Fixup shader code

* Change the WebGPU wrap modes

.clang-format

@@ -57,7 +57,7 @@ SpaceBeforeInheritanceColon: true
 SpaceBeforeParens: ControlStatements
 SpaceBeforeRangeBasedForLoopColon: false
 SpaceInEmptyParentheses: false
-SpacesBeforeTrailingComments: 0
+SpacesBeforeTrailingComments: 1
 SpacesInAngles: false
 SpacesInCStyleCastParentheses: false
 SpacesInContainerLiterals: false

.github/actions/get-commit-msg/action.yml

@@ -0,0 +1,43 @@
+name: 'Get commit message'
+outputs:
+  msg:
+    value: ${{ steps.action_output.outputs.msg }}
+runs:
+  using: "composite"
+  steps:
+    - name: Find commit message (on push)
+      if: github.event_name == 'push'
+      shell: bash
+      run: |
+        AUTHOR_NAME="${{ github.event.head_commit.author.name }}"
+        AUTHOR_EMAIL="${{ github.event.head_commit.author.email }}"
+        TSTAMP="${{ github.event.head_commit.timestamp }}"
+        echo "commit ${{ github.event.head_commit.id }}" >> /tmp/commit_msg.txt
+        echo "Author: ${AUTHOR_NAME}<${AUTHOR_EMAIL}>" >> /tmp/commit_msg.txt
+        echo "Date: ${TSTAMP}" >> /tmp/commit_msg.txt
+        echo "" >> /tmp/commit_msg.txt
+        echo "${{ github.event.head_commit.message }}" >> /tmp/commit_msg.txt
+    - name: Checkout code
+      shell: bash
+      id: checkout_code
+      if: github.event_name == 'pull_request'
+      run: |
+        echo "hash=$(git rev-parse HEAD)" >> "$GITHUB_OUTPUT"
+        git checkout ${{ github.event.pull_request.head.ref }}
+        COMMIT_MESSAGE=$(git log -1)
+        echo "$(git log -1)" >> /tmp/commit_msg.txt
+    - shell: bash
+      id: action_output
+      run: |
+        DELIMITER="EOF_FILE_CONTENT_$(date +%s)" # Using timestamp to make it more unique
+        echo "msg<<$DELIMITER" >> "$GITHUB_OUTPUT"
+        cat /tmp/commit_msg.txt >> "$GITHUB_OUTPUT"
+        echo "$DELIMITER" >> "$GITHUB_OUTPUT"
+        echo "----- got commit message ---"
+        cat /tmp/commit_msg.txt
+        echo "----------------------------"
+    - name: Uncheckout code
+      shell: bash
+      if: github.event_name == 'pull_request'
+      run: |
+        git checkout ${{ steps.checkout_code.outputs.hash }}

.github/workflows/postsubmit.yml

@@ -0,0 +1,35 @@
+name: 'Post-submit tasks'
+
+on:
+  push:
+    branches:
+      - main
+
+jobs:
+  update-renderdiff-goldens:
+    name: update-renderdiff-goldens
+    runs-on: 'ubuntu-24.04-4core'
+    steps:
+      - uses: actions/checkout@v4.1.6
+        with:
+          fetch-depth: 0
+      - uses: ./.github/actions/linux-prereq
+      - id: get_commit_msg
+        uses: ./.github/actions/get-commit-msg
+      - name: Prerequisites
+        run: pip install tifffile numpy
+      - name: Run update script
+        env:
+          GH_TOKEN: ${{ secrets.FILAMENTBOT_TOKEN }}
+        run: |
+          GOLDEN_BRANCH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | python3 test/renderdiff/src/commit_msg.py)
+          COMMIT_HASH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | head -n 1 | tr -d 'commit ')
+          if [[ "${GOLDEN_BRANCH}" != "main" ]]; then
+              git config --global user.email "filament.bot@gmail.com"
+              git config --global user.name "Filament Bot"
+              git config --global credential.helper cache
+              echo "branch==${GOLDEN_BRANCH}"
+              echo "hash==${COMMIT_HASH}"
+              python3 test/renderdiff/src/update_golden.py --branch=${GOLDEN_BRANCH} \
+                  --merge-to-main --filament-tag=${COMMIT_HASH} --golden-repo-token=${GH_TOKEN}
+          fi

.github/workflows/presubmit.yml

@@ -3,10 +3,10 @@ name: Presubmit
 on:
   push:
     branches:
-    - main
+      - main
   pull_request:
     branches:
-    - main
+      - main
 
 jobs:
   build-desktop-mac:
@@ -41,8 +41,7 @@ jobs:
 
   build-windows:
     name: build-windows
-    runs-on: win-2019-16core
-
+    runs-on: windows-2022-32core
     steps:
       - uses: actions/checkout@v4.1.6
         with:
@@ -110,31 +109,35 @@ jobs:
           fetch-depth: 0
       - name: Check for manual edits to /docs
         run: |
-           echo "${{ github.event.pull_request.head.sha }} -- ${{ github.event.pull_request.head.sha }}"
-           # disable for now
-           # bash docs_src/build/presubmit_check.sh ${{ github.event.pull_request.head.sha }}
+          echo "${{ github.event.pull_request.head.sha }} -- ${{ github.event.pull_request.head.sha }}"
+          # disable for now
+          # bash docs_src/build/presubmit_check.sh ${{ github.event.pull_request.head.sha }}
 
   test-renderdiff:
     name: test-renderdiff
     runs-on: macos-14-xlarge
     steps:
+      - uses: actions/checkout@v4.1.6
+        with:
+          fetch-depth: 0
+      - id: get_commit_msg
+        uses: ./.github/actions/get-commit-msg
       - uses: actions/checkout@v4.1.6
         with:
           fetch-depth: 0
       - uses: ./.github/actions/mac-prereq
       - name: Cache Mesa and deps
-        id: mesa-cache
         uses: actions/cache@v4
         with:
           path: mesa
           key: ${{ runner.os }}-mesa-deps-2-${{ vars.MESA_VERSION }}
       - name: Prerequisites
-        id: prereqs
         run: |
           bash build/common/get-mesa.sh
           pip install tifffile numpy
       - name: Run Test
-        run: bash test/renderdiff/test.sh
+        run: |
+          echo "${{ steps.get_commit_msg.outputs.msg }}" | bash test/renderdiff/test.sh
       - uses: actions/upload-artifact@v4
         with:
           name: presubmit-renderdiff-result

.github/workflows/release.yml

@@ -226,7 +226,7 @@ jobs:
 
   build-windows:
     name: build-windows
-    runs-on: windows-2019-32core
+    runs-on: windows-2022-32core
     if: github.event_name == 'release' || github.event.inputs.platform == 'windows'
 
     steps:

.github/workflows/windows-continuous.yml

@@ -10,7 +10,7 @@ on:
 jobs:
   build-windows:
     name: build-windows
-    runs-on: windows-2019-32core
+    runs-on: windows-2022-32core
 
     steps:
       - uses: actions/checkout@v4.1.6

BUILDING.md

@@ -97,6 +97,10 @@ Make sure you've installed the following dependencies:
 - `libxcomposite-dev` (`libXcomposite-devel` on Fedora)
 - `libxxf86vm-dev` (`libXxf86vm-devel` on Fedora)
 
+```shell
+sudo apt install clang-14 libglu1-mesa-dev libc++-14-dev libc++abi-14-dev ninja-build libxi-dev libxcomposite-dev libxxf86vm-dev -y
+```
+
 After dependencies have been installed, we highly recommend using the [easy build](#easy-build)
 script.
 

CMakeLists.txt

@@ -49,6 +49,10 @@ option(FILAMENT_SUPPORTS_OSMESA "Enable OSMesa (headless GL context) for Filamen
 
 option(FILAMENT_ENABLE_FGVIEWER "Enable the frame graph viewer" OFF)
 
+# This is to disable GTAO for the short-term while we investigate a way to better manage size increases.
+# On the regular filament build (where size is of less concern), we enable GTAO by default.
+option(FILAMENT_DISABLE_GTAO "Disable GTAO" OFF)
+
 set(FILAMENT_NDK_VERSION "" CACHE STRING
     "Android NDK version or version prefix to be used when building for Android."
 )

NEW_RELEASE_NOTES.md

@@ -1,9 +1,13 @@
 # Filament Release Notes log
 
 **If you are merging a PR into main**: please add the release note below, under the *Release notes
-for next branch cut* header.
+We are chaning the way Vulkan buffers are handled. We need to switch over to a managed (or view-based) model where the data stored inside the object is a proxy to a Vulkan object that can dynamically be swapped around.
 
 **If you are cherry-picking a commit into an rc/ branch**: add the release note under the
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).
 
 ## Release notes for next branch cut
+
+- samples: samples now have a CLI to select backend api
+- materials: sampler now export their type in the material binary [⚠️ **New Material Version**]
+- materials: new mutable specialization constants feature. See the [materials documentation](https://google.github.io/filament/Materials.html) for details. [⚠️ **New Material Version**]

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.60.0'
+    implementation 'com.google.android.filament:filament-android:1.61.0'
 }
 ```
 
@@ -51,7 +51,7 @@ Here are all the libraries available in the group `com.google.android.filament`:
 iOS projects can use CocoaPods to install the latest release:
 
 ```shell
-pod 'Filament', '~> 1.60.0'
+pod 'Filament', '~> 1.61.0'
 ```
 
 ## Documentation

RELEASE_NOTES.md

@@ -7,6 +7,15 @@ 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).
 
+## v1.61.1
+
+
+## v1.61.0
+
+- materials: sampler now export their type in the material binary [⚠️ **New Material Version**]
+- samples/texturedquad.cpp now has CLI to select backend api
+- samples/hellopbr.cpp CLI now allows for selecting webgpu
+
 ## v1.60.1
 
 

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

@@ -60,6 +60,14 @@ static void setParameter(JNIEnv* env, jlong nativeMaterialInstance, jstring name
     env->ReleaseStringUTFChars(name_, name);
 }
 
+template<typename T>
+static void setConstant(JNIEnv* env, jlong nativeMaterialInstance, jstring name_, T v) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    instance->setConstant(name, v);
+    env->ReleaseStringUTFChars(name_, name);
+}
+
 extern "C"
 JNIEXPORT void JNICALL
 Java_com_google_android_filament_MaterialInstance_nSetParameterBool(JNIEnv *env, jclass,
@@ -264,6 +272,13 @@ Java_com_google_android_filament_MaterialInstance_nSetParameterTexture(
     env->ReleaseStringUTFChars(name_, name);
 }
 
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_MaterialInstance_nSetConstantBool(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_, jboolean x) {
+    setConstant(env, nativeMaterialInstance, name_, bool(x));
+}
+
 extern "C"
 JNIEXPORT void JNICALL
 Java_com_google_android_filament_MaterialInstance_nSetScissor(

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

@@ -402,6 +402,16 @@ public class MaterialInstance {
         nSetParameterFloat4(getNativeObject(), name, color[0], color[1], color[2], color[3]);
     }
 
+    /**
+     * Sets the value of a bool constant.
+     *
+     * @param name the name of the material constant
+     * @param x    the value of the material constant
+     */
+    public void setConstant(@NonNull String name, boolean x) {
+        nSetConstantBool(getNativeObject(), name, x);
+    }
+
     /**
      * Set-up a custom scissor rectangle; by default it is disabled.
      *
@@ -921,6 +931,9 @@ public class MaterialInstance {
             @NonNull String name, int element, @NonNull @Size(min = 1) float[] v,
             @IntRange(from = 0) int offset, @IntRange(from = 1) int count);
 
+    private static native void nSetConstantBool(long nativeMaterialInstance,
+            @NonNull String name, boolean x);
+
     private static native void nSetParameterTexture(long nativeMaterialInstance,
             @NonNull String name, long nativeTexture, long sampler);
 

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

@@ -1801,6 +1801,22 @@ public class View {
      * @see setAmbientOcclusionOptions()
      */
     public static class AmbientOcclusionOptions {
+        public enum AmbientOcclusionType {
+            /**
+             * use Scalable Ambient Occlusion
+             */
+            SAO,
+            /**
+             * use Ground Truth-Based Ambient Occlusion
+             */
+            GTAO,
+        }
+
+        /**
+         * Type of ambient occlusion algorithm.
+         */
+        @NonNull
+        public AmbientOcclusionOptions.AmbientOcclusionType aoType = AmbientOcclusionOptions.AmbientOcclusionType.SAO;
         /**
          * Ambient Occlusion radius in meters, between 0 and ~10.
          */
@@ -1810,7 +1826,8 @@ public class View {
          */
         public float power = 1.0f;
         /**
-         * Self-occlusion bias in meters. Use to avoid self-occlusion. Between 0 and a few mm.
+         * Self-occlusion bias in meters. Use to avoid self-occlusion.
+         * Between 0 and a few mm. No effect when aoType set to GTAO
          */
         public float bias = 0.0005f;
         /**
@@ -1826,12 +1843,12 @@ public class View {
          */
         public float bilateralThreshold = 0.05f;
         /**
-         * affects # of samples used for AO.
+         * affects # of samples used for AO and params for filtering
          */
         @NonNull
         public QualityLevel quality = QualityLevel.LOW;
         /**
-         * affects AO smoothness
+         * affects AO smoothness. Recommend setting to HIGH when aoType set to GTAO.
          */
         @NonNull
         public QualityLevel lowPassFilter = QualityLevel.MEDIUM;
@@ -1849,7 +1866,7 @@ public class View {
          */
         public boolean bentNormals = false;
         /**
-         * min angle in radian to consider
+         * min angle in radian to consider. No effect when aoType set to GTAO.
          */
         public float minHorizonAngleRad = 0.0f;
         /**
@@ -1904,6 +1921,19 @@ public class View {
          */
         public boolean ssctEnabled = false;
 
+        /**
+         * Ground Truth-base Ambient Occlusion (GTAO) options
+         */
+        public int gtaoSampleSliceCount = 4;
+        /**
+         * Ground Truth-base Ambient Occlusion (GTAO) options
+         */
+        public int gtaoSampleStepsPerSlice = 3;
+        /**
+         * Ground Truth-base Ambient Occlusion (GTAO) options
+         */
+        public float gtaoThicknessHeuristic = 0.004f;
+
     }
 
     /**

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.60.0
+VERSION_NAME=1.61.0
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

build/common/get-mesa.sh

@@ -41,6 +41,8 @@ for cmd in "${NEEDED_PYTHON_DEPS[@]}"; do
 done
 deactivate
 
+LOCAL_PKG_CONFIG_PATH=
+
 # Install system deps
 if [[ "$OS_NAME" == "Linux" ]]; then
     if [[ "$GITHUB_WORKFLOW" ]]; then
@@ -82,6 +84,9 @@ elif [[ "$OS_NAME" == "Darwin" ]]; then
         fi
     fi
     HOMEBREW_NO_INSTALLED_DEPENDENTS_CHECK=true brew install autoconf automake libx11 libxext libxrandr llvm@${LLVM_VERSION} ninja meson pkg-config libxshmfence
+
+    # For reasons unknown, this is necessary for pkg-config to find homebrew's packages
+    LOCAL_PKG_CONFIG_PATH="/opt/homebrew/lib/pkgconfig:$PKG_CONFIG_PATH"
 fi # [[ "$OS_NAME" == x ]]
 
 LOCAL_LDFLAGS=${LDFLAGS}
@@ -134,9 +139,11 @@ fi
 # -Dgallium-drivers=swrast  => builds GL software rasterizer
 # -Dvulkan-drivers=swrast   => builds VK software rasterizer
 # -Dgallium-drivers=llvmpipe is needed for GL >= 4.1 pipe-screen (see src/gallium/auxiliary/target-helpers/inline_sw_helper.h)
-CXX=${LOCAL_CXX} CC=${LOCAL_CC} PATH=${LOCAL_PATH} LDFLAGS=${LOCAL_LDFLAGS} CPPFLAGS=${LOCAL_CPPFLAGS} \
+PKG_CONFIG_PATH=${LOCAL_PKG_CONFIG_PATH} PATH=${LOCAL_PATH} \
+CXX=${LOCAL_CXX} CC=${LOCAL_CC} LDFLAGS=${LOCAL_LDFLAGS} CPPFLAGS=${LOCAL_CPPFLAGS} \
    meson setup --wipe builddir/ -Dprefix="${MESA_DIR}/out" -Dglx=xlib -Dosmesa=true -Dgallium-drivers=llvmpipe,swrast -Dvulkan-drivers=swrast
-CXX=${LOCAL_CXX} CC=${LOCAL_CC} PATH=${LOCAL_PATH} LDFLAGS=${LOCAL_LDFLAGS} CPPFLAGS=${LOCAL_CPPFLAGS} \
+PKG_CONFIG_PATH=${LOCAL_PKG_CONFIG_PATH} PATH=${LOCAL_PATH} \
+CXX=${LOCAL_CXX} CC=${LOCAL_CC} LDFLAGS=${LOCAL_LDFLAGS} CPPFLAGS=${LOCAL_CPPFLAGS} \
    meson install -C builddir/
 
 # Disable python venv

build/windows/build-github.bat

@@ -47,7 +47,11 @@ if "%RUNNING_LOCALLY%" == "1" (
     set "PATH=%PATH%;C:\Program Files\7-Zip"
 )
 
-call "C:\Program Files (x86)\Microsoft Visual Studio\2019\%VISUAL_STUDIO_VERSION%\VC\Auxiliary\Build\vcvars64.bat"
+:: Outdated windows-2019 pattern
+:: call "C:\Program Files (x86)\Microsoft Visual Studio\2019\%VISUAL_STUDIO_VERSION%\VC\Auxiliary\Build\vcvars64.bat"
+
+call "C:\Program Files\Microsoft Visual Studio\2022\%VISUAL_STUDIO_VERSION%\VC\Auxiliary\Build\vcvars64.bat"
+echo Passed vcvars64.bat
 if errorlevel 1 exit /b %errorlevel%
 
 msbuild /version
@@ -107,7 +111,7 @@ cd out\cmake-%variant%
 if errorlevel 1 exit /b %errorlevel%
 
 cmake ..\.. ^
-    -G "Visual Studio 16 2019" ^
+    -G "Visual Studio 17 2022" ^
     -A x64 ^
     %flag% ^
     -DCMAKE_INSTALL_PREFIX=..\%variant% ^

docs_src/markdeep/Materials.md.html

@@ -211,7 +211,7 @@ when using textures.
 This property can dramatically change the appearance of a surface. Non-metallic surfaces have
 chromatic diffuse reflection and achromatic specular reflection (reflected light does not change
 color). Metallic surfaces do not have any diffuse reflection and chromatic specular reflection
-(reflected light takes on the color of the surfaced as defined by `baseColor`). 
+(reflected light takes on the color of the surfaced as defined by `baseColor`).
 
 The effect of `metallic` is shown in figure [metallicProperty] (click on the image to see a
 larger version).
@@ -247,7 +247,7 @@ The effect of `roughness` on metallic surfaces is shown in figure [roughnessCond
 When refraction through an object is enabled (using a `refractonType` of `thin` or `solid`), the
 `roughness` property will also affect the refractions, as shown in figure
 [roughnessRefractionProperty] (click on the image to see a larger version).
-                             
+
 ![Figure [roughnessRefractionProperty]: Refractive sphere with `roughness` varying from 0.0
  (left) to 1.0 (right)](images/materials/refraction_roughness.png)
 
@@ -306,7 +306,7 @@ The sheen color controls the color appearance and strength of an optional sheen
 base layer described by the properties above. The sheen layer always sits below the clear coat layer
 if such a layer is present.
 
-The sheen layer can be used to represent cloth and fabric materials. Please refer to 
+The sheen layer can be used to represent cloth and fabric materials. Please refer to
 section [Cloth model] for more information about cloth and fabric materials.
 
 The effect of `sheenColor` is shown in figure [materialSheenColor]
@@ -519,14 +519,14 @@ light to bend further away from the initial path.
 
 Table [commonMatIOR] describes acceptable refractive indices for various types of materials.
 
-Material                   |        IOR       
+Material                   |        IOR
 --------------------------:|:-----------------
 Air                        | 1.0
-Water                      | 1.33             
-Common liquids             | 1.33 to 1.5      
-Common gemstones           | 1.58 to 2.33     
-Plastics, glass            | 1.5 to 1.58      
-Other dielectric materials | 1.33 to 1.58     
+Water                      | 1.33
+Common liquids             | 1.33 to 1.5
+Common gemstones           | 1.58 to 2.33
+Plastics, glass            | 1.5 to 1.58
+Other dielectric materials | 1.33 to 1.58
 [Table [commonMatIOR]: Index of refraction of common materials]
 
 The appearance of a refractive material will greatly depend on the `refractionType` and
@@ -1091,30 +1091,37 @@ Value
     `bool` or `number`, depending on the `type` of the constant. The type must be one of the types
     described in table [materialConstantsTypes].
 
-         Type          |            Description                   |      Default
-:----------------------|:-----------------------------------------|:------------------
-int                    | A signed, 32 bit GLSL int                |         0
-float                  | A single-precision GLSL float            |         0.0
-bool                   | A GLSL bool                              |         false
+    Constants may also be specified as mutable by setting the `mutable` property to `true`. Only
+    `bool` spec constants may be specified as mutable.
+
+| Type  | Description                   | Default | May be mutable? |
+|:------|:------------------------------|:--------|-----------------|
+| int   | A signed, 32 bit GLSL int     |         | no              |
+| float | A single-precision GLSL float | 0.0     | no              |
+| bool  | A GLSL bool                   | false   | yes             |
 [Table [materialConstantsTypes]: Material constants types]
 
 Description
 :   Lists the constant parameters accepted by your material. These constants can be set, or
     "specialized", at runtime when loading a material package. Multiple materials can be loaded from
-    the same material package with differing constant parameter specializations. Once a material is
-    loaded from a material package, its constant parameters cannot be changed. Compared to regular
-    parameters, constant parameters allow the compiler to generate more efficient code. Access
-    constant parameters from the shader by prefixing the name with `materialConstant_`. For example,
-    a constant parameter named `myConstant` is accessed in the shader as
-    `materialConstant_myConstant`. If a constant parameter is not set at runtime, the default is
-    used.
+    the same material package with differing constant parameter specializations. If a constant
+    parameter is specialized as mutable, it may be changed at any time on a per-instance basis via
+    `MaterialInstance::setConstant(name, value)`. Otherwise, once a material is loaded from a
+    material package, its constant parameters cannot be changed.
+
+    Compared to regular parameters, constant parameters allow the compiler to generate more
+    efficient code. Access constant parameters from the shader by prefixing the name with
+    `materialConstant_`. For example, a constant parameter named `myConstant` is accessed in the
+    shader as `materialConstant_myConstant`. If a constant parameter is not set at runtime, the
+    default is used.
 
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ JSON
 material {
     constants : [
         {
            name : overrideAlpha,
-           type : bool
+           type : bool,
+           mutable : true,
         },
         {
            name : customAlpha,
@@ -2553,7 +2560,7 @@ standard skybox material. It produces a list of 2 parameters, named `showSun` an
 respectively a boolean and a cubemap texture.
 
 ```text
-$ matc --reflect parameters filament/src/materials/skybox.mat 
+$ matc --reflect parameters filament/src/materials/skybox.mat
 {
   "parameters": [
     {
@@ -2569,7 +2576,7 @@ $ matc --reflect parameters filament/src/materials/skybox.mat
     }
   ]
 }
-``` 
+```
 
 ### --variant-filter
 

filament/CMakeLists.txt

@@ -265,6 +265,13 @@ set(MATERIAL_SRCS
         src/materials/vsmMipmap.mat
 )
 
+if (NOT FILAMENT_DISABLE_GTAO)
+    list(APPEND MATERIAL_SRCS
+        src/materials/ssao/gtao.mat
+        src/materials/ssao/gtaoBentNormals.mat
+    )
+endif()
+
 set(MATERIAL_FL0_SRCS
         src/materials/defaultMaterial.mat
         src/materials/skybox.mat
@@ -314,12 +321,14 @@ if (FILAMENT_FORCE_PROFILING_MODE)
     add_definitions(-DFILAMENT_FORCE_PROFILING_MODE)
 endif()
 
+if (FILAMENT_DISABLE_GTAO)
+    add_definitions(-DFILAMENT_DISABLE_GTAO)
+endif()
+
 # ==================================================================================================
 # Definitions
 # ==================================================================================================
 
-# "2" corresponds to SYSTRACE_TAG_FILEMENT (See: utils/Systrace.h)
-add_definitions(-DSYSTRACE_TAG=2)
 add_definitions(-DFILAMENT_DFG_LUT_SIZE=${DFG_LUT_SIZE})
 add_definitions(
     -DFILAMENT_PER_RENDER_PASS_ARENA_SIZE_IN_MB=${FILAMENT_PER_RENDER_PASS_ARENA_SIZE_IN_MB}
@@ -463,6 +472,30 @@ add_custom_command(
         APPEND
 )
 
+if (NOT FILAMENT_DISABLE_GTAO)
+    add_custom_command(
+            OUTPUT "${MATERIAL_DIR}/gtao.filamat"
+            DEPENDS src/materials/ssao/ssaoUtils.fs
+            DEPENDS src/materials/ssao/ssct.fs
+            DEPENDS src/materials/utils/depthUtils.fs
+            DEPENDS src/materials/utils/geometry.fs
+            DEPENDS src/materials/ssao/gtaoImpl.fs
+            DEPENDS src/materials/ssao/ssctImpl.fs
+            APPEND
+    )
+
+    add_custom_command(
+            OUTPUT "${MATERIAL_DIR}/gtaoBentNormals.filamat"
+            DEPENDS src/materials/ssao/ssaoUtils.fs
+            DEPENDS src/materials/ssao/ssct.fs
+            DEPENDS src/materials/utils/depthUtils.fs
+            DEPENDS src/materials/utils/geometry.fs
+            DEPENDS src/materials/ssao/gtaoImpl.fs
+            DEPENDS src/materials/ssao/ssctImpl.fs
+            APPEND
+    )
+endif()
+
 add_custom_command(
         OUTPUT "${MATERIAL_DIR}/bilateralBlur.filamat"
         DEPENDS src/materials/ssao/ssaoUtils.fs

filament/backend/CMakeLists.txt

@@ -180,8 +180,11 @@ if (FILAMENT_SUPPORTS_VULKAN)
             src/vulkan/VulkanAsyncHandles.h
             src/vulkan/VulkanBlitter.cpp
             src/vulkan/VulkanBlitter.h
-            src/vulkan/VulkanBuffer.cpp
             src/vulkan/VulkanBuffer.h
+            src/vulkan/VulkanBufferCache.h
+            src/vulkan/VulkanBufferCache.cpp
+            src/vulkan/VulkanBufferProxy.h
+            src/vulkan/VulkanBufferProxy.cpp
             src/vulkan/VulkanCommands.cpp
             src/vulkan/VulkanCommands.h
             src/vulkan/VulkanConstants.h
@@ -254,16 +257,36 @@ if (FILAMENT_SUPPORTS_WEBGPU)
     list(APPEND SRCS
             include/backend/platforms/WebGPUPlatform.h
             src/webgpu/platform/WebGPUPlatform.cpp
+            src/webgpu/WebGPUBufferBase.cpp
+            src/webgpu/WebGPUBufferBase.h
+            src/webgpu/WebGPUBufferObject.cpp
+            src/webgpu/WebGPUBufferObject.h
             src/webgpu/WebGPUConstants.h
+            src/webgpu/WebGPUDescriptorSet.cpp
+            src/webgpu/WebGPUDescriptorSet.h
+            src/webgpu/WebGPUDescriptorSetLayout.cpp
+            src/webgpu/WebGPUDescriptorSetLayout.h
             src/webgpu/WebGPUDriver.cpp
             src/webgpu/WebGPUDriver.h
-            src/webgpu/WebGPUHandles.cpp
-            src/webgpu/WebGPUHandles.h
+            src/webgpu/WebGPUIndexBuffer.cpp
+            src/webgpu/WebGPUIndexBuffer.h
             src/webgpu/WebGPUPipelineCreation.cpp
             src/webgpu/WebGPUPipelineCreation.h
+            src/webgpu/WebGPUProgram.cpp
+            src/webgpu/WebGPUProgram.h
+            src/webgpu/WebGPURenderPrimitive.h
+            src/webgpu/WebGPURenderTarget.cpp
+            src/webgpu/WebGPURenderTarget.h
+            src/webgpu/WebGPUStrings.h
             src/webgpu/WebGPUSwapChain.cpp
             src/webgpu/WebGPUSwapChain.h
-            src/webgpu/WGPUProgram.cpp
+            src/webgpu/WebGPUTexture.cpp
+            src/webgpu/WebGPUTexture.h
+            src/webgpu/WebGPUVertexBuffer.cpp
+            src/webgpu/WebGPUVertexBuffer.h
+            src/webgpu/WebGPUVertexBufferInfo.cpp
+            src/webgpu/WebGPUVertexBufferInfo.h
+            src/webgpu/SpdMipmapGenerator.cpp
     )
     if (WIN32)
         list(APPEND SRCS src/webgpu/platform/WebGPUPlatformWindows.cpp)
@@ -290,12 +313,6 @@ if (ANDROID)
     list(APPEND SRCS src/BackendUtilsAndroid.cpp)
 endif()
 
-# ==================================================================================================
-# Definitions
-# ==================================================================================================
-# "2" corresponds to SYSTRACE_TAG_FILEMENT (See: utils/Systrace.h)
-add_definitions(-DSYSTRACE_TAG=2 )
-
 # ==================================================================================================
 # Includes & target definition
 # ==================================================================================================

filament/backend/include/backend/DriverEnums.h

@@ -36,6 +36,7 @@
 #include <array>
 #include <type_traits>
 #include <variant>
+#include <string_view>
 
 #include <stddef.h>
 #include <stdint.h>
@@ -162,7 +163,7 @@ enum class TimerQueryResult : int8_t {
     AVAILABLE = 1,  // result is available
 };
 
-static constexpr const char* backendToString(Backend backend) {
+constexpr std::string_view to_string(Backend const backend) noexcept {
     switch (backend) {
         case Backend::NOOP:
             return "Noop";
@@ -174,9 +175,10 @@ static constexpr const char* backendToString(Backend backend) {
             return "Metal";
         case Backend::WEBGPU:
             return "WebGPU";
-        default:
-            return "Unknown";
+        case Backend::DEFAULT:
+            return "Default";
     }
+    return "Unknown";
 }
 
 /**
@@ -193,7 +195,7 @@ enum class ShaderLanguage {
     WGSL = 5,
 };
 
-static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguage) {
+constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguage) noexcept {
     switch (shaderLanguage) {
         case ShaderLanguage::ESSL1:
             return "ESSL 1.0";
@@ -208,6 +210,7 @@ static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguag
         case ShaderLanguage::WGSL:
             return "WGSL";
     }
+    return "UNKNOWN";
 }
 
 enum class ShaderStage : uint8_t {
@@ -225,7 +228,7 @@ enum class ShaderStageFlags : uint8_t {
     ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
 };
 
-static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
+constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
     switch (type) {
         case ShaderStage::VERTEX:
             return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
@@ -236,14 +239,239 @@ static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage t
     }
 }
 
-enum class DescriptorType : uint8_t {
-    UNIFORM_BUFFER,
-    SHADER_STORAGE_BUFFER,
-    SAMPLER,
-    INPUT_ATTACHMENT,
-    SAMPLER_EXTERNAL
+enum class TextureType : uint8_t {
+    FLOAT,
+    INT,
+    UINT,
+    DEPTH,
+    STENCIL,
+    DEPTH_STENCIL
 };
 
