsedCount(mask.sampler),
-                .inputAttachment = collapsedCount(mask.inputAttachment),
+                    .ubo = countBits(collapseStages(mask.ubo)),
+                    .dynamicUbo = countBits(collapseStages(mask.dynamicUbo)),
+                    .sampler = countBits(collapseStages(mask.sampler)),
+                    .inputAttachment = countBits(collapseStages(mask.inputAttachment)),
             };
         }
 
@@ -114,50 +97,89 @@ struct VulkanDescriptorSetLayout : public VulkanResource, HwDescriptorSetLayout
         }
     };
 
-    VulkanDescriptorSetLayout(DescriptorSetLayout const& layout);
+    static_assert(sizeof(Bitmask) % 8 == 0);
 
-    ~VulkanDescriptorSetLayout() = default;
+    explicit VulkanDescriptorSetLayout(VkDevice device, VkDescriptorSetLayoutCreateInfo const& info,
+            Bitmask const& bitmask);
 
-    VkDescriptorSetLayout getVkLayout() const { return mVkLayout; }
-    void setVkLayout(VkDescriptorSetLayout vklayout) { mVkLayout = vklayout; }
+    ~VulkanDescriptorSetLayout();
 
+    VkDevice const mDevice;
+    VkDescriptorSetLayout const vklayout;
     Bitmask const bitmask;
+
+    // This is a convenience struct so that we don't have to iterate through all the bits of the
+    // bitmask (which correspondings to binding indices).
+    struct _Bindings {
+        utils::FixedCapacityVector<uint8_t> const ubo;
+        utils::FixedCapacityVector<uint8_t> const dynamicUbo;
+        utils::FixedCapacityVector<uint8_t> const sampler;
+        utils::FixedCapacityVector<uint8_t> const inputAttachment;
+    } bindings;
+
     Count const count;
 
 private:
-    VkDescriptorSetLayout mVkLayout = VK_NULL_HANDLE;
+
+    template <typename MaskType>
+    utils::FixedCapacityVector<uint8_t> bits(MaskType mask) {
+        utils::FixedCapacityVector<uint8_t> ret =
+                utils::FixedCapacityVector<uint8_t>::with_capacity(countBits(mask));
+        for (uint8_t i = 0; i < sizeof(mask) * 4; ++i) {
+            if (mask & (1 << i)) {
+                ret.push_back(i);
+            }
+        }
+        return ret;
+    }
+
+    _Bindings getBindings(Bitmask const& bitmask) {
+        auto const uboCollapsed = collapseStages(bitmask.ubo);
+        auto const dynamicUboCollapsed = collapseStages(bitmask.dynamicUbo);
+        auto const samplerCollapsed = collapseStages(bitmask.sampler);
+        auto const inputAttachmentCollapsed = collapseStages(bitmask.inputAttachment);
+        return {
+            bits(uboCollapsed),
+            bits(dynamicUboCollapsed),
+            bits(samplerCollapsed),
+            bits(inputAttachmentCollapsed),
+        };
+    }
 };
 
-struct VulkanDescriptorSet : public VulkanResource, HwDescriptorSet {
+using VulkanDescriptorSetLayoutList = std::array<Handle<VulkanDescriptorSetLayout>,
+        VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
+struct VulkanDescriptorSet : public VulkanResource {
 public:
     // Because we need to recycle descriptor sets not used, we allow for a callback that the "Pool"
     // can use to repackage the vk handle.
-    using OnRecycle = std::function<void(VulkanDescriptorSet*)>;
+    using OnRecycle = std::function<void()>;
 
-    VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
-            OnRecycle&& onRecycleFn)
+    VulkanDescriptorSet(VulkanResourceAllocator* allocator,
+            VkDescriptorSet rawSet, OnRecycle&& onRecycleFn)
         : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
+          resources(allocator),
           vkSet(rawSet),
-          mResources(allocator),
           mOnRecycleFn(std::move(onRecycleFn)) {}
 
     ~VulkanDescriptorSet() {
         if (mOnRecycleFn) {
-            mOnRecycleFn(this);
+            mOnRecycleFn();
         }
     }
 
-    void acquire(VulkanTexture* texture);
-
-    void acquire(VulkanBufferObject* texture);
-
+    // TODO: maybe change to fixed size for performance.
+    VulkanAcquireOnlyResourceManager resources;
     VkDescriptorSet const vkSet;
 
 private:
-    VulkanAcquireOnlyResourceManager mResources;
     OnRecycle mOnRecycleFn;
 };
 
+using VulkanDescriptorSetList = std::array<Handle<VulkanDescriptorSet>,
+        VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
 using PushConstantNameArray = utils::FixedCapacityVector<char const*>;
 using PushConstantNameByStage = std::array<PushConstantNameArray, Program::SHADER_TYPE_COUNT>;
 
@@ -194,10 +216,24 @@ struct VulkanProgram : public HwProgram, VulkanResource {
 
     inline VkShaderModule getFragmentShader() const { return mInfo->shaders[1]; }
 
+    inline utils::FixedCapacityVector<uint16_t> const& getBindingToSamplerIndex() const {
+        return mInfo->bindingToSamplerIndex;
+    }
+
     // Get a list of the sampler binding indices so that we don't have to loop through all possible
     // samplers.
     inline BindingList const& getBindings() const { return mInfo->bindings; }
 
+    // TODO: this is currently not used. This will replace getLayoutDescriptionList below.
+    // inline descset::DescriptorSetLayout const& getLayoutDescription() const {
+    //    return mInfo->layout;
+    // }
+    // In the usual case, we would have just one layout per program. But in the current setup, we
+    // have a set/layout for each descriptor type. This will be changed in the future.
+    using LayoutDescriptionList = std::array<descset::DescriptorSetLayout,
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+    inline LayoutDescriptionList const& getLayoutDescriptionList() const { return mInfo->layouts; }
+
     inline uint32_t getPushConstantRangeCount() const {
         return mInfo->pushConstantDescription.getVkRangeCount();
     }
@@ -224,7 +260,8 @@ struct VulkanProgram : public HwProgram, VulkanResource {
 private:
     struct PipelineInfo {
         explicit PipelineInfo(backend::Program const& program) noexcept
-            : pushConstantDescription(program)
+            : bindingToSamplerIndex(MAX_SAMPLER_COUNT, 0xffff),
+              pushConstantDescription(program)
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
             , bindingToName(MAX_SAMPLER_COUNT, "")
 #endif
@@ -232,8 +269,14 @@ private:
 
         BindingList bindings;
 
+        // We store the samplerGroupIndex as the top 8-bit and the index within each group as the lower 8-bit.
+        utils::FixedCapacityVector<uint16_t> bindingToSamplerIndex;
         VkShaderModule shaders[MAX_SHADER_MODULES] = { VK_NULL_HANDLE };
 
+        // TODO: Use this instead of `layouts` after Filament-side Descriptor Set API is in place.
+        // descset::DescriptorSetLayout layout;
+        LayoutDescriptionList layouts;
+
         PushConstantDescription pushConstantDescription;
 
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
@@ -259,11 +302,9 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
     // Creates an offscreen render target.
     VulkanRenderTarget(VkDevice device, VkPhysicalDevice physicalDevice,
             VulkanContext const& context, VmaAllocator allocator,
-            VulkanCommands* commands,
-            VulkanResourceAllocator* handleAllocator,
-            uint32_t width, uint32_t height,
+            VulkanCommands* commands, uint32_t width, uint32_t height,
             uint8_t samples, VulkanAttachment color[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
-            VulkanAttachment depthStencil[2], VulkanStagePool& stagePool, uint8_t layerCount);
+            VulkanAttachment depthStencil[2], VulkanStagePool& stagePool);
 
     // Creates a special "default" render target (i.e. associated with the swap chain)
     explicit VulkanRenderTarget();
@@ -278,7 +319,6 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
     VulkanAttachment& getMsaaDepth();
     uint8_t getColorTargetCount(const VulkanRenderPass& pass) const;
     uint8_t getSamples() const { return mSamples; }
-    uint8_t getLayerCount() const { return mLayerCount; }
     bool hasDepth() const { return mDepth.texture; }
     bool isSwapChain() const { return !mOffscreen; }
     void bindToSwapChain(VulkanSwapChain& surf);
@@ -290,7 +330,6 @@ private:
     VulkanAttachment mMsaaDepthAttachment = {};
     const bool mOffscreen : 1;
     uint8_t mSamples : 7;
-    uint8_t mLayerCount = 1;
 };
 
 struct VulkanBufferObject;
@@ -443,7 +482,6 @@ private:
     utils::Mutex mFenceMutex;
 };
 
-
 inline constexpr VkBufferUsageFlagBits getBufferObjectUsage(
         BufferObjectBinding bindingType) noexcept {
     switch(bindingType) {

filament/backend/src/vulkan/VulkanImageUtility.cpp

@@ -131,18 +131,14 @@ getVkTransition(const VulkanLayoutTransition& transition) {
 
 }// anonymous namespace
 
-bool transitionLayout(VkCommandBuffer cmdbuffer,
+void transitionLayout(VkCommandBuffer cmdbuffer,
         VulkanLayoutTransition transition) {
     if (transition.oldLayout == transition.newLayout) {
-        return false;
+        return;
     }
     auto [srcAccessMask, dstAccessMask, srcStage, dstStage, oldLayout, newLayout]
             = getVkTransition(transition);
 
