Capture command callstacks for debugging

2023-10-30 14:42:24 -07:00
114 changed files with 1303 additions and 2343 deletions
@@ -206,7 +206,8 @@ jobs:
          TAG: ${{ steps.git_ref.outputs.tag }}
        run: |
          build\windows\build-github.bat release
-          move out\filament-windows.tgz out\filament-$Env:TAG-windows.tgz
+          cd ..\..
+          move out\filament-windows.tgz out\filament-%TAG%-windows.tgz
        shell: cmd
      - uses: actions/github-script@v6
        env:
@@ -40,20 +40,20 @@ inside the Filament source tree.

 To trigger an incremental debug build:

-```shell
-./build.sh debug
+```
+$ ./build.sh debug
 ```

 To trigger an incremental release build:

-```shell
-./build.sh release
+```
+$ ./build.sh release
 ```

 To trigger both incremental debug and release builds:

-```shell
-./build.sh debug release
+```
+$ ./build.sh debug release
 ```

 To install the libraries and executables in `out/debug/` and `out/release/`, add the `-i` flag.
@@ -76,9 +76,9 @@ The following CMake options are boolean options specific to Filament:

 To turn an option on or off:

-```shell
-cd <cmake-build-directory>
-cmake . -DOPTION=ON       # Replace OPTION with the option name, set to ON / OFF
+```
+$ cd <cmake-build-directory>
+$ cmake . -DOPTION=ON       # Replace OPTION with the option name, set to ON / OFF
 ```

 Options can also be set with the CMake GUI.
@@ -102,38 +102,38 @@ script.
 If you'd like to run `cmake` directly rather than using the build script, it can be invoked as
 follows, with some caveats that are explained further down.

-```shell
-mkdir out/cmake-release
-cd out/cmake-release
-cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../release/filament ../..
+```
+$ mkdir out/cmake-release
+$ cd out/cmake-release
+$ cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../release/filament ../..
 ```

 Your Linux distribution might default to `gcc` instead of `clang`, if that's the case invoke
 `cmake` with the following command:

-```shell
-mkdir out/cmake-release
-cd out/cmake-release
+```
+$ mkdir out/cmake-release
+$ cd out/cmake-release
 # Or use a specific version of clang, for instance /usr/bin/clang-14
-CC=/usr/bin/clang CXX=/usr/bin/clang++ CXXFLAGS=-stdlib=libc++ \
-  cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../release/filament ../..
+$ CC=/usr/bin/clang CXX=/usr/bin/clang++ CXXFLAGS=-stdlib=libc++ \
+    cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../release/filament ../..
 ```

 You can also export the `CC` and `CXX` environment variables to always point to `clang`. Another
 solution is to use `update-alternatives` to both change the default compiler, and point to a
 specific version of clang:

-```shell
-update-alternatives --install /usr/bin/clang clang /usr/bin/clang-14 100
-update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-14 100
-update-alternatives --install /usr/bin/cc cc /usr/bin/clang 100
-update-alternatives --install /usr/bin/c++ c++ /usr/bin/clang++ 100
+```
+$ update-alternatives --install /usr/bin/clang clang /usr/bin/clang-14 100
+$ update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-14 100
+$ update-alternatives --install /usr/bin/cc cc /usr/bin/clang 100
+$ update-alternatives --install /usr/bin/c++ c++ /usr/bin/clang++ 100
 ```

 Finally, invoke `ninja`:

-```shell
-ninja
+```
+$ ninja
 ```

 This will build Filament, its tests and samples, and various host tools.
@@ -143,8 +143,8 @@ This will build Filament, its tests and samples, and various host tools.
 To compile Filament you must have the most recent version of Xcode installed and you need to
 make sure the command line tools are setup by running:

-```shell
-xcode-select --install
+```
+$ xcode-select --install
 ```

 If you wish to run the Vulkan backend instead of the default Metal backend, you must install
@@ -152,11 +152,11 @@ the LunarG SDK, enable "System Global Components", and reboot your machine.

 Then run `cmake` and `ninja` to trigger a build:

-```shell
-mkdir out/cmake-release
-cd out/cmake-release
-cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../release/filament ../..
-ninja
+```
+$ mkdir out/cmake-release
+$ cd out/cmake-release
+$ cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../release/filament ../..
+$ ninja
 ```

 ### iOS
@@ -164,8 +164,8 @@ ninja
 The easiest way to build Filament for iOS is to use `build.sh` and the
 `-p ios` flag. For instance to build the debug target:

-```shell
-./build.sh -p ios debug
+```
+$ ./build.sh -p ios debug
 ```

 See [ios/samples/README.md](./ios/samples/README.md) for more information.
@@ -191,10 +191,10 @@ using `fsutil.exe file queryCaseSensitiveInfo`.
 Next, open `x64 Native Tools Command Prompt for VS 2019`, create a working directory, and run
 CMake in it:

-```bat
-mkdir out
-cd out
-cmake ..
+```
+> mkdir out
+> cd out
+> cmake ..
 ```

 Open the generated solution file `TNT.sln` in Visual Studio.
@@ -204,15 +204,15 @@ target in the _Solution Explorer_ and choose _Build_ to build a specific target.

 For example, build the `material_sandbox` sample and run it from the `out` directory with:

-```bat
-samples\Debug\material_sandbox.exe ..\assets\models\monkey\monkey.obj
+```
+> samples\Debug\material_sandbox.exe ..\assets\models\monkey\monkey.obj
 ```

 You can also use CMake to invoke the build without opening Visual Studio. For example, from the
 `out` folder run the following command.

-```bat
-cmake --build . --target gltf_viewer --config Release
+```
+> cmake --build . --target gltf_viewer --config Release
 ```

 ### Android
@@ -237,8 +237,8 @@ To build Android on Windows machines, see [android/Windows.md](android/Windows.m
 The easiest way to build Filament for Android is to use `build.sh` and the
 `-p android` flag. For instance to build the release target:

-```shell
-./build.sh -p android release
+```
+$ ./build.sh -p android release
 ```

 Run `build.sh -h` for more information.
@@ -248,23 +248,23 @@ Run `build.sh -h` for more information.
 Invoke CMake in a build directory of your choice, inside of filament's directory. The commands
 below show how to build Filament for ARM 64-bit (`aarch64`).

-```shell
-mkdir out/android-build-release-aarch64
-cd out/android-build-release-aarch64
-cmake -G Ninja -DCMAKE_TOOLCHAIN_FILE=../../build/toolchain-aarch64-linux-android.cmake \
-      -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../android-release/filament ../..
+```
+$ mkdir out/android-build-release-aarch64
+$ cd out/android-build-release-aarch64
+$ cmake -G Ninja -DCMAKE_TOOLCHAIN_FILE=../../build/toolchain-aarch64-linux-android.cmake \
+        -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../android-release/filament ../..
 ```

 And then invoke `ninja`:

-```shell
-ninja install
+```
+$ ninja install
 ```

 or

-```shell
-ninja install/strip
+```
+$ ninja install/strip
 ```

 This will generate Filament's Android binaries in `out/android-release`. This location is important
@@ -296,8 +296,8 @@ AAR.
 Alternatively you can build the AAR from the command line by executing the following in the
 `android/` directory:

-```shell
-./gradlew -Pcom.google.android.filament.dist-dir=../../out/android-release/filament assembleRelease
+```
+$ ./gradlew -Pcom.google.android.filament.dist-dir=../../out/android-release/filament assembleRelease
 ```

 The `-Pcom.google.android.filament.dist-dir` can be used to specify a different installation
@@ -311,7 +311,7 @@ sure to add the newly created module as a dependency to your application.
 If you do not wish to include all supported ABIs, make sure to create the appropriate flavors in
 your Gradle build file. For example:

-```gradle
+```
 flavorDimensions 'cpuArch'
 productFlavors {
    arm8 {
@@ -353,7 +353,7 @@ started, follow the instructions for building Filament on your platform ([macOS]
 Next, you need to install the Emscripten SDK. The following instructions show how to install the
 same version that our continuous builds use.

-```shell
+```
 cd <your chosen parent folder for the emscripten SDK>
 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
@@ -364,7 +364,7 @@ source ./emsdk_env.sh

 After this you can invoke the [easy build](#easy-build) script as follows:

-```shell
+```
 export EMSDK=<your chosen home for the emscripten SDK>
 ./build.sh -p webgl release
 ```
@@ -374,7 +374,7 @@ creates a `samples` folder that can be used as the root of a simple static web s
 cannot open the HTML directly from the filesystem due to CORS. We recommend using the emrun tool
 to create a quick localhost server:

-```shell
+```
 emrun out/cmake-webgl-release/web/samples --no_browser --port 8000
 ```

@@ -395,7 +395,7 @@ Some of the samples accept FBX/OBJ meshes while others rely on the `filamesh` fi
 generate a `filamesh ` file from an FBX/OBJ asset, run the `filamesh` tool
 (`./tools/filamesh/filamesh` in your build directory):

-```shell
+```
 filamesh ./assets/models/monkey/monkey.obj monkey.filamesh
 ```

@@ -405,7 +405,7 @@ files for the IBL (which are PNGs containing `R11F_G11F_B10F` data) or a path to
 containing two `.ktx` files (one for the IBL itself, one for the skybox). To generate an IBL
 simply use this command:

-```shell
+```
 cmgen -f ktx -x ./ibls/ my_ibl.exr
 ```

@@ -427,9 +427,9 @@ value is the desired roughness between 0 and 1.
 To generate the documentation you must first install `doxygen` and `graphviz`, then run the 
 following commands:

-```shell
-cd filament/filament
-doxygen docs/doxygen/filament.doxygen
+```
+$ cd filament/filament
+$ doxygen docs/doxygen/filament.doxygen
 ```

 Finally simply open `docs/html/index.html` in your web browser.
@@ -439,7 +439,7 @@ Finally simply open `docs/html/index.html` in your web browser.
 To try out Filament's Vulkan support with SwiftShader, first build SwiftShader and set the
 `SWIFTSHADER_LD_LIBRARY_PATH` variable to the folder that contains `libvk_swiftshader.dylib`:

-```shell
+```
 git clone https://github.com/google/swiftshader.git
 cd swiftshader/build
 cmake .. &&  make -j
@@ -454,7 +454,7 @@ Continuous testing turnaround can be quite slow if you need to build SwiftShader
 provide an Ubuntu-based Docker image that has it already built. The Docker image also includes
 everything necessary for building Filament. You can fetch and run the image as follows:

-```shell
+```
 docker pull ghcr.io/filament-assets/swiftshader
 docker run -it ghcr.io/filament-assets/swiftshader
 ```
@@ -3,14 +3,6 @@
 # ==================================================================================================
 cmake_minimum_required(VERSION 3.19)

-# ==================================================================================================
-# Toolchain configuration
-# ==================================================================================================
-if (APPLE AND NOT IOS)
-    # This must be set before project() is called
-    set(CMAKE_OSX_DEPLOYMENT_TARGET 10.15 CACHE STRING "")
-endif()
-
 # ==================================================================================================
 # Project declaration
 # ==================================================================================================
@@ -27,7 +27,7 @@ again.

 ## Code Style

-See [CODE_STYLE.md](/CODE_STYLE.md)
+See [CodeStyle.md](/CODE_STYLE.md)

 ## Code reviews

@@ -7,3 +7,9 @@ for next branch cut* header.
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).

 ## Release notes for next branch cut
+
+- Fix possible NPE when updating fog options from Java/Kotlin
+- The `emissive` property was not applied properly to `MASKED` materials, and could cause
+  dark fringes to appear (recompile materials)
+- Allow glTF materials with transmission/volume extensions to choose their alpha mode
+  instead of forcing `MASKED`
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.43.1'
+    implementation 'com.google.android.filament:filament-android:1.42.0'
 }
 ```

@@ -50,8 +50,8 @@ 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', '~> 1.43.1'
+```
+pod 'Filament', '~> 1.42.0'
 ```

 ### Snapshots
@@ -7,30 +7,8 @@ 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).

-## v1.43.1
-
-## v1.43.0
-
- gltfio: Fix possible change of scale sign when decomposing transform matrix for animation
- engine: Fixes "stable" shadows (see b/299310624)
-
-## v1.42.2
-
- Fix possible NPE when updating fog options from Java/Kotlin
- The `emissive` property was not applied properly to `MASKED` materials, and could cause
-  dark fringes to appear (recompile materials)
- Allow glTF materials with transmission/volume extensions to choose their alpha mode
-  instead of forcing `MASKED`
- Fix a crash in gltfio when not using ubershaders
- Use flatmat for mat parameter in jsbinding
- Fix TextureFlags for sheenRoughnessMap when textures of sheenRoughnessMap and sheenColorMap is same
- Directional shadows can now be transformed (b/297095805)
-
 ## v1.42.1

- Fix potential `EXC_BAD_ACCESS` with Metal backend: b/297059776
- `setFrameCompletedCallback` now takes a `backend::CallbackHandler`.
-
 ## v1.42.0

 - engine: add preliminary support for instanced stereoscopic rendering [⚠️ **Recompile materials**]
@@ -79,8 +79,7 @@ Java_com_google_android_filament_LightManager_nBuilderShadowOptions(JNIEnv* env,
        jfloat shadowFarHint, jboolean stable, jboolean lispsm,
        jfloat polygonOffsetConstant, jfloat polygonOffsetSlope,
        jboolean screenSpaceContactShadows, jint stepCount,
-        jfloat maxShadowDistance, jboolean elvsm, jfloat blurWidth, jfloat shadowBulbRadius,
-        jfloatArray transform) {
+        jfloat maxShadowDistance, jboolean elvsm, jfloat blurWidth, jfloat shadowBulbRadius) {
    LightManager::Builder *builder = (LightManager::Builder *) nativeBuilder;
    LightManager::ShadowOptions shadowOptions {
            .mapSize = (uint32_t)mapSize,
@@ -103,18 +102,12 @@ Java_com_google_android_filament_LightManager_nBuilderShadowOptions(JNIEnv* env,
            },
            .shadowBulbRadius = shadowBulbRadius
    };
-
    jfloat *nativeSplits = env->GetFloatArrayElements(splitPositions, NULL);
    const jsize splitCount = std::min((jsize) 3, env->GetArrayLength(splitPositions));
-    std::copy_n(nativeSplits, splitCount, shadowOptions.cascadeSplitPositions);
+    for (jsize i = 0; i < splitCount; i++) {
+        shadowOptions.cascadeSplitPositions[i] = nativeSplits[i];
+    }
    env->ReleaseFloatArrayElements(splitPositions, nativeSplits, 0);
-
-    jfloat* nativeTransform = env->GetFloatArrayElements(transform, NULL);
-    std::copy_n(nativeTransform,
-                std::min(4, env->GetArrayLength(transform)),
-                shadowOptions.transform.xyzw.v);
-    env->ReleaseFloatArrayElements(transform, nativeTransform, 0);
-
    builder->shadowOptions(shadowOptions);
 }

@@ -258,7 +258,6 @@ public class LightManager {
         * shadows that are too far and wouldn't contribute to the scene much, improving
         * performance and quality. This value is always positive.
         * Use 0.0f to use the camera far distance.
-         * This only affect directional lights.
         */
        public float shadowFar = 0.0f;

@@ -369,17 +368,6 @@ public class LightManager {
         * enabled. (2cm by default).
         */
        public float shadowBulbRadius = 0.02f;
-
-        /**
-         * Transforms the shadow direction. Must be a unit quaternion.
-         * The default is identity.
-         * Ignored if the light type isn't directional. For artistic use. Use with caution.
-         * The quaternion is stored as the imaginary part in the first 3 elements and the real
-         * part in the last element of the transform array.
-         */
-        @NonNull
-        @Size(min = 4, max = 4)
-        public float[] transform = { 0.0f, 0.0f, 0.0f, 1.0f };
    }

    public static class ShadowCascades {
@@ -518,7 +506,7 @@ public class LightManager {
                    options.polygonOffsetConstant, options.polygonOffsetSlope,
                    options.screenSpaceContactShadows,
                    options.stepCount, options.maxShadowDistance,
-                    options.elvsm, options.blurWidth, options.shadowBulbRadius, options.transform);
+                    options.elvsm, options.blurWidth, options.shadowBulbRadius);
            return this;
        }

@@ -1181,7 +1169,7 @@ public class LightManager {
             boolean stable, boolean lispsm,
             float polygonOffsetConstant, float polygonOffsetSlope,
             boolean screenSpaceContactShadows, int stepCount, float maxShadowDistance,
-             boolean elvsm, float blurWidth, float shadowBulbRadius, float[] transform);
+             boolean elvsm, float blurWidth, float shadowBulbRadius);
    private static native void nBuilderCastLight(long nativeBuilder, boolean enabled);
    private static native void nBuilderPosition(long nativeBuilder, float x, float y, float z);
    private static native void nBuilderDirection(long nativeBuilder, float x, float y, float z);
@@ -1418,17 +1418,6 @@ public class View {
         * limit highlights to this value before bloom [10, +inf]
         */
        public float highlight = 1000.0f;
-        /**
-         * Bloom quality level.
-         * LOW (default): use a more optimized down-sampling filter, however there can be artifacts
-         *      with dynamic resolution, this can be alleviated by using the homogenous mode.
-         * MEDIUM: Good balance between quality and performance.
-         * HIGH: In this mode the bloom resolution is automatically increased to avoid artifacts.
-         *      This mode can be significantly slower on mobile, especially at high resolution.
-         *      This mode greatly improves the anamorphic bloom.
-         */
-        @NonNull
-        public QualityLevel quality = QualityLevel.LOW;
        /**
         * enable screen-space lens flare
         */
@@ -52,7 +52,6 @@ set(GLTFIO_SRCS
        ${GLTFIO_DIR}/include/gltfio/FilamentInstance.h
        ${GLTFIO_DIR}/include/gltfio/MaterialProvider.h
        ${GLTFIO_DIR}/include/gltfio/NodeManager.h
-        ${GLTFIO_DIR}/include/gltfio/TrsTransformManager.h
        ${GLTFIO_DIR}/include/gltfio/ResourceLoader.h
        ${GLTFIO_DIR}/include/gltfio/TextureProvider.h
        ${GLTFIO_DIR}/include/gltfio/math.h
@@ -70,12 +69,10 @@ set(GLTFIO_SRCS
        ${GLTFIO_DIR}/src/FilamentAsset.cpp
        ${GLTFIO_DIR}/src/FilamentInstance.cpp
        ${GLTFIO_DIR}/src/FNodeManager.h
-        ${GLTFIO_DIR}/src/FTrsTransformManager.h
        ${GLTFIO_DIR}/src/GltfEnums.h
        ${GLTFIO_DIR}/src/Ktx2Provider.cpp
        ${GLTFIO_DIR}/src/MaterialProvider.cpp
        ${GLTFIO_DIR}/src/NodeManager.cpp
-        ${GLTFIO_DIR}/src/TrsTransformManager.cpp
        ${GLTFIO_DIR}/src/ResourceLoader.cpp
        ${GLTFIO_DIR}/src/StbProvider.cpp
        ${GLTFIO_DIR}/src/TangentsJob.cpp
@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.43.1
+VERSION_NAME=1.42.0

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

@@ -87,9 +87,9 @@ compile Filament's native library and Filament's AAR for this project. The easie
 is to install all the required dependencies and to run the following commands at the root of the
 source tree:

-```shell
-./build.sh -p desktop -i release
-./build.sh -p android release
+```
+$ ./build.sh -p desktop -i release
+$ ./build.sh -p android release
 ```

 This will build all the native components and the AAR required by this sample application.
@@ -100,8 +100,8 @@ distribution/install directory for desktop (produced by make/ninja install). Thi
 contain `bin/matc` and `bin/cmgen`.

 Example:
-```shell
-./gradlew -Pfilament_tools_dir=../../dist-release assembleDebug
+```
+$ ./gradlew -Pfilament_tools_dir=../../dist-release assembleDebug
 ```

 ## Important: SDK location
@@ -110,24 +110,14 @@ Either ensure your `ANDROID_HOME` environment variable is set or make sure the r
 contains a `local.properties` file with the `sdk.dir` property pointing to your installation of
 the Android SDK.

-## Compiling
-
-### Android Studio
+## Android Studio

 You must use the latest stable release of Android Studio. To open the project, point Studio to the
 `android` folder. After opening the project and syncing to gradle, select the sample of your choice
 using the drop-down widget in the toolbar.

+## Compiling
+
 To compile and run each sample make sure you have selected the appropriate build variant
 (arm7, arm8, x86 or x86_64). If you are not sure you can simply select the "universal"
 variant which includes all the other ones.
-
-### Command Line
-
-From the `android` directory in the project root:
-
-```shell
-./gradlew :samples:sample-hello-triangle:installDebug
-```
-
-Replace `sample-hello-triangle` with your preferred project.
@@ -210,10 +210,13 @@ function build_desktop_target {
    echo "Building ${lc_target} in out/cmake-${lc_target}..."
    mkdir -p "out/cmake-${lc_target}"

-    pushd "out/cmake-${lc_target}" > /dev/null
+    cd "out/cmake-${lc_target}"

+    # On macOS, set the deployment target to 10.15.
    local lc_name=$(echo "${UNAME}" | tr '[:upper:]' '[:lower:]')
    if [[ "${lc_name}" == "darwin" ]]; then
+        local deployment_target="-DCMAKE_OSX_DEPLOYMENT_TARGET=10.15"
+
        if [[ "${BUILD_UNIVERSAL_LIBRARIES}" == "true" ]]; then
            local architectures="-DCMAKE_OSX_ARCHITECTURES=arm64;x86_64"
        fi
@@ -230,6 +233,7 @@ function build_desktop_target {
            ${MATDBG_OPTION} \
            ${MATOPT_OPTION} \
            ${ASAN_UBSAN_OPTION} \
+            ${deployment_target} \
            ${architectures} \
            ../..
    fi
@@ -243,13 +247,12 @@ function build_desktop_target {
    if [[ -d "../${lc_target}/filament" ]]; then
        if [[ "${ISSUE_ARCHIVES}" == "true" ]]; then
            echo "Generating out/filament-${lc_target}-${LC_UNAME}.tgz..."
-            pushd "../${lc_target}" > /dev/null
+            cd "../${lc_target}"
            tar -czvf "../filament-${lc_target}-${LC_UNAME}.tgz" filament
-            popd > /dev/null
        fi
    fi

-    popd > /dev/null
+    cd ../..
 }

 function build_desktop {
@@ -267,7 +270,7 @@ function build_webgl_with_target {

    echo "Building WebGL ${lc_target}..."
    mkdir -p "out/cmake-webgl-${lc_target}"
-    pushd "out/cmake-webgl-${lc_target}" > /dev/null
+    cd "out/cmake-webgl-${lc_target}"

    if [[ ! "${BUILD_TARGETS}" ]]; then
        BUILD_TARGETS=${BUILD_CUSTOM_TARGETS}
@@ -304,17 +307,17 @@ function build_webgl_with_target {

        if [[ "${ISSUE_ARCHIVES}" == "true" ]]; then
            echo "Generating out/filament-${lc_target}-web.tgz..."
-            pushd web/filament-js > /dev/null
+            cd web/filament-js
            tar -cvf "../../../filament-${lc_target}-web.tar" filament.js
            tar -rvf "../../../filament-${lc_target}-web.tar" filament.wasm
            tar -rvf "../../../filament-${lc_target}-web.tar" filament.d.ts
-            popd > /dev/null
+            cd -
            gzip -c "../filament-${lc_target}-web.tar" > "../filament-${lc_target}-web.tgz"
            rm "../filament-${lc_target}-web.tar"
        fi
    fi

-    popd > /dev/null
+    cd ../..
 }

 function build_webgl {
@@ -345,7 +348,7 @@ function build_android_target {
    echo "Building Android ${lc_target} (${arch})..."
    mkdir -p "out/cmake-android-${lc_target}-${arch}"

-    pushd "out/cmake-android-${lc_target}-${arch}" > /dev/null
+    cd "out/cmake-android-${lc_target}-${arch}"

    if [[ ! -d "CMakeFiles" ]] || [[ "${ISSUE_CMAKE_ALWAYS}" == "true" ]]; then
        cmake \
@@ -364,7 +367,7 @@ function build_android_target {
    # We must always install Android libraries to build the AAR
    ${BUILD_COMMAND} install

-    popd > /dev/null
+    cd ../..
 }

 function build_android_arch {
@@ -385,9 +388,9 @@ function archive_android {
    if [[ -d "out/android-${lc_target}/filament" ]]; then
        if [[ "${ISSUE_ARCHIVES}" == "true" ]]; then
            echo "Generating out/filament-android-${lc_target}-${LC_UNAME}.tgz..."
-            pushd "out/android-${lc_target}" > /dev/null
+            cd "out/android-${lc_target}"
            tar -czvf "../filament-android-${lc_target}-${LC_UNAME}.tgz" filament
-            popd > /dev/null
+            cd ../..
        fi
    fi
 }
@@ -465,7 +468,7 @@ function build_android {
        archive_android "Release"
    fi

-    pushd android > /dev/null
+    cd android

    if [[ "${ISSUE_DEBUG_BUILD}" == "true" ]]; then
        ./gradlew \
@@ -566,7 +569,7 @@ function build_android {
        fi
    fi

-    popd > /dev/null
+    cd ..
 }

 function build_ios_target {
@@ -577,7 +580,7 @@ function build_ios_target {
    echo "Building iOS ${lc_target} (${arch}) for ${platform}..."
    mkdir -p "out/cmake-ios-${lc_target}-${arch}"

-    pushd "out/cmake-ios-${lc_target}-${arch}" > /dev/null
+    cd "out/cmake-ios-${lc_target}-${arch}"

    if [[ ! -d "CMakeFiles" ]] || [[ "${ISSUE_CMAKE_ALWAYS}" == "true" ]]; then
        cmake \
@@ -601,7 +604,7 @@ function build_ios_target {
        ${BUILD_COMMAND} ${INSTALL_COMMAND}
    fi

-    popd > /dev/null
+    cd ../..
 }

 function archive_ios {
@@ -610,9 +613,9 @@ function archive_ios {
    if [[ -d "out/ios-${lc_target}/filament" ]]; then
        if [[ "${ISSUE_ARCHIVES}" == "true" ]]; then
            echo "Generating out/filament-${lc_target}-ios.tgz..."
-            pushd "out/ios-${lc_target}" > /dev/null
+            cd "out/ios-${lc_target}"
            tar -czvf "../filament-${lc_target}-ios.tgz" filament
-            popd > /dev/null
+            cd ../..
        fi
    fi
 }
@@ -671,14 +674,14 @@ function build_web_docs {

    mkdir -p out/web-docs
    cp -f docs/web-docs-package.json out/web-docs/package.json
-    pushd out/web-docs > /dev/null
+    cd out/web-docs

    npm install > /dev/null

    # Generate documents
    npx markdeep-rasterizer ../../docs/Filament.md.html ../../docs/Materials.md.html  ../../docs/

-    popd > /dev/null
+    cd ../..
 }

 function validate_build_command {
@@ -751,16 +754,6 @@ function run_tests {
    fi
 }

-function check_debug_release_build {
-    if [[ "${ISSUE_DEBUG_BUILD}" == "true" || "${ISSUE_RELEASE_BUILD}" == "true" ]]; then
-        "$@";
-    else
-        echo "You must declare a debug or release target for $@ builds."
-        echo ""
-        exit 1
-    fi
-}
-
 # Beginning of the script

 pushd "$(dirname "$0")" > /dev/null
@@ -769,7 +762,7 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
    case ${opt} in
        h)
            print_help
-            exit 0
+            exit 1
            ;;
        a)
            ISSUE_ARCHIVES=true
@@ -805,7 +798,7 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
            platforms=$(echo "${OPTARG}" | tr ',' '\n')
            for platform in ${platforms}
            do
-                case $(echo "${platform}" | tr '[:upper:]' '[:lower:]') in
+                case ${platform} in
                    desktop)
                        ISSUE_DESKTOP_BUILD=true
                    ;;
@@ -824,12 +817,6 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
                        ISSUE_DESKTOP_BUILD=true
                        ISSUE_WEBGL_BUILD=false
                    ;;
-                    *)
-                        echo "Unknown platform ${platform}"
-                        echo "Platform must be one of [desktop|android|ios|webgl|all]"
-                        echo ""
-                        exit 1
-                    ;;    
                esac
            done
            ;;
@@ -842,7 +829,7 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
            abis=$(echo "${OPTARG}" | tr ',' '\n')
            for abi in ${abis}
            do
-                case $(echo "${abi}" | tr '[:upper:]' '[:lower:]') in
+                case ${abi} in
                    armeabi-v7a)
                        ABI_ARMEABI_V7A=true
                    ;;
@@ -861,12 +848,6 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
                        ABI_X86=true
                        ABI_X86_64=true
                    ;;
-                    *)
-                        echo "Unknown abi ${abi}"
-                        echo "ABI must be one of [armeabi-v7a|arm64-v8a|x86|x86_64|all]"
-                        echo ""
-                        exit 1
-                    ;;
                esac
            done
            ;;
@@ -930,9 +911,9 @@ fi
 shift $((OPTIND - 1))

 for arg; do
-    if [[ $(echo "${arg}" | tr '[:upper:]' '[:lower:]') == "release" ]]; then
+    if [[ "${arg}" == "release" ]]; then
        ISSUE_RELEASE_BUILD=true
-    elif [[ $(echo "${arg}" | tr '[:upper:]' '[:lower:]') == "debug" ]]; then
+    elif [[ "${arg}" == "debug" ]]; then
        ISSUE_DEBUG_BUILD=true
    else
        BUILD_CUSTOM_TARGETS="${BUILD_CUSTOM_TARGETS} ${arg}"
@@ -950,19 +931,19 @@ if [[ "${ISSUE_CLEAN_AGGRESSIVE}" == "true" ]]; then
 fi

 if [[ "${ISSUE_DESKTOP_BUILD}" == "true" ]]; then
-    check_debug_release_build build_desktop
+    build_desktop
 fi

 if [[ "${ISSUE_ANDROID_BUILD}" == "true" ]]; then
-    check_debug_release_build build_android
+    build_android
 fi

 if [[ "${ISSUE_IOS_BUILD}" == "true" ]]; then
-    check_debug_release_build build_ios
+    build_ios
 fi

 if [[ "${ISSUE_WEBGL_BUILD}" == "true" ]]; then
-    check_debug_release_build build_webgl
+    build_webgl
 fi

 if [[ "${ISSUE_WEB_DOCS}" == "true" ]]; then
@@ -230,8 +230,6 @@ set(MATERIAL_SRCS
        src/materials/flare/flare.mat
        src/materials/blitLow.mat
        src/materials/bloom/bloomDownsample.mat
-        src/materials/bloom/bloomDownsample2x.mat
-        src/materials/bloom/bloomDownsample9.mat
        src/materials/bloom/bloomUpsample.mat
        src/materials/ssao/bilateralBlur.mat
        src/materials/ssao/bilateralBlurBentNormals.mat
@@ -61,7 +61,7 @@ with the platform name, for example, `filament-20181009-linux.tgz`.

 Create a file, `main.cpp`, in the same directory with the following contents:

-```c++
+```
 #include <filament/FilamentAPI.h>
 #include <filament/Engine.h>

@@ -91,7 +91,7 @@ Copy your platform's Makefile below into a `Makefile` inside the same directory.

 ### Linux

-```make
+```
 FILAMENT_LIBS=-lfilament -lbackend -lbluegl -lbluevk -lfilabridge -lfilaflat -lutils -lgeometry -lsmol-v -lvkshaders -libl
 CC=clang++

@@ -109,7 +109,7 @@ clean:

 ### macOS

-```make
+```
 FILAMENT_LIBS=-lfilament -lbackend -lbluegl -lbluevk -lfilabridge -lfilaflat -lutils -lgeometry -lsmol-v -lvkshaders -libl
 FRAMEWORKS=-framework Cocoa -framework Metal -framework CoreVideo
 CC=clang++
@@ -137,7 +137,7 @@ be sure to also include `matdbg.lib` in `FILAMENT_LIBS`.
 When building Filament from source, the `USE_STATIC_CRT` CMake option can be
 used to change the run-time library version.

-```make
+```
 FILAMENT_LIBS=filament.lib backend.lib bluegl.lib bluevk.lib filabridge.lib filaflat.lib \
              utils.lib geometry.lib smol-v.lib ibl.lib vkshaders.lib
 CC=cl.exe
@@ -171,12 +171,12 @@ and invoke `nmake` instead of `make`.

 ### Generating C++ documentation

-To generate the documentation you must first install `doxygen` and `graphviz`, then run the
+To generate the documentation you must first install `doxygen` and `graphviz`, then run the 
 following commands:

-```shell
-cd filament/filament
-doxygen docs/doxygen/filament.doxygen
+```
+$ cd filament/filament
+$ doxygen docs/doxygen/filament.doxygen
 ```