+constexpr std::string_view to_string(TextureType type) noexcept {
+    switch (type) {
+        case TextureType::FLOAT:            return "FLOAT";
+        case TextureType::INT:              return "INT";
+        case TextureType::UINT:             return "UINT";
+        case TextureType::DEPTH:            return "DEPTH";
+        case TextureType::STENCIL:          return "STENCIL";
+        case TextureType::DEPTH_STENCIL:    return "DEPTH_STENCIL";
+    }
+    return "UNKNOWN";
+}
+
+ enum class DescriptorType : uint8_t {
+     SAMPLER_2D_FLOAT,
+     SAMPLER_2D_INT,
+     SAMPLER_2D_UINT,
+     SAMPLER_2D_DEPTH,
+
+     SAMPLER_2D_ARRAY_FLOAT,
+     SAMPLER_2D_ARRAY_INT,
+     SAMPLER_2D_ARRAY_UINT,
+     SAMPLER_2D_ARRAY_DEPTH,
+
+     SAMPLER_CUBE_FLOAT,
+     SAMPLER_CUBE_INT,
+     SAMPLER_CUBE_UINT,
+     SAMPLER_CUBE_DEPTH,
+
+     SAMPLER_CUBE_ARRAY_FLOAT,
+     SAMPLER_CUBE_ARRAY_INT,
+     SAMPLER_CUBE_ARRAY_UINT,
+     SAMPLER_CUBE_ARRAY_DEPTH,
+
+     SAMPLER_3D_FLOAT,
+     SAMPLER_3D_INT,
+     SAMPLER_3D_UINT,
+
+     SAMPLER_2D_MS_FLOAT,
+     SAMPLER_2D_MS_INT,
+     SAMPLER_2D_MS_UINT,
+
+     SAMPLER_2D_MS_ARRAY_FLOAT,
+     SAMPLER_2D_MS_ARRAY_INT,
+     SAMPLER_2D_MS_ARRAY_UINT,
+
+     SAMPLER_EXTERNAL,
+     UNIFORM_BUFFER,
+     SHADER_STORAGE_BUFFER,
+     INPUT_ATTACHMENT,
+ };
+
+constexpr bool isDepthDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_DEPTH:
+        case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+        case DescriptorType::SAMPLER_CUBE_DEPTH:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool isFloatDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_FLOAT:
+        case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+        case DescriptorType::SAMPLER_CUBE_FLOAT:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+        case DescriptorType::SAMPLER_3D_FLOAT:
+        case DescriptorType::SAMPLER_2D_MS_FLOAT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool isIntDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_INT:
+        case DescriptorType::SAMPLER_2D_ARRAY_INT:
+        case DescriptorType::SAMPLER_CUBE_INT:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+        case DescriptorType::SAMPLER_3D_INT:
+        case DescriptorType::SAMPLER_2D_MS_INT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool isUnsignedIntDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_UINT:
+        case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+        case DescriptorType::SAMPLER_CUBE_UINT:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+        case DescriptorType::SAMPLER_3D_UINT:
+        case DescriptorType::SAMPLER_2D_MS_UINT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool is3dTypeDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_3D_FLOAT:
+        case DescriptorType::SAMPLER_3D_INT:
+        case DescriptorType::SAMPLER_3D_UINT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool is2dTypeDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_FLOAT:
+        case DescriptorType::SAMPLER_2D_INT:
+        case DescriptorType::SAMPLER_2D_UINT:
+        case DescriptorType::SAMPLER_2D_DEPTH:
+        case DescriptorType::SAMPLER_2D_MS_FLOAT:
+        case DescriptorType::SAMPLER_2D_MS_INT:
+        case DescriptorType::SAMPLER_2D_MS_UINT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool is2dArrayTypeDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+        case DescriptorType::SAMPLER_2D_ARRAY_INT:
+        case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+        case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool isCubeTypeDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_CUBE_FLOAT:
+        case DescriptorType::SAMPLER_CUBE_INT:
+        case DescriptorType::SAMPLER_CUBE_UINT:
+        case DescriptorType::SAMPLER_CUBE_DEPTH:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool isCubeArrayTypeDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+        case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr bool isMultiSampledTypeDescriptor(DescriptorType const type) noexcept {
+    switch (type) {
+        case DescriptorType::SAMPLER_2D_MS_FLOAT:
+        case DescriptorType::SAMPLER_2D_MS_INT:
+        case DescriptorType::SAMPLER_2D_MS_UINT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+        case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
+            return true;
+        default: ;
+    }
+    return false;
+}
+
+constexpr std::string_view to_string(DescriptorType type) noexcept {
+    #define DESCRIPTOR_TYPE_CASE(TYPE)  case DescriptorType::TYPE: return #TYPE;
+    switch (type) {
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_DEPTH)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_ARRAY_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_ARRAY_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_ARRAY_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_ARRAY_DEPTH)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_DEPTH)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_ARRAY_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_ARRAY_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_ARRAY_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_CUBE_ARRAY_DEPTH)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_3D_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_3D_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_3D_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_MS_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_MS_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_MS_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_MS_ARRAY_FLOAT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_MS_ARRAY_INT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_2D_MS_ARRAY_UINT)
+        DESCRIPTOR_TYPE_CASE(SAMPLER_EXTERNAL)
+        DESCRIPTOR_TYPE_CASE(UNIFORM_BUFFER)
+        DESCRIPTOR_TYPE_CASE(SHADER_STORAGE_BUFFER)
+        DESCRIPTOR_TYPE_CASE(INPUT_ATTACHMENT)
+    }
+    return "UNKNOWN";
+    #undef DESCRIPTOR_TYPE_CASE
+}
+
 enum class DescriptorFlags : uint8_t {
     NONE = 0x00,
     DYNAMIC_OFFSET = 0x01
@@ -254,6 +482,13 @@ using descriptor_set_t = uint8_t;
 using descriptor_binding_t = uint8_t;
 
 struct DescriptorSetLayoutBinding {
+    static bool isSampler(DescriptorType type) noexcept {
+        return int(type) <= int(DescriptorType::SAMPLER_EXTERNAL);
+    }
+    static bool isBuffer(DescriptorType type) noexcept {
+        return type == DescriptorType::UNIFORM_BUFFER ||
+               type == DescriptorType::SHADER_STORAGE_BUFFER;
+    }
     DescriptorType type;
     ShaderStageFlags stageFlags;
     descriptor_binding_t binding;
@@ -264,7 +499,7 @@ struct DescriptorSetLayoutBinding {
     //  no uninitialized padding bytes.
     //    uint8_t externalSamplerDataIndex = EXTERNAL_SAMPLER_DATA_INDEX_UNUSED;
 
-    friend inline bool operator==(DescriptorSetLayoutBinding const& lhs,
+    friend bool operator==(DescriptorSetLayoutBinding const& lhs,
             DescriptorSetLayoutBinding const& rhs) noexcept {
         return lhs.type == rhs.type &&
                lhs.flags == rhs.flags &&
@@ -297,7 +532,7 @@ enum class TargetBufferFlags : uint32_t {
     ALL = COLOR_ALL | DEPTH | STENCIL       //!< Color, depth and stencil buffer selected.
 };
 
-inline constexpr TargetBufferFlags getTargetBufferFlagsAt(size_t index) noexcept {
+constexpr TargetBufferFlags getTargetBufferFlagsAt(size_t index) noexcept {
     if (index == 0u) return TargetBufferFlags::COLOR0;
     if (index == 1u) return TargetBufferFlags::COLOR1;
     if (index == 2u) return TargetBufferFlags::COLOR2;
@@ -454,6 +689,24 @@ enum class SamplerType : uint8_t {
     SAMPLER_CUBEMAP_ARRAY,  //!< Cube map array texture (feature level 2)
 };
 
+constexpr std::string_view to_string(SamplerType const type) noexcept {
+    switch (type) {
+        case SamplerType::SAMPLER_2D:
+            return "SAMPLER_2D";
+        case SamplerType::SAMPLER_2D_ARRAY:
+            return "SAMPLER_2D_ARRAY";
+        case SamplerType::SAMPLER_CUBEMAP:
+            return "SAMPLER_CUBEMAP";
+        case SamplerType::SAMPLER_EXTERNAL:
+            return "SAMPLER_EXTERNAL";
+        case SamplerType::SAMPLER_3D:
+            return "SAMPLER_3D";
+        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
+            return "SAMPLER_CUBEMAP_ARRAY";
+    }
+    return "Unknown";
+}
+
 //! Subpass type
 enum class SubpassType : uint8_t {
     SUBPASS_INPUT
@@ -467,6 +720,20 @@ enum class SamplerFormat : uint8_t {
     SHADOW = 3      //!< shadow sampler (PCF)
 };
 
+constexpr std::string_view to_string(SamplerFormat const format) noexcept {
+    switch (format) {
+        case SamplerFormat::INT:
+            return "INT";
+        case SamplerFormat::UINT:
+            return "UINT";
+        case SamplerFormat::FLOAT:
+            return "FLOAT";
+        case SamplerFormat::SHADOW:
+            return "SHADOW";
+    }
+    return "Unknown";
+}
+
 /**
  * Supported element types
  */
@@ -506,6 +773,15 @@ enum class BufferObjectBinding : uint8_t {
     SHADER_STORAGE
 };
 
+constexpr std::string_view to_string(BufferObjectBinding type) noexcept {
+    switch (type) {
+        case BufferObjectBinding::VERTEX:           return "VERTEX";
+        case BufferObjectBinding::UNIFORM:          return "UNIFORM";
+        case BufferObjectBinding::SHADER_STORAGE:   return "SHADER_STORAGE";
+    }
+    return "UNKNOWN";
+}
+
 //! Face culling Mode
 enum class CullingMode : uint8_t {
     NONE,               //!< No culling, front and back faces are visible
@@ -767,6 +1043,8 @@ enum class TextureFormat : uint16_t {
     SRGB_ALPHA_BPTC_UNORM,  // BC7 sRGB
 };
 
+TextureType getTextureType(TextureFormat format) noexcept;
+
 //! Bitmask describing the intended Texture Usage
 enum class TextureUsage : uint16_t {
     NONE                = 0x0000,
@@ -794,7 +1072,7 @@ enum class TextureSwizzle : uint8_t {
 };
 
 //! returns whether this format a depth format
-static constexpr bool isDepthFormat(TextureFormat format) noexcept {
+constexpr bool isDepthFormat(TextureFormat format) noexcept {
     switch (format) {
         case TextureFormat::DEPTH32F:
         case TextureFormat::DEPTH24:
@@ -807,7 +1085,7 @@ static constexpr bool isDepthFormat(TextureFormat format) noexcept {
     }
 }
 
-static constexpr bool isStencilFormat(TextureFormat format) noexcept {
+constexpr bool isStencilFormat(TextureFormat format) noexcept {
     switch (format) {
         case TextureFormat::STENCIL8:
         case TextureFormat::DEPTH24_STENCIL8:
@@ -818,7 +1096,7 @@ static constexpr bool isStencilFormat(TextureFormat format) noexcept {
     }
 }
 
-inline constexpr bool isColorFormat(TextureFormat format) noexcept {
+constexpr bool isColorFormat(TextureFormat format) noexcept {
     switch (format) {
         // Standard color formats
         case TextureFormat::R8:
@@ -845,7 +1123,7 @@ inline constexpr bool isColorFormat(TextureFormat format) noexcept {
     return false;
 }
 
-static constexpr bool isUnsignedIntFormat(TextureFormat format) {
+constexpr bool isUnsignedIntFormat(TextureFormat format) {
     switch (format) {
         case TextureFormat::R8UI:
         case TextureFormat::R16UI:
@@ -866,7 +1144,7 @@ static constexpr bool isUnsignedIntFormat(TextureFormat format) {
     }
 }
 
-static constexpr bool isSignedIntFormat(TextureFormat format) {
+constexpr bool isSignedIntFormat(TextureFormat format) {
     switch (format) {
         case TextureFormat::R8I:
         case TextureFormat::R16I:
@@ -888,35 +1166,35 @@ static constexpr bool isSignedIntFormat(TextureFormat format) {
 }
 
 //! returns whether this format is a compressed format
-static constexpr bool isCompressedFormat(TextureFormat format) noexcept {
+constexpr bool isCompressedFormat(TextureFormat format) noexcept {
     return format >= TextureFormat::EAC_R11;
 }
 
 //! returns whether this format is an ETC2 compressed format
-static constexpr bool isETC2Compression(TextureFormat format) noexcept {
+constexpr bool isETC2Compression(TextureFormat format) noexcept {
     return format >= TextureFormat::EAC_R11 && format <= TextureFormat::ETC2_EAC_SRGBA8;
 }
 
 //! returns whether this format is an S3TC compressed format
-static constexpr bool isS3TCCompression(TextureFormat format) noexcept {
+constexpr bool isS3TCCompression(TextureFormat format) noexcept {
     return format >= TextureFormat::DXT1_RGB && format <= TextureFormat::DXT5_SRGBA;
 }
 
-static constexpr bool isS3TCSRGBCompression(TextureFormat format) noexcept {
+constexpr bool isS3TCSRGBCompression(TextureFormat format) noexcept {
     return format >= TextureFormat::DXT1_SRGB && format <= TextureFormat::DXT5_SRGBA;
 }
 
 //! returns whether this format is an RGTC compressed format
-static constexpr bool isRGTCCompression(TextureFormat format) noexcept {
+constexpr bool isRGTCCompression(TextureFormat format) noexcept {
     return format >= TextureFormat::RED_RGTC1 && format <= TextureFormat::SIGNED_RED_GREEN_RGTC2;
 }
 
 //! returns whether this format is an BPTC compressed format
-static constexpr bool isBPTCCompression(TextureFormat format) noexcept {
+constexpr bool isBPTCCompression(TextureFormat format) noexcept {
     return format >= TextureFormat::RGB_BPTC_SIGNED_FLOAT && format <= TextureFormat::SRGB_ALPHA_BPTC_UNORM;
 }
 
-static constexpr bool isASTCCompression(TextureFormat format) noexcept {
+constexpr bool isASTCCompression(TextureFormat format) noexcept {
     return format >= TextureFormat::RGBA_ASTC_4x4 && format <= TextureFormat::SRGB8_ALPHA8_ASTC_12x12;
 }
 
@@ -1037,19 +1315,23 @@ struct SamplerParams {             // NOLINT
             assert_invariant(lhs.padding2 == 0);
             auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
             auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
-            return *pLhs == *pRhs;
+            return *pLhs < *pRhs;
         }
     };
 
+    bool isFiltered() const noexcept {
+        return filterMag != SamplerMagFilter::NEAREST || filterMin != SamplerMinFilter::NEAREST;
+    }
+
 private:
-    friend inline bool operator == (SamplerParams lhs, SamplerParams rhs) noexcept {
-        return SamplerParams::EqualTo{}(lhs, rhs);
+    friend bool operator == (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return EqualTo{}(lhs, rhs);
     }
-    friend inline bool operator != (SamplerParams lhs, SamplerParams rhs) noexcept {
-        return  !SamplerParams::EqualTo{}(lhs, rhs);
+    friend bool operator != (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return  !EqualTo{}(lhs, rhs);
     }
-    friend inline bool operator < (SamplerParams lhs, SamplerParams rhs) noexcept {
-        return SamplerParams::LessThan{}(lhs, rhs);
+    friend bool operator < (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return LessThan{}(lhs, rhs);
     }
 };
 
@@ -1099,15 +1381,15 @@ struct SamplerYcbcrConversion {// NOLINT
     };
 
 private:
-    friend inline bool operator == (SamplerYcbcrConversion lhs, SamplerYcbcrConversion rhs)
+    friend bool operator == (SamplerYcbcrConversion lhs, SamplerYcbcrConversion rhs)
             noexcept {
         return SamplerYcbcrConversion::EqualTo{}(lhs, rhs);
     }
-    friend inline bool operator != (SamplerYcbcrConversion lhs, SamplerYcbcrConversion rhs)
+    friend bool operator != (SamplerYcbcrConversion lhs, SamplerYcbcrConversion rhs)
             noexcept {
         return  !SamplerYcbcrConversion::EqualTo{}(lhs, rhs);
     }
-    friend inline bool operator < (SamplerYcbcrConversion lhs, SamplerYcbcrConversion rhs)
+    friend bool operator < (SamplerYcbcrConversion lhs, SamplerYcbcrConversion rhs)
             noexcept {
         return SamplerYcbcrConversion::LessThan{}(lhs, rhs);
     }

filament/backend/include/backend/PixelBufferDescriptor.h

@@ -204,23 +204,15 @@ public:
                 }, new T(std::forward<T>(functor))
         };
     }
-
-    // --------------------------------------------------------------------------------------------
-
     /**
-     * Computes the size in bytes needed to fit an image of given dimensions and format
+     * Computes the size in bytes for a pixel of given dimensions and format
      *
      * @param format    Format of the image pixels
      * @param type      Type of the image pixels
-     * @param stride    Stride of a row in pixels
-     * @param height    Height of the image in rows
-     * @param alignment Alignment in bytes of pixel rows
-     * @return The buffer size needed to fit this image in bytes
+     * @return The size of the specified pixel in bytes
      */
-    static constexpr size_t computeDataSize(PixelDataFormat format, PixelDataType type,
-            size_t stride, size_t height, size_t alignment) noexcept {
-        assert_invariant(alignment);
 
+    static constexpr size_t computePixelSize(PixelDataFormat format, PixelDataType type) noexcept {
         if (type == PixelDataType::COMPRESSED) {
             return 0;
         }
@@ -242,7 +234,7 @@ public:
             case PixelDataFormat::RGB_INTEGER:
                 n = 3;
                 break;
-            case PixelDataFormat::UNUSED: // shouldn't happen (used to be rgbm)
+            case PixelDataFormat::UNUSED:// shouldn't happen (used to be rgbm)
             case PixelDataFormat::RGBA:
             case PixelDataFormat::RGBA_INTEGER:
                 n = 4;
@@ -251,7 +243,7 @@ public:
 
         size_t bpp = n;
         switch (type) {
-            case PixelDataType::COMPRESSED: // Impossible -- to squash the IDE warnings
+            case PixelDataType::COMPRESSED:// Impossible -- to squash the IDE warnings
             case PixelDataType::UBYTE:
             case PixelDataType::BYTE:
                 // nothing to do
@@ -282,16 +274,35 @@ public:
                 bpp = 2;
                 break;
         }
+        return bpp;
+    }
 
+    // --------------------------------------------------------------------------------------------
+
+    /**
+     * Computes the size in bytes needed to fit an image of given dimensions and format
+     *
+     * @param format    Format of the image pixels
+     * @param type      Type of the image pixels
+     * @param stride    Stride of a row in pixels
+     * @param height    Height of the image in rows
+     * @param alignment Alignment in bytes of pixel rows
+     * @return The buffer size needed to fit this image in bytes
+     */
+    static constexpr size_t computeDataSize(PixelDataFormat format, PixelDataType type,
+            size_t stride, size_t height, size_t alignment) noexcept {
+        assert_invariant(alignment);
+
+        size_t bpp = computePixelSize(format, type);
         size_t const bpr = bpp * stride;
         size_t const bprAligned = (bpr + (alignment - 1)) & (~alignment + 1);
         return bprAligned * height;
     }
 
     //! left coordinate in pixels
-    uint32_t left   = 0;
+    uint32_t left = 0;
     //! top coordinate in pixels
-    uint32_t top    = 0;
+    uint32_t top = 0;
     union {
         struct {
             //! stride in pixels

filament/backend/include/backend/Program.h

@@ -20,6 +20,7 @@
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Invocable.h>
+#include <utils/bitset.h>
 
 #include <backend/DriverEnums.h>
 
@@ -66,6 +67,7 @@ public:
     using DescriptorBindingsInfo = utils::FixedCapacityVector<Descriptor>;
     using DescriptorSetInfo = std::array<DescriptorBindingsInfo, MAX_DESCRIPTOR_SET_COUNT>;
     using SpecializationConstantsInfo = utils::FixedCapacityVector<SpecializationConstant>;
+    using MutableSpecConstantsInfo = utils::bitset8;
     using ShaderBlob = utils::FixedCapacityVector<uint8_t>;
     using ShaderSource = std::array<ShaderBlob, SHADER_TYPE_COUNT>;
 
@@ -102,7 +104,8 @@ public:
     Program& descriptorBindings(backend::descriptor_set_t set,
             DescriptorBindingsInfo descriptorBindings) noexcept;
 
-    Program& specializationConstants(SpecializationConstantsInfo specConstants) noexcept;
+    Program& specializationConstants(SpecializationConstantsInfo specConstants,
+            uint32_t firstMutableId, MutableSpecConstantsInfo mutableSpecConstants) noexcept;
 
     struct PushConstant {
         utils::CString name;

filament/backend/include/backend/platforms/OpenGLPlatform.h

@@ -199,7 +199,7 @@ public:
      */
     virtual bool makeCurrent(ContextType type,
             SwapChain* UTILS_NONNULL drawSwapChain,
-            SwapChain* UTILS_NONNULL readSwapChain) noexcept = 0;
+            SwapChain* UTILS_NONNULL readSwapChain) = 0;
 
     /**
      * Called by the driver to make the OpenGL context active on the calling thread and bind
@@ -219,7 +219,7 @@ public:
             SwapChain* UTILS_NONNULL drawSwapChain,
             SwapChain* UTILS_NONNULL readSwapChain,
             utils::Invocable<void()> preContextChange,
-            utils::Invocable<void(size_t index)> postContextChange) noexcept;
+            utils::Invocable<void(size_t index)> postContextChange);
 
     /**
      * Called by the backend just before calling commit()

filament/backend/include/backend/platforms/PlatformCocoaGL.h

@@ -58,7 +58,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) override;
     void commit(SwapChain* swapChain) noexcept override;
     ExternalTexture* createExternalImageTexture() noexcept override;
     void destroyExternalImageTexture(ExternalTexture* texture) noexcept override;

filament/backend/include/backend/platforms/PlatformCocoaTouchGL.h

@@ -55,7 +55,7 @@ public:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) override;
     void commit(SwapChain* swapChain) noexcept override;
 
     ExternalTexture* createExternalImageTexture() noexcept override;

filament/backend/include/backend/platforms/PlatformEGL.h

@@ -109,11 +109,11 @@ protected:
 
     bool makeCurrent(ContextType type,
             SwapChain* drawSwapChain,
-            SwapChain* readSwapChain) noexcept override;
+            SwapChain* readSwapChain) override;
 
     void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain,
             utils::Invocable<void()> preContextChange,
-            utils::Invocable<void(size_t index)> postContextChange) noexcept override;
+            utils::Invocable<void(size_t index)> postContextChange) override;
 
     void commit(SwapChain* swapChain) noexcept override;
 
@@ -148,12 +148,12 @@ protected:
     EGLContext getContextForType(ContextType type) const noexcept;
 
     // makes the draw and read surface current without changing the current context
-    EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
+    EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) {
         return egl.makeCurrent(drawSurface, readSurface);
     }
 
     // makes context current and set draw and read surfaces to EGL_NO_SURFACE
-    EGLBoolean makeCurrent(EGLContext context) noexcept {
+    EGLBoolean makeCurrent(EGLContext context) {
         return egl.makeCurrent(context, mEGLDummySurface, mEGLDummySurface);
     }
 
@@ -211,9 +211,9 @@ private:
     public:
         explicit EGL(EGLDisplay& dpy) : mEGLDisplay(dpy) {}
         EGLBoolean makeCurrent(EGLContext context,
-                EGLSurface drawSurface, EGLSurface readSurface) noexcept;
+                EGLSurface drawSurface, EGLSurface readSurface);
 
-        EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
+        EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) {
             return makeCurrent(mCurrentContext, drawSurface, readSurface);
         }
     } egl{ mEGLDisplay };

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

@@ -127,7 +127,7 @@ protected:
 protected:
     bool makeCurrent(ContextType type,
             SwapChain* drawSwapChain,
-            SwapChain* readSwapChain) noexcept override;
+            SwapChain* readSwapChain) override;
 
 private:
     struct InitializeJvmForPerformanceManagerIfNeeded {

filament/backend/include/backend/platforms/PlatformGLX.h

@@ -51,7 +51,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) override;
     void commit(SwapChain* swapChain) noexcept override;
 
 private:

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

@@ -56,7 +56,7 @@ protected:
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
     bool makeCurrent(ContextType type, SwapChain* drawSwapChain,
-            SwapChain* readSwapChain) noexcept override;
+            SwapChain* readSwapChain) override;
     void commit(SwapChain* swapChain) noexcept override;
 
 private:

filament/backend/include/backend/platforms/PlatformWGL.h

@@ -53,7 +53,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) override;
     void commit(SwapChain* swapChain) noexcept override;
 
 protected:

filament/backend/include/backend/platforms/PlatformWebGL.h

@@ -46,7 +46,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) override;
     void commit(SwapChain* swapChain) noexcept override;
 };
 

filament/backend/include/backend/platforms/WebGPUPlatform.h

@@ -22,6 +22,7 @@
 #include <webgpu/webgpu_cpp.h>
 
 #include <cstdint>
+#include <vector>
 
 namespace filament::backend {
 
@@ -56,6 +57,10 @@ protected:
             const Platform::DriverConfig& driverConfig) noexcept override;
 
 private:
+    // returns adapter request option variations applicable for the particular
+    // platform
+    [[nodiscard]] static std::vector<wgpu::RequestAdapterOptions> getAdapterOptions();
+
     // we may consider having the driver own this in the future
     wgpu::Instance mInstance;
 };

filament/backend/include/private/backend/CircularBuffer.h

@@ -72,7 +72,7 @@ public:
     Range getBuffer() noexcept;
 
 private:
-    void* alloc(size_t size) noexcept;
+    void* alloc(size_t size);
     void dealloc() noexcept;
 
     // pointer to the beginning of the circular buffer (constant)

filament/backend/include/private/backend/CommandBufferQueue.h

@@ -76,7 +76,7 @@ public:
 
     // all commands buffers (Slices) written to this point are returned by waitForCommand(). This
     // call blocks until the CircularBuffer has at least mRequiredSize bytes available.
-    void flush() noexcept;
+    void flush();
 
     // returns from waitForCommands() immediately.
     void requestExit();

filament/backend/include/private/backend/VirtualMachineEnv.h

@@ -27,13 +27,13 @@ namespace filament {
 class VirtualMachineEnv {
 public:
     // must be called before VirtualMachineEnv::get() from a thread that is attached to the JavaVM
-    static jint JNI_OnLoad(JavaVM* vm) noexcept;
+    static jint JNI_OnLoad(JavaVM* vm);
 
     // must be called on backend thread
     static VirtualMachineEnv& get() noexcept;
 
     // can be called from any thread that already has a JniEnv
-    static JNIEnv* getThreadEnvironment() noexcept;
+    static JNIEnv* getThreadEnvironment();
 
     // must be called from the backend thread
     JNIEnv* getEnvironment() noexcept {
@@ -49,7 +49,7 @@ public:
 private:
     explicit VirtualMachineEnv(JavaVM* vm) noexcept;
     ~VirtualMachineEnv() noexcept;
-    JNIEnv* getEnvironmentSlow() noexcept;
+    JNIEnv* getEnvironmentSlow();
 
     static utils::Mutex sLock;
     static JavaVM* sVirtualMachine;

filament/backend/src/BackendUtils.cpp

@@ -18,6 +18,8 @@
 
 #include "DataReshaper.h"
 
+#include <backend/DriverEnums.h>
+
 #include <utils/CString.h>
 
 #include <string_view>
@@ -223,6 +225,143 @@ size_t getFormatSize(TextureFormat format) noexcept {
     }
 }
 
+TextureType getTextureType(TextureFormat const format) noexcept {
+    switch (format) {
+        case TextureFormat::UNUSED:
+            // should not happen
+            return TextureType::FLOAT;
+
+        case TextureFormat::R8:
+        case TextureFormat::R16F:
+        case TextureFormat::RG8:
+        case TextureFormat::RGB565:
+        case TextureFormat::RGB5_A1:
+        case TextureFormat::RGBA4:
+        case TextureFormat::R32F:
+        case TextureFormat::RGB8:
+        case TextureFormat::SRGB8:
+        case TextureFormat::RG16F:
+        case TextureFormat::R11F_G11F_B10F:
+        case TextureFormat::RGB9_E5:
+        case TextureFormat::RGBA8:
+        case TextureFormat::SRGB8_A8:
+        case TextureFormat::RGB10_A2:
+        case TextureFormat::RGB16F:
+        case TextureFormat::RG32F:
+        case TextureFormat::RGBA16F:
+        case TextureFormat::RGB32F:
+        case TextureFormat::RGBA32F:
+            return TextureType::FLOAT;
+
+        case TextureFormat::R8_SNORM:
+        case TextureFormat::RG8_SNORM:
+        case TextureFormat::RGB8_SNORM:
+        case TextureFormat::RGBA8_SNORM:
+            // SNORM are treated as float
+            return TextureType::FLOAT;
+
+        case TextureFormat::R8UI:
+        case TextureFormat::R16UI:
+        case TextureFormat::RG8UI:
+        case TextureFormat::RGB8UI:
+        case TextureFormat::R32UI:
+        case TextureFormat::RG16UI:
+        case TextureFormat::RGBA8UI:
+        case TextureFormat::RGB16UI:
+        case TextureFormat::RG32UI:
+        case TextureFormat::RGBA16UI:
+        case TextureFormat::RGB32UI:
+        case TextureFormat::RGBA32UI:
+            return TextureType::UINT;
+
+        case TextureFormat::R8I:
+        case TextureFormat::R16I:
+        case TextureFormat::RG8I:
+        case TextureFormat::RGB8I:
+        case TextureFormat::R32I:
+        case TextureFormat::RG16I:
+        case TextureFormat::RGBA8I:
+        case TextureFormat::RGB16I:
+        case TextureFormat::RG32I:
+        case TextureFormat::RGBA16I:
+        case TextureFormat::RGB32I:
+        case TextureFormat::RGBA32I:
+            return TextureType::INT;
+
+        case TextureFormat::DEPTH16:
+        case TextureFormat::DEPTH24:
+        case TextureFormat::DEPTH32F:
+            return TextureType::DEPTH;
+
+        case TextureFormat::STENCIL8:
+            return TextureType::STENCIL;
+
+        case TextureFormat::DEPTH24_STENCIL8:
+        case TextureFormat::DEPTH32F_STENCIL8:
+            return TextureType::DEPTH_STENCIL;
+
+        // Compressed formats ---------------------------------------------------------------------
+
+        case TextureFormat::EAC_RG11:
+        case TextureFormat::EAC_RG11_SIGNED:
+        case TextureFormat::ETC2_EAC_RGBA8:
+        case TextureFormat::ETC2_EAC_SRGBA8:
+        case TextureFormat::EAC_R11:
+        case TextureFormat::EAC_R11_SIGNED:
+        case TextureFormat::ETC2_RGB8:
+        case TextureFormat::ETC2_SRGB8:
+        case TextureFormat::ETC2_RGB8_A1:
+        case TextureFormat::ETC2_SRGB8_A1:
+        case TextureFormat::DXT1_RGB:
+        case TextureFormat::DXT1_RGBA:
+        case TextureFormat::DXT1_SRGB:
+        case TextureFormat::DXT1_SRGBA:
+        case TextureFormat::DXT3_RGBA:
+        case TextureFormat::DXT3_SRGBA:
+        case TextureFormat::DXT5_RGBA:
+        case TextureFormat::DXT5_SRGBA:
+        case TextureFormat::RED_RGTC1:
+        case TextureFormat::SIGNED_RED_RGTC1:
+        case TextureFormat::RED_GREEN_RGTC2:
+        case TextureFormat::SIGNED_RED_GREEN_RGTC2:
+        case TextureFormat::RGB_BPTC_SIGNED_FLOAT:
+        case TextureFormat::RGB_BPTC_UNSIGNED_FLOAT:
+        case TextureFormat::RGBA_BPTC_UNORM:
+        case TextureFormat::SRGB_ALPHA_BPTC_UNORM:
+        case TextureFormat::RGBA_ASTC_4x4:
+        case TextureFormat::RGBA_ASTC_5x4:
+        case TextureFormat::RGBA_ASTC_5x5:
+        case TextureFormat::RGBA_ASTC_6x5:
+        case TextureFormat::RGBA_ASTC_6x6:
+        case TextureFormat::RGBA_ASTC_8x5:
+        case TextureFormat::RGBA_ASTC_8x6:
+        case TextureFormat::RGBA_ASTC_8x8:
+        case TextureFormat::RGBA_ASTC_10x5:
+        case TextureFormat::RGBA_ASTC_10x6:
+        case TextureFormat::RGBA_ASTC_10x8:
+        case TextureFormat::RGBA_ASTC_10x10:
+        case TextureFormat::RGBA_ASTC_12x10:
+        case TextureFormat::RGBA_ASTC_12x12:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_4x4:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_5x4:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_5x5:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_6x5:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_6x6:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_8x5:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_8x6:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_8x8:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_10x5:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_10x6:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_10x8:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_10x10:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_12x10:
+        case TextureFormat::SRGB8_ALPHA8_ASTC_12x12:
+            return TextureType::FLOAT;
+    }
+
+    return TextureType::FLOAT;
+}
+
 size_t getFormatComponentCount(TextureFormat format) noexcept {
     switch (format) {
         case TextureFormat::R8:

filament/backend/src/CircularBuffer.cpp

@@ -65,7 +65,7 @@ CircularBuffer::~CircularBuffer() noexcept {
 // to each others and a special case in circularize()
 
 UTILS_NOINLINE
-void* CircularBuffer::alloc(size_t size) noexcept {
+void* CircularBuffer::alloc(size_t size) {
 #if HAS_MMAP
     void* data = nullptr;
     void* vaddr = MAP_FAILED;

filament/backend/src/CommandBufferQueue.cpp

@@ -18,12 +18,13 @@
 #include "private/backend/CircularBuffer.h"
 #include "private/backend/CommandStream.h"
 
+#include <private/utils/Tracing.h>
+
 #include <utils/compiler.h>
 #include <utils/Log.h>
 #include <utils/Mutex.h>
 #include <utils/ostream.h>
 #include <utils/Panic.h>
-#include <utils/Systrace.h>
 #include <utils/debug.h>
 
 #include <algorithm>
@@ -52,18 +53,18 @@ CommandBufferQueue::~CommandBufferQueue() {
 }
 
 void CommandBufferQueue::requestExit() {
-    std::lock_guard<utils::Mutex> const lock(mLock);
+    std::lock_guard const lock(mLock);
     mExitRequested = EXIT_REQUESTED;
     mCondition.notify_one();
 }
 
 bool CommandBufferQueue::isPaused() const noexcept {
-    std::lock_guard<utils::Mutex> const lock(mLock);
+    std::lock_guard const lock(mLock);
     return mPaused;
 }
 
 void CommandBufferQueue::setPaused(bool paused) {
-    std::lock_guard<utils::Mutex> const lock(mLock);
+    std::lock_guard const lock(mLock);
     if (paused) {
         mPaused = true;
     } else {
@@ -73,13 +74,13 @@ void CommandBufferQueue::setPaused(bool paused) {
 }
 
 bool CommandBufferQueue::isExitRequested() const {
-    std::lock_guard<utils::Mutex> const lock(mLock);
-    return (bool)mExitRequested;
+    std::lock_guard const lock(mLock);
+    return bool(mExitRequested);
 }
 
 
-void CommandBufferQueue::flush() noexcept {
-    SYSTRACE_CALL();
+void CommandBufferQueue::flush() {
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     CircularBuffer& circularBuffer = mCircularBuffer;
     if (circularBuffer.empty()) {
@@ -102,7 +103,7 @@ void CommandBufferQueue::flush() noexcept {
             static_cast<char const*>(begin), static_cast<char const*>(end));
 
 
-    std::unique_lock<utils::Mutex> lock(mLock);
+    std::unique_lock lock(mLock);
 
     // circular buffer is too small, we corrupted the stream
     FILAMENT_CHECK_POSTCONDITION(used <= mFreeSpace) <<
@@ -129,7 +130,7 @@ void CommandBufferQueue::flush() noexcept {
         mHighWatermark = std::max(mHighWatermark, totalUsed);
 #endif
 
-        SYSTRACE_NAME("waiting: CircularBuffer::flush()");
+        FILAMENT_TRACING_NAME(FILAMENT_TRACING_CATEGORY_FILAMENT, "waiting: CircularBuffer::flush()");
 
         FILAMENT_CHECK_POSTCONDITION(!mPaused) <<
                 "CommandStream is full, but since the rendering thread is paused, "
@@ -146,7 +147,7 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
     if (!UTILS_HAS_THREADING) {
         return std::move(mCommandBuffersToExecute);
     }
-    std::unique_lock<utils::Mutex> lock(mLock);
+    std::unique_lock lock(mLock);
     while ((mCommandBuffersToExecute.empty() || mPaused) && !mExitRequested) {
         mCondition.wait(lock);
     }
@@ -156,7 +157,7 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
 void CommandBufferQueue::releaseBuffer(CommandBufferQueue::Range const& buffer) {
     size_t const used = std::distance(
             static_cast<char const*>(buffer.begin), static_cast<char const*>(buffer.end));
-    std::lock_guard<utils::Mutex> const lock(mLock);
+    std::lock_guard const lock(mLock);
     mFreeSpace += used;
     mCondition.notify_one();
 }

filament/backend/src/CommandStream.cpp

@@ -16,6 +16,8 @@
 
 #include "private/backend/CommandStream.h"
 
+#include <private/utils/Tracing.h>
+
 #if DEBUG_COMMAND_STREAM
 #include <utils/CallStack.h>
 #endif
@@ -24,7 +26,6 @@
 #include <utils/Log.h>
 #include <utils/ostream.h>
 #include <utils/Profiler.h>
-#include <utils/Systrace.h>
 
 #include <cstddef>
 #include <functional>
@@ -79,12 +80,13 @@ CommandStream::CommandStream(Driver& driver, CircularBuffer& buffer) noexcept
 }
 
 void CommandStream::execute(void* buffer) {
-    // NOTE: we can't use SYSTRACE_CALL() or similar here because, execute() below, also
+    // NOTE: we can't use FILAMENT_TRACING_CALL() or similar here because, execute() below, also
     // uses systrace BEGIN/END and the END is not guaranteed to be happening in this scope.
 