-    if (oldLayout == newLayout) {
-        return false;
-    }
-
     assert_invariant(transition.image != VK_NULL_HANDLE && "No image for transition");
     VkImageMemoryBarrier barrier = {
             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
@@ -156,7 +152,6 @@ bool transitionLayout(VkCommandBuffer cmdbuffer,
             .subresourceRange = transition.subresources,
     };
     vkCmdPipelineBarrier(cmdbuffer, srcStage, dstStage, 0, 0, nullptr, 0, nullptr, 1, &barrier);
-    return true;
 }
 
 }// namespace filament::backend

filament/backend/src/vulkan/VulkanImageUtility.h

@@ -76,6 +76,36 @@ inline VkImageViewType getViewType(SamplerType target) {
     }
 }
 
+inline VulkanLayout getDefaultLayout(TextureUsage usage) {
+    if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
+        if (any(usage & TextureUsage::SAMPLEABLE)) {
+            return VulkanLayout::DEPTH_SAMPLER;
+        } else {
+            return VulkanLayout::DEPTH_ATTACHMENT;
+        }
+    }
+
+    if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
+        return VulkanLayout::COLOR_ATTACHMENT;
+    }
+    // Finally, the layout for an immutable texture is optimal read-only.
+    return VulkanLayout::READ_ONLY;
+}
+
+inline VulkanLayout getDefaultLayout(VkImageUsageFlags vkusage) {
+    TextureUsage usage{};
+    if (vkusage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+        usage = usage | TextureUsage::DEPTH_ATTACHMENT;
+    }
+    if (vkusage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+        usage = usage | TextureUsage::COLOR_ATTACHMENT;
+    }
+    if (vkusage & VK_IMAGE_USAGE_SAMPLED_BIT) {
+        usage = usage | TextureUsage::SAMPLEABLE;
+    }
+    return getDefaultLayout(usage);
+}
+
 constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
     switch (layout) {
         case VulkanLayout::UNDEFINED:
@@ -105,9 +135,7 @@ constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
     }
 }
 
-// Returns true if a transition has been added to the command buffer, false otherwis (where there is
-// no transition necessary).
-bool transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
+void transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
 
 } // namespace imgutil
 

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -98,7 +98,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     VkPipelineColorBlendStateCreateInfo colorBlendState;
     colorBlendState = VkPipelineColorBlendStateCreateInfo{};
     colorBlendState.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO;
-    colorBlendState.attachmentCount = mPipelineRequirements.rasterState.colorTargetCount;    
+    colorBlendState.attachmentCount = 1;
     colorBlendState.pAttachments = colorBlendAttachments;
 
     // If we reach this point, we need to create and stash a brand new pipeline object.
@@ -184,7 +184,6 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     vkRaster.polygonMode = VK_POLYGON_MODE_FILL;
     vkRaster.cullMode = raster.cullMode;
     vkRaster.frontFace = raster.frontFace;
-    vkRaster.depthClampEnable = raster.depthClamp;
     vkRaster.depthBiasEnable = raster.depthBiasEnable;
     vkRaster.depthBiasConstantFactor = raster.depthBiasConstantFactor;
     vkRaster.depthBiasClamp = 0.0f;
@@ -210,8 +209,8 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     pipelineCreateInfo.pDynamicState = &dynamicState;
 
     // Filament assumes consistent blend state across all color attachments.
-    for (uint8_t i = 0; i < colorBlendState.attachmentCount; ++i) {
-        auto& target = colorBlendAttachments[i];
+    colorBlendState.attachmentCount = mPipelineRequirements.rasterState.colorTargetCount;
+    for (auto& target : colorBlendAttachments) {
         target.blendEnable = mPipelineRequirements.rasterState.blendEnable;
         target.srcColorBlendFactor = mPipelineRequirements.rasterState.srcColorBlendFactor;
         target.dstColorBlendFactor = mPipelineRequirements.rasterState.dstColorBlendFactor;

filament/backend/src/vulkan/VulkanPipelineCache.h

@@ -90,8 +90,7 @@ public:
         VkBlendFactor         srcAlphaBlendFactor : 5;
         VkBlendFactor         dstAlphaBlendFactor : 5;
         VkColorComponentFlags colorWriteMask : 4;
-        uint8_t               rasterizationSamples : 4;// offset = 4 bytes
-        uint8_t               depthClamp : 4;
+        uint8_t               rasterizationSamples;    // offset = 4 bytes
         uint8_t               colorTargetCount;        // offset = 5 bytes
         BlendEquation         colorBlendOp : 4;        // offset = 6 bytes
         BlendEquation         alphaBlendOp : 4;

filament/backend/src/vulkan/VulkanReadPixels.cpp

@@ -268,7 +268,7 @@ void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint
             imgutil::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);
 
     // Restore the source image layout.
-    srcTexture->transitionLayout(cmdbuffer, srcRange, srcTexture->getDefaultLayout());
+    srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::COLOR_ATTACHMENT);
 
     vkEndCommandBuffer(cmdbuffer);
 

filament/backend/src/vulkan/VulkanResourceAllocator.h

@@ -105,10 +105,6 @@ public:
         mHandleAllocatorImpl.deallocate(handle, obj);
     }
 
-    inline void associateHandle(HandleBase::HandleId id, utils::CString&& tag) noexcept {
-        mHandleAllocatorImpl.associateTagToHandle(id, std::move(tag));
-    }
-
 private:
     AllocatorImpl mHandleAllocatorImpl;
 

filament/backend/src/vulkan/VulkanSwapChain.cpp

@@ -26,14 +26,12 @@ using namespace utils;
 namespace filament::backend {
 
 VulkanSwapChain::VulkanSwapChain(VulkanPlatform* platform, VulkanContext const& context,
-        VmaAllocator allocator, VulkanCommands* commands, VulkanResourceAllocator* handleAllocator,
-        VulkanStagePool& stagePool,
+        VmaAllocator allocator, VulkanCommands* commands, VulkanStagePool& stagePool,
         void* nativeWindow, uint64_t flags, VkExtent2D extent)
     : VulkanResource(VulkanResourceType::SWAP_CHAIN),
       mPlatform(platform),
       mCommands(commands),
       mAllocator(allocator),
-      mHandleAllocator(handleAllocator),
       mStagePool(stagePool),
       mHeadless(extent.width != 0 && extent.height != 0 && !nativeWindow),
       mFlushAndWaitOnResize(platform->getCustomization().flushAndWaitOnWindowResize),
@@ -64,12 +62,12 @@ void VulkanSwapChain::update() {
     VkDevice const device = mPlatform->getDevice();
 
     for (auto const color: bundle.colors) {
-        mColors.push_back(std::make_unique<VulkanTexture>(device, mAllocator, mCommands, mHandleAllocator,
-                color, bundle.colorFormat, 1, bundle.extent.width, bundle.extent.height,
+        mColors.push_back(std::make_unique<VulkanTexture>(device, mAllocator, mCommands, color,
+                bundle.colorFormat, 1, bundle.extent.width, bundle.extent.height,
                 TextureUsage::COLOR_ATTACHMENT, mStagePool, true /* heap allocated */));
     }
-    mDepth = std::make_unique<VulkanTexture>(device, mAllocator, mCommands, mHandleAllocator,
-            bundle.depth, bundle.depthFormat, 1, bundle.extent.width, bundle.extent.height,
+    mDepth = std::make_unique<VulkanTexture>(device, mAllocator, mCommands, bundle.depth,
+            bundle.depthFormat, 1, bundle.extent.width, bundle.extent.height,
             TextureUsage::DEPTH_ATTACHMENT, mStagePool, true /* heap allocated */);
 
     mExtent = bundle.extent;
@@ -77,7 +75,7 @@ void VulkanSwapChain::update() {
 
 void VulkanSwapChain::present() {
     if (!mHeadless && mTransitionSwapChainImageLayoutForPresent) {
-        VulkanCommandBuffer& commands = mCommands->get();
+        VkCommandBuffer const cmdbuf = mCommands->get().buffer();
         VkImageSubresourceRange const subresources{
                 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
                 .baseMipLevel = 0,
@@ -85,7 +83,7 @@ void VulkanSwapChain::present() {
                 .baseArrayLayer = 0,
                 .layerCount = 1,
         };
-        mColors[mCurrentSwapIndex]->transitionLayout(&commands, subresources, VulkanLayout::PRESENT);
+        mColors[mCurrentSwapIndex]->transitionLayout(cmdbuf, subresources, VulkanLayout::PRESENT);
     }
 
     mCommands->flush();

filament/backend/src/vulkan/VulkanSwapChain.h

@@ -36,13 +36,11 @@ namespace filament::backend {
 
 struct VulkanHeadlessSwapChain;
 struct VulkanSurfaceSwapChain;
-class VulkanResourceAllocator;
 
 // A wrapper around the platform implementation of swapchain.
 struct VulkanSwapChain : public HwSwapChain, VulkanResource {
     VulkanSwapChain(VulkanPlatform* platform, VulkanContext const& context, VmaAllocator allocator,
-            VulkanCommands* commands, VulkanResourceAllocator* handleAllocator,
-            VulkanStagePool& stagePool,
+            VulkanCommands* commands, VulkanStagePool& stagePool,
             void* nativeWindow, uint64_t flags, VkExtent2D extent = {0, 0});
 