 Finally simply open `docs/html/index.html` in your web browser.
@@ -412,7 +412,6 @@ if (APPLE)
        test/test_RenderExternalImage.cpp
        test/test_StencilBuffer.cpp
        test/test_Scissor.cpp
-        test/test_MipLevels.cpp
        )

    target_link_libraries(backend_test PRIVATE
@@ -813,22 +813,22 @@ struct SamplerParams { // NOLINT
    struct Hasher {
        size_t operator()(SamplerParams p) const noexcept {
            // we don't use std::hash<> here, so we don't have to include <functional>
-            return *reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&p));
+            return *reinterpret_cast<uint64_t const*>(reinterpret_cast<char const*>(&p));
        }
    };

    struct EqualTo {
        bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
-            auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
-            auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
+            auto* pLhs = reinterpret_cast<uint64_t const*>(reinterpret_cast<char const*>(&lhs));
+            auto* pRhs = reinterpret_cast<uint64_t const*>(reinterpret_cast<char const*>(&rhs));
            return *pLhs == *pRhs;
        }
    };

    struct LessThan {
        bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
-            auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
-            auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
+            auto* pLhs = reinterpret_cast<uint64_t const*>(reinterpret_cast<char const*>(&lhs));
+            auto* pRhs = reinterpret_cast<uint64_t const*>(reinterpret_cast<char const*>(&rhs));
            return *pLhs == *pRhs;
        }
    };
@@ -838,7 +838,6 @@ private:
        return SamplerParams::LessThan{}(lhs, rhs);
    }
 };
-static_assert(sizeof(SamplerParams) == 4);

 // The limitation to 64-bits max comes from how we store a SamplerParams in our JNI code
 // see android/.../TextureSampler.cpp
@@ -67,7 +67,7 @@ public:
    Program& operator=(const Program& rhs) = delete;

    Program(Program&& rhs) noexcept;
-    Program& operator=(Program&& rhs) noexcept = delete;
+    Program& operator=(Program&& rhs) noexcept;

    ~Program() noexcept;

@@ -104,9 +104,8 @@ public:
            Sampler const* samplers, size_t count) noexcept;

    struct SpecializationConstant {
-        using Type = std::variant<int32_t, float, bool>;
        uint32_t id;                                // id set in glsl
-        Type value;                                 // value and type
+        std::variant<int32_t, float, bool> value;   // value and type
    };

    Program& specializationConstants(
@@ -73,14 +73,32 @@ public:
        // a cost here (writing and reading the stack at each iteration), in the end it's
        // probably better to pay the cost at just one location.
        intptr_t next;
+        driver.mCurrentExecutingCommand = this;
        mExecute(driver, this, &next);
        return reinterpret_cast<CommandBase*>(reinterpret_cast<intptr_t>(this) + next);
    }

+    inline void captureCallstack() noexcept {
+        auto c = utils::CallStack::unwind(4);
+        size_t i = 0;
+        for (; i < c.getFrameCount() && i < 16; i++) {
+            mCallstack[i] = c[i];
+        }
+        for (; i < 16; i++) {
+            mCallstack[i] = 0;
+        }
+    }
+
+    void printCallstack() noexcept {
+        auto c = utils::CallStack(mCallstack);
+        utils::slog.d << c << utils::io::endl;
+    }
+
    inline ~CommandBase() noexcept = default;

 private:
    Execute mExecute;
+    std::array<intptr_t, 16> mCallstack = {0};
 };

 // ------------------------------------------------------------------------------------------------
@@ -218,6 +236,7 @@ public:
        using Cmd = COMMAND_TYPE(methodName);                                                   \
        void* const p = allocateCommand(CommandBase::align(sizeof(Cmd)));                       \
        new(p) Cmd(mDispatcher.methodName##_, APPLY(std::move, params));                        \
+        ((Cmd*)p)->captureCallstack();                                                          \
        DEBUG_COMMAND_END(methodName, false);                                                   \
    }

@@ -237,6 +256,7 @@ public:
        using Cmd = COMMAND_TYPE(methodName##R);                                                \
        void* const p = allocateCommand(CommandBase::align(sizeof(Cmd)));                       \
        new(p) Cmd(mDispatcher.methodName##_, RetType(result), APPLY(std::move, params));       \
+        ((Cmd*)p)->captureCallstack();                                                          \
        DEBUG_COMMAND_END(methodName, false);                                                   \
        return result;                                                                          \
    }
@@ -53,6 +53,7 @@ template<typename T>
 class ConcreteDispatcher;
 class Dispatcher;
 class CommandStream;
+class CommandBase;

 class Driver {
 public:
@@ -83,6 +84,8 @@ public:
    virtual void debugCommandEnd(CommandStream* cmds,
            bool synchronous, const char* methodName) noexcept = 0;

+    CommandBase* mCurrentExecutingCommand = nullptr;
+
    /*
     * Asynchronous calls here only to provide a type to CommandStream. They must be non-virtual
     * so that calling the concrete implementation won't go through a vtable.
@@ -26,6 +26,17 @@ Program::Program() noexcept { // NOLINT(modernize-use-equals-default)

 Program::Program(Program&& rhs) noexcept = default;

+Program& Program::operator=(Program&& rhs) noexcept {
+    mUniformBlocks.operator=(rhs.mUniformBlocks);
+    mSamplerGroups.operator=(std::move(rhs.mSamplerGroups));
+    mShadersSource.operator=(std::move(rhs.mShadersSource));
+    mName.operator=(std::move(rhs.mName));
+    mLogger.operator=(std::move(rhs.mLogger));
+    mSpecializationConstants.operator=(std::move(rhs.mSpecializationConstants));
+    mBindingUniformInfo.operator=(std::move(rhs.mBindingUniformInfo));
+    return *this;
+}
+
 Program::~Program() noexcept = default;

 Program& Program::priorityQueue(CompilerPriorityQueue priorityQueue) noexcept {
@@ -345,7 +345,7 @@ void MetalDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
        auto colorTexture = handle_cast<MetalTexture>(buffer.handle);
        ASSERT_PRECONDITION(colorTexture->getMtlTextureForWrite(),
                "Color texture passed to render target has no texture allocation");
-        colorTexture->extendLodRangeTo(buffer.level);
+        colorTexture->updateLodRange(buffer.level);
        colorAttachments[i] = { colorTexture, color[i].level, color[i].layer };
    }

@@ -356,7 +356,7 @@ void MetalDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
        auto depthTexture = handle_cast<MetalTexture>(depth.handle);
        ASSERT_PRECONDITION(depthTexture->getMtlTextureForWrite(),
                "Depth texture passed to render target has no texture allocation.");
-        depthTexture->extendLodRangeTo(depth.level);
+        depthTexture->updateLodRange(depth.level);
        depthAttachment = { depthTexture, depth.level, depth.layer };
    }

@@ -367,7 +367,7 @@ void MetalDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
        auto stencilTexture = handle_cast<MetalTexture>(stencil.handle);
        ASSERT_PRECONDITION(stencilTexture->getMtlTextureForWrite(),
                "Stencil texture passed to render target has no texture allocation.");
-        stencilTexture->extendLodRangeTo(stencil.level);
+        stencilTexture->updateLodRange(stencil.level);
        stencilAttachment = { stencilTexture, stencil.level, stencil.layer };
    }

@@ -789,8 +789,6 @@ void MetalDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t ind
 }

 void MetalDriver::setMinMaxLevels(Handle<HwTexture> th, uint32_t minLevel, uint32_t maxLevel) {
-    auto tex = handle_cast<MetalTexture>(th);
-    tex->setLodRange(minLevel, maxLevel);
 }

 void MetalDriver::update3DImage(Handle<HwTexture> th, uint32_t level,
@@ -902,24 +900,29 @@ void MetalDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh, BufferDescripto
    // 2. LOD-clamped textures
    //
    // Both of these cases prevent us from knowing the final id<MTLTexture> that will be bound into
-    // the argument buffer representing the sampler group. So, we wait until draw call time to bind
-    // textures (done in finalizeSamplerGroup).
+    // the argument buffer representing the sampler group. So, we bind what we can now and wait
+    // until draw call time to bind any special cases (done in finalizeSamplerGroup).
    // The good news is that once a render pass has started, the texture bindings won't change.
    // A SamplerGroup is "finalized" when all of its textures have been set and is ready for use in
    // a draw call.
-    // finalizeSamplerGroup has one additional responsibility: to call useResources for all the
-    // textures, which is required by Metal.
+    // Even if we do know all the final textures at this point, we still wait until draw call time
+    // to call finalizeSamplerGroup, which has one additional responsibility: to call useResources
+    // for all the textures, which is required by Metal.
    for (size_t s = 0; s < data.size / sizeof(SamplerDescriptor); s++) {
        if (!samplers[s].t) {
-            // Assign a default sampler to empty slots.
+            // Assign a default texture / sampler to empty slots.
            // Metal requires all samplers referenced in shaders to be bound.
-            // An empty texture will be assigned inside finalizeSamplerGroup.
+            id<MTLTexture> empty = getOrCreateEmptyTexture(mContext);
+            sb->setFinalizedTexture(s, empty);
+
            id<MTLSamplerState> sampler = mContext->samplerStateCache.getOrCreateState({});
            sb->setFinalizedSampler(s, sampler);
+
            continue;
        }

-        // Bind the sampler state. We always know the full sampler state at updateSamplerGroup time.
+        // First, bind the sampler state. We always know the full sampler state at
+        // updateSamplerGroup time.
        SamplerState samplerState {
                .samplerParams = samplers[s].s,
        };
@@ -927,6 +930,27 @@ void MetalDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh, BufferDescripto
        sb->setFinalizedSampler(s, sampler);

        sb->setTextureHandle(s, samplers[s].t);
+
+        auto* t = handle_cast<MetalTexture>(samplers[s].t);
+        assert_invariant(t);
+
+        // If this texture is an external texture, we defer binding the texture until draw call time
+        // (in finalizeSamplerGroup).
+        if (t->target == SamplerType::SAMPLER_EXTERNAL) {
+            continue;
+        }
+
+        if (!t->allLodsValid()) {
+            // The texture doesn't have all of its LODs loaded, and this could change by the time we
+            // issue a draw call with this sampler group. So, we defer binding the texture until
+            // draw call time (in finalizeSamplerGroup).
+            continue;
+        }
+
+        // If we get here, we know we have a valid MTLTexture that's guaranteed not to change.
+        id<MTLTexture> mtlTexture = t->getMtlTextureForRead();
+        assert_invariant(mtlTexture);
+        sb->setFinalizedTexture(s, mtlTexture);
    }

    scheduleDestroy(std::move(data));
@@ -1352,8 +1376,8 @@ void MetalDriver::blit(TargetBufferFlags buffers,
 }

 void MetalDriver::finalizeSamplerGroup(MetalSamplerGroup* samplerGroup) {
-    // All the id<MTLSamplerState> objects have already been bound to the argument buffer.
-    // Here we bind all the textures.
+    // All of the id<MTLSamplerState> objects have already been bound to the argument buffer.
+    // Here we bind any textures that were unable to be bound in updateSamplerGroup.

    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(mContext);

@@ -1375,38 +1399,28 @@ void MetalDriver::finalizeSamplerGroup(MetalSamplerGroup* samplerGroup) {
    }
 #endif

-    utils::FixedCapacityVector<id<MTLTexture>> newTextures(samplerGroup->size, nil);
    for (size_t binding = 0; binding < samplerGroup->size; binding++) {
-        auto [th, _] = samplerGroup->getFinalizedTexture(binding);
+        auto [th, t] = samplerGroup->getFinalizedTexture(binding);

-        if (!th) {
-            // Bind an empty texture.
-            newTextures[binding] = getOrCreateEmptyTexture(mContext);
+        // This may be an external texture, in which case we can't cache the id<MTLTexture>, we
+        // need to refetch it in case the external image has changed.
+        bool isExternalImage = false;
+        if (th) {
+            auto* texture = handle_cast<MetalTexture>(th);
+            isExternalImage = texture->target == SamplerType::SAMPLER_EXTERNAL;
+        }
+
+        // If t is non-nil, then we've already finalized this texture.
+        if (t && !isExternalImage) {
            continue;
        }

+        // It's possible that some texture handles are null, but we should have already handled
+        // these inside updateSamplerGroup by binding an "empty" texture.
        assert_invariant(th);
        auto* texture = handle_cast<MetalTexture>(th);

-        // External images
-        if (texture->target == SamplerType::SAMPLER_EXTERNAL) {
-            if (texture->externalImage.isValid()) {
-                id<MTLTexture> mtlTexture = texture->externalImage.getMetalTextureForDraw();
-                assert_invariant(mtlTexture);
-                newTextures[binding] = mtlTexture;
-            } else {
-                // Bind an empty texture.
-                newTextures[binding] = getOrCreateEmptyTexture(mContext);
-            }
-            continue;
-        }
-
-        newTextures[binding] = texture->getMtlTextureForRead();
-    }
-
-    if (!std::equal(newTextures.begin(), newTextures.end(), samplerGroup->textures.begin())) {
-        // One or more of the id<MTLTexture>s has changed.
-        // First, determine if this SamplerGroup needs mutation.
+        // Determine if this SamplerGroup needs mutation.
        // We can't just simply mutate the SamplerGroup, since it could currently be in use by the
        // GPU from a prior render pass.
        // If the SamplerGroup does need mutation, then there's two cases:
@@ -1414,18 +1428,30 @@ void MetalDriver::finalizeSamplerGroup(MetalSamplerGroup* samplerGroup) {
        //    draw call). We're free to mutate it.
        // 2. The SamplerGroup is finalized. We must call mutate(), which will create a new argument
        //    buffer that we can then mutate freely.
+        // TODO: don't just always call mutate, check to see if the texture is actually different.

        if (samplerGroup->isFinalized()) {
            samplerGroup->mutate(cmdBuffer);
        }

-        for (size_t binding = 0; binding < samplerGroup->size; binding++) {
-            samplerGroup->setFinalizedTexture(binding, newTextures[binding]);
+        // External images
+        if (texture->target == SamplerType::SAMPLER_EXTERNAL) {
+            if (texture->externalImage.isValid()) {
+                id<MTLTexture> mtlTexture = texture->externalImage.getMetalTextureForDraw();
+                assert_invariant(mtlTexture);
+                samplerGroup->setFinalizedTexture(binding, mtlTexture);
+            } else {
+                // Bind an empty texture.
+                samplerGroup->setFinalizedTexture(binding, getOrCreateEmptyTexture(mContext));
+            }
+            continue;
        }

-        samplerGroup->finalize();
+        samplerGroup->setFinalizedTexture(binding, texture->getMtlTextureForRead());
    }

+    samplerGroup->finalize();
+
    // At this point, all the id<MTLTextures> should be set to valid textures. Some of them will be
    // the "empty" texture. Per Apple documentation, the useResource method must be called once per
    // render pass.
@@ -217,14 +217,21 @@ public:
    void generateMipmaps() noexcept;

    // A texture starts out with none of its mip levels (also referred to as LODs) available for
-    // reading. 4 actions update the range of LODs available:
+    // reading. 3 actions update the range of LODs available:
    // - calling loadImage
    // - calling generateMipmaps
    // - using the texture as a render target attachment
-    // - calling setMinMaxLevels
+    // The range of available mips can only increase, never decrease.
    // A texture's available mips are consistent throughout a render pass.
-    void setLodRange(uint32_t minLevel, uint32_t maxLevel);
-    void extendLodRangeTo(uint32_t level);
+    void updateLodRange(uint32_t level);
+    void updateLodRange(uint32_t minLevel, uint32_t maxLevel);
+
+    // Returns true if the texture has all of its mip levels accessible for reading.
+    // For any MetalTexture, once this is true, will always return true.
+    // The value returned will remain consistent for an entire render pass.
+    bool allLodsValid() const {
+        return minLod == 0 && maxLod == levels - 1;
+    }

    static MTLPixelFormat decidePixelFormat(MetalContext* context, TextureFormat format);

@@ -513,7 +513,7 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
        externalImage(context) {
    texture = metalTexture;
-    setLodRange(0, levels - 1);
+    updateLodRange(0, levels - 1);
 }

 MetalTexture::~MetalTexture() {
@@ -658,14 +658,14 @@ void MetalTexture::loadImage(uint32_t level, MTLRegion region, PixelBufferDescri
        }
    }

-    extendLodRangeTo(level);
+    updateLodRange(level);
 }

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

 void MetalTexture::loadSlice(uint32_t level, MTLRegion region, uint32_t byteOffset, uint32_t slice,
@@ -788,18 +788,18 @@ void MetalTexture::loadWithBlit(uint32_t level, uint32_t slice, MTLRegion region
    context.blitter->blit(getPendingCommandBuffer(&context), args, "Texture upload blit");
 }

-void MetalTexture::extendLodRangeTo(uint32_t level) {
+void MetalTexture::updateLodRange(uint32_t level) {
    assert_invariant(!isInRenderPass(&context));
    minLod = std::min(minLod, level);
    maxLod = std::max(maxLod, level);
    lodTextureView = nil;
 }

-void MetalTexture::setLodRange(uint32_t min, uint32_t max) {
+void MetalTexture::updateLodRange(uint32_t min, uint32_t max) {
    assert_invariant(!isInRenderPass(&context));
    assert_invariant(min <= max);
-    minLod = min;
-    maxLod = max;
+    minLod = std::min(minLod, min);
+    maxLod = std::max(maxLod, max);
    lodTextureView = nil;
 }

@@ -103,11 +103,6 @@ struct ShaderCompilerService::OpenGLProgramToken : ProgramToken {
        return programData;
    }

-    void wait() const noexcept {
-        std::unique_lock l(lock);
-        cond.wait(l, [this](){ return signaled; });
-    }
-
    // Checks if the programBinary is ready.
    // This is similar to std::future::wait_for(0s)
    bool isReady() const noexcept {
@@ -341,23 +336,7 @@ GLuint ShaderCompilerService::getProgram(ShaderCompilerService::program_token_t&

    token->canceled = true;

-    bool canceled = token->compiler.cancelTickOp(token);
-
-    if (token->compiler.mShaderCompilerThreadCount) {
-        auto job = token->compiler.mCompilerThreadPool.dequeue(token);
-        if (!job) {
-            // The job is being executed right now. We need to wait for it to finish to avoid a
-            // race.
-            token->wait();
-        } else {
-            // The job has not been executed, but we still need to inform the callback manager in
-            // order for future callbacks to be successfully called.
-            token->compiler.mCallbackManager.put(token->handle);
-        }
-    } else if (canceled) {
-        // Since the tick op was canceled, we need to .put the token here.
-        token->compiler.mCallbackManager.put(token->handle);
-    }
+    token->compiler.cancelTickOp(token);

    for (GLuint& shader: token->gl.shaders) {
        if (shader) {
@@ -683,7 +662,7 @@ void ShaderCompilerService::runAtNextTick(CompilerPriorityQueue priority,
    SYSTRACE_VALUE32("ShaderCompilerService Jobs", mRunAtNextTickOps.size());
 }

-bool ShaderCompilerService::cancelTickOp(program_token_t token) noexcept {
+void ShaderCompilerService::cancelTickOp(program_token_t token) noexcept {
    // We do a linear search here, but this is rare, and we know the list is pretty small.
    auto& ops = mRunAtNextTickOps;
    auto pos = std::find_if(ops.begin(), ops.end(), [&](const auto& item) {
@@ -691,11 +670,9 @@ bool ShaderCompilerService::cancelTickOp(program_token_t token) noexcept {
    });
    if (pos != ops.end()) {
        ops.erase(pos);
-        return true;
    }
    SYSTRACE_CONTEXT();
    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
-    return false;
 }

 void ShaderCompilerService::executeTickOps() noexcept {
@@ -141,7 +141,7 @@ private:
    void runAtNextTick(CompilerPriorityQueue priority,
            const program_token_t& token, Job job) noexcept;
    void executeTickOps() noexcept;
-    bool cancelTickOp(program_token_t token) noexcept;
+    void cancelTickOp(program_token_t token) noexcept;
    // order of insertion is important

    using ContainerType = std::tuple<CompilerPriorityQueue, program_token_t, Job>;
@@ -28,6 +28,8 @@

 #include "generated/vkshaders/vkshaders.h"

+#define FILAMENT_VULKAN_CHECK_BLIT_FORMAT 0
+
 using namespace bluevk;
 using namespace utils;

@@ -67,7 +69,7 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
            .layerCount = 1,
    };

-    if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
+    if constexpr (FILAMENT_VULKAN_VERBOSE) {
        utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                      << "layout=" << src.getLayout()
                      << " to=" << dst.texture->getVkImage() << ",level=" << (int) dst.level
@@ -130,7 +132,7 @@ void VulkanBlitter::blitColor(BlitArgs args) {
    const VulkanAttachment dst = args.dstTarget->getColor(0);
    const VkImageAspectFlags aspect = VK_IMAGE_ASPECT_COLOR_BIT;

-#if FVK_ENABLED(FVK_DEBUG_BLIT_FORMAT)
+#if FILAMENT_VULKAN_CHECK_BLIT_FORMAT
    VkPhysicalDevice const gpu = mPhysicalDevice;
    VkFormatProperties info;
    vkGetPhysicalDeviceFormatProperties(gpu, src.getFormat(), &info);
@@ -145,7 +147,7 @@ void VulkanBlitter::blitColor(BlitArgs args) {
    }
 #endif
    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
+    VkCommandBuffer const cmdbuffer = commands.cmdbuffer;
    commands.acquire(src.texture);
    commands.acquire(dst.texture);

@@ -158,7 +160,7 @@ void VulkanBlitter::blitDepth(BlitArgs args) {
    const VulkanAttachment dst = args.dstTarget->getDepth();
    const VkImageAspectFlags aspect = VK_IMAGE_ASPECT_DEPTH_BIT;

-#if FVK_ENABLED(FVK_DEBUG_BLIT_FORMAT)
+#if FILAMENT_VULKAN_CHECK_BLIT_FORMAT
    VkPhysicalDevice const gpu = mPhysicalDevice;
    VkFormatProperties info;
    vkGetPhysicalDeviceFormatProperties(gpu, src.getFormat(), &info);
@@ -182,7 +184,7 @@ void VulkanBlitter::blitDepth(BlitArgs args) {
    }

    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
+    VkCommandBuffer const cmdbuffer = commands.cmdbuffer;
    commands.acquire(src.texture);
    commands.acquire(dst.texture);
    blitFast(cmdbuffer, aspect, args.filter, args.srcTarget->getExtent(), src, dst, args.srcRectPair,
@@ -195,11 +197,13 @@ void VulkanBlitter::terminate() noexcept {
        mDepthResolveProgram = nullptr;

        if (mTriangleBuffer) {
+            mTriangleBuffer->terminate();
            delete mTriangleBuffer;
            mTriangleBuffer = nullptr;
        }

        if (mParamsBuffer) {
+            mParamsBuffer->terminate();
            delete mParamsBuffer;
            mParamsBuffer = nullptr;
        }
@@ -239,7 +243,7 @@ void VulkanBlitter::lazyInit() noexcept {
    mDepthResolveProgram->samplerGroupInfo[0].samplers.reserve(1);
    mDepthResolveProgram->samplerGroupInfo[0].samplers.resize(1);

-    if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
+    if constexpr (FILAMENT_VULKAN_VERBOSE) {
        utils::slog.d << "Created Shader Module for VulkanBlitter "
                    << "shaders = (" << vertexShader << ", " << fragmentShader << ")"
                    << utils::io::endl;
@@ -253,7 +257,7 @@ void VulkanBlitter::lazyInit() noexcept {
    };

    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
+    VkCommandBuffer const cmdbuffer = commands.cmdbuffer;

    mTriangleBuffer = new VulkanBuffer(mAllocator, mStagePool, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
            sizeof(kTriangleVertices));
@@ -274,7 +278,7 @@ void VulkanBlitter::blitSlowDepth(VkFilter filter, const VkExtent2D srcExtent, V
    lazyInit();

    VulkanCommandBuffer* commands = &mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands->buffer();
+    VkCommandBuffer const cmdbuffer = commands->cmdbuffer;
    commands->acquire(src.texture);
    commands->acquire(dst.texture);

@@ -45,7 +45,14 @@ VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
 }

 VulkanBuffer::~VulkanBuffer() {
+    assert_invariant(mGpuMemory == VK_NULL_HANDLE);
+    assert_invariant(mGpuBuffer == VK_NULL_HANDLE);
+}
+
+void VulkanBuffer::terminate() {
    vmaDestroyBuffer(mAllocator, mGpuBuffer, mGpuMemory);
+    mGpuMemory = VK_NULL_HANDLE;
+    mGpuBuffer = VK_NULL_HANDLE;
 }

 void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
@@ -28,6 +28,7 @@ public:
    VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool, VkBufferUsageFlags usage,
            uint32_t numBytes);
    ~VulkanBuffer();
+    void terminate();
    void loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
            uint32_t numBytes) const;
    VkBuffer getGpuBuffer() const {
@@ -33,53 +33,43 @@ using namespace utils;

 namespace filament::backend {

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

-VulkanCmdFence::VulkanCmdFence(VkFence ifence)
-    : fence(ifence) {
+VulkanCmdFence::VulkanCmdFence(VkDevice device, bool signaled) : device(device) {
+    VkFenceCreateInfo fenceCreateInfo { .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO };
+    if (signaled) {
+        fenceCreateInfo.flags = VK_FENCE_CREATE_SIGNALED_BIT;
+    }
+    vkCreateFence(device, &fenceCreateInfo, VKALLOC, &fence);
+
    // 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);
 }

-VulkanCommandBuffer::VulkanCommandBuffer(VulkanResourceAllocator* allocator, VkDevice device,
-        VkCommandPool pool)
-    : mResourceManager(allocator) {
-    // 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);
+VulkanCmdFence::~VulkanCmdFence() {
+    vkDestroyFence(device, fence, VKALLOC);
 }

 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,
+        .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);
-
-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+#if FILAMENT_VULKAN_VERBOSE
    mTimestamps.push_back(start.time_since_epoch().count() > 0.0
                                  ? start
                                  : std::chrono::high_resolution_clock::now());
@@ -90,7 +80,7 @@ std::pair<std::string, Timestamp> VulkanGroupMarkers::pop() noexcept {
    auto const marker = mMarkers.back();
    mMarkers.pop_back();

-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+#if FILAMENT_VULKAN_VERBOSE
    auto const timestamp = mTimestamps.back();
    mTimestamps.pop_back();
    return std::make_pair(marker, timestamp);
@@ -103,7 +93,7 @@ std::pair<std::string, Timestamp> VulkanGroupMarkers::pop_bottom() noexcept {
    auto const marker = mMarkers.front();
    mMarkers.pop_front();

-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+#if FILAMENT_VULKAN_VERBOSE
    auto const timestamp = mTimestamps.front();
    mTimestamps.pop_front();
    return std::make_pair(marker, timestamp);
@@ -115,8 +105,8 @@ std::pair<std::string, Timestamp> VulkanGroupMarkers::pop_bottom() noexcept {
 std::pair<std::string, Timestamp> VulkanGroupMarkers::top() const {
    assert_invariant(!empty());
    auto const marker = mMarkers.back();
-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
-    auto const topTimestamp = mTimestamps.front();
+#if FILAMENT_VULKAN_VERBOSE
+    auto const topTimestamp = mTimestamps.top();
    return std::make_pair(marker, topTimestamp);
 #else
    return std::make_pair(marker, Timestamp{});
@@ -127,8 +117,6 @@ bool VulkanGroupMarkers::empty() const noexcept {
    return mMarkers.empty();
 }

-#endif // FVK_DEBUG_GROUP_MARKERS
-
 VulkanCommands::VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
        VulkanContext* context, VulkanResourceAllocator* allocator)
    : mDevice(device),
@@ -141,30 +129,18 @@ VulkanCommands::VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFam
        vkCreateSemaphore(mDevice, &sci, nullptr, &semaphore);
    }

-    VkFenceCreateInfo fenceCreateInfo{.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO};
-    for (auto& fence: mFences) {
-        vkCreateFence(device, &fenceCreateInfo, VKALLOC, &fence);
-    }
-
    for (size_t i = 0; i < CAPACITY; ++i) {
-        mStorage[i] = std::make_unique<VulkanCommandBuffer>(allocator, mDevice, mPool);
+        mStorage[i] = std::make_unique<VulkanCommandBuffer>(allocator);
    }
-
-#if !FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    (void) mContext;
-#endif
 }

 VulkanCommands::~VulkanCommands() {
    wait();
    gc();
    vkDestroyCommandPool(mDevice, mPool, VKALLOC);
-    for (VkSemaphore sema: mSubmissionSignals) {
+    for (VkSemaphore sema : mSubmissionSignals) {
        vkDestroySemaphore(mDevice, sema, VKALLOC);
    }
-    for (VkFence fence: mFences) {
-        vkDestroyFence(mDevice, fence, VKALLOC);
-    }
 }

 VulkanCommandBuffer& VulkanCommands::get() {
@@ -175,11 +151,11 @@ VulkanCommandBuffer& VulkanCommands::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)
-        slog.i << "VulkanCommands has stalled. "
-               << "If this occurs frequently, consider increasing VK_MAX_COMMAND_BUFFERS."
-               << io::endl;
+    while (mAvailableCount == 0) {
+#if VK_REPORT_STALLS
+        slog.i  << "VulkanCommands has stalled. "
+                << "If this occurs frequently, consider increasing VK_MAX_COMMAND_BUFFERS."
+                << io::endl;
 #endif
        wait();
        gc();
@@ -189,7 +165,7 @@ VulkanCommandBuffer& VulkanCommands::get() {
    // Find an available slot.
    for (size_t i = 0; i < CAPACITY; ++i) {
        auto wrapper = mStorage[i].get();
-        if (wrapper->buffer() == VK_NULL_HANDLE) {
+        if (wrapper->cmdbuffer == VK_NULL_HANDLE) {
            mCurrentCommandBufferIndex = static_cast<int8_t>(i);
            currentbuf = wrapper;
            break;
@@ -197,33 +173,41 @@ VulkanCommandBuffer& VulkanCommands::get() {
    }

    assert_invariant(currentbuf);
-    mAvailableBufferCount--;
+    --mAvailableCount;
+
+    // Create the low-level command buffer.
+    const VkCommandBufferAllocateInfo allocateInfo {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
+        .commandPool = mPool,
+        .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
+        .commandBufferCount = 1
+    };
+    vkAllocateCommandBuffers(mDevice, &allocateInfo, &currentbuf->cmdbuffer);

    // 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]);
+    currentbuf->fence = std::make_shared<VulkanCmdFence>(mDevice);

    // 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,
+    const VkCommandBufferBeginInfo binfo {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
    };
-    vkBeginCommandBuffer(currentbuf->buffer(), &binfo);
+    vkBeginCommandBuffer(currentbuf->cmdbuffer, &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;
 }

@@ -233,25 +217,11 @@ bool VulkanCommands::flush() {
        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();
-    }
-#endif
-
    int8_t const index = mCurrentCommandBufferIndex;
    VulkanCommandBuffer const* currentbuf = mStorage[index].get();
    VkSemaphore const renderingFinished = mSubmissionSignals[index];

-    vkEndCommandBuffer(currentbuf->buffer());
+    vkEndCommandBuffer(currentbuf->cmdbuffer);

    // 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
@@ -260,26 +230,24 @@ bool VulkanCommands::flush() {
    // 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,
+        VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+        VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
    };

    VkSemaphore signals[2] = {
-            VK_NULL_HANDLE,
-            VK_NULL_HANDLE,
+        VK_NULL_HANDLE,
+        VK_NULL_HANDLE,
    };

-    VkCommandBuffer const cmdbuffer = currentbuf->buffer();
-
-    VkSubmitInfo submitInfo{
-            .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
-            .waitSemaphoreCount = 0,
-            .pWaitSemaphores = signals,
-            .pWaitDstStageMask = waitDestStageMasks,
-            .commandBufferCount = 1,
-            .pCommandBuffers = &cmdbuffer,
-            .signalSemaphoreCount = 1u,
-            .pSignalSemaphores = &renderingFinished,
+    VkSubmitInfo submitInfo {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+        .waitSemaphoreCount = 0,
+        .pWaitSemaphores = signals,
+        .pWaitDstStageMask = waitDestStageMasks,
+        .commandBufferCount = 1,
+        .pCommandBuffers = &currentbuf->cmdbuffer,
+        .signalSemaphoreCount = 1u,
+        .pSignalSemaphores = &renderingFinished,
    };

    if (mSubmissionSignal) {
@@ -295,14 +263,24 @@ bool VulkanCommands::flush() {
        submitInfo.pWaitSemaphores = VK_NULL_HANDLE;
    }

-#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-    slog.i << "Submitting cmdbuffer=" << cmdbuffer
-           << " wait=(" << signals[0] << ", " << signals[1] << ") "
-           << " signal=" << renderingFinished
-           << " fence=" << currentbuf->fence->fence
-           << utils::io::endl;
+#if FILAMENT_VULKAN_VERBOSE
+    slog.i << "Submitting cmdbuffer=" << currentbuf->cmdbuffer
+        << " wait=(" << signals[0] << ", " << signals[1] << ") "
+        << " signal=" << renderingFinished
+        << io::endl;
 #endif