     Profiler profiler;
 
-    if (SYSTRACE_TAG) {
+
+    if constexpr (FILAMENT_TRACING_ENABLED) {
         if (UTILS_UNLIKELY(mUsePerformanceCounter)) {
             // we want to remove all this when tracing is completely disabled
             profiler.resetEvents(Profiler::EV_CPU_CYCLES  | Profiler::EV_BPU_MISSES);
@@ -100,17 +102,17 @@ void CommandStream::execute(void* buffer) {
         }
     });
 
-    if (SYSTRACE_TAG) {
+    if constexpr (FILAMENT_TRACING_ENABLED) {
         if (UTILS_UNLIKELY(mUsePerformanceCounter)) {
             // we want to remove all this when tracing is completely disabled
             profiler.stop();
             UTILS_UNUSED Profiler::Counters const counters = profiler.readCounters();
-            SYSTRACE_CONTEXT();
-            SYSTRACE_VALUE32("GLThread (I)", counters.getInstructions());
-            SYSTRACE_VALUE32("GLThread (C)", counters.getCpuCycles());
-            SYSTRACE_VALUE32("GLThread (CPI x10)", counters.getCPI() * 10);
-            SYSTRACE_VALUE32("GLThread (BPU miss)", counters.getBranchMisses());
-            SYSTRACE_VALUE32("GLThread (I / BPU miss)",
+            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+            FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "GLThread (I)", counters.getInstructions());
+            FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "GLThread (C)", counters.getCpuCycles());
+            FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "GLThread (CPI x10)", counters.getCPI() * 10);
+            FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "GLThread (BPU miss)", counters.getBranchMisses());
+            FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "GLThread (I / BPU miss)",
                     counters.getInstructions() / counters.getBranchMisses());
         }
     }

filament/backend/src/CommandStreamDispatcher.h

@@ -20,8 +20,9 @@
 #include "private/backend/Driver.h"
 #include "private/backend/CommandStream.h"
 
+#include <private/utils/Tracing.h>
+
 #include <utils/compiler.h>
-#include <utils/Systrace.h>
 
 #include <utility>
 
@@ -38,7 +39,7 @@
 #if DEBUG_LEVEL == DEBUG_LEVEL_NONE
 #   define SYSTRACE()
 #elif DEBUG_LEVEL == DEBUG_LEVEL_SYSTRACE
-#   define SYSTRACE() SYSTRACE_CALL();
+#   define SYSTRACE() FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 #else
 #   error "invalid debug level"
 #endif

filament/backend/src/CompilerThreadPool.cpp

@@ -16,7 +16,7 @@
 
 #include "CompilerThreadPool.h"
 
-#include <utils/Systrace.h>
+#include <private/utils/Tracing.h>
 
 #include <memory>
 
@@ -41,7 +41,7 @@ void CompilerThreadPool::init(uint32_t threadCount,
 
     for (size_t i = 0; i < threadCount; i++) {
         mCompilerThreads.emplace_back([this, setup, cleanup]() {
-            SYSTRACE_CONTEXT();
+            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
             (*setup)();
 
@@ -54,7 +54,7 @@ void CompilerThreadPool::init(uint32_t threadCount,
                                     [](auto&& q) { return q.empty(); }));
                 });
 
-                SYSTRACE_VALUE32("CompilerThreadPool Jobs",
+                FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "CompilerThreadPool Jobs",
                         mQueues[0].size() + mQueues[1].size());
 
                 if (UTILS_LIKELY(!mExitRequested)) {

filament/backend/src/Driver.cpp

@@ -23,11 +23,12 @@
 #include <backend/BufferDescriptor.h>
 #include <backend/DriverEnums.h>
 
+#include <private/utils/Tracing.h>
+
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/Log.h>
 #include <utils/ostream.h>
-#include <utils/Systrace.h>
 
 #include <math/half.h>
 #include <math/vec2.h>
@@ -151,30 +152,31 @@ void DriverBase::debugCommandBegin(CommandStream* cmds, bool synchronous, const
             utils::slog.d << methodName << utils::io::endl;
         }
         if constexpr (bool(FILAMENT_DEBUG_COMMANDS & FILAMENT_DEBUG_COMMANDS_SYSTRACE)) {
-            SYSTRACE_CONTEXT();
-            SYSTRACE_NAME_BEGIN(methodName);
+            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+            FILAMENT_TRACING_NAME_BEGIN(FILAMENT_TRACING_CATEGORY_FILAMENT, methodName);
 
             if (!synchronous) {
                 cmds->queueCommand([=]() {
-                    SYSTRACE_CONTEXT();
-                    SYSTRACE_NAME_BEGIN(methodName);
+                    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+                    FILAMENT_TRACING_NAME_BEGIN(FILAMENT_TRACING_CATEGORY_FILAMENT, methodName);
                 });
             }
         }
     }
 }
 
-void DriverBase::debugCommandEnd(CommandStream* cmds, bool synchronous, const char* methodName) noexcept {
+void DriverBase::debugCommandEnd(CommandStream* cmds, bool synchronous,
+        const char* methodName) noexcept {
     if constexpr (bool(FILAMENT_DEBUG_COMMANDS > FILAMENT_DEBUG_COMMANDS_NONE)) {
         if constexpr (bool(FILAMENT_DEBUG_COMMANDS & FILAMENT_DEBUG_COMMANDS_SYSTRACE)) {
             if (!synchronous) {
                 cmds->queueCommand([]() {
-                    SYSTRACE_CONTEXT();
-                    SYSTRACE_NAME_END();
+                    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+                    FILAMENT_TRACING_NAME_END(FILAMENT_TRACING_CATEGORY_FILAMENT);
                 });
             }
-            SYSTRACE_CONTEXT();
-            SYSTRACE_NAME_END();
+            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+            FILAMENT_TRACING_NAME_END(FILAMENT_TRACING_CATEGORY_FILAMENT);
         }
     }
 }

filament/backend/src/PlatformFactory.cpp

@@ -16,7 +16,8 @@
 
 #include <private/backend/PlatformFactory.h>
 
-#include <utils/Systrace.h>
+#include <private/utils/Tracing.h>
+
 #include <utils/debug.h>
 
 // We need to keep this up top for the linux (X11) name collisions.
@@ -85,7 +86,7 @@ namespace filament::backend {
 // responsible for destroying it. Initialization of the backend API is deferred until
 // createDriver(). The passed-in backend hint is replaced with the resolved backend.
 Platform* PlatformFactory::create(Backend* backend) noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
     assert_invariant(backend);
 
 #if defined(__ANDROID__)

filament/backend/src/Program.cpp

@@ -80,8 +80,18 @@ Program& Program::attributes(AttributesInfo attributes) noexcept {
     return *this;
 }
 
-Program& Program::specializationConstants(SpecializationConstantsInfo specConstants) noexcept {
+Program& Program::specializationConstants(SpecializationConstantsInfo specConstants,
+        uint32_t firstMutableId, MutableSpecConstantsInfo mutableSpecConstants) noexcept {
+    // String the two lists together.
     mSpecializationConstants = std::move(specConstants);
+    uint32_t firstMutableIndex = specConstants.size();
+    mSpecializationConstants.reserve(specConstants.size() + mutableSpecConstants.size());
+    for (uint32_t i = 0; i < mutableSpecConstants.size(); i++) {
+        mSpecializationConstants[i + firstMutableIndex] = SpecializationConstant {
+                .id = i + firstMutableId,
+                .value = mutableSpecConstants[i],
+        };
+    }
     return *this;
 }
 

filament/backend/src/SystraceProfile.h

@@ -17,7 +17,7 @@
 #ifndef TNT_FILAMENT_BACKEND_SYSTRACEPROFILE_H
 #define TNT_FILAMENT_BACKEND_SYSTRACEPROFILE_H
 
-#include <utils/Systrace.h>
+#include <private/utils/Tracing.h>
 
 #define PROFILE_SCOPE(marker)       SYSTRACE_NAME(marker)
 

filament/backend/src/VirtualMachineEnv.cpp

@@ -50,7 +50,7 @@ JavaVM* VirtualMachineEnv::getVirtualMachine() {
  */
 UTILS_PUBLIC
 UTILS_NOINLINE
-jint VirtualMachineEnv::JNI_OnLoad(JavaVM* vm) noexcept {
+jint VirtualMachineEnv::JNI_OnLoad(JavaVM* vm) {
     std::lock_guard const lock(sLock);
     if (sVirtualMachine) {
         // It doesn't make sense for JNI_OnLoad() to be called more than once
@@ -77,7 +77,7 @@ VirtualMachineEnv& VirtualMachineEnv::get() noexcept {
 }
 
 UTILS_NOINLINE
-JNIEnv* VirtualMachineEnv::getThreadEnvironment() noexcept {
+JNIEnv* VirtualMachineEnv::getThreadEnvironment() {
     JavaVM* const vm = getVirtualMachine();
     JNIEnv* env = nullptr;
     jint const result = vm->GetEnv(reinterpret_cast<void**>(&env), JNI_VERSION_1_6);
@@ -101,7 +101,7 @@ VirtualMachineEnv::~VirtualMachineEnv() noexcept {
 }
 
 UTILS_NOINLINE
-JNIEnv* VirtualMachineEnv::getEnvironmentSlow() noexcept {
+JNIEnv* VirtualMachineEnv::getEnvironmentSlow() {
     FILAMENT_CHECK_PRECONDITION(mVirtualMachine)
             << "JNI_OnLoad() has not been called";
 

filament/backend/src/metal/MetalDriver.mm

@@ -672,22 +672,8 @@ void MetalDriver::createTimerQueryR(Handle<HwTimerQuery> tqh, int) {
     // nothing to do, timer query was constructed in createTimerQueryS
 }
 
-const char* toString(DescriptorType type) {
-    switch (type) {
-        case DescriptorType::UNIFORM_BUFFER:
-            return "UNIFORM_BUFFER";
-        case DescriptorType::SHADER_STORAGE_BUFFER:
-            return "SHADER_STORAGE_BUFFER";
-        case DescriptorType::SAMPLER:
-            return "SAMPLER";
-        case DescriptorType::INPUT_ATTACHMENT:
-            return "INPUT_ATTACHMENT";
-        case DescriptorType::SAMPLER_EXTERNAL:
-            return "SAMPLER_EXTERNAL";
-    }
-}
-
-const char* toString(ShaderStageFlags flags) {
+UTILS_UNUSED
+static const char* toString(ShaderStageFlags flags) {
     std::vector<const char*> stages;
     if (any(flags & ShaderStageFlags::VERTEX)) {
         stages.push_back("VERTEX");

filament/backend/src/metal/MetalHandles.mm

@@ -1356,7 +1356,31 @@ id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoderSlow(id<MTLDe
                 [arguments addObject:bufferArgument];
                 break;
             }
-            case DescriptorType::SAMPLER:
+            case DescriptorType::SAMPLER_2D_FLOAT:
+            case DescriptorType::SAMPLER_2D_INT:
+            case DescriptorType::SAMPLER_2D_UINT:
+            case DescriptorType::SAMPLER_2D_DEPTH:
+            case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+            case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_INT:
+            case DescriptorType::SAMPLER_CUBE_UINT:
+            case DescriptorType::SAMPLER_CUBE_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_3D_FLOAT:
+            case DescriptorType::SAMPLER_3D_INT:
+            case DescriptorType::SAMPLER_3D_UINT:
+            case DescriptorType::SAMPLER_2D_MS_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_INT:
+            case DescriptorType::SAMPLER_2D_MS_UINT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
             case DescriptorType::SAMPLER_EXTERNAL: {
                 MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
                 textureArgument.index = binding.binding * 2;
@@ -1478,7 +1502,31 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
                            atIndex:binding.binding * 2];
                 break;
             }
-            case DescriptorType::SAMPLER:
+            case DescriptorType::SAMPLER_2D_FLOAT:
+            case DescriptorType::SAMPLER_2D_INT:
+            case DescriptorType::SAMPLER_2D_UINT:
+            case DescriptorType::SAMPLER_2D_DEPTH:
+            case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+            case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_INT:
+            case DescriptorType::SAMPLER_CUBE_UINT:
+            case DescriptorType::SAMPLER_CUBE_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_3D_FLOAT:
+            case DescriptorType::SAMPLER_3D_INT:
+            case DescriptorType::SAMPLER_3D_UINT:
+            case DescriptorType::SAMPLER_2D_MS_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_INT:
+            case DescriptorType::SAMPLER_2D_MS_UINT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
             case DescriptorType::SAMPLER_EXTERNAL: {
                 auto found = textures.find(binding.binding);
                 if (found == textures.end()) {

filament/backend/src/opengl/BindingMap.h

@@ -69,9 +69,8 @@ public:
         assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
         assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
         assert_invariant(entry.binding < 128); // we reserve 1 bit for the type right now
-        mStorage[set][binding] = { (uint8_t)entry.binding,
-                                   entry.type == DescriptorType::SAMPLER ||
-                                   entry.type == DescriptorType::SAMPLER_EXTERNAL };
+        mStorage[set][binding] = { uint8_t(entry.binding),
+                                   DescriptorSetLayoutBinding::isSampler(entry.type) };
         mActiveDescriptors[set].set(binding);
     }
 

filament/backend/src/opengl/GLDescriptorSet.cpp

@@ -95,7 +95,32 @@ GLDescriptorSet::GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle ds
                 }
                 break;
             }
-            case DescriptorType::SAMPLER:
+
+            case DescriptorType::SAMPLER_2D_FLOAT:
+            case DescriptorType::SAMPLER_2D_INT:
+            case DescriptorType::SAMPLER_2D_UINT:
+            case DescriptorType::SAMPLER_2D_DEPTH:
+            case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+            case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_INT:
+            case DescriptorType::SAMPLER_CUBE_UINT:
+            case DescriptorType::SAMPLER_CUBE_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_3D_FLOAT:
+            case DescriptorType::SAMPLER_3D_INT:
+            case DescriptorType::SAMPLER_3D_UINT:
+            case DescriptorType::SAMPLER_2D_MS_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_INT:
+            case DescriptorType::SAMPLER_2D_MS_UINT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
             case DescriptorType::SAMPLER_EXTERNAL:
                 if (UTILS_UNLIKELY(gl.isES2())) {
                     desc.emplace<SamplerGLES2>();
@@ -137,8 +162,11 @@ void GLDescriptorSet::update(OpenGLContext&,
     }, descriptors[binding].desc);
 }
 
-void GLDescriptorSet::update(OpenGLContext& gl,
-        descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept {
+void GLDescriptorSet::update(OpenGLContext& gl, HandleAllocatorGL& handleAllocator,
+        descriptor_binding_t binding, TextureHandle th, SamplerParams params) noexcept {
+
+    GLTexture* t = th ? handleAllocator.handle_cast<GLTexture*>(th) : nullptr;
+
     assert_invariant(binding < descriptors.size());
     std::visit([=, &gl](auto&& arg) mutable {
         using T = std::decay_t<decltype(arg)>;
@@ -171,20 +199,12 @@ void GLDescriptorSet::update(OpenGLContext& gl,
                 }
             }
 
-            arg.target = t ? t->gl.target : 0;
-            arg.id = t ? t->gl.id : 0;
-            arg.external = t ? t->gl.external :  false;
+            arg.handle = th;
             if constexpr (std::is_same_v<T, Sampler> ||
                           std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
                 if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
                     arg.anisotropy = float(1u << params.anisotropyLog2);
                 }
-                if (t) {
-                    arg.ref = t->ref;
-                    arg.baseLevel = t->gl.baseLevel;
-                    arg.maxLevel = t->gl.maxLevel;
-                    arg.swizzle = t->gl.swizzle;
-                }
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
                 arg.sampler = gl.getSampler(params);
 #else
@@ -200,39 +220,39 @@ void GLDescriptorSet::update(OpenGLContext& gl,
     }, descriptors[binding].desc);
 }
 
-template<typename T>
 void GLDescriptorSet::updateTextureView(OpenGLContext& gl,
-        HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept {
+        HandleAllocatorGL& handleAllocator, GLuint unit, GLTexture const* t) noexcept {
     // The common case is that we don't have a ref handle (we only have one if
     // the texture ever had a View on it).
-    assert_invariant(desc.ref);
-    GLTextureRef* const ref = handleAllocator.handle_cast<GLTextureRef*>(desc.ref);
-    if (UTILS_UNLIKELY((desc.baseLevel != ref->baseLevel || desc.maxLevel != ref->maxLevel))) {
+    assert_invariant(t);
+    assert_invariant(t->ref);
+    GLTextureRef* const ref = handleAllocator.handle_cast<GLTextureRef*>(t->ref);
+    if (UTILS_UNLIKELY((t->gl.baseLevel != ref->baseLevel || t->gl.maxLevel != ref->maxLevel))) {
         // If we have views, then it's still uncommon that we'll switch often
         // handle the case where we reset to the original texture
-        GLint baseLevel = GLint(desc.baseLevel); // NOLINT(*-signed-char-misuse)
-        GLint maxLevel = GLint(desc.maxLevel); // NOLINT(*-signed-char-misuse)
+        GLint baseLevel = GLint(t->gl.baseLevel); // NOLINT(*-signed-char-misuse)
+        GLint maxLevel = GLint(t->gl.maxLevel); // NOLINT(*-signed-char-misuse)
         if (baseLevel > maxLevel) {
             baseLevel = 0;
             maxLevel = 1000; // per OpenGL spec
         }
         // that is very unfortunate that we have to call activeTexture here
         gl.activeTexture(unit);
-        glTexParameteri(desc.target, GL_TEXTURE_BASE_LEVEL, baseLevel);
-        glTexParameteri(desc.target, GL_TEXTURE_MAX_LEVEL,  maxLevel);
-        ref->baseLevel = desc.baseLevel;
-        ref->maxLevel = desc.maxLevel;
+        glTexParameteri(t->gl.target, GL_TEXTURE_BASE_LEVEL, baseLevel);
+        glTexParameteri(t->gl.target, GL_TEXTURE_MAX_LEVEL,  maxLevel);
+        ref->baseLevel = t->gl.baseLevel;
+        ref->maxLevel = t->gl.maxLevel;
     }
-    if (UTILS_UNLIKELY(desc.swizzle != ref->swizzle)) {
+    if (UTILS_UNLIKELY(t->gl.swizzle != ref->swizzle)) {
         using namespace GLUtils;
         gl.activeTexture(unit);
 #if !defined(__EMSCRIPTEN__)  && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_R, (GLint)getSwizzleChannel(desc.swizzle[0]));
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_G, (GLint)getSwizzleChannel(desc.swizzle[1]));
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_B, (GLint)getSwizzleChannel(desc.swizzle[2]));
-        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_A, (GLint)getSwizzleChannel(desc.swizzle[3]));
+        glTexParameteri(t->gl.target, GL_TEXTURE_SWIZZLE_R, (GLint)getSwizzleChannel(t->gl.swizzle[0]));
+        glTexParameteri(t->gl.target, GL_TEXTURE_SWIZZLE_G, (GLint)getSwizzleChannel(t->gl.swizzle[1]));
+        glTexParameteri(t->gl.target, GL_TEXTURE_SWIZZLE_B, (GLint)getSwizzleChannel(t->gl.swizzle[2]));
+        glTexParameteri(t->gl.target, GL_TEXTURE_SWIZZLE_A, (GLint)getSwizzleChannel(t->gl.swizzle[3]));
 #endif
-        ref->swizzle = desc.swizzle;
+        ref->swizzle = t->gl.swizzle;
     }
 }
 
@@ -285,27 +305,31 @@ void GLDescriptorSet::bind(
                 }
             } else if constexpr (std::is_same_v<T, Sampler>) {
                 GLuint const unit = p.getTextureUnit(set, binding);
-                if (arg.target) {
-                    gl.bindTexture(unit, arg.target, arg.id, arg.external);
+
+
+                if (arg.handle) {
+                    GLTexture const* const t = handleAllocator.handle_cast<GLTexture*>(arg.handle);
+                    gl.bindTexture(unit, t->gl.target, t->gl.id, t->gl.external);
                     gl.bindSampler(unit, arg.sampler);
-                    if (UTILS_UNLIKELY(arg.ref)) {
-                        updateTextureView(gl, handleAllocator, unit, arg);
+                    if (UTILS_UNLIKELY(t->ref)) {
+                        updateTextureView(gl, handleAllocator, unit, t);
                     }
                 } else {
                     gl.unbindTextureUnit(unit);
                 }
             } else if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
                 GLuint const unit = p.getTextureUnit(set, binding);
-                if (arg.target) {
-                    gl.bindTexture(unit, arg.target, arg.id, arg.external);
+                if (arg.handle) {
+                    GLTexture const* const t = handleAllocator.handle_cast<GLTexture*>(arg.handle);
+                    gl.bindTexture(unit, t->gl.target, t->gl.id, t->gl.external);
                     gl.bindSampler(unit, arg.sampler);
-                    if (UTILS_UNLIKELY(arg.ref)) {
-                        updateTextureView(gl, handleAllocator, unit, arg);
+                    if (UTILS_UNLIKELY(t->ref)) {
+                        updateTextureView(gl, handleAllocator, unit, t);
                     }
 #if defined(GL_EXT_texture_filter_anisotropic)
                     // Driver claims to support anisotropic filtering, but it fails when set on
                     // the sampler, we have to set it on the texture instead.
-                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                    glTexParameterf(t->gl.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
                             std::min(gl.gets.max_anisotropy, float(arg.anisotropy)));
 #endif
                 } else {
@@ -314,19 +338,20 @@ void GLDescriptorSet::bind(
             } else if constexpr (std::is_same_v<T, SamplerGLES2>) {
                 // in ES2 the sampler parameters need to be set on the texture itself
                 GLuint const unit = p.getTextureUnit(set, binding);
-                if (arg.target) {
-                    gl.bindTexture(unit, arg.target, arg.id, arg.external);
+                if (arg.handle) {
+                    GLTexture const* const t = handleAllocator.handle_cast<GLTexture*>(arg.handle);
+                    gl.bindTexture(unit, t->gl.target, t->gl.id, t->gl.external);
                     SamplerParams const params = arg.params;
-                    glTexParameteri(arg.target, GL_TEXTURE_MIN_FILTER,
+                    glTexParameteri(t->gl.target, GL_TEXTURE_MIN_FILTER,
                             (GLint)GLUtils::getTextureFilter(params.filterMin));
-                    glTexParameteri(arg.target, GL_TEXTURE_MAG_FILTER,
+                    glTexParameteri(t->gl.target, GL_TEXTURE_MAG_FILTER,
                             (GLint)GLUtils::getTextureFilter(params.filterMag));
-                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_S,
+                    glTexParameteri(t->gl.target, GL_TEXTURE_WRAP_S,
                             (GLint)GLUtils::getWrapMode(params.wrapS));
-                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_T,
+                    glTexParameteri(t->gl.target, GL_TEXTURE_WRAP_T,
                             (GLint)GLUtils::getWrapMode(params.wrapT));
 #if defined(GL_EXT_texture_filter_anisotropic)
-                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                    glTexParameterf(t->gl.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
                             std::min(gl.gets.max_anisotropy, arg.anisotropy));
 #endif
                 } else {

filament/backend/src/opengl/GLDescriptorSet.h

@@ -59,8 +59,8 @@ struct GLDescriptorSet : public HwDescriptorSet {
             descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept;
 
     // update a sampler descriptor in the set
-    void update(OpenGLContext& gl,
-            descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept;
+    void update(OpenGLContext& gl, HandleAllocatorGL& handleAllocator,
+            descriptor_binding_t binding, TextureHandle th, SamplerParams params) noexcept;
 
     // conceptually bind the set to the command buffer
     void bind(
@@ -111,46 +111,19 @@ private:
 
     // A sampler descriptor
     struct Sampler {
-        uint16_t target;                        // 2 (GLenum)
-        bool external = false;                  // 1
-        bool reserved = false;                  // 1
-        GLuint id = 0;                          // 4
+        TextureHandle handle;                   // 4
         GLuint sampler = 0;                     // 4
-        Handle<GLTextureRef> ref;               // 4
-        int8_t baseLevel = 0x7f;                // 1
-        int8_t maxLevel = -1;                   // 1
-        std::array<TextureSwizzle, 4> swizzle{  // 4
-                TextureSwizzle::CHANNEL_0,
-                TextureSwizzle::CHANNEL_1,
-                TextureSwizzle::CHANNEL_2,
-                TextureSwizzle::CHANNEL_3
-        };
     };
 
     struct SamplerWithAnisotropyWorkaround {
-        uint16_t target;                        // 2 (GLenum)
-        bool external = false;                  // 1
-        bool reserved = false;                  // 1
-        GLuint id = 0;                          // 4
+        TextureHandle handle;                   // 4
         GLuint sampler = 0;                     // 4
-        Handle<GLTextureRef> ref;               // 4
         math::half anisotropy = 1.0f;           // 2
-        int8_t baseLevel = 0x7f;                // 1
-        int8_t maxLevel = -1;                   // 1
-        std::array<TextureSwizzle, 4> swizzle{  // 4
-                TextureSwizzle::CHANNEL_0,
-                TextureSwizzle::CHANNEL_1,
-                TextureSwizzle::CHANNEL_2,
-                TextureSwizzle::CHANNEL_3
-        };
     };
 
     // A sampler descriptor for ES2
     struct SamplerGLES2 {
-        uint16_t target;                        // 2 (GLenum)
-        bool external = false;                  // 1
-        bool reserved = false;                  // 1
-        GLuint id = 0;                          // 4
+        TextureHandle handle;                   // 4
         SamplerParams params{};                 // 4
         float anisotropy = 1.0f;                // 4
     };
@@ -165,9 +138,8 @@ private:
     };
     static_assert(sizeof(Descriptor) <= 32);
 
-    template<typename T>
     static void updateTextureView(OpenGLContext& gl,
-            HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept;
+            HandleAllocatorGL& handleAllocator, GLuint unit, GLTexture const* t) noexcept;
 
     utils::FixedCapacityVector<Descriptor> descriptors;     // 16
     utils::bitset64 dynamicBuffers;                         // 8

filament/backend/src/opengl/OpenGLBlobCache.cpp

@@ -21,7 +21,7 @@
 #include <backend/Platform.h>
 #include <backend/Program.h>
 
-#include <utils/Systrace.h>
+#include <private/utils/Tracing.h>
 
 namespace filament::backend {
 
@@ -36,7 +36,7 @@ OpenGLBlobCache::OpenGLBlobCache(OpenGLContext& gl) noexcept
 
 GLuint OpenGLBlobCache::retrieve(BlobCacheKey* outKey, Platform& platform,
         Program const& program) const noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
     if (!mCachingSupported || !platform.hasRetrieveBlobFunc()) {
         // the key is never updated in that case
         return 0;
@@ -68,7 +68,7 @@ GLuint OpenGLBlobCache::retrieve(BlobCacheKey* outKey, Platform& platform,
         programId = glCreateProgram();
 
         { // scope for systrace
-            SYSTRACE_NAME("glProgramBinary");
+            FILAMENT_TRACING_NAME(FILAMENT_TRACING_CATEGORY_FILAMENT, "glProgramBinary");
             glProgramBinary(programId, blob->format, blob->data, programBinarySize);
         }
 
@@ -90,7 +90,7 @@ GLuint OpenGLBlobCache::retrieve(BlobCacheKey* outKey, Platform& platform,
 
 void OpenGLBlobCache::insert(Platform& platform,
         BlobCacheKey const& key, GLuint program) noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
     if (!mCachingSupported || !platform.hasInsertBlobFunc()) {
         // the key is never updated in that case
         return;
@@ -100,7 +100,7 @@ void OpenGLBlobCache::insert(Platform& platform,
     GLenum format;
     GLint programBinarySize = 0;
     { // scope for systrace
-        SYSTRACE_NAME("glGetProgramiv");
+        FILAMENT_TRACING_NAME(FILAMENT_TRACING_CATEGORY_FILAMENT, "glGetProgramiv");
         glGetProgramiv(program, GL_PROGRAM_BINARY_LENGTH, &programBinarySize);
     }
     if (programBinarySize) {
@@ -108,7 +108,7 @@ void OpenGLBlobCache::insert(Platform& platform,
         std::unique_ptr<Blob, decltype(&::free)> blob{ (Blob*)malloc(size), &::free };
         if (UTILS_LIKELY(blob)) {
             { // scope for systrace
-                SYSTRACE_NAME("glGetProgramBinary");
+                FILAMENT_TRACING_NAME(FILAMENT_TRACING_CATEGORY_FILAMENT, "glGetProgramBinary");
                 glGetProgramBinary(program, programBinarySize,
                         &programBinarySize, &format, blob->data);
             }

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -44,6 +44,8 @@
 #include "private/backend/Dispatcher.h"
 #include "private/backend/DriverApi.h"
 
+#include <private/utils/Tracing.h>
+
 #include <type_traits>
 #include <utils/BitmaskEnum.h>
 #include <utils/FixedCapacityVector.h>
@@ -51,7 +53,6 @@
 #include <utils/Invocable.h>
 #include <utils/Log.h>
 #include <utils/Panic.h>
-#include <utils/Systrace.h>
 #include <utils/Slice.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -98,13 +99,13 @@
 
 #define DEBUG_GROUP_MARKER_NONE       0x00    // no debug marker
 #define DEBUG_GROUP_MARKER_OPENGL     0x01    // markers in the gl command queue (req. driver support)
-#define DEBUG_GROUP_MARKER_BACKEND    0x02    // markers on the backend side (systrace)
+#define DEBUG_GROUP_MARKER_BACKEND    0x02    // markers on the backend side (perfetto)
 #define DEBUG_GROUP_MARKER_ALL        0xFF    // all markers
 
 #define DEBUG_MARKER_NONE             0x00    // no debug marker
 #define DEBUG_MARKER_OPENGL           0x01    // markers in the gl command queue (req. driver support)
-#define DEBUG_MARKER_BACKEND          0x02    // markers on the backend side (systrace)
-#define DEBUG_MARKER_PROFILE          0x04    // profiling on the backend side (systrace)
+#define DEBUG_MARKER_BACKEND          0x02    // markers on the backend side (perfetto)
+#define DEBUG_MARKER_PROFILE          0x04    // profiling on the backend side (perfetto)
 #define DEBUG_MARKER_ALL              (0xFF & ~DEBUG_MARKER_PROFILE) // all markers
 