     ~VulkanSwapChain();
@@ -82,7 +80,6 @@ private:
     VulkanPlatform* mPlatform;
     VulkanCommands* mCommands;
     VmaAllocator mAllocator;
-    VulkanResourceAllocator* const mHandleAllocator;
     VulkanStagePool& mStagePool;
     bool const mHeadless;
     bool const mFlushAndWaitOnResize;

filament/backend/src/vulkan/VulkanTexture.cpp

@@ -15,7 +15,6 @@
  */
 
 #include "VulkanMemory.h"
-#include "VulkanResourceAllocator.h"
 #include "VulkanTexture.h"
 #include "VulkanUtility.h"
 
@@ -29,185 +28,50 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-namespace {
-
-inline uint8_t getLayerCount(SamplerType const target, uint32_t const depth) {
-    switch (target) {
-        case SamplerType::SAMPLER_2D:
-        case SamplerType::SAMPLER_3D:
-        case SamplerType::SAMPLER_EXTERNAL:
-            return 1;
-        case SamplerType::SAMPLER_CUBEMAP:
-            return 6;
-        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
-            return depth * 6;
-        case SamplerType::SAMPLER_2D_ARRAY:
-            return depth;
-    }
-}
-
-VkComponentMapping composeSwizzle(VkComponentMapping const& prev, VkComponentMapping const& next) {
-    static constexpr VkComponentSwizzle IDENTITY[] = {
-        VK_COMPONENT_SWIZZLE_R,
-        VK_COMPONENT_SWIZZLE_G,
-        VK_COMPONENT_SWIZZLE_B,
-        VK_COMPONENT_SWIZZLE_A,
-    };
-
-    auto const compose = [](VkComponentSwizzle out, VkComponentMapping const& prev,
-                                 uint8_t channelIndex) {
-        // We need to first change all identities to its equivalent channel.
-        if (out == VK_COMPONENT_SWIZZLE_IDENTITY) {
-            out = IDENTITY[channelIndex];
-        }
-        switch (out) {
-            case VK_COMPONENT_SWIZZLE_R:
-                out = prev.r;
-                break;
-            case VK_COMPONENT_SWIZZLE_G:
-                out = prev.g;
-                break;
-            case VK_COMPONENT_SWIZZLE_B:
-                out = prev.b;
-                break;
-            case VK_COMPONENT_SWIZZLE_A:
-                out = prev.a;
-                break;
-            case VK_COMPONENT_SWIZZLE_IDENTITY:
-            case VK_COMPONENT_SWIZZLE_ZERO:
-            case VK_COMPONENT_SWIZZLE_ONE:
-                return out;
-            // Below is not exposed in Vulkan's API, but needs to be there for compilation.
-            case VK_COMPONENT_SWIZZLE_MAX_ENUM:
-                break;
-        }
-        // If the result correctly corresponds to the identity, just return identity.
-        if (IDENTITY[channelIndex] == out) {
-            return VK_COMPONENT_SWIZZLE_IDENTITY;
-        }
-        return out;
-    };
-
-    auto const identityToChannel = [](VkComponentSwizzle val, uint8_t channelIndex) {
-        if (val != VK_COMPONENT_SWIZZLE_IDENTITY) {
-            return val;
-        }
-        return IDENTITY[channelIndex];
-    };
-
-    // We make sure all all identities are mapped into respective channels so that actual channel
-    // mapping will be passed onto the output.
-    VkComponentMapping const prevExplicit = {
-            identityToChannel(prev.r, 0),
-            identityToChannel(prev.g, 1),
-            identityToChannel(prev.b, 2),
-            identityToChannel(prev.a, 3),
-    };
-
-    // Note that the channel index corresponds to the VkComponentMapping struct layout.
-    return {
-        compose(next.r, prevExplicit, 0),
-        compose(next.g, prevExplicit, 1),
-        compose(next.b, prevExplicit, 2),
-        compose(next.a, prevExplicit, 3),
-    };
-}
-
-inline VulkanLayout getDefaultLayoutImpl(TextureUsage usage) {
-    if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
-        if (any(usage & TextureUsage::SAMPLEABLE)) {
-            return VulkanLayout::DEPTH_SAMPLER;
-        } else {
-            return VulkanLayout::DEPTH_ATTACHMENT;
-        }
-    }
-
-    if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
-        return VulkanLayout::COLOR_ATTACHMENT;
-    }
-    // Finally, the layout for an immutable texture is optimal read-only.
-    return VulkanLayout::READ_ONLY;
-}
-
-inline VulkanLayout getDefaultLayoutImpl(VkImageUsageFlags vkusage) {
-    TextureUsage usage{};
-    if (vkusage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
-        usage = usage | TextureUsage::DEPTH_ATTACHMENT;
-    }
-    if (vkusage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
-        usage = usage | TextureUsage::COLOR_ATTACHMENT;
-    }
-    if (vkusage & VK_IMAGE_USAGE_SAMPLED_BIT) {
-        usage = usage | TextureUsage::SAMPLEABLE;
-    }
-    return getDefaultLayoutImpl(usage);
-}
-
-} // anonymous namespace
-
-VulkanTextureState::VulkanTextureState(VkDevice device, VmaAllocator allocator,
-        VulkanCommands* commands, VulkanStagePool& stagePool, VkFormat format,
-        VkImageViewType viewType, uint8_t levels, uint8_t layerCount, VulkanLayout defaultLayout)
-    : VulkanResource(VulkanResourceType::HEAP_ALLOCATED),
+VulkanTexture::VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
+        VkImage image, VkFormat format, uint8_t samples, uint32_t width, uint32_t height,
+        TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
+    : HwTexture(SamplerType::SAMPLER_2D, 1, samples, width, height, 1, TextureFormat::UNUSED,
+            tusage),
+      VulkanResource(
+              heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
       mVkFormat(format),
-      mViewType(viewType),
-      mFullViewRange{filament::backend::getImageAspect(format), 0, levels, 0, layerCount},
-      mDefaultLayout(defaultLayout),
+      mViewType(imgutil::getViewType(target)),
+      mSwizzle({}),
+      mTextureImage(image),
+      mFullViewRange{
+              .aspectMask = getImageAspect(),
+              .baseMipLevel = 0,
+              .levelCount = 1,
+              .baseArrayLayer = 0,
+              .layerCount = 1,
+      },
+      mPrimaryViewRange(mFullViewRange),
       mStagePool(stagePool),
       mDevice(device),
       mAllocator(allocator),
-      mCommands(commands),
-      mIsTransientAttachment(false) {}
+      mCommands(commands) {}
 
-VulkanTextureState* VulkanTexture::getSharedState() {
-    VulkanTextureState* state = mAllocator->handle_cast<VulkanTextureState*>(mState);
-    return state;
-}
-
-VulkanTextureState const* VulkanTexture::getSharedState() const {
-    VulkanTextureState const* state = mAllocator->handle_cast<VulkanTextureState const*>(mState);
-    return state;
-}
-
-// Constructor for internally passed VkImage
-VulkanTexture::VulkanTexture(
-        VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
-        VulkanResourceAllocator* handleAllocator,
-        VkImage image, VkFormat format, uint8_t samples, uint32_t width, uint32_t height,
-        TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
-        : HwTexture(SamplerType::SAMPLER_2D, 1, samples, width, height, 1, TextureFormat::UNUSED,
-                tusage),
-          VulkanResource(
-                  heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
-          mAllocator(handleAllocator),
-          mState(handleAllocator->initHandle<VulkanTextureState>(
-                  device, allocator, commands, stagePool,
-                  format, imgutil::getViewType(SamplerType::SAMPLER_2D), 1, 1,
-                  getDefaultLayoutImpl(tusage))) {
-    auto* const state = getSharedState();
-    state->mTextureImage = image;
-    mPrimaryViewRange = state->mFullViewRange;
-}
-
-// Constructor for user facing texture
 VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
-        VulkanResourceAllocator* handleAllocator, SamplerType target, uint8_t levels,
-        TextureFormat tformat, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
-        TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
+        SamplerType target, uint8_t levels, TextureFormat tformat, uint8_t samples, uint32_t w,
+        uint32_t h, uint32_t depth, TextureUsage tusage, VulkanStagePool& stagePool,
+        bool heapAllocated, VkComponentMapping swizzle)
     : HwTexture(target, levels, samples, w, h, depth, tformat, tusage),
       VulkanResource(
               heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
-      mAllocator(handleAllocator),
-      mState(handleAllocator->initHandle<VulkanTextureState>(device, allocator, commands, stagePool,
-              backend::getVkFormat(tformat), imgutil::getViewType(target), levels,
-              getLayerCount(target, depth), VulkanLayout::UNDEFINED)) {
-    auto* const state = getSharedState();
+      mVkFormat(backend::getVkFormat(tformat)),
+      mViewType(imgutil::getViewType(target)),
+      mSwizzle(swizzle),
+      mStagePool(stagePool),
+      mDevice(device),
+      mAllocator(allocator),
+      mCommands(commands) {
 
     // Create an appropriately-sized device-only VkImage, but do not fill it yet.
     VkImageCreateInfo imageInfo{.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
             .imageType = target == SamplerType::SAMPLER_3D ? VK_IMAGE_TYPE_3D : VK_IMAGE_TYPE_2D,
-            .format = state->mVkFormat,
+            .format = mVkFormat,
             .extent = {w, h, depth},
             .mipLevels = levels,
             .arrayLayers = 1,
@@ -233,34 +97,23 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         imageInfo.extent.depth = 1;
     }
 
