Check for invalid texture inside MaterialInstance

vk: reduce map look-up in descriptor set hot loop
DescriptorSetManager's commit() is the inner most of the rendering loop. We want to avoid any hash map look up here. We just store a pointer to the history object where appropriate.
2024-09-12 12:11:01 -07:00 · 2024-09-12 08:48:25 -07:00 · 2024-09-12 08:48:25 -07:00 · 2024-09-11 18:51:15 -07:00 · 2024-09-10 13:36:45 -07:00 · 2024-09-10 12:13:24 -07:00
246 changed files with 4033 additions and 7606 deletions
@@ -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
@@ -8,15 +8,35 @@ on:
      - rc/**

 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
+  build-android-armv7:
+    name: build-android-armv7
+    runs-on: macos-14

    steps:
      - uses: actions/checkout@v4.1.6
      - name: Run Android Continuous
        uses: ./.github/actions/android-continuous
        with:
-          build-abi: armeabi-v7a,arm64-v8a,x86_64
+          build-abi: armeabi-v7a
+
+  build-android-armv8a:
+    name: build-android-armv8a
+    runs-on: macos-14
+
+    steps:
+      - uses: actions/checkout@v4.1.6
+      - name: Run Android Continuous
+        uses: ./.github/actions/android-continuous
+        with:
+          build-abi: arm64-v8a
+
+  build-android-x86_64:
+    name: build-android-x86_64
+    runs-on: macos-14
+
+    steps:
+      - uses: actions/checkout@v4.1.6
+      - name: Run Android Continuous
+        uses: ./.github/actions/android-continuous
+        with:
+          build-abi: x86_64
@@ -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
@@ -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
@@ -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
@@ -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
@@ -56,7 +56,7 @@ jobs:

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

    steps:
      - uses: actions/checkout@v4.1.6
@@ -69,25 +69,10 @@ 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
@@ -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:
@@ -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
@@ -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
@@ -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
@@ -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)
@@ -7,3 +7,5 @@ for next branch cut* header.
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).

 ## Release notes for next branch cut
+
+- Add support for multi-layered render target with array textures.
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.56.5'
+    implementation 'com.google.android.filament:filament-android:1.54.2'
 }
 ```

@@ -51,7 +51,7 @@ 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', '~> 1.56.5'
+pod 'Filament', '~> 1.54.2'
 ```

 ## Documentation
@@ -7,38 +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).

-## v1.56.5
-
-
-## v1.56.4
-
-
-## v1.56.3
-
-
-## v1.56.2
-
- vk: fix stage pool gc logic
-
-## v1.56.1
-
-## v1.56.0
-
- backend: descriptor layouts distinguish samplers and external samplers (b/376089915) [⚠️ **New Material Version**]
-
-## 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


@@ -673,7 +641,7 @@ Instead, if you are authoring a PR for the main branch, add your release note to

 - engine: Binary size improvements.
 - engine: Add basic support for instanced renderables [**NEW API**].
- engine: Fix, first imaged passed to `Stream::SetAcquiredImage` is ignored and leaked.
+- engine: Fix, first imaged passsed to `Stream::SetAcquiredImage` is ignored and leaked.
 - Vulkan: Robustness improvements.
 - Java: Fix, lookAt z axis negated.
 - gltfio: Be graceful when model has > 4 weights per vert.
@@ -83,12 +83,12 @@ buildscript {
        'minSdk': 21,
        'targetSdk': 34,
        'compileSdk': 34,
-        'kotlin': '2.0.21',
-        'kotlin_coroutines': '1.9.0',
-        'buildTools': '35.0.0',
+        'kotlin': '2.0.0',
+        'kotlin_coroutines': '1.9.0-RC',
+        'buildTools': '34.0.0',
        'ndk': '27.0.11718014',
        'androidx_core': '1.13.1',
-        'androidx_annotations': '1.9.0'
+        'androidx_annotations': '1.8.0'
    ]

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

    dependencies {
-        classpath 'com.android.tools.build:gradle:8.6.1'
+        classpath 'com.android.tools.build:gradle:8.4.1'
        classpath "org.jetbrains.kotlin:kotlin-gradle-plugin:${versions.kotlin}"
    }

@@ -499,37 +499,6 @@ Java_com_google_android_filament_Engine_nGetActiveFeatureLevel(JNIEnv *, jclass,
    return (jint)engine->getActiveFeatureLevel();
 }

-extern "C"
-JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_Engine_nHasFeatureFlag(JNIEnv *env, jclass clazz,
-        jlong nativeEngine, jstring name_) {
-    Engine* engine = (Engine*) nativeEngine;
-    const char *name = env->GetStringUTFChars(name_, 0);
-    std::optional<bool> result = engine->getFeatureFlag(name);
-    env->ReleaseStringUTFChars(name_, name);
-    return result.has_value();
-}
-extern "C"
-JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_Engine_nSetFeatureFlag(JNIEnv *env, jclass clazz,
-        jlong nativeEngine, jstring name_, jboolean value) {
-    Engine* engine = (Engine*) nativeEngine;
-    const char *name = env->GetStringUTFChars(name_, 0);
-    jboolean result = engine->setFeatureFlag(name, (bool)value);
-    env->ReleaseStringUTFChars(name_, name);
-    return result;
-}
-extern "C"
-JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_Engine_nGetFeatureFlag(JNIEnv *env, jclass clazz,
-        jlong nativeEngine, jstring name_) {
-    Engine* engine = (Engine*) nativeEngine;
-    const char *name = env->GetStringUTFChars(name_, 0);
-    std::optional<bool> result = engine->getFeatureFlag(name);
-    env->ReleaseStringUTFChars(name_, name);
-    return result.value_or(false); // we should never fail here
-}
-
 extern "C" JNIEXPORT jlong JNICALL Java_com_google_android_filament_Engine_nCreateBuilder(JNIEnv*,
        jclass) {
    Engine::Builder* builder = new Engine::Builder{};
@@ -551,12 +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,
@@ -565,6 +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,
@@ -573,7 +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);
 }
@@ -596,16 +566,6 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
    builder->paused((bool) paused);
 }

-extern "C"
-JNIEXPORT void JNICALL
-Java_com_google_android_filament_Engine_nSetBuilderFeature(JNIEnv *env, jclass clazz,
-        jlong nativeBuilder, jstring name_, jboolean value) {
-    Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
-    const char *name = env->GetStringUTFChars(name_, 0);
-    builder->feature(name, (bool)value);
-    env->ReleaseStringUTFChars(name_, name);
-}
-
 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Engine_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder) {
    Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
@@ -38,7 +38,7 @@ Java_com_google_android_filament_EntityManager_nCreateArray(JNIEnv* env, jclass,
    // (which it is), but still.
    em->create((size_t) n, reinterpret_cast<Entity *>(entities));

-    env->ReleaseIntArrayElements(entities_, entities, 0);
+    env->ReleaseIntArrayElements(entities_, entities, JNI_ABORT);
 }

 extern "C" JNIEXPORT jint JNICALL
@@ -167,7 +167,7 @@ Java_com_google_android_filament_IndirectLight_nGetDirectionEstimateStatic(JNIEn
    jfloat *outDirection = env->GetFloatArrayElements(outDirection_, NULL);
    *reinterpret_cast<filament::math::float3*>(outDirection) = IndirectLight::getDirectionEstimate((filament::math::float3*)sh);
    env->ReleaseFloatArrayElements(outDirection_, outDirection, 0);
-    env->ReleaseFloatArrayElements(sh_, sh, 0);
+    env->ReleaseFloatArrayElements(sh_, sh, JNI_ABORT);
 }

 extern "C" JNIEXPORT void JNICALL
@@ -178,5 +178,5 @@ Java_com_google_android_filament_IndirectLight_nGetColorEstimateStatic(JNIEnv *e
    *reinterpret_cast<filament::math::float4*>(outColor) =
            IndirectLight::getColorEstimate((filament::math::float3*)sh, math::float3{x, y, z});
    env->ReleaseFloatArrayElements(outColor_, outColor, 0);
-    env->ReleaseFloatArrayElements(sh_, sh, 0);
+    env->ReleaseFloatArrayElements(sh_, sh, JNI_ABORT);
 }
@@ -25,14 +25,13 @@ using namespace filament;

 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Material_nBuilderBuild(JNIEnv *env, jclass,
-        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount, jint shadowQuality) {
+        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount) {
    Engine* engine = (Engine*) nativeEngine;
    AutoBuffer buffer(env, buffer_, size);
    auto builder = Material::Builder();
    if (shBandCount) {
        builder.sphericalHarmonicsBandCount(shBandCount);
    }
-    builder.shadowSamplingQuality((Material::Builder::ShadowSamplingQuality)shadowQuality);
    Material* material = builder
            .package(buffer.getData(), buffer.getSize())
            .build(*engine);
@@ -222,20 +222,6 @@ Java_com_google_android_filament_View_nIsFrontFaceWindingInverted(JNIEnv*,
    return static_cast<jboolean>(view->isFrontFaceWindingInverted());
 }

-extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_View_nSetTransparentPickingEnabled(JNIEnv*,
-        jclass, jlong nativeView, jboolean enabled) {
-    View* view = (View*) nativeView;
-    view->setTransparentPickingEnabled(enabled);
-}
-
-extern "C" JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_View_nIsTransparentPickingEnabled(JNIEnv*,
-        jclass, jlong nativeView) {
-    View* view = (View*) nativeView;
-    return static_cast<jboolean>(view->isTransparentPickingEnabled());
-}
-
 extern "C" JNIEXPORT void JNICALL
 Java_com_google_android_filament_View_nSetAmbientOcclusion(JNIEnv*, jclass, jlong nativeView, jint ordinal) {
    View* view = (View*) nativeView;
@@ -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;
        }

@@ -258,17 +259,6 @@ public class Engine {
            return this;
        }

-        /**
-         * Set a feature flag value. This is the only way to set constant feature flags.
-         * @param name feature name
-         * @param value true to enable, false to disable
-         * @return A reference to this Builder for chaining calls.
-         */
-        public Builder feature(@NonNull String name, boolean value) {
-            nSetBuilderFeature(mNativeBuilder, name, value);
-            return this;
-        }
-
        /**
         * Creates an instance of Engine
         *
@@ -404,7 +394,6 @@ public class Engine {
        /**
         * Set to `true` to forcibly disable parallel shader compilation in the backend.
         * Currently only honored by the GL backend.
-         * @Deprecated use "backend.disable_parallel_shader_compile" feature flag instead
         */
        public boolean disableParallelShaderCompile = false;

@@ -430,12 +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.
@@ -443,13 +432,12 @@ public class Engine {
         */
        public long resourceAllocatorCacheMaxAge = 1;

-        /**
+        /*
         * Disable backend handles use-after-free checks.
-         * @Deprecated use "backend.disable_handle_use_after_free_check" feature flag instead
         */
        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
@@ -471,20 +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
-         * @Deprecated use "backend.opengl.assert_native_window_is_valid" feature flag instead
-         */
-        public boolean assertNativeWindowIsValid = false;
    }

    private Engine(long nativeEngine, Config config) {
@@ -707,11 +687,11 @@ public class Engine {

    /**
     * Returns the maximum number of stereoscopic eyes supported by Filament. The actual number of
-     * eyes rendered is set at Engine creation time with the {@link Config#stereoscopicEyeCount}
-     * setting.
+     * eyes rendered is set at Engine creation time with the {@link
+     * Engine#Config#stereoscopicEyeCount} setting.
     *
     * @return the max number of stereoscopic eyes supported
-     * @see Config#stereoscopicEyeCount
+     * @see Engine#Config#stereoscopicEyeCount
     */
    public long getMaxStereoscopicEyes() {
        return nGetMaxStereoscopicEyes(getNativeObject());
@@ -908,8 +888,7 @@ public class Engine {

    /**
     * Returns whether the object is valid.
-     * @param ma Material
-     * @param mi MaterialInstance to check for validity
+     * @param object Object to check for validity
     * @return returns true if the specified object is valid.
     */
    public boolean isValidMaterialInstance(@NonNull Material ma, MaterialInstance mi) {
@@ -1331,39 +1310,6 @@ public class Engine {
     */
    public static native long getSteadyClockTimeNano();

-
-    /**
-     * Checks if a feature flag exists
-     * @param name name of the feature flag to check
-     * @return true if it exists false otherwise
-     */
-    public boolean hasFeatureFlag(@NonNull String name) {
-        return nHasFeatureFlag(mNativeObject, name);
-    }
-
-    /**
-     * Set the value of a non-constant feature flag.
-     * @param name name of the feature flag to set
-     * @param value value to set
-     * @return true if the value was set, false if the feature flag is constant or doesn't exist.
-     */
-    public boolean setFeatureFlag(@NonNull String name, boolean value) {
-        return nSetFeatureFlag(mNativeObject, name, value);
-    }
-
-    /**
-     * Retrieves the value of any feature flag.
-     * @param name name of the feature flag
-     * @return the value of the flag if it exists
-     * @exception IllegalArgumentException is thrown if the feature flag doesn't exist
-     */
-    public boolean getFeatureFlag(@NonNull String name) {
-        if (!hasFeatureFlag(name)) {
-            throw new IllegalArgumentException("The feature flag \"" + name + "\" doesn't exist");
-        }
-        return nGetFeatureFlag(mNativeObject, name);
-    }
-
    @UsedByReflection("TextureHelper.java")
    public long getNativeObject() {
        if (mNativeObject == 0) {
@@ -1453,9 +1399,6 @@ public class Engine {
    private static native int nGetSupportedFeatureLevel(long nativeEngine);
    private static native int nSetActiveFeatureLevel(long nativeEngine, int ordinal);
    private static native int nGetActiveFeatureLevel(long nativeEngine);
-    private static native boolean nHasFeatureFlag(long nativeEngine, String name);
-    private static native boolean nSetFeatureFlag(long nativeEngine, String name, boolean value);
-    private static native boolean nGetFeatureFlag(long nativeEngine, String name);

    private static native long nCreateBuilder();
    private static native void nDestroyBuilder(long nativeBuilder);
@@ -1463,14 +1406,14 @@ 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);
-    private static native void nSetBuilderFeature(long nativeBuilder, String name, boolean value);
    private static native long nBuilderBuild(long nativeBuilder);
 }
@@ -344,18 +344,9 @@ public class Material {
    }

    public static class Builder {
-        public enum ShadowSamplingQuality {
-            /** 2x2 PCF */
-            HARD,
-            /** 3x3 gaussian filter */
-            LOW,
-        }
-
        private Buffer mBuffer;
        private int mSize;
        private int mShBandCount = 0;
-        private ShadowSamplingQuality mShadowSamplingQuality = ShadowSamplingQuality.LOW;
-

        /**
         * Specifies the material data. The material data is a binary blob produced by
@@ -387,18 +378,6 @@ public class Material {
            return this;
        }

-        /**
-         * Set the quality of shadow sampling. This is only taken into account
-         * if this material is lit and in the surface domain.
-         * @param quality
-         * @return Reference to this Builder for chaining calls.
-         */
-        @NonNull
-        public Builder shadowSamplingQuality(ShadowSamplingQuality quality) {
-            mShadowSamplingQuality = quality;
-            return this;
-        }
-
        /**
         * Creates and returns the Material object.
         *
@@ -411,7 +390,7 @@ public class Material {
        @NonNull
        public Material build(@NonNull Engine engine) {
            long nativeMaterial = nBuilderBuild(engine.getNativeObject(),
-                mBuffer, mSize, mShBandCount, mShadowSamplingQuality.ordinal());
+                mBuffer, mSize, mShBandCount);
            if (nativeMaterial == 0) throw new IllegalStateException("Couldn't create Material");
            return new Material(nativeMaterial);
        }
@@ -1062,7 +1041,7 @@ public class Material {
        mNativeObject = 0;
    }

-    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount, int shadowQuality);
+    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount);
    private static native long nCreateInstance(long nativeMaterial);
    private static native long nCreateInstanceWithName(long nativeMaterial, @NonNull String name);
    private static native long nGetDefaultInstance(long nativeMaterial);
@@ -745,33 +745,6 @@ public class View {
        nSetFrontFaceWindingInverted(getNativeObject(), inverted);
    }

-    /**
-     * Returns true if transparent picking is enabled.
-     *
-     * @see #setTransparentPickingEnabled
-     */
-    public boolean isTransparentPickingEnabled() {
-        return nIsTransparentPickingEnabled(getNativeObject());
-    }
-
-    /**
-     * Enables or disables transparent picking. Disabled by default.
-     *
-     * When transparent picking is enabled, View::pick() will pick from both
-     * transparent and opaque renderables. When disabled, View::pick() will only
-     * pick from opaque renderables.
-     *
-     * <p>
-     * Transparent picking will create an extra pass for rendering depth
-     * from both transparent and opaque renderables. 
-     * </p>
-     *
-     * @param enabled true enables transparent picking, false disables it.
-     */
-    public void setTransparentPickingEnabled(boolean enabled) {
-        nSetTransparentPickingEnabled(getNativeObject(), enabled);
-    }
-
    /**
     * Sets options relative to dynamic lighting for this view.
     *
@@ -1308,8 +1281,6 @@ public class View {
    private static native boolean nIsPostProcessingEnabled(long nativeView);
    private static native void nSetFrontFaceWindingInverted(long nativeView, boolean inverted);
    private static native boolean nIsFrontFaceWindingInverted(long nativeView);
-    private static native void nSetTransparentPickingEnabled(long nativeView, boolean enabled);
-    private static native boolean nIsTransparentPickingEnabled(long nativeView);
    private static native void nSetAmbientOcclusion(long nativeView, int ordinal);
    private static native int nGetAmbientOcclusion(long nativeView);
    private static native void nSetAmbientOcclusionOptions(long nativeView, float radius, float bias, float power, float resolution, float intensity, float bilateralThreshold, int quality, int lowPassFilter, int upsampling, boolean enabled, boolean bentNormals, float minHorizonAngleRad);
@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.56.5
+VERSION_NAME=1.54.2

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

@@ -1,6 +1,6 @@
 #Wed Nov 17 10:40:18 PST 2021
 distributionBase=GRADLE_USER_HOME
 distributionPath=wrapper/dists
-distributionUrl=https\://services.gradle.org/distributions/gradle-8.8-bin.zip
+distributionUrl=https\://services.gradle.org/distributions/gradle-8.6-bin.zip
 zipStoreBase=GRADLE_USER_HOME
 zipStorePath=wrapper/dists
@@ -6,10 +6,6 @@ plugins {

 project.ext.isSample = true

-kotlin {
-    jvmToolchain(versions.jdk)
-}
-
 filamentTools {
    cmgenArgs = "-q --format=ktx --size=256 --extract-blur=0.1 --deploy=src/main/assets/envs/default_env"
    iblInputFile = project.layout.projectDirectory.file("../../../third_party/environments/lightroom_14b.hdr")
@@ -17,7 +13,7 @@ filamentTools {
 }

 // don't forget to update MainACtivity.kt when/if changing this.
-tasks.register('copyMesh', Copy) {
+task copyMesh(type: Copy) {
    from "../../../third_party/models/BusterDrone"
    into "src/main/assets/models"
 }
@@ -112,13 +112,7 @@ class MainActivity : Activity() {
    }

    private fun setupFilament() {
-        val config = Engine.Config()
-        //config.forceGLES2Context = true
-
-        engine = Engine.Builder()
-            .config(config)
-            .featureLevel(Engine.FeatureLevel.FEATURE_LEVEL_0)
-            .build()
+        engine = Engine.Builder().featureLevel(Engine.FeatureLevel.FEATURE_LEVEL_0).build()
        renderer = engine.createRenderer()
        scene = engine.createScene()
        view = engine.createView()
@@ -129,9 +123,7 @@ class MainActivity : Activity() {
        scene.skybox = Skybox.Builder().color(0.035f, 0.035f, 0.035f, 1.0f).build(engine)

        // post-processing is not supported at feature level 0
-        if (engine.activeFeatureLevel == Engine.FeatureLevel.FEATURE_LEVEL_0) {
-            view.isPostProcessingEnabled = false
-        }
+        view.isPostProcessingEnabled = false

        // Tell the view which camera we want to use
        view.camera = camera
@@ -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 ""
@@ -1,6 +1,7 @@
 #!/bin/bash
 if [ `uname` == "Linux" ];then
-    source `dirname $0`/../linux/ci-common.sh
+    curl -OL https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-linux.zip
+    unzip -q ninja-linux.zip
 elif [ `uname` == "Darwin" ];then
    curl -OL https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-mac.zip
    unzip -q ninja-mac.zip
@@ -12,6 +13,9 @@ fi
 chmod +x ninja
 export PATH="$PWD:$PATH"

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

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

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

 echo Disk info after building variant: %variant%
@@ -1787,12 +1787,7 @@ non-shader data.
      declare a variable called <code>eyeDirection</code> you can access it in the fragment shader using
      <code>variable_eyeDirection</code>. In the vertex shader, the interpolant name is simply a member of
      the <code>MaterialVertexInputs</code> structure (<code>material.eyeDirection</code> in your example). Each
-      interpolant is of type <code>float4</code> (<code>vec4</code>) in the shaders. By default the precision of the
-      interpolant is <code>highp</code> in <em class="asterisk">both</em> 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 <code>default</code> is specified the default precision is used is the fragment shader
-      (<code>mediump</code>) and in the vertex shader (<code>highp</code>).
+      interpolant is of type <code>float4</code> (<code>vec4</code>) in the shaders.

 </p></dd></dl><p></p><pre class="listing tilde"><code><span class="line">material {</span>
 <span class="line">    name : Skybox,</span>
@@ -1803,11 +1798,7 @@ non-shader data.
 <span class="line">        }</span>
 <span class="line">    ],</span>
 <span class="line">    variables : [</span>
-<span class="line">         eyeDirection,</span>
-<span class="line">         {</span>
-<span class="line">            name : eyeColor,</span>
-<span class="line">            precision : medium</span>
-<span class="line">         }</span>
+<span class="line">         eyeDirection</span>
 <span class="line">    ],</span>
 <span class="line">    vertexDomain : device,</span>
 <span class="line">    depthWrite : <span class="hljs-literal">false</span>,</span>
@@ -66,7 +66,6 @@ set(SRCS
        src/FrameSkipper.cpp
        src/Froxelizer.cpp
        src/Frustum.cpp
-        src/HwDescriptorSetLayoutFactory.cpp
        src/HwRenderPrimitiveFactory.cpp
        src/HwVertexBufferInfoFactory.cpp
        src/IndexBuffer.cpp
@@ -154,7 +153,6 @@ set(PRIVATE_HDRS
        src/FrameInfo.h
        src/FrameSkipper.h
        src/Froxelizer.h
-        src/HwDescriptorSetLayoutFactory.h
        src/HwRenderPrimitiveFactory.h
        src/HwVertexBufferInfoFactory.h
        src/Intersections.h
@@ -308,11 +306,6 @@ if (FILAMENT_ENABLE_MULTIVIEW)
    add_definitions(-DFILAMENT_ENABLE_MULTIVIEW)
 endif()

-# Whether to force the profiling mode.
-if (FILAMENT_FORCE_PROFILING_MODE)
-    add_definitions(-DFILAMENT_FORCE_PROFILING_MODE)
-endif()
-
 # ==================================================================================================
 # Definitions
 # ==================================================================================================
@@ -612,13 +605,6 @@ else()
            -Wover-aligned
            -Werror
    )
-    if (CMAKE_CXX_STANDARD EQUAL 20)
-        # The lambdas for passes in PostProcessManager.cpp capture this
-        # implicitly in a way that's deprecated in c++20, but can't easily be
-        # fixed in a way that's backwards compatible with c++17:
-        # https://www.nextptr.com/tutorial/ta1430524603/capture-this-in-lambda-expression-timeline-of-change
-        list(APPEND FILAMENT_WARNINGS -Wno-deprecated-this-capture)
-    endif()
 endif()

 target_compile_options(${TARGET} PRIVATE
@@ -119,8 +119,6 @@ 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)
@@ -179,18 +177,11 @@ if (FILAMENT_SUPPORTS_VULKAN)
            src/vulkan/caching/VulkanDescriptorSetManager.h
            src/vulkan/caching/VulkanPipelineLayoutCache.cpp
            src/vulkan/caching/VulkanPipelineLayoutCache.h
-            src/vulkan/memory/ResourceManager.cpp
-            src/vulkan/memory/ResourceManager.h
-            src/vulkan/memory/ResourcePointer.h
-            src/vulkan/memory/Resource.cpp
-            src/vulkan/memory/Resource.h
            src/vulkan/platform/VulkanPlatform.cpp
            src/vulkan/platform/VulkanPlatformSwapChainImpl.cpp
            src/vulkan/platform/VulkanPlatformSwapChainImpl.h
            src/vulkan/spirv/VulkanSpirvUtils.cpp
            src/vulkan/spirv/VulkanSpirvUtils.h
-            src/vulkan/VulkanAsyncHandles.cpp
-            src/vulkan/VulkanAsyncHandles.h
            src/vulkan/VulkanBlitter.cpp
            src/vulkan/VulkanBlitter.h
            src/vulkan/VulkanBuffer.cpp
@@ -221,17 +212,18 @@ if (FILAMENT_SUPPORTS_VULKAN)
            src/vulkan/VulkanSwapChain.h
            src/vulkan/VulkanReadPixels.cpp
            src/vulkan/VulkanReadPixels.h
+            src/vulkan/VulkanResourceAllocator.h
+            src/vulkan/VulkanResources.cpp
+            src/vulkan/VulkanResources.h
            src/vulkan/VulkanTexture.cpp
            src/vulkan/VulkanTexture.h
            src/vulkan/VulkanUtility.cpp
            src/vulkan/VulkanUtility.h
    )
-    if (LINUX OR WIN32)
-        list(APPEND SRCS src/vulkan/platform/VulkanPlatformLinuxWindows.cpp)
+    if (ANDROID OR LINUX OR WIN32)
+        list(APPEND SRCS src/vulkan/platform/VulkanPlatformAndroidLinuxWindows.cpp)
    elseif (APPLE OR IOS)
        list(APPEND SRCS src/vulkan/platform/VulkanPlatformApple.mm)
-    elseif (ANDROID)
-        list(APPEND SRCS src/vulkan/platform/VulkanPlatformAndroid.cpp)
    endif()
 endif()