 // set to the desired debug marker level (for user markers [default: All])
@@ -248,8 +249,8 @@ OpenGLDriver::DebugMarker::DebugMarker(OpenGLDriver& driver, const char* string)
 #endif
 
 #if DEBUG_MARKER_LEVEL & DEBUG_MARKER_BACKEND
-    SYSTRACE_CONTEXT();
-    SYSTRACE_NAME_BEGIN(string);
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_NAME_BEGIN(FILAMENT_TRACING_CATEGORY_FILAMENT, string);
 #endif
 #endif
 }
@@ -265,8 +266,8 @@ OpenGLDriver::DebugMarker::~DebugMarker() noexcept {
 #endif
 
 #if DEBUG_MARKER_LEVEL & DEBUG_MARKER_BACKEND
-    SYSTRACE_CONTEXT();
-    SYSTRACE_NAME_END();
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_NAME_END(FILAMENT_TRACING_CATEGORY_FILAMENT);
 #endif
 #endif
 }
@@ -931,7 +932,12 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
 #if defined(BACKEND_OPENGL_LEVEL_GLES31)
                 if (gl.features.multisample_texture) {
                     // multi-sample texture on GL 3.2 / GLES 3.1 and above
-                    t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
+                    if (depth <= 1) {
+                        // We forcibly change the target to 2D-multisample only for flat texture.
+                        // A depth value greater than 1 may indicate multiview usage, which requires
+                        // GL_TEXTURE_2D_ARRAY. Also 2D MSAA won't work with non-flat texture anyway.
+                        t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
+                    }
                 } else {
                     // Turn off multi-sampling for that texture. It's just not supported.
                 }
@@ -1207,7 +1213,12 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
 #if defined(BACKEND_OPENGL_LEVEL_GLES31)
         if (gl.features.multisample_texture) {
             // multi-sample texture on GL 3.2 / GLES 3.1 and above
-            t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
+            if (depth <= 1) {
+                // We forcibly change the target to 2D-multisample only for flat texture.
+                // A depth value greater than 1 may indicate multiview usage, which requires
+                // GL_TEXTURE_2D_ARRAY. Also 2D MSAA won't work with non-flat texture anyway.
+                t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
+            }
         } else {
             // Turn off multi-sampling for that texture. It's just not supported.
         }
@@ -1452,8 +1463,15 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
 #if !defined(__EMSCRIPTEN__) && !defined(FILAMENT_IOS)
                 if (layerCount > 1) {
                     // if layerCount > 1, it means we use the multiview extension.
-                    glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, attachment,
-                        t->gl.id, 0, binfo.layer, layerCount);
+                    if (rt->gl.samples > 1) {
+                        // For MSAA
+                        glFramebufferTextureMultisampleMultiviewOVR(GL_FRAMEBUFFER, attachment,
+                                t->gl.id, 0, rt->gl.samples, binfo.layer, layerCount);
+                    }
+                    else {
+                        glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, attachment, t->gl.id, 0,
+                                binfo.layer, layerCount);
+                    }
                 } else
 #endif // !defined(__EMSCRIPTEN__) && !defined(FILAMENT_IOS)
                 {
@@ -2513,10 +2531,33 @@ void OpenGLDriver::makeCurrent(Handle<HwSwapChain> schDraw, Handle<HwSwapChain>
 
     mPlatform.makeCurrent(scDraw->swapChain, scRead->swapChain,
             [this]() {
+                for (auto t: mTexturesWithStreamsAttached) {
+                    if (t->hwStream->streamType == StreamType::NATIVE) {
+                        mPlatform.detach(t->hwStream->stream);
+                    }
+                }
                 // OpenGL context is about to change, unbind everything
                 mContext.unbindEverything();
             },
             [this](size_t index) {
+                for (auto t: mTexturesWithStreamsAttached) {
+                    if (t->hwStream->streamType == StreamType::NATIVE) {
+                        if (t->externalTexture) {
+                            glGenTextures(1, &t->externalTexture->id);
+                            t->gl.id = t->externalTexture->id;
+                        } else {
+                            glGenTextures(1, &t->gl.id);
+                        }
+                        mPlatform.attach(t->hwStream->stream, t->gl.id);
+                        mContext.updateTexImage(GL_TEXTURE_EXTERNAL_OES, t->gl.id);
+                    }
+                }
+
+                // force invalidation of all bound descriptor sets
+                decltype(mInvalidDescriptorSetBindings) changed;
+                changed.setValue((1 << MAX_DESCRIPTOR_SET_COUNT) - 1);
+                mInvalidDescriptorSetBindings |= changed;
+
                 // OpenGL context has changed, resynchronize the state with the cache
                 mContext.synchronizeStateAndCache(index);
                 slog.d << "*** OpenGL context change : " << (index ? "protected" : "default") << io::endl;
@@ -2991,6 +3032,7 @@ void OpenGLDriver::attachStream(GLTexture* t, GLStream* hwStream) noexcept {
     switch (hwStream->streamType) {
         case StreamType::NATIVE:
             mPlatform.attach(hwStream->stream, t->gl.id);
+            mContext.updateTexImage(GL_TEXTURE_EXTERNAL_OES, t->gl.id);
             break;
         case StreamType::ACQUIRED:
             break;
@@ -3019,7 +3061,12 @@ void OpenGLDriver::detachStream(GLTexture* t) noexcept {
             break;
     }
 
-    glGenTextures(1, &t->gl.id);
+    if (t->externalTexture) {
+        glGenTextures(1, &t->externalTexture->id);
+        t->gl.id = t->externalTexture->id;
+    } else {
+        glGenTextures(1, &t->gl.id);
+    }
 
     t->hwStream = nullptr;
 }
@@ -3043,8 +3090,14 @@ void OpenGLDriver::replaceStream(GLTexture* texture, GLStream* newStream) noexce
 
     switch (newStream->streamType) {
         case StreamType::NATIVE:
-            glGenTextures(1, &texture->gl.id);
+            if (texture->externalTexture) {
+                glGenTextures(1, &texture->externalTexture->id);
+                texture->gl.id = texture->externalTexture->id;
+            } else {
+                glGenTextures(1, &texture->gl.id);
+            }
             mPlatform.attach(newStream->stream, texture->gl.id);
+            mContext.updateTexImage(GL_TEXTURE_EXTERNAL_OES, texture->gl.id);
             break;
         case StreamType::ACQUIRED:
             // Just re-use the old texture id.
@@ -3348,8 +3401,8 @@ void OpenGLDriver::pushGroupMarker(char const* string) {
 #endif
 
 #if DEBUG_GROUP_MARKER_LEVEL & DEBUG_GROUP_MARKER_BACKEND
-    SYSTRACE_CONTEXT();
-    SYSTRACE_NAME_BEGIN(string);
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_NAME_BEGIN(FILAMENT_TRACING_CATEGORY_FILAMENT, string);
 #endif
 #endif
 }
@@ -3365,8 +3418,8 @@ void OpenGLDriver::popGroupMarker(int) {
 #endif
 
 #if DEBUG_GROUP_MARKER_LEVEL & DEBUG_GROUP_MARKER_BACKEND
-    SYSTRACE_CONTEXT();
-    SYSTRACE_NAME_END();
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_NAME_END(FILAMENT_TRACING_CATEGORY_FILAMENT);
 #endif
 #endif
 }
@@ -3685,7 +3738,7 @@ void OpenGLDriver::endFrame(UTILS_UNUSED uint32_t frameId) {
     gl.depthFunc(GL_LESS);
     gl.disable(GL_SCISSOR_TEST);
 #endif
-    //SYSTRACE_NAME("glFinish");
+    //FILAMENT_TRACING_NAME(FILAMENT_TRACING_CATEGORY_FILAMENT, "glFinish");
     //glFinish();
     mPlatform.endFrame(frameId);
     insertEventMarker("endFrame");
@@ -3707,8 +3760,7 @@ void OpenGLDriver::updateDescriptorSetTexture(
         TextureHandle th,
         SamplerParams params) {
     GLDescriptorSet* ds = handle_cast<GLDescriptorSet*>(dsh);
-    GLTexture* t = th ? handle_cast<GLTexture*>(th) : nullptr;
-    ds->update(mContext, binding, t, params);
+    ds->update(mContext, mHandleAllocator, binding, th, params);
 }
 
 void OpenGLDriver::flush(int) {

filament/backend/src/opengl/OpenGLPlatform.cpp

@@ -64,7 +64,7 @@ utils::CString OpenGLPlatform::getRendererString(Driver const* driver) {
 }
 
 void OpenGLPlatform::makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain,
-        utils::Invocable<void()>, utils::Invocable<void(size_t)>) noexcept {
+        utils::Invocable<void()>, utils::Invocable<void(size_t)>) {
     makeCurrent(getCurrentContextType(), drawSwapChain, readSwapChain);
 }
 

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -25,12 +25,13 @@
 #include <backend/Program.h>
 #include <backend/Handle.h>
 
+#include <private/utils/Tracing.h>
+
 #include <utils/BitmaskEnum.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Log.h>
-#include <utils/Systrace.h>
 
 #include <algorithm>
 #include <array>
@@ -96,8 +97,7 @@ OpenGLProgram::~OpenGLProgram() noexcept {
 }
 
 void OpenGLProgram::initialize(OpenGLDriver& gld) {
-
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     assert_invariant(gl.program == 0);
     assert_invariant(mToken);
@@ -122,8 +122,7 @@ void OpenGLProgram::initialize(OpenGLDriver& gld) {
  */
 void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint program,
         LazyInitializationData& lazyInitializationData) noexcept {
-
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     // from the pipeline layout we compute a mapping from {set, binding} to {binding}
     // for both buffers and textures
@@ -175,7 +174,31 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
                     }
                     break;
                 }
-                case DescriptorType::SAMPLER:
+                case DescriptorType::SAMPLER_2D_FLOAT:
+                case DescriptorType::SAMPLER_2D_INT:
+                case DescriptorType::SAMPLER_2D_UINT:
+                case DescriptorType::SAMPLER_2D_DEPTH:
+                case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+                case DescriptorType::SAMPLER_2D_ARRAY_INT:
+                case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+                case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+                case DescriptorType::SAMPLER_CUBE_FLOAT:
+                case DescriptorType::SAMPLER_CUBE_INT:
+                case DescriptorType::SAMPLER_CUBE_UINT:
+                case DescriptorType::SAMPLER_CUBE_DEPTH:
+                case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+                case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+                case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+                case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+                case DescriptorType::SAMPLER_3D_FLOAT:
+                case DescriptorType::SAMPLER_3D_INT:
+                case DescriptorType::SAMPLER_3D_UINT:
+                case DescriptorType::SAMPLER_2D_MS_FLOAT:
+                case DescriptorType::SAMPLER_2D_MS_INT:
+                case DescriptorType::SAMPLER_2D_MS_UINT:
+                case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+                case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+                case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT:
                 case DescriptorType::SAMPLER_EXTERNAL: {
                     if (!entry.name.empty()) {
                         GLint const loc = glGetUniformLocation(program, entry.name.c_str());

filament/backend/src/opengl/OpenGLTimerQuery.cpp

@@ -23,12 +23,13 @@
 #include <backend/platforms/OpenGLPlatform.h>
 #include <backend/DriverEnums.h>
 
+#include <private/utils/Tracing.h>
+
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/JobSystem.h>
 #include <utils/Log.h>
 #include <utils/Mutex.h>
-#include <utils/Systrace.h>
 
 #include <atomic>
 #include <memory>
@@ -240,8 +241,8 @@ void TimerQueryFenceFactory::beginTimeElapsedQuery(GLTimerQuery* tq) {
         if (state) {
             platform.waitFence(fence, FENCE_WAIT_FOR_EVER);
             state->then = clock::now().time_since_epoch().count();
-            SYSTRACE_CONTEXT();
-            SYSTRACE_ASYNC_BEGIN("OpenGLTimerQueryFence", intptr_t(state.get()));
+            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+            FILAMENT_TRACING_ASYNC_BEGIN(FILAMENT_TRACING_CATEGORY_FILAMENT, "OpenGLTimerQueryFence", intptr_t(state.get()));
         }
         platform.destroyFence(fence);
     });
@@ -257,8 +258,8 @@ void TimerQueryFenceFactory::endTimeElapsedQuery(OpenGLDriver&, GLTimerQuery* tq
             platform.waitFence(fence, FENCE_WAIT_FOR_EVER);
             int64_t const now = clock::now().time_since_epoch().count();
             state->elapsed.store(now - state->then, std::memory_order_relaxed);
-            SYSTRACE_CONTEXT();
-            SYSTRACE_ASYNC_END("OpenGLTimerQueryFence", intptr_t(state.get()));
+            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+            FILAMENT_TRACING_ASYNC_END(FILAMENT_TRACING_CATEGORY_FILAMENT, "OpenGLTimerQueryFence", intptr_t(state.get()));
         }
         platform.destroyFence(fence);
     });

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -23,11 +23,14 @@
 #include "OpenGLDriver.h"
 
 #include <iterator>
+#include <optional>
 #include <private/backend/BackendUtils.h>
 
 #include <backend/DriverEnums.h>
 #include <backend/Program.h>
 
+#include <private/utils/Tracing.h>
+
 #include <utils/compiler.h>
 #include <utils/CString.h>
 #include <utils/debug.h>
@@ -36,7 +39,6 @@
 #include <utils/Log.h>
 #include <utils/ostream.h>
 #include <utils/Panic.h>
-#include <utils/Systrace.h>
 
 #include <algorithm>
 #include <array>
@@ -109,7 +111,17 @@ struct ShaderCompilerService::OpenGLProgramToken : ProgramToken {
         cond.wait(l, [this] { return signaled; });
     }
 
-    CallbackManager::Handle handle{};
+    // This is invoked upon token completion, which occurs after a successful `gl.program`
+    // population or upon cancellation. In either scenario, the callback handle must be submitted
+    // to notify the caller that resource loading has concluded.
+    void trySubmittingCallback() noexcept {
+        if (handle) {
+            compiler.submitCallbackHandle(*handle);
+            handle = std::nullopt;
+        }
+    }
+
+    std::optional<CallbackManager::Handle> handle{};
     BlobCacheKey key;
 
     // Used for the `THREAD_POOL` mode.
@@ -119,7 +131,7 @@ struct ShaderCompilerService::OpenGLProgramToken : ProgramToken {
 };
 
 ShaderCompilerService::OpenGLProgramToken::~OpenGLProgramToken() {
-    compiler.submitCallbackHandle(handle);
+    trySubmittingCallback();
 }
 
 /* static */ void ShaderCompilerService::setUserData(const program_token_t& token,
@@ -338,7 +350,7 @@ GLuint ShaderCompilerService::getProgram(program_token_t& token) {
 
     // Cleanup the token.
     token->compiler.cancelTickOp(token);
-    token = nullptr;// This will submit a callback condition (handle) to the callback manager.
+    token = nullptr; // This will try submitting a callback handle to the callback manager.
 }
 
 void ShaderCompilerService::tick() {
@@ -367,7 +379,7 @@ void ShaderCompilerService::notifyWhenAllProgramsAreReady(
 
 GLuint ShaderCompilerService::initialize(program_token_t& token) {
 
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     assert_invariant(token);// This function should be called when the token is still alive.
 
@@ -391,7 +403,7 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) {
 
     // Cleanup the token.
     token->compiler.cancelTickOp(token);
-    token = nullptr;// This will submit a callback condition (handle) to the callback manager.
+    token = nullptr;
 
     return program;
 }
@@ -454,8 +466,8 @@ void ShaderCompilerService::runAtNextTick(CompilerPriorityQueue priority,
             });
     ops.emplace(pos, priority, token, std::move(job));
 
-    SYSTRACE_CONTEXT();
-    SYSTRACE_VALUE32("ShaderCompilerService Jobs", mRunAtNextTickOps.size());
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "ShaderCompilerService Jobs", mRunAtNextTickOps.size());
 }
 
 bool ShaderCompilerService::cancelTickOp(program_token_t const& token) noexcept {
@@ -468,8 +480,8 @@ bool ShaderCompilerService::cancelTickOp(program_token_t const& token) noexcept
         ops.erase(pos);
         return true;
     }
-    SYSTRACE_CONTEXT();
-    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "ShaderCompilerService Jobs", ops.size());
     return false;
 }
 
@@ -485,15 +497,15 @@ void ShaderCompilerService::executeTickOps() noexcept {
             ++it;
         }
     }
-    SYSTRACE_CONTEXT();
-    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
+    FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
+    FILAMENT_TRACING_VALUE(FILAMENT_TRACING_CATEGORY_FILAMENT, "ShaderCompilerService Jobs", ops.size());
 }
 
 /* static */ void ShaderCompilerService::compileShaders(OpenGLContext& context,
         Program::ShaderSource shadersSource,
         FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
         bool multiview, program_token_t const& token) noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     auto const appendSpecConstantString = +[](std::string& s, Program::SpecializationConstant const& sc) {
         s += "#define SPIRV_CROSS_CONSTANT_ID_" + std::to_string(sc.id) + ' ';
@@ -616,7 +628,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
 }
 
 /* static */ void ShaderCompilerService::checkCompileStatus(program_token_t const& token) noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     UTILS_NOUNROLL
     for (size_t i = 0; i < Program::SHADER_TYPE_COUNT; i++) {
@@ -639,7 +651,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
 
 /* static */ void ShaderCompilerService::linkProgram(OpenGLContext const& context,
         program_token_t const& token) noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
 
     // Shader compilation should be completed by now. Check the status and log errors on failure.
     checkCompileStatus(token);
@@ -658,6 +670,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
     }
     glLinkProgram(program);
     token->gl.program = program;
+    token->trySubmittingCallback();
 }
 
 /* static */ bool ShaderCompilerService::isLinkCompleted(program_token_t const& token) noexcept {
@@ -675,7 +688,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
 
 /* static */ bool ShaderCompilerService::checkLinkStatusAndCleanupShaders(
         program_token_t const& token) noexcept {
-    SYSTRACE_CALL();
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
     assert_invariant(token->gl.program);
 
     bool linked = true;

filament/backend/src/opengl/gl_headers.cpp

@@ -70,6 +70,9 @@ PFNGLMAXSHADERCOMPILERTHREADSKHRPROC glMaxShaderCompilerThreadsKHR;
 #ifdef GL_OVR_multiview
 PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
 #endif
+#ifdef GL_OVR_multiview_multisampled_render_to_texture
+PFNGLFRAMEBUFFERTEXTUREMULTISAMPLEMULTIVIEWOVRPROC glFramebufferTextureMultisampleMultiviewOVR;
+#endif
 
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
 // On Android, If we want to support a build system less than ANDROID_API 21, we need to
@@ -123,6 +126,9 @@ void importGLESExtensionsEntryPoints() {
 #ifdef GL_OVR_multiview
         getProcAddress(glFramebufferTextureMultiviewOVR, "glFramebufferTextureMultiviewOVR");
 #endif
+#ifdef GL_OVR_multiview_multisampled_render_to_texture
+        getProcAddress(glFramebufferTextureMultisampleMultiviewOVR, "glFramebufferTextureMultisampleMultiviewOVR");
+#endif
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
         getProcAddress(glDispatchCompute, "glDispatchCompute");
 #endif

filament/backend/src/opengl/gl_headers.h

@@ -154,6 +154,9 @@ extern PFNGLMAXSHADERCOMPILERTHREADSKHRPROC glMaxShaderCompilerThreadsKHR;
 #ifdef GL_OVR_multiview
 extern PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
 #endif
+#ifdef GL_OVR_multiview_multisampled_render_to_texture
+extern PFNGLFRAMEBUFFERTEXTUREMULTISAMPLEMULTIVIEWOVRPROC glFramebufferTextureMultisampleMultiviewOVR;
+#endif
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
 extern PFNGLDISPATCHCOMPUTEPROC glDispatchCompute;
 #endif

filament/backend/src/opengl/platforms/PlatformCocoaGL.mm

@@ -255,7 +255,7 @@ void PlatformCocoaGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept
 }
 
 bool PlatformCocoaGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
             "ContextManagerCocoa does not support using distinct draw/read swap chains.");
     CocoaGLSwapChain* swapChain = (CocoaGLSwapChain*)drawSwapChain;

filament/backend/src/opengl/platforms/PlatformCocoaTouchGL.mm

@@ -155,7 +155,7 @@ uint32_t PlatformCocoaTouchGL::getDefaultFramebufferObject() noexcept {
 }
 
 bool PlatformCocoaTouchGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
             "PlatformCocoaTouchGL does not support using distinct draw/read swap chains.");
     CAEAGLLayer* const glLayer = (__bridge CAEAGLLayer*) drawSwapChain;

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

@@ -577,18 +577,18 @@ OpenGLPlatform::ContextType PlatformEGL::getCurrentContextType() const noexcept
 }
 
 bool PlatformEGL::makeCurrent(ContextType type,
-        SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept {
+        SwapChain* drawSwapChain, SwapChain* readSwapChain) {
     SwapChainEGL const* const dsc = static_cast<SwapChainEGL const*>(drawSwapChain);
     SwapChainEGL const* const rsc = static_cast<SwapChainEGL const*>(readSwapChain);
     EGLContext context = getContextForType(type);
     EGLBoolean const success = egl.makeCurrent(context, dsc->sur, rsc->sur);
-    return success == EGL_TRUE ? true : false;
+    return success == EGL_TRUE;
 }
 
 void PlatformEGL::makeCurrent(SwapChain* drawSwapChain,
         SwapChain* readSwapChain,
         Invocable<void()> preContextChange,
-        Invocable<void(size_t index)> postContextChange) noexcept {
+        Invocable<void(size_t index)> postContextChange) {
 
     assert_invariant(drawSwapChain);
     assert_invariant(readSwapChain);
@@ -796,7 +796,7 @@ void PlatformEGL::Config::erase(EGLint name) noexcept {
 // ------------------------------------------------------------------------------------------------
 
 EGLBoolean PlatformEGL::EGL::makeCurrent(EGLContext context, EGLSurface drawSurface,
-        EGLSurface readSurface) noexcept {
+        EGLSurface readSurface) {
     if (UTILS_UNLIKELY((
             mCurrentContext != context ||
             drawSurface != mCurrentDrawSurface || readSurface != mCurrentReadSurface))) {

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

@@ -135,7 +135,7 @@ static constexpr const std::string_view kNativeWindowInvalidMsg =
 
 bool PlatformEGLAndroid::makeCurrent(ContextType type,
         SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
 
     // fast & safe path
     if (UTILS_LIKELY(!mAssertNativeWindowIsValid)) {
@@ -255,6 +255,7 @@ Platform::ExternalImageHandle PlatformEGLAndroid::createExternalImage(AHardwareB
         p->height = hardwareBufferDescription.height;
         p->width = hardwareBufferDescription.width;
         auto textureFormat = mapToFilamentFormat(hardwareBufferDescription.format, sRGB);
+        p->format = textureFormat;
         p->usage = mapToFilamentUsage(hardwareBufferDescription.usage, textureFormat);
         return ExternalImageHandle{ p };
     }

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

@@ -266,7 +266,7 @@ void PlatformGLX::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
 }
 
 bool PlatformGLX::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
     g_glx.setCurrentContext(mGLXDisplay,
             (GLXDrawable)drawSwapChain, (GLXDrawable)readSwapChain, mGLXContext);
     return true;

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

@@ -169,7 +169,7 @@ void PlatformOSMesa::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
 }
 
 bool PlatformOSMesa::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
     OSMesaAPI* api = (OSMesaAPI*) mOsMesaApi;
     OSMesaSwapchain* impl = (OSMesaSwapchain*) drawSwapChain;
 

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

@@ -262,7 +262,7 @@ void PlatformWGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
 }
 
 bool PlatformWGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
                                   "PlatformWGL does not support distinct draw/read swap chains.");
 

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

@@ -47,7 +47,7 @@ void PlatformWebGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
 }
 
 bool PlatformWebGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
-        SwapChain* readSwapChain) noexcept {
+        SwapChain* readSwapChain) {
     return true;
 }
 

filament/backend/src/vulkan/VulkanBuffer.h

@@ -17,35 +17,36 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANBUFFER_H
 #define TNT_FILAMENT_BACKEND_VULKANBUFFER_H
 
-#include "VulkanContext.h"
-#include "VulkanStagePool.h"
 #include "VulkanMemory.h"
+#include "memory/Resource.h"
+
+#include <functional>
 
 namespace filament::backend {
 
-// Encapsulates a Vulkan buffer, its attached DeviceMemory and a staging area.
-class VulkanBuffer {
+class VulkanBuffer : public fvkmemory::Resource {
 public:
-    VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool, VkBufferUsageFlags usage,
-            uint32_t numBytes);
-    ~VulkanBuffer();
-    void loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
-            uint32_t numBytes);
-    VkBuffer getGpuBuffer() const {
-        return mGpuBuffer;
+    // Because we need to recycle the unused `VulkanGpuBuffer`, we allow for a callback that the
+    // "Pool" can use to acquire the buffer back.
+    using OnRecycle = std::function<void(VulkanGpuBuffer const*)>;
+
+    VulkanBuffer(VulkanGpuBuffer const* gpuBuffer, OnRecycle&& onRecycleFn)
+        : mGpuBuffer(gpuBuffer),
+          mOnRecycleFn(onRecycleFn) {}
+
+    ~VulkanBuffer() {
+        if (mOnRecycleFn) {
+            mOnRecycleFn(mGpuBuffer);
+        }
     }
 
-private:
-    VmaAllocator mAllocator;
-    VulkanStagePool& mStagePool;
+    VulkanGpuBuffer const* getGpuBuffer() const { return mGpuBuffer; }
 
-    VmaAllocation mGpuMemory = VK_NULL_HANDLE;
-    VkBuffer mGpuBuffer = VK_NULL_HANDLE;
-    VkBufferUsageFlags mUsage = {};
-	uint32_t mUpdatedOffset = 0;
-    uint32_t mUpdatedBytes = 0;
+private:
+    VulkanGpuBuffer const* mGpuBuffer;
+    OnRecycle mOnRecycleFn;
 };
 
-} // namespace filament::backend
+}// namespace filament::backend
 
-#endif // TNT_FILAMENT_BACKEND_VULKANBUFFER_H
+#endif// TNT_FILAMENT_BACKEND_VULKANBUFFER_H

filament/backend/src/vulkan/VulkanBufferCache.cpp

@@ -0,0 +1,207 @@
+/*
+ * Copyright (C) 2025 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 "VulkanBufferCache.h"
+
+#include "VulkanBuffer.h"
+#include "VulkanConstants.h"
+#include "VulkanMemory.h"
+#include "memory/Resource.h"
+#include "memory/ResourceManager.h"
+
+#include <utility>
+
+namespace filament::backend {
+
+namespace {
+
+VkBufferUsageFlags getVkBufferUsage(VulkanBufferUsage usage) {
+    switch (usage) {
+        case VulkanBufferUsage::VERTEX:
+            return VK_BUFFER_USAGE_VERTEX_BUFFER_BIT;
+        case VulkanBufferUsage::INDEX:
+            return VK_BUFFER_USAGE_INDEX_BUFFER_BIT;
+        case VulkanBufferUsage::UNIFORM:
+            return VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
+        case VulkanBufferUsage::SHADER_STORAGE:
+            return VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+        case VulkanBufferUsage::UNKNOWN:
+            return 0;
+    }
+
+    return 0;
+}
+
+}// namespace
+
+VulkanBufferCache::VulkanBufferCache(VulkanContext const& context,
+        fvkmemory::ResourceManager& resourceManager, VmaAllocator allocator)
+    : mContext(context),
+      mResourceManager(resourceManager),
+      mAllocator(allocator) {}
+
+fvkmemory::resource_ptr<VulkanBuffer> VulkanBufferCache::acquire(VulkanBufferUsage usage,
+        uint32_t numBytes) noexcept {
+    assert_invariant(usage != VulkanBufferUsage::UNKNOWN);
+
+    BufferPool& bufferPool = getPool(usage);
+
+    // First check if an allocation exists whose capacity is greater than or equal to the requested
+    // size.
+    auto iter = bufferPool.lower_bound(numBytes);
+    if (iter != bufferPool.end()) {
+        VulkanGpuBuffer const* gpuBuffer = iter->second.gpuBuffer;
+        bufferPool.erase(iter);
+        return fvkmemory::resource_ptr<VulkanBuffer>::construct(&mResourceManager, gpuBuffer,
+                [this](VulkanGpuBuffer const* gpuBuffer) { this->release(gpuBuffer); });
+    }
+
+    // We were not able to find a sufficiently large allocation, so create a new one that is
+    // recycled after being yielded.
+    VulkanGpuBuffer const* gpuBuffer = allocate(usage, numBytes);
+    return fvkmemory::resource_ptr<VulkanBuffer>::construct(&mResourceManager, gpuBuffer,
+            [this](VulkanGpuBuffer const* gpuBuffer) { this->release(gpuBuffer); });
+}
+
+void VulkanBufferCache::gc() noexcept {
+    FVK_SYSTRACE_CONTEXT();
+    FVK_SYSTRACE_START("VulkanBufferCache::gc");
+
+    // If this is one of the first few frames, return early to avoid wrapping unsigned integers.
+    constexpr uint32_t TIME_BEFORE_EVICTION = 3;
+    if (++mCurrentFrame <= TIME_BEFORE_EVICTION) {
+        return;
+    }
+    const uint64_t evictionTime = mCurrentFrame - TIME_BEFORE_EVICTION;
+
+    // Destroy buffers that have not been used for several frames.
+    for (auto& bufferPool: mGpuBufferPools) {
+        for (auto poolIter = bufferPool.begin(); poolIter != bufferPool.end();) {
+            if (poolIter->second.lastAccessed < evictionTime) {
+#if FVK_ENABLED(FVK_DEBUG_VULKAN_BUFFER_CACHE)
+                FVK_LOGD << "VulkanBufferCache - Destroyed vkBuffer "
+                         << poolIter->second.gpuBuffer->vkbuffer << " with usage "
+                         << static_cast<int>(poolIter->second.gpuBuffer->usage) << utils::io::endl;
+#endif// FVK_DEBUG_VULKAN_BUFFER_CACHE
+
+                destroy(poolIter->second.gpuBuffer);
+                poolIter = bufferPool.erase(poolIter);
+            } else {
+                ++poolIter;
+            }
+        }
+    }
+
+    FVK_SYSTRACE_END();
+}
+
+void VulkanBufferCache::terminate() noexcept {
+    for (auto& bufferPool: mGpuBufferPools) {
+        for (auto& poolEntry: bufferPool) {
+            destroy(poolEntry.second.gpuBuffer);
+        }
+        bufferPool.clear();
+    }
+}
+
+void VulkanBufferCache::release(VulkanGpuBuffer const* gpuBuffer) noexcept {
+    assert_invariant(gpuBuffer != nullptr);
+
+    BufferPool& bufferPool = getPool(gpuBuffer->usage);
+    bufferPool.insert(std::make_pair(gpuBuffer->numBytes, UnusedGpuBuffer{
+                                                              .lastAccessed = mCurrentFrame,
+                                                              .gpuBuffer = gpuBuffer,
+                                                          }));
+}
+
+VulkanGpuBuffer const* VulkanBufferCache::allocate(VulkanBufferUsage usage,
+        uint32_t numBytes) noexcept {
+    VkBufferCreateInfo const bufferInfo{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .size = numBytes,
+        // `VK_BUFFER_USAGE_TRANSFER_DST_BIT` is needed to allow updating the buffer through
+        // a staging using `vkCmdCopyBuffer`.
+        .usage = getVkBufferUsage(usage) | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+    };
+
+    VmaAllocationCreateFlags vmaFlags = 0;
+    if (usage == VulkanBufferUsage::UNIFORM) {
+        // In the case of UMA, the uniform buffers will always be mappable
+        if (mContext.isUnifiedMemoryArchitecture()) {
+            vmaFlags |= VMA_ALLOCATION_CREATE_MAPPED_BIT |
+                        VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
+        }
+    }
+
+    VulkanGpuBuffer* gpuBuffer = new VulkanGpuBuffer{
+        .numBytes = numBytes,
+        .usage = usage,
+    };
+    VmaAllocationCreateInfo const allocInfo{
+        .flags = vmaFlags,
+        .usage = VMA_MEMORY_USAGE_AUTO,
+        .requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+    };
+    UTILS_UNUSED_IN_RELEASE VkResult result = vmaCreateBuffer(mAllocator, &bufferInfo, &allocInfo,
+            &gpuBuffer->vkbuffer, &gpuBuffer->vmaAllocation, &gpuBuffer->allocationInfo);
+
+#if FVK_ENABLED(FVK_DEBUG_VULKAN_BUFFER_CACHE)
+    if (result != VK_SUCCESS) {
+        FVK_LOGE << "VulkanBufferCache - failed to allocate a new vkBuffer of size " << numBytes
+                 << " and usage " << static_cast<int>(usage) << ", error: " << result
+                 << utils::io::endl;
+    } else {
+        FVK_LOGD << "VulkanBufferCache - allocated a vkBuffer " << gpuBuffer->vkbuffer
+                 << " of size " << numBytes << " and usage = " << static_cast<int>(usage)
+                 << "  successfully" << utils::io::endl;
+    }
+#endif// FVK_DEBUG_VULKAN_BUFFER_CACHE
+
+    return gpuBuffer;
+}
+
+void VulkanBufferCache::destroy(VulkanGpuBuffer const* gpuBuffer) noexcept {
+    vmaDestroyBuffer(mAllocator, gpuBuffer->vkbuffer, gpuBuffer->vmaAllocation);
+    delete gpuBuffer;
+    gpuBuffer = nullptr;
+}
+
+VulkanBufferCache::BufferPool& VulkanBufferCache::getPool(VulkanBufferUsage usage) noexcept {
+
+    int poolIndex = -1;
+    switch (usage) {
+        case VulkanBufferUsage::VERTEX:
+            poolIndex = 0;
+            break;
+        case VulkanBufferUsage::INDEX:
+            poolIndex = 1;
+            break;
+        case VulkanBufferUsage::UNIFORM:
+            poolIndex = 2;
+            break;
+        case VulkanBufferUsage::SHADER_STORAGE:
+            poolIndex = 3;
+            break;
+        case VulkanBufferUsage::UNKNOWN:
+            PANIC_LOG("There's no pool for buffers with unkown usage.");
+            break;
+    }
+
+    assert_invariant(poolIndex >= 0 && poolIndex < MAX_POOL_COUNT);
+    return mGpuBufferPools[poolIndex];
+}
+
+}// namespace filament::backend

filament/backend/src/vulkan/VulkanBufferCache.h

@@ -0,0 +1,84 @@
+/*
+ * Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_VULKANBUFFERCACHE_H
+#define TNT_FILAMENT_BACKEND_VULKANBUFFERCACHE_H
+
+#include "VulkanBuffer.h"
+#include "VulkanContext.h"
+#include "VulkanMemory.h"
+#include "memory/Resource.h"
+#include "memory/ResourceManager.h"
+
+#include <map>
+
+namespace filament::backend {
+
+class VulkanBufferCache {
+public:
+    VulkanBufferCache(VulkanContext const& context, fvkmemory::ResourceManager& resourceManager,
+            VmaAllocator allocator);
+
+    // `VulkanBufferCache` is not copyable.
+    VulkanBufferCache(const VulkanBufferCache&) = delete;
+    VulkanBufferCache& operator=(const VulkanBufferCache&) = delete;
+
+    // Allocates or reuse a new VkBuffer that is device local.
+    // In the case of Unified memory architecture, uniform buffers are also host visible.
+    fvkmemory::resource_ptr<VulkanBuffer> acquire(VulkanBufferUsage usage,
+            uint32_t numBytes) noexcept;
+
+    // Evicts old unused `VulkanGpuBuffer` and bumps the current frame number
+    void gc() noexcept;
+
+    // Destroys all unused `VulkanGpuBuffer`.
+    // This should be called while the context's VkDevice is still alive.
+    void terminate() noexcept;
+
+private:
+    struct UnusedGpuBuffer {
+        uint64_t lastAccessed;
+        VulkanGpuBuffer const* gpuBuffer;
+    };
+
+    using BufferPool = std::multimap<uint32_t, UnusedGpuBuffer>;
+
+    // Return a `VulkanGpuBuffer` back to its corresponding pool
+    void release(VulkanGpuBuffer const* gpuBuffer) noexcept;
+
+    // Allocate a new VkBuffer from the VMA pool of the corresponding `numBytes` and `usage`.
+    VulkanGpuBuffer const* allocate(VulkanBufferUsage usage, uint32_t numBytes) noexcept;
+
+    // Destroy the corresponding VkBuffer and return the VkDeviceMemory to the VMA pool.
+    void destroy(VulkanGpuBuffer const* gpuBuffer) noexcept;
+
+    BufferPool& getPool(VulkanBufferUsage usage) noexcept;
+
+    VulkanContext const& mContext;
+    fvkmemory::ResourceManager& mResourceManager;
+    VmaAllocator mAllocator;
+
+    // Buffers can be recycled, after they are released. Each type of buffer have its own pool
+    static constexpr int MAX_POOL_COUNT = 4;
+    BufferPool mGpuBufferPools[MAX_POOL_COUNT];
+
+    // Store the current "time" (really just a frame count) and LRU eviction parameters.
+    uint64_t mCurrentFrame = 0;
+};
+
+}// namespace filament::backend
+
+#endif// TNT_FILAMENT_BACKEND_VULKANBUFFERCACHE_H

filament/backend/src/vulkan/VulkanBufferProxy.cpp

@@ -14,88 +14,75 @@
  * limitations under the License.
  */
 
-#include "VulkanBuffer.h"
+#include "VulkanBufferProxy.h"
+#include "VulkanCommands.h"
 #include "VulkanMemory.h"
 