-    if (any(usage & TextureUsage::BLIT_SRC)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
-    }
-    if (any(usage & TextureUsage::BLIT_DST)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
-    }
+    // Filament expects blit() to work with any texture, so we almost always set these usage flags.
+    // TODO: investigate performance implications of setting these flags.
+    constexpr VkImageUsageFlags blittable = VK_IMAGE_USAGE_TRANSFER_DST_BIT |
+            VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
 
-    // Determine if we can use the transient usage flag combined with lazily allocated memory.
-    const bool useTransientAttachment =
-        // Lazily allocated memory is available.
-        context.isLazilyAllocatedMemorySupported() &&
-        // Usage consists of attachment flags only.
-        none(tusage & ~TextureUsage::ALL_ATTACHMENTS) &&
-        // Usage contains at least one attachment flag.
-        any(tusage & TextureUsage::ALL_ATTACHMENTS);
-    state->mIsTransientAttachment = useTransientAttachment;
-
-    const VkImageUsageFlags transientFlag =
-       useTransientAttachment ? VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT : 0U;
+    if (any(usage & (TextureUsage::BLIT_DST | TextureUsage::BLIT_SRC))) {
+        imageInfo.usage |= blittable;
+    }
 
     if (any(usage & TextureUsage::SAMPLEABLE)) {
 
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
         // Validate that the format is actually sampleable.
         VkFormatProperties props;
-        vkGetPhysicalDeviceFormatProperties(physicalDevice, state->mVkFormat, &props);
+        vkGetPhysicalDeviceFormatProperties(physicalDevice, mVkFormat, &props);
         if (!(props.optimalTilingFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT)) {
-            FVK_LOGW << "Texture usage is SAMPLEABLE but format " << state->mVkFormat << " is not "
+            FVK_LOGW << "Texture usage is SAMPLEABLE but format " << mVkFormat << " is not "
                     "sampleable with optimal tiling." << utils::io::endl;
         }
 #endif
@@ -268,19 +121,20 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         imageInfo.usage |= VK_IMAGE_USAGE_SAMPLED_BIT;
     }
     if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | transientFlag;
+        imageInfo.usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | blittable;
         if (any(usage & TextureUsage::SUBPASS_INPUT)) {
             imageInfo.usage |= VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT;
         }
     }
     if (any(usage & TextureUsage::STENCIL_ATTACHMENT)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT | transientFlag;
+        imageInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
     }
     if (any(usage & TextureUsage::UPLOADABLE)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
+        imageInfo.usage |= blittable;
     }
     if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT | transientFlag;
+        imageInfo.usage |= blittable;
+        imageInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
 
         // Depth resolves uses a custom shader and therefore needs to be sampleable.
         if (samples > 1) {
@@ -293,7 +147,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
     // any kind of attachment (color or depth).
     const auto& limits = context.getPhysicalDeviceLimits();
     if (imageInfo.usage & VK_IMAGE_USAGE_SAMPLED_BIT) {
-        samples = reduceSampleCount(samples, isVkDepthFormat(state->mVkFormat)
+        samples = reduceSampleCount(samples, isVkDepthFormat(mVkFormat)
                                                      ? limits.sampledImageDepthSampleCounts
                                                      : limits.sampledImageColorSampleCounts);
     }
@@ -307,12 +161,12 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
     this->samples = samples;
     imageInfo.samples = (VkSampleCountFlagBits) samples;
 
-    VkResult error = vkCreateImage(state->mDevice, &imageInfo, VKALLOC, &state->mTextureImage);
+    VkResult error = vkCreateImage(mDevice, &imageInfo, VKALLOC, &mTextureImage);
     if (error || FVK_ENABLED(FVK_DEBUG_TEXTURE)) {
         FVK_LOGD << "vkCreateImage: "
-            << "image = " << state->mTextureImage << ", "
+            << "image = " << mTextureImage << ", "
             << "result = " << error << ", "
-            << "handle = " << utils::io::hex << state->mTextureImage << utils::io::dec << ", "
+            << "handle = " << utils::io::hex << mTextureImage << utils::io::dec << ", "
             << "extent = " << w << "x" << h << "x"<< depth << ", "
             << "mipLevels = " << int(levels) << ", "
             << "TextureUsage = " << static_cast<int>(usage) << ", "
@@ -321,19 +175,16 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
             << "type = " << imageInfo.imageType << ", "
             << "flags = " << imageInfo.flags << ", "
             << "target = " << static_cast<int>(target) <<", "
-            << "format = " << state->mVkFormat << utils::io::endl;
+            << "format = " << mVkFormat << utils::io::endl;
     }
     FILAMENT_CHECK_POSTCONDITION(!error) << "Unable to create image.";
 
     // Allocate memory for the VkImage and bind it.
     VkMemoryRequirements memReqs = {};
-    vkGetImageMemoryRequirements(state->mDevice, state->mTextureImage, &memReqs);
+    vkGetImageMemoryRequirements(mDevice, mTextureImage, &memReqs);
 
-    const VkFlags requiredMemoryFlags =
-        VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
-        (useTransientAttachment ? VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT : 0U);
     uint32_t memoryTypeIndex
-            = context.selectMemoryType(memReqs.memoryTypeBits, requiredMemoryFlags);
+            = context.selectMemoryType(memReqs.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
 
     FILAMENT_CHECK_POSTCONDITION(memoryTypeIndex < VK_MAX_MEMORY_TYPES)
             << "VulkanTexture: unable to find a memory type that meets requirements.";
@@ -343,70 +194,58 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         .allocationSize = memReqs.size,
         .memoryTypeIndex = memoryTypeIndex,
     };
-    error = vkAllocateMemory(state->mDevice, &allocInfo, nullptr, &state->mTextureImageMemory);
+    error = vkAllocateMemory(mDevice, &allocInfo, nullptr, &mTextureImageMemory);
     FILAMENT_CHECK_POSTCONDITION(!error) << "Unable to allocate image memory.";
-    error = vkBindImageMemory(state->mDevice, state->mTextureImage, state->mTextureImageMemory, 0);
+    error = vkBindImageMemory(mDevice, mTextureImage, mTextureImageMemory, 0);
     FILAMENT_CHECK_POSTCONDITION(!error) << "Unable to bind image.";
 
+    uint32_t layerCount = 0;
+    if (target == SamplerType::SAMPLER_CUBEMAP) {
+        layerCount = 6;
+    } else if (target == SamplerType::SAMPLER_CUBEMAP_ARRAY) {
+        layerCount = depth * 6;
+    } else if (target == SamplerType::SAMPLER_2D_ARRAY) {
+        layerCount = depth;
+    } else if (target == SamplerType::SAMPLER_3D) {
+        layerCount = 1;
+    } else {
+        layerCount = 1;
+    }
+
+    mFullViewRange = {
+        .aspectMask = getImageAspect(),
+        .baseMipLevel = 0,
+        .levelCount = levels,
+        .baseArrayLayer = 0,
+        .layerCount = layerCount,
+    };
+
     // Spec out the "primary" VkImageView that shaders use to sample from the image.
-    mPrimaryViewRange = state->mFullViewRange;
+    mPrimaryViewRange = mFullViewRange;
 
     // Go ahead and create the primary image view.
-    getImageView(mPrimaryViewRange, state->mViewType, mSwizzle);
+    getImageView(mPrimaryViewRange, mViewType, mSwizzle);
 
-    VulkanCommandBuffer& commandsBuf = state->mCommands->get();
-    commandsBuf.acquire(this);
-
-    auto const defaultLayout = state->mDefaultLayout = getDefaultLayoutImpl(imageInfo.usage);
-    transitionLayout(&commandsBuf, mPrimaryViewRange, defaultLayout);
-}
-
-// Constructor for creating a texture view
-VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
-        VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
-        VulkanResourceAllocator* handleAllocator, VulkanTexture const* src, uint8_t baseLevel,
-        uint8_t levelCount)
-    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
-              src->format, src->usage),
-      VulkanResource(VulkanResourceType::TEXTURE),
-      mAllocator(handleAllocator) {
-    mState = src->mState;
-    auto* state = getSharedState();
-
-    state->refs++;
-    mPrimaryViewRange = src->mPrimaryViewRange;
-    mPrimaryViewRange.baseMipLevel = src->mPrimaryViewRange.baseMipLevel + baseLevel;
-    mPrimaryViewRange.levelCount = levelCount;
-}
-
-// Constructor for creating a texture view with swizzle
-VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
-        VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
-        VulkanResourceAllocator* handleAllocator, VulkanTexture const* src,
-        VkComponentMapping swizzle)
-    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
-              src->format, src->usage),
-      VulkanResource(VulkanResourceType::TEXTURE),
-      mAllocator(handleAllocator) {
-    mState = src->mState;
-    auto* state = getSharedState();
-    state->refs++;
-    mPrimaryViewRange = src->mPrimaryViewRange;
-    mSwizzle = composeSwizzle(src->mSwizzle, swizzle);
+    // Transition the layout of each image slice that might be used as a render target.
+    // We do not transition images that are merely SAMPLEABLE, this is deferred until upload time
+    // because we do not know how many layers and levels will actually be used.
+    if (imageInfo.usage
+        & (VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
+           | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) {
+        VulkanCommandBuffer& commands = mCommands->get();
+        VkCommandBuffer const cmdbuf = commands.buffer();
+        commands.acquire(this);
+        transitionLayout(cmdbuf, mFullViewRange, imgutil::getDefaultLayout(imageInfo.usage));
+    }
 }
 