@@ -374,11 +366,6 @@ 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()
@@ -390,9 +377,6 @@ else()
            -Wover-aligned
            -Werror
    )
-    if (CMAKE_CXX_STANDARD EQUAL 20)
-        list(APPEND FILAMENT_WARNINGS -Wno-deprecated-this-capture)
-    endif()
 endif()

 if (APPLE)
@@ -402,7 +386,6 @@ endif()

 target_compile_options(${TARGET} PRIVATE
        ${FILAMENT_WARNINGS}
-        ${OSMESA_COMPILE_FLAGS}
        $<$<CONFIG:Release>:${OPTIMIZATION_FLAGS}>
        $<$<AND:$<PLATFORM_ID:Darwin>,$<CONFIG:Release>>:${DARWIN_OPTIMIZATION_FLAGS}>
 )
@@ -412,7 +395,6 @@ if (FILAMENT_SUPPORTS_METAL)
 endif()

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

@@ -447,7 +429,6 @@ 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
@@ -545,8 +526,6 @@ if (APPLE AND NOT IOS)

 add_executable(metal_utils_test test/MetalTest.mm)

-target_compile_options(metal_utils_test PRIVATE "-fobjc-arc")
-
 target_link_libraries(metal_utils_test PRIVATE
        backend
        getopt
@@ -227,7 +227,6 @@ enum class DescriptorType : uint8_t {
    SHADER_STORAGE_BUFFER,
    SAMPLER,
    INPUT_ATTACHMENT,
-    SAMPLER_EXTERNAL
 };

 enum class DescriptorFlags : uint8_t {
@@ -243,17 +242,8 @@ 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;
-    }
+    DescriptorFlags flags;
+    uint16_t count;
 };

 struct DescriptorSetLayout {
@@ -319,20 +309,9 @@ struct Viewport {
    int32_t right() const noexcept { return left + int32_t(width); }
    //! get the top coordinate in window space of the viewport
    int32_t top() const noexcept { return bottom + int32_t(height); }
-
-    friend bool operator==(Viewport const& lhs, Viewport const& rhs) noexcept {
-        // clang can do this branchless with xor/or
-        return lhs.left == rhs.left && lhs.bottom == rhs.bottom &&
-               lhs.width == rhs.width && lhs.height == rhs.height;
-    }
-
-    friend bool operator!=(Viewport const& lhs, Viewport const& rhs) noexcept {
-        // clang is being dumb and uses branches
-        return bool(((lhs.left ^ rhs.left) | (lhs.bottom ^ rhs.bottom)) |
-                    ((lhs.width ^ rhs.width) | (lhs.height ^ rhs.height)));
-    }
 };

+
 /**
 * Specifies the mapping of the near and far clipping plane to window coordinates.
 */
@@ -777,8 +756,7 @@ 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) {
@@ -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>

@@ -107,18 +106,8 @@ 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) { }

@@ -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;
@@ -30,6 +30,10 @@ namespace filament::backend {

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

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

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

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

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

@@ -89,11 +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
@@ -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
@@ -88,7 +88,6 @@ public:
        VkFormat colorFormat = VK_FORMAT_UNDEFINED;
        VkFormat depthFormat = VK_FORMAT_UNDEFINED;
        VkExtent2D extent = {0, 0};
-        bool isProtected = false;
    };

    struct ImageSyncData {
@@ -200,13 +199,6 @@ public:
     */
    virtual bool hasResized(SwapChainPtr handle);

-    /**
-     * Check if the surface is protected.
-     * @param handle             The handle returned by createSwapChain()
-     * @return                   Whether the swapchain is protected
-     */
-    virtual bool isProtected(SwapChainPtr handle);
-
    /**
     * Carry out a recreation of the swapchain.
     * @param handle             The handle returned by createSwapChain()
@@ -275,81 +267,8 @@ public:
     */
    VkQueue getGraphicsQueue() const noexcept;

-    /**
-    * @return The family index of the protected graphics queue selected for the
-    *          Vulkan backend.
-    */
-    uint32_t getProtectedGraphicsQueueFamilyIndex() const noexcept;
-
-    /**
-     * @return The index of the protected graphics queue (if there are multiple
-     *          graphics queues) selected for the Vulkan backend.
-     */
-    uint32_t getProtectedGraphicsQueueIndex() const noexcept;
-
-    /**
-     * @return The protected queue that was selected for the Vulkan backend.
-     */
-    VkQueue getProtectedGraphicsQueue() const noexcept;
-
-    struct ExternalImageMetadata {
-        /**
-         * The width of the external image
-         */
-        uint32_t width;
-
-        /**
-         * The height of the external image
-         */
-        uint32_t height;
-
-        /**
-         * The layer count of the external image
-         */
-        uint32_t layers;
-
-        /**
-         * The format of the external image
-         */
-        VkFormat format;
-
-        /**
-         * An external buffer can be protected. This tells you if it is.
-         */
-        bool isProtected;
-
-        /**
-         * The type of external format (opaque int) if used.
-         */
-        uint64_t externalFormat;
-
-        /**
-         * Image usage
-         */
-        VkImageUsageFlags usage;
-
-        /**
-         * Allocation size
-         */
-        VkDeviceSize allocationSize;
-
-        /**
-         * Heap information
-         */
-        uint32_t memoryTypeBits;
-    };
-    virtual ExternalImageMetadata getExternalImageMetadata(void* externalImage);
-
-    using ImageData = std::pair<VkImage, VkDeviceMemory>;
-    virtual ImageData createExternalImage(void* externalImage,
-            const ExternalImageMetadata& metadata);
-
 private:
    static ExtensionSet getSwapchainInstanceExtensions();
-    static ExternalImageMetadata getExternalImageMetadataImpl(void* externalImage,
-            VkDevice device);
-    static ImageData createExternalImageImpl(void* externalImage, VkDevice device,
-            const VkAllocationCallbacks* allocator, const ExternalImageMetadata& metadata);

    // Platform dependent helper methods
    using SurfaceBundle = std::tuple<VkSurfaceKHR, VkExtent2D>;
@@ -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,
@@ -169,7 +169,7 @@ 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);
@@ -179,22 +179,6 @@ public:
                        << "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();
-                }
-            }
        }

        return static_cast<Dp>(p);
@@ -202,18 +186,14 @@ 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>
@@ -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
@@ -1,28 +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_SYSTRACEPROFILE_H
-#define TNT_FILAMENT_BACKEND_SYSTRACEPROFILE_H
-
-#include <utils/Systrace.h>
-
-#define PROFILE_SCOPE(marker)       SYSTRACE_NAME(marker)
-
-#define PROFILE_NAME_BEGINFRAME    "backend::beginFrame"
-#define PROFILE_NAME_ENDFRAME      "backend::endFrame"
-
-#endif // TNT_FILAMENT_BACKEND_SYSTRACEPROFILE_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;
        }
    }

@@ -174,10 +171,8 @@ 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
@@ -216,10 +211,8 @@ 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;
@@ -66,20 +66,15 @@ 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) {
@@ -89,18 +84,17 @@ 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 {
@@ -206,13 +200,9 @@ 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 +220,10 @@ 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);
@@ -118,6 +118,9 @@ private:
 };

 struct MetalContext {
+    explicit MetalContext(size_t metalFreedTextureListSize)
+        : texturesToDestroy(metalFreedTextureListSize) {}
+
    MetalDriver* driver;
    id<MTLDevice> device = nullptr;
    id<MTLCommandQueue> commandQueue = nullptr;
@@ -135,7 +138,6 @@ struct MetalContext {
    bool supportsTextureSwizzling = false;
    bool supportsAutoDepthResolve = false;
    bool supportsMemorylessRenderTargets = false;
-    bool supportsDepthClamp = false;
    uint8_t maxColorRenderTargets = 4;
    struct {
        uint8_t common;
@@ -162,7 +164,6 @@ struct MetalContext {
    CullModeStateTracker cullModeState;
    WindingStateTracker windingState;
    DepthClampStateTracker depthClampState;
-    ScissorRectStateTracker scissorRectState;
    Handle<HwRenderPrimitive> currentRenderPrimitive;

    // State caches.
@@ -186,6 +187,14 @@ struct MetalContext {
    // Keeps track of all alive textures.
    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;

@@ -19,8 +19,6 @@
 #include "CommandStreamDispatcher.h"
 #include "metal/MetalDriver.h"

-#include <filament/SwapChain.h>
-
 #include "MetalBlitter.h"
 #include "MetalBufferPool.h"
 #include "MetalContext.h"
@@ -40,7 +38,6 @@
 #include <utils/Log.h>
 #include <utils/Panic.h>
 #include <utils/sstream.h>
-#include <utils/Invocable.h>

 #include <algorithm>

@@ -108,13 +105,13 @@ Dispatcher MetalDriver::getDispatcher() const noexcept {
    return ConcreteDispatcher<MetalDriver>::make();
 }

-MetalDriver::MetalDriver(
-        MetalPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept
-    : mPlatform(*platform),
-      mContext(new MetalContext),
-      mHandleAllocator(
-              "Handles", driverConfig.handleArenaSize, driverConfig.disableHandleUseAfterFreeCheck),
-      mStereoscopicType(driverConfig.stereoscopicType) {
+MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept
+        : mPlatform(*platform),
+          mContext(new MetalContext(driverConfig.textureUseAfterFreePoolSize)),
+          mHandleAllocator("Handles",
+                  driverConfig.handleArenaSize,
+                  driverConfig.disableHandleUseAfterFreeCheck),
+          mStereoscopicType(driverConfig.stereoscopicType) {
    mContext->driver = this;

    TrackedMetalBuffer::setPlatform(platform);
@@ -151,11 +148,6 @@ MetalDriver::MetalDriver(
            mContext->highestSupportedGpuFamily.mac   >= 2;     // newer macOS GPUs
    }

-    mContext->supportsDepthClamp = false;
-    if (@available(macOS 10.11, iOS 11.0, *)) {
-        mContext->supportsDepthClamp = true;
-    }
-
    // In order to support resolve store action on depth attachment, the GPU needs to support it.
    // Note that support for depth resolve implies support for stencil resolve using .sample0 resolve filter.
    // (Other resolve filters are supported starting .apple5 and .mac2 families).
@@ -259,19 +251,13 @@ void MetalDriver::beginFrame(int64_t monotonic_clock_ns,
                TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::RING));
        mPlatform.debugUpdateStat("filament.metal.alive_buffers.staging",
                TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::STAGING));
-        mPlatform.debugUpdateStat("filament.metal.alive_buffers.descriptor_set",
-                TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::DESCRIPTOR_SET));
    }
 }

-void MetalDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch, CallbackHandler* handler,
-        FrameScheduledCallback&& callback, uint64_t flags) {
-    // Turn off the CALLBACK_DEFAULT_USE_METAL_COMPLETION_HANDLER flag if a custom handler is provided.
-    if (handler) {
-        flags &= ~SwapChain::CALLBACK_DEFAULT_USE_METAL_COMPLETION_HANDLER;
-    }
+void MetalDriver::setFrameScheduledCallback(
+        Handle<HwSwapChain> sch, CallbackHandler* handler, FrameScheduledCallback&& callback) {
    auto* swapChain = handle_cast<MetalSwapChain>(sch);
-    swapChain->setFrameScheduledCallback(handler, std::move(callback), flags);
+    swapChain->setFrameScheduledCallback(handler, std::move(callback));
 }

 void MetalDriver::setFrameCompletedCallback(
@@ -531,11 +517,6 @@ void MetalDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
 }

 void MetalDriver::createProgramR(Handle<HwProgram> rph, Program&& program) {
-#if FILAMENT_METAL_DEBUG_LOG
-    auto handleId = rph.getId();
-    DEBUG_LOG("createProgramR(rph = %d, program = ", handleId);
-    utils::slog.d << program << utils::io::endl;
-#endif
    construct_handle<MetalProgram>(rph, *mContext, std::move(program));
 }

@@ -628,8 +609,6 @@ const char* toString(DescriptorType type) {
            return "SAMPLER";
        case DescriptorType::INPUT_ATTACHMENT:
            return "INPUT_ATTACHMENT";
-        case DescriptorType::SAMPLER_EXTERNAL:
-            return "SAMPLER_EXTERNAL";
    }
 }

@@ -671,9 +650,10 @@ void MetalDriver::createDescriptorSetLayoutR(
            [](const auto& a, const auto& b) { return a.binding < b.binding; });
    DEBUG_LOG("createDescriptorSetLayoutR(dslh = %d, info = {\n", dslh.getId());
    for (size_t i = 0; i < info.bindings.size(); i++) {
-        DEBUG_LOG("    {binding = %d, type = %s, count = %d, stage = %s, flags = %s},\n",
+        DEBUG_LOG("    {binding = %d, type = %s, count = %d, stage = %s, flags = %s}",
                info.bindings[i].binding, toString(info.bindings[i].type), info.bindings[i].count,
                toString(info.bindings[i].stageFlags), toString(info.bindings[i].flags));
+        DEBUG_LOG(",\n");
    }
    DEBUG_LOG("})\n");
    construct_handle<MetalDescriptorSetLayout>(dslh, std::move(info));
@@ -802,20 +782,9 @@ void MetalDriver::destroyRenderPrimitive(Handle<HwRenderPrimitive> rph) {
 }

 void MetalDriver::destroyProgram(Handle<HwProgram> ph) {
-    if (UTILS_UNLIKELY(!ph)) {
-        return;
+    if (ph) {
+        destruct_handle<MetalProgram>(ph);
    }
-    // Remove any cached pipeline states that refer to these programs.
-    auto* metalProgram = handle_cast<MetalProgram>(ph);
-    const auto& functions = metalProgram->getFunctionsIfPresent();
-    if (UTILS_LIKELY(functions.isRaster())) { // only raster pipelines are cached
-        const auto& raster = functions.getRasterFunctions();
-        mContext->pipelineStateCache.removeIf([&](const MetalPipelineState& state) {
-            const auto& [fragment, vertex] = raster;
-            return state.fragmentFunction == fragment || state.vertexFunction == vertex;
-        });
-    }
-    destruct_handle<MetalProgram>(ph);
 }

 void MetalDriver::destroyTexture(Handle<HwTexture> th) {
@@ -827,7 +796,15 @@ void MetalDriver::destroyTexture(Handle<HwTexture> th) {
    auto* metalTexture = handle_cast<MetalTexture>(th);
    mContext->textures.erase(metalTexture);

-    destruct_handle<MetalTexture>(th);
+    // Free memory from the texture and mark it as freed.
+    metalTexture->terminate();
+
+    // Add this texture handle to our texturesToDestroy queue to be destroyed later.
+    if (auto handleToFree = mContext->texturesToDestroy.push(th)) {
+        // If texturesToDestroy is full, then .push evicts the oldest texture handle in the
+        // queue (or simply th, if use-after-free detection is disabled).
+        destruct_handle<MetalTexture>(handleToFree.value());
+    }
 }

 void MetalDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
@@ -837,24 +814,17 @@ void MetalDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
 }

 void MetalDriver::destroySwapChain(Handle<HwSwapChain> sch) {
-    if (UTILS_UNLIKELY(!sch)) {
-        return;
-    }
-    auto* swapChain = handle_cast<MetalSwapChain>(sch);
-    if (mContext->currentDrawSwapChain == swapChain) {
-        mContext->currentDrawSwapChain = nullptr;
-    }
-    if (mContext->currentReadSwapChain == swapChain) {
-        mContext->currentReadSwapChain = nullptr;
-    }
-    // If the SwapChain is a pixel buffer, we need to wait for the current command buffer to
-    // complete before destroying it. This is because pixel buffer SwapChains hold a
-    // MetalExternalImage that could still being rendered into.
-    if (UTILS_UNLIKELY(swapChain->isPixelBuffer())) {
-        executeAfterCurrentCommandBufferCompletes(
-                [this, sch]() mutable { destruct_handle<MetalSwapChain>(sch); });
-    } else {
-        destruct_handle<MetalSwapChain>(sch);
+    if (sch) {
+        auto* swapChain = handle_cast<MetalSwapChain>(sch);
+        // If the SwapChain is a pixel buffer, we need to wait for the current command buffer to
+        // complete before destroying it. This is because pixel buffer SwapChains hold a
+        // MetalExternalImage that could still being rendered into.
+        if (UTILS_UNLIKELY(swapChain->isPixelBuffer())) {
+            executeAfterCurrentCommandBufferCompletes(
+                    [this, sch]() mutable { destruct_handle<MetalSwapChain>(sch); });
+        } else {
+            destruct_handle<MetalSwapChain>(sch);
+        }
    }
 }

@@ -877,23 +847,19 @@ void MetalDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> dslh)

 void MetalDriver::destroyDescriptorSet(Handle<HwDescriptorSet> dsh) {
    DEBUG_LOG("destroyDescriptorSet(dsh = %d)\n", dsh.getId());
-    if (!dsh) {
-        return;
+    if (dsh) {
+        executeAfterCurrentCommandBufferCompletes(
+                [this, dsh]() mutable { destruct_handle<MetalDescriptorSet>(dsh); });
    }
-
-    // Unbind this descriptor set.
-    auto* descriptorSet = handle_cast<MetalDescriptorSet>(dsh);
-    for (size_t i = 0; i < MAX_DESCRIPTOR_SET_COUNT; i++) {
-        if (UTILS_UNLIKELY(mContext->currentDescriptorSets[i] == descriptorSet)) {
-            mContext->currentDescriptorSets[i] = nullptr;
-        }
-    }
-
-    executeAfterCurrentCommandBufferCompletes(
-            [this, dsh]() mutable { destruct_handle<MetalDescriptorSet>(dsh); });
 }

 void MetalDriver::terminate() {
+    // Terminate any outstanding MetalTextures.
+    while (!mContext->texturesToDestroy.empty()) {
+        Handle<HwTexture> toDestroy = mContext->texturesToDestroy.pop();
+        destruct_handle<MetalTexture>(toDestroy);
+    }
+
    // finish() will flush the pending command buffer and will ensure all GPU work has finished.
    // This must be done before calling bufferPool->reset() to ensure no buffers are in flight.
    finish();
@@ -1068,7 +1034,7 @@ bool MetalDriver::isProtectedTexturesSupported() {
 }

 bool MetalDriver::isDepthClampSupported() {
-    return mContext->supportsDepthClamp;
+    return true;
 }

 bool MetalDriver::isWorkaroundNeeded(Workaround workaround) {
@@ -1117,34 +1083,24 @@ size_t MetalDriver::getMaxUniformBufferSize() {

 void MetalDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&& data,
        uint32_t byteOffset) {
-    FILAMENT_CHECK_PRECONDITION(data.buffer)
-            << "updateIndexBuffer called with a null buffer.";
    auto* ib = handle_cast<MetalIndexBuffer>(ibh);
-    ib->buffer.copyIntoBuffer(data.buffer, data.size, byteOffset,
-            [&]() { return mHandleAllocator.getHandleTag(ibh.getId()).c_str_safe(); });
+    ib->buffer.copyIntoBuffer(data.buffer, data.size, byteOffset);
    scheduleDestroy(std::move(data));
 }

 void MetalDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& data,
        uint32_t byteOffset) {
    FILAMENT_CHECK_PRECONDITION(!isInRenderPass(mContext))
-            << "updateBufferObject must be called outside of a render pass. tag="
-            << mHandleAllocator.getHandleTag(boh.getId()).c_str_safe();
-    FILAMENT_CHECK_PRECONDITION(data.buffer)
-            << "updateBufferObject called with a null buffer. tag="
-            << mHandleAllocator.getHandleTag(boh.getId()).c_str_safe();
+            << "updateBufferObject must be called outside of a render pass.";
    auto* bo = handle_cast<MetalBufferObject>(boh);
-
-    bo->updateBuffer(data.buffer, data.size, byteOffset,
-            [&]() { return mHandleAllocator.getHandleTag(boh.getId()).c_str_safe(); });
+    bo->updateBuffer(data.buffer, data.size, byteOffset);
    scheduleDestroy(std::move(data));
 }

 void MetalDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> boh,
        BufferDescriptor&& data, uint32_t byteOffset) {
    auto* bo = handle_cast<MetalBufferObject>(boh);
-    bo->updateBufferUnsynchronized(data.buffer, data.size, byteOffset,
-            [&]() { return mHandleAllocator.getHandleTag(boh.getId()).c_str_safe(); });
+    bo->updateBufferUnsynchronized(data.buffer, data.size, byteOffset);
    scheduleDestroy(std::move(data));
 }

@@ -1262,8 +1218,6 @@ void MetalDriver::beginRenderPass(Handle<HwRenderTarget> rth,
    mContext->depthStencilState.invalidate();
    mContext->cullModeState.invalidate();
    mContext->windingState.invalidate();
-    mContext->scissorRectState.invalidate();
-    mContext->depthClampState.invalidate();
    mContext->currentPolygonOffset = {0.0f, 0.0f};

    mContext->finalizedDescriptorSets.clear();
@@ -1642,7 +1596,6 @@ void MetalDriver::blitDEPRECATED(TargetBufferFlags buffers,
 void MetalDriver::bindPipeline(PipelineState const& ps) {
    FILAMENT_CHECK_PRECONDITION(mContext->currentRenderPassEncoder != nullptr)
            << "bindPipeline() without a valid command encoder.";
-    DEBUG_LOG("bindPipeline(ps = { program = %d }))\n", ps.program.getId());

    MetalVertexBufferInfo const* const vbi =
            handle_cast<MetalVertexBufferInfo>(ps.vertexBufferInfo);
@@ -1735,12 +1688,10 @@ void MetalDriver::bindPipeline(PipelineState const& ps) {
    }

    // depth clip mode
-    if (mContext->supportsDepthClamp) {
-        MTLDepthClipMode depthClipMode = rs.depthClamp ? MTLDepthClipModeClamp : MTLDepthClipModeClip;
-        mContext->depthClampState.updateState(depthClipMode);
-        if (mContext->depthClampState.stateChanged()) {
-            [mContext->currentRenderPassEncoder setDepthClipMode:depthClipMode];
-        }
+    MTLDepthClipMode depthClipMode = rs.depthClamp ? MTLDepthClipModeClamp : MTLDepthClipModeClip;
+    mContext->depthClampState.updateState(depthClipMode);
+    if (mContext->depthClampState.stateChanged()) {
+        [mContext->currentRenderPassEncoder setDepthClipMode:depthClipMode];
    }

    // Set the depth-stencil state, if a state change is needed.
@@ -1832,21 +1783,11 @@ void MetalDriver::bindDescriptorSet(
        backend::DescriptorSetHandle dsh,
        backend::descriptor_set_t set,
        backend::DescriptorSetOffsetArray&& offsets) {
-
-    if (UTILS_UNLIKELY(!dsh)) {
-        DEBUG_LOG("bindDescriptorSet(dsh = null, set = %d, offsets = [])\n", set);
-        mContext->currentDescriptorSets[set] = nullptr;
-        mContext->vertexDescriptorBindings.setBuffer(nil, 0, set);
-        mContext->fragmentDescriptorBindings.setBuffer(nil, 0, set);
-        mContext->dynamicOffsets.setOffsets(set, nullptr, 0);
-        return;
-    }
-
    auto descriptorSet = handle_cast<MetalDescriptorSet>(dsh);
    const size_t dynamicBindings = descriptorSet->layout->getDynamicOffsetCount();
    utils::FixedCapacityVector<size_t> offsetsVector(dynamicBindings, 0);
 #if FILAMENT_METAL_DEBUG_LOG == 1
-    printf("[METAL DEBUG] bindDescriptorSet(dsh = %d, set = %d, offsets = [", dsh.getId(), set);
+    DEBUG_LOG("bindDescriptorSet(dsh = %d, set = %d, offsets = [", dsh.getId(), set);
    for (size_t i = 0; i < dynamicBindings; i++) {
        printf("%d", offsets[i]);
        if (i < dynamicBindings - 1) {
@@ -1855,7 +1796,7 @@ void MetalDriver::bindDescriptorSet(

        offsetsVector[i] = offsets[i];
    }
-    printf("])\n");
+    DEBUG_LOG("])\n");
 #endif

    // Bind the descriptor set.
@@ -2002,11 +1943,7 @@ void MetalDriver::scissor(Viewport scissorBox) {
            .height = static_cast<NSUInteger>(bottom - top)
    };

-    auto& srs = mContext->scissorRectState;
-    srs.updateState(scissorRect);
-    if (srs.stateChanged()) {
-        [mContext->currentRenderPassEncoder setScissorRect:scissorRect];
-    }
+    [mContext->currentRenderPassEncoder setScissorRect:scissorRect];
 }

 void MetalDriver::beginTimerQuery(Handle<HwTimerQuery> tqh) {
@@ -71,8 +71,7 @@ constexpr inline MTLIndexType getIndexType(size_t elementSize) noexcept {
    } else if (elementSize == 4) {
        return MTLIndexTypeUInt32;
    }
-    assert_invariant(false);
-    return MTLIndexTypeUInt16;
+    FILAMENT_CHECK_POSTCONDITION(false) << "Index element size not supported.";
 }

 constexpr inline MTLVertexFormat getMetalFormat(ElementType type, bool normalized) noexcept {
@@ -74,8 +74,7 @@ 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);