-#include <utils/Panic.h>
+#include "VulkanBufferCache.h"
+#include "VulkanMemory.h"
 
 using namespace bluevk;
 
 namespace filament::backend {
 
-VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
-        VkBufferUsageFlags usage, uint32_t numBytes)
+VulkanBufferProxy::VulkanBufferProxy(VmaAllocator allocator, VulkanStagePool& stagePool,
+        VulkanBufferCache& bufferCache, VulkanBufferUsage usage, uint32_t numBytes)
     : mAllocator(allocator),
       mStagePool(stagePool),
-      mUsage(usage),
-	  mUpdatedOffset(0),
-      mUpdatedBytes(0) {
-    // for now make sure that only 1 bit is set in usage
-    // (because loadFromCpu() assumes that somewhat)
-    assert_invariant(usage && !(usage & (usage - 1)));
+      mBufferCache(bufferCache),
+      mBuffer(mBufferCache.acquire(usage, numBytes)),
+      mUpdatedOffset(0),
+      mUpdatedBytes(0) {}
 
-    // Create the VkBuffer.
-    VkBufferCreateInfo bufferInfo {
-        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
-        .size = numBytes,
-        .usage = usage | VK_BUFFER_USAGE_TRANSFER_DST_BIT
-    };
-
-    VmaAllocationCreateInfo allocInfo { .usage = VMA_MEMORY_USAGE_GPU_ONLY };
-    vmaCreateBuffer(mAllocator, &bufferInfo, &allocInfo, &mGpuBuffer, &mGpuMemory, nullptr);
-}
-
-VulkanBuffer::~VulkanBuffer() {
-    vmaDestroyBuffer(mAllocator, mGpuBuffer, mGpuMemory);
-}
-
-void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
-        uint32_t numBytes) {
-    VulkanStage const* stage = mStagePool.acquireStage(numBytes);
-    void* mapped;
-    vmaMapMemory(mAllocator, stage->memory, &mapped);
-    memcpy(mapped, cpuData, numBytes);
-    vmaUnmapMemory(mAllocator, stage->memory);
-    vmaFlushAllocation(mAllocator, stage->memory, byteOffset, numBytes);
+void VulkanBufferProxy::loadFromCpu(VulkanCommandBuffer& commands, const void* cpuData,
+        uint32_t byteOffset, uint32_t numBytes) {
+    // Note: this should be stored within the command buffer before going out of
+    // scope, so that the command buffer can manage its lifecycle.
+    fvkmemory::resource_ptr<VulkanStage::Segment> stage = mStagePool.acquireStage(numBytes);
+    assert_invariant(stage->memory());
+    commands.acquire(stage);
+    memcpy(stage->mapping(), cpuData, numBytes);
+    vmaFlushAllocation(mAllocator, stage->memory(), stage->offset(), numBytes);
 
     // If there was a previous update, then we need to make sure the following write is properly
     // synced with the previous read.
-    if (mUpdatedBytes > 0) {
+    if (mUpdatedBytes > 0 &&
+            (byteOffset >= mUpdatedOffset && byteOffset <= (mUpdatedOffset + mUpdatedBytes))) {
         VkAccessFlags srcAccess = 0;
         VkPipelineStageFlags srcStage = 0;
-        if (mUsage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT) {
+        if (getUsage() == VulkanBufferUsage::UNIFORM) {
             srcAccess = VK_ACCESS_SHADER_READ_BIT;
             srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
-        } else if (mUsage & VK_BUFFER_USAGE_VERTEX_BUFFER_BIT) {
+        } else if (getUsage() == VulkanBufferUsage::VERTEX) {
             srcAccess = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
             srcStage = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
-        } else if (mUsage & VK_BUFFER_USAGE_INDEX_BUFFER_BIT) {
+        } else if (getUsage() == VulkanBufferUsage::INDEX) {
             srcAccess = VK_ACCESS_INDEX_READ_BIT;
             srcStage = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
         }
 
-        VkBufferMemoryBarrier barrier{
+        VkBufferMemoryBarrier barrier = {
             .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
             .srcAccessMask = srcAccess,
             .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
             .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .buffer = mGpuBuffer,
-			.offset = mUpdatedOffset,
-            .size = mUpdatedBytes,
+            .buffer = getVkBuffer(),
+            .offset = byteOffset,
+            .size = numBytes,
         };
-        vkCmdPipelineBarrier(cmdbuf, srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 1,
-                &barrier, 0, nullptr);
+        vkCmdPipelineBarrier(commands.buffer(), srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0,
+                nullptr, 1, &barrier, 0, nullptr);
     }
 
-    VkBufferCopy region {
-            .srcOffset = 0,
-            .dstOffset = byteOffset,
-            .size = numBytes,
+    VkBufferCopy region = {
+        .srcOffset = stage->offset(),
+        .dstOffset = byteOffset,
+        .size = numBytes,
     };
-    vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);
+    vkCmdCopyBuffer(commands.buffer(), stage->buffer(), getVkBuffer(), 1, &region);
 
-	mUpdatedOffset = byteOffset;
+    mUpdatedOffset = byteOffset;
     mUpdatedBytes = numBytes;
 
     // Firstly, ensure that the copy finishes before the next draw call.
@@ -105,32 +92,40 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
     VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
     VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT;
 
-    if (mUsage & VK_BUFFER_USAGE_VERTEX_BUFFER_BIT) {
+    if (getUsage() == VulkanBufferUsage::VERTEX) {
         dstAccessMask |= VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
         dstStageMask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
-    } else if (mUsage & VK_BUFFER_USAGE_INDEX_BUFFER_BIT) {
+    } else if (getUsage() == VulkanBufferUsage::INDEX) {
         dstAccessMask |= VK_ACCESS_INDEX_READ_BIT;
         dstStageMask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
-    } else if (mUsage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT) {
-        dstAccessMask |= VK_ACCESS_UNIFORM_READ_BIT;
-        dstStageMask |=
-                (VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
-    } else if (mUsage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT) {
+    } else if (getUsage() == VulkanBufferUsage::UNIFORM) {
+        dstAccessMask |= VK_ACCESS_SHADER_READ_BIT;
+        dstStageMask |= VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT;
+    } else if (getUsage() == VulkanBufferUsage::SHADER_STORAGE) {
         // TODO: implement me
     }
 
-    VkBufferMemoryBarrier barrier{
+    VkBufferMemoryBarrier barrier = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
         .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
         .dstAccessMask = dstAccessMask,
         .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
         .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .buffer = mGpuBuffer,
-        .size = VK_WHOLE_SIZE,
+        .buffer = getVkBuffer(),
+        .offset = byteOffset,
+        .size = numBytes,
     };
 
-    vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask, 0, 0, nullptr, 1,
-            &barrier, 0, nullptr);
+    vkCmdPipelineBarrier(commands.buffer(), VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask, 0, 0,
+            nullptr, 1, &barrier, 0, nullptr);
 }
 
-} // namespace filament::backend
+VkBuffer VulkanBufferProxy::getVkBuffer() const noexcept {
+    return mBuffer->getGpuBuffer()->vkbuffer;
+}
+
+VulkanBufferUsage VulkanBufferProxy::getUsage() const noexcept {
+    return mBuffer->getGpuBuffer()->usage;
+}
+
+}// namespace filament::backend

filament/backend/src/vulkan/VulkanBufferProxy.h

@@ -0,0 +1,54 @@
+/*
+ * Copyright (C) 2018 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_VULKANBUFFERPROXY_H
+#define TNT_FILAMENT_BACKEND_VULKANBUFFERPROXY_H
+
+#include "VulkanBufferCache.h"
+#include "VulkanCommands.h"
+#include "VulkanContext.h"
+#include "VulkanMemory.h"
+#include "VulkanStagePool.h"
+
+namespace filament::backend {
+
+// This class acts as a dynamic wrapper for a `VulkanBuffer`. It allows you to modify the
+// `VulkanBuffer` it references at runtime, wihtout affecting any external objects.
+class VulkanBufferProxy {
+public:
+    VulkanBufferProxy(VmaAllocator allocator, VulkanStagePool& stagePool,
+            VulkanBufferCache& bufferCache, VulkanBufferUsage usage, uint32_t numBytes);
+
+    void loadFromCpu(VulkanCommandBuffer& commands, const void* cpuData, uint32_t byteOffset,
+            uint32_t numBytes);
+
+    VkBuffer getVkBuffer() const noexcept;
+
+    VulkanBufferUsage getUsage() const noexcept;
+
+private:
+    VmaAllocator mAllocator;
+    VulkanStagePool& mStagePool;
+    VulkanBufferCache& mBufferCache;
+
+    fvkmemory::resource_ptr<VulkanBuffer> mBuffer;
+    uint32_t mUpdatedOffset = 0;
+    uint32_t mUpdatedBytes = 0;
+};
+
+}// namespace filament::backend
+
+#endif// TNT_FILAMENT_BACKEND_VULKANBUFFERPROXY_H

filament/backend/src/vulkan/VulkanConstants.h

@@ -81,6 +81,8 @@
 // All other debug features must be disabled.
 #define FVK_DEBUG_PROFILING               0x00040000
 
+#define FVK_DEBUG_VULKAN_BUFFER_CACHE     0x00080000
+
 // Useful default combinations
 #define FVK_DEBUG_EVERYTHING              (0xFFFFFFFF & ~FVK_DEBUG_PROFILING)
 #define FVK_DEBUG_PERFORMANCE     \
@@ -151,13 +153,13 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 
 #elif FVK_ENABLED(FVK_DEBUG_SYSTRACE)
 
-    #include <utils/Systrace.h>
-    
-    #define FVK_SYSTRACE_CONTEXT()      SYSTRACE_CONTEXT()
-    #define FVK_SYSTRACE_START(marker)  SYSTRACE_NAME_BEGIN(marker)
-    #define FVK_SYSTRACE_END()          SYSTRACE_NAME_END()
-    #define FVK_SYSTRACE_SCOPE()        SYSTRACE_NAME(__func__)
-    #define FVK_PROFILE_MARKER(marker)  FVK_SYSTRACE_SCOPE()
+    #include <private/utils/Tracing.h>
+
+    #define FVK_SYSTRACE_CONTEXT()      FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT)
+    #define FVK_SYSTRACE_START(marker)  FILAMENT_TRACING_NAME_BEGIN(FILAMENT_TRACING_CATEGORY_FILAMENT, marker)
+    #define FVK_SYSTRACE_END()          FILAMENT_TRACING_NAME_END(FILAMENT_TRACING_CATEGORY_FILAMENT)
+    #define FVK_SYSTRACE_SCOPE()        FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT)
+    #define FVK_PROFILE_MARKER(marker)  FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT)
 
 #else
     #define FVK_SYSTRACE_CONTEXT()

filament/backend/src/vulkan/VulkanContext.h

@@ -71,23 +71,23 @@ struct VulkanRenderPass {
 struct VulkanContext {
 public:
     static uint32_t selectMemoryType(VkPhysicalDeviceMemoryProperties const& memoryProperties,
-            uint32_t flags, VkFlags reqs) {
+            uint32_t types, VkFlags reqs) {
         for (uint32_t i = 0; i < VK_MAX_MEMORY_TYPES; i++) {
-            if (flags & 1) {
+            if (types & 1) {
                 if ((memoryProperties.memoryTypes[i].propertyFlags & reqs) == reqs) {
                     return i;
                 }
             }
-            flags >>= 1;
+            types >>= 1;
         }
         return (uint32_t) VK_MAX_MEMORY_TYPES;
     }
 
-    inline uint32_t selectMemoryType(uint32_t flags, VkFlags reqs) const {
+    inline uint32_t selectMemoryType(uint32_t types, VkFlags reqs) const {
         if ((reqs & VK_MEMORY_PROPERTY_PROTECTED_BIT) != 0) {
             assert_invariant(isProtectedMemorySupported());
         }
-        return selectMemoryType(mMemoryProperties, flags, reqs);
+        return selectMemoryType(mMemoryProperties, types, reqs);
     }
 
     inline fvkutils::VkFormatList const& getAttachmentDepthStencilFormats() const {

filament/backend/src/vulkan/VulkanDescriptorSetCache.cpp

@@ -320,7 +320,7 @@ void VulkanDescriptorSetCache::updateBuffer(fvkmemory::resource_ptr<VulkanDescri
         uint8_t binding, fvkmemory::resource_ptr<VulkanBufferObject> bufferObject,
         VkDeviceSize offset, VkDeviceSize size) noexcept {
     VkDescriptorBufferInfo const info = {
-        .buffer = bufferObject->buffer.getGpuBuffer(),
+        .buffer = bufferObject->buffer.getVkBuffer(),
         .offset = offset,
         .range = size,
     };
@@ -358,12 +358,11 @@ void VulkanDescriptorSetCache::updateSamplerImpl(VkDescriptorSet vkset, uint8_t
         range.levelCount = 1;
         range.layerCount = 1;
     }
-    VkDescriptorImageInfo info{
+    VkDescriptorImageInfo info = {
         .sampler = sampler,
         .imageView = texture->getView(range),
-        .imageLayout = fvkutils::getVkLayout(texture->getDefaultLayout()),
+        .imageLayout = fvkutils::getVkLayout(texture->getSamplerLayout()),
     };
-
     VkWriteDescriptorSet descriptorWrite = {
         .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
         .pNext = nullptr,

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -19,7 +19,8 @@
 #include "CommandStreamDispatcher.h"
 #include "SystraceProfile.h"
 #include "VulkanAsyncHandles.h"
-#include "VulkanBuffer.h"
+#include "VulkanBufferCache.h"
+#include "VulkanBufferProxy.h"
 #include "VulkanCommands.h"
 #include "VulkanDriverFactory.h"
 #include "VulkanHandles.h"
@@ -84,7 +85,11 @@ VmaAllocator createAllocator(VkInstance instance, VkPhysicalDevice physicalDevic
         .vkGetImageMemoryRequirements2KHR = vkGetImageMemoryRequirements2KHR
 #endif
     };
-    VmaAllocatorCreateInfo const allocatorInfo {
+    VmaAllocatorCreateInfo const allocatorInfo{
+        // Disable the internal VMA synchronization because the backend is singled threaded.
+        // Improve CPU performance when using VMA functions. The backend will guarantee that all
+        // access to VMA is done in a thread safe way.
+        .flags = VMA_ALLOCATOR_CREATE_EXTERNALLY_SYNCHRONIZED_BIT,
         .physicalDevice = physicalDevice,
         .device = device,
         .pVulkanFunctions = &funcs,
@@ -209,7 +214,8 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
               mPlatform->getProtectedGraphicsQueueFamilyIndex(), &mContext),
       mPipelineLayoutCache(mPlatform->getDevice()),
       mPipelineCache(mPlatform->getDevice()),
-      mStagePool(mAllocator, &mCommands),
+      mStagePool(mAllocator, &mResourceManager, &mCommands, &mContext.getPhysicalDeviceLimits()),
+      mBufferCache(context, mResourceManager, mAllocator),
       mFramebufferCache(mPlatform->getDevice()),
       mYcbcrConversionCache(mPlatform->getDevice()),
       mSamplerCache(mPlatform->getDevice()),
@@ -328,7 +334,6 @@ void VulkanDriver::terminate() {
     // descriptorSetLayoutCache
     mExternalImageManager.terminate();
 
-    mStagePool.terminate();
     mPipelineCache.terminate();
     mFramebufferCache.terminate();
     mSamplerCache.terminate();
@@ -339,6 +344,15 @@ void VulkanDriver::terminate() {
     // Before terminating ResourceManager, we must make sure all of the resource_ptrs have been unset.
     mResourceManager.terminate();
 
+    // Must come after `mResourceManager`.
+    // Before terminating the memory pool, we must make sure all the VulkanBufferMemory are yielded
+    // back to the pool.
+    mBufferCache.terminate();
+
+    // Before terminating stagePool, we need all resources to have been
+    // reclaimed, as they perform cleanup within the stage pool.
+    mStagePool.terminate();
+
 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
     mResourceManager.print();
 #endif
@@ -371,6 +385,7 @@ void VulkanDriver::collectGarbage() {
     mCommands.gc();
     mDescriptorSetCache.gc();
     mStagePool.gc();
+    mBufferCache.gc();
     mFramebufferCache.gc();
     mPipelineCache.gc();
 
@@ -514,7 +529,7 @@ void VulkanDriver::createIndexBufferR(Handle<HwIndexBuffer> ibh, ElementType ele
     FVK_SYSTRACE_SCOPE();
     auto elementSize = (uint8_t) getElementTypeSize(elementType);
     auto ib = resource_ptr<VulkanIndexBuffer>::make(&mResourceManager, ibh, mAllocator, mStagePool,
-            elementSize, indexCount);
+            mBufferCache, elementSize, indexCount);
     ib.inc();
 }
 
@@ -531,7 +546,7 @@ void VulkanDriver::createBufferObjectR(Handle<HwBufferObject> boh, uint32_t byte
         BufferObjectBinding bindingType, BufferUsage usage) {
     FVK_SYSTRACE_SCOPE();
     auto bo = resource_ptr<VulkanBufferObject>::make(&mResourceManager, boh, mAllocator, mStagePool,
-            byteCount, bindingType);
+            mBufferCache, byteCount, bindingType);
     bo.inc();
 }
 
@@ -622,7 +637,7 @@ void VulkanDriver::createTextureExternalImage2R(Handle<HwTexture> th, backend::S
     auto& commands = mCommands.get();
     // Unlike uploaded textures or swapchains, we need to explicit transition this
     // texture into the read layout.
-    texture->transitionLayout(&commands, texture->getPrimaryViewRange(), VulkanLayout::READ_ONLY);
+    texture->transitionLayout(&commands, texture->getPrimaryViewRange(), VulkanLayout::FRAG_READ);
 
     if (imgData.external.valid()) {
         mExternalImageManager.addExternallySampledTexture(texture, externalImage);
@@ -1223,7 +1238,7 @@ void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor
     VulkanCommandBuffer& commands = mCommands.get();
     auto ib = resource_ptr<VulkanIndexBuffer>::cast(&mResourceManager, ibh);
     commands.acquire(ib);
-    ib->buffer.loadFromCpu(commands.buffer(), p.buffer, byteOffset, p.size);
+    ib->buffer.loadFromCpu(commands, p.buffer, byteOffset, p.size);
 
     scheduleDestroy(std::move(p));
 }
@@ -1238,7 +1253,7 @@ void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescript
 
     auto bo = resource_ptr<VulkanBufferObject>::cast(&mResourceManager, boh);
     commands.acquire(bo);
-    bo->buffer.loadFromCpu(commands.buffer(), bd.buffer, byteOffset, bd.size);
+    bo->buffer.loadFromCpu(commands, bd.buffer, byteOffset, bd.size);
 
     scheduleDestroy(std::move(bd));
 }
@@ -1249,7 +1264,7 @@ void VulkanDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> boh,
     auto bo = resource_ptr<VulkanBufferObject>::cast(&mResourceManager, boh);
     commands.acquire(bo);
     // TODO: implement unsynchronized version
-    bo->buffer.loadFromCpu(commands.buffer(), bd.buffer, byteOffset, bd.size);
+    bo->buffer.loadFromCpu(commands, bd.buffer, byteOffset, bd.size);
     scheduleDestroy(std::move(bd));
 }
 
@@ -1622,8 +1637,8 @@ void VulkanDriver::readPixels(Handle<HwRenderTarget> src, uint32_t x, uint32_t y
     mReadPixels.run(
             srcTarget, x, y, width, height, mPlatform->getGraphicsQueueFamilyIndex(),
             std::move(pbd),
-            [&context = mContext](uint32_t reqs, VkFlags flags) {
-                return context.selectMemoryType(reqs, flags);
+            [&context = mContext](uint32_t types, VkFlags reqs) {
+                return context.selectMemoryType(types, reqs);
             },
             [this](PixelBufferDescriptor&& pbd) {
                 scheduleDestroy(std::move(pbd));
@@ -1898,7 +1913,7 @@ void VulkanDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
     // avoid rebinding these if they are already bound, but since we do not (yet) support subranges
     // it would be rare for a client to make consecutive draw calls with the same render primitive.
     vkCmdBindVertexBuffers(cmdbuffer, 0, bufferCount, buffers, offsets);
-    vkCmdBindIndexBuffer(cmdbuffer, prim->indexBuffer->buffer.getGpuBuffer(), 0,
+    vkCmdBindIndexBuffer(cmdbuffer, prim->indexBuffer->buffer.getVkBuffer(), 0,
             prim->indexBuffer->indexType);
 }
 

filament/backend/src/vulkan/VulkanDriver.h

@@ -18,15 +18,17 @@
 #define TNT_FILAMENT_BACKEND_VULKANDRIVER_H
 
 #include "VulkanBlitter.h"
+#include "VulkanBufferCache.h"
 #include "VulkanConstants.h"
 #include "VulkanContext.h"
 #include "VulkanFboCache.h"
 #include "VulkanHandles.h"
+#include "VulkanMemory.h"
 #include "VulkanPipelineCache.h"
+#include "VulkanQueryManager.h"
 #include "VulkanReadPixels.h"
 #include "VulkanSamplerCache.h"
 #include "VulkanStagePool.h"
-#include "VulkanQueryManager.h"
 #include "VulkanYcbcrConversionCache.h"
 #include "vulkan/VulkanDescriptorSetCache.h"
 #include "vulkan/VulkanDescriptorSetLayoutCache.h"
@@ -138,6 +140,7 @@ private:
     VulkanPipelineLayoutCache mPipelineLayoutCache;
     VulkanPipelineCache mPipelineCache;
     VulkanStagePool mStagePool;
+    VulkanBufferCache mBufferCache;
     VulkanFboCache mFramebufferCache;
     VulkanYcbcrConversionCache mYcbcrConversionCache;
     VulkanSamplerCache mSamplerCache;

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -93,7 +93,32 @@ BitmaskGroup fromBackendLayout(DescriptorSetLayout const& layout) {
             case DescriptorType::SAMPLER_EXTERNAL:
                 fromStageFlags(binding.stageFlags, binding.binding, mask.externalSampler);
                 UTILS_FALLTHROUGH;
-            case DescriptorType::SAMPLER: {
+
+            case DescriptorType::SAMPLER_2D_FLOAT:
+            case DescriptorType::SAMPLER_2D_INT:
+            case DescriptorType::SAMPLER_2D_UINT:
+            case DescriptorType::SAMPLER_2D_DEPTH:
+            case DescriptorType::SAMPLER_2D_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_ARRAY_UINT:
+            case DescriptorType::SAMPLER_2D_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_INT:
+            case DescriptorType::SAMPLER_CUBE_UINT:
+            case DescriptorType::SAMPLER_CUBE_DEPTH:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_INT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_UINT:
+            case DescriptorType::SAMPLER_CUBE_ARRAY_DEPTH:
+            case DescriptorType::SAMPLER_3D_FLOAT:
+            case DescriptorType::SAMPLER_3D_INT:
+            case DescriptorType::SAMPLER_3D_UINT:
+            case DescriptorType::SAMPLER_2D_MS_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_INT:
+            case DescriptorType::SAMPLER_2D_MS_UINT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_FLOAT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_INT:
+            case DescriptorType::SAMPLER_2D_MS_ARRAY_UINT: {
                 fromStageFlags(binding.stageFlags, binding.binding, mask.sampler);
                 break;
             }
@@ -495,7 +520,8 @@ void VulkanRenderTarget::emitBarriersEndRenderPass(VulkanCommandBuffer& commands
             }
         } else {
             texture->setLayout(range, VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT);
-            if (!texture->transitionLayout(&commands, range, VulkanLayout::READ_WRITE)) {
+            if (texture->isSampleable() &&
+                    !texture->transitionLayout(&commands, range, VulkanLayout::FRAG_READ)) {
                 texture->attachmentToSamplerBarrier(&commands, range);
             }
         }
@@ -557,22 +583,22 @@ void VulkanVertexBuffer::setBuffer(fvkmemory::resource_ptr<VulkanBufferObject> b
     int8_t const* const attribToBuffer = vbi->getAttributeToBuffer();
     for (uint8_t attribIndex = 0; attribIndex < count; attribIndex++) {
         if (attribToBuffer[attribIndex] == static_cast<int8_t>(index)) {
-            vkbuffers[attribIndex] = bufferObject->buffer.getGpuBuffer();
+            vkbuffers[attribIndex] = bufferObject->buffer.getVkBuffer();
         }
     }
     mResources.push_back(bufferObject);
 }
 
 VulkanBufferObject::VulkanBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool,
-        uint32_t byteCount, BufferObjectBinding bindingType)
+        VulkanBufferCache& bufferCache, uint32_t byteCount, BufferObjectBinding bindingType)
     : HwBufferObject(byteCount),
-      buffer(allocator, stagePool, getBufferObjectUsage(bindingType), byteCount),
+      buffer(allocator, stagePool, bufferCache, getBufferObjectUsage(bindingType), byteCount),
       bindingType(bindingType) {}
 
 VulkanRenderPrimitive::VulkanRenderPrimitive(PrimitiveType pt,
         fvkmemory::resource_ptr<VulkanVertexBuffer> vb,
         fvkmemory::resource_ptr<VulkanIndexBuffer> ib)
-    : HwRenderPrimitive{.type = pt},
+    : HwRenderPrimitive{ .type = pt },
       vertexBuffer(vb),
       indexBuffer(ib) {}
 

filament/backend/src/vulkan/VulkanHandles.h

@@ -21,20 +21,21 @@
 #include "DriverBase.h"
 
 #include "VulkanAsyncHandles.h"
-#include "VulkanBuffer.h"
+#include "VulkanBufferCache.h"
+#include "VulkanBufferProxy.h"
 #include "VulkanFboCache.h"
 #include "VulkanSwapChain.h"
 #include "VulkanTexture.h"
 #include "vulkan/memory/Resource.h"
-#include "vulkan/utils/StaticVector.h"
 #include "vulkan/utils/Definitions.h"
+#include "vulkan/utils/StaticVector.h"
 
 #include <backend/Program.h>
 
-#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Mutex.h>
 #include <utils/StructureOfArrays.h>
+#include <utils/bitset.h>
 
 #include <array>
 
@@ -428,21 +429,22 @@ private:
 };
 
 struct VulkanIndexBuffer : public HwIndexBuffer, fvkmemory::Resource {
-    VulkanIndexBuffer(VmaAllocator allocator, VulkanStagePool& stagePool, uint8_t elementSize,
-            uint32_t indexCount)
+    VulkanIndexBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
+            VulkanBufferCache& bufferCache, uint8_t elementSize, uint32_t indexCount)
         : HwIndexBuffer(elementSize, indexCount),
-          buffer(allocator, stagePool, VK_BUFFER_USAGE_INDEX_BUFFER_BIT, elementSize * indexCount),
+          buffer(allocator, stagePool, bufferCache, VulkanBufferUsage::INDEX,
+                  elementSize * indexCount),
           indexType(elementSize == 2 ? VK_INDEX_TYPE_UINT16 : VK_INDEX_TYPE_UINT32) {}
 
-    VulkanBuffer buffer;
+    VulkanBufferProxy buffer;
     const VkIndexType indexType;
 };
 
 struct VulkanBufferObject : public HwBufferObject, fvkmemory::Resource {
-    VulkanBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool, uint32_t byteCount,
-            BufferObjectBinding bindingType);
+    VulkanBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool,
+            VulkanBufferCache& bufferCache, uint32_t byteCount, BufferObjectBinding bindingType);
 
-    VulkanBuffer buffer;
+    VulkanBufferProxy buffer;
     const BufferObjectBinding bindingType;
 };
 
@@ -455,18 +457,19 @@ struct VulkanRenderPrimitive : public HwRenderPrimitive, fvkmemory::Resource {
     fvkmemory::resource_ptr<VulkanIndexBuffer> indexBuffer;
 };
 
-inline constexpr VkBufferUsageFlagBits getBufferObjectUsage(
-        BufferObjectBinding bindingType) noexcept {
-    switch(bindingType) {
+inline constexpr VulkanBufferUsage getBufferObjectUsage(BufferObjectBinding bindingType) noexcept {
+    switch (bindingType) {
         case BufferObjectBinding::VERTEX:
-            return VK_BUFFER_USAGE_VERTEX_BUFFER_BIT;
+            return VulkanBufferUsage::VERTEX;
         case BufferObjectBinding::UNIFORM:
-            return VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
+            return VulkanBufferUsage::UNIFORM;
         case BufferObjectBinding::SHADER_STORAGE:
-            return VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
-        // when adding more buffer-types here, make sure to update VulkanBuffer::loadFromCpu()
-        // if necessary.
+            return VulkanBufferUsage::SHADER_STORAGE;
+            // when adding more buffer-types here, make sure to update VulkanBuffer::loadFromCpu()
+            // if necessary.
     }
+
+    return VulkanBufferUsage::UNKNOWN;
 }
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanMemory.h