 VulkanTexture::~VulkanTexture() {
-    auto* const state = getSharedState();
-    state->refs--;
-    if (state->refs == 0) {
-        if (state->mTextureImageMemory != VK_NULL_HANDLE) {
-            vkDestroyImage(state->mDevice, state->mTextureImage, VKALLOC);
-            vkFreeMemory(state->mDevice, state->mTextureImageMemory, VKALLOC);
-        }
-        for (auto entry: state->mCachedImageViews) {
-            vkDestroyImageView(state->mDevice, entry.second, VKALLOC);
-        }
-        mAllocator->destruct<VulkanTextureState>(mState);
+    if (mTextureImageMemory != VK_NULL_HANDLE) {
+        vkDestroyImage(mDevice, mTextureImage, VKALLOC);
+        vkFreeMemory(mDevice, mTextureImageMemory, VKALLOC);
+    }
+    for (auto entry : mCachedImageViews) {
+        vkDestroyImageView(mDevice, entry.second, VKALLOC);
     }
 }
 
@@ -415,7 +254,7 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
     assert_invariant(width <= this->width && height <= this->height);
     assert_invariant(depth <= this->depth * ((target == SamplerType::SAMPLER_CUBEMAP ||
                         target == SamplerType::SAMPLER_CUBEMAP_ARRAY) ? 6 : 1));
-    auto* const state = getSharedState();
+
     const PixelBufferDescriptor* hostData = &data;
     PixelBufferDescriptor reshapedData;
 
@@ -428,7 +267,7 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
 
     // If format conversion is both required and supported, use vkCmdBlitImage.
     const VkFormat hostFormat = backend::getVkFormat(hostData->format, hostData->type);
-    const VkFormat deviceFormat = getVkFormatLinear(state->mVkFormat);
+    const VkFormat deviceFormat = getVkFormatLinear(mVkFormat);
     if (hostFormat != deviceFormat && hostFormat != VK_FORMAT_UNDEFINED) {
         assert_invariant(xoffset == 0 && yoffset == 0 && zoffset == 0 &&
                 "Offsets not yet supported when format conversion is required.");
@@ -440,14 +279,14 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
 
     // Otherwise, use vkCmdCopyBufferToImage.
     void* mapped = nullptr;
-    VulkanStage const* stage = state->mStagePool.acquireStage(hostData->size);
+    VulkanStage const* stage = mStagePool.acquireStage(hostData->size);
     assert_invariant(stage->memory);
-    vmaMapMemory(state->mAllocator, stage->memory, &mapped);
+    vmaMapMemory(mAllocator, stage->memory, &mapped);
     memcpy(mapped, hostData->buffer, hostData->size);
-    vmaUnmapMemory(state->mAllocator, stage->memory);
-    vmaFlushAllocation(state->mAllocator, stage->memory, 0, hostData->size);
+    vmaUnmapMemory(mAllocator, stage->memory);
+    vmaFlushAllocation(mAllocator, stage->memory, 0, hostData->size);
 
-    VulkanCommandBuffer& commands = state->mCommands->get();
+    VulkanCommandBuffer& commands = mCommands->get();
     VkCommandBuffer const cmdbuf = commands.buffer();
     commands.acquire(this);
 
@@ -490,28 +329,27 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
     VkImageLayout const newVkLayout = imgutil::getVkLayout(newLayout);
 
     if (nextLayout == VulkanLayout::UNDEFINED) {
-        nextLayout = getDefaultLayout();
+        nextLayout = imgutil::getDefaultLayout(this->usage);
     }
 
-    transitionLayout(&commands, transitionRange, newLayout);
+    transitionLayout(cmdbuf, transitionRange, newLayout);
 
-    vkCmdCopyBufferToImage(cmdbuf, stage->buffer, state->mTextureImage, newVkLayout, 1, &copyRegion);
+    vkCmdCopyBufferToImage(cmdbuf, stage->buffer, mTextureImage, newVkLayout, 1, &copyRegion);
 
-    transitionLayout(&commands, transitionRange, nextLayout);
+    transitionLayout(cmdbuf, transitionRange, nextLayout);
 }
 
 void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, uint32_t width,
         uint32_t height, uint32_t depth, uint32_t miplevel) {
-    auto* const state = getSharedState();
     void* mapped = nullptr;
     VulkanStageImage const* stage
-            = state->mStagePool.acquireImage(hostData.format, hostData.type, width, height);
-    vmaMapMemory(state->mAllocator, stage->memory, &mapped);
+            = mStagePool.acquireImage(hostData.format, hostData.type, width, height);
+    vmaMapMemory(mAllocator, stage->memory, &mapped);
     memcpy(mapped, hostData.buffer, hostData.size);
-    vmaUnmapMemory(state->mAllocator, stage->memory);
-    vmaFlushAllocation(state->mAllocator, stage->memory, 0, hostData.size);
+    vmaUnmapMemory(mAllocator, stage->memory);
+    vmaFlushAllocation(mAllocator, stage->memory, 0, hostData.size);
 
-    VulkanCommandBuffer& commands = state->mCommands->get();
+    VulkanCommandBuffer& commands = mCommands->get();
     VkCommandBuffer const cmdbuf = commands.buffer();
     commands.acquire(this);
 
@@ -533,71 +371,63 @@ void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, u
 
     VulkanLayout const newLayout = VulkanLayout::TRANSFER_DST;
     VulkanLayout const oldLayout = getLayout(layer, miplevel);
-    transitionLayout(&commands, range, newLayout);
+    transitionLayout(cmdbuf, range, newLayout);
 
     vkCmdBlitImage(cmdbuf, stage->image, imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            state->mTextureImage, imgutil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
+            mTextureImage, imgutil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
 
-    transitionLayout(&commands, range, oldLayout);
+    transitionLayout(cmdbuf, range, oldLayout);
 }
 
-VulkanLayout VulkanTexture::getDefaultLayout() const {
-    auto* const state = getSharedState();
-    return state->mDefaultLayout;
+void VulkanTexture::setPrimaryRange(uint32_t minMiplevel, uint32_t maxMiplevel) {
+    maxMiplevel = filament::math::min(int(maxMiplevel), int(this->levels - 1));
+    mPrimaryViewRange.baseMipLevel = minMiplevel;
+    mPrimaryViewRange.levelCount = maxMiplevel - minMiplevel + 1;
+    getImageView(mPrimaryViewRange, mViewType, mSwizzle);
 }
 
 VkImageView VulkanTexture::getAttachmentView(VkImageSubresourceRange range) {
+    // Attachments should only have one mipmap level and one layer.
     range.levelCount = 1;
     range.layerCount = 1;
     return getImageView(range, VK_IMAGE_VIEW_TYPE_2D, {});
 }
 
-VkImageView VulkanTexture::getMultiviewAttachmentView(VkImageSubresourceRange range) {
-    return getImageView(range, VK_IMAGE_VIEW_TYPE_2D_ARRAY, {});
-}
-
 VkImageView VulkanTexture::getViewForType(VkImageSubresourceRange const& range, VkImageViewType type) {
     return getImageView(range, type, mSwizzle);
 }
 
 VkImageView VulkanTexture::getImageView(VkImageSubresourceRange range, VkImageViewType viewType,
         VkComponentMapping swizzle) {
-    auto* const state = getSharedState();
-    VulkanTextureState::ImageViewKey const key{ range, viewType, swizzle };
-    auto iter = state->mCachedImageViews.find(key);
-    if (iter != state->mCachedImageViews.end()) {
+    ImageViewKey const key {range, viewType, swizzle};
+    auto iter = mCachedImageViews.find(key);
+    if (iter != mCachedImageViews.end()) {
         return iter->second;
     }
     VkImageViewCreateInfo viewInfo = {
         .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
-        .image = state->mTextureImage,
+        .image = mTextureImage,
         .viewType = viewType,
-        .format = state->mVkFormat,
+        .format = mVkFormat,
         .components = swizzle,
         .subresourceRange = range,
     };
     VkImageView imageView;
-    vkCreateImageView(state->mDevice, &viewInfo, VKALLOC, &imageView);
-    state->mCachedImageViews.emplace(key, imageView);
+    vkCreateImageView(mDevice, &viewInfo, VKALLOC, &imageView);
+    mCachedImageViews.emplace(key, imageView);
     return imageView;
 }
 
 VkImageAspectFlags VulkanTexture::getImageAspect() const {
     // Helper function in VulkanUtility
-    auto* const state = getSharedState();
-    return filament::backend::getImageAspect(state->mVkFormat);
+    return filament::backend::getImageAspect(mVkFormat);
 }
 
-bool VulkanTexture::transitionLayout(VulkanCommandBuffer* commands,
-        VkImageSubresourceRange const& range, VulkanLayout newLayout) {
-    return transitionLayout(commands->buffer(), range, newLayout);
-}
-
-bool VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, VkImageSubresourceRange const& range,
+void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubresourceRange& range,
         VulkanLayout newLayout) {
-    auto* const state = getSharedState();
+
     VulkanLayout const oldLayout = getLayout(range.baseArrayLayer, range.baseMipLevel);
 
     uint32_t const firstLayer = range.baseArrayLayer;
@@ -608,7 +438,7 @@ bool VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, VkImageSubresourceR
     // If we are transitioning more than one layer/level (slice), we need to know whether they are
     // all of the same layer.  If not, we need to transition slice-by-slice. Otherwise it would
     // trigger the validation layer saying that the `oldLayout` provided is incorrect.
-    // TODO: transition by multiple slices with more sophisticated range finding.
+    // TODO: transition by multiple slices with more sophiscated range finding.
     bool transitionSliceBySlice = false;
     for (uint32_t i = firstLayer; i < lastLayer; ++i) {
         for (uint32_t j = firstLevel; j < lastLevel; ++j) {
@@ -619,107 +449,50 @@ bool VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, VkImageSubresourceR
         }
     }
 