@@ -124,7 +123,6 @@ private:
    struct {
        CallbackHandler* handler = nullptr;
        std::shared_ptr<FrameScheduledCallback> callback = nullptr;
-        uint64_t flags = 0;
    } frameScheduled;

    struct {
@@ -135,15 +133,11 @@ 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; }

 private:
@@ -202,7 +196,6 @@ public:
    MetalProgram(MetalContext& context, Program&& program) noexcept;

    const MetalShaderCompiler::MetalFunctionBundle& getFunctions();
-    const MetalShaderCompiler::MetalFunctionBundle& getFunctionsIfPresent() const;

 private:
    void initialize();
@@ -272,6 +265,26 @@ 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;
@@ -288,6 +301,8 @@ private:
    // 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;
+
+    bool terminated = false;
 };

 class MetalRenderTarget : public HwRenderTarget {
@@ -438,7 +453,7 @@ struct MetalTimerQuery : public HwTimerQuery {

    struct Status {
        std::atomic<bool> available {false};
-        std::atomic<uint64_t> elapsed {0};   // only valid if available is true
+        uint64_t elapsed {0};   // only valid if available is true
    };

    std::shared_ptr<Status> status;
@@ -496,7 +511,7 @@ struct MetalDescriptorSet : public HwDescriptorSet {

    std::vector<std::shared_ptr<MetalExternalImage>> externalImages;

-    std::array<TrackedMetalBuffer, Program::SHADER_TYPE_COUNT> cachedBuffer = { nil };
+    std::array<id<MTLBuffer>, Program::SHADER_TYPE_COUNT> cachedBuffer = { nil };
 };

 } // namespace backend
@@ -229,10 +229,9 @@ 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(
@@ -261,10 +260,10 @@ 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 +271,16 @@ 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);
-            });
-        }
+        // 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); });
    }

 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 +295,6 @@ 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 +311,8 @@ 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 +327,14 @@ 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 +369,12 @@ 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,
@@ -495,10 +479,6 @@ const MetalShaderCompiler::MetalFunctionBundle& MetalProgram::getFunctions() {
    return mFunctionBundle;
 }

-const MetalShaderCompiler::MetalFunctionBundle& MetalProgram::getFunctionsIfPresent() const {
-    return mFunctionBundle;
-}
-
 void MetalProgram::initialize() {
    if (!mToken) {
        return;
@@ -654,6 +634,14 @@ MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t
    texture = externalImage->getMtlTexture();
 }

+void MetalTexture::terminate() noexcept {
+    texture = nil;
+    swizzledTextureView = nil;
+    msaaSidecar = nil;
+    externalImage = nullptr;
+    terminated = true;
+}
+
 id<MTLTexture> MetalTexture::getMtlTextureForRead() const noexcept {
    return swizzledTextureView ? swizzledTextureView : texture;
 }
@@ -1294,8 +1282,7 @@ id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoderSlow(id<MTLDe
                [arguments addObject:bufferArgument];
                break;
            }
-            case DescriptorType::SAMPLER:
-            case DescriptorType::SAMPLER_EXTERNAL: {
+            case DescriptorType::SAMPLER: {
                MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
                textureArgument.index = binding.binding * 2;
                textureArgument.dataType = MTLDataTypeTexture;
@@ -1358,7 +1345,7 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
    auto& buffer = cachedBuffer[index];

    if (buffer) {
-        return buffer.get();
+        return buffer;
    }

    // Map all the texture bindings to their respective texture types.
@@ -1381,13 +1368,9 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
    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];
+    buffer = [context.device newBufferWithLength:encoder.encodedLength
+                                         options:MTLResourceStorageModeShared];
+    [encoder setArgumentBuffer:buffer offset:0];

    for (auto const& binding : bindings) {
        if (!hasShaderType(binding.stageFlags, stage)) {
@@ -1410,8 +1393,7 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
                           atIndex:binding.binding * 2];
                break;
            }
-            case DescriptorType::SAMPLER:
-            case DescriptorType::SAMPLER_EXTERNAL: {
+            case DescriptorType::SAMPLER: {
                auto found = textures.find(binding.binding);
                if (found == textures.end()) {
                    [encoder setTexture:driver->mContext->emptyTexture atIndex:binding.binding * 2];
@@ -1435,7 +1417,7 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
        }
    }

-    return buffer.get();
+    return buffer;
 }

 } // namespace backend
@@ -83,10 +83,6 @@ public:
            return std::get<Compute>(mPrograms);
        }

-        bool isRaster() const { return std::holds_alternative<Raster>(mPrograms); }
-
-        bool isCompute() const { return std::holds_alternative<Compute>(mPrograms); }
-
        static MetalFunctionBundle none() {
            return MetalFunctionBundle(None{});
        }
@@ -226,11 +226,9 @@ MetalShaderCompiler::program_token_t MetalShaderCompiler::createProgram(
            CompilerPriorityQueue const priorityQueue = program.getPriorityQueue();
            mCompilerThreadPool.queue(priorityQueue, token,
                    [this, name, device = mDevice, program = std::move(program), token]() {
-                        @autoreleasepool {
-                            MetalFunctionBundle compiledProgram = compileProgram(program, device);
-                            token->set(compiledProgram);
-                            mCallbackManager.put(token->handle);
-                        }
+                        MetalFunctionBundle compiledProgram = compileProgram(program, device);
+                        token->set(compiledProgram);
+                        mCallbackManager.put(token->handle);
                    });

            break;
@@ -28,9 +28,7 @@

 #include <memory>
 #include <tsl/robin_map.h>
-
 #include <utils/Hash.h>
-#include <utils/Invocable.h>

 namespace filament {
 namespace backend {
@@ -162,8 +160,6 @@ template<typename StateType,
         typename HashFn = utils::hash::MurmurHashFn<StateType>>
 class StateCache {

-    using MapType = tsl::robin_map<StateType, MetalType, HashFn>;
-
 public:

    StateCache() = default;
@@ -173,18 +169,6 @@ public:

    void setDevice(id<MTLDevice> device) noexcept { mDevice = device; }

-    void removeIf(utils::Invocable<bool(const StateType&)> fn) noexcept {
-        typename MapType::const_iterator it = mStateCache.begin();
-        while (it != mStateCache.end()) {
-            const auto& [key, _] = *it;
-            if (UTILS_UNLIKELY(fn(key))) {
-                it = mStateCache.erase(it);
-            } else {
-                ++it;
-            }
-        }
-    }
-
    MetalType getOrCreateState(const StateType& state) noexcept {
        assert_invariant(mDevice);

@@ -212,7 +196,7 @@ private:
    StateCreator creator;
    id<MTLDevice> mDevice = nil;

-    MapType mStateCache;
+    tsl::robin_map<StateType, MetalType, HashFn> mStateCache;

 };

@@ -220,8 +204,9 @@ private:
 // Different kinds of state, like pipeline state, uniform buffer state, etc., are passed to the
 // current Metal command encoder and persist throughout the lifetime of the encoder (a frame).
 // StateTracker is used to prevent calling redundant state change methods.
-template <typename StateType, typename StateEqual = std::equal_to<StateType>>
+template<typename StateType>
 class StateTracker {
+
 public:

    // Call to force the state to dirty at the beginning of each frame, as all state must be
@@ -229,7 +214,7 @@ public:
    void invalidate() noexcept { mStateDirty = true; }

    void updateState(const StateType& newState) noexcept {
-        if (!StateEqual()(mCurrentState, newState)) {
+        if (mCurrentState != newState) {
            mCurrentState = newState;
            mStateDirty = true;
        }
@@ -250,6 +235,7 @@ private:

    bool mStateDirty = true;
    StateType mCurrentState = {};
+
 };

 // Pipeline state
@@ -357,16 +343,6 @@ using DepthStencilStateTracker = StateTracker<DepthStencilState>;
 using DepthStencilStateCache = StateCache<DepthStencilState, id<MTLDepthStencilState>,
        DepthStateCreator>;

-struct MtlScissorRectEqual {
-    bool operator()(const MTLScissorRect& lhs, const MTLScissorRect& rhs) const {
-        return lhs.height == rhs.height &&
-            lhs.width == rhs.width &&
-            lhs.x == rhs.x &&
-            lhs.y == rhs.y;
-    }
-};
-using ScissorRectStateTracker = StateTracker<MTLScissorRect, MtlScissorRectEqual>;
-
 // Uniform buffers

 class MetalBufferObject;
@@ -458,7 +434,7 @@ public:

 private:
    static_assert(N <= 8);
-    std::array<__weak id<MTLBuffer>, N> mBuffers = { nil };
+    std::array<__unsafe_unretained id<MTLBuffer>, N> mBuffers = { nil };
    std::array<NSUInteger, N> mOffsets = { 0 };
    utils::bitset8 mDirtyBuffers;
    utils::bitset8 mDirtyOffsets;
@@ -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) {

 }

@@ -69,9 +69,7 @@ public:
        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 ||
-                                   entry.type == DescriptorType::SAMPLER_EXTERNAL };
+        mStorage[set][binding] = { (uint8_t)entry.binding, entry.type == DescriptorType::SAMPLER };
        mActiveDescriptors[set].set(binding);
    }

@@ -67,9 +67,7 @@ GLDescriptorSet::GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle ds
                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
                dynamicBuffers.set(index, dynamicOffset);
                if (UTILS_UNLIKELY(gl.isES2())) {
-                    if (dynamicOffset) {
-                        dynamicBufferCount++;
-                    }
+                    dynamicBufferCount++;
                    desc.emplace<BufferGLES2>(dynamicOffset);
                } else {
                    auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::UNIFORM);
@@ -96,7 +94,6 @@ GLDescriptorSet::GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle ds
                break;
            }
            case DescriptorType::SAMPLER:
-            case DescriptorType::SAMPLER_EXTERNAL:
                if (UTILS_UNLIKELY(gl.isES2())) {
                    desc.emplace<SamplerGLES2>();
                } else {
@@ -78,12 +78,8 @@ struct GLDescriptorSet : public HwDescriptorSet {
 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) {}
+        Buffer() = default;
+        explicit Buffer(GLenum target) noexcept : target(target) { }
        GLenum target;                          // 4
        GLuint id = 0;                          // 4
        uint32_t offset = 0;                    // 4
@@ -500,13 +500,11 @@ void OpenGLContext::initBugs(Bugs* bugs, Extensions const& exts,
            // qualcomm seems to have no problem with this (which is good for us)
            bugs->allow_read_only_ancillary_feedback_loop = true;

-#ifndef __EMSCRIPTEN__
            // Older Adreno devices that support ES3.0 only tend to be extremely buggy, so we
            // fall back to ES2.0.
            if (major == 3 && minor == 0) {
                bugs->force_feature_level0 = true;
            }
-#endif
        } else if (strstr(renderer, "Mali")) {
            // ARM GPU
            bugs->vao_doesnt_store_element_array_buffer_binding = true;
@@ -554,14 +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;
        }
@@ -616,7 +606,7 @@ FeatureLevel OpenGLContext::resolveFeatureLevel(GLint major, GLint minor,
                featureLevel = FeatureLevel::FEATURE_LEVEL_2;
                if (gets.max_texture_image_units >= MAX_FRAGMENT_SAMPLER_COUNT &&
                    gets.max_combined_texture_image_units >=
-                            (MAX_FRAGMENT_SAMPLER_COUNT + MAX_VERTEX_SAMPLER_COUNT)) {
+                    (MAX_FRAGMENT_SAMPLER_COUNT + MAX_VERTEX_SAMPLER_COUNT)) {
                    featureLevel = FeatureLevel::FEATURE_LEVEL_3;
                }
            }
@@ -625,13 +615,15 @@ FeatureLevel OpenGLContext::resolveFeatureLevel(GLint major, GLint minor,
 #   ifndef IOS // IOS is guaranteed to have ES3.x
    else if (UTILS_UNLIKELY(major == 2)) {
        // Runtime OpenGL version is ES 2.x
-        // note: mandatory extensions (all supported by Mali-400 and Adreno 304)
-        //      OES_depth_texture
-        //      OES_depth24
-        //      OES_packed_depth_stencil
-        //      OES_rgb8_rgba8
-        //      OES_standard_derivatives
-        //      OES_texture_npot
+#       if defined(BACKEND_OPENGL_LEVEL_GLES30)
+        // mandatory extensions (all supported by Mali-400 and Adreno 304)
+        assert_invariant(exts.OES_depth_texture);
+        assert_invariant(exts.OES_depth24);
+        assert_invariant(exts.OES_packed_depth_stencil);
+        assert_invariant(exts.OES_rgb8_rgba8);
+        assert_invariant(exts.OES_standard_derivatives);
+        assert_invariant(exts.OES_texture_npot);
+#       endif
        featureLevel = FeatureLevel::FEATURE_LEVEL_0;
    }
 #   endif // IOS
@@ -937,19 +929,15 @@ 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 +946,16 @@ 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;
+                }
            }
        }
    }
@@ -190,7 +190,7 @@ 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;
@@ -316,15 +316,10 @@ 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.
@@ -559,9 +554,6 @@ private:
            {   bugs.delay_fbo_destruction,
                    "delay_fbo_destruction",
                    ""},
-            {   bugs.rebind_buffer_after_deletion,
-                    "rebind_buffer_after_deletion",
-                    ""},
            {   bugs.force_feature_level0,
                    "force_feature_level0",
                    ""},
@@ -23,7 +23,6 @@
 #include "OpenGLDriverFactory.h"
 #include "OpenGLProgram.h"
 #include "OpenGLTimerQuery.h"
-#include "SystraceProfile.h"
 #include "gl_headers.h"

 #include <backend/platforms/OpenGLPlatform.h>
@@ -92,47 +91,24 @@
 #define DEBUG_GROUP_MARKER_NONE       0x00    // no debug marker
 #define DEBUG_GROUP_MARKER_OPENGL     0x01    // markers in the gl command queue (req. driver support)
 #define DEBUG_GROUP_MARKER_BACKEND    0x02    // markers on the backend side (systrace)
-#define DEBUG_GROUP_MARKER_ALL        0xFF    // all markers
+#define DEBUG_GROUP_MARKER_ALL        0x03    // all markers

 #define DEBUG_MARKER_NONE             0x00    // no debug marker
 #define DEBUG_MARKER_OPENGL           0x01    // markers in the gl command queue (req. driver support)
 #define DEBUG_MARKER_BACKEND          0x02    // markers on the backend side (systrace)
-#define DEBUG_MARKER_PROFILE          0x04    // profiling on the backend side (systrace)
-#define DEBUG_MARKER_ALL              (0xFF & ~DEBUG_MARKER_PROFILE) // all markers
+#define DEBUG_MARKER_ALL              0x03    // all markers

 // set to the desired debug marker level (for user markers [default: All])
 #define DEBUG_GROUP_MARKER_LEVEL      DEBUG_GROUP_MARKER_ALL

 // set to the desired debug level (for internal debugging [Default: None])
-#define DEBUG_MARKER_LEVEL            DEBUG_MARKER_NONE
+#define DEBUG_MARKER_LEVEL      DEBUG_MARKER_NONE

-// Override the debug markers if we are forcing profiling mode
-#if defined(FILAMENT_FORCE_PROFILING_MODE)
-#   undef DEBUG_GROUP_MARKER_LEVEL
-#   undef DEBUG_MARKER_LEVEL
-
-#   define DEBUG_GROUP_MARKER_LEVEL   DEBUG_GROUP_MARKER_NONE
-#   define DEBUG_MARKER_LEVEL         DEBUG_MARKER_PROFILE
-#endif
-
-#if DEBUG_MARKER_LEVEL == DEBUG_MARKER_PROFILE
-#   define DEBUG_MARKER()
-#   define PROFILE_MARKER(marker) PROFILE_SCOPE(marker);
-#   if DEBUG_GROUP_MARKER_LEVEL != DEBUG_GROUP_MARKER_NONE
-#      error PROFILING is exclusive; group markers must be disabled.
-#   endif
-#   ifndef NDEBUG
-#      error PROFILING is meaningless in DEBUG mode.
-#   endif
-#elif DEBUG_MARKER_LEVEL > DEBUG_MARKER_NONE
-#   define DEBUG_MARKER() DebugMarker _debug_marker(*this, __func__);
-#   define PROFILE_MARKER(marker) DEBUG_MARKER()
-#   if DEBUG_MARKER_LEVEL & DEBUG_MARKER_PROFILE
-#      error PROFILING is exclusive; all other debug features must be disabled.
-#   endif
+#if DEBUG_MARKER_LEVEL > DEBUG_MARKER_NONE
+#   define DEBUG_MARKER() \
+        DebugMarker _debug_marker(*this, __func__);
 #else
 #   define DEBUG_MARKER()
-#   define PROFILE_MARKER(marker)
 #endif

 using namespace filament::math;
@@ -406,21 +382,24 @@ void OpenGLDriver::bindTexture(GLuint unit, GLTexture const* t) noexcept {
 }

 bool OpenGLDriver::useProgram(OpenGLProgram* p) noexcept {
-    bool success = true;
-    if (mBoundProgram != p) {
-        // compile/link the program if needed and call glUseProgram
-        success = p->use(this, mContext);
-        assert_invariant(success == p->isValid());
-        if (success) {
-            // TODO: we could even improve this if the program could tell us which of the descriptors
-            //       bindings actually changed. In practice, it is likely that set 0 or 1 might not
-            //       change often.
-            decltype(mInvalidDescriptorSetBindings) changed;
-            changed.setValue((1 << MAX_DESCRIPTOR_SET_COUNT) - 1);
-            mInvalidDescriptorSetBindings |= changed;
+    if (UTILS_UNLIKELY(mBoundProgram == p)) {
+        // program didn't change, don't do anything.
+        return true;
+    }

-            mBoundProgram = p;
-        }
+    // compile/link the program if needed and call glUseProgram
+    bool const success = p->use(this, mContext);
+    assert_invariant(success == p->isValid());
+  
+    if (success) {
+        // TODO: we could even improve this if the program could tell us which of the descriptors
+        //       bindings actually changed. In practice, it is likely that set 0 or 1 might not
+        //       change often.
+        decltype(mInvalidDescriptorSetBindings) changed;
+        changed.setValue((1 << MAX_DESCRIPTOR_SET_COUNT) - 1);
+        mInvalidDescriptorSetBindings |= changed;
+
+        mBoundProgram = p;
    }

    if (UTILS_UNLIKELY(mContext.isES2() && success)) {
@@ -1287,7 +1266,6 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
    if (any(t->usage & TextureUsage::SAMPLEABLE)) {
        switch (t->target) {
            case SamplerType::SAMPLER_2D:
-            case SamplerType::SAMPLER_3D:
            case SamplerType::SAMPLER_2D_ARRAY:
            case SamplerType::SAMPLER_CUBEMAP_ARRAY:
                // this could be GL_TEXTURE_2D_MULTISAMPLE or GL_TEXTURE_2D_ARRAY
@@ -1298,9 +1276,7 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
                target = getCubemapTarget(binfo.layer);
                // note: cubemaps can't be multi-sampled
                break;
-            case SamplerType::SAMPLER_EXTERNAL:
-                // This is an error. We have asserted in debug build.
-                target = t->gl.target;
+            default:
                break;
        }
    }
@@ -1318,7 +1294,7 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,

    if (rt->gl.samples <= 1 ||
        (rt->gl.samples > 1 && t->samples > 1 && gl.features.multisample_texture)) {
-        // On GL3.2 / GLES3.1 and above multisample is handled when creating the texture.
+        // on GL3.2 / GLES3.1 and above multisample is handled when creating the texture.
        // If multisampled textures are not supported and we end-up here, things should
        // still work, albeit without MSAA.
        gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
@@ -1342,7 +1318,6 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
                            GL_RENDERBUFFER, t->gl.id);
                }
                break;
-            case GL_TEXTURE_3D:
            case GL_TEXTURE_2D_ARRAY:
            case GL_TEXTURE_CUBE_MAP_ARRAY:
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
@@ -1352,7 +1327,7 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
                if (layerCount > 1) {
                    // if layerCount > 1, it means we use the multiview extension.
                    glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, attachment,
-                        t->gl.id, 0, binfo.layer, layerCount);
+                        t->gl.id, 0, binfo.baseViewIndex, layerCount);
                } else
 #endif // !defined(__EMSCRIPTEN__) && !defined(IOS)
                {
@@ -1456,7 +1431,6 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
                            GL_RENDERBUFFER, t->gl.id);
                }
                break;
-            case GL_TEXTURE_3D:
            case GL_TEXTURE_2D_ARRAY:
            case GL_TEXTURE_CUBE_MAP_ARRAY:
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
@@ -1682,7 +1656,7 @@ void OpenGLDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,

    // See if we need the emulated rec709 output conversion
    if (UTILS_UNLIKELY(mContext.isES2())) {
-        sc->rec709 = ((flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) &&
+        sc->rec709 = (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE &&
                !mPlatform.isSRGBSwapChainSupported());
    }
 }
@@ -1753,7 +1727,7 @@ void OpenGLDriver::destroyIndexBuffer(Handle<HwIndexBuffer> ibh) {
    if (ibh) {
        auto& gl = mContext;
        GLIndexBuffer const* ib = handle_cast<const GLIndexBuffer*>(ibh);
-        gl.deleteBuffer(ib->gl.buffer, GL_ELEMENT_ARRAY_BUFFER);
+        gl.deleteBuffers(1, &ib->gl.buffer, GL_ELEMENT_ARRAY_BUFFER);
        destruct(ibh, ib);
    }
 }
@@ -1766,7 +1740,7 @@ void OpenGLDriver::destroyBufferObject(Handle<HwBufferObject> boh) {
        if (UTILS_UNLIKELY(bo->bindingType == BufferObjectBinding::UNIFORM && gl.isES2())) {
            free(bo->gl.buffer);
        } else {
-            gl.deleteBuffer(bo->gl.id, bo->gl.binding);
+            gl.deleteBuffers(1, &bo->gl.id, bo->gl.binding);
        }
        destruct(boh, bo);
    }
@@ -2935,9 +2909,6 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
    GLuint const fbo = gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)

-    // each render-pass starts with a disabled scissor
-    gl.disable(GL_SCISSOR_TEST);
-
    if (gl.ext.EXT_discard_framebuffer
            && !gl.bugs.disable_invalidate_framebuffer) {
        AttachmentArray attachments; // NOLINT
@@ -2950,6 +2921,7 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
        // It's important to clear the framebuffer before drawing, as it resets
        // the fb to a known state (resets fb compression and possibly other things).
        // So we use glClear instead of glInvalidateFramebuffer
+        gl.disable(GL_SCISSOR_TEST);
        clearWithRasterPipe(discardFlags & ~clearFlags, { 0.0f }, 0.0f, 0);
    }

@@ -2965,6 +2937,7 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
    }

    if (any(clearFlags)) {
+        gl.disable(GL_SCISSOR_TEST);
        clearWithRasterPipe(clearFlags,
                params.clearColor, (GLfloat)params.clearDepth, (GLint)params.clearStencil);
    }
@@ -2983,6 +2956,7 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,

 #ifndef NDEBUG
    // clear the discarded (but not the cleared ones) buffers in debug builds
+    gl.disable(GL_SCISSOR_TEST);
    clearWithRasterPipe(discardFlags & ~clearFlags,
            { 1, 0, 0, 1 }, 1.0, 0);
 #endif
@@ -3469,7 +3443,7 @@ void OpenGLDriver::beginFrame(
        UTILS_UNUSED int64_t monotonic_clock_ns,
        UTILS_UNUSED int64_t refreshIntervalNs,
        UTILS_UNUSED uint32_t frameId) {
-    PROFILE_MARKER(PROFILE_NAME_BEGINFRAME)
+    DEBUG_MARKER()
    auto& gl = mContext;
    insertEventMarker("beginFrame");
    mPlatform.beginFrame(monotonic_clock_ns, refreshIntervalNs, frameId);
@@ -3488,8 +3462,8 @@ void OpenGLDriver::beginFrame(
    }
 }

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

@@ -3504,7 +3478,7 @@ void OpenGLDriver::setPresentationTime(int64_t monotonic_clock_ns) {
 }

 void OpenGLDriver::endFrame(UTILS_UNUSED uint32_t frameId) {
-    PROFILE_MARKER(PROFILE_NAME_ENDFRAME)
+    DEBUG_MARKER()
 #if defined(__EMSCRIPTEN__)
    // WebGL builds are single-threaded so users might manipulate various GL state after we're
    // done with the frame. We do NOT officially support using Filament in this way, but we can
@@ -3896,14 +3870,6 @@ void OpenGLDriver::bindDescriptorSet(
        backend::DescriptorSetHandle dsh,
        backend::descriptor_set_t set,
        backend::DescriptorSetOffsetArray&& offsets) {
-
-    if (UTILS_UNLIKELY(!dsh)) {
-        mBoundDescriptorSets[set].dsh = dsh;
-        mInvalidDescriptorSetBindings.set(set, true);
-        mInvalidDescriptorSetBindingOffsets.set(set, true);
-        return;
-    }
-
    // handle_cast<> here also serves to validate the handle (it actually cannot return nullptr)
    GLDescriptorSet const* const ds = handle_cast<GLDescriptorSet*>(dsh);
    if (ds) {
@@ -3921,7 +3887,6 @@ void OpenGLDriver::bindDescriptorSet(
        // `offsets` data's lifetime will end when this function returns. We have to make a copy.
        // (the data is allocated inside the CommandStream)
        mBoundDescriptorSets[set].dsh = dsh;
-        assert_invariant(offsets.data() != nullptr || ds->getDynamicBufferCount() == 0);
        std::copy_n(offsets.data(), ds->getDynamicBufferCount(),
                mBoundDescriptorSets[set].offsets.data());
    }
@@ -174,8 +174,7 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
                    }
                    break;
                }
-                case DescriptorType::SAMPLER:
-                case DescriptorType::SAMPLER_EXTERNAL: {
+                case DescriptorType::SAMPLER: {
                    if (!entry.name.empty()) {
                        GLint const loc = glGetUniformLocation(program, entry.name.c_str());
                        if (loc >= 0) {
@@ -255,9 +254,7 @@ void OpenGLProgram::updateUniforms(
    for (size_t i = 0, c = records.uniforms.size(); i < c; i++) {
        Program::Uniform const& u = records.uniforms[i];
        GLint const loc = records.locations[i];
-        // mRec709Location is special, it is handled by setRec709ColorSpace() and the corresponding
-        // entry in `buffer` is typically not initialized, so we skip it.
-        if (loc < 0 || loc == mRec709Location) {
+        if (loc < 0) {
            continue;
        }
        // u.offset is in 'uint32_t' units
@@ -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;
@@ -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;
 }