@@ -37,4 +37,24 @@ VK_DEFINE_HANDLE(VmaAllocator)
 VK_DEFINE_HANDLE(VmaAllocation)
 VK_DEFINE_HANDLE(VmaPool)
 
+namespace filament::backend {
+
+enum class VulkanBufferUsage : uint8_t {
+    UNKNOWN,
+    VERTEX,
+    INDEX,
+    UNIFORM,
+    SHADER_STORAGE,
+};
+
+struct VulkanGpuBuffer {
+    VkBuffer vkbuffer = VK_NULL_HANDLE;
+    VmaAllocation vmaAllocation = VK_NULL_HANDLE;
+    VmaAllocationInfo allocationInfo;
+    uint32_t numBytes = 0;
+    VulkanBufferUsage usage = VulkanBufferUsage::UNKNOWN;
+};
+
+} // namespace filament::backend
+
 #endif // TNT_FILAMENT_BACKEND_VULKANMEMORY_H

filament/backend/src/vulkan/VulkanReadPixels.cpp

@@ -143,7 +143,7 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
         mTaskHandler = std::make_unique<TaskHandler>();
     }
 
-    VkCommandPool& cmdpool = mCommandPool;
+    VkCommandPool const cmdpool = mCommandPool;
 
     fvkmemory::resource_ptr<VulkanTexture> srcTexture = srcTarget->getColor0().texture;
     assert_invariant(srcTexture);
@@ -152,17 +152,17 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
             = srcFormat == VK_FORMAT_B8G8R8A8_UNORM || srcFormat == VK_FORMAT_B8G8R8A8_SRGB;
 
     // Create a host visible, linearly tiled image as a staging area.
-    VkImageCreateInfo const imageInfo{
-            .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
-            .imageType = VK_IMAGE_TYPE_2D,
-            .format = srcFormat,
-            .extent = {width, height, 1},
-            .mipLevels = 1,
-            .arrayLayers = 1,
-            .samples = VK_SAMPLE_COUNT_1_BIT,
-            .tiling = VK_IMAGE_TILING_LINEAR,
-            .usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT,
-            .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
+    VkImageCreateInfo const imageInfo = {
+        .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+        .imageType = VK_IMAGE_TYPE_2D,
+        .format = srcFormat,
+        .extent = { width, height, 1 },
+        .mipLevels = 1,
+        .arrayLayers = 1,
+        .samples = VK_SAMPLE_COUNT_1_BIT,
+        .tiling = VK_IMAGE_TILING_LINEAR,
+        .usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT,
+        .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
     };
 
     VkImage stagingImage;
@@ -196,20 +196,20 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
             << "VulkanReadPixels: unable to find a memory type that meets requirements.";
 
     VkMemoryAllocateInfo const allocInfo = {
-            .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
-            .allocationSize = memReqs.size,
-            .memoryTypeIndex = memoryTypeIndex,
+        .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+        .allocationSize = memReqs.size,
+        .memoryTypeIndex = memoryTypeIndex,
     };
 
     vkAllocateMemory(device, &allocInfo, VKALLOC, &stagingMemory);
     vkBindImageMemory(device, stagingImage, stagingMemory, 0);
 
     VkCommandBuffer cmdbuffer;
-    VkCommandBufferAllocateInfo const allocateInfo{
-            .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
-            .commandPool = cmdpool,
-            .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
-            .commandBufferCount = 1,
+    VkCommandBufferAllocateInfo const allocateInfo = {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
+        .commandPool = cmdpool,
+        .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
+        .commandBufferCount = 1,
     };
     vkAllocateCommandBuffers(device, &allocateInfo, &cmdbuffer);
 
@@ -234,6 +234,7 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
 
     VulkanAttachment const srcAttachment = srcTarget->getColor0();
     VkImageSubresourceRange const srcRange = srcAttachment.getSubresourceRange();
+    VulkanLayout const srcLayout = srcAttachment.getLayout();
     srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
 
     VkImageCopy const imageCopyRegion = {
@@ -270,7 +271,7 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
             fvkutils::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);
 
     // Restore the source image layout.
-    srcTexture->transitionLayout(cmdbuffer, srcRange, srcTexture->getDefaultLayout());
+    srcTexture->transitionLayout(cmdbuffer, srcRange, srcLayout);
 
     vkEndCommandBuffer(cmdbuffer);
 
@@ -303,7 +304,6 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
                                  cmdpool, cmdbuffer, pUserBuffer,
                                  fence = readCompleteFence]() mutable {
         VkResult status = vkWaitForFences(device, 1, &fence, VK_TRUE, UINT64_MAX);
-        // Fence hasn't been reached. Try waiting again.
         if (status != VK_SUCCESS) {
             FVK_LOGE << "Failed to wait for readPixels fence" << utils::io::endl;
             return;

filament/backend/src/vulkan/VulkanStagePool.cpp

@@ -28,46 +28,111 @@ static constexpr uint32_t TIME_BEFORE_EVICTION = 3;
 
 namespace filament::backend {
 
-VulkanStagePool::VulkanStagePool(VmaAllocator allocator, VulkanCommands* commands)
-    : mAllocator(allocator),
-      mCommands(commands) {}
+namespace {
 
-VulkanStage const* VulkanStagePool::acquireStage(uint32_t numBytes) {
-    // First check if a stage exists whose capacity is greater than or equal to the requested size.
-    auto iter = mFreeStages.lower_bound(numBytes);
-    if (iter != mFreeStages.end()) {
-        auto stage = iter->second;
-        mFreeStages.erase(iter);
-        stage->lastAccessed = mCurrentFrame;
-        mUsedStages.push_back(stage);
-        return stage;
-    }
-    // We were not able to find a sufficiently large stage, so create a new one.
-    VulkanStage* stage = new VulkanStage({
-        .memory = VK_NULL_HANDLE,
-        .buffer = VK_NULL_HANDLE,
-        .capacity = numBytes,
-        .lastAccessed = mCurrentFrame,
+// Note: these are temporary values, they will be configurable.
+static constexpr uint32_t MAX_EMPTY_STAGES_TO_RETAIN = 1;
+constexpr uint32_t STAGE_SIZE = 1048576;
+
+}// namespace
+
+fvkmemory::resource_ptr<VulkanStage::Segment> VulkanStage::acquireSegment(
+        fvkmemory::ResourceManager* resManager, uint32_t numBytes) {
+    auto segment = fvkmemory::resource_ptr<Segment>::construct(
+        resManager, this, numBytes, mCurrentOffset, [this](uint32_t offset) {
+            mSegments.erase(offset);
     });
+    mSegments.insert({mCurrentOffset, segment.get()});
+    mCurrentOffset += numBytes;
+    return segment;
+}
 
-    // Create the VkBuffer.
-    mUsedStages.push_back(stage);
-    VkBufferCreateInfo bufferInfo {
+VulkanStagePool::VulkanStagePool(VmaAllocator allocator, fvkmemory::ResourceManager* resManager,
+        VulkanCommands* commands, const VkPhysicalDeviceLimits* deviceLimits)
+    : mAllocator(allocator),
+      mResManager(resManager),
+      mCommands(commands),
+      mDeviceLimits(deviceLimits) {}
+
+fvkmemory::resource_ptr<VulkanStage::Segment> VulkanStagePool::acquireStage(uint32_t numBytes) {
+    // Apply alignment to the byte count to ensure that, when we later flush
+    // data written by the host, we only flush the atoms that we modified, and
+    // no adjacent atoms.
+    numBytes = alignToNonCoherentAtomSize(numBytes);
+
+    // First check if a stage segment exists whose capacity is greater than or
+    // equal to the requested size.
+    auto iter = mStages.lower_bound(numBytes);
+
+    VulkanStage* pStage;
+    if (iter != mStages.end()) {
+        pStage = iter->second;
+        mStages.erase(iter);
+    } else {
+        pStage = allocateNewStage(std::max(numBytes, STAGE_SIZE));
+    }
+
+    // Note: this allocation updates `currentOffset` and `segments` within
+    // the parent stage. When destroyed, it will update `segments`.
+    fvkmemory::resource_ptr<VulkanStage::Segment> pSegment = pStage->acquireSegment(mResManager, numBytes);
+
+    // Update the stage's metadata, and reinsert it with the remaining segment
+    // capacity.
+    uint32_t spaceRemaining = pStage->capacity() - pStage->currentOffset();
+    mStages.insert({ spaceRemaining, pStage });
+
+    return pSegment;
+}
+
+uint32_t VulkanStagePool::alignToNonCoherentAtomSize(uint32_t bytes) {
+    VkDeviceSize alignment = mDeviceLimits->nonCoherentAtomSize;
+    if (alignment == 0) {
+        return bytes;
+    }
+
+    uint32_t remainder = bytes % alignment;
+    return remainder == 0 ? bytes : bytes + (alignment - remainder);
+}
+
+VulkanStage* VulkanStagePool::allocateNewStage(uint32_t capacity) {
+    VkBufferCreateInfo bufferInfo{
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
-        .size = numBytes,
+        .size = alignToNonCoherentAtomSize(capacity),
         .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
     };
     VmaAllocationCreateInfo allocInfo { .usage = VMA_MEMORY_USAGE_CPU_ONLY };
-    UTILS_UNUSED_IN_RELEASE VkResult result = vmaCreateBuffer(mAllocator, &bufferInfo,
-            &allocInfo, &stage->buffer, &stage->memory, nullptr);
+    VkBuffer buffer;
+    VmaAllocation memory;
+    VkResult result =
+            vmaCreateBuffer(mAllocator, &bufferInfo, &allocInfo, &buffer, &memory, nullptr);
 
 #if FVK_ENABLED(FVK_DEBUG_STAGING_ALLOCATION)
     if (result != VK_SUCCESS) {
         FVK_LOGE << "Allocation error: " << result << utils::io::endl;
+    } else {
+        FVK_LOGD << "Allocated stage with hndl " << buffer << utils::io::endl;
     }
 #endif
 
-    return stage;
+    void* pMapping = nullptr;
+    if (result == VK_SUCCESS) {
+        result = vmaMapMemory(mAllocator, memory, &pMapping);
+
+#if FVK_ENABLED(FVK_DEBUG_STAGING_ALLOCATION)
+        if (result != VK_SUCCESS) {
+            FVK_LOGE << "Memory mapping erryr: " << result << utils::io::endl;
+        }
+#endif
+    }
+
+    return new VulkanStage(memory, buffer, capacity, pMapping);
+}
+
+void VulkanStagePool::destroyStage(VulkanStage const*&& stage) {
+    assert(stage->isSafeToReset());  // Ensure all segments have been reset already.
+    vmaUnmapMemory(mAllocator, stage->memory());
+    vmaDestroyBuffer(mAllocator, stage->buffer(), stage->memory());
+    delete stage;
 }
 
 VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, PixelDataType type,
@@ -125,7 +190,7 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
     fvkutils::transitionLayout(cmdbuffer, {
             .image = image->image,
             .oldLayout = VulkanLayout::UNDEFINED,
-            .newLayout = VulkanLayout::READ_WRITE, // (= VK_IMAGE_LAYOUT_GENERAL)
+            .newLayout = VulkanLayout::STAGING, // (= VK_IMAGE_LAYOUT_GENERAL)
             .subresources = { aspectFlags, 0, 1, 0, 1 },
         });
     return image;
@@ -141,27 +206,34 @@ void VulkanStagePool::gc() noexcept {
     }
     const uint64_t evictionTime = mCurrentFrame - TIME_BEFORE_EVICTION;
 
-    // Destroy buffers that have not been used for several frames.
-    decltype(mFreeStages) freeStages;
-    freeStages.swap(mFreeStages);
-    for (auto pair : freeStages) {
-        if (pair.second->lastAccessed < evictionTime) {
-            vmaDestroyBuffer(mAllocator, pair.second->buffer, pair.second->memory);
-            delete pair.second;
-        } else {
-            mFreeStages.insert(pair);
-        }
-    }
+    decltype(mStages) freeStages;
+    freeStages.swap(mStages);
+    uint8_t freeStageCount = 0;  // Assuming we'll never have > 255 free stages
+    for (auto& pair : freeStages) {
+        // First, find any stages that have no segments within them.
+        if (pair.second->isSafeToReset()) {
+            if (++freeStageCount > MAX_EMPTY_STAGES_TO_RETAIN) {
+#if FVK_ENABLED(FVK_DEBUG_STAGING_ALLOCATION)
+                FVK_LOGD << "Destroying a staging buffer with hndl " << pair.second->buffer()
+                         << utils::io::endl;
+#endif
+                destroyStage(std::move(pair.second));
+                continue;
+            }
 
-    // Reclaim buffers that are no longer being used by any command buffer.
-    decltype(mUsedStages) usedStages;
-    usedStages.swap(mUsedStages);
-    for (auto stage : usedStages) {
-        if (stage->lastAccessed < evictionTime) {
-            stage->lastAccessed = mCurrentFrame;
-            mFreeStages.insert(std::make_pair(stage->capacity, stage));
+#if FVK_ENABLED(FVK_DEBUG_STAGING_ALLOCATION)
+            if (pair.first == 0) {
+                FVK_LOGD << "Recycling an unused staging buffer with hndl " << pair.second->buffer()
+                         << utils::io::endl;
+            }
+#endif
+
+            // Note - this segment is free, make sure the structure is cleared
+            // and reinsert it into our free stage list.
+            pair.second->reset();
+            mStages.insert({ pair.second->capacity(), pair.second });
         } else {
-            mUsedStages.push_back(stage);
+            mStages.insert(pair);
         }
     }
 
@@ -192,17 +264,10 @@ void VulkanStagePool::gc() noexcept {
 }
 
 void VulkanStagePool::terminate() noexcept {
-    for (auto stage : mUsedStages) {
-        vmaDestroyBuffer(mAllocator, stage->buffer, stage->memory);
-        delete stage;
+    for (auto& pair : mStages) {
+        destroyStage(std::move(pair.second));
     }
-    mUsedStages.clear();
-
-    for (auto pair : mFreeStages) {
-        vmaDestroyBuffer(mAllocator, pair.second->buffer, pair.second->memory);
-        delete pair.second;
-    }
-    mFreeStages.clear();
+    mStages.clear();
 
     for (auto image : mUsedImages) {
         vmaDestroyImage(mAllocator, image->image, image->memory);

filament/backend/src/vulkan/VulkanStagePool.h

@@ -17,8 +17,11 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANSTAGEPOOL_H
 #define TNT_FILAMENT_BACKEND_VULKANSTAGEPOOL_H
 
-#include "backend/DriverEnums.h"
 #include "VulkanMemory.h"
+#include "backend/DriverEnums.h"
+#include "vulkan/memory/Resource.h"
+#include "vulkan/memory/ResourceManager.h"
+#include "vulkan/memory/ResourcePointer.h"
 
 #include <map>
 #include <unordered_set>
@@ -28,12 +31,96 @@ namespace filament::backend {
 
 class VulkanCommands;
 
-// Immutable POD representing a shared CPU-GPU staging area.
-struct VulkanStage {
-    VmaAllocation memory;
-    VkBuffer buffer;
-    uint32_t capacity;
-    mutable uint64_t lastAccessed;
+// Object representing a shared CPU-GPU staging area, which can be subdivided
+// into smaller buffers as needed.
+class VulkanStage {
+public:
+    VulkanStage(VmaAllocation memory, VkBuffer buffer, uint32_t capacity, void* mapping)
+        : mMemory(memory),
+          mBuffer(buffer),
+          mCapacity(capacity),
+          mMapping(mapping) {}
+
+    ~VulkanStage() = default;
+    VulkanStage(const VulkanStage& other) = delete;
+    VulkanStage(VulkanStage&& other) = delete;
+    VulkanStage& operator=(const VulkanStage& other) = delete;
+    VulkanStage& operator=(VulkanStage&& other) = delete;
+
+    class Segment : public fvkmemory::Resource {
+    public:
+        using OnRecycle = std::function<void(uint32_t offset)>;
+
+        Segment(VulkanStage* parentStage, uint32_t capacity, uint32_t offset,
+                OnRecycle&& onRecycleFn)
+            : mParentStage(parentStage),
+              mCapacity(capacity),
+              mOffset(offset),
+              mOnRecycleFn(onRecycleFn) {}
+
+        ~Segment() {
+            if (mOnRecycleFn) {
+                mOnRecycleFn(offset());
+            }
+        }
+
+        // Should not be copying this around.
+        Segment(const Segment& other) = delete;
+        Segment(Segment&& other) = delete;
+        Segment& operator=(const Segment& other) = delete;
+        Segment& operator=(Segment&& other) = delete;
+
+        inline VulkanStage* parentStage() const { return mParentStage; }
+        inline VkBuffer buffer() const { return parentStage()->buffer(); }
+        inline VmaAllocation memory() const { return parentStage()->memory(); }
+        inline uint32_t capacity() const { return mCapacity; }
+        inline uint32_t offset() const { return mOffset; }
+
+        inline void* mapping() const {
+            return reinterpret_cast<void*>(
+                    reinterpret_cast<char*>(mParentStage->mapping()) + offset());
+        }
+
+    private:
+        // Ensure parent class can access the terminate method.
+        friend class VulkanStage;
+
+        VulkanStage* const mParentStage;
+        const uint32_t mCapacity;
+        const uint32_t mOffset;
+        OnRecycle mOnRecycleFn;
+    };
+
+    inline VmaAllocation memory() const { return mMemory; }
+    inline VkBuffer buffer() const { return mBuffer; }
+    inline uint32_t capacity() const { return mCapacity; }
+    inline void* mapping() const { return mMapping; }
+
+    inline uint32_t currentOffset() { return mCurrentOffset; }
+
+    inline bool isSafeToReset() const { return mSegments.empty(); }
+
+    inline void reset() { mCurrentOffset = 0; }
+
+    // Marks a region of the block as "in-use", and provides information about
+    // the allocated region to the caller. Note: this assumes that numBytes
+    // is aligned to the physical device's nonCoherentAtomSize.
+    fvkmemory::resource_ptr<Segment> acquireSegment(fvkmemory::ResourceManager* resManager,
+            uint32_t numBytes);
+
+private:
+    const VmaAllocation mMemory;
+    const VkBuffer mBuffer;
+    const uint32_t mCapacity;
+
+    void* mMapping;
+
+    uint32_t mCurrentOffset = 0;
+
+    // Maps the start offset of a vulkan stage block to the stage block,
+    // for easy deletions later. This is managed by the blocks themselves, in an
+    // RAII pattern, during construction and destruction.
+    std::unordered_map<uint32_t, Segment*> mSegments;
 };
 
 struct VulkanStageImage {
@@ -49,11 +136,15 @@ struct VulkanStageImage {
 // This class manages two types of host-mappable staging areas: buffer stages and image stages.
 class VulkanStagePool {
 public:
-    VulkanStagePool(VmaAllocator allocator, VulkanCommands* commands);
+    VulkanStagePool(VmaAllocator allocator, fvkmemory::ResourceManager* resManager,
+            VulkanCommands* commands, const VkPhysicalDeviceLimits* deviceLimits);
 
-    // Finds or creates a stage whose capacity is at least the given number of bytes.
-    // The stage is automatically released back to the pool after TIME_BEFORE_EVICTION frames.
-    VulkanStage const* acquireStage(uint32_t numBytes);
+    // Finds or creates a stage block whose capacity is at least the given
+    // number of bytes. Internally, creates and manages and subdivides large
+    // buffers so that we have less objects around that we have to keep track
+    // of.
+    // This function is NOT thread-safe.
+    fvkmemory::resource_ptr<VulkanStage::Segment> acquireStage(uint32_t numBytes);
 
     // Images have VK_IMAGE_LAYOUT_GENERAL and must not be transitioned to any other layout
     VulkanStageImage const* acquireImage(PixelDataFormat format, PixelDataType type,
@@ -64,17 +155,37 @@ public:
 
     // Destroys all unused stages and asserts that there are no stages currently in use.
     // This should be called while the context's VkDevice is still alive.
+    // Note: it is expected that all resources have been reclaimed before this
+    // is called. It is also expected that this stage pool does not hold any
+    // resource_ptrs, as this would lead to undefined behavior.
     void terminate() noexcept;
 
 private:
     VmaAllocator mAllocator;
+    fvkmemory::ResourceManager* mResManager;
     VulkanCommands* mCommands;
+    const VkPhysicalDeviceLimits* mDeviceLimits;
+
+    // Takes a number of bytes, and aligns it to the non-coherent atom size.
+    // This allows us to ensure that when we flush buffers from the host, we
+    // never flush more atoms than we need to.
+    uint32_t alignToNonCoherentAtomSize(uint32_t numBytes);
+
+    // Allocates a new stage buffer, and optionally subdivides it into stage
+    // blocks. If subdivideBlocks is true, predefined divisions will be used.
+    // Otherwise, it's expected that capacity is defined to a value, and that
+    // is the size that will be used for the buffer (as well as the only block
+    // being created).
+    VulkanStage* allocateNewStage(uint32_t capacity);
+
+    // Performs any bookkeeping required to delete a VulkanStage object; namely,
+    // unmapping memory, freeing the allocation, and deleting the VulkanStage
+    // object. Note: takes an r-value because after this call, `stage` won't
+    // exist.
+    void destroyStage(VulkanStage const*&& stage);
 
     // Use an ordered multimap for quick (capacity => stage) lookups using lower_bound().
-    std::multimap<uint32_t, VulkanStage const*> mFreeStages;
-
-    // Simple unordered set for stashing a list of in-use stages that can be reclaimed later.
-    std::vector<VulkanStage const*> mUsedStages;
+    std::multimap<uint32_t, VulkanStage*> mStages;
 
     std::unordered_set<VulkanStageImage const*> mFreeImages;
     std::vector<VulkanStageImage const*> mUsedImages;

filament/backend/src/vulkan/VulkanTexture.cpp

@@ -25,6 +25,7 @@
 #include <backend/DriverEnums.h>
 #include <private/backend/BackendUtils.h>
 
+#include <utils/compiler.h>
 #include <utils/Panic.h>
 
 using namespace bluevk;
@@ -125,7 +126,7 @@ inline VulkanLayout getDefaultLayoutImpl(TextureUsage usage) {
         return VulkanLayout::COLOR_ATTACHMENT;
     }
     // Finally, the layout for an immutable texture is optimal read-only.
-    return VulkanLayout::READ_ONLY;
+    return VulkanLayout::FRAG_READ;
 }
 
 inline VulkanLayout getDefaultLayoutImpl(VkImageUsageFlags vkusage) {
@@ -187,7 +188,7 @@ VkImageUsageFlags getUsage(VulkanContext const& context, uint8_t samples,
             VkFormatProperties props;
             vkGetPhysicalDeviceFormatProperties(physicalDevice, vkFormat, &props);
             if (!(props.optimalTilingFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT)) {
-                FVK_LOGW << "Texture usage is SAMPLEABLE but format " << mState->mVkFormat << " is not "
+                FVK_LOGW << "Texture usage is SAMPLEABLE but format " << vkFormat << " is not "
                         "sampleable with optimal tiling." << utils::io::endl;
             }
         }
@@ -218,6 +219,40 @@ VkImageUsageFlags getUsage(VulkanContext const& context, uint8_t samples,
     return usage;
 }
 
+void adjustedMemcpy(void* mapped, PixelBufferDescriptor const& p, size_t width, size_t height,
+        size_t depth) {
+    uint8_t* buf = (uint8_t*) p.buffer;
+    size_t const pixelSize = PixelBufferDescriptor::computeDataSize(p.format, p.type, 1, 1, 1);
+    size_t const pbdStride = p.stride ? p.stride : width;
+
+    // Slow path of copying row by row
+    assert_invariant(pbdStride >= width);
+    if (UTILS_UNLIKELY(p.left > 0 || p.top > 0 || pbdStride > width)) {
+        size_t const pbdRowSize =
+                PixelBufferDescriptor::computeDataSize(p.format, p.type, pbdStride, 1, p.alignment);
+        size_t const pbdHeight = p.size / pixelSize / pbdStride / depth;
+        size_t const pbdLayerSize = pbdRowSize * pbdHeight;
+
+        size_t const rowSize = width * pixelSize;
+        size_t const layerSize = width * height * pixelSize;
+
+        // Size of a row to write
+        size_t const writeSize = std::min(pbdStride - p.left, width) * pixelSize;
+
+        for (size_t z = 0; z < depth; z++) {
+            for (size_t y = p.top; y < pbdHeight; y++) {
+                uint8_t* buf = (uint8_t*) p.buffer +
+                               ((p.left * pixelSize) + (y * pbdRowSize) + (z * pbdLayerSize));
+                uint8_t* curMapped = (uint8_t*) mapped + ((y - p.top) * rowSize + z * layerSize);
+                memcpy(curMapped, buf, writeSize);
+            }
+        }
+    } else {
+        size_t const writeSize = pixelSize * (width * height * depth);
+        memcpy(mapped, buf, writeSize);
+    }
+}
+
 } // anonymous namespace
 
 VulkanTextureState::VulkanTextureState(VulkanStagePool& stagePool, VulkanCommands* commands,
@@ -480,30 +515,40 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
     assert_invariant(hostData->size > 0 && "Data is empty");
 
     // Otherwise, use vkCmdCopyBufferToImage.
-    void* mapped = nullptr;
-    VulkanStage const* stage = mState->mStagePool.acquireStage(hostData->size);
-    assert_invariant(stage->memory);
-    vmaMapMemory(mState->mAllocator, stage->memory, &mapped);
-    memcpy(mapped, hostData->buffer, hostData->size);
-    vmaUnmapMemory(mState->mAllocator, stage->memory);
-    vmaFlushAllocation(mState->mAllocator, stage->memory, 0, hostData->size);
+    size_t const bpp =
+            PixelBufferDescriptor::computeDataSize(hostData->format, hostData->type, 1, 1, 1);
+    size_t const writeSize = width * height * depth * bpp;
+
+    // Note: the following stageSegment must be stored within the command buffer
+    // before going out of scope, to ensure proper bookkeeping within the
+    // staging buffer pool.
+    fvkmemory::resource_ptr<VulkanStage::Segment> stageSegment =
+            mState->mStagePool.acquireStage(writeSize);
+    assert_invariant(stageSegment->memory());
+    adjustedMemcpy(stageSegment->mapping(), *hostData, width, height, depth);
+    vmaFlushAllocation(mState->mAllocator, stageSegment->memory(), stageSegment->offset(),
+            writeSize);
 
     VulkanCommandBuffer& commands = mState->mCommands->get();
     VkCommandBuffer const cmdbuf = commands.buffer();
+    commands.acquire(stageSegment);
     commands.acquire(fvkmemory::resource_ptr<VulkanTexture>::cast(this));
 
+    bool const isDepth = getImageAspect() & VK_IMAGE_ASPECT_DEPTH_BIT;
+
     VkBufferImageCopy copyRegion = {
-        .bufferOffset = {},
+        .bufferOffset = stageSegment->offset(),
         .bufferRowLength = {},
         .bufferImageHeight = {},
         .imageSubresource = {
-            .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            .aspectMask = VkImageAspectFlags(
+                    isDepth ? VK_IMAGE_ASPECT_DEPTH_BIT : VK_IMAGE_ASPECT_COLOR_BIT),
             .mipLevel = miplevel,
             .baseArrayLayer = 0,
-            .layerCount = 1
+            .layerCount = 1,
         },
         .imageOffset = { int32_t(xoffset), int32_t(yoffset), int32_t(zoffset) },
-        .imageExtent = { width, height, depth }
+        .imageExtent = { width, height, depth },
     };
 
     VkImageSubresourceRange transitionRange = {
@@ -511,7 +556,7 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
         .baseMipLevel = miplevel,
         .levelCount = 1,
         .baseArrayLayer = 0,
-        .layerCount = 1
+        .layerCount = 1,
     };
 
     // Vulkan specifies subregions for 3D textures differently than from 2D arrays.
@@ -536,20 +581,25 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
 
     transitionLayout(&commands, transitionRange, newLayout);
 
-    vkCmdCopyBufferToImage(cmdbuf, stage->buffer, mState->mTextureImage, newVkLayout, 1, &copyRegion);
+    vkCmdCopyBufferToImage(cmdbuf, stageSegment->buffer(), mState->mTextureImage, newVkLayout, 1,
+            &copyRegion);
 
     transitionLayout(&commands, transitionRange, nextLayout);
 }
 
-void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, uint32_t width,
+void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& data, uint32_t width,
         uint32_t height, uint32_t depth, uint32_t miplevel) {
+    // Otherwise, use vkCmdCopyBufferToImage.
+    size_t const bpp = PixelBufferDescriptor::computeDataSize(data.format, data.type, 1, 1, 1);
+    size_t const writeSize = width * height * depth * bpp;
+
     void* mapped = nullptr;
     VulkanStageImage const* stage
-            = mState->mStagePool.acquireImage(hostData.format, hostData.type, width, height);
+            = mState->mStagePool.acquireImage(data.format, data.type, width, height);
     vmaMapMemory(mState->mAllocator, stage->memory, &mapped);
-    memcpy(mapped, hostData.buffer, hostData.size);
+    adjustedMemcpy(mapped, data, width, height, depth);
     vmaUnmapMemory(mState->mAllocator, stage->memory);
-    vmaFlushAllocation(mState->mAllocator, stage->memory, 0, hostData.size);
+    vmaFlushAllocation(mState->mAllocator, stage->memory, 0, writeSize);
 
     VulkanCommandBuffer& commands = mState->mCommands->get();
     VkCommandBuffer const cmdbuf = commands.buffer();
@@ -691,6 +741,7 @@ bool VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, VkImageSubresourceR
                  << " is skipped because of no change in layout" << utils::io::endl;
 #endif
     }
+
     return hasTransitions;
 }
 
@@ -700,16 +751,15 @@ void VulkanTexture::samplerToAttachmentBarrier(VulkanCommandBuffer* commands,
     VkImageLayout const layout =
             fvkutils::getVkLayout(getLayout(range.baseArrayLayer, range.baseMipLevel));
     VkImageMemoryBarrier barrier = {
-            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-            .srcAccessMask = VK_ACCESS_SHADER_READ_BIT,
-            .dstAccessMask =
-                    VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-            .oldLayout = layout,
-            .newLayout = layout,
-            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .image = mState->mTextureImage,
-            .subresourceRange = range,
+        .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+        .srcAccessMask = VK_ACCESS_SHADER_READ_BIT,
+        .dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+        .oldLayout = layout,
+        .newLayout = layout,
+        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .image = mState->mTextureImage,
+        .subresourceRange = range,
     };
     vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
             VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
@@ -723,15 +773,15 @@ void VulkanTexture::attachmentToSamplerBarrier(VulkanCommandBuffer* commands,
     VkImageLayout const layout
             = fvkutils::getVkLayout(getLayout(range.baseArrayLayer, range.baseMipLevel));
     VkImageMemoryBarrier barrier = {
-            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-            .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
-            .oldLayout = layout,
-            .newLayout = layout,
-            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .image = mState->mTextureImage,
-            .subresourceRange = range,
+        .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+        .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
+        .oldLayout = layout,
+        .newLayout = layout,
+        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .image = mState->mTextureImage,
+        .subresourceRange = range,
     };
     vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
             VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &barrier);

filament/backend/src/vulkan/VulkanTexture.h

@@ -145,8 +145,14 @@ struct VulkanTexture : public HwTexture, fvkmemory::Resource {
 
     VkImageSubresourceRange const& getPrimaryViewRange() const { return mPrimaryViewRange; }
 
-    VulkanLayout getPrimaryImageLayout() const {
-        return getLayout(mPrimaryViewRange.baseArrayLayer, mPrimaryViewRange.baseMipLevel);
+    VulkanLayout getSamplerLayout() const {
+        if (!isSampleable()) {
+            return VulkanLayout::UNDEFINED;
+        }
+        if (mState->mUsage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+            return VulkanLayout::DEPTH_SAMPLER;
+        }
+        return VulkanLayout::FRAG_READ;
     }
 
     // Returns the layout for the intended use of this texture (and not the expected layout at the
@@ -180,6 +186,10 @@ struct VulkanTexture : public HwTexture, fvkmemory::Resource {
         return mState->mUsage & VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT;
     }
 