-    bool hasTransitions = false;
+#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
+    FVK_LOGD << "transition texture=" << mTextureImage
+                  << " (" << range.baseArrayLayer
+                  << "," << range.baseMipLevel << ")"
+                  << " count=(" << range.layerCount
+                  << "," << range.levelCount << ")"
+                  << " from=" << oldLayout << " to=" << newLayout
+                  << " format=" << mVkFormat
+                  << " depth=" << isVkDepthFormat(mVkFormat)
+                  << " slice-by-slice=" << transitionSliceBySlice
+                  << utils::io::endl;
+#endif
+
     if (transitionSliceBySlice) {
         for (uint32_t i = firstLayer; i < lastLayer; ++i) {
             for (uint32_t j = firstLevel; j < lastLevel; ++j) {
                 VulkanLayout const layout = getLayout(i, j);
-                if (layout == newLayout) {
-                    continue;
-                }
-                hasTransitions = hasTransitions || imgutil::transitionLayout(cmdbuf, {
-                    .image = state->mTextureImage,
-                    .oldLayout = layout,
-                    .newLayout = newLayout,
-                    .subresources = {
-                        .aspectMask = range.aspectMask,
-                        .baseMipLevel = j,
-                        .levelCount = 1,
-                        .baseArrayLayer = i,
-                        .layerCount = 1,
-                    },
-                });
+                imgutil::transitionLayout(cmdbuf, {
+                        .image = mTextureImage,
+                        .oldLayout = layout,
+                        .newLayout = newLayout,
+                        .subresources = {
+                            .aspectMask = range.aspectMask,
+                            .baseMipLevel = j,
+                            .levelCount = 1,
+                            .baseArrayLayer = i,
+                            .layerCount = 1,
+                        },
+                    });
             }
         }
-    } else if (newLayout != oldLayout) {
-        hasTransitions = imgutil::transitionLayout(cmdbuf, {
-            .image = state->mTextureImage,
+    } else {
+        imgutil::transitionLayout(cmdbuf, {
+            .image = mTextureImage,
             .oldLayout = oldLayout,
             .newLayout = newLayout,
             .subresources = range,
         });
     }
 
-    // Even if we didn't carry out the transition, we should assume that the new layout is defined
-    // through this call.
     setLayout(range, newLayout);
-
-    if (hasTransitions) {
-#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
-        FVK_LOGD << "transition texture=" << state->mTextureImage << " (" << range.baseArrayLayer
-                 << "," << range.baseMipLevel << ")" << " count=(" << range.layerCount << ","
-                 << range.levelCount << ")" << " from=" << oldLayout << " to=" << newLayout
-                 << " format=" << state->mVkFormat << " depth=" << isVkDepthFormat(state->mVkFormat)
-                 << " slice-by-slice=" << transitionSliceBySlice << utils::io::endl;
-#endif
-    } else {
-#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
-        FVK_LOGD << "transition texture=" << state->mTextureImage << " (" << range.baseArrayLayer
-                 << "," << range.baseMipLevel << ")" << " count=(" << range.layerCount << ","
-                 << range.levelCount << ")" << " to=" << newLayout
-                 << " is skipped because of no change in layout" << utils::io::endl;
-#endif
-    }
-    return hasTransitions;
 }
 
-void VulkanTexture::samplerToAttachmentBarrier(VulkanCommandBuffer* commands,
-        VkImageSubresourceRange const& range) {
-    VkCommandBuffer const cmdbuf = commands->buffer();
-    auto* const state = getSharedState();
-    VkImageLayout const layout =
-            imgutil::getVkLayout(getLayout(range.baseArrayLayer, range.baseMipLevel));
-    VkImageMemoryBarrier barrier = {
-            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-            .srcAccessMask = VK_ACCESS_SHADER_READ_BIT,
-            .dstAccessMask =
-                    VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-            .oldLayout = layout,
-            .newLayout = layout,
-            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .image = state->mTextureImage,
-            .subresourceRange = range,
-    };
-    vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
-            VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
-                    VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
-            0, 0, nullptr, 0, nullptr, 1, &barrier);
-}
-
-void VulkanTexture::attachmentToSamplerBarrier(VulkanCommandBuffer* commands,
-        VkImageSubresourceRange const& range) {
-    VkCommandBuffer const cmdbuf = commands->buffer();
-    auto* const state = getSharedState();
-    VkImageLayout const layout
-            = imgutil::getVkLayout(getLayout(range.baseArrayLayer, range.baseMipLevel));
-    VkImageMemoryBarrier barrier = {
-            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-            .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
-            .oldLayout = layout,
-            .newLayout = layout,
-            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .image = state->mTextureImage,
-            .subresourceRange = range,
-    };
-    vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
-            VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &barrier);
-}
-
-void VulkanTexture::setLayout(VkImageSubresourceRange const& range, VulkanLayout newLayout) {
-    auto* const state = getSharedState();
+void VulkanTexture::setLayout(const VkImageSubresourceRange& range, VulkanLayout newLayout) {
     uint32_t const firstLayer = range.baseArrayLayer;
     uint32_t const lastLayer = firstLayer + range.layerCount;
     uint32_t const firstLevel = range.baseMipLevel;
@@ -732,34 +505,32 @@ void VulkanTexture::setLayout(VkImageSubresourceRange const& range, VulkanLayout
         for (uint32_t layer = firstLayer; layer < lastLayer; ++layer) {
             uint32_t const first = (layer << 16) | firstLevel;
             uint32_t const last = (layer << 16) | lastLevel;
-            state->mSubresourceLayouts.clear(first, last);
+            mSubresourceLayouts.clear(first, last);
         }
     } else {
         for (uint32_t layer = firstLayer; layer < lastLayer; ++layer) {
             uint32_t const first = (layer << 16) | firstLevel;
             uint32_t const last = (layer << 16) | lastLevel;
-            state->mSubresourceLayouts.add(first, last, newLayout);
+            mSubresourceLayouts.add(first, last, newLayout);
         }
     }
 }
 
 VulkanLayout VulkanTexture::getLayout(uint32_t layer, uint32_t level) const {
     assert_invariant(level <= 0xffff && layer <= 0xffff);
-    auto* const state = getSharedState();
     const uint32_t key = (layer << 16) | level;
-    if (!state->mSubresourceLayouts.has(key)) {
+    if (!mSubresourceLayouts.has(key)) {
         return VulkanLayout::UNDEFINED;
     }
-    return state->mSubresourceLayouts.get(key);
+    return mSubresourceLayouts.get(key);
 }
 
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
 void VulkanTexture::print() const {
-    auto* const state = getSharedState();
     uint32_t const firstLayer = 0;
-    uint32_t const lastLayer = firstLayer + state->mFullViewRange.layerCount;
+    uint32_t const lastLayer = firstLayer + mFullViewRange.layerCount;
     uint32_t const firstLevel = 0;
-    uint32_t const lastLevel = firstLevel + state->mFullViewRange.levelCount;
+    uint32_t const lastLevel = firstLevel + mFullViewRange.levelCount;
 
     for (uint32_t layer = firstLayer; layer < lastLayer; ++layer) {
         for (uint32_t level = firstLevel; level < lastLevel; ++level) {
@@ -768,16 +539,16 @@ void VulkanTexture::print() const {
                 layer < (mPrimaryViewRange.baseArrayLayer + mPrimaryViewRange.layerCount) &&
                 level >= mPrimaryViewRange.baseMipLevel &&
                 level < (mPrimaryViewRange.baseMipLevel + mPrimaryViewRange.levelCount);
-            FVK_LOGD << "[" << state->mTextureImage << "]: (" << layer << "," << level
+            FVK_LOGD << "[" << mTextureImage << "]: (" << layer << "," << level
                           << ")=" << getLayout(layer, level)
                           << " primary=" << primary
                           << utils::io::endl;
         }
     }
 
-    for (auto view: state->mCachedImageViews) {
+    for (auto view: mCachedImageViews) {
         auto& range = view.first.range;
-        FVK_LOGD << "[" << state->mTextureImage << ", imageView=" << view.second << "]=>"
+        FVK_LOGD << "[" << mTextureImage << ", imageView=" << view.second << "]=>"
                       << " (" << range.baseArrayLayer << "," << range.baseMipLevel << ")"
                       << " count=(" << range.layerCount << "," << range.levelCount << ")"
                       << " aspect=" << range.aspectMask << " viewType=" << view.first.type

filament/backend/src/vulkan/VulkanTexture.h

@@ -30,12 +30,84 @@
 
 namespace filament::backend {
 
-class VulkanResourceAllocator;
+struct VulkanTexture : public HwTexture, VulkanResource {
+    // Standard constructor for user-facing textures.
+    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
+            VmaAllocator allocator, VulkanCommands* commands, SamplerType target, uint8_t levels,
+            TextureFormat tformat, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
+            TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated = false,
+            VkComponentMapping swizzle = {});
 