+    // Before actually submitting, we need to pop any leftover 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();
+    }
+
    auto& cmdfence = currentbuf->fence;
    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
    cmdfence->status.store(VK_NOT_READY);
@@ -310,9 +288,9 @@ bool VulkanCommands::flush() {
    cmdfence->condition.notify_all();
    lock.unlock();

-#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
+#if FILAMENT_VULKAN_VERBOSE
    if (result != VK_SUCCESS) {
-        utils::slog.d << "Failed command buffer submission result: " << result << utils::io::endl;
+        utils::slog.d <<"Failed command buffer submission result: " << result << utils::io::endl;
    }
 #endif
    assert_invariant(result == VK_SUCCESS);
@@ -326,8 +304,8 @@ bool VulkanCommands::flush() {
 VkSemaphore VulkanCommands::acquireFinishedSignal() {
    VkSemaphore semaphore = mSubmissionSignal;
    mSubmissionSignal = VK_NULL_HANDLE;
-#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
-    slog.i << "Acquiring " << semaphore << " (e.g. for vkQueuePresentKHR)" << io::endl;
+#if FILAMENT_VULKAN_VERBOSE
+     slog.i << "Acquiring " << semaphore << " (e.g. for vkQueuePresentKHR)" << io::endl;
 #endif
    return semaphore;
 }
@@ -335,7 +313,7 @@ VkSemaphore VulkanCommands::acquireFinishedSignal() {
 void VulkanCommands::injectDependency(VkSemaphore next) {
    assert_invariant(mInjectedSignal == VK_NULL_HANDLE);
    mInjectedSignal = next;
-#if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
+#if FILAMENT_VULKAN_VERBOSE
    slog.i << "Injecting " << next << " (e.g. due to vkAcquireNextImageKHR)" << io::endl;
 #endif
 }
@@ -345,49 +323,44 @@ void VulkanCommands::wait() {
    size_t count = 0;
    for (size_t i = 0; i < CAPACITY; i++) {
        auto wrapper = mStorage[i].get();
-        if (wrapper->buffer() != VK_NULL_HANDLE
+        if (wrapper->cmdbuffer != VK_NULL_HANDLE
                && mCurrentCommandBufferIndex != static_cast<int8_t>(i)) {
            fences[count++] = wrapper->fence->fence;
        }
    }
    if (count > 0) {
        vkWaitForFences(mDevice, count, fences, VK_TRUE, UINT64_MAX);
-        updateFences();
    }
 }

 void VulkanCommands::gc() {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("commands::gc");
-
-    VkFence fences[CAPACITY];
+    VkCommandBuffer buffers[CAPACITY];
    size_t count = 0;
-
    for (size_t i = 0; i < CAPACITY; i++) {
        auto wrapper = mStorage[i].get();
-        if (wrapper->buffer() == VK_NULL_HANDLE) {
+        if (wrapper->cmdbuffer == VK_NULL_HANDLE) {
            continue;
        }
-        VkResult const result = vkGetFenceStatus(mDevice, wrapper->fence->fence);
+        VkResult const result = vkWaitForFences(mDevice, 1, &wrapper->fence->fence, VK_TRUE, 0);
        if (result != VK_SUCCESS) {
            continue;
        }
-        fences[count++] = wrapper->fence->fence;
+        buffers[count++] = wrapper->cmdbuffer;
+        wrapper->cmdbuffer = VK_NULL_HANDLE;
        wrapper->fence->status.store(VK_SUCCESS);
-        wrapper->reset();
-        mAvailableBufferCount++;
+        wrapper->fence.reset();
+        wrapper->clearResources();
+        ++mAvailableCount;
    }
-
    if (count > 0) {
-        vkResetFences(mDevice, count, fences);
+        vkFreeCommandBuffers(mDevice, mPool, count, buffers);
    }
-    FVK_SYSTRACE_END();
 }

 void VulkanCommands::updateFences() {
    for (size_t i = 0; i < CAPACITY; i++) {
        auto wrapper = mStorage[i].get();
-        if (wrapper->buffer() != VK_NULL_HANDLE) {
+        if (wrapper->cmdbuffer != VK_NULL_HANDLE) {
            VulkanCmdFence* fence = wrapper->fence.get();
            if (fence) {
                VkResult status = vkGetFenceStatus(mDevice, fence->fence);
@@ -398,10 +371,8 @@ void VulkanCommands::updateFences() {
    }
 }

-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-
 void VulkanCommands::pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp) {
-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+#if FILAMENT_VULKAN_VERBOSE
    // 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) {
@@ -410,7 +381,7 @@ void VulkanCommands::pushGroupMarker(char const* str, VulkanGroupMarkers::Timest
 #endif

    // TODO: Add group marker color to the Driver API
-    VkCommandBuffer const cmdbuffer = get().buffer();
+    const VkCommandBuffer cmdbuffer = get().cmdbuffer;

    if (!mGroupMarkers) {
        mGroupMarkers = std::make_unique<VulkanGroupMarkers>();
@@ -435,19 +406,20 @@ void VulkanCommands::pushGroupMarker(char const* str, VulkanGroupMarkers::Timest
 }

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

    if (!mGroupMarkers->empty()) {
-        VkCommandBuffer const cmdbuffer = get().buffer();
-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
-        auto const [marker, startTime] = mGroupMarkers->pop();
-        auto const endTime = std::chrono::high_resolution_clock::now();
-        std::chrono::duration<double> diff = endTime - startTime;
-        utils::slog.d << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms\n"
-                      << utils::io::flush;
-#else
-        mGroupMarkers->pop();
-#endif
+        const VkCommandBuffer cmdbuffer = get().cmdbuffer;
+        #if FILAMENT_VULKAN_VERBOSE
+            auto const [marker, startTime] = mGroupMarkers->pop();
+            auto const endTime = std::chrono::high_resolution_clock::now();
+            std::chrono::duration<double> diff = endTime - startTime;
+            utils::slog.d << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms\n"
+                          << utils::io::flush;
+        #else
+            mGroupMarkers->pop();
+        #endif
+
        if (mContext->isDebugUtilsSupported()) {
            vkCmdEndDebugUtilsLabelEXT(cmdbuffer);
        } else if (mContext->isDebugMarkersSupported()) {
@@ -462,7 +434,7 @@ void VulkanCommands::popGroupMarker() {
 }

 void VulkanCommands::insertEventMarker(char const* string, uint32_t len) {
-    VkCommandBuffer const cmdbuffer = get().buffer();
+    VkCommandBuffer const cmdbuffer = get().cmdbuffer;
    if (mContext->isDebugUtilsSupported()) {
        VkDebugUtilsLabelEXT labelInfo = {
                .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
@@ -472,9 +444,9 @@ void VulkanCommands::insertEventMarker(char const* string, uint32_t len) {
        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},
+            .sType = VK_STRUCTURE_TYPE_DEBUG_MARKER_MARKER_INFO_EXT,
+            .pMarkerName = string,
+            .color = {0.0f, 1.0f, 0.0f, 1.0f},
        };
        vkCmdDebugMarkerInsertEXT(cmdbuffer, &markerInfo);
    }
@@ -486,9 +458,8 @@ std::string VulkanCommands::getTopGroupMarker() const {
    }
    return std::get<0>(mGroupMarkers->top());
 }
-#endif // FVK_DEBUG_GROUP_MARKERS

-} // namespace filament::backend
+}// namespace filament::backend

 #if defined(_MSC_VER)
 #pragma warning( pop )
@@ -39,7 +39,6 @@ namespace filament::backend {

 struct VulkanContext;

-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
 class VulkanGroupMarkers {
 public:
    using Timestamp = std::chrono::time_point<std::chrono::high_resolution_clock>;
@@ -52,17 +51,16 @@ public:

 private:
    std::list<std::string> mMarkers;
-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
+#if FILAMENT_VULKAN_VERBOSE
    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 {
-    VulkanCmdFence(VkFence ifence);
-    ~VulkanCmdFence() = default;
+    VulkanCmdFence(VkDevice device, bool signaled = false);
+    ~VulkanCmdFence();
+    const VkDevice device;
    VkFence fence;
    utils::Condition condition;
    utils::Mutex mutex;
@@ -73,10 +71,13 @@ struct VulkanCmdFence {
 // 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(VulkanResourceAllocator* allocator, VkDevice device, VkCommandPool pool);
+    VulkanCommandBuffer(VulkanResourceAllocator* allocator)
+        : mResourceManager(allocator) {}

    VulkanCommandBuffer(VulkanCommandBuffer const&) = delete;
    VulkanCommandBuffer& operator=(VulkanCommandBuffer const&) = delete;
+    VkCommandBuffer cmdbuffer = VK_NULL_HANDLE;
+    std::shared_ptr<VulkanCmdFence> fence;

    inline void acquire(VulkanResource* resource) {
        mResourceManager.acquire(resource);
@@ -86,23 +87,12 @@ struct VulkanCommandBuffer {
        mResourceManager.acquire(srcResources);
    }

-    inline void reset() {
-        fence.reset();
+    inline void clearResources() {
        mResourceManager.clear();
    }

-    inline VkCommandBuffer buffer() const {
-        if (fence) {
-            return mBuffer;
-        }
-        return VK_NULL_HANDLE;
-    }
-
-    std::shared_ptr<VulkanCmdFence> fence;
-
 private:
    VulkanAcquireOnlyResourceManager mResourceManager;
-    VkCommandBuffer mBuffer;
 };

 // Allows classes to be notified after a new command buffer has been activated.
@@ -173,7 +163,6 @@ class VulkanCommands {
        // 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();
@@ -181,10 +170,9 @@ class VulkanCommands {
        void insertEventMarker(char const* string, uint32_t len);

        std::string getTopGroupMarker() const;
-#endif

    private:
-        static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
+        static constexpr int CAPACITY = VK_MAX_COMMAND_BUFFERS;
        VkDevice const mDevice;
        VkQueue const mQueue;
        VkCommandPool const mPool;
@@ -196,15 +184,12 @@ class VulkanCommands {
        VkSemaphore mSubmissionSignal = {};
        VkSemaphore mInjectedSignal = {};
        utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>> mStorage;
-        VkFence mFences[CAPACITY] = {};
        VkSemaphore mSubmissionSignals[CAPACITY] = {};
-        uint8_t mAvailableBufferCount = CAPACITY;
+        size_t mAvailableCount = CAPACITY;
        CommandBufferObserver* mObserver = nullptr;

-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
        std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
        std::unique_ptr<VulkanGroupMarkers> mCarriedOverMarkers;
-#endif
 };

 } // namespace filament::backend
@@ -19,6 +19,13 @@

 #include <stdint.h>

+#define FILAMENT_VULKAN_VERBOSE 0
+#define FILAMENT_VULKAN_DUMP_API 0
+
+#ifndef FILAMENT_VULKAN_HANDLE_ARENA_SIZE_IN_MB
+#define FILAMENT_VULKAN_HANDLE_ARENA_SIZE_IN_MB 8
+#endif
+
 // In debug builds, we enable validation layers and set up a debug callback.
 //
 // To enable validation layers in Android, also be sure to set the jniLibs property in the gradle
@@ -36,80 +43,21 @@
 //     ndkVersion "23.1.7779620"
 //
 // Also consider changing the root `gradle.properties` to point to a debug build, although this is
-// not required for validation.
-
-// FVK is short for Filament Vulkan
-#define FVK_DEBUG_SYSTRACE                0x00000001
-#define FVK_DEBUG_GROUP_MARKERS           0x00000002
-#define FVK_DEBUG_TEXTURE                 0x00000004
-#define FVK_DEBUG_LAYOUT_TRANSITION       0x00000008
-#define FVK_DEBUG_COMMAND_BUFFER          0x00000010
-#define FVK_DEBUG_DUMP_API                0x00000020
-#define FVK_DEBUG_VALIDATION              0x00000040
-#define FVK_DEBUG_PRINT_GROUP_MARKERS     0x00000080
-#define FVK_DEBUG_BLIT_FORMAT             0x00000100
-#define FVK_DEBUG_BLITTER                 0x00000200
-#define FVK_DEBUG_FBO_CACHE               0x00000400
-#define FVK_DEBUG_SHADER_MODULE           0x00000800
-#define FVK_DEBUG_READ_PIXELS             0x00001000
-#define FVK_DEBUG_PIPELINE_CACHE          0x00002000
-#define FVK_DEBUG_ALLOCATION              0x00004000
-
-// Usefaul default combinations
-#define FVK_DEBUG_EVERYTHING              0xFFFFFFFF
-#define FVK_DEBUG_PERFORMANCE     \
-    FVK_DEBUG_SYSTRACE |          \
-    FVK_DEBUG_GROUP_MARKERS
-
-#define FVK_DEBUG_CORRECTNESS     \
-    FVK_DEBUG_VALIDATION |        \
-    FVK_DEBUG_SHADER_MODULE |     \
-    FVK_DEBUG_TEXTURE |           \
-    FVK_DEBUG_LAYOUT_TRANSITION
-
-#define FVK_DEBUG_RENDER_PASSES   \
-    FVK_DEBUG_GROUP_MARKERS |     \
-    FVK_DEBUG_PRINT_GROUP_MARKERS
-
-#ifndef NDEBUG
-#define FVK_DEBUG_FLAGS FVK_DEBUG_PERFORMANCE
+// not required for validation if you change the definition of VK_ENABLE_VALIDATION below.
+#if defined(NDEBUG)
+#define VK_ENABLE_VALIDATION 0
 #else
-#define FVK_DEBUG_FLAGS 0
+#define VK_ENABLE_VALIDATION 1
 #endif

-#define FVK_ENABLED(flags) ((FVK_DEBUG_FLAGS) & (flags))
-#define FVK_ENABLED_BOOL(flags) ((bool) FVK_ENABLED(flags))
-
-// Ensure dependencies are met between debug options
-#if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
-static_assert(FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS));
-#endif
-
-// end dependcy checks
-
-#if 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()
-#else
-#define FVK_SYSTRACE_CONTEXT()
-#define FVK_SYSTRACE_START(marker)
-#define FVK_SYSTRACE_END()
-#endif
-
-#ifndef FVK_HANDLE_ARENA_SIZE_IN_MB
-#define FVK_HANDLE_ARENA_SIZE_IN_MB 8
-#endif
+#define VK_REPORT_STALLS 0

 // All vkCreate* functions take an optional allocator. For now we select the default allocator by
 // passing in a null pointer, and we highlight the argument by using the VKALLOC constant.
 constexpr struct VkAllocationCallbacks* VKALLOC = nullptr;

-constexpr static const int FVK_REQUIRED_VERSION_MAJOR = 1;
-constexpr static const int FVK_REQUIRED_VERSION_MINOR = 1;
+constexpr static const int VK_REQUIRED_VERSION_MAJOR = 1;
+constexpr static const int VK_REQUIRED_VERSION_MINOR = 1;

 // Maximum number of VkCommandBuffer handles managed simultaneously by VulkanCommands.
 //
@@ -117,18 +65,18 @@ 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.
-constexpr static const int FVK_MAX_COMMAND_BUFFERS = 10;
+constexpr static const int VK_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 = 10;
+constexpr static const int VK_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
+// instead it simply waits for at least VK_MAX_COMMAND_BUFFERS submissions to occur before
 // destroying any unused pipeline object.
-static_assert(FVK_MAX_PIPELINE_AGE >= FVK_MAX_COMMAND_BUFFERS);
+static_assert(VK_MAX_PIPELINE_AGE >= VK_MAX_COMMAND_BUFFERS);

 #endif
@@ -117,9 +117,8 @@ void VulkanTimestamps::beginQuery(VulkanCommandBuffer const* commands,
        VulkanTimerQuery* query) {
    uint32_t const index = query->getStartingQueryIndex();

-    auto const cmdbuffer = commands->buffer();
-    vkCmdResetQueryPool(cmdbuffer, mPool, index, 2);
-    vkCmdWriteTimestamp(cmdbuffer, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, mPool, index);
+    vkCmdResetQueryPool(commands->cmdbuffer, mPool, index, 2);
+    vkCmdWriteTimestamp(commands->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->fence);
@@ -128,7 +127,7 @@ void VulkanTimestamps::beginQuery(VulkanCommandBuffer const* commands,
 void VulkanTimestamps::endQuery(VulkanCommandBuffer const* commands,
        VulkanTimerQuery const* query) {
    uint32_t const index = query->getStoppingQueryIndex();
-    vkCmdWriteTimestamp(commands->buffer(), VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, mPool, index);
+    vkCmdWriteTimestamp(commands->cmdbuffer, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, mPool, index);
 }

 VulkanTimestamps::QueryResult VulkanTimestamps::getResult(VulkanTimerQuery const* query) {
@@ -30,9 +30,10 @@
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Panic.h>
+#include <utils/Systrace.h>

 #ifndef NDEBUG
-#include <set>  // For VulkanDriver::debugCommandBegin
+#include <set>
 #endif

 using namespace bluevk;
@@ -100,7 +101,7 @@ VulkanTexture* createEmptyTexture(VkDevice device, VkPhysicalDevice physicalDevi
    return emptyTexture;
 }

-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
+#if VK_ENABLE_VALIDATION
 VKAPI_ATTR VkBool32 VKAPI_CALL debugReportCallback(VkDebugReportFlagsEXT flags,
        VkDebugReportObjectTypeEXT objectType, uint64_t object, size_t location,
        int32_t messageCode, const char* pLayerPrefix, const char* pMessage, void* pUserData) {
@@ -135,7 +136,7 @@ VKAPI_ATTR VkBool32 VKAPI_CALL debugUtilsCallback(VkDebugUtilsMessageSeverityFla
    utils::slog.e << utils::io::endl;
    return VK_FALSE;
 }
-#endif // FVK_EANBLED(FVK_DEBUG_VALIDATION)
+#endif// VK_ENABLE_VALIDATION

 } // anonymous namespace

@@ -158,7 +159,7 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
      mBlitter(mStagePool, mPipelineCache, mFramebufferCache, mSamplerCache),
      mReadPixels(mPlatform->getDevice()) {

-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
+#if VK_ENABLE_VALIDATION
    UTILS_UNUSED const PFN_vkCreateDebugReportCallbackEXT createDebugReportCallback
            = vkCreateDebugReportCallbackEXT;
    VkResult result;
@@ -213,7 +214,7 @@ UTILS_NOINLINE
 Driver* VulkanDriver::create(VulkanPlatform* platform, VulkanContext const& context,
         Platform::DriverConfig const& driverConfig) noexcept {
    assert_invariant(platform);
-    size_t defaultSize = FVK_HANDLE_ARENA_SIZE_IN_MB * 1024U * 1024U;
+    size_t defaultSize = FILAMENT_VULKAN_HANDLE_ARENA_SIZE_IN_MB * 1024U * 1024U;
    Platform::DriverConfig validConfig{
            .handleArenaSize = std::max(driverConfig.handleArenaSize, defaultSize)};
    return new VulkanDriver(platform, context, validConfig);
@@ -265,15 +266,24 @@ void VulkanDriver::tick(int) {
 // rather than the wall clock, because we must wait 3 frames after a DriverAPI-level resource has
 // been destroyed for safe destruction, due to outstanding command buffers and triple buffering.
 void VulkanDriver::collectGarbage() {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("gc");
    // Command buffers need to be submitted and completed before other resources can be gc'd. And
    // its gc() function carrys out the *wait*.
+    SYSTRACE_CONTEXT();
+    SYSTRACE_NAME_BEGIN("gc");
+
    mCommands->gc();
+
+    SYSTRACE_NAME_BEGIN("stagepool-gc");
    mStagePool.gc();
+    SYSTRACE_NAME_END();
+
+    SYSTRACE_NAME_BEGIN("framebuffercache-gc");
    mFramebufferCache.gc();
-    FVK_SYSTRACE_END();
+    SYSTRACE_NAME_END();
+
+    SYSTRACE_NAME_END(); // "gc"
 }
+
 void VulkanDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
    // Do nothing.
 }
@@ -296,21 +306,14 @@ void VulkanDriver::endFrame(uint32_t frameId) {
 }

 void VulkanDriver::flush(int) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("flush");
    mCommands->flush();
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::finish(int dummy) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("finish");
-
    mCommands->flush();
    mCommands->wait();

    mReadPixels.runUntilComplete();
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::createSamplerGroupR(Handle<HwSamplerGroup> sbh, uint32_t count) {
@@ -523,22 +526,21 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
                      << utils::io::endl;
        flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
    }
-    auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, mCommands.get(), mStagePool, nativeWindow, flags);
-    mResourceManager.acquire(swapChain);
+    mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext, mAllocator,
+            mCommands.get(), mStagePool, nativeWindow, flags);
 }

 void VulkanDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t width,
        uint32_t height, uint64_t flags) {
-    if ((flags & backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) != 0 && !isSRGBSwapChainSupported()) {
+    if ((flags & backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) != 0 &&
+        !isSRGBSwapChainSupported()) {
        utils::slog.w << "sRGB swapchain requested, but Platform does not support it"
                      << utils::io::endl;
        flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
    }
    assert_invariant(width > 0 && height > 0 && "Vulkan requires non-zero swap chain dimensions.");
-    auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, mCommands.get(), mStagePool, nullptr, flags, VkExtent2D{width, height});
-    mResourceManager.acquire(swapChain);
+    mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext, mAllocator,
+            mCommands.get(), mStagePool, nullptr, flags, VkExtent2D{width, height});
 }

 void VulkanDriver::createTimerQueryR(Handle<HwTimerQuery> tqh, int) {
@@ -688,15 +690,16 @@ FenceStatus VulkanDriver::getFenceStatus(Handle<HwFence> fh) {
    // Internally we use the VK_INCOMPLETE status to mean "not yet submitted".
    // When this fence gets submitted, its status changes to VK_NOT_READY.
    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
-    if (cmdfence->status.load() == VK_SUCCESS) {
-        return FenceStatus::CONDITION_SATISFIED;
+    if (cmdfence->status.load() == VK_INCOMPLETE) {
+        // This will obviously timeout if Filament creates a fence and immediately waits on it
+        // without calling endFrame() or commit().
+        cmdfence->condition.wait(lock);
+    } else {
+        lock.unlock();
    }
-
-    // Two other states are possible:
-    //  - VK_INCOMPLETE: the corresponding buffer has not yet been submitted.
-    //  - VK_NOT_READY: the buffer has been submitted but not yet signaled.
-    // In either case, we return TIMEOUT_EXPIRED to indicate the fence has not been signaled.
-    return FenceStatus::TIMEOUT_EXPIRED;
+    VkResult result =
+            vkWaitForFences(mPlatform->getDevice(), 1, &cmdfence->fence, VK_TRUE, 0);
+    return result == VK_SUCCESS ? FenceStatus::CONDITION_SATISFIED : FenceStatus::TIMEOUT_EXPIRED;
 }

 // We create all textures using VK_IMAGE_TILING_OPTIMAL, so our definition of "supported" is that
@@ -869,7 +872,7 @@ void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor
    VulkanCommandBuffer& commands = mCommands->get();
    auto ib = mResourceAllocator.handle_cast<VulkanIndexBuffer*>(ibh);
    commands.acquire(ib);
-    ib->buffer.loadFromCpu(commands.buffer(), p.buffer, byteOffset, p.size);
+    ib->buffer.loadFromCpu(commands.cmdbuffer, p.buffer, byteOffset, p.size);

    scheduleDestroy(std::move(p));
 }
@@ -880,7 +883,7 @@ void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescript

    auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
    commands.acquire(bo);
-    bo->buffer.loadFromCpu(commands.buffer(), bd.buffer, byteOffset, bd.size);
+    bo->buffer.loadFromCpu(commands.cmdbuffer, bd.buffer, byteOffset, bd.size);

    scheduleDestroy(std::move(bd));
 }
@@ -891,7 +894,7 @@ void VulkanDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> boh,
    auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
    commands.acquire(bo);
    // TODO: implement unsynchronized version
-    bo->buffer.loadFromCpu(commands.buffer(), bd.buffer, byteOffset, bd.size);
+    bo->buffer.loadFromCpu(commands.cmdbuffer, bd.buffer, byteOffset, bd.size);
    mResourceManager.acquire(bo);
    scheduleDestroy(std::move(bd));
 }
@@ -987,9 +990,6 @@ void VulkanDriver::compilePrograms(CompilerPriorityQueue priority,
 }

 void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassParams& params) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("beginRenderPass");
-
    VulkanRenderTarget* const rt = mResourceAllocator.handle_cast<VulkanRenderTarget*>(rth);
    const VkExtent2D extent = rt->getExtent();
    assert_invariant(extent.width > 0 && extent.height > 0);
@@ -1008,8 +1008,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
    }

    VulkanAttachment depth = rt->getSamples() == 1 ? rt->getDepth() : rt->getMsaaDepth();
-
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FILAMENT_VULKAN_VERBOSE
    if (depth.texture) {
        depth.texture->print();
    }
@@ -1021,7 +1020,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
    // the non-sampling case.
    bool samplingDepthAttachment = false;
    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
+    VkCommandBuffer const cmdbuffer = commands.cmdbuffer;

    UTILS_NOUNROLL
    for (uint8_t samplerGroupIdx = 0; samplerGroupIdx < Program::SAMPLER_BINDING_COUNT;
@@ -1154,7 +1153,6 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
    VkFramebuffer vkfb = mFramebufferCache.getFramebuffer(fbkey);

    // Assign a label to the framebuffer for debugging purposes.
-    #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
    if (UTILS_UNLIKELY(mContext.isDebugUtilsSupported())) {
        auto const topMarker = mCommands->getTopGroupMarker();
        if (!topMarker.empty()) {
@@ -1168,7 +1166,6 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
            vkSetDebugUtilsObjectNameEXT(mPlatform->getDevice(), &info);
        }
    }
-    #endif

    // The current command buffer now owns a reference to the render target and its attachments.
    // Note that we must acquire parent textures, not sidecars.
@@ -1245,15 +1242,11 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
        .params = params,
        .currentSubpass = 0,
    };
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::endRenderPass(int) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("endRenderPass");
-
    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer cmdbuffer = commands.buffer();
+    VkCommandBuffer cmdbuffer = commands.cmdbuffer;
    vkCmdEndRenderPass(cmdbuffer);

    VulkanRenderTarget* rt = mCurrentRenderPass.renderTarget;
@@ -1295,7 +1288,6 @@ void VulkanDriver::endRenderPass(int) {
    }
    mCurrentRenderPass.renderTarget = nullptr;
    mCurrentRenderPass.renderPass = VK_NULL_HANDLE;
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::nextSubpass(int) {
@@ -1306,7 +1298,7 @@ void VulkanDriver::nextSubpass(int) {
    assert_invariant(renderTarget);
    assert_invariant(mCurrentRenderPass.params.subpassMask);

-    vkCmdNextSubpass(mCommands->get().buffer(), VK_SUBPASS_CONTENTS_INLINE);
+    vkCmdNextSubpass(mCommands->get().cmdbuffer, VK_SUBPASS_CONTENTS_INLINE);

    mPipelineCache.bindRenderPass(mCurrentRenderPass.renderPass,
            ++mCurrentRenderPass.currentSubpass);
@@ -1342,9 +1334,6 @@ void VulkanDriver::setRenderPrimitiveRange(Handle<HwRenderPrimitive> rph,
 }

 void VulkanDriver::makeCurrent(Handle<HwSwapChain> drawSch, Handle<HwSwapChain> readSch) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("makeCurrent");
-
    ASSERT_PRECONDITION_NON_FATAL(drawSch == readSch,
            "Vulkan driver does not support distinct draw/read swap chains.");
    VulkanSwapChain* swapChain = mCurrentSwapChain
@@ -1360,14 +1349,9 @@ void VulkanDriver::makeCurrent(Handle<HwSwapChain> drawSch, Handle<HwSwapChain>
    if (UTILS_LIKELY(mDefaultRenderTarget)) {
        mDefaultRenderTarget->bindToSwapChain(*swapChain);
    }
-
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::commit(Handle<HwSwapChain> sch) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("commit");
-
    VulkanSwapChain* swapChain = mResourceAllocator.handle_cast<VulkanSwapChain*>(sch);

    if (mCommands->flush()) {
@@ -1376,7 +1360,6 @@ void VulkanDriver::commit(Handle<HwSwapChain> sch) {

    // Present the backbuffer after the most recent command buffer submission has finished.
    swapChain->present();
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> boh) {
@@ -1408,26 +1391,24 @@ void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
 }

 void VulkanDriver::insertEventMarker(char const* string, uint32_t len) {
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
    mCommands->insertEventMarker(string, len);
-#endif
 }

 void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
    // Turns out all the markers are 0-terminated, so we can just pass it without len.
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
    mCommands->pushGroupMarker(string);
-#endif
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START(string);
+    {
+        SYSTRACE_CONTEXT();
+        SYSTRACE_NAME_BEGIN(string);
+    }
 }

 void VulkanDriver::popGroupMarker(int) {
-#if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
    mCommands->popGroupMarker();
-#endif
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_END();
+    {
+        SYSTRACE_CONTEXT();
+        SYSTRACE_NAME_END();
+    }
 }

 void VulkanDriver::startCapture(int) {}
@@ -1458,9 +1439,6 @@ void VulkanDriver::readBufferSubData(backend::BufferObjectHandle boh,

 void VulkanDriver::blit(TargetBufferFlags buffers, Handle<HwRenderTarget> dst, Viewport dstRect,
        Handle<HwRenderTarget> src, Viewport srcRect, SamplerMagFilter filter) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("blit");
-
    assert_invariant(mCurrentRenderPass.renderPass == VK_NULL_HANDLE);

    // blit operation only support COLOR0 color buffer
@@ -1500,16 +1478,12 @@ void VulkanDriver::blit(TargetBufferFlags buffers, Handle<HwRenderTarget> dst, V
    if (any(buffers & TargetBufferFlags::COLOR0)) {
        mBlitter.blitColor({ dstTarget, dstOffsets, srcTarget, srcOffsets, vkfilter, int(0) });
    }
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> rph,
        const uint32_t instanceCount) {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("draw");
-
    VulkanCommandBuffer* commands = &mCommands->get();
-    VkCommandBuffer cmdbuffer = commands->buffer();
+    VkCommandBuffer cmdbuffer = commands->cmdbuffer;
    const VulkanRenderPrimitive& prim = *mResourceAllocator.handle_cast<VulkanRenderPrimitive*>(rph);

    Handle<HwProgram> programHandle = pipelineState.program;
@@ -1523,7 +1497,7 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
    commands->acquire(prim.vertexBuffer);

    // If this is a debug build, validate the current shader.
-#if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+#if !defined(NDEBUG)
    if (program->bundle.vertex == VK_NULL_HANDLE || program->bundle.fragment == VK_NULL_HANDLE) {
        utils::slog.e << "Binding missing shader: " << program->name.c_str() << utils::io::endl;
    }
@@ -1639,7 +1613,7 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
                // This fallback path is very flaky because the dummy texture might not have
                // matching characteristics. (e.g. if the missing texture is a 3D texture)
                if (UTILS_UNLIKELY(texture->getPrimaryImageLayout() == VulkanLayout::UNDEFINED)) {
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#ifndef NDEBUG
                    utils::slog.w << "Uninitialized texture bound to '" << sampler.name.c_str() << "'";
                    utils::slog.w << " in material '" << program->name.c_str() << "'";
                    utils::slog.w << " at binding point " << +sampler.binding << utils::io::endl;
@@ -1713,7 +1687,6 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
    const uint32_t firstInstId = 0;

    vkCmdDrawIndexed(cmdbuffer, indexCount, instanceCount, firstIndex, vertexOffset, firstInstId);
-    FVK_SYSTRACE_END();
 }

 void VulkanDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGroupCount) {
@@ -23,7 +23,7 @@

 // If any VkRenderPass or VkFramebuffer is unused for more than TIME_BEFORE_EVICTION frames, it
 // is evicted from the cache.
-static constexpr uint32_t TIME_BEFORE_EVICTION = FVK_MAX_COMMAND_BUFFERS;
+static constexpr uint32_t TIME_BEFORE_EVICTION = VK_MAX_COMMAND_BUFFERS;

 using namespace bluevk;

@@ -97,7 +97,7 @@ VkFramebuffer VulkanFboCache::getFramebuffer(FboKey config) noexcept {
        attachments[attachmentCount++] = config.depth;
    }

-    #if FVK_ENABLED(FVK_DEBUG_FBO_CACHE)
+    #if FILAMENT_VULKAN_VERBOSE
    utils::slog.d << "Creating framebuffer " << config.width << "x" << config.height << " "
        << "for render pass " << config.renderPass << ", "
        << "samples = " << int(config.samples) << ", "
@@ -312,7 +312,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
    ASSERT_POSTCONDITION(!error, "Unable to create render pass.");
    mRenderPassCache[config] = {renderPass, mCurrentTime};