@@ -1,155 +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;
-            }
-        }
-        if (mLib) {
-            OSMesaGetProcAddress = (GetProcAddressFunc) dlsym(mLib, "OSMesaGetProcAddress");
-        } else {
-            OSMesaGetProcAddress = (GetProcAddressFunc) dlsym(RTLD_LOCAL, "OSMesaGetProcAddress");
-        }
-
-        FILAMENT_CHECK_PRECONDITION(OSMesaGetProcAddress)
-                << "Unable to against libOSMesa to create a software GL context";
-
-        OSMesaCreateContext = (CreateContextFunc) OSMesaGetProcAddress("OSMesaCreateContext");
-        OSMesaDestroyContext =
-                (DestroyContextFunc) OSMesaGetProcAddress("OSMesaDestroyContext");
-        OSMesaMakeCurrent = (MakeCurrentFunc) OSMesaGetProcAddress("OSMesaMakeCurrent");
-    }
-
-    ~OSMesaAPI() {
-        if (mLib) {
-            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
@@ -19,8 +19,8 @@
 #include <Wingdi.h>

 #ifdef _MSC_VER
-    // this variable is checked in BlueGL.h (included from "gl_headers.h" right after this),
-    // and prevents duplicate definition of OpenGL apis when building this file.
+    // this variable is checked in BlueGL.h (included from "gl_headers.h" right after this), 
+    // and prevents duplicate definition of OpenGL apis when building this file. 
    // However, GL_GLEXT_PROTOTYPES need to be defined in BlueGL.h when included from other files.
    #define FILAMENT_PLATFORM_WGL
 #endif
@@ -37,8 +37,9 @@

 namespace {

-void reportWindowsError(DWORD dwError) {
+void reportLastWindowsError() {
    LPSTR lpMessageBuffer = nullptr;
+    DWORD dwError = GetLastError();

    if (dwError == 0) {
        return;
@@ -79,7 +80,6 @@ Driver* PlatformWGL::createDriver(void* const sharedGLContext,
        const Platform::DriverConfig& driverConfig) noexcept {
    int result = 0;
    int pixelFormat = 0;
-    DWORD dwError = 0;

    mPfd = {
        sizeof(PIXELFORMATDESCRIPTOR),
@@ -105,7 +105,6 @@ Driver* PlatformWGL::createDriver(void* const sharedGLContext,
    mHWnd = CreateWindowA("STATIC", "dummy", 0, 0, 0, 1, 1, NULL, NULL, NULL, NULL);
    HDC whdc = mWhdc = GetDC(mHWnd);
    if (whdc == NULL) {
-        dwError = GetLastError();
        utils::slog.e << "CreateWindowA() failed" << utils::io::endl;
        goto error;
    }
@@ -116,7 +115,6 @@ Driver* PlatformWGL::createDriver(void* const sharedGLContext,
    // We need a tmp context to retrieve and call wglCreateContextAttribsARB.
    tempContext = wglCreateContext(whdc);
    if (!wglMakeCurrent(whdc, tempContext)) {
-        dwError = GetLastError();
        utils::slog.e << "wglMakeCurrent() failed, whdc=" << whdc << ", tempContext=" <<
                tempContext << utils::io::endl;
        goto error;
@@ -138,7 +136,6 @@ Driver* PlatformWGL::createDriver(void* const sharedGLContext,
        if (mContext) {
            break;
        }
-        dwError = GetLastError();
    }

    if (!mContext) {
@@ -151,7 +148,6 @@ Driver* PlatformWGL::createDriver(void* const sharedGLContext,
    tempContext = NULL;

    if (!wglMakeCurrent(whdc, mContext)) {
-        dwError = GetLastError();
        utils::slog.e << "wglMakeCurrent() failed, whdc=" << whdc << ", mContext=" <<
                mContext << utils::io::endl;
        goto error;
@@ -166,7 +162,7 @@ error:
    if (tempContext) {
        wglDeleteContext(tempContext);
    }
-    reportWindowsError(dwError);
+    reportLastWindowsError();
    terminate();
    return NULL;
 }
@@ -209,11 +205,9 @@ Platform::SwapChain* PlatformWGL::createSwapChain(void* nativeWindow, uint64_t f
    // on Windows, the nativeWindow maps to a HWND
    swapChain->hWnd = (HWND) nativeWindow;
    swapChain->hDc = GetDC(swapChain->hWnd);
-    if (!swapChain->hDc) {
-        DWORD dwError = GetLastError();
-        ASSERT_POSTCONDITION_NON_FATAL(swapChain->hDc,
-           "Unable to create the SwapChain (nativeWindow = %p)", nativeWindow);
-        reportWindowsError(dwError);
+    if (!ASSERT_POSTCONDITION_NON_FATAL(swapChain->hDc,
+            "Unable to create the SwapChain (nativeWindow = %p)", nativeWindow)) {
+        reportLastWindowsError();
    }

 	// We have to match pixel formats across the HDC and HGLRC (mContext)
@@ -270,10 +264,8 @@ bool PlatformWGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
    HDC hdc = wglSwapChain->hDc;
    if (hdc != NULL) {
        BOOL success = wglMakeCurrent(hdc, mContext);
-        if (!success) {
-            DWORD dwError = GetLastError();
-            ASSERT_POSTCONDITION_NON_FATAL(success, "wglMakeCurrent() failed. hdc = %p", hdc);
-            reportWindowsError(dwError);
+        if (!ASSERT_POSTCONDITION_NON_FATAL(success, "wglMakeCurrent() failed. hdc = %p", hdc)) {
+            reportLastWindowsError();
            wglMakeCurrent(0, NULL);
        }
    }
@@ -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 << "}";
 }
@@ -1,43 +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 "VulkanAsyncHandles.h"
-
-namespace filament::backend {
-
-VulkanTimerQuery::VulkanTimerQuery(std::tuple<uint32_t, uint32_t> indices)
-    : mStartingQueryIndex(std::get<0>(indices)),
-      mStoppingQueryIndex(std::get<1>(indices)) {}
-
-void VulkanTimerQuery::setFence(std::shared_ptr<VulkanCmdFence> fence) noexcept {
-    std::unique_lock<utils::Mutex> lock(mFenceMutex);
-    mFence = fence;
-}
-
-bool VulkanTimerQuery::isCompleted() noexcept {
-    std::unique_lock<utils::Mutex> lock(mFenceMutex);
-    // QueryValue is a synchronous call and might occur before beginTimerQuery has written anything
-    // into the command buffer, which is an error according to the validation layer that ships in
-    // the Android NDK.  Even when AVAILABILITY_BIT is set, validation seems to require that the
-    // 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;
-}
-
-VulkanTimerQuery::~VulkanTimerQuery() = default;
-
-} // namespace filament::backend
@@ -1,81 +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_VULKANASYNCHANDLES_H
-#define TNT_FILAMENT_BACKEND_VULKANASYNCHANDLES_H
-
-#include <bluevk/BlueVK.h>
-
-#include "DriverBase.h"
-
-#include "vulkan/memory/Resource.h"
-
-#include <utils/Mutex.h>
-#include <utils/Condition.h>
-
-namespace filament::backend {
-
-// Wrapper to enable use of shared_ptr for implementing shared ownership of low-level Vulkan fences.
-struct VulkanCmdFence {
-    VulkanCmdFence(VkResult initialStatus) {
-        // Internally we use the VK_INCOMPLETE status to mean "not yet submitted". When this fence
-        // gets submitted, its status changes to VK_NOT_READY. Finally, when the GPU actually
-        // finishes executing the command buffer, the status changes to VK_SUCCESS.
-        status.store(initialStatus);
-    }
-
-    ~VulkanCmdFence() = default;
-
-    void setStatus(VkResult value) { status.store(value); }
-
-    VkResult getStatus() { return status.load(std::memory_order_acquire); }
-
-private:
-    std::atomic<VkResult> status;
-};
-
-struct VulkanFence : public HwFence, fvkmemory::ThreadSafeResource {
-    VulkanFence() {}
-    std::shared_ptr<VulkanCmdFence> fence;
-};
-
-struct VulkanTimerQuery : public HwTimerQuery, fvkmemory::ThreadSafeResource {
-    explicit VulkanTimerQuery(std::tuple<uint32_t, uint32_t> indices);
-    ~VulkanTimerQuery();
-
-    void setFence(std::shared_ptr<VulkanCmdFence> fence) noexcept;
-
-    bool isCompleted() noexcept;
-
-    uint32_t getStartingQueryIndex() const {
-        return mStartingQueryIndex;
-    }
-
-    uint32_t getStoppingQueryIndex() const {
-        return mStoppingQueryIndex;
-    }
-
-private:
-    uint32_t mStartingQueryIndex;
-    uint32_t mStoppingQueryIndex;
-
-    std::shared_ptr<VulkanCmdFence> mFence;
-    utils::Mutex mFenceMutex;
-};
-
-} // namespace filament::backend
-
-#endif // TNT_FILAMENT_BACKEND_VULKANHASYNCANDLES_H
@@ -66,10 +66,10 @@ inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, V
            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);
@@ -109,10 +109,10 @@ 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);
@@ -126,7 +126,8 @@ struct BlitterUniforms {
 }// anonymous namespace

 VulkanBlitter::VulkanBlitter(VkPhysicalDevice physicalDevice, VulkanCommands* commands) noexcept
-        : mPhysicalDevice(physicalDevice), mCommands(commands) {}
+    : mPhysicalDevice(physicalDevice),
+      mCommands(commands) {}

 void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {

@@ -150,8 +151,7 @@ void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {
    }
 #endif

-    VulkanCommandBuffer& commands = dst.texture->getIsProtected() ?
-            mCommands->getProtected() : mCommands->get();
+    VulkanCommandBuffer& commands = mCommands->get();
    commands.acquire(src.texture);
    commands.acquire(dst.texture);
    resolveFast(&commands, aspect, src, dst);
@@ -176,8 +176,7 @@ void VulkanBlitter::blit(VkFilter filter,
 #endif
    // src and dst should have the same aspect here
    VkImageAspectFlags const aspect = src.texture->getImageAspect();
-    VulkanCommandBuffer& commands = dst.texture->getIsProtected() ?
-            mCommands->getProtected() : mCommands->get();
+    VulkanCommandBuffer& commands = mCommands->get();
    commands.acquire(src.texture);
    commands.acquire(dst.texture);
    blitFast(&commands, aspect, filter, src, dst, srcRectPair, dstRectPair);
@@ -33,194 +33,275 @@ using namespace utils;

 namespace filament::backend {

-namespace {
-
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
 using Timestamp = VulkanGroupMarkers::Timestamp;
 #endif

-VkCommandBuffer createCommandBuffer(VkDevice device, VkCommandPool pool) {
-    VkCommandBuffer cmdbuffer;
-    // Create the low-level command buffer.
-    VkCommandBufferAllocateInfo const allocateInfo{
-        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
-        .commandPool = pool,
-        .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
-        .commandBufferCount = 1,
-    };
-    // The buffer allocated here will be implicitly reset when vkBeginCommandBuffer is called.
-    // We don't need to deallocate since destroying the pool will free all of the buffers.
-    vkAllocateCommandBuffers(device, &allocateInfo, &cmdbuffer);
-    return cmdbuffer;
+VulkanCmdFence::VulkanCmdFence(VkFence ifence)
+    : fence(ifence) {
+    // Internally we use the VK_INCOMPLETE status to mean "not yet submitted". When this fence gets
+    // submitted, its status changes to VK_NOT_READY. Finally, when the GPU actually finishes
+    // executing the command buffer, the status changes to VK_SUCCESS.
+    status.store(VK_INCOMPLETE);
 }

-} // anonymous namespace
+VulkanCommandBuffer::VulkanCommandBuffer(VulkanResourceAllocator* allocator, VkDevice device,
+        VkCommandPool pool)
+    : mResourceManager(allocator),
+      mPipeline(VK_NULL_HANDLE) {
+    // Create the low-level command buffer.
+    const VkCommandBufferAllocateInfo allocateInfo{
+            .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
+            .commandPool = pool,
+            .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
+            .commandBufferCount = 1,
+    };
+
+    // The buffer allocated here will be implicitly reset when vkBeginCommandBuffer is called.
+    // We don't need to deallocate since destroying the pool will free all of the buffers.
+    vkAllocateCommandBuffers(device, &allocateInfo, &mBuffer);
+}
+
+CommandBufferObserver::~CommandBufferObserver() {}
+
+static VkCommandPool createPool(VkDevice device, uint32_t queueFamilyIndex) {
+    VkCommandPoolCreateInfo createInfo = {
+            .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
+            .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT
+                     | VK_COMMAND_POOL_CREATE_TRANSIENT_BIT,
+            .queueFamilyIndex = queueFamilyIndex,
+    };
+    VkCommandPool pool;
+    vkCreateCommandPool(device, &createInfo, VKALLOC, &pool);
+    return pool;
+}

 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
 void VulkanGroupMarkers::push(std::string const& marker, Timestamp start) noexcept {
-    mMarkers.push_back({marker,
-        start.time_since_epoch().count() > 0.0
-        ? start
-        : std::chrono::high_resolution_clock::now()});
+    mMarkers.push_back(marker);
+
+#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+    mTimestamps.push_back(start.time_since_epoch().count() > 0.0
+                                  ? start
+                                  : std::chrono::high_resolution_clock::now());
+#endif
 }

 std::pair<std::string, Timestamp> VulkanGroupMarkers::pop() noexcept {
-    auto ret = mMarkers.back();
+    auto const marker = mMarkers.back();
    mMarkers.pop_back();
-    return ret;
+
+#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+    auto const timestamp = mTimestamps.back();
+    mTimestamps.pop_back();
+    return std::make_pair(marker, timestamp);
+#else
+    return std::make_pair(marker, Timestamp{});
+#endif
 }

 std::pair<std::string, Timestamp> VulkanGroupMarkers::pop_bottom() noexcept {
-    auto ret = mMarkers.front();
+    auto const marker = mMarkers.front();
    mMarkers.pop_front();
-    return ret;
+
+#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+    auto const timestamp = mTimestamps.front();
+    mTimestamps.pop_front();
+    return std::make_pair(marker, timestamp);
+#else
+    return std::make_pair(marker, Timestamp{});
+#endif
 }

-std::pair<std::string, Timestamp> const& VulkanGroupMarkers::top() const {
+std::pair<std::string, Timestamp> VulkanGroupMarkers::top() const {
    assert_invariant(!empty());
-    return mMarkers.back();
+    auto const marker = mMarkers.back();
+#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+    auto const topTimestamp = mTimestamps.front();
+    return std::make_pair(marker, topTimestamp);
+#else
+    return std::make_pair(marker, Timestamp{});
+#endif
 }

 bool VulkanGroupMarkers::empty() const noexcept {
    return mMarkers.empty();
 }
+
 #endif // FVK_DEBUG_GROUP_MARKERS

-VulkanCommandBuffer::VulkanCommandBuffer(VulkanContext* context, VkDevice device, VkQueue queue,
-        VkCommandPool pool, bool isProtected)
-    : mContext(context),
-      mMarkerCount(0),
-      isProtected(isProtected),
-      mDevice(device),
+VulkanCommands::VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
+        VulkanContext* context, VulkanResourceAllocator* allocator)
+    : mDevice(device),
      mQueue(queue),
-      mBuffer(createCommandBuffer(device, pool)),
-      mFenceStatus(std::make_shared<VulkanCmdFence>(VK_INCOMPLETE)) {
+      mPool(createPool(mDevice, queueFamilyIndex)),
+      mContext(context),
+      mStorage(CAPACITY) {
    VkSemaphoreCreateInfo sci{.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO};
-    vkCreateSemaphore(mDevice, &sci, VKALLOC, &mSubmission);
+    for (auto& semaphore: mSubmissionSignals) {
+        vkCreateSemaphore(mDevice, &sci, nullptr, &semaphore);
+    }

    VkFenceCreateInfo fenceCreateInfo{.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO};
-    vkCreateFence(device, &fenceCreateInfo, VKALLOC, &mFence);
-}
-
-VulkanCommandBuffer::~VulkanCommandBuffer() {
-    vkDestroySemaphore(mDevice, mSubmission, VKALLOC);
-    vkDestroyFence(mDevice, mFence, VKALLOC);
-}
-
-void VulkanCommandBuffer::reset() noexcept {
-    mMarkerCount = 0;
-    mResources.clear();
-    mWaitSemaphores.clear();
-
-    // Internally we use the VK_INCOMPLETE status to mean "not yet submitted". When this fence
-    // gets, gets submitted, its status changes to VK_NOT_READY. Finally, when the GPU actually
-    // finishes executing the command buffer, the status changes to VK_SUCCESS.
-    mFenceStatus = std::make_shared<VulkanCmdFence>(VK_INCOMPLETE);
-    vkResetFences(mDevice, 1, &mFence);
-}
-
-void VulkanCommandBuffer::pushMarker(char const* marker) noexcept {
-    if (mContext->isDebugUtilsSupported()) {
-        VkDebugUtilsLabelEXT labelInfo = {
-                .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
-                .pLabelName = marker,
-                .color = {0, 1, 0, 1},
-        };
-        vkCmdBeginDebugUtilsLabelEXT(mBuffer, &labelInfo);
-    } else if (mContext->isDebugMarkersSupported()) {
-        VkDebugMarkerMarkerInfoEXT markerInfo = {
-                .sType = VK_STRUCTURE_TYPE_DEBUG_MARKER_MARKER_INFO_EXT,
-                .pMarkerName = marker,
-                .color = {0.0f, 1.0f, 0.0f, 1.0f},
-        };
-        vkCmdDebugMarkerBeginEXT(mBuffer, &markerInfo);
+    for (auto& fence: mFences) {
+        vkCreateFence(device, &fenceCreateInfo, VKALLOC, &fence);
    }
-    mMarkerCount++;
-}

-void VulkanCommandBuffer::popMarker() noexcept{
-    assert_invariant(mMarkerCount > 0);
-    if (mContext->isDebugUtilsSupported()) {
-        vkCmdEndDebugUtilsLabelEXT(mBuffer);
-    } else if (mContext->isDebugMarkersSupported()) {
-        vkCmdDebugMarkerEndEXT(mBuffer);
+    for (size_t i = 0; i < CAPACITY; ++i) {
+        mStorage[i] = std::make_unique<VulkanCommandBuffer>(allocator, mDevice, mPool);
    }
-    mMarkerCount--;
+
+#if !FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
+    (void) mContext;
+#endif
 }

-void VulkanCommandBuffer::insertEvent(char const* marker) noexcept {
-    if (mContext->isDebugUtilsSupported()) {
-        VkDebugUtilsLabelEXT labelInfo = {
-                .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
-                .pLabelName = marker,
-                .color = {1, 1, 0, 1},
-        };
-        vkCmdInsertDebugUtilsLabelEXT(mBuffer, &labelInfo);
-    } else if (mContext->isDebugMarkersSupported()) {
-        VkDebugMarkerMarkerInfoEXT markerInfo = {
-                .sType = VK_STRUCTURE_TYPE_DEBUG_MARKER_MARKER_INFO_EXT,
-                .pMarkerName = marker,
-                .color = {0.0f, 1.0f, 0.0f, 1.0f},
-        };
-        vkCmdDebugMarkerInsertEXT(mBuffer, &markerInfo);
+void VulkanCommands::terminate() {
+    wait();
+    gc();
+    vkDestroyCommandPool(mDevice, mPool, VKALLOC);
+    for (VkSemaphore sema: mSubmissionSignals) {
+        vkDestroySemaphore(mDevice, sema, VKALLOC);
+    }
+    for (VkFence fence: mFences) {
+        vkDestroyFence(mDevice, fence, VKALLOC);
    }
 }

-void VulkanCommandBuffer::begin() noexcept {
-    // Begin writing into the command buffer.
-    VkCommandBufferBeginInfo const binfo{
-        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
-        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
-    };
-    vkBeginCommandBuffer(mBuffer, &binfo);
-}
-
-VkSemaphore VulkanCommandBuffer::submit() {
-    while (mMarkerCount > 0) {
-        popMarker();
-    }
-
-    vkEndCommandBuffer(mBuffer);
-
-    VkPipelineStageFlags const waitDestStageMasks[2] = {
-        VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-        VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-    };
-
-    VkSubmitInfo submitInfo{
-        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
-        .waitSemaphoreCount = mWaitSemaphores.size(),
-        .pWaitSemaphores = mWaitSemaphores.data(),
-        .pWaitDstStageMask = waitDestStageMasks,
-        .commandBufferCount = 1u,
-        .pCommandBuffers = &mBuffer,
-        .signalSemaphoreCount = 1u,
-        .pSignalSemaphores = &mSubmission,
-    };
-    // add submit protection if needed
-    VkProtectedSubmitInfo protectedSubmitInfo{
-        .sType = VK_STRUCTURE_TYPE_PROTECTED_SUBMIT_INFO,
-        .protectedSubmit = VK_TRUE,
-    };
-
-    if (isProtected) {
-        submitInfo.pNext = &protectedSubmitInfo;
+VulkanCommandBuffer& VulkanCommands::get() {
+    if (mCurrentCommandBufferIndex >= 0) {
+        return *mStorage[mCurrentCommandBufferIndex].get();
    }

+    // If we ran out of available command buffers, stall until one finishes. This is very rare.
+    // It occurs only when Filament invokes commit() or endFrame() a large number of times without
+    // presenting the swap chain or waiting on a fence.
+    while (mAvailableBufferCount == 0) {
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-    FVK_LOGI << "Submitting cmdbuffer=" << mBuffer
-             << " wait=(";
-    for (size_t s = 0, count = mWaitSemaphores.size(); s < count; ++s) {
-        FVK_LOGI << mWaitSemaphores[s] << " ";
+        FVK_LOGI << "VulkanCommands has stalled. "
+               << "If this occurs frequently, consider increasing VK_MAX_COMMAND_BUFFERS."
+               << io::endl;
+#endif
+        wait();
+        gc();
+    }
+
+    VulkanCommandBuffer* currentbuf = nullptr;
+    // Find an available slot.
+    for (size_t i = 0; i < CAPACITY; ++i) {
+        auto wrapper = mStorage[i].get();
+        if (wrapper->buffer() == VK_NULL_HANDLE) {
+            mCurrentCommandBufferIndex = static_cast<int8_t>(i);
+            currentbuf = wrapper;
+            break;
+        }
+    }
+
+    assert_invariant(currentbuf);
+    mAvailableBufferCount--;
+
+    // Note that the fence wrapper uses shared_ptr because a DriverAPI fence can also have ownership
+    // over it.  The destruction of the low-level fence occurs either in VulkanCommands::gc(), or in
+    // VulkanDriver::destroyFence(), both of which are safe spots.
+    currentbuf->fence = std::make_shared<VulkanCmdFence>(mFences[mCurrentCommandBufferIndex]);
+
+    // Begin writing into the command buffer.
+    const VkCommandBufferBeginInfo binfo{
+            .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+            .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+    };
+    vkBeginCommandBuffer(currentbuf->buffer(), &binfo);
+
+    // Notify the observer that a new command buffer has been activated.
+    if (mObserver) {
+        mObserver->onCommandBuffer(*currentbuf);
+    }
+
+#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
+    // We push the current markers onto a temporary stack. This must be placed after currentbuf is
+    // set to the new command buffer since pushGroupMarker also calls get().
+    while (mCarriedOverMarkers && !mCarriedOverMarkers->empty()) {
+        auto [marker, time] = mCarriedOverMarkers->pop();
+        pushGroupMarker(marker.c_str(), time);
+    }
+#endif
+    return *currentbuf;
+}
+
+bool VulkanCommands::flush() {
+    // It's perfectly fine to call flush when no commands have been written.
+    if (mCurrentCommandBufferIndex < 0) {
+        return false;
+    }
+
+    // Before actually submitting, we need to pop any leftover group markers.
+    // Note that this needs to occur before vkEndCommandBuffer.
+#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
+    while (mGroupMarkers && !mGroupMarkers->empty()) {
+        if (!mCarriedOverMarkers) {
+            mCarriedOverMarkers = std::make_unique<VulkanGroupMarkers>();
+        }
+        auto const [marker, time] = mGroupMarkers->top();
+        mCarriedOverMarkers->push(marker, time);
+        // We still need to call through to vkCmdEndDebugUtilsLabelEXT.
+        popGroupMarker();
    }
-    FVK_LOGI << ") "
-             << " signal=" << mSubmission
-             << " fence=" << mFence << utils::io::endl;
 #endif