-struct VulkanTextureState : public VulkanResource {
-    VulkanTextureState(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
-            VulkanStagePool& stagePool, VkFormat format, VkImageViewType viewType, uint8_t levels,
-            uint8_t layerCount, VulkanLayout defaultLayout);
+    // Specialized constructor for internally created textures (e.g. from a swap chain)
+    // The texture will never destroy the given VkImage, but it does manages its subresources.
+    VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands, VkImage image,
+            VkFormat format, uint8_t samples, uint32_t width, uint32_t height, TextureUsage tusage,
+            VulkanStagePool& stagePool, bool heapAllocated = false);
+
+    ~VulkanTexture();
+
+    // Uploads data into a subregion of a 2D or 3D texture.
+    void updateImage(const PixelBufferDescriptor& data, uint32_t width, uint32_t height,
+            uint32_t depth, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset, uint32_t miplevel);
+
+    // Returns the primary image view, which is used for shader sampling.
+    VkImageView getPrimaryImageView() {
+        return getImageView(mPrimaryViewRange, mViewType, mSwizzle);
+    }
+
+    VkImageViewType getViewType() const { return mViewType; }
+
+    // Sets the min/max range of miplevels in the primary image view.
+    void setPrimaryRange(uint32_t minMiplevel, uint32_t maxMiplevel);
+
+    VkImageSubresourceRange getPrimaryViewRange() const { return mPrimaryViewRange; }
+
+    VkImageSubresourceRange getFullViewRange() const { return mFullViewRange; }
+
+    VulkanLayout getPrimaryImageLayout() const {
+        return getLayout(mPrimaryViewRange.baseArrayLayer, mPrimaryViewRange.baseMipLevel);
+    }
+
+    // Gets or creates a cached VkImageView for a single subresource that can be used as a render
+    // target attachment.  Unlike the primary image view, this always has type VK_IMAGE_VIEW_TYPE_2D
+    // and the identity swizzle.
+    VkImageView getAttachmentView(VkImageSubresourceRange range);
+
+    // This is a workaround for the first few frames where we're waiting for the texture to actually
+    // be uploaded.  In that case, we bind the sampler to an empty texture, but the corresponding
+    // imageView needs to be of the right type. Hence, we provide an option to indicate the
+    // view type. Swizzle option does not matter in this case.
+    VkImageView getViewForType(VkImageSubresourceRange const& range, VkImageViewType type);
+
+    VkFormat getVkFormat() const { return mVkFormat; }
+    VkImage getVkImage() const { return mTextureImage; }
+
+    VulkanLayout getLayout(uint32_t layer, uint32_t level) const;
+
+    void setSidecar(VulkanTexture* sidecar) {
+        mSidecarMSAA.reset(sidecar);
+    }
+
+    VulkanTexture* getSidecar() const {
+        return mSidecarMSAA.get();
+    }
+
+    void transitionLayout(VkCommandBuffer commands, const VkImageSubresourceRange& range,
+            VulkanLayout newLayout);
+
+    // Returns the preferred data plane of interest for all image views.
+    // For now this always returns either DEPTH or COLOR.
+    VkImageAspectFlags getImageAspect() const;
+
+    // For implicit transition like the end of a render pass, we need to be able to set the layout
+    // manually (outside of calls to transitionLayout).
+    void setLayout(const VkImageSubresourceRange& range, VulkanLayout newLayout);
+
+#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+    void print() const;
+#endif
+
+private:
 
     struct ImageViewKey {
         VkImageSubresourceRange range;  // 4 * 5 bytes
@@ -58,155 +130,6 @@ struct VulkanTextureState : public VulkanResource {
 
     using ImageViewHash = utils::hash::MurmurHashFn<ImageViewKey>;
 
-    uint32_t refs = 1;
-
-    // The texture with the sidecar owns the sidecar.
-    std::unique_ptr<VulkanTexture> mSidecarMSAA;
-    VkDeviceMemory mTextureImageMemory = VK_NULL_HANDLE;
-
-    VkFormat const mVkFormat;
-    VkImageViewType const mViewType;
-    VkImageSubresourceRange const mFullViewRange;
-    VkImage mTextureImage = VK_NULL_HANDLE;
-    VulkanLayout mDefaultLayout;
-
-    // Track the image layout of each subresource using a sparse range map.
-    utils::RangeMap<uint32_t, VulkanLayout> mSubresourceLayouts;
-
-    std::unordered_map<ImageViewKey, VkImageView, ImageViewHash> mCachedImageViews;
-    VulkanStagePool& mStagePool;
-    VkDevice mDevice;
-    VmaAllocator mAllocator;
-    VulkanCommands* mCommands;
-    bool mIsTransientAttachment;
-};
-
-
-struct VulkanTexture : public HwTexture, VulkanResource {
-    // Standard constructor for user-facing textures.
-    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
-            VmaAllocator allocator, VulkanCommands* commands,
-            VulkanResourceAllocator* handleAllocator,
-            SamplerType target, uint8_t levels,
-            TextureFormat tformat, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
-            TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated = false);
-
-    // Specialized constructor for internally created textures (e.g. from a swap chain)
-    // The texture will never destroy the given VkImage, but it does manages its subresources.
-    VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
-            VulkanResourceAllocator* handleAllocator,
-            VkImage image,
-            VkFormat format, uint8_t samples, uint32_t width, uint32_t height, TextureUsage tusage,
-            VulkanStagePool& stagePool, bool heapAllocated = false);
-
-    // Constructor for creating a texture view for wrt specific mip range
-    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
-            VmaAllocator allocator, VulkanCommands* commands,
-            VulkanResourceAllocator* handleAllocator,
-            VulkanTexture const* src, uint8_t baseLevel, uint8_t levelCount);
-
-    // Constructor for creating a texture view for swizzle.
-    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
-            VmaAllocator allocator, VulkanCommands* commands,
-            VulkanResourceAllocator* handleAllocator,
-            VulkanTexture const* src, VkComponentMapping swizzle);
-
-    ~VulkanTexture();
-
-    // Uploads data into a subregion of a 2D or 3D texture.
-    void updateImage(const PixelBufferDescriptor& data, uint32_t width, uint32_t height,
-            uint32_t depth, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset, uint32_t miplevel);
-
-    // Returns the primary image view, which is used for shader sampling.
-    VkImageView getPrimaryImageView() {
-        VulkanTextureState* state = getSharedState();
-        return getImageView(mPrimaryViewRange, state->mViewType, mSwizzle);
-    }
-
-    VkImageViewType getViewType() const {
-        VulkanTextureState const* state = getSharedState();
-        return state->mViewType;
-    }
-
-    VkImageSubresourceRange getPrimaryViewRange() const { return mPrimaryViewRange; }
-
-    VulkanLayout getPrimaryImageLayout() const {
-        return getLayout(mPrimaryViewRange.baseArrayLayer, mPrimaryViewRange.baseMipLevel);
-    }
-
-    // Returns the layout for the intended use of this texture (and not the expected layout at the
-    // time of the call.
-    VulkanLayout getDefaultLayout() const;
-
-    // Gets or creates a cached VkImageView for a single subresource that can be used as a render
-    // target attachment.  Unlike the primary image view, this always has type VK_IMAGE_VIEW_TYPE_2D
-    // and the identity swizzle.
-    VkImageView getAttachmentView(VkImageSubresourceRange range);
-
-    // Gets or creates a cached VkImageView for a single subresource that can be used as a render
-    // target attachment when rendering with multiview.
-    VkImageView getMultiviewAttachmentView(VkImageSubresourceRange range);
-
-    // This is a workaround for the first few frames where we're waiting for the texture to actually
-    // be uploaded.  In that case, we bind the sampler to an empty texture, but the corresponding
-    // imageView needs to be of the right type. Hence, we provide an option to indicate the
-    // view type. Swizzle option does not matter in this case.
-    VkImageView getViewForType(VkImageSubresourceRange const& range, VkImageViewType type);
-
-    VkFormat getVkFormat() const {
-        VulkanTextureState const* state = getSharedState();
-        return state->mVkFormat;
-    }
-    VkImage getVkImage() const {
-        VulkanTextureState const* state = getSharedState();
-        return state->mTextureImage;
-    }
-
-    VulkanLayout getLayout(uint32_t layer, uint32_t level) const;
-
-    void setSidecar(VulkanTexture* sidecar) {
-        VulkanTextureState* state = getSharedState();
-        state->mSidecarMSAA.reset(sidecar);
-    }
-
-    VulkanTexture* getSidecar() const {
-        VulkanTextureState const* state = getSharedState();
-        return state->mSidecarMSAA.get();
-    }
-
-    bool isTransientAttachment() const {
-        VulkanTextureState const* state = getSharedState();
-        return state->mIsTransientAttachment;
-    }
-
-    bool transitionLayout(VulkanCommandBuffer* commands, VkImageSubresourceRange const& range,
-            VulkanLayout newLayout);
-
-    bool transitionLayout(VkCommandBuffer cmdbuf, VkImageSubresourceRange const& range,
-            VulkanLayout newLayout);
-
-    void attachmentToSamplerBarrier(VulkanCommandBuffer* commands,
-            VkImageSubresourceRange const& range);
-
-    void samplerToAttachmentBarrier(VulkanCommandBuffer* commands,
-            VkImageSubresourceRange const& range);
-
-    // Returns the preferred data plane of interest for all image views.
-    // For now this always returns either DEPTH or COLOR.
-    VkImageAspectFlags getImageAspect() const;
-
-    // For implicit transition like the end of a render pass, we need to be able to set the layout
-    // manually (outside of calls to transitionLayout).
-    void setLayout(VkImageSubresourceRange const& range, VulkanLayout newLayout);
-
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
-    void print() const;
-#endif
-
-private:
-    VulkanTextureState* getSharedState();
-    VulkanTextureState const* getSharedState() const;
-
     // Gets or creates a cached VkImageView for a range of miplevels, array layers, viewType, and
     // swizzle (or not).
     VkImageView getImageView(VkImageSubresourceRange range, VkImageViewType viewType,
@@ -215,15 +138,28 @@ private:
     void updateImageWithBlit(const PixelBufferDescriptor& hostData, uint32_t width, uint32_t height,
             uint32_t depth, uint32_t miplevel);
 