-    #if FVK_ENABLED(FVK_DEBUG_FBO_CACHE)
+    #if FILAMENT_VULKAN_VERBOSE
    utils::slog.d << "Created render pass " << renderPass << " with "
        << "samples = " << int(config.samples) << ", "
        << "depth = " << (hasDepth ? 1 : 0) << ", "
@@ -338,9 +338,6 @@ void VulkanFboCache::reset() noexcept {
 // Frees up old framebuffers and render passes, then nulls out their key.  Doesn't bother removing
 // the actual map entry since it is fairly small.
 void VulkanFboCache::gc() noexcept {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("fbocache::gc");
-
    // If this is one of the first few frames, return early to avoid wrapping unsigned integers.
    if (++mCurrentTime <= TIME_BEFORE_EVICTION) {
        return;
@@ -362,7 +359,6 @@ void VulkanFboCache::gc() noexcept {
            iter.value().handle = VK_NULL_HANDLE;
        }
    }
-    FVK_SYSTRACE_END();
 }

 } // namespace filament::backend
@@ -89,35 +89,30 @@ VulkanProgram::VulkanProgram(VkDevice device, const Program& builder) noexcept
        };

        for (size_t i = 0; i < specializationConstants.size(); i++) {
-            uint32_t const offset = uint32_t(i) * 4;
-            pEntries[i] = {
-                .constantID = specializationConstants[i].id,
-                .offset = offset,
-                // Note that bools are 4-bytes in Vulkan
-                // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkBool32.html
-                .size = 4,
-            };
-
-            using SpecConstant = Program::SpecializationConstant::Type;
-            char const* addr = (char*)pData + offset;
-            SpecConstant const& arg = specializationConstants[i].value;
-            if (std::holds_alternative<bool>(arg)) {
-                *((VkBool32*)addr) = std::get<bool>(arg) ? VK_TRUE : VK_FALSE;
-            } else if (std::holds_alternative<float>(arg)) {
-                *((float*)addr) = std::get<float>(arg);
-            } else {
-                *((int32_t*)addr) = std::get<int32_t>(arg);
-            }
+            const uint32_t offset = uint32_t(i) * 4;
+            std::visit([&](auto&& arg) {
+                using T = std::decay_t<decltype(arg)>;
+                pEntries[i] = {
+                        .constantID = specializationConstants[i].id,
+                        .offset = offset,
+                        // Turns out vulkan expects the size of bool to be 4 (verified through
+                        // validation layer). So all expected types are of 4 bytes.
+                        .size = 4,
+                };
+                T* const addr = (T*)((char*)pData + offset);
+                *addr = arg;
+            }, specializationConstants[i].value);
        }
+
        bundle.specializationInfos = pInfo;
    }

    // Make a copy of the binding map
    samplerGroupInfo = builder.getSamplerGroupInfo();
-    #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+    if constexpr (FILAMENT_VULKAN_VERBOSE) {
        utils::slog.d << "Created VulkanProgram " << builder << ", shaders = (" << bundle.vertex
                      << ", " << bundle.fragment << ")" << utils::io::endl;
-    #endif
+    }
 }

 VulkanProgram::VulkanProgram(VkDevice device, VkShaderModule vs, VkShaderModule fs) noexcept
@@ -96,6 +96,10 @@ struct VulkanVertexBuffer : public HwVertexBuffer, VulkanResource {

    void setBuffer(VulkanBufferObject* bufferObject, uint32_t index);

+    inline void terminate() {
+        mResources.clear();
+    }
+
    utils::FixedCapacityVector<VulkanBuffer const*> buffers;

 private:
@@ -110,6 +114,9 @@ struct VulkanIndexBuffer : public HwIndexBuffer, VulkanResource {
          buffer(allocator, stagePool, VK_BUFFER_USAGE_INDEX_BUFFER_BIT, elementSize * indexCount),
          indexType(elementSize == 2 ? VK_INDEX_TYPE_UINT16 : VK_INDEX_TYPE_UINT32) {}

+    void terminate() {
+        buffer.terminate();
+    }
    VulkanBuffer buffer;
    const VkIndexType indexType;
 };
@@ -117,7 +124,9 @@ struct VulkanIndexBuffer : public HwIndexBuffer, VulkanResource {
 struct VulkanBufferObject : public HwBufferObject, VulkanResource {
    VulkanBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool, uint32_t byteCount,
            BufferObjectBinding bindingType, BufferUsage usage);
-
+    void terminate() {
+        buffer.terminate();
+    }
    VulkanBuffer buffer;
    const BufferObjectBinding bindingType;
 };
@@ -218,7 +218,7 @@ bool operator<(const VkImageSubresourceRange& a, const VkImageSubresourceRange&
    return false;
 }

-#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION | FVK_DEBUG_TEXTURE)
+#if FILAMENT_VULKAN_VERBOSE
 #define CASE(VALUE)                                                                                \
    case filament::backend::VulkanLayout::VALUE: {                                                 \
        out << #VALUE;                                                                             \
@@ -166,7 +166,7 @@ bool VulkanPipelineCache::bindDescriptors(VkCommandBuffer cmdbuffer) noexcept {
 }

 bool VulkanPipelineCache::bindPipeline(VulkanCommandBuffer* commands) noexcept {
-    VkCommandBuffer const cmdbuffer = commands->buffer();
+    VkCommandBuffer const cmdbuffer = commands->cmdbuffer;

    PipelineMap::iterator pipelineIter = mPipelines.find(mPipelineRequirements);

@@ -481,10 +481,10 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n

    PipelineCacheEntry cacheEntry = {};

-    #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+    if constexpr (FILAMENT_VULKAN_VERBOSE) {
        utils::slog.d << "vkCreateGraphicsPipelines with shaders = ("
                << shaderStages[0].module << ", " << shaderStages[1].module << ")" << utils::io::endl;
-    #endif
+    }
    VkResult error = vkCreateGraphicsPipelines(mDevice, VK_NULL_HANDLE, 1, &pipelineCreateInfo,
            VKALLOC, &cacheEntry.handle);
    assert_invariant(error == VK_SUCCESS);
@@ -678,10 +678,10 @@ void VulkanPipelineCache::terminate() noexcept {
    mDummyMemory = VK_NULL_HANDLE;
 }

-void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
+void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& cmdbuffer) {
    // The timestamp associated with a given cache entry represents "time" as a count of flush
    // events since the cache was constructed. If any cache entry was most recently used over
-    // FVK_MAX_PIPELINE_AGE flush events in the past, then we can be sure that it is no longer
+    // VK_MAX_PIPELINE_AGE flush events in the past, then we can be sure that it is no longer
    // being used by the GPU, and is therefore safe to destroy or reclaim.
    ++mCurrentTime;

@@ -698,7 +698,7 @@ void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
    using ConstDescIterator = decltype(mDescriptorSets)::const_iterator;
    for (ConstDescIterator iter = mDescriptorSets.begin(); iter != mDescriptorSets.end();) {
        const DescriptorCacheEntry& cacheEntry = iter.value();
-        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+        if (cacheEntry.lastUsed + VK_MAX_PIPELINE_AGE < mCurrentTime) {
            auto& arenas = mPipelineLayouts[cacheEntry.pipelineLayout].descriptorSetArenas;
            for (uint32_t i = 0; i < DESCRIPTOR_TYPE_COUNT; ++i) {
                arenas[i].push_back(cacheEntry.handles[i]);
@@ -712,11 +712,11 @@ void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
    }

    // Evict any pipelines that have not been used in a while.
-    // Any pipeline older than FVK_MAX_COMMAND_BUFFERS can be safely destroyed.
+    // Any pipeline older than VK_MAX_COMMAND_BUFFERS can be safely destroyed.
    using ConstPipeIterator = decltype(mPipelines)::const_iterator;
    for (ConstPipeIterator iter = mPipelines.begin(); iter != mPipelines.end();) {
        const PipelineCacheEntry& cacheEntry = iter.value();
-        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+        if (cacheEntry.lastUsed + VK_MAX_PIPELINE_AGE < mCurrentTime) {
            vkDestroyPipeline(mDevice, iter->second.handle, VKALLOC);
            iter = mPipelines.erase(iter);
        } else {
@@ -728,15 +728,15 @@ void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
    using ConstLayoutIterator = decltype(mPipelineLayouts)::const_iterator;
    for (ConstLayoutIterator iter = mPipelineLayouts.begin(); iter != mPipelineLayouts.end();) {
        const PipelineLayoutCacheEntry& cacheEntry = iter.value();
-        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+        if (cacheEntry.lastUsed + VK_MAX_PIPELINE_AGE < mCurrentTime) {
            vkDestroyPipelineLayout(mDevice, iter->second.handle, VKALLOC);
            for (auto setLayout : iter->second.descriptorSetLayouts) {
-                #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
+#ifndef NDEBUG
                PipelineLayoutKey key = iter.key();
                for (auto& pair : mDescriptorSets) {
                    assert_invariant(pair.second.pipelineLayout != key);
                }
-                #endif
+#endif
                vkDestroyDescriptorSetLayout(mDevice, setLayout, VKALLOC);
            }
            auto& arenas = iter->second.descriptorSetArenas;
@@ -755,7 +755,7 @@ void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
    // destroy the extinct pools, which implicitly frees their associated descriptor sets.
    bool canPurgeExtinctPools = true;
    for (auto& bundle : mExtinctDescriptorBundles) {
-        if (bundle.lastUsed + FVK_MAX_PIPELINE_AGE >= mCurrentTime) {
+        if (bundle.lastUsed + VK_MAX_PIPELINE_AGE >= mCurrentTime) {
            canPurgeExtinctPools = false;
            break;
        }
@@ -799,7 +799,7 @@ VkDescriptorPool VulkanPipelineCache::createDescriptorPool(uint32_t size) const
 void VulkanPipelineCache::destroyLayoutsAndDescriptors() noexcept {
    // Our current descriptor set strategy can cause the # of descriptor sets to explode in certain
    // situations, so it's interesting to report the number that get stuffed into the cache.
-    #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
+    #ifndef NDEBUG
    utils::slog.d << "Destroying " << mDescriptorSets.size() << " bundles of descriptor sets."
            << utils::io::endl;
    #endif
@@ -808,7 +808,7 @@ void VulkanPipelineCache::destroyLayoutsAndDescriptors() noexcept {

    // Our current layout bundle strategy can cause the # of layout bundles to explode in certain
    // situations, so it's interesting to report the number that get stuffed into the cache.
-    #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
+    #ifndef NDEBUG
    utils::slog.d << "Destroying " << mPipelineLayouts.size() << " pipeline layouts."
                  << utils::io::endl;
    #endif
@@ -340,7 +340,7 @@ private:

    // The timestamp associated with a given cache entry represents time as a count of flush
    // events since the cache was constructed. If any cache entry was most recently used over
-    // FVK_MAX_PIPELINE_AGE flushes in the past, then we can be sure that it is no longer
+    // VK_MAX_PIPELINE_AGE flushes in the past, then we can be sure that it is no longer
    // being used by the GPU, and is therefore safe to destroy or reclaim.
    using Timestamp = uint64_t;
    Timestamp mCurrentTime = 0;
@@ -167,7 +167,7 @@ void VulkanReadPixels::run(VulkanRenderTarget const* srcTarget, uint32_t const x
    VkImage stagingImage;
    vkCreateImage(device, &imageInfo, VKALLOC, &stagingImage);

-#if FVK_ENABLED(FVK_DEBUG_READ_PIXELS)
+#if FILAMENT_VULKAN_VERBOSE
    utils::slog.d << "readPixels created image=" << stagingImage
                  << " to copy from image=" << srcTexture->getVkImage()
                  << " src-layout=" << srcTexture->getLayout(0, 0) << utils::io::endl;
@@ -101,6 +101,10 @@ public:
    template<typename D, typename B>
    inline void destruct(Handle<B> handle) noexcept {
        auto obj = handle_cast<D*>(handle);
+        if constexpr (std::is_base_of_v<VulkanIndexBuffer, D>
+                      || std::is_base_of_v<VulkanBufferObject, D>) {
+            obj->terminate();
+        }
        TRACK_DECREMENT();
        mHandleAllocatorImpl.deallocate(handle, obj);
    }
@@ -164,10 +164,6 @@ private:
 public:
    using const_iterator = FixedSizeArray::const_iterator;

-    inline ~FixedCapacityResourceSet() {
-        clear();
-    }
-
    inline const_iterator begin() {
        if (mInd == 0) {
            return mArray.cend();
@@ -22,7 +22,7 @@

 #include <utils/Panic.h>

-static constexpr uint32_t TIME_BEFORE_EVICTION = FVK_MAX_COMMAND_BUFFERS;
+static constexpr uint32_t TIME_BEFORE_EVICTION = VK_MAX_COMMAND_BUFFERS;

 namespace filament::backend {

@@ -59,7 +59,7 @@ VulkanStage const* VulkanStagePool::acquireStage(uint32_t numBytes) {
    UTILS_UNUSED_IN_RELEASE VkResult result = vmaCreateBuffer(mAllocator, &bufferInfo,
            &allocInfo, &stage->buffer, &stage->memory, nullptr);

-#if FVK_ENABLED(FVK_DEBUG_ALLOCATION)
+#ifndef NDEBUG
    if (result != VK_SUCCESS) {
        utils::slog.e << "Allocation error: " << result << utils::io::endl;
    }
@@ -111,7 +111,7 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
    assert_invariant(result == VK_SUCCESS);

    VkImageAspectFlags const aspectFlags = getImageAspect(vkformat);
-    VkCommandBuffer const cmdbuffer = mCommands->get().buffer();
+    const VkCommandBuffer cmdbuffer = mCommands->get().cmdbuffer;

    // We use VK_IMAGE_LAYOUT_GENERAL here because the spec says:
    // "Host access to image memory is only well-defined for linear images and for image
@@ -129,9 +129,6 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
 }

 void VulkanStagePool::gc() noexcept {
-    FVK_SYSTRACE_CONTEXT();
-    FVK_SYSTRACE_START("stagepool::gc");
-
    // If this is one of the first few frames, return early to avoid wrapping unsigned integers.
    if (++mCurrentFrame <= TIME_BEFORE_EVICTION) {
        return;
@@ -185,7 +182,6 @@ void VulkanStagePool::gc() noexcept {
            mUsedImages.insert(image);
        }
    }
-    FVK_SYSTRACE_END();
 }

 void VulkanStagePool::terminate() noexcept {
@@ -79,7 +79,7 @@ void VulkanSwapChain::update() {

 void VulkanSwapChain::present() {
    if (!mHeadless) {
-        VkCommandBuffer const cmdbuf = mCommands->get().buffer();
+        VkCommandBuffer const cmdbuf = mCommands->get().cmdbuffer;
        VkImageSubresourceRange const subresources{
                .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
                .baseMipLevel = 0,
@@ -106,7 +106,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,

    if (any(usage & TextureUsage::SAMPLEABLE)) {

-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if VK_ENABLE_VALIDATION
        // Validate that the format is actually sampleable.
        VkFormatProperties props;
        vkGetPhysicalDeviceFormatProperties(physicalDevice, mVkFormat, &props);
@@ -160,7 +160,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
    imageInfo.samples = (VkSampleCountFlagBits) samples;

    VkResult error = vkCreateImage(mDevice, &imageInfo, VKALLOC, &mTextureImage);
-    if (error || FVK_ENABLED_BOOL(FVK_DEBUG_TEXTURE)) {
+    if (error || FILAMENT_VULKAN_VERBOSE) {
        utils::slog.d << "vkCreateImage: "
            << "image = " << mTextureImage << ", "
            << "result = " << error << ", "
@@ -227,7 +227,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
        VkImageSubresourceRange range = { getImageAspect(), 0, levels, 0, layers };

        VulkanCommandBuffer& commands = mCommands->get();
-        VkCommandBuffer const cmdbuf = commands.buffer();
+        VkCommandBuffer const cmdbuf = commands.cmdbuffer;
        commands.acquire(this);

        transitionLayout(cmdbuf, range, ImgUtil::getDefaultLayout(imageInfo.usage));
@@ -280,7 +280,7 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
    vmaFlushAllocation(mAllocator, stage->memory, 0, hostData->size);

    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuf = commands.buffer();
+    VkCommandBuffer const cmdbuf = commands.cmdbuffer;
    commands.acquire(this);

    VkBufferImageCopy copyRegion = {
@@ -343,7 +343,7 @@ void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, u
    vmaFlushAllocation(mAllocator, stage->memory, 0, hostData.size);

    VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuf = commands.buffer();
+    VkCommandBuffer const cmdbuf = commands.cmdbuffer;
    commands.acquire(this);

    // TODO: support blit-based format conversion for 3D images and cubemaps.
@@ -417,7 +417,7 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
        VulkanLayout newLayout) {
    VulkanLayout oldLayout = getLayout(range.baseArrayLayer, range.baseMipLevel);

-    #if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
+    #if FILAMENT_VULKAN_VERBOSE
    utils::slog.i << "transition layout of " << mTextureImage << ",layer=" << range.baseArrayLayer
                  << ",level=" << range.baseMipLevel << " from=" << oldLayout << " to=" << newLayout
                  << " format=" << mVkFormat
@@ -463,7 +463,7 @@ VulkanLayout VulkanTexture::getLayout(uint32_t layer, uint32_t level) const {
    return mSubresourceLayouts.get(key);
 }

-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FILAMENT_VULKAN_VERBOSE
 void VulkanTexture::print() const {
    const uint32_t firstLayer = 0;
    const uint32_t lastLayer = firstLayer + mPrimaryViewRange.layerCount;
@@ -84,7 +84,7 @@ struct VulkanTexture : public HwTexture, VulkanResource {
    // For now this always returns either DEPTH or COLOR.
    VkImageAspectFlags getImageAspect() const;

-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FILAMENT_VULKAN_VERBOSE
    void print() const;
 #endif

@@ -45,11 +45,11 @@ constexpr uint32_t const INVALID_VK_INDEX = 0xFFFFFFFF;

 typedef std::unordered_set<std::string_view> ExtensionSet;

-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
+#if VK_ENABLE_VALIDATION
 // These strings need to be allocated outside a function stack
 const std::string_view DESIRED_LAYERS[] = {
        "VK_LAYER_KHRONOS_validation",
-#if FVK_ENABLED(FVK_DEBUG_DUMP_API)
+#if FILAMENT_VULKAN_DUMP_API
        "VK_LAYER_LUNARG_api_dump",
 #endif
 #if defined(ENABLE_RENDERDOC)
@@ -75,7 +75,7 @@ FixedCapacityVector<const char*> getEnabledLayers() {
    }
    return enabledLayers;
 }
-#endif // FVK_EANBLED(FVK_DEBUG_VALIDATION)
+#endif

 void printDeviceInfo(VkInstance instance, VkPhysicalDevice device) {
    // Print some driver or MoltenVK information if it is available.
@@ -129,30 +129,30 @@ void printDeviceInfo(VkInstance instance, VkPhysicalDevice device) {
                  << minor << ")" << utils::io::endl;
 }

-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
 void printDepthFormats(VkPhysicalDevice device) {
    // For diagnostic purposes, print useful information about available depth formats.
    // Note that Vulkan is more constrained than OpenGL ES 3.1 in this area.
-    const VkFormatFeatureFlags required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT
+    if constexpr (VK_ENABLE_VALIDATION && FILAMENT_VULKAN_VERBOSE) {
+        const VkFormatFeatureFlags required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT
                                              | VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
-    utils::slog.i << "Sampleable depth formats: ";
-    for (VkFormat format = (VkFormat) 1;;) {
-        VkFormatProperties props;
-        vkGetPhysicalDeviceFormatProperties(device, format, &props);
-        if ((props.optimalTilingFeatures & required) == required) {
-            utils::slog.i << format << " ";
+        utils::slog.i << "Sampleable depth formats: ";
+        for (VkFormat format = (VkFormat) 1;;) {
+            VkFormatProperties props;
+            vkGetPhysicalDeviceFormatProperties(device, format, &props);
+            if ((props.optimalTilingFeatures & required) == required) {
+                utils::slog.i << format << " ";
+            }
+            if (format == VK_FORMAT_ASTC_12x12_SRGB_BLOCK) {
+                utils::slog.i << utils::io::endl;
+                break;
+            }
+            format = (VkFormat) (1 + (int) format);
        }
-        if (format == VK_FORMAT_ASTC_12x12_SRGB_BLOCK) {
-            utils::slog.i << utils::io::endl;
-            break;
-        }
-        format = (VkFormat) (1 + (int) format);
    }
 }
-#endif

 ExtensionSet getInstanceExtensions() {
-    std::string_view const TARGET_EXTS[] = {
+    std::array<std::string_view, 4> const TARGET_EXTS = {
            // Request all cross-platform extensions.
            VK_KHR_SURFACE_EXTENSION_NAME,
            VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
@@ -160,10 +160,6 @@ ExtensionSet getInstanceExtensions() {
            // Request these if available.
            VK_EXT_DEBUG_UTILS_EXTENSION_NAME,
            VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME,
-
-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
-            VK_EXT_DEBUG_REPORT_EXTENSION_NAME,
-#endif
    };
    ExtensionSet exts;
    FixedCapacityVector<VkExtensionProperties> const availableExts
@@ -180,7 +176,7 @@ ExtensionSet getInstanceExtensions() {
 }

 ExtensionSet getDeviceExtensions(VkPhysicalDevice device) {
-    std::string_view const TARGET_EXTS[] = {
+    std::array<std::string_view, 5> const TARGET_EXTS = {
            VK_EXT_DEBUG_MARKER_EXTENSION_NAME,
            VK_KHR_PORTABILITY_SUBSET_EXTENSION_NAME,
            VK_KHR_MAINTENANCE1_EXTENSION_NAME,
@@ -207,7 +203,7 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
    VkInstanceCreateInfo instanceCreateInfo = {};
    bool validationFeaturesSupported = false;

-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
+#if VK_ENABLE_VALIDATION
    auto const enabledLayers = getEnabledLayers();
    if (!enabledLayers.empty()) {
        // If layers are supported, Check if VK_EXT_validation_features is supported.
@@ -229,17 +225,19 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
 #else
        utils::slog.d << "Validation layer not available; did you install the Vulkan SDK?\n"
                      << "Please ensure that VK_LAYER_PATH is set correctly." << utils::io::endl;
-#endif // __ANDROID__
-
+#endif
    }
-#endif // FVK_ENABLED(FVK_DEBUG_VALIDATION)
+#endif// VK_ENABLE_VALIDATION

    // The Platform class can require 1 or 2 instance extensions, plus we'll request at most 5
    // instance extensions here in the common code. So that's a max of 7.
-    static constexpr uint32_t MAX_INSTANCE_EXTENSION_COUNT = 8;
+    static constexpr uint32_t MAX_INSTANCE_EXTENSION_COUNT = 7;
    const char* ppEnabledExtensions[MAX_INSTANCE_EXTENSION_COUNT];
    uint32_t enabledExtensionCount = 0;

+#if VK_ENABLE_VALIDATION && defined(__ANDROID__)
+    ppEnabledExtensions[enabledExtensionCount++] = VK_EXT_DEBUG_REPORT_EXTENSION_NAME;
+#endif
    if (validationFeaturesSupported) {
        ppEnabledExtensions[enabledExtensionCount++] = VK_EXT_VALIDATION_FEATURES_EXTENSION_NAME;
    }
@@ -253,7 +251,7 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
    VkApplicationInfo appInfo = {};
    appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
    appInfo.apiVersion
-            = VK_MAKE_API_VERSION(0, FVK_REQUIRED_VERSION_MAJOR, FVK_REQUIRED_VERSION_MINOR, 0);
+            = VK_MAKE_API_VERSION(0, VK_REQUIRED_VERSION_MAJOR, VK_REQUIRED_VERSION_MINOR, 0);
    instanceCreateInfo.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO;
    instanceCreateInfo.pApplicationInfo = &appInfo;
    instanceCreateInfo.enabledExtensionCount = enabledExtensionCount;
@@ -275,8 +273,12 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
    }

    VkResult result = vkCreateInstance(&instanceCreateInfo, VKALLOC, &instance);
-    ASSERT_POSTCONDITION(result == VK_SUCCESS, "Unable to create Vulkan instance. Result=%d",
-            result);
+#ifndef NDEBUG
+    if (result != VK_SUCCESS) {
+        utils::slog.e << "Unable to create instance: " << result << utils::io::endl;
+    }
+#endif
+    ASSERT_POSTCONDITION(result == VK_SUCCESS, "Unable to create Vulkan instance.");
    return instance;
 }

@@ -440,10 +442,10 @@ VkPhysicalDevice selectPhysicalDevice(VkInstance instance,
        int const minor = VK_VERSION_MINOR(targetDeviceProperties.apiVersion);

        // Does the device support the required Vulkan level?
-        if (major < FVK_REQUIRED_VERSION_MAJOR) {
+        if (major < VK_REQUIRED_VERSION_MAJOR) {
            continue;
        }
-        if (major == FVK_REQUIRED_VERSION_MAJOR && minor < FVK_REQUIRED_VERSION_MINOR) {
+        if (major == VK_REQUIRED_VERSION_MAJOR && minor < VK_REQUIRED_VERSION_MINOR) {
            continue;
        }

@@ -510,7 +512,7 @@ VkPhysicalDevice selectPhysicalDevice(VkInstance instance,

 VkFormat findSupportedFormat(VkPhysicalDevice device) {
    VkFormatFeatureFlags const features = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
-    VkFormat const formats[] = {VK_FORMAT_D32_SFLOAT, VK_FORMAT_X8_D24_UNORM_PACK32};
+    std::array<VkFormat, 2> const formats = {VK_FORMAT_D32_SFLOAT, VK_FORMAT_X8_D24_UNORM_PACK32};
    for (VkFormat format: formats) {
        VkFormatProperties props;
        vkGetPhysicalDeviceFormatProperties(device, format, &props);
@@ -664,9 +666,7 @@ Driver* VulkanPlatform::createDriver(void* sharedContext,
    // just yet, since that would require a corollary change to the "aspect" flags for the VkImage.
    context.mDepthFormat = findSupportedFormat(mImpl->mPhysicalDevice);

-#if FVK_ENABLED(FVK_DEBUG_VALIDATION)
    printDepthFormats(mImpl->mPhysicalDevice);
-#endif

    // Keep a copy of context for swapchains.
    mImpl->mContext = context;
@@ -211,5 +211,66 @@ int runTests() {
    return RUN_ALL_TESTS();
 }

+void getPixelInfo(PixelDataFormat format, PixelDataType type, size_t& outComponents, int& outBpp) {
+    assert_invariant(type != PixelDataType::COMPRESSED);
+    switch (format) {
+        case PixelDataFormat::UNUSED:
+        case PixelDataFormat::R:
+        case PixelDataFormat::R_INTEGER:
+        case PixelDataFormat::DEPTH_COMPONENT:
+        case PixelDataFormat::ALPHA:
+            outComponents = 1;
+            break;
+        case PixelDataFormat::RG:
+        case PixelDataFormat::RG_INTEGER:
+        case PixelDataFormat::DEPTH_STENCIL:
+            outComponents = 2;
+            break;
+        case PixelDataFormat::RGB:
+        case PixelDataFormat::RGB_INTEGER:
+            outComponents = 3;
+            break;
+        case PixelDataFormat::RGBA:
+        case PixelDataFormat::RGBA_INTEGER:
+            outComponents = 4;
+            break;
+    }
+
+    outBpp = outComponents;
+    switch (type) {
+        case PixelDataType::COMPRESSED: // Impossible -- to squash the IDE warnings
+        case PixelDataType::UBYTE:
+        case PixelDataType::BYTE:
+            // nothing to do
+            break;
+        case PixelDataType::USHORT:
+        case PixelDataType::SHORT:
+        case PixelDataType::HALF:
+            outBpp *= 2;
+            break;
+        case PixelDataType::UINT:
+        case PixelDataType::INT:
+        case PixelDataType::FLOAT:
+            outBpp *= 4;
+            break;
+        case PixelDataType::UINT_10F_11F_11F_REV:
+            // Special case, format must be RGB and uses 4 bytes
+            assert_invariant(format == PixelDataFormat::RGB);
+            outBpp = 4;
+            break;
+        case PixelDataType::UINT_2_10_10_10_REV:
+            // Special case, format must be RGBA and uses 4 bytes
+            assert_invariant(format == PixelDataFormat::RGBA);
+            outBpp = 4;
+            break;
+        case PixelDataType::USHORT_565:
+            // Special case, format must be RGB and uses 2 bytes
+            assert_invariant(format == PixelDataFormat::RGB);
+            outBpp = 2;
+            break;
+    }
+}
+
+
 } // namespace test

@@ -75,6 +75,12 @@ private:
    filament::backend::Handle<filament::backend::HwBufferObject> uniform;
 };

+
+// Utilities
+
+void getPixelInfo(filament::backend::PixelDataFormat format, filament::backend::PixelDataType type,
+        size_t& outComponents, int& outBpp);
+
 } // namespace test

 #endif
@@ -1,174 +0,0 @@
-/*
- * Copyright (C) 2023 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_BACKENDTESTUTILS_H
-#define TNT_BACKENDTESTUTILS_H
-
-#include <cstddef>
-
-#include <backend/PixelBufferDescriptor.h>
-
-using namespace filament;
-using namespace filament::backend;
-
-inline void getPixelInfo(PixelDataFormat format, PixelDataType type, size_t& outComponents, int& outBpp) {
-    assert_invariant(type != PixelDataType::COMPRESSED);
-    switch (format) {
-        case PixelDataFormat::UNUSED:
-        case PixelDataFormat::R:
-        case PixelDataFormat::R_INTEGER:
-        case PixelDataFormat::DEPTH_COMPONENT:
-        case PixelDataFormat::ALPHA:
-            outComponents = 1;
-            break;
-        case PixelDataFormat::RG:
-        case PixelDataFormat::RG_INTEGER:
-        case PixelDataFormat::DEPTH_STENCIL:
-            outComponents = 2;
-            break;
-        case PixelDataFormat::RGB:
-        case PixelDataFormat::RGB_INTEGER:
-            outComponents = 3;
-            break;
-        case PixelDataFormat::RGBA:
-        case PixelDataFormat::RGBA_INTEGER:
-            outComponents = 4;
-            break;
-    }
-
-    outBpp = outComponents;
-    switch (type) {
-        case PixelDataType::COMPRESSED: // Impossible -- to squash the IDE warnings
-        case PixelDataType::UBYTE:
-        case PixelDataType::BYTE:
-            // nothing to do
-            break;
-        case PixelDataType::USHORT:
-        case PixelDataType::SHORT:
-        case PixelDataType::HALF:
-            outBpp *= 2;
-            break;
-        case PixelDataType::UINT:
-        case PixelDataType::INT:
-        case PixelDataType::FLOAT:
-            outBpp *= 4;
-            break;
-        case PixelDataType::UINT_10F_11F_11F_REV:
-            // Special case, format must be RGB and uses 4 bytes
-            assert_invariant(format == PixelDataFormat::RGB);
-            outBpp = 4;
-            break;
-        case PixelDataType::UINT_2_10_10_10_REV:
-            // Special case, format must be RGBA and uses 4 bytes
-            assert_invariant(format == PixelDataFormat::RGBA);
-            outBpp = 4;
-            break;
-        case PixelDataType::USHORT_565:
-            // Special case, format must be RGB and uses 2 bytes
-            assert_invariant(format == PixelDataFormat::RGB);
-            outBpp = 2;
-            break;
-    }
-}
-
-template<typename ComponentType>
-static void fillCheckerboard(void* buffer, size_t size, size_t stride, size_t components,
-        ComponentType value) {
-    ComponentType* row = (ComponentType*)buffer;
-    int p = 0;
-    for (int r = 0; r < size; r++) {
-        ComponentType* pixel = row;
-        for (int col = 0; col < size; col++) {
-            // Generate a checkerboard pattern.
-            if ((p & 0x0010) ^ ((p / size) & 0x0010)) {
-                // Turn on the first component (red).
-                pixel[0] = value;
-            }
-            pixel += components;
-            p++;
-        }
-        row += stride * components;
-    }
-}
-
-static PixelBufferDescriptor checkerboardPixelBuffer(PixelDataFormat format, PixelDataType type,
-        size_t size, size_t bufferPadding = 0) {
-    size_t components; int bpp;
-    getPixelInfo(format, type, components, bpp);
-
-    size_t bufferSize = size + bufferPadding * 2;
-    uint8_t* buffer = (uint8_t*) calloc(1, bufferSize * bufferSize * bpp);
-
-    uint8_t* ptr = buffer + (bufferSize * bufferPadding * bpp) + (bufferPadding * bpp);
-
-    switch (type) {
-        case PixelDataType::BYTE:
-            fillCheckerboard<int8_t>(ptr, size, bufferSize, components, 1);
-            break;
-
-        case PixelDataType::UBYTE:
-            fillCheckerboard<uint8_t>(ptr, size, bufferSize, components, 0xFF);
-            break;
-
-        case PixelDataType::SHORT:
-            fillCheckerboard<int16_t>(ptr, size, bufferSize, components, 1);
-            break;
-
-        case PixelDataType::USHORT:
-            fillCheckerboard<uint16_t>(ptr, size, bufferSize, components, 1u);
-            break;
-
-        case PixelDataType::UINT:
-            fillCheckerboard<uint32_t>(ptr, size, bufferSize, components, 1u);
-            break;
-
-        case PixelDataType::INT:
-            fillCheckerboard<int32_t>(ptr, size, bufferSize, components, 1);
-            break;
-
-        case PixelDataType::FLOAT:
-            fillCheckerboard<float>(ptr, size, bufferSize, components, 1.0f);
-            break;
-
-        case PixelDataType::HALF:
-            fillCheckerboard<math::half>(ptr, size, bufferSize, components, math::half(1.0f));
-            break;
-
-        case PixelDataType::UINT_2_10_10_10_REV:
-            fillCheckerboard<uint32_t>(ptr, size, bufferSize, 1, 0xC00003FF /* red */);
-            break;
-
-        case PixelDataType::USHORT_565:
-            fillCheckerboard<uint16_t>(ptr, size, bufferSize, 1, 0xF800 /* red */);
-            break;
-
-        case PixelDataType::UINT_10F_11F_11F_REV:
-            fillCheckerboard<uint32_t>(ptr, size, bufferSize, 1, 0x000003C0 /* red */);
-            break;
-
-        case PixelDataType::COMPRESSED:
-            break;
-    }
-
-    PixelBufferDescriptor descriptor(buffer, bufferSize * bufferSize * bpp, format, type,
-            1, bufferPadding, bufferPadding, bufferSize, [](void* buffer, size_t size, void* user) {
-                free(buffer);
-            }, nullptr);
-    return descriptor;
-}
-
-#endif  // TNT_BACKENDTESTUTILS_H
@@ -18,7 +18,6 @@