-    mFenceStatus->setStatus(VK_NOT_READY);
-    UTILS_UNUSED_IN_RELEASE VkResult result =
-        vkQueueSubmit(mQueue, 1, &submitInfo, mFence);
+    int8_t const index = mCurrentCommandBufferIndex;
+    VulkanCommandBuffer const* currentbuf = mStorage[index].get();
+    VkSemaphore const renderingFinished = mSubmissionSignals[index];
+
+    vkEndCommandBuffer(currentbuf->buffer());
+
+    // If the injected semaphore is an "image available" semaphore that has not yet been signaled,
+    // it is sometimes fine to start executing commands anyway, as along as we stall the GPU at the
+    // VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT stage. However we need to assume the worst
+    // here and use VK_PIPELINE_STAGE_ALL_COMMANDS_BIT. This is a more aggressive stall, but it is
+    // the only safe option because the previously submitted command buffer might have set up some
+    // state that the new command buffer depends on.
+    VkPipelineStageFlags waitDestStageMasks[2] = {
+            VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+            VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+    };
+
+    VkSemaphore signals[2] = {
+            VK_NULL_HANDLE,
+            VK_NULL_HANDLE,
+    };
+    uint32_t waitSemaphoreCount = 0;
+    if (mSubmissionSignal) {
+        signals[waitSemaphoreCount++] = mSubmissionSignal;
+    }
+    if (mInjectedSignal) {
+        signals[waitSemaphoreCount++] = mInjectedSignal;
+    }
+    VkCommandBuffer const cmdbuffer = currentbuf->buffer();
+    VkSubmitInfo submitInfo{
+            .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+            .waitSemaphoreCount = waitSemaphoreCount,
+            .pWaitSemaphores = waitSemaphoreCount > 0 ? signals : nullptr,
+            .pWaitDstStageMask = waitDestStageMasks,
+            .commandBufferCount = 1,
+            .pCommandBuffers = &cmdbuffer,
+            .signalSemaphoreCount = 1u,
+            .pSignalSemaphores = &renderingFinished,
+    };
+
+#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
+    FVK_LOGI << "Submitting cmdbuffer=" << cmdbuffer
+           << " wait=(" << signals[0] << ", " << signals[1] << ") "
+           << " signal=" << renderingFinished
+           << " fence=" << currentbuf->fence->fence
+           << utils::io::endl;
+#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());
+    }

 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
    if (result != VK_SUCCESS) {
@@ -228,298 +309,176 @@ VkSemaphore VulkanCommandBuffer::submit() {
    }
 #endif
    assert_invariant(result == VK_SUCCESS);
-    mWaitSemaphores.clear();
-    return mSubmission;
-}
-
-CommandBufferPool::CommandBufferPool(VulkanContext* context, VkDevice device, VkQueue queue,
-        uint8_t queueFamilyIndex, bool isProtected)
-    : mDevice(device),
-      mRecording(INVALID) {
-    VkCommandPoolCreateInfo createInfo = {
-        .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
-        .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT |
-                 VK_COMMAND_POOL_CREATE_TRANSIENT_BIT |
-                (isProtected ? VK_COMMAND_POOL_CREATE_PROTECTED_BIT : 0u),
-        .queueFamilyIndex = queueFamilyIndex,
-    };
-    vkCreateCommandPool(device, &createInfo, VKALLOC, &mPool);
-
-    for (size_t i = 0; i < CAPACITY; ++i) {
-        mBuffers.emplace_back(
-                std::make_unique<VulkanCommandBuffer>(context, device, queue, mPool, isProtected));
-    }
-}
-
-CommandBufferPool::~CommandBufferPool() {
-    wait();
-    gc();
-    vkDestroyCommandPool(mDevice, mPool, VKALLOC);
-}
-
-VulkanCommandBuffer& CommandBufferPool::getRecording() {
-    if (isRecording()) {
-        return *mBuffers[mRecording];
-    }
-
-    auto const findNext = [this]() {
-        for (int8_t i = 0; i < CAPACITY; ++i) {
-            if (!mSubmitted[i]) {
-                return i;
-            }
-        }
-        return INVALID;
-    };
-
-    while ((mRecording = findNext()) == INVALID) {
-        wait();
-        gc();
-    }
-
-    auto& recording = *mBuffers[mRecording];
-    recording.begin();
-
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    if (mGroupMarkers) {
-        std::unique_ptr<VulkanGroupMarkers> markers = std::make_unique<VulkanGroupMarkers>();
-        while (!mGroupMarkers->empty()) {
-            auto [marker, timestamp] = mGroupMarkers->pop_bottom();
-            recording.pushMarker(marker.c_str());
-            markers->push(marker, timestamp);
-        }
-        std::swap(mGroupMarkers, markers);
-    }
-#endif
-
-    return recording;
-}
-
-void CommandBufferPool::gc() {
-    ActiveBuffers reclaimed;
-    mSubmitted.forEachSetBit([this,&reclaimed] (size_t index) {
-        auto& buffer = mBuffers[index];
-        if (buffer->getStatus() == VK_SUCCESS) {
-            reclaimed.set(index, true);
-            buffer->reset();
-        }
-    });
-    mSubmitted &= ~reclaimed;
-}
-
-void CommandBufferPool::update() {
-    mSubmitted.forEachSetBit([this] (size_t index) {
-        auto& buffer = mBuffers[index];
-        VkResult status = vkGetFenceStatus(mDevice, buffer->getVkFence());
-        if (status == VK_SUCCESS) {
-            buffer->setComplete();
-        }
-    });
-}
-
-VkSemaphore CommandBufferPool::flush() {
-    // We're not recording right now.
-    if (!isRecording()) {
-        return VK_NULL_HANDLE;
-    }
-    auto submitSemaphore = mBuffers[mRecording]->submit();
-    mSubmitted.set(mRecording, true);
-    mRecording = INVALID;
-    return submitSemaphore;
-}
-
-void CommandBufferPool::wait() {
-    uint8_t count = 0;
-    VkFence fences[CAPACITY];
-    mSubmitted.forEachSetBit([this, &count, &fences] (size_t index) {
-        fences[count++] = mBuffers[index]->getVkFence();
-    });
-    vkWaitForFences(mDevice, count, fences, VK_TRUE, UINT64_MAX);
-    update();
-}
-
-void CommandBufferPool::waitFor(VkSemaphore previousAction) {
-    if (!isRecording()) {
-        return;
-    }
-    auto& recording = mBuffers[mRecording];
-    recording->insertWait(previousAction);
-}
-
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-std::string CommandBufferPool::topMarker() const {
-    if (!mGroupMarkers || mGroupMarkers->empty()) {
-        return "";
-    }
-    return std::get<0>(mGroupMarkers->top());
-}
-
-void CommandBufferPool::pushMarker(char const* marker, VulkanGroupMarkers::Timestamp timestamp) {
-    if (!mGroupMarkers) {
-        mGroupMarkers = std::make_unique<VulkanGroupMarkers>();
-    }
-    mGroupMarkers->push(marker, timestamp);
-    getRecording().pushMarker(marker);
-}
-
-std::pair<std::string, VulkanGroupMarkers::Timestamp> CommandBufferPool::popMarker() {
-    assert_invariant(mGroupMarkers && !mGroupMarkers->empty());
-    auto ret = mGroupMarkers->pop();
-
-    // Note that if we're popping a marker while not recording, we would just pop the conceptual
-    // stack of marker (i.e. mGroupMarkers) and not carry out the pop on the command buffer.
-    if (isRecording()) {
-        getRecording().popMarker();
-    }
-    return ret;
-}
-
-void CommandBufferPool::insertEvent(char const* marker) {
-    getRecording().insertEvent(marker);
-}
-#endif // FVK_DEBUG_GROUP_MARKERS
-
-VulkanCommands::VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
-        VkQueue protectedQueue, uint32_t protectedQueueFamilyIndex, VulkanContext* context)
-    : mDevice(device),
-      mProtectedQueue(protectedQueue),
-      mProtectedQueueFamilyIndex(protectedQueueFamilyIndex),
-      mContext(context),
-      mPool(std::make_unique<CommandBufferPool>(context, device, queue, queueFamilyIndex, false)) {}
-
-void VulkanCommands::terminate() {
-    mPool.reset();
-    mProtectedPool.reset();
-}
-
-VulkanCommandBuffer& VulkanCommands::get() {
-    auto& ret = mPool->getRecording();
-    return ret;
-}
-
-VulkanCommandBuffer& VulkanCommands::getProtected() {
-    assert_invariant(mProtectedQueue != VK_NULL_HANDLE);
-
-    if (!mProtectedPool) {
-        mProtectedPool = std::make_unique<CommandBufferPool>(mContext, mDevice, mProtectedQueue,
-                mProtectedQueueFamilyIndex, true);
-    }
-    auto& ret = mProtectedPool->getRecording();
-    return ret;
-}
-
-bool VulkanCommands::flush() {
-    // It's possible to call flush and wait at "terminate", in which case, we'll just return.
-    if (!mPool && !mProtectedPool) {
-        return false;
-    }
-
-    VkSemaphore dependency = mInjectedDependency;
-    VkSemaphore lastSubmit = mLastSubmit;
-    bool hasFlushed = false;
-
-    // Note that we've ordered it so that the non-protected commands are followed by the protected
-    // commands.  This assumes that the protected commands will be that one doing the rendering into
-    // the protected memory (i.e. protected render target).
-    for (auto pool: {mPool.get(), mProtectedPool.get()}) {
-        if (!pool || !pool->isRecording()) {
-            continue;
-        }
-        if (dependency != VK_NULL_HANDLE) {
-            pool->waitFor(dependency);
-        }
-        if (lastSubmit != VK_NULL_HANDLE) {
-            pool->waitFor(lastSubmit);
-            lastSubmit = VK_NULL_HANDLE;
-        }
-        dependency = pool->flush();
-        hasFlushed = true;
-    }
-
-    if (hasFlushed) {
-        mInjectedDependency = VK_NULL_HANDLE;
-        mLastSubmit = dependency;
-    }

+    mSubmissionSignal = renderingFinished;
+    mInjectedSignal = VK_NULL_HANDLE;
+    mCurrentCommandBufferIndex = -1;
    return true;
 }

+VkSemaphore VulkanCommands::acquireFinishedSignal() {
+    VkSemaphore semaphore = mSubmissionSignal;
+    mSubmissionSignal = VK_NULL_HANDLE;
+#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
+    FVK_LOGI << "Acquiring " << semaphore << " (e.g. for vkQueuePresentKHR)" << io::endl;
+#endif
+    return semaphore;
+}
+
+void VulkanCommands::injectDependency(VkSemaphore next) {
+    assert_invariant(mInjectedSignal == VK_NULL_HANDLE);
+    mInjectedSignal = next;
+#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
+    FVK_LOGI << "Injecting " << next << " (e.g. due to vkAcquireNextImageKHR)" << io::endl;
+#endif
+}
+
 void VulkanCommands::wait() {
-    // It's possible to call flush and wait at "terminate", in which case, we'll just return.
-    if (!mPool && !mProtectedPool) {
-        return;
+    VkFence fences[CAPACITY];
+    size_t count = 0;
+    for (size_t i = 0; i < CAPACITY; i++) {
+        auto wrapper = mStorage[i].get();
+        if (wrapper->buffer() != VK_NULL_HANDLE
+                && mCurrentCommandBufferIndex != static_cast<int8_t>(i)) {
+            fences[count++] = wrapper->fence->getFence();
+        }
    }
-
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("commands::wait");
-
-    mPool->wait();
-    if (mProtectedPool) {
-        mProtectedPool->wait();
+    if (count > 0) {
+        vkWaitForFences(mDevice, count, fences, VK_TRUE, UINT64_MAX);
+        updateFences();
    }
-    FVK_SYSTRACE_END();
 }

 void VulkanCommands::gc() {
    FVK_SYSTRACE_CONTEXT();
    FVK_SYSTRACE_START("commands::gc");

-    mPool->gc();
-    if (mProtectedPool) {
-        mProtectedPool->gc();
+    VkFence fences[CAPACITY];
+    size_t count = 0;
+
+    for (size_t i = 0; i < CAPACITY; i++) {
+        auto wrapper = mStorage[i].get();
+        if (wrapper->buffer() == VK_NULL_HANDLE) {
+            continue;
+        }
+        auto const vkfence = wrapper->fence->getFence();
+        VkResult const result = vkGetFenceStatus(mDevice, vkfence);
+        if (result != VK_SUCCESS) {
+            continue;
+        }
+        fences[count++] = vkfence;
+        wrapper->fence->setValue(VK_SUCCESS);
+        wrapper->reset();
+        mAvailableBufferCount++;
+    }
+
+    if (count > 0) {
+        vkResetFences(mDevice, count, fences);
    }
    FVK_SYSTRACE_END();
 }

 void VulkanCommands::updateFences() {
-    mPool->update();
-    if (mProtectedPool) {
-        mProtectedPool->update();
+    for (size_t i = 0; i < CAPACITY; i++) {
+        auto wrapper = mStorage[i].get();
+        if (wrapper->buffer() != VK_NULL_HANDLE) {
+            VulkanCmdFence* fence = wrapper->fence.get();
+            if (fence) {
+                VkResult status = vkGetFenceStatus(mDevice, fence->getFence());
+                // This is either VK_SUCCESS, VK_NOT_READY, or VK_ERROR_DEVICE_LOST.
+                fence->setValue(status);
+            }
+        }
    }
 }

 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)

 void VulkanCommands::pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp) {
-    mPool->pushMarker(str, timestamp);
-    if (mProtectedPool) {
-        mProtectedPool->pushMarker(str, timestamp);
-    }
 #if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
-    FVK_LOGD << "----> " << str << utils::io::endl;
+    // If the timestamp is not 0, then we are carrying over a marker across buffer submits.
+    // If it is 0, then this is a normal marker push and we should just print debug line as usual.
+    if (timestamp.time_since_epoch().count() == 0.0) {
+        FVK_LOGD << "----> " << str << utils::io::endl;
+    }
 #endif
+
+    // TODO: Add group marker color to the Driver API
+    VkCommandBuffer const cmdbuffer = get().buffer();
+
+    if (!mGroupMarkers) {
+        mGroupMarkers = std::make_unique<VulkanGroupMarkers>();
+    }
+    mGroupMarkers->push(str, timestamp);
+
+    if (mContext->isDebugUtilsSupported()) {
+        VkDebugUtilsLabelEXT labelInfo = {
+                .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
+                .pLabelName = str,
+                .color = {0, 1, 0, 1},
+        };
+        vkCmdBeginDebugUtilsLabelEXT(cmdbuffer, &labelInfo);
+    } else if (mContext->isDebugMarkersSupported()) {
+        VkDebugMarkerMarkerInfoEXT markerInfo = {
+                .sType = VK_STRUCTURE_TYPE_DEBUG_MARKER_MARKER_INFO_EXT,
+                .pMarkerName = str,
+                .color = {0.0f, 1.0f, 0.0f, 1.0f},
+        };
+        vkCmdDebugMarkerBeginEXT(cmdbuffer, &markerInfo);
+    }
 }

 void VulkanCommands::popGroupMarker() {
+    assert_invariant(mGroupMarkers);

+    if (!mGroupMarkers->empty()) {
+        VkCommandBuffer const cmdbuffer = get().buffer();
 #if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
-    auto ret = mPool->popMarker();
-    auto const& marker = ret.first;
-    auto const& startTime = ret.second;
-    auto const endTime = std::chrono::high_resolution_clock::now();
-    std::chrono::duration<double> diff = endTime - startTime;
-    FVK_LOGD << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms"
-             << utils::io::endl;
+        auto const [marker, startTime] = mGroupMarkers->pop();
+        auto const endTime = std::chrono::high_resolution_clock::now();
+        std::chrono::duration<double> diff = endTime - startTime;
+        FVK_LOGD << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms"
+                      << utils::io::endl;
 #else
-    mPool->popMarker();
-#endif // FVK_DEBUG_PRINT_GROUP_MARKERS
-
-    if (mProtectedPool) {
-        mProtectedPool->popMarker();
+        mGroupMarkers->pop();
+#endif
+        if (mContext->isDebugUtilsSupported()) {
+            vkCmdEndDebugUtilsLabelEXT(cmdbuffer);
+        } else if (mContext->isDebugMarkersSupported()) {
+            vkCmdDebugMarkerEndEXT(cmdbuffer);
+        }
+    } else if (mCarriedOverMarkers && !mCarriedOverMarkers->empty()) {
+        // It could be that pop is called between flush() and get() (new command buffer), in which
+        // case the marker is in "carried over" state, we'd just remove that. Since the
+        // mCarriedOverMarkers is in the opposite order, we pop the bottom instead of the top.
+        mCarriedOverMarkers->pop_bottom();
    }
 }

-void VulkanCommands::insertEventMarker(char const* str, uint32_t len) {
-    mPool->insertEvent(str);
-    if (mProtectedPool) {
-        mProtectedPool->insertEvent(str);
+void VulkanCommands::insertEventMarker(char const* string, uint32_t len) {
+    VkCommandBuffer const cmdbuffer = get().buffer();
+    if (mContext->isDebugUtilsSupported()) {
+        VkDebugUtilsLabelEXT labelInfo = {
+                .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
+                .pLabelName = string,
+                .color = {1, 1, 0, 1},
+        };
+        vkCmdInsertDebugUtilsLabelEXT(cmdbuffer, &labelInfo);
+    } else if (mContext->isDebugMarkersSupported()) {
+        VkDebugMarkerMarkerInfoEXT markerInfo = {
+                .sType = VK_STRUCTURE_TYPE_DEBUG_MARKER_MARKER_INFO_EXT,
+                .pMarkerName = string,
+                .color = {0.0f, 1.0f, 0.0f, 1.0f},
+        };
+        vkCmdDebugMarkerInsertEXT(cmdbuffer, &markerInfo);
    }
 }

 std::string VulkanCommands::getTopGroupMarker() const {
-    if (mProtectedPool) {
-        return mProtectedPool->topMarker();
+    if (!mGroupMarkers || mGroupMarkers->empty()) {
+        return "";
    }
-    return mPool->topMarker();
+    return std::get<0>(mGroupMarkers->top());
 }
 #endif // FVK_DEBUG_GROUP_MARKERS

@@ -21,10 +21,8 @@

 #include "DriverBase.h"

-#include "VulkanAsyncHandles.h"
 #include "VulkanConstants.h"
-#include "VulkanUtility.h"
-#include "vulkan/memory/ResourcePointer.h"
+#include "VulkanResources.h"

 #include <utils/Condition.h>
 #include <utils/FixedCapacityVector.h>
@@ -39,8 +37,6 @@

 namespace filament::backend {

-using namespace fvkmemory;
-
 struct VulkanContext;

 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
@@ -51,122 +47,111 @@ public:
    void push(std::string const& marker, Timestamp start = {}) noexcept;
    std::pair<std::string, Timestamp> pop() noexcept;
    std::pair<std::string, Timestamp> pop_bottom() noexcept;
-    std::pair<std::string, Timestamp> const& top() const;
+    std::pair<std::string, Timestamp> top() const;
    bool empty() const noexcept;

 private:
-    std::list<std::pair<std::string, Timestamp>> mMarkers;
+    std::list<std::string> mMarkers;
+#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+    std::list<Timestamp> mTimestamps;
+#endif
 };

 #endif // FVK_DEBUG_GROUP_MARKERS

+// 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;
+    std::atomic<VkResult> status;
+};
+
 // The submission fence has shared ownership semantics because it is potentially wrapped by a
 // DriverApi fence object and should not be destroyed until both the DriverApi object is freed and
 // we're done waiting on the most recent submission of the given command buffer.
 struct VulkanCommandBuffer {
-    VulkanCommandBuffer(VulkanContext* mContext,
-            VkDevice device, VkQueue queue, VkCommandPool pool, bool isProtected);
+    VulkanCommandBuffer(VulkanResourceAllocator* allocator, VkDevice device, VkCommandPool pool);

    VulkanCommandBuffer(VulkanCommandBuffer const&) = delete;
    VulkanCommandBuffer& operator=(VulkanCommandBuffer const&) = delete;

-    ~VulkanCommandBuffer();
-
-    inline void acquire(fvkmemory::resource_ptr<fvkmemory::Resource> resource) {
-        mResources.push_back(resource);
+    inline void acquire(VulkanResource* resource) {
+        mResourceManager.acquire(resource);
    }

-    void reset() noexcept;
-
-    inline void insertWait(VkSemaphore sem) {
-        mWaitSemaphores.insert(sem);
+    inline void acquire(VulkanAcquireOnlyResourceManager* srcResources) {
+        mResourceManager.acquireAll(srcResources);
    }

-    void pushMarker(char const* marker) noexcept;
-    void popMarker() noexcept;
-    void insertEvent(char const* marker) noexcept;
-
-    void begin() noexcept;
-    VkSemaphore submit();
-
-    inline void setComplete() {
-        mFenceStatus->setStatus(VK_SUCCESS);
+    inline void reset() {
+        fence.reset();
+        mResourceManager.clear();
+        mPipeline = VK_NULL_HANDLE;
    }

-    VkResult getStatus() {
-        return mFenceStatus->getStatus();
+    inline void setPipeline(VkPipeline pipeline) {
+        mPipeline = pipeline;
    }

-    std::shared_ptr<VulkanCmdFence> getFenceStatus() const {
-        return mFenceStatus;
+    inline VkPipeline pipeline() const {
+        return mPipeline;
    }

-    VkFence getVkFence() const {
-        return mFence;
+    inline VkCommandBuffer buffer() const {
+        if (fence) {
+            return mBuffer;
+        }
+        return VK_NULL_HANDLE;
    }

-    VkCommandBuffer buffer() const {
-        return mBuffer;
-    }
+    std::shared_ptr<VulkanCmdFence> fence;

 private:
-    VulkanContext* mContext;
-    uint8_t mMarkerCount;
-    bool const isProtected;
-    VkDevice mDevice;
-    VkQueue mQueue;
-    CappedArray<VkSemaphore, 2> mWaitSemaphores;
+    VulkanAcquireOnlyResourceManager mResourceManager;
    VkCommandBuffer mBuffer;
-    VkSemaphore mSubmission;
-    VkFence mFence;
-    std::shared_ptr<VulkanCmdFence> mFenceStatus;
-    std::vector<fvkmemory::resource_ptr<Resource>> mResources;
+    VkPipeline mPipeline;
 };

-struct CommandBufferPool {
-    using ActiveBuffers = utils::bitset64;
-    static constexpr int8_t INVALID = -1;
-
-    CommandBufferPool(VulkanContext* context, VkDevice device, VkQueue queue,
-            uint8_t queueFamilyIndex, bool isProtected);
-    ~CommandBufferPool();
-
-    VulkanCommandBuffer& getRecording();
-
-    void gc();
-    void update();
-    VkSemaphore flush();
-    void wait();
-    void waitFor(VkSemaphore previousAction);
-
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    std::string topMarker() const;
-    void pushMarker(char const* marker, VulkanGroupMarkers::Timestamp timestamp);
-    std::pair<std::string, VulkanGroupMarkers::Timestamp> popMarker();
-    void insertEvent(char const* marker);
-#endif
-
-    inline bool isRecording() const { return mRecording != INVALID; }
-
-private:
-    static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
-    // int8 only goes up to 127, therefore capacity must be less than that.
-    static_assert(CAPACITY < 128);
-
-    // The number of bits in ActiveBuffers describe the usage of the buffers in the pool, so must be
-    // larger than the size of the pool.
-    static_assert(sizeof(ActiveBuffers) * 8 >= CAPACITY);
-
-    using BufferList = utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>>;
-    VkDevice mDevice;
-    VkCommandPool mPool;
-    ActiveBuffers mSubmitted;
-    std::vector<std::unique_ptr<VulkanCommandBuffer>> mBuffers;
-    int8_t mRecording;
-
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
-#endif
+// Allows classes to be notified after a new command buffer has been activated.
+class CommandBufferObserver {
+public:
+    virtual void onCommandBuffer(const VulkanCommandBuffer& cmdbuffer) = 0;
+    virtual ~CommandBufferObserver();
 };

 // Manages a set of command buffers and semaphores, exposing an API that is significantly simpler
@@ -180,6 +165,9 @@ private:
 //    - This creates a guarantee of in-order execution.
 //    - Semaphores are recycled to prevent create / destroy churn.
 //
+// - Notifies listeners when recording begins in a new VkCommandBuffer.
+//    - Used by PipelineCache so that it knows when to clear out its shadow state.
+//
 // - Allows 1 user to inject a "dependency" semaphore that stalls the next flush.
 //    - This is used for asynchronous acquisition of a swap chain image, since the GPU
 //      might require a valid swap chain image when it starts executing the command buffer.
@@ -195,17 +183,13 @@ private:
 class VulkanCommands {
 public:
    VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
-            VkQueue protectedQueue, uint32_t protectedQueueFamilyIndex, VulkanContext* context);
+            VulkanContext* context, VulkanResourceAllocator* allocator);

    void terminate();

    // Creates a "current" command buffer if none exists, otherwise returns the current one.
    VulkanCommandBuffer& get();

-    // Creates a "current" protected capable command buffer if none exists, otherwise
-    // returns the current one.
-    VulkanCommandBuffer& getProtected();
-
    // Submits the current command buffer if it exists, then sets "current" to null.
    // If there are no outstanding commands then nothing happens and this returns false.
    bool flush();
@@ -213,17 +197,11 @@ public:
    // Returns the "rendering finished" semaphore for the most recent flush and removes
    // it from the existing dependency chain. This is especially useful for setting up
    // vkQueuePresentKHR.
-    VkSemaphore acquireFinishedSignal() {
-        VkSemaphore ret= mLastSubmit;
-        mLastSubmit = VK_NULL_HANDLE;
-        return ret;
-    }
+    VkSemaphore acquireFinishedSignal();

    // Takes a semaphore that signals when the next flush can occur. Only one injected
    // semaphore is allowed per flush. Useful after calling vkAcquireNextImageKHR.
-    void injectDependency(VkSemaphore next) {
-        mInjectedDependency = next;
-    }
+    void injectDependency(VkSemaphore next);