+    bool isSampleable() const {
+        return mState->mUsage & VK_IMAGE_USAGE_SAMPLED_BIT;
+    }
+
     bool getIsProtected() const {
         return mState->mIsProtected;
     }

filament/backend/src/vulkan/memory/Resource.cpp

@@ -26,6 +26,7 @@ template ResourceType getTypeEnum<VulkanIndexBuffer>() noexcept;
 template ResourceType getTypeEnum<VulkanProgram>() noexcept;
 template ResourceType getTypeEnum<VulkanRenderTarget>() noexcept;
 template ResourceType getTypeEnum<VulkanSwapChain>() noexcept;
+template ResourceType getTypeEnum<VulkanStage::Segment>() noexcept;
 template ResourceType getTypeEnum<VulkanRenderPrimitive>() noexcept;
 template ResourceType getTypeEnum<VulkanTexture>() noexcept;
 template ResourceType getTypeEnum<VulkanTextureState>() noexcept;
@@ -35,6 +36,7 @@ template ResourceType getTypeEnum<VulkanVertexBufferInfo>() noexcept;
 template ResourceType getTypeEnum<VulkanDescriptorSetLayout>() noexcept;
 template ResourceType getTypeEnum<VulkanDescriptorSet>() noexcept;
 template ResourceType getTypeEnum<VulkanFence>() noexcept;
+template ResourceType getTypeEnum<VulkanBuffer>() noexcept;
 
 template<typename D>
 ResourceType getTypeEnum() noexcept {
@@ -53,6 +55,9 @@ ResourceType getTypeEnum() noexcept {
     if constexpr (std::is_same_v<D, VulkanSwapChain>) {
         return ResourceType::SWAP_CHAIN;
     }
+    if constexpr (std::is_same_v<D, VulkanStage::Segment>) {
+        return ResourceType::STAGE_SEGMENT;
+    }
     if constexpr (std::is_same_v<D, VulkanRenderPrimitive>) {
         return ResourceType::RENDER_PRIMITIVE;
     }
@@ -80,6 +85,9 @@ ResourceType getTypeEnum() noexcept {
     if constexpr (std::is_same_v<D, VulkanFence>) {
         return ResourceType::FENCE;
     }
+    if constexpr (std::is_same_v<D, VulkanBuffer>) {
+        return ResourceType::VULKAN_BUFFER;
+    }
     return ResourceType::UNDEFINED_TYPE;
 }
 
@@ -95,6 +103,8 @@ std::string getTypeStr(ResourceType type) {
             return "RenderTarget";
         case ResourceType::SWAP_CHAIN:
             return "SwapChain";
+        case ResourceType::STAGE_SEGMENT:
+            return "Stage::Segment";
         case ResourceType::RENDER_PRIMITIVE:
             return "RenderPrimitive";
         case ResourceType::TEXTURE:
@@ -113,6 +123,8 @@ std::string getTypeStr(ResourceType type) {
             return "DescriptorSet";
         case ResourceType::FENCE:
             return "Fence";
+        case ResourceType::VULKAN_BUFFER:
+            return "VulkanBuffer";
         case ResourceType::UNDEFINED_TYPE:
             return "";
     }

filament/backend/src/vulkan/memory/Resource.h

@@ -49,7 +49,9 @@ enum class ResourceType : uint8_t {
     DESCRIPTOR_SET_LAYOUT = 11,
     DESCRIPTOR_SET = 12,
     FENCE = 13,
-    UNDEFINED_TYPE = 14,    // Must be the last enum because we use it for iterating over the enums.
+    VULKAN_BUFFER = 14,
+    STAGE_SEGMENT = 15,
+    UNDEFINED_TYPE = 16,    // Must be the last enum because we use it for iterating over the enums.
 };
 
 template<typename D>

filament/backend/src/vulkan/memory/ResourceManager.cpp

@@ -77,6 +77,9 @@ void ResourceManager::destroyWithType(ResourceType type, HandleId id) {
         case ResourceType::SWAP_CHAIN:
             destruct<VulkanSwapChain>(Handle<VulkanSwapChain>(id));
             break;
+        case ResourceType::STAGE_SEGMENT:
+            destruct<VulkanStage::Segment>(Handle<VulkanStage::Segment>(id));
+            break;
         case ResourceType::RENDER_PRIMITIVE:
             destruct<VulkanRenderPrimitive>(Handle<VulkanRenderPrimitive>(id));
             break;
@@ -104,6 +107,9 @@ void ResourceManager::destroyWithType(ResourceType type, HandleId id) {
         case ResourceType::FENCE:
             destruct<VulkanFence>(Handle<VulkanFence>(id));
             break;
+        case ResourceType::VULKAN_BUFFER:
+            destruct<VulkanBuffer>(Handle<VulkanBuffer>(id));
+            break;
         case ResourceType::UNDEFINED_TYPE:
             break;
     }

filament/backend/src/vulkan/platform/VulkanPlatform.cpp

@@ -92,6 +92,16 @@ StructA* chainStruct(StructA* structA, StructB* structB) {
     return structA;
 }
 
+bool shouldSkipFormat(VkFormat format) {
+    // Skip formats that require extensions.
+    for (VkFormat const extFormat: fvkutils::EXT_VK_FORMATS) {
+        if (format == extFormat) {
+            return true;
+        }
+    }
+    return false;
+}
+
 void printDeviceInfo(VkInstance instance, VkPhysicalDevice device) {
     // Print some driver or MoltenVK information if it is available.
     if (vkGetPhysicalDeviceProperties2) {
@@ -151,8 +161,14 @@ void printDepthFormats(VkPhysicalDevice device) {
     constexpr VkFormatFeatureFlags required =
             VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT | VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
     FVK_LOGI << "Sampleable depth formats: ";
-    for (VkFormat format : fvkutils::ALL_VK_FORMATS) {
+    for (VkFormat const format : fvkutils::ALL_VK_FORMATS) {
+        // Skip formats that require extensions.
+        if (shouldSkipFormat(format)) {
+            continue;
+        }
+
         VkFormatProperties props;
+
         vkGetPhysicalDeviceFormatProperties(device, format, &props);
         if ((props.optimalTilingFeatures & required) == required) {
             FVK_LOGI << format << " ";
@@ -617,7 +633,13 @@ fvkutils::VkFormatList findBlittableDepthStencilFormats(VkPhysicalDevice device)
     std::vector<VkFormat> selectedFormats;
     constexpr VkFormatFeatureFlags required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT |
             VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
-    for (VkFormat format : fvkutils::ALL_VK_FORMATS) {
+
+    for (VkFormat const format : fvkutils::ALL_VK_FORMATS) {
+        // Skip formats that require extensions.
+        if (shouldSkipFormat(format)) {
+            continue;
+        }
+
         if (fvkutils::isVkDepthFormat(format)) {
             VkFormatProperties props;
             vkGetPhysicalDeviceFormatProperties(device, format, &props);

filament/backend/src/vulkan/utils/Definitions.h

@@ -32,313 +32,325 @@ using VkFormatList = utils::FixedCapacityVector<VkFormat>;
 // Copied from
 //   https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkFormat.html
 constexpr VkFormat ALL_VK_FORMATS[] = {
-        VK_FORMAT_UNDEFINED,
-        VK_FORMAT_R4G4_UNORM_PACK8,
-        VK_FORMAT_R4G4B4A4_UNORM_PACK16,
-        VK_FORMAT_B4G4R4A4_UNORM_PACK16,
-        VK_FORMAT_R5G6B5_UNORM_PACK16,
-        VK_FORMAT_B5G6R5_UNORM_PACK16,
-        VK_FORMAT_R5G5B5A1_UNORM_PACK16,
-        VK_FORMAT_B5G5R5A1_UNORM_PACK16,
-        VK_FORMAT_A1R5G5B5_UNORM_PACK16,
-        VK_FORMAT_R8_UNORM,
-        VK_FORMAT_R8_SNORM,
-        VK_FORMAT_R8_USCALED,
-        VK_FORMAT_R8_SSCALED,
-        VK_FORMAT_R8_UINT,
-        VK_FORMAT_R8_SINT,
-        VK_FORMAT_R8_SRGB,
-        VK_FORMAT_R8G8_UNORM,
-        VK_FORMAT_R8G8_SNORM,
-        VK_FORMAT_R8G8_USCALED,
-        VK_FORMAT_R8G8_SSCALED,
-        VK_FORMAT_R8G8_UINT,
-        VK_FORMAT_R8G8_SINT,
-        VK_FORMAT_R8G8_SRGB,
-        VK_FORMAT_R8G8B8_UNORM,
-        VK_FORMAT_R8G8B8_SNORM,
-        VK_FORMAT_R8G8B8_USCALED,
-        VK_FORMAT_R8G8B8_SSCALED,
-        VK_FORMAT_R8G8B8_UINT,
-        VK_FORMAT_R8G8B8_SINT,
-        VK_FORMAT_R8G8B8_SRGB,
-        VK_FORMAT_B8G8R8_UNORM,
-        VK_FORMAT_B8G8R8_SNORM,
-        VK_FORMAT_B8G8R8_USCALED,
-        VK_FORMAT_B8G8R8_SSCALED,
-        VK_FORMAT_B8G8R8_UINT,
-        VK_FORMAT_B8G8R8_SINT,
-        VK_FORMAT_B8G8R8_SRGB,
-        VK_FORMAT_R8G8B8A8_UNORM,
-        VK_FORMAT_R8G8B8A8_SNORM,
-        VK_FORMAT_R8G8B8A8_USCALED,
-        VK_FORMAT_R8G8B8A8_SSCALED,
-        VK_FORMAT_R8G8B8A8_UINT,
-        VK_FORMAT_R8G8B8A8_SINT,
-        VK_FORMAT_R8G8B8A8_SRGB,
-        VK_FORMAT_B8G8R8A8_UNORM,
-        VK_FORMAT_B8G8R8A8_SNORM,
-        VK_FORMAT_B8G8R8A8_USCALED,
-        VK_FORMAT_B8G8R8A8_SSCALED,
-        VK_FORMAT_B8G8R8A8_UINT,
-        VK_FORMAT_B8G8R8A8_SINT,
-        VK_FORMAT_B8G8R8A8_SRGB,
-        VK_FORMAT_A8B8G8R8_UNORM_PACK32,
-        VK_FORMAT_A8B8G8R8_SNORM_PACK32,
-        VK_FORMAT_A8B8G8R8_USCALED_PACK32,
-        VK_FORMAT_A8B8G8R8_SSCALED_PACK32,
-        VK_FORMAT_A8B8G8R8_UINT_PACK32,
-        VK_FORMAT_A8B8G8R8_SINT_PACK32,
-        VK_FORMAT_A8B8G8R8_SRGB_PACK32,
-        VK_FORMAT_A2R10G10B10_UNORM_PACK32,
-        VK_FORMAT_A2R10G10B10_SNORM_PACK32,
-        VK_FORMAT_A2R10G10B10_USCALED_PACK32,
-        VK_FORMAT_A2R10G10B10_SSCALED_PACK32,
-        VK_FORMAT_A2R10G10B10_UINT_PACK32,
-        VK_FORMAT_A2R10G10B10_SINT_PACK32,
-        VK_FORMAT_A2B10G10R10_UNORM_PACK32,
-        VK_FORMAT_A2B10G10R10_SNORM_PACK32,
-        VK_FORMAT_A2B10G10R10_USCALED_PACK32,
-        VK_FORMAT_A2B10G10R10_SSCALED_PACK32,
-        VK_FORMAT_A2B10G10R10_UINT_PACK32,
-        VK_FORMAT_A2B10G10R10_SINT_PACK32,
-        VK_FORMAT_R16_UNORM,
-        VK_FORMAT_R16_SNORM,
-        VK_FORMAT_R16_USCALED,
-        VK_FORMAT_R16_SSCALED,
-        VK_FORMAT_R16_UINT,
-        VK_FORMAT_R16_SINT,
-        VK_FORMAT_R16_SFLOAT,
-        VK_FORMAT_R16G16_UNORM,
-        VK_FORMAT_R16G16_SNORM,
-        VK_FORMAT_R16G16_USCALED,
-        VK_FORMAT_R16G16_SSCALED,
-        VK_FORMAT_R16G16_UINT,
-        VK_FORMAT_R16G16_SINT,
-        VK_FORMAT_R16G16_SFLOAT,
-        VK_FORMAT_R16G16B16_UNORM,
-        VK_FORMAT_R16G16B16_SNORM,
-        VK_FORMAT_R16G16B16_USCALED,
-        VK_FORMAT_R16G16B16_SSCALED,
-        VK_FORMAT_R16G16B16_UINT,
-        VK_FORMAT_R16G16B16_SINT,
-        VK_FORMAT_R16G16B16_SFLOAT,
-        VK_FORMAT_R16G16B16A16_UNORM,
-        VK_FORMAT_R16G16B16A16_SNORM,
-        VK_FORMAT_R16G16B16A16_USCALED,
-        VK_FORMAT_R16G16B16A16_SSCALED,
-        VK_FORMAT_R16G16B16A16_UINT,
-        VK_FORMAT_R16G16B16A16_SINT,
-        VK_FORMAT_R16G16B16A16_SFLOAT,
-        VK_FORMAT_R32_UINT,
-        VK_FORMAT_R32_SINT,
-        VK_FORMAT_R32_SFLOAT,
-        VK_FORMAT_R32G32_UINT,
-        VK_FORMAT_R32G32_SINT,
-        VK_FORMAT_R32G32_SFLOAT,
-        VK_FORMAT_R32G32B32_UINT,
-        VK_FORMAT_R32G32B32_SINT,
-        VK_FORMAT_R32G32B32_SFLOAT,
-        VK_FORMAT_R32G32B32A32_UINT,
-        VK_FORMAT_R32G32B32A32_SINT,
-        VK_FORMAT_R32G32B32A32_SFLOAT,
-        VK_FORMAT_R64_UINT,
-        VK_FORMAT_R64_SINT,
-        VK_FORMAT_R64_SFLOAT,
-        VK_FORMAT_R64G64_UINT,
-        VK_FORMAT_R64G64_SINT,
-        VK_FORMAT_R64G64_SFLOAT,
-        VK_FORMAT_R64G64B64_UINT,
-        VK_FORMAT_R64G64B64_SINT,
-        VK_FORMAT_R64G64B64_SFLOAT,
-        VK_FORMAT_R64G64B64A64_UINT,
-        VK_FORMAT_R64G64B64A64_SINT,
-        VK_FORMAT_R64G64B64A64_SFLOAT,
-        VK_FORMAT_B10G11R11_UFLOAT_PACK32,
-        VK_FORMAT_E5B9G9R9_UFLOAT_PACK32,
-        VK_FORMAT_D16_UNORM,
-        VK_FORMAT_X8_D24_UNORM_PACK32,
-        VK_FORMAT_D32_SFLOAT,
-        VK_FORMAT_S8_UINT,
-        VK_FORMAT_D16_UNORM_S8_UINT,
-        VK_FORMAT_D24_UNORM_S8_UINT,
-        VK_FORMAT_D32_SFLOAT_S8_UINT,
-        VK_FORMAT_BC1_RGB_UNORM_BLOCK,
-        VK_FORMAT_BC1_RGB_SRGB_BLOCK,
-        VK_FORMAT_BC1_RGBA_UNORM_BLOCK,
-        VK_FORMAT_BC1_RGBA_SRGB_BLOCK,
-        VK_FORMAT_BC2_UNORM_BLOCK,
-        VK_FORMAT_BC2_SRGB_BLOCK,
-        VK_FORMAT_BC3_UNORM_BLOCK,
-        VK_FORMAT_BC3_SRGB_BLOCK,
-        VK_FORMAT_BC4_UNORM_BLOCK,
-        VK_FORMAT_BC4_SNORM_BLOCK,
-        VK_FORMAT_BC5_UNORM_BLOCK,
-        VK_FORMAT_BC5_SNORM_BLOCK,
-        VK_FORMAT_BC6H_UFLOAT_BLOCK,
-        VK_FORMAT_BC6H_SFLOAT_BLOCK,
-        VK_FORMAT_BC7_UNORM_BLOCK,
-        VK_FORMAT_BC7_SRGB_BLOCK,
-        VK_FORMAT_ETC2_R8G8B8_UNORM_BLOCK,
-        VK_FORMAT_ETC2_R8G8B8_SRGB_BLOCK,
-        VK_FORMAT_ETC2_R8G8B8A1_UNORM_BLOCK,
-        VK_FORMAT_ETC2_R8G8B8A1_SRGB_BLOCK,
-        VK_FORMAT_ETC2_R8G8B8A8_UNORM_BLOCK,
-        VK_FORMAT_ETC2_R8G8B8A8_SRGB_BLOCK,
-        VK_FORMAT_EAC_R11_UNORM_BLOCK,
-        VK_FORMAT_EAC_R11_SNORM_BLOCK,
-        VK_FORMAT_EAC_R11G11_UNORM_BLOCK,
-        VK_FORMAT_EAC_R11G11_SNORM_BLOCK,
-        VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
-        VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
-        VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
-        VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
-        VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
-        VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
-        VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
-        VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
-        VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
-        VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
-        VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
-        VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
-        VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
-        VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
-        VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
-        VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
-        VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
-        VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
-        VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
-        VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
-        VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
-        VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
-        VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
-        VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
-        VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
-        VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
-        VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
-        VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
-        VK_FORMAT_G8B8G8R8_422_UNORM,
-        VK_FORMAT_B8G8R8G8_422_UNORM,
-        VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM,
-        VK_FORMAT_G8_B8R8_2PLANE_420_UNORM,
-        VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM,
-        VK_FORMAT_G8_B8R8_2PLANE_422_UNORM,
-        VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM,
-        VK_FORMAT_R10X6_UNORM_PACK16,
-        VK_FORMAT_R10X6G10X6_UNORM_2PACK16,
-        VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16,
-        VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16,
-        VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16,
-        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16,
-        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16,
-        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16,
-        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16,
-        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16,
-        VK_FORMAT_R12X4_UNORM_PACK16,
-        VK_FORMAT_R12X4G12X4_UNORM_2PACK16,
-        VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16,
-        VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16,
-        VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16,
-        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16,
-        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16,
-        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16,
-        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16,
-        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16,
-        VK_FORMAT_G16B16G16R16_422_UNORM,
-        VK_FORMAT_B16G16R16G16_422_UNORM,
-        VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM,
-        VK_FORMAT_G16_B16R16_2PLANE_420_UNORM,
-        VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM,
-        VK_FORMAT_G16_B16R16_2PLANE_422_UNORM,
-        VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM,
-        VK_FORMAT_G8_B8R8_2PLANE_444_UNORM,
-        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16,
-        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16,
-        VK_FORMAT_G16_B16R16_2PLANE_444_UNORM,
-        VK_FORMAT_A4R4G4B4_UNORM_PACK16,
-        VK_FORMAT_A4B4G4R4_UNORM_PACK16,
-        VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_5x4_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_5x5_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_6x5_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_6x6_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_8x5_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_8x6_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_8x8_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_10x5_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_10x6_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_10x8_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_10x10_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_12x10_SFLOAT_BLOCK,
-        VK_FORMAT_ASTC_12x12_SFLOAT_BLOCK,
-        VK_FORMAT_PVRTC1_2BPP_UNORM_BLOCK_IMG,
-        VK_FORMAT_PVRTC1_4BPP_UNORM_BLOCK_IMG,
-        VK_FORMAT_PVRTC2_2BPP_UNORM_BLOCK_IMG,
-        VK_FORMAT_PVRTC2_4BPP_UNORM_BLOCK_IMG,
-        VK_FORMAT_PVRTC1_2BPP_SRGB_BLOCK_IMG,
-        VK_FORMAT_PVRTC1_4BPP_SRGB_BLOCK_IMG,
-        VK_FORMAT_PVRTC2_2BPP_SRGB_BLOCK_IMG,
-        VK_FORMAT_PVRTC2_4BPP_SRGB_BLOCK_IMG,
-// Not supported (yet) by bluevk
-//        VK_FORMAT_R16G16_SFIXED5_NV,
-//        VK_FORMAT_A1B5G5R5_UNORM_PACK16_KHR,
-//        VK_FORMAT_A8_UNORM_KHR,
-//        VK_FORMAT_A8_UNORM,
-        VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_5x4_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_5x5_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_6x5_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_6x6_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_8x5_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_8x6_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_8x8_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_10x5_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_10x6_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_10x8_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_10x10_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_12x10_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_ASTC_12x12_SFLOAT_BLOCK_EXT,
-        VK_FORMAT_G8B8G8R8_422_UNORM_KHR,
-        VK_FORMAT_B8G8R8G8_422_UNORM_KHR,
-        VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM_KHR,
-        VK_FORMAT_G8_B8R8_2PLANE_420_UNORM_KHR,
-        VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM_KHR,
-        VK_FORMAT_G8_B8R8_2PLANE_422_UNORM_KHR,
-        VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM_KHR,
-        VK_FORMAT_R10X6_UNORM_PACK16_KHR,
-        VK_FORMAT_R10X6G10X6_UNORM_2PACK16_KHR,
-        VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16_KHR,
-        VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16_KHR,
-        VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16_KHR,
-        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16_KHR,
-        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16_KHR,
-        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16_KHR,
-        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16_KHR,
-        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16_KHR,
-        VK_FORMAT_R12X4_UNORM_PACK16_KHR,
-        VK_FORMAT_R12X4G12X4_UNORM_2PACK16_KHR,
-        VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16_KHR,
-        VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16_KHR,
-        VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16_KHR,
-        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16_KHR,
-        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16_KHR,
-        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16_KHR,
-        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16_KHR,
-        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16_KHR,
-        VK_FORMAT_G16B16G16R16_422_UNORM_KHR,
-        VK_FORMAT_B16G16R16G16_422_UNORM_KHR,
-        VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM_KHR,
-        VK_FORMAT_G16_B16R16_2PLANE_420_UNORM_KHR,
-        VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM_KHR,
-        VK_FORMAT_G16_B16R16_2PLANE_422_UNORM_KHR,
-        VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM_KHR,
-        VK_FORMAT_G8_B8R8_2PLANE_444_UNORM_EXT,
-        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16_EXT,
-        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16_EXT,
-        VK_FORMAT_G16_B16R16_2PLANE_444_UNORM_EXT,
-        VK_FORMAT_A4R4G4B4_UNORM_PACK16_EXT,
-        VK_FORMAT_A4B4G4R4_UNORM_PACK16_EXT,
-        VK_FORMAT_R16G16_S10_5_NV,
+    VK_FORMAT_UNDEFINED,
+    VK_FORMAT_R4G4_UNORM_PACK8,
+    VK_FORMAT_R4G4B4A4_UNORM_PACK16,
+    VK_FORMAT_B4G4R4A4_UNORM_PACK16,
+    VK_FORMAT_R5G6B5_UNORM_PACK16,
+    VK_FORMAT_B5G6R5_UNORM_PACK16,
+    VK_FORMAT_R5G5B5A1_UNORM_PACK16,
+    VK_FORMAT_B5G5R5A1_UNORM_PACK16,
+    VK_FORMAT_A1R5G5B5_UNORM_PACK16,
+    VK_FORMAT_R8_UNORM,
+    VK_FORMAT_R8_SNORM,
+    VK_FORMAT_R8_USCALED,
+    VK_FORMAT_R8_SSCALED,
+    VK_FORMAT_R8_UINT,
+    VK_FORMAT_R8_SINT,
+    VK_FORMAT_R8_SRGB,
+    VK_FORMAT_R8G8_UNORM,
+    VK_FORMAT_R8G8_SNORM,
+    VK_FORMAT_R8G8_USCALED,
+    VK_FORMAT_R8G8_SSCALED,
+    VK_FORMAT_R8G8_UINT,
+    VK_FORMAT_R8G8_SINT,
+    VK_FORMAT_R8G8_SRGB,
+    VK_FORMAT_R8G8B8_UNORM,
+    VK_FORMAT_R8G8B8_SNORM,
+    VK_FORMAT_R8G8B8_USCALED,
+    VK_FORMAT_R8G8B8_SSCALED,
+    VK_FORMAT_R8G8B8_UINT,
+    VK_FORMAT_R8G8B8_SINT,
+    VK_FORMAT_R8G8B8_SRGB,
+    VK_FORMAT_B8G8R8_UNORM,
+    VK_FORMAT_B8G8R8_SNORM,
+    VK_FORMAT_B8G8R8_USCALED,
+    VK_FORMAT_B8G8R8_SSCALED,
+    VK_FORMAT_B8G8R8_UINT,
+    VK_FORMAT_B8G8R8_SINT,
+    VK_FORMAT_B8G8R8_SRGB,
+    VK_FORMAT_R8G8B8A8_UNORM,
+    VK_FORMAT_R8G8B8A8_SNORM,
+    VK_FORMAT_R8G8B8A8_USCALED,
+    VK_FORMAT_R8G8B8A8_SSCALED,
+    VK_FORMAT_R8G8B8A8_UINT,
+    VK_FORMAT_R8G8B8A8_SINT,
+    VK_FORMAT_R8G8B8A8_SRGB,
+    VK_FORMAT_B8G8R8A8_UNORM,
+    VK_FORMAT_B8G8R8A8_SNORM,
+    VK_FORMAT_B8G8R8A8_USCALED,
+    VK_FORMAT_B8G8R8A8_SSCALED,
+    VK_FORMAT_B8G8R8A8_UINT,
+    VK_FORMAT_B8G8R8A8_SINT,
+    VK_FORMAT_B8G8R8A8_SRGB,
+    VK_FORMAT_A8B8G8R8_UNORM_PACK32,
+    VK_FORMAT_A8B8G8R8_SNORM_PACK32,
+    VK_FORMAT_A8B8G8R8_USCALED_PACK32,
+    VK_FORMAT_A8B8G8R8_SSCALED_PACK32,
+    VK_FORMAT_A8B8G8R8_UINT_PACK32,
+    VK_FORMAT_A8B8G8R8_SINT_PACK32,
+    VK_FORMAT_A8B8G8R8_SRGB_PACK32,
+    VK_FORMAT_A2R10G10B10_UNORM_PACK32,
+    VK_FORMAT_A2R10G10B10_SNORM_PACK32,
+    VK_FORMAT_A2R10G10B10_USCALED_PACK32,
+    VK_FORMAT_A2R10G10B10_SSCALED_PACK32,
+    VK_FORMAT_A2R10G10B10_UINT_PACK32,
+    VK_FORMAT_A2R10G10B10_SINT_PACK32,
+    VK_FORMAT_A2B10G10R10_UNORM_PACK32,
+    VK_FORMAT_A2B10G10R10_SNORM_PACK32,
+    VK_FORMAT_A2B10G10R10_USCALED_PACK32,
+    VK_FORMAT_A2B10G10R10_SSCALED_PACK32,
+    VK_FORMAT_A2B10G10R10_UINT_PACK32,
+    VK_FORMAT_A2B10G10R10_SINT_PACK32,
+    VK_FORMAT_R16_UNORM,
+    VK_FORMAT_R16_SNORM,
+    VK_FORMAT_R16_USCALED,
+    VK_FORMAT_R16_SSCALED,
+    VK_FORMAT_R16_UINT,
+    VK_FORMAT_R16_SINT,
+    VK_FORMAT_R16_SFLOAT,
+    VK_FORMAT_R16G16_UNORM,
+    VK_FORMAT_R16G16_SNORM,
+    VK_FORMAT_R16G16_USCALED,
+    VK_FORMAT_R16G16_SSCALED,
+    VK_FORMAT_R16G16_UINT,
+    VK_FORMAT_R16G16_SINT,
+    VK_FORMAT_R16G16_SFLOAT,
+    VK_FORMAT_R16G16B16_UNORM,
+    VK_FORMAT_R16G16B16_SNORM,
+    VK_FORMAT_R16G16B16_USCALED,
+    VK_FORMAT_R16G16B16_SSCALED,
+    VK_FORMAT_R16G16B16_UINT,
+    VK_FORMAT_R16G16B16_SINT,
+    VK_FORMAT_R16G16B16_SFLOAT,
+    VK_FORMAT_R16G16B16A16_UNORM,
+    VK_FORMAT_R16G16B16A16_SNORM,
+    VK_FORMAT_R16G16B16A16_USCALED,
+    VK_FORMAT_R16G16B16A16_SSCALED,
+    VK_FORMAT_R16G16B16A16_UINT,
+    VK_FORMAT_R16G16B16A16_SINT,
+    VK_FORMAT_R16G16B16A16_SFLOAT,
+    VK_FORMAT_R32_UINT,
+    VK_FORMAT_R32_SINT,
+    VK_FORMAT_R32_SFLOAT,
+    VK_FORMAT_R32G32_UINT,
+    VK_FORMAT_R32G32_SINT,
+    VK_FORMAT_R32G32_SFLOAT,
+    VK_FORMAT_R32G32B32_UINT,
+    VK_FORMAT_R32G32B32_SINT,
+    VK_FORMAT_R32G32B32_SFLOAT,
+    VK_FORMAT_R32G32B32A32_UINT,
+    VK_FORMAT_R32G32B32A32_SINT,
+    VK_FORMAT_R32G32B32A32_SFLOAT,
+    VK_FORMAT_R64_UINT,
+    VK_FORMAT_R64_SINT,
+    VK_FORMAT_R64_SFLOAT,
+    VK_FORMAT_R64G64_UINT,
+    VK_FORMAT_R64G64_SINT,
+    VK_FORMAT_R64G64_SFLOAT,
+    VK_FORMAT_R64G64B64_UINT,
+    VK_FORMAT_R64G64B64_SINT,
+    VK_FORMAT_R64G64B64_SFLOAT,
+    VK_FORMAT_R64G64B64A64_UINT,
+    VK_FORMAT_R64G64B64A64_SINT,
+    VK_FORMAT_R64G64B64A64_SFLOAT,
+    VK_FORMAT_B10G11R11_UFLOAT_PACK32,
+    VK_FORMAT_E5B9G9R9_UFLOAT_PACK32,
+    VK_FORMAT_D16_UNORM,
+    VK_FORMAT_X8_D24_UNORM_PACK32,
+    VK_FORMAT_D32_SFLOAT,
+    VK_FORMAT_S8_UINT,
+    VK_FORMAT_D16_UNORM_S8_UINT,
+    VK_FORMAT_D24_UNORM_S8_UINT,
+    VK_FORMAT_D32_SFLOAT_S8_UINT,
+    VK_FORMAT_BC1_RGB_UNORM_BLOCK,
+    VK_FORMAT_BC1_RGB_SRGB_BLOCK,
+    VK_FORMAT_BC1_RGBA_UNORM_BLOCK,
+    VK_FORMAT_BC1_RGBA_SRGB_BLOCK,
+    VK_FORMAT_BC2_UNORM_BLOCK,
+    VK_FORMAT_BC2_SRGB_BLOCK,
+    VK_FORMAT_BC3_UNORM_BLOCK,
+    VK_FORMAT_BC3_SRGB_BLOCK,
+    VK_FORMAT_BC4_UNORM_BLOCK,
+    VK_FORMAT_BC4_SNORM_BLOCK,
+    VK_FORMAT_BC5_UNORM_BLOCK,
+    VK_FORMAT_BC5_SNORM_BLOCK,
+    VK_FORMAT_BC6H_UFLOAT_BLOCK,
+    VK_FORMAT_BC6H_SFLOAT_BLOCK,
+    VK_FORMAT_BC7_UNORM_BLOCK,
+    VK_FORMAT_BC7_SRGB_BLOCK,
+    VK_FORMAT_ETC2_R8G8B8_UNORM_BLOCK,
+    VK_FORMAT_ETC2_R8G8B8_SRGB_BLOCK,
+    VK_FORMAT_ETC2_R8G8B8A1_UNORM_BLOCK,
+    VK_FORMAT_ETC2_R8G8B8A1_SRGB_BLOCK,
+    VK_FORMAT_ETC2_R8G8B8A8_UNORM_BLOCK,
+    VK_FORMAT_ETC2_R8G8B8A8_SRGB_BLOCK,
+    VK_FORMAT_EAC_R11_UNORM_BLOCK,
+    VK_FORMAT_EAC_R11_SNORM_BLOCK,
+    VK_FORMAT_EAC_R11G11_UNORM_BLOCK,
+    VK_FORMAT_EAC_R11G11_SNORM_BLOCK,
+    VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
+    VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
+    VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
+    VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
+    VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
+    VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
+    VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
+    VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
+    VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
+    VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
+    VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
+    VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
+    VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
+    VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
+    VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
+    VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
+    VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
+    VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
+    VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
+    VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
+    VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
+    VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
+    VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
+    VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
+    VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
+    VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
+    VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
+    VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
+    VK_FORMAT_G8B8G8R8_422_UNORM,
+    VK_FORMAT_B8G8R8G8_422_UNORM,
+    VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM,
+    VK_FORMAT_G8_B8R8_2PLANE_420_UNORM,
+    VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM,
+    VK_FORMAT_G8_B8R8_2PLANE_422_UNORM,
+    VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM,
+    VK_FORMAT_R10X6_UNORM_PACK16,
+    VK_FORMAT_R10X6G10X6_UNORM_2PACK16,
+    VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16,
+    VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16,
+    VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16,
+    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16,
+    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16,
+    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16,
+    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16,
+    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16,
+    VK_FORMAT_R12X4_UNORM_PACK16,
+    VK_FORMAT_R12X4G12X4_UNORM_2PACK16,
+    VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16,
+    VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16,
+    VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16,
+    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16,
+    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16,
+    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16,
+    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16,
+    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16,
+    VK_FORMAT_G16B16G16R16_422_UNORM,
+    VK_FORMAT_B16G16R16G16_422_UNORM,
+    VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM,
+    VK_FORMAT_G16_B16R16_2PLANE_420_UNORM,
+    VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM,
+    VK_FORMAT_G16_B16R16_2PLANE_422_UNORM,
+    VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM,
+    VK_FORMAT_G8_B8R8_2PLANE_444_UNORM,
+    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16,
+    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16,
+    VK_FORMAT_G16_B16R16_2PLANE_444_UNORM,
+    VK_FORMAT_A4R4G4B4_UNORM_PACK16,
+    VK_FORMAT_A4B4G4R4_UNORM_PACK16,
+    VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_5x4_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_5x5_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_6x5_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_6x6_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_8x5_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_8x6_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_8x8_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_10x5_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_10x6_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_10x8_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_10x10_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_12x10_SFLOAT_BLOCK,
+    VK_FORMAT_ASTC_12x12_SFLOAT_BLOCK,
+    VK_FORMAT_PVRTC1_2BPP_UNORM_BLOCK_IMG,
+    VK_FORMAT_PVRTC1_4BPP_UNORM_BLOCK_IMG,
+    VK_FORMAT_PVRTC2_2BPP_UNORM_BLOCK_IMG,
+    VK_FORMAT_PVRTC2_4BPP_UNORM_BLOCK_IMG,
+    VK_FORMAT_PVRTC1_2BPP_SRGB_BLOCK_IMG,
+    VK_FORMAT_PVRTC1_4BPP_SRGB_BLOCK_IMG,
+    VK_FORMAT_PVRTC2_2BPP_SRGB_BLOCK_IMG,
+    VK_FORMAT_PVRTC2_4BPP_SRGB_BLOCK_IMG,
+    // Not supported (yet) by bluevk
+    //        VK_FORMAT_R16G16_SFIXED5_NV,
+    //        VK_FORMAT_A1B5G5R5_UNORM_PACK16_KHR,
+    //        VK_FORMAT_A8_UNORM_KHR,
+    //        VK_FORMAT_A8_UNORM,
+    VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_5x4_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_5x5_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_6x5_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_6x6_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_8x5_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_8x6_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_8x8_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_10x5_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_10x6_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_10x8_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_10x10_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_12x10_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_ASTC_12x12_SFLOAT_BLOCK_EXT,
+    VK_FORMAT_G8B8G8R8_422_UNORM_KHR,
+    VK_FORMAT_B8G8R8G8_422_UNORM_KHR,
+    VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM_KHR,
+    VK_FORMAT_G8_B8R8_2PLANE_420_UNORM_KHR,
+    VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM_KHR,
+    VK_FORMAT_G8_B8R8_2PLANE_422_UNORM_KHR,
+    VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM_KHR,
+    VK_FORMAT_R10X6_UNORM_PACK16_KHR,
+    VK_FORMAT_R10X6G10X6_UNORM_2PACK16_KHR,
+    VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16_KHR,
+    VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16_KHR,
+    VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16_KHR,
+    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16_KHR,
+    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16_KHR,
+    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16_KHR,
+    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16_KHR,
+    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16_KHR,
+    VK_FORMAT_R12X4_UNORM_PACK16_KHR,
+    VK_FORMAT_R12X4G12X4_UNORM_2PACK16_KHR,
+    VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16_KHR,
+    VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16_KHR,
+    VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16_KHR,
+    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16_KHR,
+    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16_KHR,
+    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16_KHR,
+    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16_KHR,
+    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16_KHR,
+    VK_FORMAT_G16B16G16R16_422_UNORM_KHR,
+    VK_FORMAT_B16G16R16G16_422_UNORM_KHR,
+    VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM_KHR,
+    VK_FORMAT_G16_B16R16_2PLANE_420_UNORM_KHR,
+    VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM_KHR,
+    VK_FORMAT_G16_B16R16_2PLANE_422_UNORM_KHR,
+    VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM_KHR,
+    VK_FORMAT_G8_B8R8_2PLANE_444_UNORM_EXT,
+    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16_EXT,
+    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16_EXT,
+    VK_FORMAT_G16_B16R16_2PLANE_444_UNORM_EXT,
+    VK_FORMAT_A4R4G4B4_UNORM_PACK16_EXT,
+    VK_FORMAT_A4B4G4R4_UNORM_PACK16_EXT,
+    VK_FORMAT_R16G16_S10_5_NV,
+};
+
+constexpr VkFormat EXT_VK_FORMATS[] = {
+   // VK_EXT_ycbcr_2plane_444_formats
+   VK_FORMAT_G16_B16R16_2PLANE_444_UNORM,
+   VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16,
+   VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16,
+   VK_FORMAT_G8_B8R8_2PLANE_444_UNORM,
+   VK_FORMAT_G16_B16R16_2PLANE_444_UNORM,
+
+   // VK_NV_optical_flow
+   VK_FORMAT_R16G16_S10_5_NV,
 };
 
 using UniformBufferBitmask = utils::bitset64;

filament/backend/src/vulkan/utils/Helper.h

@@ -21,6 +21,7 @@
 
 #include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
+#include <utils/Panic.h>
 