 #include "ShaderGenerator.h"
 #include "TrianglePrimitive.h"
-#include "BackendTestUtils.h"

 #include "private/filament/SamplerInterfaceBlock.h"
 #include "private/backend/SamplerGroup.h"
@@ -114,7 +113,91 @@ namespace test {

 template<typename componentType> inline componentType getMaxValue();

+template<typename ComponentType>
+static void fillCheckerboard(void* buffer, size_t size, size_t stride, size_t components,
+        ComponentType value) {
+    ComponentType* row = (ComponentType*)buffer;
+    int p = 0;
+    for (int r = 0; r < size; r++) {
+        ComponentType* pixel = row;
+        for (int col = 0; col < size; col++) {
+            // Generate a checkerboard pattern.
+            if ((p & 0x0010) ^ ((p / size) & 0x0010)) {
+                // Turn on the first component (red).
+                pixel[0] = value;
+            }
+            pixel += components;
+            p++;
+        }
+        row += stride * components;
+    }
+}

+static PixelBufferDescriptor checkerboardPixelBuffer(PixelDataFormat format, PixelDataType type,
+        size_t size, size_t bufferPadding = 0) {
+    size_t components; int bpp;
+    getPixelInfo(format, type, components, bpp);
+
+    size_t bufferSize = size + bufferPadding * 2;
+    uint8_t* buffer = (uint8_t*) calloc(1, bufferSize * bufferSize * bpp);
+
+    uint8_t* ptr = buffer + (bufferSize * bufferPadding * bpp) + (bufferPadding * bpp);
+
+    switch (type) {
+        case PixelDataType::BYTE:
+            fillCheckerboard<int8_t>(ptr, size, bufferSize, components, 1);
+            break;
+
+        case PixelDataType::UBYTE:
+            fillCheckerboard<uint8_t>(ptr, size, bufferSize, components, 0xFF);
+            break;
+
+        case PixelDataType::SHORT:
+            fillCheckerboard<int16_t>(ptr, size, bufferSize, components, 1);
+            break;
+
+        case PixelDataType::USHORT:
+            fillCheckerboard<uint16_t>(ptr, size, bufferSize, components, 1u);
+            break;
+
+        case PixelDataType::UINT:
+            fillCheckerboard<uint32_t>(ptr, size, bufferSize, components, 1u);
+            break;
+
+        case PixelDataType::INT:
+            fillCheckerboard<int32_t>(ptr, size, bufferSize, components, 1);
+            break;
+
+        case PixelDataType::FLOAT:
+            fillCheckerboard<float>(ptr, size, bufferSize, components, 1.0f);
+            break;
+
+        case PixelDataType::HALF:
+            fillCheckerboard<math::half>(ptr, size, bufferSize, components, math::half(1.0f));
+            break;
+
+        case PixelDataType::UINT_2_10_10_10_REV:
+            fillCheckerboard<uint32_t>(ptr, size, bufferSize, 1, 0xC00003FF /* red */);
+            break;
+
+        case PixelDataType::USHORT_565:
+            fillCheckerboard<uint16_t>(ptr, size, bufferSize, 1, 0xF800 /* red */);
+            break;
+
+        case PixelDataType::UINT_10F_11F_11F_REV:
+            fillCheckerboard<uint32_t>(ptr, size, bufferSize, 1, 0x000003C0 /* red */);
+            break;
+
+        case PixelDataType::COMPRESSED:
+            break;
+    }
+
+    PixelBufferDescriptor descriptor(buffer, bufferSize * bufferSize * bpp, format, type,
+            1, bufferPadding, bufferPadding, bufferSize, [](void* buffer, size_t size, void* user) {
+                free(buffer);
+            }, nullptr);
+    return descriptor;
+}

 inline std::string stringReplace(const std::string& find, const std::string& replace,
        std::string source) {
@@ -1,245 +0,0 @@
-/*
- * Copyright (C) 2023 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 "BackendTest.h"
-
-#include "ShaderGenerator.h"
-#include "TrianglePrimitive.h"
-#include "BackendTestUtils.h"
-
-#include "private/backend/SamplerGroup.h"
-
-namespace {
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Shaders
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-std::string vertex (R"(#version 450 core
-
-layout(location = 0) in vec4 mesh_position;
-layout(location = 0) out vec2 uv;
-
-void main() {
-    gl_Position = vec4(mesh_position.xy, 0.0, 1.0);
-    uv = (mesh_position.xy * 0.5 + 0.5);
-}
-)");
-
-std::string fragment (R"(#version 450 core
-
-layout(location = 0) out vec4 fragColor;
-layout(location = 0) in vec2 uv;
-
-layout(location = 0, set = 1) uniform sampler2D backend_test_sib_tex;
-
-void main() {
-    fragColor = textureLod(backend_test_sib_tex, uv, 1);
-}
-)");
-
-std::string whiteFragment (R"(#version 450 core
-
-layout(location = 0) out vec4 fragColor;
-layout(location = 0) in vec2 uv;
-
-layout(location = 0, set = 1) uniform sampler2D backend_test_sib_tex;
-
-void main() {
-    fragColor = vec4(1.0);
-}
-)");
-
-}
-
-namespace test {
-
-using namespace filament;
-using namespace filament::backend;
-
-TEST_F(BackendTest, SetMinMaxLevel) {
-    auto& api = getDriverApi();
-    api.startCapture(0);
-
-    // The test is executed within this block scope to force destructors to run before
-    // executeCommands().
-    {
-        // Create a SwapChain and make it current.
-        auto swapChain = createSwapChain();
-        api.makeCurrent(swapChain, swapChain);
-
-        // Create a program that draws only white.
-        Handle<HwProgram> whiteProgram;
-        {
-            ShaderGenerator shaderGen(vertex, whiteFragment, sBackend, sIsMobilePlatform);
-            Program p = shaderGen.getProgram(api);
-            Program::Sampler sampler{utils::CString("backend_test_sib_tex"), 0};
-            p.setSamplerGroup(0, ShaderStageFlags::FRAGMENT, &sampler, 1);
-            whiteProgram = api.createProgram(std::move(p));
-        }
-
-        // Create a program that samples a texture.
-        Handle<HwProgram> textureProgram;
-        {
-            SamplerInterfaceBlock sib = filament::SamplerInterfaceBlock::Builder()
-                    .name("backend_test_sib")
-                    .stageFlags(backend::ShaderStageFlags::FRAGMENT)
-                    .add( {{"tex", SamplerType::SAMPLER_2D, SamplerFormat::FLOAT, Precision::HIGH }} )
-                    .build();
-            ShaderGenerator shaderGen(vertex, fragment, sBackend, sIsMobilePlatform, &sib);
-            Program p = shaderGen.getProgram(api);
-            Program::Sampler sampler{utils::CString("backend_test_sib_tex"), 0};
-            p.setSamplerGroup(0, ShaderStageFlags::FRAGMENT, &sampler, 1);
-            textureProgram = api.createProgram(std::move(p));
-        }
-
-        // Create a texture that has 4 mip levels. Each level is a different color.
-        // Level 0: 128x128 (red)
-        // Level 1:   64x64 (green)
-        // Level 2:   32x32 (blue)
-        // Level 3:   16x16 (yellow)
-        const size_t kTextureSize = 128;
-        const size_t kMipLevels = 4;
-        Handle<HwTexture> texture = api.createTexture(SamplerType::SAMPLER_2D, kMipLevels,
-                TextureFormat::RGBA8, 1, kTextureSize, kTextureSize, 1,
-                TextureUsage::SAMPLEABLE | TextureUsage::COLOR_ATTACHMENT | TextureUsage::UPLOADABLE);
-
-        // Create image data.
-        auto pixelFormat = PixelDataFormat::RGBA;
-        auto pixelType = PixelDataType::UBYTE;
-        size_t components; int bpp;
-        getPixelInfo(pixelFormat, pixelType, components, bpp);
-        uint32_t colors[] = {
-                0xFF0000FF, /* red */
-                0xFF00FF00, /* green */
-                0xFFFF0000, /* blue */
-                0xFF00FFFF, /* yellow */
-        };
-        for (int l = 0; l < kMipLevels; l++) {
-            size_t mipSize = kTextureSize >> l;
-            auto* buffer = (uint8_t*)calloc(1, mipSize * mipSize * bpp);
-            fillCheckerboard<uint32_t>(buffer, mipSize, mipSize, 1, colors[l]);
-            PixelBufferDescriptor descriptor(
-                    buffer, mipSize * mipSize * bpp, pixelFormat, pixelType, 1, 0, 0,
-                    mipSize, [](void* buffer, size_t size, void* user) { free(buffer); },
-                    nullptr);
-            api.update3DImage(
-                    texture, l, 0, 0, 0, mipSize, mipSize, 1, std::move(descriptor));
-        }
-
-        TrianglePrimitive triangle(api);
-
-        api.beginFrame(0, 0);
-
-        // We set the base mip to 1, and the max mip to 3
-        // Level 0: 128x128 (red)
-        // Level 1:   64x64 (green)             <-- base
-        // Level 2:   32x32 (blue)              <--- white triangle rendered
-        // Level 3:   16x16 (yellow)            <-- max
-        api.setMinMaxLevels(texture, 1, 3);
-
-        // Render a white triangle into level 2.
-        // We specify mip level 2, because minMaxLevels has no effect when rendering into a texture.
-        Handle<HwRenderTarget> renderTarget = api.createRenderTarget(
-                TargetBufferFlags::COLOR, 32, 32, 1,
-                {texture, 2 /* level */, 0 /* layer */}, {}, {});
-        {
-            RenderPassParams params = {};
-            fullViewport(params);
-            params.flags.clear = TargetBufferFlags::NONE;
-            params.flags.discardStart = TargetBufferFlags::NONE;
-            params.flags.discardEnd = TargetBufferFlags::NONE;
-            PipelineState ps = {};
-            ps.program = whiteProgram;
-            ps.rasterState.colorWrite = true;
-            ps.rasterState.depthWrite = false;
-            api.beginRenderPass(renderTarget, params);
-            api.draw(ps, triangle.getRenderPrimitive(), 1);
-            api.endRenderPass();
-        }
-
-        backend::Handle<HwRenderTarget> defaultRenderTarget = api.createDefaultRenderTarget(0);
-
-        RenderPassParams params = {};
-        fullViewport(params);
-        params.flags.clear = TargetBufferFlags::COLOR;
-        params.clearColor = {0.f, 0.f, 0.5f, 1.f};
-        params.flags.discardStart = TargetBufferFlags::ALL;
-        params.flags.discardEnd = TargetBufferFlags::NONE;
-
-        PipelineState state;
-        state.scissor = params.viewport;
-        state.program = textureProgram;
-        state.rasterState.colorWrite = true;
-        state.rasterState.depthWrite = false;
-        state.rasterState.depthFunc = SamplerCompareFunc::A;
-        state.rasterState.culling = CullingMode::NONE;
-
-        SamplerGroup samplers(1);
-        SamplerParams samplerParams {};
-        samplerParams.filterMag = SamplerMagFilter::NEAREST;
-        samplerParams.filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST;
-        samplers.setSampler(0, { texture, samplerParams });
-        backend::Handle<HwSamplerGroup> samplerGroup = api.createSamplerGroup(1);
-        api.updateSamplerGroup(samplerGroup, samplers.toBufferDescriptor(api));
-        api.bindSamplers(0, samplerGroup);
-
-        // Render a triangle to the screen, sampling from mip level 1.
-        // Because the min level is 1, the result color should be the white triangle drawn in the
-        // previous pass.
-        api.beginRenderPass(defaultRenderTarget, params);
-        api.draw(state, triangle.getRenderPrimitive(), 1);
-        api.endRenderPass();
-
-        // Adjust the base mip to 2.
-        // Note that this is done without another call to updateSamplerGroup.
-        api.setMinMaxLevels(texture, 2, 3);
-
-        // Render a second, smaller, triangle, again sampling from mip level 1.
-        // This triangle should be yellow striped.
-        static filament::math::float2 vertices[3] = {
-                { -0.5, -0.5 },
-                {  0.5, -0.5 },
-                { -0.5,  0.5 }
-        };
-        triangle.updateVertices(vertices);
-        params.flags.clear = TargetBufferFlags::NONE;
-        params.flags.discardStart = TargetBufferFlags::NONE;
-        api.beginRenderPass(defaultRenderTarget, params);
-        api.draw(state, triangle.getRenderPrimitive(), 1);
-        api.endRenderPass();
-
-        api.commit(swapChain);
-        api.endFrame(0);
-
-        api.stopCapture(0);
-
-        // Cleanup.
-        api.destroySwapChain(swapChain);
-        api.destroyRenderTarget(renderTarget);
-        api.destroyTexture(texture);
-        api.destroyProgram(whiteProgram);
-        api.destroyProgram(textureProgram);
-    }
-
-    api.finish();
-
-    executeCommands();
-    getDriver().purge();
-}
-
-} // namespace test
@@ -18,7 +18,6 @@

 #include "ShaderGenerator.h"
 #include "TrianglePrimitive.h"
-#include "BackendTestUtils.h"

 #include <utils/Hash.h>

@@ -25,7 +25,6 @@
 #include <utils/EntityInstance.h>

 #include <math/mathfwd.h>
-#include <math/quat.h>

 namespace utils {
    class Entity;
@@ -245,7 +244,6 @@ public:
         * shadows that are too far and wouldn't contribute to the scene much, improving
         * performance and quality. This value is always positive.
         * Use 0.0f to use the camera far distance.
-         * This only affect directional lights.
         */
        float shadowFar = 0.0f;

@@ -362,13 +360,6 @@ public:
         * enabled. (2cm by default).
         */
        float shadowBulbRadius = 0.02f;
-
-        /**
-         * Transforms the shadow direction. Must be a unit quaternion.
-         * The default is identity.
-         * Ignored if the light type isn't directional. For artistic use. Use with caution.
-         */
-        math::quatf transform{ 1.0f };
    };

    struct ShadowCascades {
@@ -680,7 +671,7 @@ public:
     * @return      true is this light is a type of directional light
     */
    inline bool isDirectional(Instance i) const noexcept {
-        Type const type = getType(i);
+        Type type = getType(i);
        return type == Type::DIRECTIONAL || type == Type::SUN;
    }

@@ -701,7 +692,7 @@ public:
     * @return      true is this light is a type of spot light
     */
    inline bool isSpotLight(Instance i) const noexcept {
-        Type const type = getType(i);
+        Type type = getType(i);
        return type == Type::SPOT || type == Type::FOCUSED_SPOT;
    }

@@ -141,17 +141,6 @@ struct BloomOptions {
    bool enabled = false;                   //!< enable or disable bloom
    float highlight = 1000.0f;              //!< limit highlights to this value before bloom [10, +inf]

-    /**
-     * Bloom quality level.
-     * LOW (default): use a more optimized down-sampling filter, however there can be artifacts
-     *      with dynamic resolution, this can be alleviated by using the homogenous mode.
-     * MEDIUM: Good balance between quality and performance.
-     * HIGH: In this mode the bloom resolution is automatically increased to avoid artifacts.
-     *      This mode can be significantly slower on mobile, especially at high resolution.
-     *      This mode greatly improves the anamorphic bloom.
-     */
-    QualityLevel quality = QualityLevel::LOW;
-
    bool lensFlare = false;                 //!< enable screen-space lens flare
    bool starburst = true;                  //!< enable starburst effect on lens flare
    float chromaticAberration = 0.005f;     //!< amount of chromatic aberration
@@ -285,7 +285,7 @@ void PerViewUniforms::prepareDirectionalLight(FEngine& engine,
        float exposure,
        float3 const& sceneSpaceDirection,
        PerViewUniforms::LightManagerInstance directionalLight) noexcept {
-    FLightManager const& lcm = engine.getLightManager();
+    FLightManager& lcm = engine.getLightManager();
    auto& s = mUniforms.edit();

    const float3 l = -sceneSpaceDirection; // guaranteed normalized
@@ -302,11 +302,11 @@ void PerViewUniforms::prepareDirectionalLight(FEngine& engine,
        // The last parameter must be < 0.0f for regular directional lights
        float4 sun{ 0.0f, 0.0f, 0.0f, -1.0f };
        if (UTILS_UNLIKELY(isSun && colorIntensity.w > 0.0f)) {
-            // Currently we have only a single directional light, so it's probably likely that it's
+            // currently we have only a single directional light, so it's probably likely that it's
            // also the Sun. However, conceptually, most directional lights won't be sun lights.
-            float const radius = lcm.getSunAngularRadius(directionalLight);
-            float const haloSize = lcm.getSunHaloSize(directionalLight);
-            float const haloFalloff = lcm.getSunHaloFalloff(directionalLight);
+            float radius = lcm.getSunAngularRadius(directionalLight);
+            float haloSize = lcm.getSunHaloSize(directionalLight);
+            float haloFalloff = lcm.getSunHaloFalloff(directionalLight);
            sun.x = std::cos(radius);
            sun.y = std::sin(radius);
            sun.z = 1.0f / (std::cos(radius * haloSize) - sun.x);
@@ -212,8 +212,6 @@ static const PostProcessManager::MaterialInfo sMaterialList[] = {
        { "bilateralBlurBentNormals",   MATERIAL(BILATERALBLURBENTNORMALS) },
        { "blitLow",                    MATERIAL(BLITLOW) },
        { "bloomDownsample",            MATERIAL(BLOOMDOWNSAMPLE) },
-        { "bloomDownsample2x",          MATERIAL(BLOOMDOWNSAMPLE2X) },
-        { "bloomDownsample9",           MATERIAL(BLOOMDOWNSAMPLE9) },
        { "bloomUpsample",              MATERIAL(BLOOMUPSAMPLE) },
        { "colorGrading",               MATERIAL(COLORGRADING) },
        { "colorGradingAsSubpass",      MATERIAL(COLORGRADINGASSUBPASS) },
@@ -452,6 +450,7 @@ PostProcessManager::StructurePassOutput PostProcessManager::structure(FrameGraph
                for (size_t level = 0; level < levelCount - 1; level++) {
                    auto out = resources.getRenderPassInfo(level);
                    driver.setMinMaxLevels(in, level, level);
+                    mi->setParameter("level", uint32_t(level));
                    commitAndRender(out, material, driver);
                }
                driver.setMinMaxLevels(in, 0, levelCount - 1);
@@ -1597,7 +1596,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::dof(FrameGraph& fg,
                    auto const& out = resources.getRenderPassInfo(data.rp[level]);
                    driver.setMinMaxLevels(inOutColor, level, level);
                    driver.setMinMaxLevels(inOutCoc, level, level);
-                    mi->setParameter("weightScale", 0.5f / float(1u << level));   // FIXME: halfres?
+                    mi->setParameter("mip", uint32_t(level));
+                    mi->setParameter("weightScale", 0.5f / float(1u<<level));   // FIXME: halfres?
                    mi->setParameter("texelSize", float2{ 1.0f / w, 1.0f / h });
                    mi->commit(driver);
                    render(out, pipeline, driver);
@@ -1870,53 +1870,15 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
    return bloomPass(fg, input, outFormat, inoutBloomOptions, scale);
 }

-FrameGraphId<FrameGraphTexture> PostProcessManager::downscalePass(FrameGraph& fg,
-        FrameGraphId<FrameGraphTexture> input,
-        FrameGraphTexture::Descriptor const& outDesc,
-        bool threshold, float highlight, bool fireflies) noexcept {
-    struct DownsampleData {
-        FrameGraphId<FrameGraphTexture> input;
-        FrameGraphId<FrameGraphTexture> output;
-    };
-    auto& downsamplePass = fg.addPass<DownsampleData>("Downsample",
-            [&](FrameGraph::Builder& builder, auto& data) {
-                data.input = builder.sample(input);
-                data.output = builder.createTexture("Downsample-output", outDesc);
-                builder.declareRenderPass(data.output);
-            },
-            [=](FrameGraphResources const& resources,
-                    auto const& data, DriverApi& driver) {
-                auto const& material = getPostProcessMaterial("bloomDownsample2x");
-                auto* mi = material.getMaterialInstance(mEngine);
-                mi->setParameter("source", resources.getTexture(data.input), {
-                        .filterMag = SamplerMagFilter::LINEAR,
-                        .filterMin = SamplerMinFilter::LINEAR
-                });
-                mi->setParameter("level", 0);
-                mi->setParameter("threshold", threshold ? 1.0f : 0.0f);
-                mi->setParameter("fireflies", fireflies ? 1.0f : 0.0f);
-                mi->setParameter("invHighlight", std::isinf(highlight) ? 0.0f : 1.0f / highlight);
-                commitAndRender(resources.getRenderPassInfo(), material, driver);
-            });
-    return downsamplePass->output;
-}
-
 PostProcessManager::BloomPassOutput PostProcessManager::bloomPass(FrameGraph& fg,
        FrameGraphId<FrameGraphTexture> input, TextureFormat outFormat,
        BloomOptions& inoutBloomOptions, float2 scale) noexcept {
-
-    // Figure out a good size for the bloom buffer. We must use a fixed bloom buffer size so
-    // that the size/strength of the bloom doesn't vary much with the resolution, otherwise
-    // dynamic resolution would affect the bloom effect too much.
-    auto desc = fg.getDescriptor(input);
+    // Figure out a good size for the bloom buffer.
+    auto const& desc = fg.getDescriptor(input);

    // width and height after dynamic resolution upscaling
    const float aspect = (float(desc.width) * scale.y) / (float(desc.height) * scale.x);

-    // FIXME: don't allow inoutBloomOptions.resolution to be larger than input's resolution
-    //        (avoid upscale) but how does this affect dynamic resolution
-    // FIXME: check what happens on WebGL and intel's processors
-
    // compute the desired bloom buffer size
    float bloomHeight = float(inoutBloomOptions.resolution);
    float bloomWidth  = bloomHeight * aspect;
@@ -1931,73 +1893,38 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloomPass(FrameGraph& fg
        bloomHeight *= inoutBloomOptions.anamorphism;
    }

+    // convert back to integer width/height
+    const uint32_t width  = std::max(1u, uint32_t(std::floor(bloomWidth)));
+    const uint32_t height = std::max(1u, uint32_t(std::floor(bloomHeight)));
+
    // we might need to adjust the max # of levels
    const uint32_t major = uint32_t(std::max(bloomWidth,  bloomHeight));
    const uint8_t maxLevels = FTexture::maxLevelCount(major);
    inoutBloomOptions.levels = std::min(inoutBloomOptions.levels, maxLevels);
    inoutBloomOptions.levels = std::min(inoutBloomOptions.levels, kMaxBloomLevels);

-    if (inoutBloomOptions.quality == QualityLevel::LOW) {
-        // In low quality mode, we adjust the bloom buffer size so that both dimensions
-        // have enough exact mip levels. This can slightly affect the aspect ratio causing
-        // some artifacts:
-        // - add some anamorphism (experimentally not visible)
-        // - visible bloom size changes with dynamic resolution in non-homogenous mode
-        // This allows us to use the 9 sample downsampling filter (instead of 13)
-        // for at least 4 levels.
-        uint32_t width  = std::max(1u, uint32_t(std::floor(bloomWidth)));
-        uint32_t height = std::max(1u, uint32_t(std::floor(bloomHeight)));
-        width  &= ~((1 << 4) - 1);  // at least 4 levels
-        height &= ~((1 << 4) - 1);
-        bloomWidth  = float(width);
-        bloomHeight = float(height);
+    if (2 * width < desc.width || 2 * height < desc.height) {
+        // if we're scaling down by more than 2x, prescale the image with a blit to improve
+        // performance. This is important on mobile/tilers.
+        input = opaqueBlit(fg, input, { 0, 0, desc.width, desc.height }, {
+                .width = std::max(1u, desc.width / 2),
+                .height = std::max(1u, desc.height / 2),
+                .format = outFormat
+        });
    }

-    bool threshold = inoutBloomOptions.threshold;
-
-    while (2 * bloomWidth < float(desc.width) || 2 * bloomHeight < float(desc.height)) {
-        if (inoutBloomOptions.quality == QualityLevel::LOW ||
-            inoutBloomOptions.quality == QualityLevel::MEDIUM) {
-            input = downscalePass(fg, input, {
-                            .width  = (desc.width  = std::max(1u, desc.width  / 2)),
-                            .height = (desc.height = std::max(1u, desc.height / 2)),
-                            .format = outFormat
-                    },
-                    threshold, inoutBloomOptions.highlight, threshold);
-            threshold = false; // we do the thresholding only once during down sampling
-        } else if (inoutBloomOptions.quality == QualityLevel::HIGH ||
-                   inoutBloomOptions.quality == QualityLevel::ULTRA) {
-            // In high quality mode, we increase the size of the bloom buffer such that the
-            // first scaling is less than 2x, and we increase the number of levels accordingly.
-            if (bloomWidth * 2.0f > 2048.0f || bloomHeight * 2.0f > 2048.0f) {
-                // but we can't scale above the h/w guaranteed minspec
-                break;
-            }
-            bloomWidth  *= 2.0f;
-            bloomHeight *= 2.0f;
-            inoutBloomOptions.levels++;
-        }
-    }
-
-    // convert back to integer width/height
-    uint32_t const width  = std::max(1u, uint32_t(std::floor(bloomWidth)));
-    uint32_t const height = std::max(1u, uint32_t(std::floor(bloomHeight)));
-
-    input = downscalePass(fg, input,
-            { .width = width, .height = height, .format = outFormat },
-            threshold, inoutBloomOptions.highlight, threshold);
-
    struct BloomPassData {
+        FrameGraphId<FrameGraphTexture> in;
        FrameGraphId<FrameGraphTexture> out;
+        FrameGraphId<FrameGraphTexture> stage;
        uint32_t outRT[kMaxBloomLevels];
+        uint32_t stageRT[kMaxBloomLevels];
    };

-    // Creating a mip-chain poses a "feedback" loop problem on some GPU. We will disable
-    // Bloom on these.
-    // See: https://github.com/google/filament/issues/2338
-
+    // downsample phase
    auto& bloomDownsamplePass = fg.addPass<BloomPassData>("Bloom Downsample",
            [&](FrameGraph::Builder& builder, auto& data) {
+                data.in = builder.sample(input);
                data.out = builder.createTexture("Bloom Out Texture", {
                        .width = width,
                        .height = height,
@@ -2006,107 +1933,165 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloomPass(FrameGraph& fg
                });
                data.out = builder.sample(data.out);

+                data.stage = builder.createTexture("Bloom Stage Texture", {
+                        .width = width,
+                        .height = height,
+                        .levels = inoutBloomOptions.levels,
+                        .format = outFormat
+                });
+                data.stage = builder.sample(data.stage);
+
                for (size_t i = 0; i < inoutBloomOptions.levels; i++) {
                    auto out = builder.createSubresource(data.out, "Bloom Out Texture mip",
                            { .level = uint8_t(i) });
-                    if (i == 0) {
-                        // this causes the last blit above to render into this mip
-                       fg.forwardResource(out, input);
-                    }
+                    auto stage = builder.createSubresource(data.stage,
+                            "Bloom Stage Texture mip", { .level = uint8_t(i) });
                    builder.declareRenderPass(out, &data.outRT[i]);
+                    builder.declareRenderPass(stage, &data.stageRT[i]);
                }
            },
            [=](FrameGraphResources const& resources,
                    auto const& data, DriverApi& driver) {

-                // TODO: if downsampling is not exactly a multiple of two, use the 13 samples
-                //       filter. This is generally the accepted solution, however, the 13 samples
-                //       filter is not correct either when we don't sample at integer coordinates,
-                //       but it seems ot create less artifacts.
-                //       A better solution might be to use the filter described in
-                //       Castaño, 2013, "Shadow Mapping Summary Part 1", which is 5x5 filter with
-                //       9 samples, but works at all coordinates.
-
+                auto hwIn = resources.getTexture(data.in);
                auto hwOut = resources.getTexture(data.out);
+                auto hwStage = resources.getTexture(data.stage);

-                auto const& material9 = getPostProcessMaterial("bloomDownsample9");
-                auto const& material13 = getPostProcessMaterial("bloomDownsample");
+                auto const& material = getPostProcessMaterial("bloomDownsample");
+                auto const* ma = material.getMaterial(mEngine);

-                auto* mi9 = material9.getMaterialInstance(mEngine);
-                auto* mi13 = material13.getMaterialInstance(mEngine);
+                FMaterialInstance* mis[] = {
+                        ma->createInstance("bloomDownsample-ping"),
+                        ma->createInstance("bloomDownsample-pong"),
+                        ma->createInstance("bloomDownsample-first"),
+                };

-                mi9->setParameter("source", hwOut, {
+                mis[0]->setParameter("source", hwOut, {
                        .filterMag = SamplerMagFilter::LINEAR,
-                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST });
+                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
+                });

-                mi13->setParameter("source", hwOut, {
+                mis[1]->setParameter("source", hwStage, {
                        .filterMag = SamplerMagFilter::LINEAR,
-                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST });
+                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
+                });
+                mis[2]->setParameter("source", hwIn, {
+                        .filterMag = SamplerMagFilter::LINEAR,
+                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
+                });

-                mi9->commit(driver);
-                mi13->commit(driver);
+                for (auto* mi : mis) {
+                    mi->setParameter("level", 0.0f);
+                    mi->setParameter("threshold", inoutBloomOptions.threshold ? 1.0f : 0.0f);
+                    mi->setParameter("invHighlight", std::isinf(inoutBloomOptions.highlight)
+                            ? 0.0f : 1.0f / inoutBloomOptions.highlight);
+                    mi->commit(driver);
+                }

-                // PipelineState for both materials should be the same
-                const PipelineState pipeline(material9.getPipelineState(mEngine));
+                const PipelineState pipeline(material.getPipelineState(mEngine));

-                for (size_t i = 1; i < inoutBloomOptions.levels; i++) {
-                    auto hwDstRT = resources.getRenderPassInfo(data.outRT[i]);
+                { // first iteration
+                    auto hwDstRT = resources.getRenderPassInfo(data.outRT[0]);
                    hwDstRT.params.flags.discardStart = TargetBufferFlags::COLOR;
                    hwDstRT.params.flags.discardEnd = TargetBufferFlags::NONE;
-
-                    // if downsampling is a multiple of 2 in each dimension we can use the
-                    // 9 samples filter.
-                    auto vp = resources.getRenderPassInfo(data.outRT[i-1]).params.viewport;
-                    auto* const mi = (vp.width & 1 || vp.height & 1) ? mi13 : mi9;
-                    mi->use(driver);
-                    driver.setMinMaxLevels(hwOut, i - 1, i - 1); // this offsets baseLevel to i-1
+                    mis[2]->use(driver);
                    render(hwDstRT, pipeline, driver);
                }
-                driver.setMinMaxLevels(hwOut, 0, inoutBloomOptions.levels - 1);
+
+                for (size_t i = 1; i < inoutBloomOptions.levels; i++) {
+                    const size_t parity = 1u - (i & 0x1u);
+                    auto hwDstRT = resources.getRenderPassInfo(parity ? data.outRT[i] : data.stageRT[i]);
+                    hwDstRT.params.flags.discardStart = TargetBufferFlags::COLOR;
+                    hwDstRT.params.flags.discardEnd = TargetBufferFlags::NONE;
+                    mis[parity]->setParameter("level", float(i - 1));
+                    mis[parity]->commit(driver);
+                    mis[parity]->use(driver);
+                    render(hwDstRT, pipeline, driver);
+                }
+
+                for (auto& mi : mis) {
+                    mEngine.destroy(mi);
+                }
            });