    // Destroys all command buffers that are no longer in use.
    void gc();
@@ -234,25 +212,42 @@ public:
    // Updates the atomic "status" variable in every extant fence.
    void updateFences();

+    // Sets an observer who is notified every time a new command buffer has been made "current".
+    // The observer's event handler can only be called during get().
+    void setObserver(CommandBufferObserver* observer) { mObserver = observer; }
+
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
    void pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp = {});
+
    void popGroupMarker();
+
    void insertEventMarker(char const* string, uint32_t len);
+
    std::string getTopGroupMarker() const;
 #endif

 private:
+    static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
    VkDevice const mDevice;
-    VkQueue const mProtectedQueue;
-    // For defered initialization if/when we need protected content
-    uint32_t const mProtectedQueueFamilyIndex;
-    VulkanContext* mContext;
+    VkQueue const mQueue;
+    VkCommandPool const mPool;
+    VulkanContext const* mContext;

-    std::unique_ptr<CommandBufferPool> mPool;
-    std::unique_ptr<CommandBufferPool> mProtectedPool;
+    // int8 only goes up to 127, therefore capacity must be less than that.
+    static_assert(CAPACITY < 128);
+    int8_t mCurrentCommandBufferIndex = -1;
+    VkSemaphore mSubmissionSignal = {};
+    VkSemaphore mInjectedSignal = {};
+    utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>> mStorage;
+    VkFence mFences[CAPACITY] = {};
+    VkSemaphore mSubmissionSignals[CAPACITY] = {};
+    uint8_t mAvailableBufferCount = CAPACITY;
+    CommandBufferObserver* mObserver = nullptr;

-    VkSemaphore mInjectedDependency = VK_NULL_HANDLE;
-    VkSemaphore mLastSubmit = VK_NULL_HANDLE;
+#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
+    std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
+    std::unique_ptr<VulkanGroupMarkers> mCarriedOverMarkers;
+#endif
 };

 } // namespace filament::backend
@@ -1,7 +1,7 @@
 /*
 * Copyright (C) 2021 The Android Open Source Project
 *
- * Licensed under the Apache License, Version 2.0 (the "License");
+* 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
 *
@@ -63,7 +63,7 @@
 #define FVK_DEBUG_SHADER_MODULE           0x00000800
 #define FVK_DEBUG_READ_PIXELS             0x00001000
 #define FVK_DEBUG_PIPELINE_CACHE          0x00002000
-#define FVK_DEBUG_STAGING_ALLOCATION      0x00004000
+#define FVK_DEBUG_ALLOCATION              0x00004000

 // Enable the debug utils extension if it is available.
 #define FVK_DEBUG_DEBUG_UTILS             0x00008000
@@ -77,12 +77,8 @@
 // order of calls).
 #define FVK_DEBUG_FORCE_LOG_TO_I          0x00020000

-// Enable a minimal set of traces to assess the performance of the backend.
-// All other debug features must be disabled.
-#define FVK_DEBUG_PROFILING               0x00040000
-
 // Useful default combinations
-#define FVK_DEBUG_EVERYTHING              (0xFFFFFFFF & ~FVK_DEBUG_PROFILING)
+#define FVK_DEBUG_EVERYTHING              0xFFFFFFFF
 #define FVK_DEBUG_PERFORMANCE     \
    FVK_DEBUG_SYSTRACE

@@ -98,12 +94,6 @@
 #define FVK_DEBUG_FLAGS 0
 #endif

-// Override the debug flags if we are forcing profiling mode
-#if defined(FILAMENT_FORCE_PROFILING_MODE)
-#undef FVK_DEBUG_FLAGS
-#define FVK_DEBUG_FLAGS (FVK_DEBUG_PROFILING)
-#endif
-
 #define FVK_ENABLED(flags) (((FVK_DEBUG_FLAGS) & (flags)) == (flags))

 // Group marker only works only if validation or debug utils is enabled since it uses
@@ -122,10 +112,6 @@ static_assert(FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS));
 static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 #endif

-#if FVK_ENABLED(FVK_DEBUG_PROFILING) && FVK_DEBUG_FLAGS != FVK_DEBUG_PROFILING
-#error PROFILING is exclusive; all other debug features must be disabled.
-#endif
-
 // end dependcy checks

 // Shorthand for combination of enabled debug flags
@@ -137,34 +123,17 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));

 // end shorthands

-#if FVK_DEBUG_FLAGS == FVK_DEBUG_PROFILING
+#if FVK_ENABLED(FVK_DEBUG_SYSTRACE)

-    #ifndef NDEBUG
-        #error PROFILING is meaningless in DEBUG mode.
-    #endif
-
-    #define FVK_SYSTRACE_CONTEXT()
-    #define FVK_SYSTRACE_START(marker)
-    #define FVK_SYSTRACE_END()
-    #define FVK_SYSTRACE_SCOPE()
-    #define FVK_PROFILE_MARKER(marker)  PROFILE_SCOPE(marker)
-
-#elif FVK_ENABLED(FVK_DEBUG_SYSTRACE)
-
-    #include <utils/Systrace.h>
-    
-    #define FVK_SYSTRACE_CONTEXT()      SYSTRACE_CONTEXT()
-    #define FVK_SYSTRACE_START(marker)  SYSTRACE_NAME_BEGIN(marker)
-    #define FVK_SYSTRACE_END()          SYSTRACE_NAME_END()
-    #define FVK_SYSTRACE_SCOPE()        SYSTRACE_NAME(__func__)
-    #define FVK_PROFILE_MARKER(marker)  FVK_SYSTRACE_SCOPE()
+#include <utils/Systrace.h>

+#define FVK_SYSTRACE_CONTEXT()      SYSTRACE_CONTEXT()
+#define FVK_SYSTRACE_START(marker)  SYSTRACE_NAME_BEGIN(marker)
+#define FVK_SYSTRACE_END()          SYSTRACE_NAME_END()
 #else
-    #define FVK_SYSTRACE_CONTEXT()
-    #define FVK_SYSTRACE_START(marker)
-    #define FVK_SYSTRACE_END()
-    #define FVK_SYSTRACE_SCOPE()
-    #define FVK_PROFILE_MARKER(marker)
+#define FVK_SYSTRACE_CONTEXT()
+#define FVK_SYSTRACE_START(marker)
+#define FVK_SYSTRACE_END()
 #endif

 #ifndef FVK_HANDLE_ARENA_SIZE_IN_MB
@@ -196,16 +165,14 @@ constexpr static const int FVK_REQUIRED_VERSION_MINOR = 1;
 // buffers that have been submitted but have not yet finished rendering. Note that Filament can
 // issue multiple commit calls in a single frame, and that we use a triple buffered swap chain on
 // some platforms.
-//
-// Heuristic: Triple Buffering (3) multiplied by maximum number of renderpasses (15).
-constexpr static const int FVK_MAX_COMMAND_BUFFERS = 3 * 15;
+constexpr static const int FVK_MAX_COMMAND_BUFFERS = 10;

 // Number of command buffer submissions that should occur before an unused pipeline is removed
 // from the cache.
 //
 // If this number is low, VkPipeline construction will occur frequently, which can
 // be extremely slow. If this number is high, the memory footprint will be large.
-constexpr static const int FVK_MAX_PIPELINE_AGE = FVK_MAX_COMMAND_BUFFERS;
+constexpr static const int FVK_MAX_PIPELINE_AGE = 10;

 // VulkanPipelineCache does not track which command buffers contain references to which pipelines,
 // instead it simply waits for at least FVK_MAX_COMMAND_BUFFERS submissions to occur before
@@ -25,11 +25,15 @@
 #include <backend/PixelBufferDescriptor.h>

 #include <utils/Panic.h>
+#include <utils/FixedCapacityVector.h>

 #include <algorithm> // for std::max

 using namespace bluevk;

+using utils::FixedCapacityVector;
+
+
 namespace {

 } // end anonymous namespace
@@ -82,7 +86,7 @@ VulkanTimestamps::VulkanTimestamps(VkDevice device) : mDevice(device) {
    VkQueryPoolCreateInfo tqpCreateInfo = {
        .sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
        .queryType = VK_QUERY_TYPE_TIMESTAMP,
-    };
+    };  
    std::unique_lock<utils::Mutex> lock(mMutex);
    tqpCreateInfo.queryCount = mUsed.size() * 2;
    VkResult result = vkCreateQueryPool(mDevice, &tqpCreateInfo, VKALLOC, &mPool);
@@ -97,7 +101,7 @@ std::tuple<uint32_t, uint32_t> VulkanTimestamps::getNextQuery() {
    for (size_t timerIndex = 0; timerIndex < maxTimers; ++timerIndex) {
        if (!mUsed.test(timerIndex)) {
            mUsed.set(timerIndex);
-            return std::make_tuple(timerIndex * 2, timerIndex * 2 + 1);
+	    return std::make_tuple(timerIndex * 2, timerIndex * 2 + 1);
        }
    }
    FVK_LOGE << "More than " << maxTimers << " timers are not supported." << utils::io::endl;
@@ -109,7 +113,7 @@ void VulkanTimestamps::clearQuery(uint32_t queryIndex) {
 }

 void VulkanTimestamps::beginQuery(VulkanCommandBuffer const* commands,
-        fvkmemory::resource_ptr<VulkanTimerQuery> query) {
+        VulkanTimerQuery* query) {
    uint32_t const index = query->getStartingQueryIndex();

    auto const cmdbuffer = commands->buffer();
@@ -117,24 +121,23 @@ void VulkanTimestamps::beginQuery(VulkanCommandBuffer const* commands,
    vkCmdWriteTimestamp(cmdbuffer, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, mPool, index);

    // We stash this because getResult might come before the query is actually processed.
-    query->setFence(commands->getFenceStatus());
+    query->setFence(commands->fence);
 }

 void VulkanTimestamps::endQuery(VulkanCommandBuffer const* commands,
-        fvkmemory::resource_ptr<VulkanTimerQuery> query) {
+        VulkanTimerQuery const* query) {
    uint32_t const index = query->getStoppingQueryIndex();
    vkCmdWriteTimestamp(commands->buffer(), VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, mPool, index);
 }

-VulkanTimestamps::QueryResult VulkanTimestamps::getResult(
-        fvkmemory::resource_ptr<VulkanTimerQuery> query) {
+VulkanTimestamps::QueryResult VulkanTimestamps::getResult(VulkanTimerQuery const* query) {
    uint32_t const index = query->getStartingQueryIndex();
    QueryResult result;
    size_t const dataSize = result.size() * sizeof(uint64_t);
    VkDeviceSize const stride = sizeof(uint64_t) * 2;
    VkResult vkresult =
-            vkGetQueryPoolResults(mDevice, mPool, index, 2, dataSize, (void*) result.data(), stride,
-                    VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WITH_AVAILABILITY_BIT);
+            vkGetQueryPoolResults(mDevice, mPool, index, 2, dataSize, (void*) result.data(),
+                    stride, VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WITH_AVAILABILITY_BIT);
    FILAMENT_CHECK_POSTCONDITION(vkresult == VK_SUCCESS || vkresult == VK_NOT_READY)
            << "vkGetQueryPoolResults error: " << static_cast<int32_t>(vkresult);
    if (vkresult == VK_NOT_READY) {
@@ -21,8 +21,6 @@
 #include "VulkanImageUtility.h"
 #include "VulkanUtility.h"

-#include "vulkan/memory/ResourcePointer.h"
-
 #include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Mutex.h>
@@ -43,10 +41,11 @@ struct VulkanTimerQuery;
 struct VulkanCommandBuffer;

 struct VulkanAttachment {
-    fvkmemory::resource_ptr<VulkanTexture> texture;
+    VulkanTexture* texture = nullptr;
    uint8_t level = 0;
+    uint8_t baseViewIndex = 0;
    uint8_t layerCount = 1;
-    uint8_t layer = 0;
+    uint16_t layer = 0;

    bool isDepth() const;
    VkImage getImage() const;
@@ -72,11 +71,9 @@ public:
    std::tuple<uint32_t, uint32_t> getNextQuery();
    void clearQuery(uint32_t queryIndex);

-    void beginQuery(VulkanCommandBuffer const* commands,
-            fvkmemory::resource_ptr<VulkanTimerQuery> query);
-    void endQuery(VulkanCommandBuffer const* commands,
-            fvkmemory::resource_ptr<VulkanTimerQuery> query);
-    QueryResult getResult(fvkmemory::resource_ptr<VulkanTimerQuery> query);
+    void beginQuery(VulkanCommandBuffer const* commands, VulkanTimerQuery* query);
+    void endQuery(VulkanCommandBuffer const* commands, VulkanTimerQuery const* query);
+    QueryResult getResult(VulkanTimerQuery const* query);

 private:
    VkDevice mDevice;
@@ -86,9 +83,7 @@ private:
 };

 struct VulkanRenderPass {
-    // Between the begin and end command render pass we cache the command buffer
-    VulkanCommandBuffer* commandBuffer;
-    fvkmemory::resource_ptr<VulkanRenderTarget> renderTarget;
+    VulkanRenderTarget* renderTarget;
    VkRenderPass renderPass;
    RenderPassParams params;
    int currentSubpass;
@@ -99,9 +94,6 @@ struct VulkanRenderPass {
 struct VulkanContext {
 public:
    inline uint32_t selectMemoryType(uint32_t flags, VkFlags reqs) const {
-        if ((reqs & VK_MEMORY_PROPERTY_PROTECTED_BIT) != 0) {
-            assert_invariant(isProtectedMemorySupported() == true);
-        }
        for (uint32_t i = 0; i < VK_MAX_MEMORY_TYPES; i++) {
            if (flags & 1) {
                if ((mMemoryProperties.memoryTypes[i].propertyFlags & reqs) == reqs) {
@@ -122,19 +114,19 @@ public:
    }

    inline VkPhysicalDeviceLimits const& getPhysicalDeviceLimits() const noexcept {
-        return mPhysicalDeviceProperties.properties.limits;
+        return mPhysicalDeviceProperties.limits;
    }

    inline uint32_t getPhysicalDeviceVendorId() const noexcept {
-        return mPhysicalDeviceProperties.properties.vendorID;
+        return mPhysicalDeviceProperties.vendorID;
    }

    inline bool isImageCubeArraySupported() const noexcept {
-        return mPhysicalDeviceFeatures.features.imageCubeArray == VK_TRUE;
+        return mPhysicalDeviceFeatures.imageCubeArray == VK_TRUE;
    }

    inline bool isDepthClampSupported() const noexcept {
-        return mPhysicalDeviceFeatures.features.depthClamp == VK_TRUE;
+        return mPhysicalDeviceFeatures.depthClamp == VK_TRUE;
    }

    inline bool isDebugMarkersSupported() const noexcept {
@@ -150,30 +142,16 @@ public:
    }

    inline bool isClipDistanceSupported() const noexcept {
-        return mPhysicalDeviceFeatures.features.shaderClipDistance == VK_TRUE;
-    }
-
-    inline bool isLazilyAllocatedMemorySupported() const noexcept {
-        return mLazilyAllocatedMemorySupported;
-    }
-
-    inline bool isProtectedMemorySupported() const noexcept {
-        return mProtectedMemorySupported;
+        return mPhysicalDeviceFeatures.shaderClipDistance == VK_TRUE;
    }

 private:
    VkPhysicalDeviceMemoryProperties mMemoryProperties = {};
-    VkPhysicalDeviceProperties2 mPhysicalDeviceProperties = {
-        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2
-    };
-    VkPhysicalDeviceFeatures2 mPhysicalDeviceFeatures = {
-        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
-    };
+    VkPhysicalDeviceProperties mPhysicalDeviceProperties = {};
+    VkPhysicalDeviceFeatures mPhysicalDeviceFeatures = {};
    bool mDebugMarkersSupported = false;
    bool mDebugUtilsSupported = false;
    bool mMultiviewEnabled = false;
-    bool mLazilyAllocatedMemorySupported = false;
-    bool mProtectedMemorySupported = false;

    VkFormatList mDepthStencilFormats;
    VkFormatList mBlittableDepthStencilFormats;
@@ -24,14 +24,13 @@
 #include "VulkanHandles.h"
 #include "VulkanPipelineCache.h"
 #include "VulkanReadPixels.h"
+#include "VulkanResourceAllocator.h"
 #include "VulkanSamplerCache.h"
 #include "VulkanStagePool.h"
 #include "VulkanUtility.h"
 #include "backend/DriverEnums.h"
 #include "caching/VulkanDescriptorSetManager.h"
 #include "caching/VulkanPipelineLayoutCache.h"
-#include "memory/ResourceManager.h"
-#include "memory/ResourcePointer.h"

 #include "DriverBase.h"
 #include "private/backend/Driver.h"
@@ -48,6 +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,
@@ -78,9 +92,6 @@ public:
 #endif // FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)

 private:
-    template<typename D>
-    using resource_ptr = fvkmemory::resource_ptr<D>;
-
    static constexpr uint8_t MAX_SAMPLER_BINDING_COUNT = Program::SAMPLER_BINDING_COUNT;

    void debugCommandBegin(CommandStream* cmds, bool synchronous,
@@ -117,16 +128,25 @@ private:
    void collectGarbage();

    VulkanPlatform* mPlatform = nullptr;
-    fvkmemory::ResourceManager mResourceManager;
    std::unique_ptr<VulkanTimestamps> mTimestamps;

-    resource_ptr<VulkanSwapChain> mCurrentSwapChain;
-    resource_ptr<VulkanRenderTarget> mDefaultRenderTarget;
+    // Placeholder resources
+    VulkanTexture* mEmptyTexture;
+    VulkanBufferObject* mEmptyBufferObject;
+
+    VulkanSwapChain* mCurrentSwapChain = nullptr;
+    VulkanRenderTarget* mDefaultRenderTarget = nullptr;
    VulkanRenderPass mCurrentRenderPass = {};
    VmaAllocator mAllocator = VK_NULL_HANDLE;
    VkDebugReportCallbackEXT mDebugCallback = VK_NULL_HANDLE;

    VulkanContext mContext = {};
+    VulkanResourceAllocator mResourceAllocator;
+    VulkanResourceManager mResourceManager;
+
+    // Used for resources that are created synchronously and used and destroyed on the backend
+    // thread.
+    VulkanThreadSafeResourceManager mThreadSafeResourceManager;

    VulkanCommands mCommands;
    VulkanPipelineLayoutCache mPipelineLayoutCache;
@@ -140,18 +160,13 @@ private:
    VulkanDescriptorSetManager mDescriptorSetManager;

    // This is necessary for us to write to push constants after binding a pipeline.
-    struct {
-        resource_ptr<VulkanProgram> program;
+    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;
@@ -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;
@@ -41,7 +42,6 @@ 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;
@@ -69,8 +69,8 @@ VulkanFboCache::~VulkanFboCache() {
            << "Please explicitly call terminate() while the VkDevice is still alive.";
 }

-VkFramebuffer VulkanFboCache::getFramebuffer(FboKey const& config) noexcept {
-    FboMap::iterator iter = mFramebufferCache.find(config);
+VkFramebuffer VulkanFboCache::getFramebuffer(FboKey config) noexcept {
+    auto iter = mFramebufferCache.find(config);
    if (UTILS_LIKELY(iter != mFramebufferCache.end() && iter->second.handle != VK_NULL_HANDLE)) {
        iter.value().timestamp = mCurrentTime;
        return iter->second.handle;
@@ -121,7 +121,7 @@ VkFramebuffer VulkanFboCache::getFramebuffer(FboKey const& config) noexcept {
    return framebuffer;
 }

-VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey const& config) noexcept {
+VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
    auto iter = mRenderPassCache.find(config);
    if (UTILS_LIKELY(iter != mRenderPassCache.end() && iter->second.handle != VK_NULL_HANDLE)) {
        iter.value().timestamp = mCurrentTime;
@@ -255,10 +255,12 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey const& 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),
        };
    }
@@ -362,7 +364,7 @@ void VulkanFboCache::gc() noexcept {
    }
    const uint32_t evictTime = mCurrentTime - TIME_BEFORE_EVICTION;

-    for (FboMap::iterator iter = mFramebufferCache.begin(); iter != mFramebufferCache.end(); ++iter) {
+    for (auto iter = mFramebufferCache.begin(); iter != mFramebufferCache.end(); ++iter) {
        const FboVal fbo = iter->second;
        if (fbo.timestamp < evictTime && fbo.handle) {
            mRenderPassRefCount[iter->first.renderPass]--;
@@ -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];
    };
    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.");
@@ -98,10 +109,10 @@ public:
    ~VulkanFboCache();

    // Retrieves or creates a VkFramebuffer handle.
-    VkFramebuffer getFramebuffer(FboKey const& config) noexcept;
+    VkFramebuffer getFramebuffer(FboKey config) noexcept;

    // Retrieves or creates a VkRenderPass handle.
-    VkRenderPass getRenderPass(RenderPassKey const& config) noexcept;
+    VkRenderPass getRenderPass(RenderPassKey config) noexcept;

    // Evicts old unused Vulkan objects. Call this once per frame.
    void gc() noexcept;
@@ -111,9 +122,7 @@ public:

 private:
    VkDevice mDevice;
-    using FboMap = tsl::robin_map<FboKey, FboVal, FboKeyHashFn, FboKeyEqualFn>;
-    FboMap mFramebufferCache;
-
+    tsl::robin_map<FboKey, FboVal, FboKeyHashFn, FboKeyEqualFn> mFramebufferCache;
    tsl::robin_map<RenderPassKey, RenderPassVal, RenderPassHash, RenderPassEq> mRenderPassCache;
    tsl::robin_map<VkRenderPass, uint32_t> mRenderPassRefCount;
    uint32_t mCurrentTime = 0;
@@ -17,13 +17,9 @@
 #include "VulkanHandles.h"

 #include "VulkanConstants.h"
-
-// TODO: remove this by moving DebugUtils out of VulkanDriver
-#include "VulkanDriver.h"
-
 #include "VulkanMemory.h"
+#include "VulkanResourceAllocator.h"
 #include "VulkanUtility.h"
-#include "vulkan/memory/ResourcePointer.h"
 #include "spirv/VulkanSpirvUtils.h"

 #include <backend/platforms/VulkanPlatform.h>
@@ -36,12 +32,15 @@ namespace filament::backend {

 namespace {

+void flipVertically(VkRect2D* rect, uint32_t framebufferHeight) {
+    rect->offset.y = framebufferHeight - rect->offset.y - rect->extent.height;
+}
+
 void flipVertically(VkViewport* rect, uint32_t framebufferHeight) {
    rect->y = framebufferHeight - rect->y - rect->height;
 }

 void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
-    rect->offset.y = fbHeight - rect->offset.y - rect->extent.height;
    int32_t x = std::max(rect->offset.x, 0);
    int32_t y = std::max(rect->offset.y, 0);
    int32_t right = std::min(rect->offset.x + (int32_t) rect->extent.width, (int32_t) fbWidth);
@@ -63,6 +62,13 @@ inline void fromStageFlags(backend::ShaderStageFlags stage, descriptor_binding_t
    }
 }

+inline VkDescriptorSetLayout createDescriptorSetLayout(VkDevice device,
+        VkDescriptorSetLayoutCreateInfo const& info) {
+    VkDescriptorSetLayout layout;
+    vkCreateDescriptorSetLayout(device, &info, VKALLOC, &layout);
+    return layout;
+}
+
 inline VkShaderStageFlags getVkStage(backend::ShaderStage stage) {
    switch(stage) {
        case backend::ShaderStage::VERTEX:
@@ -74,6 +80,61 @@ inline VkShaderStageFlags getVkStage(backend::ShaderStage stage) {
    }
 }

+inline VkDescriptorSetLayoutCreateInfo getLayoutCreateInfo(DescriptorSetLayout const& layout) {
+    // Note that the following *needs* to be static so that VkDescriptorSetLayoutCreateInfo will not
+    // refer to stack memory.
+    static VkDescriptorSetLayoutBinding toBind[VulkanDescriptorSetLayout::MAX_BINDINGS];
+    uint32_t count = 0;
+
+    for (auto const& binding: layout.bindings) {
+        VkShaderStageFlags stages = 0;
+        VkDescriptorType type;
+
+        if ((binding.stageFlags & ShaderStageFlags::VERTEX) != ShaderStageFlags::NONE) {
+            stages |= VK_SHADER_STAGE_VERTEX_BIT;
+        }
+        if ((binding.stageFlags & ShaderStageFlags::FRAGMENT) != ShaderStageFlags::NONE) {
+            stages |= VK_SHADER_STAGE_FRAGMENT_BIT;
+        }
+        assert_invariant(stages != 0);
+
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER: {
+                type = (binding.flags & DescriptorFlags::DYNAMIC_OFFSET) != DescriptorFlags::NONE
+                               ? VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
+                               : VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT: {
+                type = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
+                break;
+            }
+            case DescriptorType::SHADER_STORAGE_BUFFER:
+                PANIC_POSTCONDITION("Shader storage is not supported.");
+                break;
+        }
+        toBind[count++] = {
+                .binding = binding.binding,
+                .descriptorType = type,
+                .descriptorCount = 1,
+                .stageFlags = stages,
+        };
+    }
+
+    assert_invariant(count != 0 && "Need at least one binding for descriptor set layout.");
+    VkDescriptorSetLayoutCreateInfo dlinfo = {
+            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+            .pNext = nullptr,
+            .bindingCount = count,
+            .pBindings = toBind,
+    };
+    return dlinfo;
+}
+
 using BitmaskGroup = VulkanDescriptorSetLayout::Bitmask;
 BitmaskGroup fromBackendLayout(DescriptorSetLayout const& layout) {
    BitmaskGroup mask;
@@ -87,8 +148,6 @@ BitmaskGroup fromBackendLayout(DescriptorSetLayout const& layout) {
                }
                break;
            }
-            // TODO: properly handle external sampler
-            case DescriptorType::SAMPLER_EXTERNAL:
            case DescriptorType::SAMPLER: {
                fromStageFlags(binding.stageFlags, binding.binding, mask.sampler);
                break;
@@ -105,40 +164,28 @@ BitmaskGroup fromBackendLayout(DescriptorSetLayout const& layout) {
    return mask;
 }