-    VulkanResourceAllocator* const mAllocator;
+    // The texture with the sidecar owns the sidecar.
+    std::unique_ptr<VulkanTexture> mSidecarMSAA;
+    const VkFormat mVkFormat;
+    const VkImageViewType mViewType;
+    const VkComponentMapping mSwizzle;
+    VkImage mTextureImage = VK_NULL_HANDLE;
+    VkDeviceMemory mTextureImageMemory = VK_NULL_HANDLE;
 
-    Handle<VulkanTextureState> mState;
+    // Track the image layout of each subresource using a sparse range map.
+    utils::RangeMap<uint32_t, VulkanLayout> mSubresourceLayouts;
+
+    VkImageSubresourceRange mFullViewRange;
 
     // Track the range of subresources that define the "primary" image view, which is the special
     // image view that gets bound to an actual texture sampler.
     VkImageSubresourceRange mPrimaryViewRange;
 
-    VkComponentMapping mSwizzle {};
+    std::unordered_map<ImageViewKey, VkImageView, ImageViewHash> mCachedImageViews;
+    VulkanStagePool& mStagePool;
+    VkDevice mDevice;
+    VmaAllocator mAllocator;
+    VulkanCommands* mCommands;
 };
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanUtility.cpp

@@ -577,7 +577,7 @@ uint32_t getComponentCount(VkFormat format) {
     return {};
 }
 
-VkComponentMapping getSwizzleMap(TextureSwizzle const swizzle[4]) {
+VkComponentMapping getSwizzleMap(TextureSwizzle swizzle[4]) {
     VkComponentMapping map;
     VkComponentSwizzle* dst = &map.r;
     for (int i = 0; i < 4; ++i, ++dst) {

filament/backend/src/vulkan/VulkanUtility.h

@@ -19,7 +19,6 @@
 
 #include <backend/DriverEnums.h>
 
-#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 
 #include <bluevk/BlueVK.h>
@@ -39,7 +38,7 @@ VkCullModeFlags getCullMode(CullingMode mode);
 VkFrontFace getFrontFace(bool inverseFrontFaces);
 PixelDataType getComponentType(VkFormat format);
 uint32_t getComponentCount(VkFormat format);
-VkComponentMapping getSwizzleMap(TextureSwizzle const swizzle[4]);
+VkComponentMapping getSwizzleMap(TextureSwizzle swizzle[4]);
 VkShaderStageFlags getShaderStageFlags(ShaderStageFlags stageFlags);
 
 bool equivalent(const VkRect2D& a, const VkRect2D& b);
@@ -406,13 +405,12 @@ constexpr VkFormat ALL_VK_FORMATS[] = {
         VK_FORMAT_R16G16_S10_5_NV,
 };
 
-// An Array that will be statically fixed in capacity, but the "size" (as in user added elements) is
-// variable. Note that this class is movable.
+// An Array that will be fixed capacity, but the "size" (as in user added elements) is variable.
+// Note that this class is movable.
 template<typename T, uint16_t CAPACITY>
 class CappedArray {
 private:
     using FixedSizeArray = std::array<T, CAPACITY>;
-
 public:
     using const_iterator = typename FixedSizeArray::const_iterator;
     using iterator = typename FixedSizeArray::iterator;
@@ -450,20 +448,6 @@ public:
         return mArray.cend();
     }
 
-    inline iterator begin() {
-        if (mInd == 0) {
-            return mArray.end();
-        }
-        return mArray.begin();
-    }
-
-    inline iterator end() {
-        if (mInd > 0 && mInd < CAPACITY) {
-            return mArray.begin() + mInd;
-        }
-        return mArray.end();
-    }
-
     inline T back() {
         assert_invariant(mInd > 0);
         return *(mArray.begin() + mInd);
@@ -528,28 +512,125 @@ private:
     uint32_t mInd = 0;
 };
 
-using UniformBufferBitmask = utils::bitset64;
-using SamplerBitmask = utils::bitset64;
+// TODO: ok to remove once Filament-side API is complete
+namespace descset {
+
+// Used to describe the descriptor binding in shader stages. We assume that the binding index does
+// not exceed 31. We also assume that we have two shader stages - vertex and fragment.  The below
+// types and struct are used across VulkanDescriptorSet and VulkanProgram.
+using UniformBufferBitmask = uint32_t;
+using SamplerBitmask = uint64_t;
 
 // We only have at most one input attachment, so this bitmask exists only to make the code more
 // general.
-using InputAttachmentBitmask = utils::bitset64;
+using InputAttachmentBitmask = uint8_t;
+
+constexpr UniformBufferBitmask UBO_VERTEX_STAGE = 0x1;
+constexpr UniformBufferBitmask UBO_FRAGMENT_STAGE = (0x1ULL << (sizeof(UniformBufferBitmask) * 4));
+constexpr SamplerBitmask SAMPLER_VERTEX_STAGE = 0x1;
+constexpr SamplerBitmask SAMPLER_FRAGMENT_STAGE = (0x1ULL << (sizeof(SamplerBitmask) * 4));
+constexpr InputAttachmentBitmask INPUT_ATTACHMENT_VERTEX_STAGE = 0x1;
+constexpr InputAttachmentBitmask INPUT_ATTACHMENT_FRAGMENT_STAGE =
+        (0x1ULL << (sizeof(InputAttachmentBitmask) * 4));
 
 template<typename Bitmask>
-static constexpr uint8_t getVertexStageShift() noexcept {
-    // We assume the bottom half of bits are for vertex stages.
-    return 0;
+static constexpr Bitmask getVertexStage() noexcept {
+    if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
+        return UBO_VERTEX_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
+        return SAMPLER_VERTEX_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
+        return INPUT_ATTACHMENT_VERTEX_STAGE;
+    }
 }
 
 template<typename Bitmask>
-static constexpr uint8_t getFragmentStageShift() noexcept {
-    // We assume the top half of bits are for fragment stages.
-    return sizeof(Bitmask) * 4;
+static constexpr Bitmask getFragmentStage() noexcept {
+    if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
+        return UBO_FRAGMENT_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
+        return SAMPLER_FRAGMENT_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
+        return INPUT_ATTACHMENT_FRAGMENT_STAGE;
+    }
 }
 
-// We have at most 4 descriptor sets. This is to indicate which ones are active.
-using DescriptorSetMask = utils::bitset8;
+typedef enum ShaderStageFlags2 : uint8_t {
+    NONE        =    0,
+    VERTEX      =    0x1,
+    FRAGMENT    =    0x2,
+} ShaderStageFlags2;
 
+enum class DescriptorType : uint8_t {
+    UNIFORM_BUFFER,
+    SAMPLER,
+    INPUT_ATTACHMENT,
+};
+
+enum class DescriptorFlags : uint8_t {
+    NONE = 0x00,
+    DYNAMIC_OFFSET = 0x01
+};
+
+struct DescriptorSetLayoutBinding {
+    DescriptorType type;
+    ShaderStageFlags2 stageFlags;
+    uint8_t binding;
+    DescriptorFlags flags;
+    uint16_t count;
+};
+
+struct DescriptorSetLayout {
+    utils::FixedCapacityVector<DescriptorSetLayoutBinding> bindings;
+};
+
+} // namespace descset
+
+namespace {
+// Use constexpr to statically generate a bit count table for 8-bit numbers.
+struct _BitCountHelper {
+    constexpr _BitCountHelper() : data{} {
+        for (uint16_t i = 0; i < 256; ++i) {
+            data[i] = 0;
+            for (auto j = i; j > 0; j /= 2) {
+                if (j & 1) {
+                    data[i]++;
+                }
+            }
+        }
+    }
+
+    template<typename MaskType>
+    constexpr uint8_t count(MaskType num) {
+        uint8_t count = 0;
+        for (uint8_t i = 0; i < sizeof(MaskType) * 8; i+=8) {
+            count += data[(num >> i) & 0xFF];
+        }
+        return count;
+    }
+
+private:
+    uint8_t data[256];
+};
+} // namespace anonymous
+
+template<typename MaskType>
+inline uint8_t countBits(MaskType num) {
+    static _BitCountHelper BitCounter = {};
+    return BitCounter.count(num);
+}
+
+// This is useful for counting the total number of descriptors for both vertex and fragment stages.
+template<typename MaskType>
+inline MaskType collapseStages(MaskType mask) {
+    constexpr uint8_t NBITS_DIV_2 = sizeof(MaskType) * 4;
+    // First zero out the top-half and then or the bottom-half against the original top-half.
+    return ((mask << NBITS_DIV_2) >> NBITS_DIV_2) | (mask >> NBITS_DIV_2);
+}
 
 } // namespace filament::backend