-    // output of bloom downsample pass becomes input of next (flare) pass
-    input = bloomDownsamplePass->out;
+    FrameGraphId<FrameGraphTexture> output = bloomDownsamplePass->out;
+    FrameGraphId<FrameGraphTexture> stage = bloomDownsamplePass->stage;

    // flare pass
-    auto flare = flarePass(fg, input, width, height, outFormat, inoutBloomOptions);
+    auto flare = flarePass(fg, bloomDownsamplePass->out, width, height, outFormat, inoutBloomOptions);

+    // upsample phase
    auto& bloomUpsamplePass = fg.addPass<BloomPassData>("Bloom Upsample",
            [&](FrameGraph::Builder& builder, auto& data) {
-                data.out = builder.sample(input);
+                data.out = builder.sample(output);
+                data.stage = builder.sample(stage);
                for (size_t i = 0; i < inoutBloomOptions.levels; i++) {
                    auto out = builder.createSubresource(data.out, "Bloom Out Texture mip",
                            { .level = uint8_t(i) });
+                    auto staging = builder.createSubresource(data.stage,
+                            "Bloom Stage Texture mip", { .level = uint8_t(i) });
                    builder.declareRenderPass(out, &data.outRT[i]);
+                    builder.declareRenderPass(staging, &data.stageRT[i]);
                }
            },
            [=](FrameGraphResources const& resources, auto const& data, DriverApi& driver) {
+
                auto hwOut = resources.getTexture(data.out);
+                auto hwStage = resources.getTexture(data.stage);
                auto const& outDesc = resources.getDescriptor(data.out);

                auto const& material = getPostProcessMaterial("bloomUpsample");
-                auto* mi = material.getMaterialInstance(mEngine);
-                mi->setParameter("source", hwOut, {
+                auto const* ma = material.getMaterial(mEngine);
+
+                FMaterialInstance* mis[] = {
+                        ma->createInstance("bloomUpsample-ping"),
+                        ma->createInstance("bloomUpsample-pong"),
+                };
+
+                mis[0]->setParameter("source", hwOut, {
                        .filterMag = SamplerMagFilter::LINEAR,
-                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST});
-                mi->use(driver);
+                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
+                });
+
+                mis[1]->setParameter("source", hwStage, {
+                        .filterMag = SamplerMagFilter::LINEAR,
+                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
+                });

                PipelineState pipeline(material.getPipelineState(mEngine));
                pipeline.rasterState.blendFunctionSrcRGB = BlendFunction::ONE;
                pipeline.rasterState.blendFunctionDstRGB = BlendFunction::ONE;

                for (size_t j = inoutBloomOptions.levels, i = j - 1; i >= 1; i--, j++) {
-                    auto hwDstRT = resources.getRenderPassInfo(data.outRT[i - 1]);
+                    const size_t parity = 1u - (j % 2u);
+
+                    auto hwDstRT = resources.getRenderPassInfo(
+                            parity ? data.outRT[i - 1] : data.stageRT[i - 1]);
                    hwDstRT.params.flags.discardStart = TargetBufferFlags::NONE; // b/c we'll blend
                    hwDstRT.params.flags.discardEnd = TargetBufferFlags::NONE;
+
                    auto w = FTexture::valueForLevel(i - 1, outDesc.width);
                    auto h = FTexture::valueForLevel(i - 1, outDesc.height);
-                    mi->setParameter("resolution", float4{ w, h, 1.0f / w, 1.0f / h });
-                    mi->commit(driver);
-                    driver.setMinMaxLevels(hwOut, i, i); // this offsets baseLevel to i
+                    mis[parity]->setParameter("resolution", float4{ w, h, 1.0f / w, 1.0f / h });
+                    mis[parity]->setParameter("level", float(i));
+                    mis[parity]->commit(driver);
+                    mis[parity]->use(driver);
                    render(hwDstRT, pipeline, driver);
                }
-                driver.setMinMaxLevels(hwOut, 0, inoutBloomOptions.levels - 1);
+
+                for (auto& mi : mis) {
+                    mEngine.destroy(mi);
+                }
+
+                // Every other level is missing from the out texture, so we need to do
+                // blits to complete the chain.
+                const SamplerMagFilter filter = SamplerMagFilter::NEAREST;
+                for (size_t i = 1; i < inoutBloomOptions.levels; i += 2) {
+                    auto in = resources.getRenderPassInfo(data.stageRT[i]);
+                    auto out = resources.getRenderPassInfo(data.outRT[i]);
+                    driver.blit(TargetBufferFlags::COLOR, out.target, out.params.viewport,
+                            in.target, in.params.viewport, filter);
+                }
            });

    return { bloomUpsamplePass->out, flare };
@@ -3051,6 +3036,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::vsmMipmapPass(FrameGraph& fg
                        .filterMag = SamplerMagFilter::LINEAR,
                        .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
                });
+                mi->setParameter("level", uint32_t(level));
                mi->setParameter("layer", uint32_t(layer));
                mi->setParameter("uvscale", 1.0f / float(dim));
                mi->commit(driver);
@@ -292,11 +292,6 @@ private:
            FrameGraphId<FrameGraphTexture> input, backend::TextureFormat outFormat,
            BloomOptions& inoutBloomOptions, math::float2 scale) noexcept;

-    FrameGraphId<FrameGraphTexture> downscalePass(FrameGraph& fg,
-            FrameGraphId<FrameGraphTexture> input,
-            FrameGraphTexture::Descriptor const& outDesc,
-            bool threshold, float highlight, bool fireflies) noexcept;
-
    void commitAndRender(FrameGraphResources::RenderPassInfo const& out,
            PostProcessMaterial const& material, uint8_t variant,
            backend::DriverApi& driver) const noexcept;
@@ -110,7 +110,7 @@ math::mat4f ShadowMap::getPointLightViewMatrix(backend::TextureCubemapFace face,
 }

 ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,
-        FScene::LightSoa const& lightData, size_t index,
+        const FScene::LightSoa& lightData, size_t index,
        filament::CameraInfo const& camera,
        ShadowMapInfo const& shadowMapInfo,
        SceneInfo const& sceneInfo) noexcept {
@@ -131,7 +131,20 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,
    else            params.options.shadowFarHint = dzf * dz + camera.zf;
 #endif

-    const auto direction = params.options.transform * lightData.elementAt<FScene::DIRECTION>(index);
+    // Adjust the camera's projection for the light's shadowFar
+    const mat4f cullingProjection{ [&](auto p) {
+        if (params.options.shadowFar > 0.0f) {
+            float const n = camera.zn;
+            float const f = params.options.shadowFar;
+            // orthographic projection
+            assert_invariant(std::abs(p[2].w) <= std::numeric_limits<float>::epsilon());
+            p[2].z =    2.0f / (n - f);
+            p[3].z = (f + n) / (n - f);
+        }
+        return p;
+    }(camera.cullingProjection) };
+
+    const auto direction = lightData.elementAt<FScene::DIRECTION>(index);

    /*
     * Compute the light's model matrix
@@ -151,7 +164,7 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,

    // view frustum vertices in world-space
    float3 wsViewFrustumVertices[8];
-    const mat4f worldToClipMatrix = camera.cullingProjection * camera.view;
+    const mat4f worldToClipMatrix = cullingProjection * camera.view;
    const Frustum wsFrustum(worldToClipMatrix);
    computeFrustumCorners(wsViewFrustumVertices, inverse(worldToClipMatrix), sceneInfo.csNearFar);

@@ -203,7 +216,7 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,
    }

    // Now that we know the znear (-lsLightFrustumBounds.max.z), adjust the light's position such
-    // that znear = 0, this is only needed for VSM, but doesn't hurt PCF.
+    // that znear = 0, this is only need for VSM, but doesn't hurt PCF.
    const mat4f Mv = getDirectionalLightViewMatrix(direction, direction * -lsLightFrustumBounds.max.z);

    // near / far planes are specified relative to the direction the eye is looking at
@@ -227,11 +240,8 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,
        const float4 shadowReceiverVolumeBoundingSphere = computeBoundingSphere(
                wsShadowReceiversVolume.getCorners().data(), 8);

-        // in stable mode we simply take the view volume bounding sphere, but we calculate it
-        // in view space, so that it's perfectly stable.
-        float3 vertices[8];
-        computeFrustumCorners(vertices, inverse(camera.cullingProjection), sceneInfo.csNearFar);
-        viewVolumeBoundingSphere = computeBoundingSphere(vertices, 8);
+        // in stable mode we simply take the view volume, bounding sphere
+        viewVolumeBoundingSphere = computeBoundingSphere(wsViewFrustumVertices, 8);

        if (shadowReceiverVolumeBoundingSphere.w < viewVolumeBoundingSphere.w) {

@@ -310,66 +320,45 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,
    //
    //   In LiPSM mode, we're using the warped space here.

-    float2 s, o;
+    Aabb bounds;
+    if (params.options.stable && viewVolumeBoundingSphere.w > 0) {
+        bounds = compute2DBounds(Mv, viewVolumeBoundingSphere);
+    } else {
+        bounds = compute2DBounds(WLMpMv, wsClippedShadowReceiverVolume.data(), vertexCount);
+    }
+    lsLightFrustumBounds.min.xy = bounds.min.xy;
+    lsLightFrustumBounds.max.xy = bounds.max.xy;
+
    if (params.options.stable) {
-        if (viewVolumeBoundingSphere.w > 0) {
-            s = 1.0f / viewVolumeBoundingSphere.w;
-            o = mat4f::project(LMpMv * camera.model, viewVolumeBoundingSphere.xyz).xy;
-        } else {
-            Aabb const bounds = compute2DBounds(LMpMv,
-                    wsClippedShadowReceiverVolume.data(), vertexCount);
-            if (UTILS_UNLIKELY((bounds.min.x >= bounds.max.x) || (bounds.min.y >= bounds.max.y))) {
-                // this could happen if the only thing visible is a perfectly horizontal or
-                // vertical thin line
-                mHasVisibleShadows = false;
-                return {};
-            }
-            assert_invariant(bounds.min.x < bounds.max.x);
-            assert_invariant(bounds.min.y < bounds.max.y);
-
-            s = 2.0f / float2(bounds.max.xy - bounds.min.xy);
-            o = float2(bounds.max.xy + bounds.min.xy) * 0.5f;
-
-            // Quantize the scale in world-space units. This value can be very small because
-            // if it wasn't for floating-point imprecision, the scale would be a constant.
-            double2 const quantizer = 0.0625;
-            s = 1.0 / (ceil(1.0 / (s * quantizer)) * quantizer);
-        }
+        // in stable mode we can't do anything that can change the scaling of the texture
    } else {
-        Aabb const bounds = compute2DBounds(WLMpMv,
-                wsClippedShadowReceiverVolume.data(), vertexCount);
-        lsLightFrustumBounds.min.xy = bounds.min.xy;
-        lsLightFrustumBounds.max.xy = bounds.max.xy;
        // For directional lights, we further constraint the light frustum to the
        // intersection of the shadow casters & shadow receivers in light-space.
        // ** This relies on the 1-texel shadow map border **
-
        if (engine.debug.shadowmap.focus_shadowcasters) {
-            intersectWithShadowCasters(&lsLightFrustumBounds, WLMpMv, wsShadowCastersVolume);
+            intersectWithShadowCasters(lsLightFrustumBounds, WLMpMv, wsShadowCastersVolume);
        }
-        if (UTILS_UNLIKELY((lsLightFrustumBounds.min.x >= lsLightFrustumBounds.max.x) ||
-                           (lsLightFrustumBounds.min.y >= lsLightFrustumBounds.max.y))) {
-            // this could happen if the only thing visible is a perfectly horizontal or
-            // vertical thin line
-            mHasVisibleShadows = false;
-            return {};
-        }
-        assert_invariant(lsLightFrustumBounds.min.x < lsLightFrustumBounds.max.x);
-        assert_invariant(lsLightFrustumBounds.min.y < lsLightFrustumBounds.max.y);
-
-        s = 2.0f / float2(lsLightFrustumBounds.max.xy - lsLightFrustumBounds.min.xy);
-        o = float2(lsLightFrustumBounds.max.xy + lsLightFrustumBounds.min.xy) * 0.5f;
-
-        // TODO: we could quantize `s` here to give some stability when lispsm is disabled,
-        //       however, the quantization paramater should probably be user settable.
    }

-    // adjust offset for scale
-    o = -s * o;
+    if (UTILS_UNLIKELY((lsLightFrustumBounds.min.x >= lsLightFrustumBounds.max.x) ||
+                       (lsLightFrustumBounds.min.y >= lsLightFrustumBounds.max.y))) {
+        // this could happen if the only thing visible is a perfectly horizontal or
+        // vertical thin line
+        mHasVisibleShadows = false;
+        return {};
+    }
+
+    assert_invariant(lsLightFrustumBounds.min.x < lsLightFrustumBounds.max.x);
+    assert_invariant(lsLightFrustumBounds.min.y < lsLightFrustumBounds.max.y);
+
+    // compute focus scale and offset
+    float2 s = 2.0f / float2(lsLightFrustumBounds.max.xy - lsLightFrustumBounds.min.xy);
+    float2 o =   -s * float2(lsLightFrustumBounds.max.xy + lsLightFrustumBounds.min.xy) * 0.5f;

    if (params.options.stable) {
-        snapLightFrustum(s, o, LMpMv,
-                sceneInfo.wsShadowCastersVolume.center(), shadowMapInfo.shadowDimension);
+        // Use the world origin as reference point, fixed w.r.t. the camera
+        snapLightFrustum(s, o, Mv, camera.worldOrigin[3].xyz,
+                1.0f / float(shadowMapInfo.shadowDimension));
    }

    const mat4f F(mat4f::row_major_init {
@@ -388,7 +377,8 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,

    // Computes St the transform to use in the shader to access the shadow map texture
    // i.e. it transforms a world-space vertex to a texture coordinate in the shadowmap
-    const auto [Mt, Mn] = ShadowMap::getTextureCoordsMapping(shadowMapInfo, getViewport());
+    const backend::Viewport viewport = getViewport();
+    const auto [Mt, Mn] = ShadowMap::getTextureCoordsMapping(shadowMapInfo, viewport);
    const mat4f St = math::highPrecisionMultiply(Mt, S);

    ShadowMap::ShaderParameters shaderParameters;
@@ -410,7 +400,7 @@ ShadowMap::ShaderParameters ShadowMap::updateDirectional(FEngine& engine,
        shaderParameters.lightSpace = computeVsmLightSpaceMatrix(St, Mv, znear, zfar);
    }

-    shaderParameters.scissorNormalized = getClampToEdgeCoords(shadowMapInfo);
+    shaderParameters.scissorNormalized = getViewportNormalized(shadowMapInfo);

    // We apply the constant bias in world space (as opposed to light-space) to account
    // for perspective and lispsm shadow maps. This also allows us to do this at zero-cost
@@ -443,8 +433,9 @@ ShadowMap::ShaderParameters ShadowMap::updatePunctual(
    assert_invariant(shadowMapInfo.textureDimension == mOptions->mapSize);

    // Final shadow transform
+    const backend::Viewport viewport = getViewport();
    const mat4f S = math::highPrecisionMultiply(Mp, Mv);
-    const auto [Mt, Mn] = ShadowMap::getTextureCoordsMapping(shadowMapInfo, getViewport());
+    const auto [Mt, Mn] = ShadowMap::getTextureCoordsMapping(shadowMapInfo, viewport);
    const mat4f St = math::highPrecisionMultiply(Mt, S);

    // TODO: focus projection
@@ -471,7 +462,7 @@ ShadowMap::ShaderParameters ShadowMap::updatePunctual(
        shaderParameters.lightSpace = computeVsmLightSpaceMatrix(St, Mv, nearPlane, farPlane);
    }

-    shaderParameters.scissorNormalized = getClampToEdgeCoords(shadowMapInfo);
+    shaderParameters.scissorNormalized = getViewportNormalized(shadowMapInfo);

    const float3 direction = -transpose(Mv)[2].xyz;
    const float constantBias = shadowMapInfo.vsm ? 0.0f : params.options.constantBias;
@@ -768,12 +759,20 @@ Aabb ShadowMap::compute2DBounds(const mat4f& lightView,
    return bounds;
 }

-void ShadowMap::intersectWithShadowCasters(Aabb* UTILS_RESTRICT lightFrustum,
+Aabb ShadowMap::compute2DBounds(const mat4f& lightView, float4 const& sphere) noexcept {
+    // this assumes a rigid body transform
+    float4 s;
+    s.xyz = (lightView * float4{sphere.xyz, 1.0f}).xyz;
+    s.w = sphere.w;
+    return Aabb{s.xyz - s.w, s.xyz + s.w};
+}
+
+void ShadowMap::intersectWithShadowCasters(Aabb& UTILS_RESTRICT lightFrustum,
        mat4f const& lightView, Aabb const& wsShadowCastersVolume) noexcept {

    // construct the Focus transform (scale + offset)
-    const float2 s = 2.0f / float2(lightFrustum->max.xy - lightFrustum->min.xy);
-    const float2 o =   -s * float2(lightFrustum->max.xy + lightFrustum->min.xy) * 0.5f;
+    const float2 s = 2.0f / float2(lightFrustum.max.xy - lightFrustum.min.xy);
+    const float2 o =   -s * float2(lightFrustum.max.xy + lightFrustum.min.xy) * 0.5f;
    const mat4f F(mat4f::row_major_init {
            s.x,  0.0f, 0.0f,    o.x,
            0.0f,  s.y, 0.0f,    o.y,
@@ -799,8 +798,8 @@ void ShadowMap::intersectWithShadowCasters(Aabb* UTILS_RESTRICT lightFrustum,
    Aabb const box = compute2DBounds(lightView, wsClippedShadowCasterVolumeVertices.data(), vertexCount);

    // intersect shadow-caster and current light frustum bounds
-    lightFrustum->min.xy = max(box.min.xy, lightFrustum->min.xy);
-    lightFrustum->max.xy = min(box.max.xy, lightFrustum->max.xy);
+    lightFrustum.min.xy = max(box.min.xy, lightFrustum.min.xy);
+    lightFrustum.max.xy = min(box.max.xy, lightFrustum.max.xy);
 }

 void ShadowMap::computeFrustumCorners(float3* UTILS_RESTRICT out,
@@ -826,39 +825,21 @@ void ShadowMap::computeFrustumCorners(float3* UTILS_RESTRICT out,
 }

 void ShadowMap::snapLightFrustum(float2& s, float2& o,
-        mat4f const& Mv, double3 wsSnapCoords, int2 resolution) noexcept {
+        mat4f const& Mv, float3 worldOrigin, float2 shadowMapResolution) noexcept {

-    auto proj = [](mat4 m, double3 v) -> double3 {
-        // for directional light p.w == 1, exactly
-        auto p = m * v;
-        assert_invariant(p.w == 1.0);
-        return p.xyz;
+    auto fmod = [](float2 x, float2 y) -> float2 {
+        auto mod = [](float x, float y) -> float { return std::fmod(x, y); };
+        return float2{ mod(x[0], y[0]), mod(x[1], y[1]) };
    };

-    auto fract = [](auto v) {
-        using namespace std;
-        using T = decltype(v);
-        return fmod(v, T{1});
-    };
-
-    const mat4 F(mat4::row_major_init {
-            s.x, 0.0, 0.0, o.x,
-            0.0, s.y, 0.0, o.y,
-            0.0, 0.0, 1.0, 0.0,
-            0.0, 0.0, 0.0, 1.0,
-    });
-
-    // The (resolution * 0.5) comes from Mv having a NDC in the range -1,1 (so a range of 2).
-
-    // focused light-space
-    mat4 const FMv{ F * Mv };
+    // This snaps the shadow map bounds to texels.
+    // The 2.0 comes from Mv having a NDC in the range -1,1 (so a range of 2).
+    const float2 r = 2.0f * shadowMapResolution;
+    o -= fmod(o, r);

    // This offsets the texture coordinates, so it has a fixed offset w.r.t the world
-    double2 const lsOrigin = proj(FMv, wsSnapCoords).xy;
-    double2 const d = (fract(lsOrigin * resolution * 0.5) * 2.0) / resolution;
-
-    // adjust offset
-    o -= d;
+    const float2 lsOrigin = mat4f::project(Mv, worldOrigin).xy * s;
+    o -= fmod(lsOrigin, r);
 }

 size_t ShadowMap::intersectFrustumWithBox(
@@ -1061,7 +1042,7 @@ bool ShadowMap::intersectSegmentWithPlanarQuad(float3& UTILS_RESTRICT p,
 }

 float ShadowMap::texelSizeWorldSpace(const mat3f& worldToShadowTexture,
-        uint16_t shadowDimension) noexcept {
+        uint16_t shadowDimension) const noexcept {
    // The Jacobian of the transformation from texture-to-world is the matrix itself for
    // orthographic projections. We just need to inverse worldToShadowTexture,
    // which is guaranteed to be orthographic.
@@ -1076,7 +1057,7 @@ float ShadowMap::texelSizeWorldSpace(const mat3f& worldToShadowTexture,
 }

 float ShadowMap::texelSizeWorldSpace(const mat4f& Wp, const mat4f& MbMtF,
-        uint16_t shadowDimension) noexcept {
+        uint16_t shadowDimension) const noexcept {
    // Here we compute the Jacobian of inverse(MbMtF * Wp).
    // The expression below has been computed with Mathematica. However, it's not very hard,
    // albeit error-prone, to do it by hand because MbMtF is a linear transform.
@@ -1097,7 +1078,7 @@ float ShadowMap::texelSizeWorldSpace(const mat4f& Wp, const mat4f& MbMtF,

    constexpr bool JACOBIAN_ESTIMATE = false;
    if constexpr (JACOBIAN_ESTIMATE) {
-        // This estimates the Jacobian -- this is a lot heavier. This is mostly for reference
+        // this estimates the Jacobian -- this is a lot heavier. This is mostly for reference
        // and testing.
        const mat4f Si(inverse(MbMtF * Wp));
        const float3 p0 = mat4f::project(Si, p);
@@ -1236,58 +1217,33 @@ void ShadowMap::updateSceneInfoSpot(mat4f const& Mv, FScene const& scene,
 }

 backend::Viewport ShadowMap::getViewport() const noexcept {
-    // We set a viewport with a 1-texel border for when we index outside the texture.
-    // This happens only for directional lights when "focus shadow casters" is used,
-    // or when shadowFar is smaller than the camera far.
-    // For spot- and point-lights we also use a 1-texel border, so that bilinear filtering
-    // can work properly if the shadowmap is in an atlas (and we can't rely on h/w clamp).
+    // We set a viewport with a 1-texel border for when we index outside the
+    // texture. This can only happen for the directional light when "focus shadow casters is used".
    const uint32_t dim = mOptions->mapSize;
    const uint16_t border = 1u;
    return { border, border, dim - 2u * border, dim - 2u * border };
 }

 backend::Viewport ShadowMap::getScissor() const noexcept {
-    // We set a viewport with a 1-texel border for when we index outside the texture.
-    // This happens only for directional lights when "focus shadow casters" is used,
-    // or when shadowFar is smaller than the camera far.
-    // For spot- and point-lights we also use a 1-texel border, so that bilinear filtering
-    // can work properly if the shadowmap is in an atlas (and we can't rely on h/w clamp), so we
-    // don't scissor the border, so it gets filled with correct neighboring texels.
+    // We set a viewport with a 1-texel border for when we index outside the
+    // texture. This can only happen for the directional light when "focus shadow casters is used".
    const uint32_t dim = mOptions->mapSize;
    const uint16_t border = 1u;
+
    switch (mShadowType) {
        case ShadowType::DIRECTIONAL:
            return { border, border, dim - 2u * border, dim - 2u * border };
        case ShadowType::SPOT:
        case ShadowType::POINT:
+        default:
            return { 0, 0, dim, dim };
    }
 }

-math::float4 ShadowMap::getClampToEdgeCoords(ShadowMapInfo const& shadowMapInfo) const noexcept {
-    float border; // shadowmap border in texels
-    switch (mShadowType) {
-        case ShadowType::DIRECTIONAL:
-            // For directional lights, we need to allow the sampling to reach the border, it
-            // happens when "focus shadow casters" is used for instance.
-            border = 0.5f;
-            break;
-        case ShadowType::SPOT:
-        case ShadowType::POINT:
-            // For spot and point light, this is equal to the viewport. i.e. the valid
-            // texels are inside the viewport (w/ 1-texel border), the border will be used
-            // for bilinear filtering.
-            border = 1.0f;
-            break;
-    }
-
-    float const texel = 1.0f / float(shadowMapInfo.atlasDimension);
-    float const dim = float(mOptions->mapSize);
-    float const l = border;
-    float const b = border;
-    float const w = dim - 2.0f * border;
-    float const h = dim - 2.0f * border;
-    float4 const v = float4{ l, b, l + w, b + h } * texel;
+math::float4 ShadowMap::getViewportNormalized(ShadowMapInfo const& shadowMapInfo) const noexcept {
+    const auto [l, b, w, h] = getViewport();
+    const float texel = 1.0f / float(shadowMapInfo.atlasDimension);
+    const float4 v = float4{ l, b, l + w, b + h } * texel;
    if (shadowMapInfo.textureSpaceFlipped) {
        // this is equivalent to calling uvToRenderTargetUV() in the shader *after* clamping
        // texture coordinates to this normalized viewport.
@@ -222,7 +222,7 @@ private:
            const math::float3& dir);

    static inline void snapLightFrustum(math::float2& s, math::float2& o,
-            math::mat4f const& Mv, math::double3 wsSnapCoords, math::int2 resolution) noexcept;
+            math::mat4f const& Mv, math::float3 worldOrigin, math::float2 shadowMapResolution) noexcept;

    static inline void computeFrustumCorners(math::float3* out,
            const math::mat4f& projectionViewInverse, math::float2 csNearFar = { -1.0f, 1.0f }) noexcept;
@@ -240,8 +240,10 @@ private:
    static inline Aabb compute2DBounds(const math::mat4f& lightView,
            math::float3 const* wsVertices, size_t count) noexcept;

-    static inline void intersectWithShadowCasters(Aabb* lightFrustum,
-            math::mat4f const& lightView,
+    static inline Aabb compute2DBounds(const math::mat4f& lightView,
+            math::float4 const& sphere) noexcept;
+
+    static inline void intersectWithShadowCasters(Aabb& lightFrustum, const math::mat4f& lightView,
            Aabb const& wsShadowCastersVolume) noexcept;

    static inline math::float2 computeNearFarOfWarpSpace(math::mat4f const& lightView,
@@ -279,13 +281,13 @@ private:
    static math::mat4f computeVsmLightSpaceMatrix(const math::mat4f& lightSpacePcf,
            const math::mat4f& Mv, float znear, float zfar) noexcept;

-    math::float4 getClampToEdgeCoords(ShadowMapInfo const& shadowMapInfo) const noexcept;
+    math::float4 getViewportNormalized(ShadowMapInfo const& shadowMapInfo) const noexcept;

-    static float texelSizeWorldSpace(const math::mat3f& worldToShadowTexture,
-            uint16_t shadowDimension) noexcept;
+    float texelSizeWorldSpace(const math::mat3f& worldToShadowTexture,
+            uint16_t shadowDimension) const noexcept;

-    static float texelSizeWorldSpace(const math::mat4f& W, const math::mat4f& MbMtF,
-            uint16_t shadowDimension) noexcept;
+    float texelSizeWorldSpace(const math::mat4f& W, const math::mat4f& MbMtF,
+            uint16_t shadowDimension) const noexcept;

    static constexpr const Segment sBoxSegments[12] = {
            { 0, 1 }, { 1, 3 }, { 3, 2 }, { 2, 0 },
@@ -435,9 +435,8 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::render(FEngine& engine, FrameG
 }

 ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEngine& engine,
-        FView& view, CameraInfo cameraInfo, FScene::RenderableSoa& renderableData,
+        FView& view, CameraInfo const& cameraInfo, FScene::RenderableSoa& renderableData,
        FScene::LightSoa const& lightData, ShadowMap::SceneInfo sceneInfo) noexcept {
-
    FScene* scene = view.getScene();
    auto& lcm = engine.getLightManager();

@@ -445,24 +444,6 @@ ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEng
    FLightManager::ShadowOptions const& options = lcm.getShadowOptions(directionalLight);
    FLightManager::ShadowParams const& params = lcm.getShadowParams(directionalLight);

-    // Adjust the camera's projection for the light's shadowFar
-
-    cameraInfo.zf = params.options.shadowFar > 0.0f ? params.options.shadowFar : cameraInfo.zf;
-    if (UTILS_UNLIKELY(params.options.shadowFar > 0.0f)) {
-        cameraInfo.zf = params.options.shadowFar;
-        float const n = cameraInfo.zn;
-        float const f = cameraInfo.zf;
-        if (std::abs(cameraInfo.cullingProjection[2].w) > std::numeric_limits<float>::epsilon()) {
-            // perspective projection
-            cameraInfo.cullingProjection[2].z = (f + n) / (n - f);
-            cameraInfo.cullingProjection[3].z = (2 * f * n) / (n - f);
-        } else {
-            // orthographic projection
-            cameraInfo.cullingProjection[2].z = 2.0f / (n - f);
-            cameraInfo.cullingProjection[3].z = (f + n) / (n - f);
-        }
-    }
-
    const ShadowMap::ShadowMapInfo shadowMapInfo{
            .atlasDimension      = mTextureAtlasRequirements.size,
            .textureDimension    = uint16_t(options.mapSize),
@@ -478,7 +459,7 @@ ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEng
        // entire camera frustum, as if we only had a single cascade.
        ShadowMap& shadowMap = *mCascadeShadowMaps[0];

-        const auto direction = options.transform * lightData.elementAt<FScene::DIRECTION>(0);
+        const auto direction = lightData.elementAt<FScene::DIRECTION>(0);

        // We compute the directional light's model matrix using the origin's as the light position.
        // The choice of the light's origin initially doesn't matter for a directional light.
@@ -523,13 +504,19 @@ ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEng
            splitPercentages[i] = options.cascadeSplitPositions[i - 1];
        }

-        const CascadeSplits splits({
+        const CascadeSplits::Params p{
                .proj = cameraInfo.cullingProjection,
                .near = vsNear,
                .far = vsFar,
                .cascadeCount = cascadeCount,
                .splitPositions = splitPercentages
-        });
+        };
+        if (p != mCascadeSplitParams) {
+            mCascadeSplits = CascadeSplits{ p };
+            mCascadeSplitParams = p;
+        }
+
+        const CascadeSplits& splits = mCascadeSplits;

        // The split positions uniform is a float4. To save space, we chop off the first split position
        // (which is the near plane, and doesn't need to be communicated to the shaders).
@@ -543,7 +530,7 @@ ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEng

        mShadowMappingUniforms.cascadeSplits = wsSplitPositionUniform;

-        // When computing the required bias we need a half-texel size, so we multiply by 0.5 here.
+        // when computing the required bias we need a half-texel size, so we multiply by 0.5 here.
        // note: normalBias is set to zero for VSM
        const float normalBias = shadowMapInfo.vsm ? 0.0f : 0.5f * lcm.getShadowNormalBias(0);

@@ -109,7 +109,7 @@ public:

 private:
    ShadowMapManager::ShadowTechnique updateCascadeShadowMaps(FEngine& engine,
-            FView& view, CameraInfo cameraInfo, FScene::RenderableSoa& renderableData,
+            FView& view, CameraInfo const& cameraInfo, FScene::RenderableSoa& renderableData,
            FScene::LightSoa const& lightData, ShadowMap::SceneInfo sceneInfo) noexcept;

    ShadowMapManager::ShadowTechnique updateSpotShadowMaps(FEngine& engine,
@@ -145,8 +145,17 @@ private:
            float far = 0.0f;
            size_t cascadeCount = 1;
            std::array<float, SPLIT_COUNT> splitPositions = { 0.0f };
+
+            bool operator!=(const Params& rhs) const {
+                return proj != rhs.proj ||
+                       near != rhs.near ||
+                       far != rhs.far ||
+                       cascadeCount != rhs.cascadeCount ||
+                       splitPositions != rhs.splitPositions;
+            }
        };

+        CascadeSplits() noexcept : CascadeSplits(Params{}) {}
        explicit CascadeSplits(Params const& params) noexcept;

        // Split positions in world-space.
@@ -177,6 +186,9 @@ private:

    SoftShadowOptions mSoftShadowOptions;

+    CascadeSplits::Params mCascadeSplitParams;
+    CascadeSplits mCascadeSplits;
+
    mutable TypedUniformBuffer<ShadowUib> mShadowUb;
    backend::Handle<backend::HwBufferObject> mShadowUbh;