-fvkmemory::resource_ptr<VulkanTexture> initMsaaTexture(
-        fvkmemory::resource_ptr<VulkanTexture> texture, VkDevice device,
-        VkPhysicalDevice physicalDevice, VulkanContext const& context, VmaAllocator allocator,
-        VulkanCommands* commands, fvkmemory::ResourceManager* resManager, uint8_t levels,
-        uint8_t samples, VulkanStagePool& stagePool) {
-    assert_invariant(texture);
-    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);
-
-        msTexture = resource_ptr<VulkanTexture>::construct(resManager, device, physicalDevice,
-                context, allocator, resManager, commands, texture->target, levels, texture->format,
-                samples, texture->width, texture->height, texture->depth, usage, stagePool);
-        texture->setSidecar(msTexture);
-    }
-    return msTexture;
-}
-
 } // anonymous namespace

-void VulkanDescriptorSet::acquire(fvkmemory::resource_ptr<VulkanTexture> texture) {
-    mResources.push_back(texture);
+VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(VkDevice device,
+        DescriptorSetLayout const& layout)
+    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET_LAYOUT), mDevice(device),
+      vklayout(createDescriptorSetLayout(device, getLayoutCreateInfo(layout))),
+      bitmask(fromBackendLayout(layout)),
+      count(Count::fromLayoutBitmask(bitmask)) {
 }

-void VulkanDescriptorSet::acquire(fvkmemory::resource_ptr<VulkanBufferObject> obj) {
-    mResources.push_back(obj);
+VulkanDescriptorSetLayout::~VulkanDescriptorSetLayout() {
+    vkDestroyDescriptorSetLayout(mDevice, vklayout, VKALLOC);
 }

-VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(DescriptorSetLayout const& layout)
-    : bitmask(fromBackendLayout(layout)),
-      count(Count::fromLayoutBitmask(bitmask)) {}
+void VulkanDescriptorSet::acquire(VulkanTexture* texture) {
+    mResources.acquire(texture);
+    mTextures[mTextureCount++] = texture;
+}
+
+void VulkanDescriptorSet::acquire(VulkanBufferObject* bufferObject) {
+    mResources.acquire(bufferObject);
+}

 PushConstantDescription::PushConstantDescription(backend::Program const& program) noexcept {
    mRangeCount = 0;
@@ -163,9 +210,9 @@ PushConstantDescription::PushConstantDescription(backend::Program const& program
    }
 }

-void PushConstantDescription::write(VkCommandBuffer cmdbuf, VkPipelineLayout layout,
+void PushConstantDescription::write(VulkanCommands* commands, VkPipelineLayout layout,
        backend::ShaderStage stage, uint8_t index, backend::PushConstantVariant const& value) {
-
+    VulkanCommandBuffer* cmdbuf = &(commands->get());
    uint32_t binaryValue = 0;
    UTILS_UNUSED_IN_RELEASE auto const& types = mTypes[(uint8_t) stage];
    if (std::holds_alternative<bool>(value)) {
@@ -181,12 +228,13 @@ void PushConstantDescription::write(VkCommandBuffer cmdbuf, VkPipelineLayout lay
        int const ival = std::get<int>(value);
        binaryValue = *reinterpret_cast<uint32_t const*>(&ival);
    }
-    vkCmdPushConstants(cmdbuf, layout, getVkStage(stage), index * ENTRY_SIZE, ENTRY_SIZE,
+    vkCmdPushConstants(cmdbuf->buffer(), layout, getVkStage(stage), index * ENTRY_SIZE, ENTRY_SIZE,
            &binaryValue);
 }

 VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
    : HwProgram(builder.getName()),
+      VulkanResource(VulkanResourceType::PROGRAM),
      mInfo(new(std::nothrow) PipelineInfo(builder)),
      mDevice(device) {

@@ -252,239 +300,153 @@ VulkanProgram::~VulkanProgram() {
 }

 // Creates a special "default" render target (i.e. associated with the swap chain)
-VulkanRenderTarget::VulkanRenderTarget()
-    : HwRenderTarget(0, 0),
-      mOffscreen(false),
-      mProtected(false),
-      mInfo(std::make_unique<Auxiliary>()) {
-    mInfo->rpkey.samples = mInfo->fbkey.samples = 1;
-}
+VulkanRenderTarget::VulkanRenderTarget() :
+    HwRenderTarget(0, 0),
+    VulkanResource(VulkanResourceType::RENDER_TARGET),
+    mOffscreen(false), mSamples(1) {}

-VulkanRenderTarget::~VulkanRenderTarget() = default;
-
-void VulkanRenderTarget::bindToSwapChain(fvkmemory::resource_ptr<VulkanSwapChain> swapchain) {
+void VulkanRenderTarget::bindToSwapChain(VulkanSwapChain& swapChain) {
    assert_invariant(!mOffscreen);
-
-    VkExtent2D const extent = swapchain->getExtent();
+    VkExtent2D const extent = swapChain.getExtent();
+    mColor[0] = { .texture = swapChain.getCurrentColor() };
+    mDepth = { .texture = swapChain.getDepth() };
    width = extent.width;
    height = extent.height;
-    mProtected = swapchain->isProtected();
-
-    VulkanAttachment color = {};
-    color.texture = swapchain->getCurrentColor();
-    mInfo->attachments = {color};
-
-    auto& fbkey = mInfo->fbkey;
-    auto& rpkey = mInfo->rpkey;
-
-    rpkey.colorFormat[0] = color.getFormat();
-    fbkey.width = width;
-    fbkey.height = height;
-    fbkey.color[0] = color.getImageView();
-    fbkey.resolve[0] = VK_NULL_HANDLE;
-
-    if (swapchain->getDepth()) {
-        VulkanAttachment depth = {};
-        depth.texture = swapchain->getDepth();
-        mInfo->attachments.push_back(depth);
-        mInfo->depthIndex = 1;
-
-        rpkey.depthFormat = depth.getFormat();
-        fbkey.depth = depth.getImageView();
-    } else {
-        rpkey.depthFormat = VK_FORMAT_UNDEFINED;
-        fbkey.depth = VK_NULL_HANDLE;
-    }
-    mInfo->colors.set(0);
 }

 VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physicalDevice,
-        VulkanContext const& context, fvkmemory::ResourceManager* resourceManager,
-        VmaAllocator allocator, VulkanCommands* commands, uint32_t width, uint32_t height,
-        uint8_t samples, VulkanAttachment color[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
+        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)
    : HwRenderTarget(width, height),
+      VulkanResource(VulkanResourceType::RENDER_TARGET),
      mOffscreen(true),
-      mProtected(false),
-      mInfo(std::make_unique<Auxiliary>()) {
-    auto& depth = depthStencil[0];
+      mSamples(samples),
+      mLayerCount(layerCount) {
+    for (int index = 0; index < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; index++) {
+        mColor[index] = color[index];
+    }
+    mDepth = depthStencil[0];
+    VulkanTexture* depthTexture = (VulkanTexture*) mDepth.texture;
+
+    if (samples == 1) {
+        return;
+    }

    // Constrain the sample count according to both kinds of sample count masks obtained from
    // VkPhysicalDeviceProperties. This is consistent with the VulkanTexture constructor.
    auto const& limits = context.getPhysicalDeviceLimits();
-    samples = reduceSampleCount(samples, limits.framebufferDepthSampleCounts &
+    mSamples = samples = reduceSampleCount(samples, limits.framebufferDepthSampleCounts &
            limits.framebufferColorSampleCounts);

-    auto& rpkey = mInfo->rpkey;
-    rpkey.samples = samples;
-    rpkey.depthFormat = depth.getFormat();
-    rpkey.viewCount = layerCount;
-
-    auto& fbkey = mInfo->fbkey;
-    fbkey.width = width;
-    fbkey.height = height;
-    fbkey.samples = samples;
-
-    std::vector<VulkanAttachment>& attachments = mInfo->attachments;
-    std::vector<VulkanAttachment> msaa;
-
+    // Create sidecar MSAA textures for color attachments if they don't already exist.
    for (int index = 0; index < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; index++) {
-        VulkanAttachment& attachment = color[index];
-        auto texture = attachment.texture;
-        if (!texture) {
-            rpkey.colorFormat[index] = VK_FORMAT_UNDEFINED;
-            continue;
-        }
-
-        mProtected |= texture->getIsProtected();
-
-        attachments.push_back(attachment);
-        mInfo->colors.set(index);
-
-        rpkey.colorFormat[index] = attachment.getFormat();
-        fbkey.color[index] = attachment.getImageView();
-        fbkey.resolve[index] = VK_NULL_HANDLE;
-
-        if (samples > 1) {
-            VulkanAttachment msaaAttachment = {};
-            if (texture->samples == 1) {
-                auto msaaTexture = initMsaaTexture(texture, device, physicalDevice, context,
-                        allocator, commands, resourceManager, texture->levels, samples, stagePool);
-                if (msaaTexture && msaaTexture->isTransientAttachment()) {
-                    rpkey.usesLazilyAllocatedMemory |= (1 << index);
-                }
-                if (attachment.texture->samples == 1) {
-                    rpkey.needsResolveMask |= (1 << index);
-                }
-                msaaAttachment = {
-                    .texture = msaaTexture,
-                    .layerCount = layerCount,
-                };
-
-                fbkey.resolve[index] = attachment.getImageView();
-            } else {
-                msaaAttachment = {
-                    .texture = texture,
-                    .layerCount = layerCount,
-                };
+        VulkanAttachment const& spec = color[index];
+        VulkanTexture* texture = (VulkanTexture*) spec.texture;
+        if (texture && texture->samples == 1) {
+            auto msTexture = texture->getSidecar();
+            if (UTILS_UNLIKELY(!msTexture)) {
+                // 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, texture->usage,
+                        stagePool, true /* heap allocated */);
+                texture->setSidecar(msTexture);
            }
-            fbkey.color[index] = msaaAttachment.getImageView();
-            msaa.push_back(msaaAttachment);
+            mMsaaAttachments[index] = {.texture = msTexture};
+        }
+        if (texture && texture->samples > 1) {
+            mMsaaAttachments[index] = mColor[index];
        }
    }

-    if (attachments.size() > 0 && samples > 1 && msaa.size() > 0) {
-        mInfo->msaaIndex = (uint8_t) attachments.size();
-        attachments.insert(attachments.end(), msaa.begin(), msaa.end());
+    if (!depthTexture) {
+        return;
    }

-    if (depth.texture) {
-        auto depthTexture = depth.texture;
-        mInfo->depthIndex = (uint8_t) attachments.size();
-        attachments.push_back(depth);
-        fbkey.depth = depth.getImageView();
-        if (samples > 1) {
-            mInfo->msaaDepthIndex = mInfo->depthIndex;
-            if (depthTexture->samples == 1) {
-                // MSAA depth texture must have the mipmap count of 1
-                uint8_t const msLevel = 1;
-                // Create sidecar MSAA texture for the depth attachment if it does not already
-                // exist.
-                auto msaa = initMsaaTexture(depthTexture, device, physicalDevice, context,
-                        allocator, commands, resourceManager, msLevel, samples, stagePool);
-                mInfo->msaaDepthIndex = (uint8_t) attachments.size();
-                attachments.push_back({ .texture = msaa, .layerCount = layerCount });
-            }
-        }
+    // There is no need for sidecar depth if the depth texture is already MSAA.
+    if (depthTexture->samples > 1) {
+        mMsaaDepthAttachment = mDepth;
+        return;
    }
+
+    // MSAA depth texture must have the mipmap count of 1
+    uint8_t const msLevel = 1;
+
+    // Create sidecar MSAA texture for the depth attachment if it does not already exist.
+    VulkanTexture* msTexture = depthTexture->getSidecar();
+    if (UTILS_UNLIKELY(!msTexture)) {
+        msTexture = new VulkanTexture(device, physicalDevice, context, allocator,
+                commands, handleAllocator,
+                depthTexture->target, msLevel, depthTexture->format, samples,
+                depthTexture->width, depthTexture->height, depthTexture->depth, depthTexture->usage,
+                stagePool, true /* heap allocated */);
+        depthTexture->setSidecar(msTexture);
+    }
+
+    mMsaaDepthAttachment = {
+        .texture = msTexture,
+        .level = msLevel,
+        .layer = mDepth.layer,
+    };
 }

 void VulkanRenderTarget::transformClientRectToPlatform(VkRect2D* bounds) const {
-    auto const& extent = getExtent();
+    const auto& extent = getExtent();
+    flipVertically(bounds, extent.height);
    clampToFramebuffer(bounds, extent.width, extent.height);
 }

-void VulkanRenderTarget::transformViewportToPlatform(VkViewport* bounds) const {
+void VulkanRenderTarget::transformClientRectToPlatform(VkViewport* bounds) const {
    flipVertically(bounds, getExtent().height);
 }

+VkExtent2D VulkanRenderTarget::getExtent() const {
+    return {width, height};
+}
+
+VulkanAttachment& VulkanRenderTarget::getColor(int target) {
+    return mColor[target];
+}
+
+VulkanAttachment& VulkanRenderTarget::getMsaaColor(int target) {
+    return mMsaaAttachments[target];
+}
+
+VulkanAttachment& VulkanRenderTarget::getDepth() {
+    return mDepth;
+}
+
+VulkanAttachment& VulkanRenderTarget::getMsaaDepth() {
+    return mMsaaDepthAttachment;
+}
+
 uint8_t VulkanRenderTarget::getColorTargetCount(const VulkanRenderPass& pass) const {
    if (!mOffscreen) {
        return 1;
    }
-    if (pass.currentSubpass == 1) {
-        return mInfo->colors.count();
-    }
    uint8_t count = 0;
-    mInfo->colors.forEachSetBit([&count, &pass](size_t index) {
-        if (!(pass.params.subpassMask & (1 << index))) {
+    for (uint8_t i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
+        if (!mColor[i].texture) {
+            continue;
+        }
+        // NOTE: This must be consistent with VkRenderPass construction (see VulkanFboCache).
+        if (!(pass.params.subpassMask & (1 << i)) || pass.currentSubpass == 1) {
            count++;
        }
-    });
+    }
    return count;
 }

-void VulkanRenderTarget::emitBarriersBeginRenderPass(VulkanCommandBuffer& commands) {
-    auto& attachments = mInfo->attachments;
-    auto samples = mInfo->fbkey.samples;
-    auto barrier = [&commands](VulkanAttachment& attachment, VulkanLayout const layout) {
-        auto tex = attachment.texture;
-        auto const& range = attachment.getSubresourceRange();
-        if (tex->getLayout(range.baseMipLevel, range.baseArrayLayer) != layout &&
-                !tex->transitionLayout(&commands, range, layout)) {
-            // If the layout transition did not emit a barrier, we do it manually here.
-            tex->samplerToAttachmentBarrier(&commands, range);
-        }
-    };
-
-    for (size_t i = 0, count = mInfo->colors.count(); i < count; ++i) {
-        auto& attachment = attachments[i];
-        auto tex = attachment.texture;
-        if (samples == 1 || tex->samples == 1) {
-            barrier(attachment, VulkanLayout::COLOR_ATTACHMENT);
-        }
-    }
-    if (mInfo->msaaIndex != Auxiliary::UNDEFINED_INDEX) {
-        for (size_t i = mInfo->msaaIndex, count = mInfo->msaaIndex + mInfo->colors.count();
-                i < count; ++i) {
-            barrier(attachments[i], VulkanLayout::COLOR_ATTACHMENT);
-        }
-    }
-    if (mInfo->depthIndex != Auxiliary::UNDEFINED_INDEX) {
-        barrier(attachments[mInfo->depthIndex], VulkanLayout::DEPTH_ATTACHMENT);
-    }
-    if (mInfo->msaaDepthIndex != Auxiliary::UNDEFINED_INDEX) {
-        barrier(attachments[mInfo->msaaDepthIndex], VulkanLayout::DEPTH_ATTACHMENT);
-    }
-}
-
-void VulkanRenderTarget::emitBarriersEndRenderPass(VulkanCommandBuffer& commands) {
-    if (isSwapChain()) {
-        return;
-    }
-
-    for (auto& attachment: mInfo->attachments) {
-        auto const& range = attachment.getSubresourceRange();
-        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);
-            }
-        }
-    }
-}
-
 VulkanVertexBufferInfo::VulkanVertexBufferInfo(
        uint8_t bufferCount, uint8_t attributeCount, AttributeArray const& attributes)
    : HwVertexBufferInfo(bufferCount, attributeCount),
+      VulkanResource(VulkanResourceType::VERTEX_BUFFER_INFO),
      mInfo(attributes.size()) {
+
    auto attribDesc = mInfo.mSoa.data<PipelineInfo::ATTRIBUTE_DESCRIPTION>();
    auto bufferDesc = mInfo.mSoa.data<PipelineInfo::BUFFER_DESCRIPTION>();
    auto offsets = mInfo.mSoa.data<PipelineInfo::OFFSETS>();
@@ -522,15 +484,19 @@ VulkanVertexBufferInfo::VulkanVertexBufferInfo(
 }

 VulkanVertexBuffer::VulkanVertexBuffer(VulkanContext& context, VulkanStagePool& stagePool,
-        uint32_t vertexCount, fvkmemory::resource_ptr<VulkanVertexBufferInfo> vbi)
+        VulkanResourceAllocator* allocator,
+        uint32_t vertexCount, Handle<HwVertexBufferInfo> vbih)
    : HwVertexBuffer(vertexCount),
-      vbi(vbi),
-      // TODO: Seems a bit wasteful. can we do better here?
-      mBuffers(MAX_VERTEX_BUFFER_COUNT) {
+      VulkanResource(VulkanResourceType::VERTEX_BUFFER),
+      vbih(vbih),
+      mBuffers(MAX_VERTEX_BUFFER_COUNT), // TODO: can we do better here?
+      mResources(allocator) {
 }

-void VulkanVertexBuffer::setBuffer(fvkmemory::resource_ptr<VulkanBufferObject> bufferObject,
-        uint32_t index) {
+void VulkanVertexBuffer::setBuffer(VulkanResourceAllocator const& allocator,
+        VulkanBufferObject* bufferObject, uint32_t index) {
+    VulkanVertexBufferInfo const* const vbi =
+            const_cast<VulkanResourceAllocator&>(allocator).handle_cast<VulkanVertexBufferInfo*>(vbih);
    size_t const count = vbi->getAttributeCount();
    VkBuffer* const vkbuffers = getVkBuffers();
    int8_t const* const attribToBuffer = vbi->getAttributeToBuffer();
@@ -539,20 +505,48 @@ void VulkanVertexBuffer::setBuffer(fvkmemory::resource_ptr<VulkanBufferObject> b
            vkbuffers[attribIndex] = bufferObject->buffer.getGpuBuffer();
        }
    }
-    mResources.push_back(bufferObject);
+    mResources.acquire(bufferObject);
 }

 VulkanBufferObject::VulkanBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool,
        uint32_t byteCount, BufferObjectBinding bindingType)
    : HwBufferObject(byteCount),
+      VulkanResource(VulkanResourceType::BUFFER_OBJECT),
      buffer(allocator, stagePool, getBufferObjectUsage(bindingType), byteCount),
      bindingType(bindingType) {}

-VulkanRenderPrimitive::VulkanRenderPrimitive(PrimitiveType pt,
-        fvkmemory::resource_ptr<VulkanVertexBuffer> vb,
-        fvkmemory::resource_ptr<VulkanIndexBuffer> ib)
-    : HwRenderPrimitive{.type = pt},
-      vertexBuffer(vb),
-      indexBuffer(ib) {}
+VulkanTimerQuery::VulkanTimerQuery(std::tuple<uint32_t, uint32_t> indices)
+    : VulkanThreadSafeResource(VulkanResourceType::TIMER_QUERY),
+      mStartingQueryIndex(std::get<0>(indices)),
+      mStoppingQueryIndex(std::get<1>(indices)) {}
+
+void VulkanTimerQuery::setFence(std::shared_ptr<VulkanCmdFence> fence) noexcept {
+    std::unique_lock<utils::Mutex> lock(mFenceMutex);
+    mFence = fence;
+}
+
+bool VulkanTimerQuery::isCompleted() noexcept {
+    std::unique_lock<utils::Mutex> lock(mFenceMutex);
+    // QueryValue is a synchronous call and might occur before beginTimerQuery has written anything
+    // into the command buffer, which is an error according to the validation layer that ships in
+    // the Android NDK.  Even when AVAILABILITY_BIT is set, validation seems to require that the
+    // 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;
+}
+
+VulkanTimerQuery::~VulkanTimerQuery() = default;
+
+VulkanRenderPrimitive::VulkanRenderPrimitive(VulkanResourceAllocator* resourceAllocator,
+        PrimitiveType pt, Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh)
+        : VulkanResource(VulkanResourceType::RENDER_PRIMITIVE),
+          mResources(resourceAllocator) {
+    type = pt;
+    vertexBuffer = resourceAllocator->handle_cast<VulkanVertexBuffer*>(vbh);
+    indexBuffer = resourceAllocator->handle_cast<VulkanIndexBuffer*>(ibh);
+    mResources.acquire(vertexBuffer);
+    mResources.acquire(indexBuffer);
+}

 } // namespace filament::backend
@@ -20,14 +20,13 @@
 // This needs to be at the top
 #include "DriverBase.h"

-#include "VulkanAsyncHandles.h"
 #include "VulkanBuffer.h"
-#include "VulkanFboCache.h"
+#include "VulkanResources.h"
 #include "VulkanSwapChain.h"
 #include "VulkanTexture.h"
 #include "VulkanUtility.h"
-#include "vulkan/memory/Resource.h"

+#include <private/backend/SamplerGroup.h>
 #include <backend/Program.h>

 #include <utils/bitset.h>
@@ -35,8 +34,6 @@
 #include <utils/Mutex.h>
 #include <utils/StructureOfArrays.h>

-#include <array>
-
 namespace filament::backend {

 namespace {
@@ -44,11 +41,11 @@ namespace {
 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>());
+    constexpr uint64_t VERTEX_MASK = (1ULL << getVertexStageShift<Bitmask>()) - 1ULL;
+    constexpr uint64_t FRAGMENT_MASK = (1ULL << getFragmentStageShift<Bitmask>()) - 1ULL;
    uint64_t val = mask.getValue();
-    val = ((val & VERTEX_MASK) >> getVertexStageShift<Bitmask>()) |
-        ((val & FRAGMENT_MASK) >> getFragmentStageShift<Bitmask>());
+    val = ((val | VERTEX_MASK) >> getVertexStageShift<Bitmask>()) |
+        ((val | FRAGMENT_MASK) >> getFragmentStageShift<Bitmask>());
    return (uint8_t) Bitmask(val).count();
 }

@@ -57,7 +54,7 @@ inline uint8_t collapsedCount(Bitmask const& mask) {
 class VulkanTimestamps;
 struct VulkanBufferObject;

-struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Resource {
+struct VulkanDescriptorSetLayout : public VulkanResource, HwDescriptorSetLayout {
    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 4;
    static constexpr uint8_t MAX_BINDINGS = 25;

@@ -117,59 +114,51 @@ struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Reso
        }
    };

-    VulkanDescriptorSetLayout(DescriptorSetLayout const& layout);
+    VulkanDescriptorSetLayout(VkDevice device, DescriptorSetLayout const& layout);

-    ~VulkanDescriptorSetLayout() = default;
-
-    VkDescriptorSetLayout getVkLayout() const { return mVkLayout; }
-    void setVkLayout(VkDescriptorSetLayout vklayout) { mVkLayout = vklayout; }
+    ~VulkanDescriptorSetLayout();

+    VkDevice const mDevice;
+    VkDescriptorSetLayout const vklayout;
    Bitmask const bitmask;
    Count const count;
-
-private:
-    VkDescriptorSetLayout mVkLayout = VK_NULL_HANDLE;
 };

-struct VulkanDescriptorSet : public HwDescriptorSet, fvkmemory::Resource {
+struct VulkanDescriptorSet : public VulkanResource, HwDescriptorSet {
 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(VkDescriptorSet rawSet,
-            UniformBufferBitmask const& dynamicUboMask,
-            uint8_t uniqueDynamicUboCount,
+    VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
            OnRecycle&& onRecycleFn)
-        : vkSet(rawSet),
-          dynamicUboMask(dynamicUboMask),
-          uniqueDynamicUboCount(uniqueDynamicUboCount),
+        : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
+          vkSet(rawSet),
+          mResources(allocator),
          mOnRecycleFn(std::move(onRecycleFn)) {}

    ~VulkanDescriptorSet() {
        if (mOnRecycleFn) {
-            mOnRecycleFn(this);
+            mOnRecycleFn();
        }
    }

-    void setOffsets(backend::DescriptorSetOffsetArray&& offsets) noexcept {
-        mOffsets = std::move(offsets);
+    void acquire(VulkanTexture* texture);
+
+    void acquire(VulkanBufferObject* texture);
+
+    bool hasTexture(VulkanTexture* texture) {
+        return std::any_of(mTextures.begin(), mTextures.end(),
+                [texture](auto t) { return t == texture; });
    }