@@ -192,8 +204,8 @@ private:
            utils::FixedCapacityVector<ShadowMap*>::with_capacity(
                    CONFIG_MAX_SHADOWMAPS - CONFIG_MAX_SHADOW_CASCADES) };

-    // Inline storage for all our ShadowMap objects, we can't easily use a std::array<> directly.
-    // Because ShadowMap doesn't have a default ctor, and we avoid out-of-line allocations.
+    // inline storage for all our ShadowMap objects, we can't easily use a std::array<> directly.
+    // because ShadowMap doesn't have a default ctor, and we avoid out-of-line allocations.
    // Each ShadowMap is currently 40 bytes (total of 2.5KB for 64 shadow maps)
    using ShadowMapStorage = std::aligned_storage<sizeof(ShadowMap), alignof(ShadowMap)>::type;
    std::array<ShadowMapStorage, CONFIG_MAX_SHADOWMAPS> mShadowMapCache;
@@ -127,12 +127,12 @@ public:
    }

    bool isSpotLight(Instance i) const noexcept {
-        Type const type = getType(i);
+        Type type = getType(i);
        return type == Type::FOCUSED_SPOT || type == Type::SPOT;
    }

    bool isDirectionalLight(Instance i) const noexcept {
-        Type const type = getType(i);
+        Type type = getType(i);
        return type == Type::DIRECTIONAL || type == Type::SUN;
    }

@@ -1080,9 +1080,8 @@ void FView::setSoftShadowOptions(SoftShadowOptions options) noexcept {

 void FView::setBloomOptions(BloomOptions options) noexcept {
    options.dirtStrength = math::saturate(options.dirtStrength);
-    options.resolution = math::clamp(options.resolution, 2u, 2048u);
-    options.levels = math::clamp(options.levels, uint8_t(1),
-            FTexture::maxLevelCount(options.resolution));
+    options.levels = math::clamp(options.levels, uint8_t(1), uint8_t(11));
+    options.resolution = math::clamp(options.resolution, 1u << options.levels, 2048u);
    options.anamorphism = math::clamp(options.anamorphism, 1.0f/32.0f, 32.0f);
    options.highlight = std::max(10.0f, options.highlight);
    mBloomOptions = options;
@@ -5,6 +5,18 @@ material {
            type : sampler2d,
            name : source,
            precision: medium
+        },
+        {
+            type : float,
+            name : level
+        },
+        {
+            type : float,
+            name : threshold
+        },
+        {
+            type : float,
+            name : invHighlight
        }
    ],
    variables : [
@@ -25,43 +37,85 @@ fragment {

    void dummy(){}

+    void threshold(inout vec3 c) {
+        // threshold everything below 1.0
+        c = max(vec3(0.0), c - 1.0);
+        // crush everything above 1
+        highp float f = max3(c);
+        c *= 1.0 / (1.0 + f * materialParams.invHighlight);
+    }
+
    vec3 box4x4(vec3 s0, vec3 s1, vec3 s2, vec3 s3) {
        return (s0 + s1 + s2 + s3) * 0.25;
    }

+    vec3 box4x4Reinhard(vec3 s0, vec3 s1, vec3 s2, vec3 s3) {
+        float w0 = 1.0 / (1.0 + max3(s0));
+        float w1 = 1.0 / (1.0 + max3(s1));
+        float w2 = 1.0 / (1.0 + max3(s2));
+        float w3 = 1.0 / (1.0 + max3(s3));
+        return (s0 * w0 + s1 * w1 + s2 * w2 + s3 * w3) * (1.0 / (w0 + w1 + w2 + w3));
+    }
+
    void postProcess(inout PostProcessInputs postProcess) {
+        float lod = materialParams.level;
        highp vec2 uv = variable_vertex.xy;

        // see SIGGRAPH 2014: Advances in Real-Time Rendering
        //     "Next Generation Post-Processing in Call of Duty Advanced Warfare"
        //      Jorge Jimenez
-        vec3 c = textureLod(materialParams_source, uv, 0.0).rgb;
+        vec3 c = textureLod(materialParams_source, uv, lod).rgb;

        // The offsets below are in "source" texture space
-        vec3 lt  = textureLodOffset(materialParams_source, uv, 0.0, ivec2(-1, -1)).rgb;
-        vec3 rt  = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 1, -1)).rgb;
-        vec3 rb  = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 1,  1)).rgb;
-        vec3 lb  = textureLodOffset(materialParams_source, uv, 0.0, ivec2(-1,  1)).rgb;
+        vec3 lt  = textureLodOffset(materialParams_source, uv, lod, ivec2(-1, -1)).rgb;
+        vec3 rt  = textureLodOffset(materialParams_source, uv, lod, ivec2( 1, -1)).rgb;
+        vec3 rb  = textureLodOffset(materialParams_source, uv, lod, ivec2( 1,  1)).rgb;
+        vec3 lb  = textureLodOffset(materialParams_source, uv, lod, ivec2(-1,  1)).rgb;

-        vec3 lt2 = textureLodOffset(materialParams_source, uv, 0.0, ivec2(-2, -2)).rgb;
-        vec3 rt2 = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 2, -2)).rgb;
-        vec3 rb2 = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 2,  2)).rgb;
-        vec3 lb2 = textureLodOffset(materialParams_source, uv, 0.0, ivec2(-2,  2)).rgb;
+        vec3 lt2 = textureLodOffset(materialParams_source, uv, lod, ivec2(-2, -2)).rgb;
+        vec3 rt2 = textureLodOffset(materialParams_source, uv, lod, ivec2( 2, -2)).rgb;
+        vec3 rb2 = textureLodOffset(materialParams_source, uv, lod, ivec2( 2,  2)).rgb;
+        vec3 lb2 = textureLodOffset(materialParams_source, uv, lod, ivec2(-2,  2)).rgb;

-        vec3 l   = textureLodOffset(materialParams_source, uv, 0.0, ivec2(-2,  0)).rgb;
-        vec3 t   = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 0, -2)).rgb;
-        vec3 r   = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 2,  0)).rgb;
-        vec3 b   = textureLodOffset(materialParams_source, uv, 0.0, ivec2( 0,  2)).rgb;
+        vec3 l   = textureLodOffset(materialParams_source, uv, lod, ivec2(-2,  0)).rgb;
+        vec3 t   = textureLodOffset(materialParams_source, uv, lod, ivec2( 0, -2)).rgb;
+        vec3 r   = textureLodOffset(materialParams_source, uv, lod, ivec2( 2,  0)).rgb;
+        vec3 b   = textureLodOffset(materialParams_source, uv, lod, ivec2( 0,  2)).rgb;

        // five h4x4 boxes
        vec3 c0, c1;

-        // common case
-        c0  = box4x4(lt, rt, rb, lb);
-        c1  = box4x4(c, l, t, lt2);
-        c1 += box4x4(c, r, t, rt2);
-        c1 += box4x4(c, r, b, rb2);
-        c1 += box4x4(c, l, b, lb2);
+        if (materialParams.level <= 0.5) {
+            if (materialParams.threshold > 0.0) {
+                // Threshold the first level blur
+                threshold(c);
+                threshold(lt);
+                threshold(rt);
+                threshold(rb);
+                threshold(lb);
+                threshold(lt2);
+                threshold(rt2);
+                threshold(rb2);
+                threshold(lb2);
+                threshold(l);
+                threshold(t);
+                threshold(r);
+                threshold(b);
+            }
+            // Also apply fireflies (flickering) filtering
+            c0  = box4x4Reinhard(lt, rt, rb, lb);
+            c1  = box4x4Reinhard(c, l, t, lt2);
+            c1 += box4x4Reinhard(c, r, t, rt2);
+            c1 += box4x4Reinhard(c, r, b, rb2);
+            c1 += box4x4Reinhard(c, l, b, lb2);
+        } else {
+            // common case
+            c0  = box4x4(lt, rt, rb, lb);
+            c1  = box4x4(c, l, t, lt2);
+            c1 += box4x4(c, r, t, rt2);
+            c1 += box4x4(c, r, b, rb2);
+            c1 += box4x4(c, l, b, lb2);
+        }

        // weighted average of the five boxes
        postProcess.color.rgb = c0 * 0.5 + c1 * 0.125;
@@ -1,98 +0,0 @@
-material {
-    name : bloomDownsample2x,
-    parameters : [
-        {
-            type : sampler2d,
-            name : source,
-            precision: medium
-        },
-        {
-            type : float,
-            name : level
-        },
-        {
-            type : float,
-            name : threshold
-        },
-        {
-            type : float,
-            name : fireflies
-        },
-        {
-            type : float,
-            name : invHighlight
-        }
-    ],
-    variables : [
-        vertex
-    ],
-    domain : postprocess,
-    depthWrite : false,
-    depthCulling : false
-}
-
-vertex {
-    void postProcessVertex(inout PostProcessVertexInputs postProcess) {
-        postProcess.vertex.xy = uvToRenderTargetUV(postProcess.normalizedUV);
-    }
-}
-
-fragment {
-
-    void dummy(){}
-
-    void threshold(inout vec3 c) {
-        // threshold everything below 1.0
-        c = max(vec3(0.0), c - 1.0);
-        // crush everything above 1
-        highp float f = max3(c);
-        c *= 1.0 / (1.0 + f * materialParams.invHighlight);
-    }
-
-    void postProcess(inout PostProcessInputs postProcess) {
-        float lod = materialParams.level;
-
-        highp vec2 size = vec2(textureSize(materialParams_source, int(lod)));
-        highp vec2 texelSize = vec2(1.0) / size;
-
-        // Castaño, 2013, "Shadow Mapping Summary Part 1"
-        // 3x3 gaussian filter with 4 linear samples
-        vec2 offset = vec2(0.5);
-        highp vec2 uv = (variable_vertex.xy * size) + offset;
-        highp vec2 base = (floor(uv) - offset) * texelSize;
-        highp vec2 st = fract(uv);
-        vec2 uw = vec2(3.0 - 2.0 * st.x, 1.0 + 2.0 * st.x);
-        vec2 vw = vec2(3.0 - 2.0 * st.y, 1.0 + 2.0 * st.y);
-        highp vec2 u = vec2((2.0 - st.x) / uw.x - 1.0, st.x / uw.y + 1.0) * texelSize.x;
-        highp vec2 v = vec2((2.0 - st.y) / vw.x - 1.0, st.y / vw.y + 1.0) * texelSize.y;
-        vec3 c0 = textureLod(materialParams_source, base + vec2(u.x, v.x), lod).rgb;
-        vec3 c1 = textureLod(materialParams_source, base + vec2(u.y, v.x), lod).rgb;
-        vec3 c2 = textureLod(materialParams_source, base + vec2(u.x, v.y), lod).rgb;
-        vec3 c3 = textureLod(materialParams_source, base + vec2(u.y, v.y), lod).rgb;
-
-        float w0 = uw.x * vw.x * (1.0 / 16.0);
-        float w1 = uw.y * vw.x * (1.0 / 16.0);
-        float w2 = uw.x * vw.y * (1.0 / 16.0);
-        float w3 = uw.y * vw.y * (1.0 / 16.0);
-
-        if (materialParams.fireflies > 0.0) {
-            w0 /= (1.0 + max3(c0));
-            w1 /= (1.0 + max3(c1));
-            w2 /= (1.0 + max3(c2));
-            w3 /= (1.0 + max3(c3));
-            float w = 1.0 / (w0 + w1 + w2 + w3);
-            w0 *= w;
-            w1 *= w;
-            w2 *= w;
-            w3 *= w;
-        }
-
-        vec3 c = c0 * w0 + c1 * w1 + c2 * w2 + c3 * w3;
-
-        if (materialParams.threshold > 0.0) {
-            threshold(c);
-        }
-
-        postProcess.color.rgb = c;
-    }
-}
@@ -1,65 +0,0 @@
-material {
-    name : bloomDownsample9,
-    parameters : [
-        {
-            type : sampler2d,
-            name : source,
-            precision: medium
-        }
-    ],
-    variables : [
-        vertex
-    ],
-    domain : postprocess,
-    depthWrite : false,
-    depthCulling : false
-}
-
-vertex {
-    void postProcessVertex(inout PostProcessVertexInputs postProcess) {
-        postProcess.vertex.xy = uvToRenderTargetUV(postProcess.normalizedUV);
-    }
-}
-
-fragment {
-
-    void dummy(){}
-
-    // see https://www.shadertoy.com/view/cslczj
-    // 6x6 downsampling kernel implemented via 9 bilinear samples
-
-    void postProcess(inout PostProcessInputs postProcess) {
-        highp vec2 uv = variable_vertex.xy;
-        highp vec2 size = vec2(1.0) / vec2(textureSize(materialParams_source, 0));
-
-        float o  = 1.5 + 0.261629;
-        float wa = 7.46602 / 32.0;
-        float wb = 1.0 - wa * 2.0;
-        float wab = wa * wb;
-        float waa = wa * wa;
-        float wbb = wb * wb;
-
-        size *= o;
-
-        vec3 c  = textureLod(materialParams_source, uv + vec2(0.0)            , 0.0).rgb;
-        vec3 l  = textureLod(materialParams_source, uv + vec2(-size.x,    0.0), 0.0).rgb;
-        vec3 r  = textureLod(materialParams_source, uv + vec2( size.x,    0.0), 0.0).rgb;
-        vec3 b  = textureLod(materialParams_source, uv + vec2(    0.0,-size.y), 0.0).rgb;
-        vec3 t  = textureLod(materialParams_source, uv + vec2(    0.0, size.y), 0.0).rgb;
-        vec3 lb = textureLod(materialParams_source, uv + vec2(-size.x,-size.y), 0.0).rgb;
-        vec3 rb = textureLod(materialParams_source, uv + vec2( size.x,-size.y), 0.0).rgb;
-        vec3 lt = textureLod(materialParams_source, uv + vec2(-size.x, size.y), 0.0).rgb;
-        vec3 rt = textureLod(materialParams_source, uv + vec2( size.x, size.y), 0.0).rgb;
-
-        postProcess.color.rgb =
-                (c * wbb +
-                (l * wab +
-                (r * wab +
-                (b * wab +
-                (t * wab +
-                (lb * waa +
-                (rb * waa +
-                (lt * waa +
-                (rt * waa)))))))));
-    }
-}
@@ -10,6 +10,10 @@ material {
            type : float4,
            name : resolution,
            precision: high
+        },
+        {
+            type : float,
+            name : level
        }
    ],
    variables : [
@@ -30,15 +34,16 @@ fragment {
    void dummy(){}

    void postProcess(inout PostProcessInputs postProcess) {
+        float lod = materialParams.level;
        highp vec2 uv = variable_vertex.xy;

 #if FILAMENT_QUALITY < FILAMENT_QUALITY_HIGH
        highp vec4 d = vec4(materialParams.resolution.zw, -materialParams.resolution.zw) * 0.5;
        vec3 c;
-        c  = textureLod(materialParams_source, uv + d.zw, 0.0).rgb;
-        c += textureLod(materialParams_source, uv + d.xw, 0.0).rgb;
-        c += textureLod(materialParams_source, uv + d.xy, 0.0).rgb;
-        c += textureLod(materialParams_source, uv + d.zy, 0.0).rgb;
+        c  = textureLod(materialParams_source, uv + d.zw, lod).rgb;
+        c += textureLod(materialParams_source, uv + d.xw, lod).rgb;
+        c += textureLod(materialParams_source, uv + d.xy, lod).rgb;
+        c += textureLod(materialParams_source, uv + d.zy, lod).rgb;
        postProcess.color.rgb = c * 0.25;
 #else
        // see SIGGRAPH 2014: Advances in Real-Time Rendering
@@ -47,15 +52,15 @@ fragment {
        const float radius = 1.0;
        highp vec4 d = vec4(materialParams.resolution.zw, -materialParams.resolution.zw) * radius;
        vec3 c0, c1;
-        c0  = textureLod(materialParams_source, uv + d.zw, 0.0).rgb;
-        c0 += textureLod(materialParams_source, uv + d.xw, 0.0).rgb;
-        c0 += textureLod(materialParams_source, uv + d.xy, 0.0).rgb;
-        c0 += textureLod(materialParams_source, uv + d.zy, 0.0).rgb;
-        c0 += 4.0 * textureLod(materialParams_source, uv, 0.0).rgb;
-        c1  = textureLod(materialParams_source, uv + vec2(d.z,  0.0), 0.0).rgb;
-        c1 += textureLod(materialParams_source, uv + vec2(0.0,  d.w), 0.0).rgb;
-        c1 += textureLod(materialParams_source, uv + vec2(d.x,  0.0), 0.0).rgb;
-        c1 += textureLod(materialParams_source, uv + vec2( 0.0, d.y), 0.0).rgb;
+        c0  = textureLod(materialParams_source, uv + d.zw, lod).rgb;
+        c0 += textureLod(materialParams_source, uv + d.xw, lod).rgb;
+        c0 += textureLod(materialParams_source, uv + d.xy, lod).rgb;
+        c0 += textureLod(materialParams_source, uv + d.zy, lod).rgb;
+        c0 += 4.0 * textureLod(materialParams_source, uv, lod).rgb;
+        c1  = textureLod(materialParams_source, uv + vec2(d.z,  0.0), lod).rgb;
+        c1 += textureLod(materialParams_source, uv + vec2(0.0,  d.w), lod).rgb;
+        c1 += textureLod(materialParams_source, uv + vec2(d.x,  0.0), lod).rgb;
+        c1 += textureLod(materialParams_source, uv + vec2( 0.0, d.y), lod).rgb;
        postProcess.color.rgb = (c0 + 2.0 * c1) * (1.0 / 16.0);
 #endif
    }
@@ -11,6 +11,10 @@ material {
            name : coc,
            precision: medium
        },
+        {
+            type : int,
+            name : mip
+        },
        {
            type : float,
            name : weightScale
@@ -61,18 +65,19 @@ void postProcess(inout PostProcessInputs postProcess) {

    // the bilateral weights need to be scaled by to match the lower resolution
    float weightScale = materialParams.weightScale;
+    float mip = float(materialParams.mip);

-    vec4 s01 = textureLodOffset(materialParams_color, uv, 0.0, ivec2(0, 1));
-    vec4 s11 = textureLodOffset(materialParams_color, uv, 0.0, ivec2(1, 1));
-    vec4 s10 = textureLodOffset(materialParams_color, uv, 0.0, ivec2(1, 0));
-    vec4 s00 = textureLodOffset(materialParams_color, uv, 0.0, ivec2(0, 0));
+    vec4 s01 = textureLodOffset(materialParams_color, uv, mip, ivec2(0, 1));
+    vec4 s11 = textureLodOffset(materialParams_color, uv, mip, ivec2(1, 1));
+    vec4 s10 = textureLodOffset(materialParams_color, uv, mip, ivec2(1, 0));
+    vec4 s00 = textureLodOffset(materialParams_color, uv, mip, ivec2(0, 0));

    // fetch the 4 corresponding CoC (textureGather with LOD doesn't exist)
    vec4 c;
-    c[0] = textureLodOffset(materialParams_coc, uv, 0.0, ivec2(0, 1)).r;
-    c[1] = textureLodOffset(materialParams_coc, uv, 0.0, ivec2(1, 1)).r;
-    c[2] = textureLodOffset(materialParams_coc, uv, 0.0, ivec2(1, 0)).r;
-    c[3] = textureLodOffset(materialParams_coc, uv, 0.0, ivec2(0, 0)).r;
+    c[0] = textureLodOffset(materialParams_coc, uv, mip, ivec2(0, 1)).r;
+    c[1] = textureLodOffset(materialParams_coc, uv, mip, ivec2(1, 1)).r;
+    c[2] = textureLodOffset(materialParams_coc, uv, mip, ivec2(1, 0)).r;
+    c[3] = textureLodOffset(materialParams_coc, uv, mip, ivec2(0, 0)).r;

    float outCoc = downsampleCoC(c);
    vec4 w = downsampleCocWeights(c, outCoc, weightScale);
@@ -5,6 +5,10 @@ material {
            type : sampler2d,
            name : depth,
            precision: high
+        },
+        {
+            type : int,
+            name : level
        }
    ],
    variables : [
@@ -26,8 +30,9 @@ fragment {
    // We use a rotated grid sub-sample as it's cheap and gives good results
    // See Scalable Ambient Obscurance by McGuire and al.
    void postProcess(inout PostProcessInputs postProcess) {
+        int level = materialParams.level;
        ivec2 icoord = ivec2(gl_FragCoord.xy);
        postProcess.depth = texelFetch(materialParams_depth,
-                2 * icoord + ivec2(icoord.y & 1, icoord.x & 1), 0).r;
+                2 * icoord + ivec2(icoord.y & 1, icoord.x & 1), level).r;
    }
 }
@@ -10,6 +10,10 @@ material {
            type : int,
            name : layer
        },
+        {
+            type : int,
+            name : level
+        },
        {
            type : float,
            name : uvscale
@@ -32,6 +36,7 @@ fragment {
    void postProcess(inout PostProcessInputs postProcess) {
        highp vec2 uv = gl_FragCoord.xy * materialParams.uvscale;
        postProcess.color = textureLod(materialParams_color,
-                vec3(uv, materialParams.layer), 0.0);
+                vec3(uv, materialParams.layer),
+                float(materialParams.level));
    }
 }
@@ -1,12 +1,12 @@
 Pod::Spec.new do |spec|
  spec.name = "Filament"
-  spec.version = "1.43.1"
+  spec.version = "1.42.0"
  spec.license = { :type => "Apache 2.0", :file => "LICENSE" }
  spec.homepage = "https://google.github.io/filament"
  spec.authors = "Google LLC."
  spec.summary = "Filament is a real-time physically based rendering engine for Android, iOS, Windows, Linux, macOS, and WASM/WebGL."
  spec.platform = :ios, "11.0"
-  spec.source = { :http => "https://github.com/google/filament/releases/download/v1.43.1/filament-v1.43.1-ios.tgz" }
+  spec.source = { :http => "https://github.com/google/filament/releases/download/v1.42.0/filament-v1.42.0-ios.tgz" }

  # Fix linking error with Xcode 12; we do not yet support the simulator on Apple silicon.
  spec.pod_target_xcconfig = {
@@ -17,8 +17,6 @@
 #ifndef TNT_FILAMENT_SAMPLE_CAMERA_MANIPULATOR_H
 #define TNT_FILAMENT_SAMPLE_CAMERA_MANIPULATOR_H

-#include <functional>
-
 #include <math/vec2.h>
 #include <math/vec3.h>
 #include <math/vec4.h>
@@ -17,8 +17,6 @@
 #ifndef TNT_FILAMENT_SAMPLE_CAMERA_MANIPULATOR_H
 #define TNT_FILAMENT_SAMPLE_CAMERA_MANIPULATOR_H

-#include <functional>
-
 #include <math/vec2.h>
 #include <math/vec3.h>
 #include <math/vec4.h>
@@ -28,7 +28,7 @@
 namespace filament {

 // update this when a new version of filament wouldn't work with older materials
-static constexpr size_t MATERIAL_VERSION = 43;
+static constexpr size_t MATERIAL_VERSION = 42;

 /**
 * Supported shading models
@@ -35,7 +35,7 @@ for example, `filament-20181009-linux.tgz`.

 Create a file, `main.cpp`, in the same directory with the following contents:

-```c++
+```
 #include <filamat/MaterialBuilder.h>

 #include <iostream>
@@ -85,7 +85,7 @@ Copy your platform's Makefile below into a `Makefile` inside the same directory.

 ### Linux

-```make
+```
 FILAMENT_LIBS=-lfilamat -lfilabridge -lshaders -lutils -lsmol-v
 CC=clang++

@@ -103,7 +103,7 @@ clean:

 ### macOS

-```make
+```
 FILAMENT_LIBS=-lfilamat -lfilabridge -lshaders -lutils -lsmol-v
 CC=clang++

@@ -129,7 +129,7 @@ flags](https://docs.microsoft.com/en-us/cpp/build/reference/md-mt-ld-use-run-tim
 When building Filamat from source, the `USE_STATIC_CRT` CMake option can be
 used to change the run-time library version.

-```make
+```
 FILAMENT_LIBS=lib/x86_64/mt/filamat.lib lib/x86_64/mt/filabridge.lib lib/x86_64/mt/shaders.lib \
              lib/x86_64/mt/utils.lib lib/x86_64/mt/smol-v.lib
 CC=clang-cl.exe
@@ -164,7 +164,7 @@ and invoke `nmake` instead of `make`.
 For simplicity, this demo doesn't do anything useful with the built material package. To use the
 material with Filament, pass the material package's data into a Filament Material builder:

-```c++
+```
    Package package = builder.build();
    filament::Material* myMaterial = Material::Builder()
        .package(package.getData(), package.getSize())
@@ -186,7 +186,7 @@ In addition, `filamat_lite` only performs a simple text match to determine which
 `MaterialInputs` structure are set. The `material` input variable must also always be refered to by
 the name `material`.

-```glsl
+```
 void anotherFunction(inout MaterialInputs m) {
    // Incorrect! The MaterialInputs is being referred to by the name "m".
    m.metallic = 0.0;
@@ -16,7 +16,6 @@ set(PUBLIC_HDRS
        include/gltfio/FilamentInstance.h
        include/gltfio/MaterialProvider.h
        include/gltfio/NodeManager.h
-        include/gltfio/TrsTransformManager.h
        include/gltfio/ResourceLoader.h
        include/gltfio/TextureProvider.h
        include/gltfio/math.h
@@ -36,12 +35,10 @@ set(SRCS
        src/FilamentAsset.cpp
        src/FilamentInstance.cpp
        src/FNodeManager.h
-        src/FTrsTransformManager.h
        src/GltfEnums.h
        src/Ktx2Provider.cpp
        src/MaterialProvider.cpp
        src/NodeManager.cpp
-        src/TrsTransformManager.cpp
        src/ResourceLoader.cpp
        src/StbProvider.cpp
        src/TangentsJob.cpp
@@ -189,7 +186,8 @@ if (NOT WEBGL AND NOT ANDROID AND NOT IOS)
    # ==================================================================================================
    # Compiler flags
    # ==================================================================================================
-    if (NOT MSVC)
+    if (MSVC)
+    else()
        target_compile_options(${TARGET} PRIVATE -Wno-deprecated-register)
    endif()

@@ -227,9 +225,7 @@ if (TNT_DEV AND NOT WEBGL AND NOT ANDROID AND NOT IOS)
    add_dependencies(${TEST_TARGET} test_gltfio_files)

    target_link_libraries(${TEST_TARGET} PRIVATE ${TARGET} filament filabridge gtest uberarchive)
-    if (NOT MSVC)
-        target_compile_options(${TEST_TARGET} PRIVATE -Wno-deprecated-register)
-    endif()
+    target_compile_options(${TEST_TARGET} PRIVATE -Wno-deprecated-register)
    set_target_properties(${TEST_TARGET} PROPERTIES FOLDER Tests)
 endif()

@@ -1,114 +0,0 @@
-/*
- * Copyright (C) 2023 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 GLTFIO_TRSTRANSFORMMANAGER_H
-#define GLTFIO_TRSTRANSFORMMANAGER_H
-
-#include <filament/FilamentAPI.h>
-
-#include <utils/compiler.h>
-#include <utils/EntityInstance.h>
-#include <math/quat.h>
-#include <math/vec3.h>
-#include <math/mat4.h>
-
-using namespace filament::math;
-
-namespace utils {
-class Entity;
-} // namespace utils
-
-namespace filament::gltfio {
-
-class FTrsTransformManager;
-
-/**
- * TrsTransformManager is used to add entities with glTF-specific trs information.
- *
- * Trs information here just used for Animation, DON'T use for transform.
- */
-class UTILS_PUBLIC TrsTransformManager {
-public:
-    using Instance = utils::EntityInstance<TrsTransformManager>;
-    using Entity = utils::Entity;
-
-    /**
-     * Returns whether a particular Entity is associated with a component of this TrsTransformManager
-     * @param e An Entity.
-     * @return true if this Entity has a component associated with this manager.
-     */
-    bool hasComponent(Entity e) const noexcept;
-
-    /**
-     * Gets an Instance representing the trs transform component associated with the given Entity.
-     * @param e An Entity.
-     * @return An Instance object, which represents the trs transform component associated with the Entity e.
-     * @note Use Instance::isValid() to make sure the component exists.
-     * @see hasComponent()
-     */
-    Instance getInstance(Entity e) const noexcept;
-
-    /**
-     * Creates a trs transform component and associates it with the given entity.
-     * @param entity            An Entity to associate a trs transform component with.
-     * @param translation       The translation to initialize the trs transform component with.
-     * @param rotation          The rotation to initialize the trs transform component with.
-     * @param scale             The scale to initialize the trs transform component with.
-     *
-     * If this component already exists on the given entity, it is first destroyed as if
-     * destroy(Entity e) was called.
-     *
-     * @see destroy()
-     */
-    void create(Entity entity);
-    void create(Entity entity, const float3& translation, const quatf& rotation,
-            const float3& scale);  //!< \overload
-
-    /**
-     * Destroys this component from the given entity.
-     * @param e An entity.
-     *
-     * @see create()
-     */
-    void destroy(Entity e) noexcept;
-
-    void setTranslation(Instance ci, const float3& translation) noexcept;
-    const float3& getTranslation(Instance ci) const noexcept;
-
-    void setRotation(Instance ci, const quatf& rotation) noexcept;
-    const quatf& getRotation(Instance ci) const noexcept;
-
-    void setScale(Instance ci, const float3& scale) noexcept;
-    const float3& getScale(Instance ci) const noexcept;
-
-    void setTrs(Instance ci, const float3& translation, const quatf& rotation,
-            const float3& scale) noexcept;
-    const mat4f getTransform(Instance ci) const noexcept;
-
-protected:
-    TrsTransformManager() noexcept = default;
-    ~TrsTransformManager() = default;
-
-public:
-    TrsTransformManager(TrsTransformManager const&) = delete;
-    TrsTransformManager(TrsTransformManager&&) = delete;
-    TrsTransformManager& operator=(TrsTransformManager const&) = delete;
-    TrsTransformManager& operator=(TrsTransformManager&&) = delete;
-};
-
-} // namespace filament::gltfio
-
-#endif // GLTFIO_TRSTRANSFORMMANAGER_H
@@ -19,7 +19,6 @@

 #include "FFilamentAsset.h"
 #include "FFilamentInstance.h"
-#include "FTrsTransformManager.h"
 #include "downcast.h"

 #include <filament/VertexBuffer.h>
@@ -75,7 +74,6 @@ struct AnimatorImpl {
    FFilamentInstance* instance = nullptr;
    RenderableManager* renderableManager;
    TransformManager* transformManager;
-    TrsTransformManager* trsTransformManager;
    vector<float> weights;
    FixedCapacityVector<mat4f> crossFade;
    void addChannels(const FixedCapacityVector<Entity>& nodeMap, const cgltf_animation& srcAnim,
@@ -192,7 +190,6 @@ Animator::Animator(FFilamentAsset const* asset, FFilamentInstance* instance) {
    mImpl->instance = instance;
    mImpl->renderableManager = &asset->mEngine->getRenderableManager();
    mImpl->transformManager = &asset->mEngine->getTransformManager();
-    mImpl->trsTransformManager = asset->getTrsTransformManager();

    const cgltf_data* srcAsset = asset->mSourceAsset->hierarchy;
    const cgltf_animation* srcAnims = srcAsset->animations;
@@ -432,13 +429,20 @@ void AnimatorImpl::applyAnimation(const Channel& channel, float t, size_t prevIn
        size_t nextIndex) {
    const Sampler* sampler = channel.sourceData;
    const TimeValues& times = sampler->times;
-    TrsTransformManager::Instance trsNode = trsTransformManager->getInstance(channel.targetEntity);
    TransformManager::Instance node = transformManager->getInstance(channel.targetEntity);

+    // Perform the interpolation. This is a simple but inefficient implementation; Filament
+    // stores transforms as mat4's but glTF animation is based on TRS (translation rotation
+    // scale).
+    mat4f xform = transformManager->getTransform(node);
+    float3 scale;
+    quatf rotation;
+    float3 translation;
+    decomposeMatrix(xform, &translation, &rotation, &scale);
+
    switch (channel.transformType) {

        case Channel::SCALE: {
-            float3 scale;
            const float3* srcVec3 = (const float3*) sampler->values.data();
            if (sampler->interpolation == Sampler::CUBIC) {
                float3 vert0 = srcVec3[prevIndex * 3 + 1];
@@ -449,12 +453,10 @@ void AnimatorImpl::applyAnimation(const Channel& channel, float t, size_t prevIn
            } else {
                scale = ((1 - t) * srcVec3[prevIndex]) + (t * srcVec3[nextIndex]);
            }
-            trsTransformManager->setScale(trsNode, scale);
            break;
        }

        case Channel::TRANSLATION: {
-            float3 translation;
            const float3* srcVec3 = (const float3*) sampler->values.data();
            if (sampler->interpolation == Sampler::CUBIC) {
                float3 vert0 = srcVec3[prevIndex * 3 + 1];
@@ -465,12 +467,10 @@ void AnimatorImpl::applyAnimation(const Channel& channel, float t, size_t prevIn
            } else {
                translation = ((1 - t) * srcVec3[prevIndex]) + (t * srcVec3[nextIndex]);
            }
-            trsTransformManager->setTranslation(trsNode, translation);
            break;
        }

        case Channel::ROTATION: {
-            quatf rotation;
            const quatf* srcQuat = (const quatf*) sampler->values.data();
            if (sampler->interpolation == Sampler::CUBIC) {
                quatf vert0 = srcQuat[prevIndex * 3 + 1];
@@ -481,7 +481,6 @@ void AnimatorImpl::applyAnimation(const Channel& channel, float t, size_t prevIn
            } else {
                rotation = slerp(srcQuat[prevIndex], srcQuat[nextIndex], t);
            }
-            trsTransformManager->setRotation(trsNode, rotation);
            break;
        }