-    backend::DescriptorSetOffsetArray const* getOffsets() {
-        return &mOffsets;
-    }
-
-    void acquire(fvkmemory::resource_ptr<VulkanTexture> texture);
-    void acquire(fvkmemory::resource_ptr<VulkanBufferObject> buffer);
-
+    // TODO: maybe change to fixed size for performance.
    VkDescriptorSet const vkSet;
-    UniformBufferBitmask const dynamicUboMask;
-    uint8_t const uniqueDynamicUboCount;

 private:
-    backend::DescriptorSetOffsetArray mOffsets;
-    std::vector<fvkmemory::resource_ptr<fvkmemory::Resource>> mResources;
+    std::array<VulkanTexture*, 16> mTextures = { nullptr };
+    uint8_t mTextureCount = 0;
+    VulkanAcquireOnlyResourceManager mResources;
    OnRecycle mOnRecycleFn;
 };

@@ -177,11 +166,14 @@ using PushConstantNameArray = utils::FixedCapacityVector<char const*>;
 using PushConstantNameByStage = std::array<PushConstantNameArray, Program::SHADER_TYPE_COUNT>;

 struct PushConstantDescription {
+
    explicit PushConstantDescription(backend::Program const& program) noexcept;

    VkPushConstantRange const* getVkRanges() const noexcept { return mRanges; }
+
    uint32_t getVkRangeCount() const noexcept { return mRangeCount; }
-    void write(VkCommandBuffer cmdbuf, VkPipelineLayout layout, backend::ShaderStage stage,
+
+    void write(VulkanCommands* commands, VkPipelineLayout layout, backend::ShaderStage stage,
            uint8_t index, backend::PushConstantVariant const& value);

 private:
@@ -192,10 +184,12 @@ private:
    uint32_t mRangeCount;
 };

-struct VulkanProgram : public HwProgram, fvkmemory::Resource {
+struct VulkanProgram : public HwProgram, VulkanResource {
+
    using BindingList = CappedArray<uint16_t, MAX_SAMPLER_COUNT>;

    VulkanProgram(VkDevice device, Program const& builder) noexcept;
+
    ~VulkanProgram();

    inline VkShaderModule getVertexShader() const {
@@ -204,6 +198,14 @@ struct VulkanProgram : public HwProgram, fvkmemory::Resource {

    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; }
+
    inline uint32_t getPushConstantRangeCount() const {
        return mInfo->pushConstantDescription.getVkRangeCount();
    }
@@ -212,9 +214,9 @@ struct VulkanProgram : public HwProgram, fvkmemory::Resource {
        return mInfo->pushConstantDescription.getVkRanges();
    }

-    inline void writePushConstant(VkCommandBuffer cmdbuf, VkPipelineLayout layout,
+    inline void writePushConstant(VulkanCommands* commands, VkPipelineLayout layout,
            backend::ShaderStage stage, uint8_t index, backend::PushConstantVariant const& value) {
-        mInfo->pushConstantDescription.write(cmdbuf, layout, stage, index, value);
+        mInfo->pushConstantDescription.write(commands, layout, stage, index, value);
    }

 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
@@ -230,11 +232,26 @@ struct VulkanProgram : public HwProgram, fvkmemory::Resource {
 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
            {}

+        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 };
+
        PushConstantDescription pushConstantDescription;
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+        // We store the sampler name mapped from binding index (only for debug purposes).
+        utils::FixedCapacityVector<std::string> bindingToName;
+#endif
+
    };

    PipelineInfo* mInfo;
@@ -249,88 +266,47 @@ private:
 //
 // We use private inheritance to shield clients from the width / height fields in HwRenderTarget,
 // which are not representative when this is the default render target.
-struct VulkanRenderTarget : private HwRenderTarget, fvkmemory::Resource {
+struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
    // Creates an offscreen render target.
    VulkanRenderTarget(VkDevice device, VkPhysicalDevice physicalDevice,
-            VulkanContext const& context, fvkmemory::ResourceManager* resourceManager,
-            VmaAllocator allocator, VulkanCommands* commands, uint32_t width, uint32_t height,
+            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);

-    ~VulkanRenderTarget();
-
    // Creates a special "default" render target (i.e. associated with the swap chain)
    explicit VulkanRenderTarget();

    void transformClientRectToPlatform(VkRect2D* bounds) const;
-
-    void transformViewportToPlatform(VkViewport* bounds) const;
-
-    inline VkExtent2D getExtent() const {
-        return {width, height};
-    }
-
-    inline VulkanAttachment& getColor0() const {
-        assert_invariant(mInfo->colors[0]);
-        return mInfo->attachments[0];
-    }
-
-    inline VulkanAttachment& getDepth() const {
-        assert_invariant(hasDepth());
-        if (mInfo->fbkey.samples == 1) {
-            return mInfo->attachments[mInfo->depthIndex];
-        }
-        return mInfo->attachments[mInfo->msaaDepthIndex];
-    }
-
-    inline VulkanFboCache::RenderPassKey const& getRenderPassKey() const {
-        return mInfo->rpkey;
-    }
-
-    inline VulkanFboCache::FboKey const& getFboKey() const {
-        return mInfo->fbkey;
-    }
-
-    inline uint8_t getSamples() const {
-        return mInfo->fbkey.samples;
-    }
-
-    uint8_t getColorTargetCount(VulkanRenderPass const& pass) const;
-
-    inline bool hasDepth() const { return mInfo->depthIndex != Auxiliary::UNDEFINED_INDEX; }
-
-    inline bool isSwapChain() const { return !mOffscreen; }
-    inline bool isProtected() const { return mProtected; }
-
-    void bindToSwapChain(fvkmemory::resource_ptr<VulkanSwapChain> swapchain);
-
-    void emitBarriersBeginRenderPass(VulkanCommandBuffer& commands);
-
-    void emitBarriersEndRenderPass(VulkanCommandBuffer& commands);
+    void transformClientRectToPlatform(VkViewport* bounds) const;
+    VkExtent2D getExtent() const;
+    // We return references in the following methods to avoid a copy.
+    VulkanAttachment& getColor(int target);
+    VulkanAttachment& getMsaaColor(int target);
+    VulkanAttachment& getDepth();
+    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);

 private:
-    struct Auxiliary {
-        static constexpr int8_t UNDEFINED_INDEX = -1;
-
-        explicit Auxiliary() noexcept = default;
-
-        VulkanFboCache::RenderPassKey rpkey = {};
-        VulkanFboCache::FboKey fbkey = {};
-        std::vector<VulkanAttachment> attachments;
-        utils::bitset32 colors;
-        int8_t depthIndex = UNDEFINED_INDEX;
-        int8_t msaaDepthIndex = UNDEFINED_INDEX;
-        int8_t msaaIndex = UNDEFINED_INDEX;
-    };
-    bool const mOffscreen;
-    bool mProtected;
-
-    std::unique_ptr<Auxiliary> mInfo;
+    VulkanAttachment mColor[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT] = {};
+    VulkanAttachment mDepth = {};
+    VulkanAttachment mMsaaAttachments[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT] = {};
+    VulkanAttachment mMsaaDepthAttachment = {};
+    const bool mOffscreen : 1;
+    uint8_t mSamples : 7;
+    uint8_t mLayerCount = 1;
 };

 struct VulkanBufferObject;

-struct VulkanVertexBufferInfo : public HwVertexBufferInfo, fvkmemory::Resource {
+struct VulkanVertexBufferInfo : public HwVertexBufferInfo, VulkanResource {
    VulkanVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
            AttributeArray const& attributes);

@@ -377,24 +353,33 @@ private:
    PipelineInfo mInfo;
 };

-struct VulkanVertexBuffer : public HwVertexBuffer, fvkmemory::Resource {
-    VulkanVertexBuffer(VulkanContext& context, VulkanStagePool& stagePool, uint32_t vertexCount,
-            fvkmemory::resource_ptr<VulkanVertexBufferInfo> vbi);
-    void setBuffer(fvkmemory::resource_ptr<VulkanBufferObject> bufferObject, uint32_t index);
+struct VulkanVertexBuffer : public HwVertexBuffer, VulkanResource {
+    VulkanVertexBuffer(VulkanContext& context, VulkanStagePool& stagePool,
+            VulkanResourceAllocator* allocator,
+            uint32_t vertexCount, Handle<HwVertexBufferInfo> vbih);

-    inline VkBuffer const* getVkBuffers() const { return mBuffers.data(); }
-    inline VkBuffer* getVkBuffers() { return mBuffers.data(); }
-    fvkmemory::resource_ptr<VulkanVertexBufferInfo> vbi;
+    void setBuffer(VulkanResourceAllocator const& allocator,
+            VulkanBufferObject* bufferObject, uint32_t index);

+    inline VkBuffer const* getVkBuffers() const {
+        return mBuffers.data();
+    }
+
+    inline VkBuffer* getVkBuffers() {
+        return mBuffers.data();
+    }
+
+    Handle<HwVertexBufferInfo> vbih;
 private:
    utils::FixedCapacityVector<VkBuffer> mBuffers;
-    std::vector<fvkmemory::resource_ptr<VulkanBufferObject>> mResources;
+    FixedSizeVulkanResourceManager<MAX_VERTEX_BUFFER_COUNT> mResources;
 };

-struct VulkanIndexBuffer : public HwIndexBuffer, fvkmemory::Resource {
+struct VulkanIndexBuffer : public HwIndexBuffer, VulkanResource {
    VulkanIndexBuffer(VmaAllocator allocator, VulkanStagePool& stagePool, uint8_t elementSize,
            uint32_t indexCount)
        : HwIndexBuffer(elementSize, indexCount),
+          VulkanResource(VulkanResourceType::INDEX_BUFFER),
          buffer(allocator, stagePool, VK_BUFFER_USAGE_INDEX_BUFFER_BIT, elementSize * indexCount),
          indexType(elementSize == 2 ? VK_INDEX_TYPE_UINT16 : VK_INDEX_TYPE_UINT32) {}

@@ -402,7 +387,7 @@ struct VulkanIndexBuffer : public HwIndexBuffer, fvkmemory::Resource {
    const VkIndexType indexType;
 };

-struct VulkanBufferObject : public HwBufferObject, fvkmemory::Resource {
+struct VulkanBufferObject : public HwBufferObject, VulkanResource {
    VulkanBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool, uint32_t byteCount,
            BufferObjectBinding bindingType);

@@ -410,15 +395,66 @@ struct VulkanBufferObject : public HwBufferObject, fvkmemory::Resource {
    const BufferObjectBinding bindingType;
 };

-struct VulkanRenderPrimitive : public HwRenderPrimitive, fvkmemory::Resource {
-    VulkanRenderPrimitive(PrimitiveType pt, fvkmemory::resource_ptr<VulkanVertexBuffer> vb,
-            fvkmemory::resource_ptr<VulkanIndexBuffer> ib);
-    ~VulkanRenderPrimitive() = default;
-
-    fvkmemory::resource_ptr<VulkanVertexBuffer> vertexBuffer;
-    fvkmemory::resource_ptr<VulkanIndexBuffer> indexBuffer;
+struct VulkanSamplerGroup : public HwSamplerGroup, VulkanResource {
+    // NOTE: we have to use out-of-line allocation here because the size of a Handle<> is limited
+    std::unique_ptr<SamplerGroup> sb;// FIXME: this shouldn't depend on filament::SamplerGroup
+    explicit VulkanSamplerGroup(size_t size) noexcept
+        : VulkanResource(VulkanResourceType::SAMPLER_GROUP),
+          sb(new SamplerGroup(size)) {}
 };

+struct VulkanRenderPrimitive : public HwRenderPrimitive, VulkanResource {
+    VulkanRenderPrimitive(VulkanResourceAllocator* resourceAllocator,
+            PrimitiveType pt, Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh);
+
+    ~VulkanRenderPrimitive() {
+        mResources.clear();
+    }
+
+    VulkanVertexBuffer* vertexBuffer = nullptr;
+    VulkanIndexBuffer* indexBuffer = nullptr;
+
+private:
+    // Keep references to the vertex buffer and the index buffer.
+    FixedSizeVulkanResourceManager<2> mResources;
+};
+
+struct VulkanFence : public HwFence, VulkanResource {
+    VulkanFence()
+        : VulkanResource(VulkanResourceType::FENCE) {}
+
+    explicit VulkanFence(std::shared_ptr<VulkanCmdFence> fence)
+        : VulkanResource(VulkanResourceType::FENCE),
+          fence(fence) {}
+
+    std::shared_ptr<VulkanCmdFence> fence;
+};
+
+struct VulkanTimerQuery : public HwTimerQuery, VulkanThreadSafeResource {
+    explicit VulkanTimerQuery(std::tuple<uint32_t, uint32_t> indices);
+    ~VulkanTimerQuery();
+
+    void setFence(std::shared_ptr<VulkanCmdFence> fence) noexcept;
+
+    bool isCompleted() noexcept;
+
+    uint32_t getStartingQueryIndex() const {
+        return mStartingQueryIndex;
+    }
+
+    uint32_t getStoppingQueryIndex() const {
+        return mStoppingQueryIndex;
+    }
+
+private:
+    uint32_t mStartingQueryIndex;
+    uint32_t mStoppingQueryIndex;
+
+    std::shared_ptr<VulkanCmdFence> mFence;
+    utils::Mutex mFenceMutex;
+};
+
+
 inline constexpr VkBufferUsageFlagBits getBufferObjectUsage(
        BufferObjectBinding bindingType) noexcept {
    switch(bindingType) {
@@ -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:
@@ -31,7 +31,6 @@
 #pragma clang diagnostic ignored "-Wnullability-completeness"
 #pragma clang diagnostic ignored "-Wextra-semi"
 #pragma clang diagnostic ignored "-Wc++98-compat-extra-semi"
-#pragma clang diagnostic ignored "-Wthread-safety-analysis"

 static const PFN_vkGetInstanceProcAddr& vkGetInstanceProcAddr = bluevk::vkGetInstanceProcAddr;
 static const PFN_vkGetDeviceProcAddr& vkGetDeviceProcAddr = bluevk::vkGetDeviceProcAddr;
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Benjamin Doherty	69d5ac2107	Check for invalid texture inside MaterialInstance	2024-09-12 12:11:01 -07:00
Powei Feng	c01d2c09b6	vk: reduce map look-up in descriptor set hot loop DescriptorSetManager's commit() is the inner most of the rendering loop. We want to avoid any hash map look up here. We just store a pointer to the history object where appropriate.	2024-09-12 08:48:25 -07:00
Powei Feng	f37d8b2fc2	vk: refactor descriptor bitmask to use bitset Using bitset enables simpler counting of bits and the use of the handy 'forEachSetBit'. Also cleaned up various unused code.	2024-09-12 08:48:25 -07:00
Sungun Park	6397fa3fe1	Fix a crash for color grading (#8118 ) Set postprocess descriptor set for color grading	2024-09-11 18:51:15 -07:00
Benjamin Doherty	37d1baf0e8	Metal debug logging improvements	2024-09-10 13:36:45 -07:00
Powei Feng	1596a5ec71	vk: Fix descriptor set bitmask padding The sizing of each field in Bitmask has changed, and the description was no longer accurate. This caused msan to fail because the mask is also used as a key to cache.	2024-09-10 12:13:24 -07:00
Mathias Agopian	22e62b92a6	disable compute test to fix build this test is no important for now, since we don't have a finished compute api. the test breaks because it still use the old non-descriptor based API.	2024-09-09 23:09:01 -07:00
Mathias Agopian	8f979dacaf	Merge branch 'main' into ma/descriptor-sets	2024-09-09 22:58:33 -07:00
Benjamin Doherty	e3cb67045b	Remove some more magic numbers	2024-09-09 17:01:53 -07:00
Benjamin Doherty	f4cd9f712a	GLSLPostProcessor clean up, remove magic numbers	2024-09-09 16:51:11 -07:00
Benjamin Doherty	99b3587dd3	Update comment	2024-09-09 12:28:35 -07:00
Benjamin Doherty	8e30b1bfc9	Rename toString	2024-09-09 12:26:53 -07:00
Benjamin Doherty	c0b1daf339	MetalDescriptorSet move implementation into MetalHandles.mm	2024-09-09 12:23:45 -07:00
Benjamin Doherty	efb94da111	Remove deprecated driver APIS	2024-09-09 12:17:27 -07:00
Benjamin Doherty	c538bded6f	Get rid of unused MetalSamplerGroup and related code	2024-09-09 12:17:27 -07:00
Benjamin Doherty	f6d9db14c3	Better DEBUG_LOG	2024-09-09 12:17:27 -07:00
Benjamin Doherty	91ce3a9375	Smaller empty buffer	2024-09-09 12:17:27 -07:00
Mathias Agopian	3aeee1630a	Merge branch 'main' into ma/descriptor-sets	2024-09-06 18:49:04 -07:00
Benjamin Doherty	15d77ed1d7	Metal: respect the shader stage for each descriptor	2024-09-05 11:38:10 -07:00
Mathias Agopian	6a64cdb059	fix build due to a bad merge	2024-09-03 12:14:10 -07:00
Mathias Agopian	cf0cb66b28	Merge branch 'main' into ma/descriptor-sets	2024-09-03 11:25:40 -07:00
Mathias Agopian	3a2333d709	Merge branch 'main' into ma/descriptor-sets	2024-08-30 16:31:23 -07:00
Mathias Agopian	6790a1238b	fix webgl build	2024-08-30 16:28:14 -07:00
Powei Feng	fad5d57053	vk: properly compose swizzle	2024-08-30 15:17:03 -07:00
Powei Feng	e95f0ed6a0	vk: implement createTextureViewSwizzled()	2024-08-30 15:17:03 -07:00
Mathias Agopian	73b0be799c	Merge branch 'main' into ma/descriptor-sets	2024-08-30 15:13:43 -07:00
Mathias Agopian	6de6a2123d	implement createTextureViewSwizzled() in the GL backend	2024-08-28 12:53:50 -07:00
Ben Doherty	d161ef24f5	Update createTextureView API to include swizzling, implement on Metal (#8077 )	2024-08-28 11:52:49 -07:00
Mathias Agopian	c3ea1e45f9	Merge branch 'main' into ma/descriptor-sets	2024-08-28 11:45:02 -07:00
Mathias Agopian	c6c20df474	Merge branch 'main' into ma/descriptor-sets	2024-08-23 23:04:57 -07:00
Ben Doherty	bbb45218ac	Fix, assertion when an external image is used but not yet set (#8073 )	2024-08-21 15:47:07 -07:00
Mathias Agopian	02bdddf50e	fix a too aggressive assert	2024-08-21 15:19:56 -07:00
Powei Feng	21d93ce49d	Fix missing duplication of descriptor set in material instance (#8068 )	2024-08-21 12:07:24 -07:00
Powei Feng	cab5fc561c	GLDescriptorSet: less restrictive size assert (#8070 ) For some platform, the compilation fails because the size of the struct is not exactly 32 bits (but is always less).	2024-08-21 11:51:59 -07:00
Mathias Agopian	c780804fd6	Merge branch 'main' into ma/descriptor-sets	2024-08-21 10:27:31 -07:00
Mathias Agopian	b6d6f8f6e4	Merge branch 'main' into ma/descriptor-sets	2024-08-20 12:08:32 -07:00
Mathias Agopian	f098d08c48	Merge branch 'main' into ma/descriptor-sets	2024-08-19 17:13:44 -07:00
Mathias Agopian	5c08f34e8c	Merge branch 'main' into ma/descriptor-sets	2024-08-19 16:40:36 -07:00
Mathias Agopian	a37d03101e	Merge branch 'main' into ma/descriptor-sets	2024-08-19 11:47:19 -07:00
Mathias Agopian	c0819b781a	Merge branch 'main' into ma/descriptor-sets	2024-08-16 16:03:54 -07:00
Powei Feng	d9bf68ed06	vk: fix missing read-after-write barrier (#8046 ) We relied on layout transition to emit a barrier, but sometimes it's skipped. In that case, we explicitly emit a barrier for the transition from attachment -> sampler. Without this, we'll hit read-after-write sync validation error and produce various artifacts (e.g a fade to white effect) on different platforms.	2024-08-16 15:59:38 -07:00
Powei Feng	32ac10161b	Move morphing buffer offset out of pipeline check (#8047 ) The push constant call should not be tied to the caching of the pipeline.	2024-08-16 13:17:37 -07:00
Benjamin Doherty	d073fb5342	Implement external images / descriptor sets for Metal	2024-08-16 13:08:24 -07:00
Mathias Agopian	8fe7c1ff30	Merge branch 'main' into ma/descriptor-sets	2024-08-14 22:47:16 -07:00
Mathias Agopian	3b12e54573	Merge branch 'main' into ma/descriptor-sets	2024-08-12 15:09:26 -07:00
Mathias Agopian	a1f146ebea	Merge branch 'main' into ma/descriptor-sets	2024-08-12 15:05:43 -07:00
Mathias Agopian	9401d6be9e	fix warning about missing case this is to allow -Werror to work.	2024-08-11 23:53:16 -07:00
Mathias Agopian	111be6ef6d	Merge branch 'main' into ma/descriptor-sets	2024-08-09 15:25:58 -07:00
Mathias Agopian	bd0ca7a9bd	Merge branch 'main' into ma/descriptor-sets	2024-08-08 22:48:25 -07:00
Mathias Agopian	8797f45bd0	Merge branch 'main' into ma/descriptor-sets	2024-08-08 13:27:20 -07:00
Mathias Agopian	f04a25028e	Merge branch 'main' into ma/descriptor-sets	2024-08-01 22:19:16 -07:00
Powei Feng	bb1f58e97c	Fix broken merge for vk (#8007 )	2024-07-31 10:40:09 +08:00
Powei Feng	189e36d539	Merge branch 'main' into ma/descriptor-sets	2024-07-31 10:11:51 +08:00
Mathias Agopian	cba819be8f	Merge branch 'main' into ma/descriptor-sets	2024-07-25 17:25:08 -07:00
Mathias Agopian	41e1e7e54b	Merge branch 'main' into ma/descriptor-sets	2024-07-24 15:38:12 -07:00
Ben Doherty	d8580d93ad	New external image API (#7981 )	2024-07-23 13:06:48 -06:00
Mathias Agopian	fef40a8c5d	Merge branch 'main' into ma/descriptor-sets	2024-07-19 10:09:28 -07:00
Mathias Agopian	d299b1f00e	Merge branch 'main' into ma/descriptor-sets	2024-06-28 13:20:28 -07:00
Mathias Agopian	daa2790c4e	Merge branch 'main' into ma/descriptor-sets	2024-06-28 12:54:46 -07:00
Mathias Agopian	d2cb53e39d	add minimal support for Texture Views in the backends Texture Views are currently limited to the min/max lod, and it's not allowed to have several views of the same texture in a single shader. Only SAMPLEABLE textures can have a texture view. Additionally, setMinMaxLevels() is removed as well as all automatic min/max lods adjustments. This is all handled by filament itself from now on, using Texture Views.	2024-06-28 12:23:58 -07:00
Mathias Agopian	315c8d75bc	vk: repair transmission + MSAA That code was changed as part as what looked like a refactoring unrelated to the descriptor-set change itself. Reverting that thunk fixes the problem, however the revert doesn't have the layout transitions that were added. This PR reverts back to the old behavior, but adds the layout transitions. It looks like the refactored code didn't handle "auto resolve" anymore.	2024-06-28 12:12:27 -07:00
Ben Doherty	c19785b72f	Merge branch 'main' into ma/descriptor-sets	2024-06-27 17:48:34 -07:00
Benjamin Doherty	38b3074835	Fix validation error, bind descriptor sets at bind time	2024-06-27 17:42:24 -07:00
Benjamin Doherty	929ad27606	Pass the correct stage flags	2024-06-27 14:44:59 -07:00
Benjamin Doherty	c12bcbb21e	Bind the actual resource at updateDescriptorSet time	2024-06-27 14:44:57 -07:00
Benjamin Doherty	07f07e0700	Remove log spew	2024-06-27 14:35:20 -07:00
Benjamin Doherty	7e0181b2cf	Merge remote-tracking branch 'origin/main' into ma/descriptor-sets	2024-06-25 18:19:39 -07:00
Mathias Agopian	b4162eb313	fix windows build	2024-06-24 22:57:09 -07:00
Mathias Agopian	a0dc7104b6	fix backend tests The pipeline layout wasn't set properly	2024-06-24 22:08:02 -07:00
Mathias Agopian	80f151ccf7	Merge branch 'main' into ma/descriptor-sets	2024-06-24 16:32:18 -07:00
Benjamin Doherty	26bb99c64e	Pass descriptor set info to spirvToMsl more concicely	2024-06-24 15:19:09 -07:00
Benjamin Doherty	1612cdb353	WIP: Implement Metal descriptor sets	2024-06-21 17:00:47 -07:00
Benjamin Doherty	e75882e4d4	Update MetalArgumentBuffer to support buffers	2024-06-21 16:42:20 -07:00
Benjamin Doherty	2e763149e7	MetalArgumentBuffer support buffers	2024-06-21 16:42:19 -07:00
Mathias Agopian	e40a4c69a3	Merge branch 'main' into ma/descriptor-sets	2024-06-21 09:50:29 -07:00
Mathias Agopian	9a1ee1ff67	fix bad merge	2024-06-18 14:11:04 -07:00
Mathias Agopian	9fb692b2f6	Merge branch 'main' into ma/descriptor-sets	2024-06-18 13:39:02 -07:00
Ben Doherty	ac7c679783	Merge branch 'main' into ma/descriptor-sets	2024-06-18 13:21:36 -07:00
Mathias Agopian	38a3c52629	vk: build new descriptor set framework	2024-06-17 23:19:24 -07:00
Mathias Agopian	29ecfd5fda	generate the proper per-view descriptor-set layout in the shader the per-view descriptor-set layout depends on both the material and the variant.	2024-06-17 23:19:24 -07:00
Mathias Agopian	f11a04722b	new descriptor set api	2024-06-17 23:19:24 -07:00