@@ -520,7 +519,8 @@ void AnimatorImpl::applyAnimation(const Channel& channel, float t, size_t prevIn
        }
    }

-    transformManager->setTransform(node, trsTransformManager->getTransform(trsNode));
+    xform = composeMatrix(translation, rotation, scale);
+    transformManager->setTransform(node, xform);
 }

 void AnimatorImpl::resetBoneMatrices(FFilamentInstance* instance) {
@@ -21,7 +21,6 @@

 #include "FFilamentAsset.h"
 #include "FNodeManager.h"
-#include "FTrsTransformManager.h"
 #include "GltfEnums.h"

 #include <filament/Box.h>
@@ -296,7 +295,6 @@ public:
    MaterialProvider& mMaterials;
    Engine& mEngine;
    FNodeManager mNodeManager;
-    FTrsTransformManager mTrsTransformManager;

    // Transient state used only for the asset currently being loaded:
    FFilamentAsset* mAsset;
@@ -402,8 +400,7 @@ void FAssetLoader::createRootAsset(const cgltf_data* srcAsset) {
    #endif

    mDummyBufferObject = nullptr;
-    mAsset = new FFilamentAsset(&mEngine, mNameManager, &mEntityManager, &mNodeManager,
-            &mTrsTransformManager, srcAsset);
+    mAsset = new FFilamentAsset(&mEngine, mNameManager, &mEntityManager, &mNodeManager, srcAsset);

    // It is not an error for a glTF file to have zero scenes.
    mAsset->mScenes.clear();
@@ -566,9 +563,7 @@ void FAssetLoader::recurseEntities(const cgltf_data* srcAsset, const cgltf_node*
        quatf* rotation = (quatf*) &node->rotation[0];
        float3* scale = (float3*) &node->scale[0];
        float3* translation = (float3*) &node->translation[0];
-        mTrsTransformManager.create(entity, *translation, *rotation, *scale);
-        localTransform = mTrsTransformManager.getTransform(
-                mTrsTransformManager.getInstance(entity));
+        localTransform = composeMatrix(*translation, *rotation, *scale);
    }

    auto parentTransform = mTransformManager.getInstance(parent);
@@ -1472,9 +1467,8 @@ MaterialInstance* FAssetLoader::createMaterialInstance(const cgltf_data* srcAsse
            }
        }
        if (matkey.hasSheenRoughnessTexture) {
-            bool sameTexture = shConfig.sheen_color_texture.texture == shConfig.sheen_roughness_texture.texture;
            mAsset->addTextureBinding(mi, "sheenRoughnessMap",
-                    shConfig.sheen_roughness_texture.texture, sameTexture ? sRGB : LINEAR);
+                    shConfig.sheen_roughness_texture.texture, LINEAR);
            if (matkey.hasTextureTransforms) {
                const cgltf_texture_transform& uvt = shConfig.sheen_roughness_texture.transform;
                auto uvmat = matrixFromUvTransform(uvt.offset, uvt.rotation, uvt.scale);
@@ -19,7 +19,6 @@

 #include <gltfio/FilamentAsset.h>
 #include <gltfio/NodeManager.h>
-#include <gltfio/TrsTransformManager.h>

 #include <filament/Engine.h>
 #include <filament/IndexBuffer.h>
@@ -111,9 +110,9 @@ using MeshCache = utils::FixedCapacityVector<utils::FixedCapacityVector<Primitiv
 struct FFilamentAsset : public FilamentAsset {
    FFilamentAsset(Engine* engine, utils::NameComponentManager* names,
            utils::EntityManager* entityManager, NodeManager* nodeManager,
-            TrsTransformManager* trsTransformManager, const cgltf_data* srcAsset) :
+            const cgltf_data* srcAsset) :
            mEngine(engine), mNameManager(names), mEntityManager(entityManager),
-            mNodeManager(nodeManager), mTrsTransformManager(trsTransformManager),
+            mNodeManager(nodeManager),
            mSourceAsset(new SourceAsset {(cgltf_data*)srcAsset}),
            mTextures(srcAsset->textures_count),
            mMeshCache(srcAsset->meshes_count) {}
@@ -196,10 +195,6 @@ struct FFilamentAsset : public FilamentAsset {
        return mEngine;
    }

-    TrsTransformManager* getTrsTransformManager() const noexcept {
-        return mTrsTransformManager;
-    }
-
    void releaseSourceData() noexcept;

    const void* getSourceAsset() const noexcept {
@@ -247,7 +242,6 @@ struct FFilamentAsset : public FilamentAsset {
    utils::NameComponentManager* const mNameManager;
    utils::EntityManager* const mEntityManager;
    NodeManager* const mNodeManager;
-    TrsTransformManager* const mTrsTransformManager;
    std::vector<utils::Entity> mEntities; // sorted such that renderables come first
    std::vector<utils::Entity> mLightEntities;
    std::vector<utils::Entity> mCameraEntities;
@@ -1,165 +0,0 @@
-/*
- * Copyright (C) 2023 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 GLTFIO_FTRSTRANSFORMMANAGER_H
-#define GLTFIO_FTRSTRANSFORMMANAGER_H
-
-#include "downcast.h"
-#include "gltfio/math.h"
-#include "math/quat.h"
-#include "utils/debug.h"
-
-#include <gltfio/TrsTransformManager.h>
-
-#include <utils/compiler.h>
-#include <utils/SingleInstanceComponentManager.h>
-#include <utils/Entity.h>
-#include <utils/FixedCapacityVector.h>
-#include <utils/Slice.h>
-
-namespace filament::gltfio {
-
-class UTILS_PRIVATE FTrsTransformManager : public TrsTransformManager {
-public:
-    using Instance = TrsTransformManager::Instance;
-
-    FTrsTransformManager() noexcept {}
-
-    ~FTrsTransformManager() noexcept {
-        assert_invariant(mManager.getComponentCount() == 0);
-    }
-
-    void terminate() noexcept;
-
-    bool hasComponent(utils::Entity e) const noexcept {
-        return mManager.hasComponent(e);
-    }
-
-    Instance getInstance(utils::Entity e) const noexcept {
-        return Instance(mManager.getInstance(e));
-    }
-
-    void create(utils::Entity entity) {
-        create(entity, float3{}, quatf{}, float3{1});
-    }
-
-    void create(utils::Entity entity, const float3& translation,
-                const quatf& rotation, const float3& scale) {
-        if (UTILS_UNLIKELY(mManager.hasComponent(entity))) {
-            destroy(entity);
-        }
-        UTILS_UNUSED_IN_RELEASE Instance ci = mManager.addComponent(entity);
-        assert_invariant(ci);
-
-        if (ci) {
-            setTrs(ci, translation, rotation, scale);
-        }
-    }
-
-    void destroy(utils::Entity e) noexcept {
-        if (Instance ci = mManager.getInstance(e); ci) {
-            mManager.removeComponent(e);
-        }
-    }
-
-    void gc(utils::EntityManager& em) noexcept {
-        mManager.gc(em);
-    }
-
-    void setTranslation(Instance ci, const float3& translation) noexcept {
-        assert_invariant(ci.isValid());
-        mManager[ci].translation = translation;
-    }
-
-    const float3& getTranslation(Instance ci) const noexcept {
-        return mManager[ci].translation;
-    }
-
-    void setRotation(Instance ci, const quatf& rotation) noexcept {
-        assert_invariant(ci.isValid());
-        mManager[ci].rotation = rotation;
-    }
-
-    const quatf& getRotation(Instance ci) const noexcept {
-        return mManager[ci].rotation;
-    }
-
-    void setScale(Instance ci, const float3& scale) noexcept {
-        assert_invariant(ci.isValid());
-        mManager[ci].scale = scale;
-    }
-
-    const float3& getScale(Instance ci) const noexcept {
-        return mManager[ci].scale;
-    }
-
-    void setTrs(Instance ci, const float3& translation,
-            const quatf& rotation, const float3& scale) noexcept {
-        setTranslation(ci, translation);
-        setRotation(ci, rotation);
-        setScale(ci, scale);
-    }
-
-    const mat4f getTransform(Instance ci) const noexcept {
-        return composeMatrix(getTranslation(ci), getRotation(ci), getScale(ci));
-    }
-
-private:
-    enum {
-        TRANSLATION,
-        ROTATION,
-        SCALE,
-    };
-
-    using Base = utils::SingleInstanceComponentManager<
-            float3,
-            quatf,
-            float3>;
-
-    struct Sim : public Base {
-        using Base::gc;
-        using Base::swap;
-
-        typename Base::SoA& getSoA() { return mData; }
-
-        struct Proxy {
-            UTILS_ALWAYS_INLINE
-            Proxy(Base& sim, utils::EntityInstanceBase::Type i) noexcept :
-                    translation{ sim, i } { }
-
-            union {
-                Field<TRANSLATION>   translation;
-                Field<ROTATION>      rotation;
-                Field<SCALE>         scale;
-            };
-        };
-
-        UTILS_ALWAYS_INLINE Proxy operator[](Instance i) noexcept {
-            return { *this, i };
-        }
-        UTILS_ALWAYS_INLINE const Proxy operator[](Instance i) const noexcept {
-            return { const_cast<Sim&>(*this), i };
-        }
-    };
-
-    Sim mManager;
-};
-
-FILAMENT_DOWNCAST(TrsTransformManager)
-
-} // namespace filament::gltfio
-
-#endif // GLTFIO_FTRSTRANSFORMMANAGER_H
@@ -60,11 +60,6 @@ FFilamentAsset::~FFilamentAsset() {

    }

-    // Destroy gltfio trs transform components.
-    for (auto entity : mEntities) {
-        mTrsTransformManager->destroy(entity);
-    }
-
    // Destroy all renderable, light, transform, and camera components,
    // then destroy the actual entities. This includes instances.
    if (!mDetachedFilamentComponents) {
@@ -287,7 +287,7 @@ std::string shaderFromKey(const MaterialKey& config) {

                // TODO: Provided by Filament, but this should really be provided/computed by gltfio
                // TODO: This scale is per renderable and should include the scale of the mesh node
-                float scale = getObjectUserData();
+                float scale = object_uniforms.userData;
                material.thickness = materialParams.volumeThicknessFactor * scale;
            )SHADER";

@@ -1,98 +0,0 @@
-/*
- * Copyright (C) 2023 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 "FTrsTransformManager.h"
-
-#include <utils/Log.h>
-
-#include "downcast.h"
-#include "gltfio/TrsTransformManager.h"
-
-using namespace utils;
-
-namespace filament::gltfio {
-
-using Instance = TrsTransformManager::Instance;
-
-void FTrsTransformManager::terminate() noexcept {
-    auto& manager = mManager;
-    if (!manager.empty()) {
-#ifndef NDEBUG
-        utils::slog.d << "cleaning up " << manager.getComponentCount()
-            << " leaked trs transform components" << utils::io::endl;
-#endif
-        while (!manager.empty()) {
-            Instance ci = manager.end() - 1;
-            manager.removeComponent(manager.getEntity(ci));
-        }
-    }
-}
-
-bool TrsTransformManager::hasComponent(Entity e) const noexcept {
-    return downcast(this)->hasComponent(e);
-}
-
-Instance TrsTransformManager::getInstance(Entity e) const noexcept {
-    return downcast(this)->getInstance(e);
-}
-
-void TrsTransformManager::create(Entity entity) {
-    downcast(this)->create(entity);
-}
-
-void TrsTransformManager::create(Entity entity, const float3& translation,
-        const quatf& rotation, const float3& scale) {
-    downcast(this)->create(entity, translation, rotation, scale);
-}
-
-void TrsTransformManager::destroy(Entity e) noexcept {
-    downcast(this)->destroy(e);
-}
-
-void TrsTransformManager::setTranslation(Instance ci, const float3& translation) noexcept {
-    downcast(this)->setTranslation(ci, translation);
-}
-
-const float3& TrsTransformManager::getTranslation(Instance ci) const noexcept {
-    return downcast(this)->getTranslation(ci);
-}
-
-void TrsTransformManager::setRotation(Instance ci, const quatf& rotation) noexcept {
-    downcast(this)->setRotation(ci, rotation);
-}
-
-const quatf& TrsTransformManager::getRotation(Instance ci) const noexcept {
-    return downcast(this)->getRotation(ci);
-}
-
-void TrsTransformManager::setScale(Instance ci, const float3& scale) noexcept {
-    downcast(this)->setScale(ci, scale);
-}
-
-const float3& TrsTransformManager::getScale(Instance ci) const noexcept {
-    return downcast(this)->getScale(ci);
-}
-
-void TrsTransformManager::setTrs(Instance ci, const float3& translation,
-        const quatf& rotation, const float3& scale) noexcept {
-    downcast(this)->setTrs(ci, translation, rotation, scale);
-}
-
-const mat4f TrsTransformManager::getTransform(Instance ci) const noexcept {
-    return downcast(this)->getTransform(ci);
-}
-
-} // namespace filament::gltfio
@@ -25,7 +25,10 @@
 #include <stdint.h>
 #include <sys/types.h>

-namespace filament::math::details {
+namespace filament {
+namespace math {
+namespace details {
+// -------------------------------------------------------------------------------------

 /*
 * No user serviceable parts here.
@@ -46,7 +49,7 @@ namespace filament::math::details {
 template<template<typename T> class QUATERNION, typename T>
 class TQuatProductOperators {
 public:
-    /* compound assignment from another quaternion of the same size but different
+    /* compound assignment from a another quaternion of the same size but different
     * element type.
     */
    template<typename OTHER>
@@ -82,7 +85,7 @@ public:
     * (the first one, BASE<T> being known).
     */

-    /* The operators below handle operation between quaternions of the same size
+    /* The operators below handle operation between quaternion of the same size
     * but of a different element type.
     */
    template<typename RT>
@@ -124,19 +127,19 @@ public:
     */
    friend inline
    constexpr QUATERNION<T> MATH_PURE operator*(QUATERNION<T> q, T scalar) {
-        // don't pass q by reference because we need a copy anyway
+        // don't pass q by reference because we need a copy anyways
        return q *= scalar;
    }

    friend inline
    constexpr QUATERNION<T> MATH_PURE operator*(T scalar, QUATERNION<T> q) {
-        // don't pass q by reference because we need a copy anyway
+        // don't pass q by reference because we need a copy anyways
        return q *= scalar;
    }

    friend inline
    constexpr QUATERNION<T> MATH_PURE operator/(QUATERNION<T> q, T scalar) {
-        // don't pass q by reference because we need a copy anyway
+        // don't pass q by reference because we need a copy anyways
        return q /= scalar;
    }
 };
@@ -280,6 +283,9 @@ public:
    }
 };

-}  // namespace filament::math::details
+// -------------------------------------------------------------------------------------
+}  // namespace details
+}  // namespace math
+}  // namespace filament

 #endif  // TNT_MATH_TQUATHELPERS_H
@@ -443,37 +443,6 @@ private:
        return v;
    }

-    template<typename U>
-    friend inline
-    VECTOR<T> MATH_PURE fmod(VECTOR<T> const& x, VECTOR<U> const& y) {
-        VECTOR<T> r;
-        for (size_t i = 0; i < r.size(); i++) {
-            r[i] = std::fmod(x[i], y[i]);
-        }
-        return r;
-    }
-
-    template<typename U>
-    friend inline
-    VECTOR<T> MATH_PURE remainder(VECTOR<T> const& x, VECTOR<U> const& y) {
-        VECTOR<T> r;
-        for (size_t i = 0; i < r.size(); i++) {
-            r[i] = std::remainder(x[i], y[i]);
-        }
-        return r;
-    }
-
-    template<typename U>
-    friend inline
-    VECTOR<T> MATH_PURE remquo(VECTOR<T> const& x, VECTOR<U> const& y,
-            VECTOR<int>* q) {
-        VECTOR<T> r;
-        for (size_t i = 0; i < r.size(); i++) {
-            r[i] = std::remquo(x[i], y[i], &((*q)[i]));
-        }
-        return r;
-    }
-
    friend inline VECTOR<T> MATH_PURE inversesqrt(VECTOR<T> v) {
        for (size_t i = 0; i < v.size(); i++) {
            v[i] = T(1) / std::sqrt(v[i]);
@@ -160,7 +160,7 @@ constexpr inline half makeHalf(uint16_t bits) noexcept {

 #endif // __ARM_NEON

-inline constexpr half operator""_h(long double v) {
+inline constexpr half operator "" _h(long double v) {
    return half( static_cast<float>(v) );
 }

@@ -33,7 +33,8 @@

 #include <stdint.h>

-namespace filament::math {
+namespace filament {
+namespace math {
 namespace details {

 template<typename T> class TVec2;
@@ -44,8 +45,6 @@ template<typename T> class TMat22;
 template<typename T> class TMat33;
 template<typename T> class TMat44;

-template<typename T> class TQuaternion;
-
 }  // namespace details

 using double2   = details::TVec2<double>;
@@ -87,10 +86,8 @@ using mat3f     = details::TMat33<float>;
 using mat4      = details::TMat44<double>;
 using mat4f     = details::TMat44<float>;

-using quat      = details::TQuaternion<double>;
-using quatf     = details::TQuaternion<float>;
-
-}  // namespace filament::math
+}  // namespace math
+}  // namespace filament

 #endif // _MSC_VER

@@ -26,7 +26,10 @@
 #include <stdint.h>
 #include <sys/types.h>

-namespace filament::math {
+namespace filament {
+namespace math {
+// -------------------------------------------------------------------------------------
+
 namespace details {

 template<typename T>
@@ -86,15 +89,15 @@ public:

    // constructors

-    // Leaves object uninitialized. Use with caution.
+    // leaves object uninitialized. use with caution.
    explicit constexpr TQuaternion(no_init) {}

    // default constructor. sets all values to zero.
    constexpr TQuaternion() : x(0), y(0), z(0), w(0) {}

-    // Handles implicit conversion to a quat. Must not be explicit.
+    // handles implicit conversion to a quat. must not be explicit.
    template<typename A, typename = enable_if_arithmetic_t<A>>
-    constexpr TQuaternion(A w) : x(0), y(0), z(0), w(w) {} // NOLINT(google-explicit-constructor)
+    constexpr TQuaternion(A w) : x(0), y(0), z(0), w(w) {}

    // initialize from 4 values to w + xi + yj + zk
    template<typename A, typename B, typename C, typename D,
@@ -172,29 +175,31 @@ typedef details::TQuaternion<float> quatf;
 typedef details::TQuaternion<half> quath;

 constexpr inline quat operator "" _i(long double v) {
-    return { 0.0, double(v), 0.0, 0.0 };
+    return quat(0.0, double(v), 0.0, 0.0);
 }

 constexpr inline quat operator "" _j(long double v) {
-    return { 0.0, 0.0, double(v), 0.0 };
+    return quat(0.0, 0.0, double(v), 0.0);
 }

 constexpr inline quat operator "" _k(long double v) {
-    return { 0.0, 0.0, 0.0, double(v) };
+    return quat(0.0, 0.0, 0.0, double(v));
 }

 constexpr inline quat operator "" _i(unsigned long long v) {
-    return { 0.0, double(v), 0.0, 0.0 };
+    return quat(0.0, double(v), 0.0, 0.0);
 }

 constexpr inline quat operator "" _j(unsigned long long v) {
-    return { 0.0, 0.0, double(v), 0.0 };
+    return quat(0.0, 0.0, double(v), 0.0);
 }

 constexpr inline quat operator "" _k(unsigned long long v) {
-    return { 0.0, 0.0, 0.0, double(v) };
+    return quat(0.0, 0.0, 0.0, double(v));
 }

-}  // namespace filament::math
+// ----------------------------------------------------------------------------------------
+}  // namespace math
+}  // namespace filament

 #endif  // TNT_MATH_QUAT_H
@@ -36,6 +36,13 @@ public:
     * @see CallStack::capture()
     */
    CallStack() = default;
+    template <unsigned long N>
+    explicit CallStack(const std::array<intptr_t, N>& symbols)
+        : m_frame_count(N) {
+        for (size_t i = 0; i < N; i++) {
+            m_stack[i] = symbols[i];
+        }
+    }
    CallStack(const CallStack&) = default;
    ~CallStack() = default;

@@ -114,12 +121,10 @@ private:

    static constexpr size_t NUM_FRAMES = 20;

-    struct StackFrameInfo {
-        intptr_t pc;
-    };
+    using StackFrameInfo = intptr_t;

    size_t m_frame_count = 0;
-    StackFrameInfo m_stack[NUM_FRAMES];
+    StackFrameInfo m_stack[NUM_FRAMES] = {0};
 };

 } // namespace utils
@@ -70,7 +70,7 @@ intptr_t CallStack::operator[](size_t index) const {
 #endif
        std::abort();
    }
-    return m_stack[index].pc;
+    return m_stack[index];
 }

 size_t CallStack::getFrameCount() const noexcept {
@@ -91,7 +91,7 @@ void CallStack::update_gcc(size_t ignore) noexcept {
    size -= ignore;
 #endif
    for (ssize_t i = 0; i < size; i++) {
-        m_stack[i].pc = intptr_t(array[ignore + i]);
+        m_stack[i] = intptr_t(array[ignore + i]);
    }
    size--; // the last one seems to always be 0x0

@@ -362,13 +362,19 @@ static int parse(jsmntok_t const* tokens, int i, const char* jsonChunk,
 static int parse(jsmntok_t const* tokens, int i, const char* jsonChunk,
        LightManager::ShadowOptions* out) {
    CHECK_TOKTYPE(tokens[i], JSMN_OBJECT);
-    int const size = tokens[i++].size;
+    int size = tokens[i++].size;
    math::float3 splitsVector;
    for (int j = 0; j < size; ++j) {
        const jsmntok_t tok = tokens[i];
        CHECK_KEY(tok);
        if (compare(tok, jsonChunk, "mapSize") == 0) {
            i = parse(tokens, i + 1, jsonChunk, &out->mapSize);
+        } else if (compare(tok, jsonChunk, "stable") == 0) {
+            i = parse(tokens, i + 1, jsonChunk, &out->stable);
+        } else if (compare(tok, jsonChunk, "lispsm") == 0) {
+            i = parse(tokens, i + 1, jsonChunk, &out->lispsm);
+        } else if (compare(tok, jsonChunk, "screenSpaceContactShadows") == 0) {
+            i = parse(tokens, i + 1, jsonChunk, &out->screenSpaceContactShadows);
        } else if (compare(tok, jsonChunk, "shadowCascades") == 0) {
            i = parse(tokens, i + 1, jsonChunk, &out->shadowCascades);
        } else if (compare(tok, jsonChunk, "cascadeSplitPositions") == 0) {
@@ -376,27 +382,8 @@ static int parse(jsmntok_t const* tokens, int i, const char* jsonChunk,
            out->cascadeSplitPositions[0] = splitsVector[0];
            out->cascadeSplitPositions[1] = splitsVector[1];
            out->cascadeSplitPositions[2] = splitsVector[2];
-        // TODO: constantBias
-        // TODO: normalBias
-        // TODO: shadowFar
-        // TODO: shadowNearHint
-        // TODO: shadowFarHint
-        } else if (compare(tok, jsonChunk, "stable") == 0) {
-            i = parse(tokens, i + 1, jsonChunk, &out->stable);
-        } else if (compare(tok, jsonChunk, "lispsm") == 0) {
-            i = parse(tokens, i + 1, jsonChunk, &out->lispsm);
-        // TODO: polygonOffsetConstant
-        // TODO: polygonOffsetSlope
-        } else if (compare(tok, jsonChunk, "screenSpaceContactShadows") == 0) {
-            i = parse(tokens, i + 1, jsonChunk, &out->screenSpaceContactShadows);
-        // TODO: stepCount
-        // TODO: maxShadowDistance
        } else if (compare(tok, jsonChunk, "vsm") == 0) {
            i = parse(tokens, i + 1, jsonChunk, &out->vsm);
-        } else if (compare(tok, jsonChunk, "shadowBulbRadius") == 0) {
-            i = parse(tokens, i + 1, jsonChunk, &out->shadowBulbRadius);
-        } else if (compare(tok, jsonChunk, "transform") == 0) {
-            i = parse(tokens, i + 1, jsonChunk, &out->transform.xyzw);
        } else {
            slog.w << "Invalid shadow options key: '" << STR(tok, jsonChunk) << "'" << io::endl;
            i = parse(tokens, i + 1);
@@ -578,9 +565,11 @@ void applySettings(Engine* engine, const LightSettings& settings, IndirectLight*
        ibl->setRotation(math::mat3f::rotation(settings.iblRotation, math::float3 { 0, 1, 0 }));
    }
    for (size_t i = 0; i < sceneLightCount; i++) {
-        auto const li = lm->getInstance(sceneLights[i]);
-        lm->setShadowCaster(li, settings.enableShadows);
-        lm->setShadowOptions(li, settings.shadowOptions);
+        light = lm->getInstance(sceneLights[i]);
+        if (lm->isSpotLight(light)) {
+            lm->setShadowCaster(light, settings.enableShadows);
+        }
+        lm->setShadowOptions(light, settings.shadowOptions);
    }
    view->setSoftShadowOptions(settings.softShadowOptions);
 }
@@ -723,20 +712,18 @@ static std::ostream& operator<<(std::ostream& out, const ColorGradingSettings& i

 static std::ostream& operator<<(std::ostream& out, const LightManager::ShadowOptions& in) {
    const float* splits = in.cascadeSplitPositions;
-    math::float3 const splitsVector = { splits[0], splits[1], splits[2] };
+    math::float3 splitsVector = { splits[0], splits[1], splits[2] };
    return out << "{\n"
        << "\"vsm\": {\n"
        << "\"elvsm\": " << to_string(in.vsm.elvsm) << ",\n"
        << "\"blurWidth\": " << in.vsm.blurWidth << "\n"
        << "},\n"
        << "\"mapSize\": " << in.mapSize << ",\n"
-        << "\"shadowCascades\": " << int(in.shadowCascades) << ",\n"
-        << "\"cascadeSplitPositions\": " << (splitsVector) << "\n"
        << "\"stable\": " << to_string(in.stable) << ",\n"
        << "\"lispsm\": " << to_string(in.lispsm) << ",\n"
        << "\"screenSpaceContactShadows\": " << to_string(in.screenSpaceContactShadows) << ",\n"
-        << "\"shadowBulbRadius\": " << in.shadowBulbRadius << ",\n"
-        << "\"transform\": " << in.transform.xyzw << ",\n"
+        << "\"shadowCascades\": " << int(in.shadowCascades) << ",\n"
+        << "\"cascadeSplitPositions\": " << (splitsVector) << "\n"
        << "}";
 }

@@ -249,8 +249,6 @@ int parse(jsmntok_t const* tokens, int i, const char* jsonChunk, BloomOptions* o
            i = parse(tokens, i + 1, jsonChunk, &out->enabled);
        } else if (compare(tok, jsonChunk, "highlight") == 0) {
            i = parse(tokens, i + 1, jsonChunk, &out->highlight);
-        } else if (compare(tok, jsonChunk, "quality") == 0) {
-            i = parse(tokens, i + 1, jsonChunk, &out->quality);
        } else if (compare(tok, jsonChunk, "lensFlare") == 0) {
            i = parse(tokens, i + 1, jsonChunk, &out->lensFlare);
        } else if (compare(tok, jsonChunk, "starburst") == 0) {
@@ -293,7 +291,6 @@ std::ostream& operator<<(std::ostream& out, const BloomOptions& in) {
        << "\"threshold\": " << to_string(in.threshold) << ",\n"
        << "\"enabled\": " << to_string(in.enabled) << ",\n"
        << "\"highlight\": " << (in.highlight) << ",\n"
-        << "\"quality\": " << (in.quality) << ",\n"
        << "\"lensFlare\": " << to_string(in.lensFlare) << ",\n"
        << "\"starburst\": " << to_string(in.starburst) << ",\n"
        << "\"chromaticAberration\": " << (in.chromaticAberration) << ",\n"
@@ -777,10 +777,6 @@ void ViewerGui::updateUserInterface() {
        ImGui::SliderInt("Levels", &levels, 3, 11);
        mSettings.view.bloom.levels = levels;

-        int quality = (int) mSettings.view.bloom.quality;
-        ImGui::SliderInt("Bloom Quality", &quality, 0, 3);
-        mSettings.view.bloom.quality = (View::QualityLevel) quality;
-
        ImGui::Checkbox("Lens Flare", &mSettings.view.bloom.lensFlare);
    }

@@ -857,14 +853,6 @@ void ViewerGui::updateUserInterface() {
            ImGui::SliderFloat("Halo falloff", &light.sunlightHaloFalloff, 4.0f, 1024.0f);
            ImGui::SliderFloat("Sun radius", &light.sunlightAngularRadius, 0.1f, 10.0f);
            ImGuiExt::DirectionWidget("Sun direction", light.sunlightDirection.v);
-
-            float3 shadowDirection = light.shadowOptions.transform * light.sunlightDirection;
-            ImGuiExt::DirectionWidget("Shadow direction", shadowDirection.v);
-            light.shadowOptions.transform = normalize(quatf{
-                    cross(light.sunlightDirection, shadowDirection),
-                    sqrt(length2(light.sunlightDirection) * length2(shadowDirection))
-                            + dot(light.sunlightDirection, shadowDirection)
-            });
        }
        if (ImGui::CollapsingHeader("All lights")) {
            ImGui::Checkbox("Enable shadows", &light.enableShadows);
@@ -101,7 +101,6 @@ struct App {

    bool actualSize = false;
    bool originIsFarAway = false;
-    float originDistance = 6378137; // Earth's radius in [m]

    struct Scene {
        Entity groundPlane;
@@ -760,7 +759,6 @@ int main(int argc, char** argv) {
                ImGui::Checkbox("Disable buffer padding", debug.getPropertyAddress<bool>("d.renderer.disable_buffer_padding"));
                ImGui::Checkbox("Camera at origin", debug.getPropertyAddress<bool>("d.view.camera_at_origin"));
                ImGui::Checkbox("Far Origin", &app.originIsFarAway);
-                ImGui::SliderFloat("Origin", &app.originDistance, 0, 10000000);
                auto dataSource = debug.getDataSource("d.view.frame_info");
                if (dataSource.data) {
                    ImGuiExt::PlotLinesSeries("FrameInfo", 6,
@@ -958,7 +956,7 @@ int main(int argc, char** argv) {
        tcm.setParent(tcm.getInstance(camera.getEntity()), root);
        tcm.setParent(tcm.getInstance(app.asset->getRoot()), root);
        tcm.setParent(tcm.getInstance(view->getFogEntity()), root);
-        tcm.setTransform(root, mat4f::translation(float3{ app.originIsFarAway ? app.originDistance : 0.0f }));
+        tcm.setTransform(root, mat4f::translation(float3{ app.originIsFarAway ? 1e6f : 0.0f }));

        // Check if color grading has changed.
        ColorGradingSettings& options = app.viewer->getSettings().view.colorGrading;
@@ -109,7 +109,7 @@ highp vec4 getSpotLightSpacePosition(int index, highp vec3 dir, highp float zLig
    // for spotlights, the bias depends on z
    float bias = shadowUniforms.shadows[index].normalBias * zLight;

-    return computeLightSpacePosition(getWorldPosition(), getWorldGeometricNormalVector(),
+    return computeLightSpacePosition(getWorldPosition(), getWorldNormalVector(),
            dir, bias, lightFromWorldMatrix);
 }
 #endif
@@ -141,7 +141,7 @@ highp vec4 getCascadeLightSpacePosition(int cascade) {
        return vertex_lightSpacePosition;
    }

-    return computeLightSpacePosition(getWorldPosition(), getWorldGeometricNormalVector(),
+    return computeLightSpacePosition(getWorldPosition(), getWorldNormalVector(),
        frameUniforms.lightDirection,
        shadowUniforms.shadows[cascade].normalBias,
        shadowUniforms.shadows[cascade].lightFromWorldMatrix);
@@ -45,8 +45,8 @@ end

 Then run:

-```shell
-pod install
+```
+$ pod install
 ```

 Close the project and then re-open the newly created HelloCocoaPods.xcworkspace file.
--- a/Show More
+++ b/Show More