 #include <bluevk/BlueVK.h>
 

filament/backend/src/vulkan/utils/Image.cpp

@@ -42,20 +42,22 @@ getVkTransition(const VulkanLayoutTransition& transition) {
             srcStage = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
             break;
         case VulkanLayout::COLOR_ATTACHMENT:
-            srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT |
-                            VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+            srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
                             VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-                       VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+            srcStage = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
             break;
-        case VulkanLayout::READ_WRITE:
+        case VulkanLayout::STAGING:
             srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
             srcStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
             break;
-        case VulkanLayout::READ_ONLY:
+        case VulkanLayout::FRAG_READ:
             srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
             srcStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
             break;
+        case VulkanLayout::VERT_READ:
+            srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
+            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT;
+            break;
         case VulkanLayout::TRANSFER_SRC:
             srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
             srcStage = VK_PIPELINE_STAGE_TRANSFER_BIT;
@@ -69,7 +71,7 @@ getVkTransition(const VulkanLayoutTransition& transition) {
             srcStage = VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
             break;
         case VulkanLayout::DEPTH_SAMPLER:
-            srcAccessMask = VK_ACCESS_MEMORY_READ_BIT;
+            srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
             srcStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
             break;
         case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:
@@ -84,19 +86,21 @@ getVkTransition(const VulkanLayoutTransition& transition) {
 
     switch (transition.newLayout) {
         case VulkanLayout::COLOR_ATTACHMENT:
-            dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT
-                            | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
+            dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
                             | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-            dstStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT
-                       | VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+            dstStage = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
             break;
-        case VulkanLayout::READ_WRITE:
+        case VulkanLayout::STAGING:
             dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
             dstStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
             break;
-        case VulkanLayout::READ_ONLY:
+        case VulkanLayout::FRAG_READ:
             dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
-            dstStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT;
+            dstStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
+            break;
+        case VulkanLayout::VERT_READ:
+            dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
+            dstStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT;
             break;
         case VulkanLayout::TRANSFER_SRC:
             dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
@@ -112,10 +116,8 @@ getVkTransition(const VulkanLayoutTransition& transition) {
             dstStage = VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
             break;
         case VulkanLayout::DEPTH_SAMPLER:
-            dstAccessMask =
-                    VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
-            dstStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-                       VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
+            dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
+            dstStage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
             break;
         case VulkanLayout::PRESENT:
         case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:
@@ -138,7 +140,6 @@ bool transitionLayout(VkCommandBuffer cmdbuffer,
     }
     auto [srcAccessMask, dstAccessMask, srcStage, dstStage, oldLayout, newLayout]
             = getVkTransition(transition);
-
     if (oldLayout == newLayout) {
         return false;
     }
@@ -155,7 +156,7 @@ bool transitionLayout(VkCommandBuffer cmdbuffer,
             .image = transition.image,
             .subresourceRange = transition.subresources,
     };
-    vkCmdPipelineBarrier(cmdbuffer, srcStage, dstStage, 0, 0, nullptr, 0, nullptr, 1, &barrier);
+   vkCmdPipelineBarrier(cmdbuffer, srcStage, dstStage, 0, 0, nullptr, 0, nullptr, 1, &barrier);
     return true;
 }
 
@@ -264,8 +265,9 @@ utils::io::ostream& operator<<(utils::io::ostream& out,
         const filament::backend::VulkanLayout& layout) {
     switch (layout) {
         CASE(UNDEFINED)
-        CASE(READ_ONLY)
-        CASE(READ_WRITE)
+        CASE(FRAG_READ)
+        CASE(VERT_READ)
+        CASE(STAGING)
         CASE(TRANSFER_SRC)
         CASE(TRANSFER_DST)
         CASE(DEPTH_ATTACHMENT)

filament/backend/src/vulkan/utils/Image.h

@@ -32,9 +32,11 @@ enum class VulkanLayout : uint8_t {
     // any transition.
     UNDEFINED,
     // Fragment/vertex shader accessible layout for reading and writing.
-    READ_WRITE,
-    // Fragment/vertex shader accessible layout for reading only.
-    READ_ONLY,
+    STAGING,
+    // Fragment shader accessible layout for reading only.
+    FRAG_READ,
+    // Vertex shader accessible layout for reading only.
+    VERT_READ,
     // For the source of a copy operation.
     TRANSFER_SRC,
     // For the destination of a copy operation.
@@ -65,9 +67,10 @@ constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
     switch (layout) {
         case VulkanLayout::UNDEFINED:
             return VK_IMAGE_LAYOUT_UNDEFINED;
-        case VulkanLayout::READ_WRITE:
+        case VulkanLayout::STAGING:
             return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::READ_ONLY:
+        case VulkanLayout::FRAG_READ:
+        case VulkanLayout::VERT_READ:
             return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
         case VulkanLayout::TRANSFER_SRC:
             return VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
@@ -80,9 +83,8 @@ constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
         case VulkanLayout::PRESENT:
             return VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
         // Filament sometimes samples from one miplevel while writing to another level in the
-        // same texture (e.g. bloom does this). Moreover we'd like to avoid lots of expensive
-        // layout transitions. So, keep it simple and use GENERAL for all color-attachable
-        // textures.
+        // same texture (e.g. bloom does this). So, keep it simple and use GENERAL for all
+        // color-attachable textures.
         case VulkanLayout::COLOR_ATTACHMENT:
             return VK_IMAGE_LAYOUT_GENERAL;
         case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:

filament/backend/src/vulkan/utils/StaticVector.h

@@ -20,6 +20,7 @@
 // An Array that will be statically fixed in capacity, but the "size" (as in user added elements) is
 // variable. Note that this class is movable.
 
+#include <utils/debug.h>
 #include <utils/Panic.h>
 
 #include <array>

filament/backend/src/webgpu/SpdMipmapGenerator.cpp

@@ -0,0 +1,795 @@
+#include "SpdMipmapGenerator.h"
+#include <sstream>
+#include <stdexcept>
+// C++ port of https://github.com/JolifantoBambla/webgpu-spd for early experiments- do not merge like this
+namespace spd {
+
+    // Helper to convert enums to strings for map keys or shader code
+    const char* to_string(wgpu::TextureFormat format) {
+        switch (format) {
+            case wgpu::TextureFormat::RGBA8Unorm: return "rgba8unorm";
+            case wgpu::TextureFormat::BGRA8Unorm: return "bgra8unorm";
+            case wgpu::TextureFormat::R32Float: return "r32float";
+            case wgpu::TextureFormat::RG32Float: return "rg32float";
+            case wgpu::TextureFormat::RGBA32Float: return "rgba32float";
+            case wgpu::TextureFormat::R16Float: return "r16float";
+            case wgpu::TextureFormat::RG16Float: return "rg16float";
+            case wgpu::TextureFormat::RGBA16Float: return "rgba16float";
+            // Add other formats as needed
+            default: return "rgba8unorm";
+        }
+    }
+
+
+    std::string MipmapGenerator::GetFilterCode(SPDFilter filter) {
+        switch (filter) {
+            case SPDFilter::Min:
+                return R"(
+    fn spd_reduce_4(v0: vec4<SPDScalar>, v1: vec4<SPDScalar>, v2: vec4<SPDScalar>, v3: vec4<SPDScalar>) -> vec4<SPDScalar> {
+        return min(min(v0, v1), min(v2, v3));
+    }
+    )";
+            case SPDFilter::Max:
+                return R"(
+    fn spd_reduce_4(v0: vec4<SPDScalar>, v1: vec4<SPDScalar>, v2: vec4<SPDScalar>, v3: vec4<SPDScalar>) -> vec4<SPDScalar> {
+        return max(max(v0, v1), max(v2, v3));
+    }
+    )";
+            case SPDFilter::MinMax:
+                 return R"(
+    fn spd_reduce_4(v0: vec4<SPDScalar>, v1: vec4<SPDScalar>, v2: vec4<SPDScalar>, v3: vec4<SPDScalar>) -> vec4<SPDScalar> {
+        let max4 = max(max(v0.xy, v1.xy), max(v2.xy, v3.xy));
+        return vec4<SPDScalar>(min(min(v0.x, v1.x), min(v2.x, v3.x)), max(max4.x, max4.y), 0.0, 0.0);
+    }
+    )";
+            case SPDFilter::Average:
+            default:
+                return R"(
+    fn spd_reduce_4(v0: vec4<SPDScalar>, v1: vec4<SPDScalar>, v2: vec4<SPDScalar>, v3: vec4<SPDScalar>) -> vec4<SPDScalar> {
+        return (v0 + v1 + v2 + v3) * 0.25;
+    }
+    )";
+        }
+    }
+
+
+    MipmapGenerator::MipmapGenerator(const wgpu::Device& device) : m_device(device) {
+        wgpu::BindGroupLayoutEntry bglEntry{};
+        bglEntry.binding = 0;
+        bglEntry.visibility = wgpu::ShaderStage::Compute;
+        bglEntry.buffer.type = wgpu::BufferBindingType::Uniform;
+        bglEntry.buffer.minBindingSize = 16;
+
+        wgpu::BindGroupLayoutDescriptor bglDesc{};
+        bglDesc.entryCount = 1;
+        bglDesc.entries = &bglEntry;
+        m_internalResourcesBindGroupLayout = m_device.CreateBindGroupLayout(&bglDesc);
+    }
+
+    void MipmapGenerator::PreparePipelines(wgpu::TextureFormat format, SPDFilter filter, bool halfPrecision) {
+        SPDScalarType scalarType = SanitizeScalarType(format, halfPrecision);
+        // Prepare for a reasonable number of mips
+        for (uint32_t i = 1; i <= 4; ++i) {
+            GetOrCreatePipeline(format, filter, i, scalarType);
+        }
+    }
+    
+    SPDScalarType MipmapGenerator::SanitizeScalarType(wgpu::TextureFormat format, bool halfPrecision) {
+        std::string formatStr = to_string(format);
+        std::transform(formatStr.begin(), formatStr.end(), formatStr.begin(), ::tolower);
+
+        SPDScalarType texelType = SPDScalarType::F32;
+        if (formatStr.find("sint") != std::string::npos) {
+            texelType = SPDScalarType::I32;
+        } else if (formatStr.find("uint") != std::string::npos) {
+            texelType = SPDScalarType::U32;
+        }
+
+        if (halfPrecision) {
+            bool hasF16 = false;
+            // In a real Dawn app, you would check device.GetSupportedFeatures()
+            // For now, let's assume it's available if requested.
+             hasF16 = true; 
+
+            if (!hasF16) {
+                // Log warning: half precision requested but not supported
+            }
+            if (texelType != SPDScalarType::F32) {
+                // Log warning: half precision for non-float format
+            }
+            if (hasF16 && texelType == SPDScalarType::F32) {
+                return SPDScalarType::F16;
+            }
+        }
+        return texelType;
+    }
+
+    SPDPipeline& MipmapGenerator::GetOrCreatePipeline(wgpu::TextureFormat format, SPDFilter filter, uint32_t numMips, SPDScalarType scalarType) {
+        if (m_pipelines[format][scalarType][filter].count(numMips) == 0) {
+            // Create the pipeline
+            SPDPipeline spdPipeline;
+
+            std::vector<wgpu::BindGroupLayoutEntry> mipsBglEntries;
+            for (uint32_t i = 0; i <= numMips; ++i) {
+                wgpu::BindGroupLayoutEntry entry{};
+                entry.binding = i;
+                entry.visibility = wgpu::ShaderStage::Compute;
+                if (i == 0) {
+                    entry.texture.sampleType = (scalarType == SPDScalarType::I32) ? wgpu::TextureSampleType::Sint :
+                                               (scalarType == SPDScalarType::U32) ? wgpu::TextureSampleType::Uint :
+                                               wgpu::TextureSampleType::UnfilterableFloat;
+                    entry.texture.viewDimension = wgpu::TextureViewDimension::e2DArray;
+                } else {
+                    entry.storageTexture.access = wgpu::StorageTextureAccess::WriteOnly;
+                    entry.storageTexture.format = format;
+                    entry.storageTexture.viewDimension = wgpu::TextureViewDimension::e2DArray;
+                }
+                mipsBglEntries.push_back(entry);
+            }
+            
+            wgpu::BindGroupLayoutDescriptor mipsBglDesc{};
+            mipsBglDesc.entryCount = mipsBglEntries.size();
+            mipsBglDesc.entries = mipsBglEntries.data();
+            spdPipeline.mipsBindGroupLayout = m_device.CreateBindGroupLayout(&mipsBglDesc);
+
+            std::string shaderCode = MakeShaderCode(format, GetFilterCode(filter), numMips, scalarType);
+            wgpu::ShaderModuleWGSLDescriptor wgslDesc{};
+            wgslDesc.code = shaderCode.c_str();
+
+            wgpu::ShaderModuleDescriptor shaderModuleDesc{};
+            shaderModuleDesc.nextInChain = &wgslDesc;
+            wgpu::ShaderModule shaderModule = m_device.CreateShaderModule(&shaderModuleDesc);
+            
+            wgpu::BindGroupLayout bgls[] = { spdPipeline.mipsBindGroupLayout, m_internalResourcesBindGroupLayout };
+            wgpu::PipelineLayoutDescriptor layoutDesc{};
+            layoutDesc.bindGroupLayoutCount = 2;
+            layoutDesc.bindGroupLayouts = bgls;
+
+            wgpu::ComputePipelineDescriptor pipelineDesc{};
+            pipelineDesc.layout = m_device.CreatePipelineLayout(&layoutDesc);
+            pipelineDesc.compute.module = shaderModule;
+            pipelineDesc.compute.entryPoint = "downsample";
+
+            spdPipeline.pipeline = m_device.CreateComputePipeline(&pipelineDesc);
+            m_pipelines[format][scalarType][filter][numMips] = std::move(spdPipeline);
+        }
+        return m_pipelines[format][scalarType][filter][numMips];
+    }
+    
+    void MipmapGenerator::Generate(
+        wgpu::CommandEncoder& commandEncoder,
+        wgpu::Texture srcTexture,
+        const SPDPassConfig& config)
+    {
+        uint32_t width = srcTexture.GetWidth();
+        uint32_t height = srcTexture.GetHeight();
+        uint32_t arrayLayerCount = srcTexture.GetDepthOrArrayLayers();
+
+        wgpu::Texture target = config.targetTexture ? config.targetTexture : srcTexture;
+        uint32_t numMips = config.numMips > 0 ? config.numMips : target.GetMipLevelCount() - 1;
+
+        if (numMips == 0) return;
+
+        SPDScalarType scalarType = SanitizeScalarType(srcTexture.GetFormat(), config.halfPrecision);
+        SPDPipeline& spdPipeline = GetOrCreatePipeline(target.GetFormat(), config.filter, numMips, scalarType);
+        
+        // --- Create Bind Group 0 (Mips) ---
+        std::vector<wgpu::BindGroupEntry> mipEntries;
+
+        wgpu::TextureViewDescriptor srcViewDesc{};
+        srcViewDesc.dimension = wgpu::TextureViewDimension::e2DArray;
+        srcViewDesc.baseMipLevel = config.sourceMipLevel;
+        srcViewDesc.mipLevelCount = 1;
+        srcViewDesc.baseArrayLayer = 0;
+        srcViewDesc.arrayLayerCount = arrayLayerCount;
+        
+        wgpu::BindGroupEntry srcEntry{};
+        srcEntry.binding = 0;
+        srcEntry.textureView = srcTexture.CreateView(&srcViewDesc);
+        mipEntries.push_back(srcEntry);
+
+        for (uint32_t i = 0; i < numMips; ++i) {
+            wgpu::TextureViewDescriptor dstViewDesc{};
+            dstViewDesc.dimension = wgpu::TextureViewDimension::e2DArray;
+            dstViewDesc.baseMipLevel = config.sourceMipLevel + i + 1;
+            dstViewDesc.mipLevelCount = 1;
+            dstViewDesc.baseArrayLayer = 0;
+            dstViewDesc.arrayLayerCount = arrayLayerCount;
+           
+            wgpu::BindGroupEntry dstEntry{};
+            dstEntry.binding = i + 1;
+            dstEntry.textureView = target.CreateView(&dstViewDesc);
+            mipEntries.push_back(dstEntry);
+        }
+
+        wgpu::BindGroupDescriptor mipBindGroupDesc{};
+        mipBindGroupDesc.layout = spdPipeline.mipsBindGroupLayout;
+        mipBindGroupDesc.entryCount = mipEntries.size();
+        mipBindGroupDesc.entries = mipEntries.data();
+        wgpu::BindGroup mipsBindGroup = m_device.CreateBindGroup(&mipBindGroupDesc);
+
+        // --- Create Bind Group 1 (Internal Resources) ---
+        uint32_t numWorkGroupsX = (width + 63) / 64;
+        uint32_t numWorkGroupsY = (height + 63) / 64;
+        uint32_t numWorkGroups = numWorkGroupsX * numWorkGroupsY;
+
+        struct DownsamplePassMeta {
+            uint32_t work_group_offset[2] = {0, 0};
+            uint32_t num_work_groups;
+            uint32_t mips;
+            uint32_t padding[12]; // Ensure size is multiple of 16
+        } meta;
+        meta.num_work_groups = numWorkGroups;
+        meta.mips = numMips;
+        
+        wgpu::BufferDescriptor metaBufferDesc{};
+        metaBufferDesc.size = sizeof(DownsamplePassMeta);
+        metaBufferDesc.usage = wgpu::BufferUsage::Uniform | wgpu::BufferUsage::CopyDst;
+        wgpu::Buffer metaBuffer = m_device.CreateBuffer(&metaBufferDesc);
+        m_device.GetQueue().WriteBuffer(metaBuffer, 0, &meta, sizeof(meta));
+
+        wgpu::BindGroupEntry metaEntry{};
+        metaEntry.binding = 0;
+        metaEntry.buffer = metaBuffer;
+        
+        wgpu::BindGroupDescriptor internalBindGroupDesc{};
+        internalBindGroupDesc.layout = m_internalResourcesBindGroupLayout;
+        internalBindGroupDesc.entryCount = 1;
+        internalBindGroupDesc.entries = &metaEntry;
+        wgpu::BindGroup internalBindGroup = m_device.CreateBindGroup(&internalBindGroupDesc);
+        
+        // --- Dispatch ---
+        wgpu::ComputePassEncoder pass = commandEncoder.BeginComputePass();
+        pass.SetPipeline(spdPipeline.pipeline);
+        pass.SetBindGroup(0, mipsBindGroup);
+        pass.SetBindGroup(1, internalBindGroup);
+        pass.DispatchWorkgroups(numWorkGroupsX, numWorkGroupsY, arrayLayerCount);
+        pass.End();
+    }
+
+    // Main shader generation logic
+
+// Helper function to check if a string is in a vector of strings
+bool includes(const std::vector<std::string>& vec, const std::string& str) {
+    for (const auto& s : vec) {
+        if (s == str) {
+            return true;
+        }
+    }
+    return false;
+}
+
+std::string MakeShaderCode(wgpu::TextureFormat outputFormat,
+                           const std::string& filterOp,
+                           unsigned int numMips, // Assuming a default value for numMips
+                           SPDScalarType scalarType) { // Default scalarType
+    std::stringstream ss;
+
+    std::string texelType;
+    if (scalarType == SPDScalarType::I32) {
+        texelType = "i32";
+    } else if (scalarType == SPDScalarType::U32) {
+        texelType = "u32";
+    } else {
+        texelType = "f32";
+    }
+
+    bool useF16 = (scalarType == SPDScalarType::F16);
+
+    std::string filterCode = filterOp;
+    if (filterOp == SPD_FILTER_AVERAGE && !includes({"f32", "f16"}, texelType)) {
+        // Replace "* 0.25" with "/ 4"
+        size_t pos = filterCode.find("* 0.25");
+        if (pos != std::string::npos) {
+            filterCode.replace(pos, std::string("* 0.25").length(), "/ 4");
+        }
+    }
+
+    // Generate mipsBindings
+    std::string mipsBindings;
+    for (unsigned int i = 0; i < numMips; ++i) {
+        mipsBindings += "@group(0) @binding(" + std::to_string(i + 1) + ") var dst_mip_" + std::to_string(i + 1) + ": texture_storage_2d_array<" + to_string(outputFormat)  + ", write>;\n";
+    }
+
+    // Generate mipsAccessorBody
+    std::string mipsAccessorBody;
+    for (unsigned int i = 0; i < numMips; ++i) {
+        if (i == 5 && numMips > 6) {
+            mipsAccessorBody += " else if mip == 6 {\n";
+            mipsAccessorBody += "                textureStore(dst_mip_6, uv, slice, " + (useF16 ? "vec4<" + texelType + ">(value)" : "value") + ");\n";
+            mipsAccessorBody += "                mip_dst_6_buffer[slice][uv.y][uv.x] = value;\n";
+            mipsAccessorBody += "            }";
+        } else {
+            if (i != 0) {
+                mipsAccessorBody += " else ";
+            }
+            mipsAccessorBody += "if mip == " + std::to_string(i + 1) + " {\n";
+            mipsAccessorBody += "                textureStore(dst_mip_" + std::to_string(i + 1) + ", uv, slice, " + (useF16 ? "vec4<" + texelType + ">(value)" : "value") + ");\n";
+            mipsAccessorBody += "            }";
+        }
+    }
+
+    std::string mipsAccessor = "fn store_dst_mip(value: vec4<SPDScalar>, uv: vec2<u32>, slice: u32, mip: u32) {\n" + mipsAccessorBody + "\n}";
+    std::string midMipAccessor = "return mip_dst_6_buffer[slice][uv.y][uv.x];";
+
+    // Start building the final shader code string
+    ss << R"(
+// This file is part of the FidelityFX SDK.
+//
+// Copyright (C) 2023 Advanced Micro Devices, Inc.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a copy
+// of this software and associated documentation files(the “Software”), to deal
+// in the Software without restriction, including without limitation the rights
+// to use, copy, modify, merge, publish, distribute, sublicense, and /or sell
+// copies of the Software, and to permit persons to whom the Software is
+// furnished to do so, subject to the following conditions :
+//
+// The above copyright notice and this permission notice shall be included in
+// all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE
+// AUTHORS COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+// THE SOFTWARE.
+
+
+// Definitions --------------------------------------------------------------------------------------------------------
+
+)";
+    if (useF16) {
+        ss << "enable f16;\n";
+    }
+    ss << "alias SPDScalar = " << texelType << ";\n\n"; // Using texelType here, assuming SPDScalar maps to it.
+
+    ss << R"(
+// Helpers ------------------------------------------------------------------------------------------------------------
+
+/**
+ * A helper function performing a remap 64x1 to 8x8 remapping which is necessary for 2D wave reductions.
+ * * The 64-wide lane indices to 8x8 remapping is performed as follows:
+ * 00 01 08 09 10 11 18 19
+ * 02 03 0a 0b 12 13 1a 1b
+ * 04 05 0c 0d 14 15 1c 1d
+ * 06 07 0e 0f 16 17 1e 1f
+ * 20 21 28 29 30 31 38 39
+ * 22 23 2a 2b 32 33 3a 3b
+ * 24 25 2c 2d 34 35 3c 3d
+ * 26 27 2e 2f 36 37 3e 3f
+ * * @param a: The input 1D coordinate to remap.
+ *
+ * @returns The remapped 2D coordinates.
+ */
+fn remap_for_wave_reduction(a: u32) -> vec2<u32> {
+    return vec2<u32>(
+        insertBits(extractBits(a, 2u, 3u), a, 0u, 1u),
+        insertBits(extractBits(a, 3u, 3u), extractBits(a, 1u, 2u), 0u, 2u)
+    );
+}
+
+fn map_to_xy(local_invocation_index: u32) -> vec2<u32> {
+    let sub_xy: vec2<u32> = remap_for_wave_reduction(local_invocation_index % 64);
+    return vec2<u32>(
+        sub_xy.x + 8 * ((local_invocation_index >> 6) % 2),
+        sub_xy.y + 8 * ((local_invocation_index >> 7))
+    );
+}
+
+/*
+ * Compute a linear value from a SRGB value.
+ * * @param value: The value to convert to linear from SRGB.
+ * * @returns A value in SRGB space.
+ */
+/*
+fn srgb_to_linear(value: SPDScalar) -> SPDScalar {
+    let j = vec3<SPDScalar>(0.0031308 * 12.92, 12.92, 1.0 / 2.4);
+    let k = vec2<SPDScalar>(1.055, -0.055);
+    return clamp(j.x, value * j.y, pow(value, j.z) * k.x + k.y);
+}
+*/
+
+// Resources & Accessors -----------------------------------------------------------------------------------------------
+struct DownsamplePassMeta {
+    work_group_offset: vec2<u32>,
+    num_work_groups: u32,
+    mips: u32,
+}
+
+// In the original version dst_mip_i is an image2Darray [SPD_MAX_MIP_LEVELS+1], i.e., 12+1, but WGSL doesn't support arrays of textures yet
+// Also these are read_write because for mips 7-13, the workgroup reads from mip level 6 - since most formats don't support read_write access in WGSL yet, we use a single read_write buffer in such cases instead
+@group(0) @binding(0) var src_mip_0: texture_2d_array<)" << texelType << R"(>;
+)" << mipsBindings << R"(
+
+@group(1) @binding(0) var<uniform> downsample_pass_meta : DownsamplePassMeta;
+@group(1) @binding(1) var<storage, read_write> spd_global_counter: array<atomic<u32>>;
+@group(1) @binding(2) var<storage, read_write> mip_dst_6_buffer: array<array<array<vec4<f32>, 64>, 64>>;
+
+fn get_mips() -> u32 {
+    return downsample_pass_meta.mips;
+}
+
+fn get_num_work_groups() -> u32 {
+    return downsample_pass_meta.num_work_groups;
+}
+
+fn get_work_group_offset() -> vec2<u32> {
+    return downsample_pass_meta.work_group_offset;
+}
+
+fn load_src_image(uv: vec2<u32>, slice: u32) -> vec4<SPDScalar> {
+    return vec4<SPDScalar>(textureLoad(src_mip_0, uv, slice, 0));
+}
+
+fn load_mid_mip_image(uv: vec2<u32>, slice: u32) -> vec4<SPDScalar> {
+    )";
+    if (numMips > 6) {
+        ss << midMipAccessor;
+    } else {
+        ss << "return vec4<SPDScalar>();";
+    }
+    ss << R"(
+}
+
+)" << mipsAccessor << R"(
+
+// Workgroup -----------------------------------------------------------------------------------------------------------
+
+var<workgroup> spd_intermediate: array<array<vec4<SPDScalar>, 16>, 16>;
+var<workgroup> spd_counter: atomic<u32>;
+
+fn spd_increase_atomic_counter(slice: u32) {
+    atomicStore(&spd_counter, atomicAdd(&spd_global_counter[slice], 1));
+}
+
+fn spd_get_atomic_counter() -> u32 {
+    return atomicLoad(&spd_counter);
+}
+
+fn spd_reset_atomic_counter(slice: u32) {
+    atomicStore(&spd_global_counter[slice], 0);
+}
+
+// Cotnrol flow --------------------------------------------------------------------------------------------------------
+
+fn spd_barrier() {
+    // in glsl this does: groupMemoryBarrier(); barrier();
+    workgroupBarrier();
+}
+
+// Only last active workgroup should proceed
+fn spd_exit_workgroup(num_work_groups: u32, local_invocation_index: u32, slice: u32) -> bool {
+    // global atomic counter
+    if (local_invocation_index == 0) {
+        spd_increase_atomic_counter(slice);
+    }
+    spd_barrier();
+    return spd_get_atomic_counter() != (num_work_groups - 1);
+}
+
+// Pixel access --------------------------------------------------------------------------------------------------------
+
+)" << filterCode << R"(
+
+fn spd_store(pix: vec2<u32>, out_value: vec4<SPDScalar>, mip: u32, slice: u32) {
+    store_dst_mip(out_value, pix, slice, mip + 1);
+}
+
+fn spd_load_intermediate(x: u32, y: u32) -> vec4<SPDScalar> {
+    return spd_intermediate[x][y];
+}
+
+fn spd_store_intermediate(x: u32, y: u32, value: vec4<SPDScalar>) {
+    spd_intermediate[x][y] = value;
+}
+
+fn spd_reduce_intermediate(i0: vec2<u32>, i1: vec2<u32>, i2: vec2<u32>, i3: vec2<u32>) -> vec4<SPDScalar> {
+    let v0 = spd_load_intermediate(i0.x, i0.y);
+    let v1 = spd_load_intermediate(i1.x, i1.y);
+    let v2 = spd_load_intermediate(i2.x, i2.y);
+    let v3 = spd_load_intermediate(i3.x, i3.y);
+    return spd_reduce_4(v0, v1, v2, v3);
+}
+
+fn spd_reduce_load_4(base: vec2<u32>, slice: u32) -> vec4<SPDScalar> {
+    let v0 = load_src_image(base + vec2<u32>(0, 0), slice);
+    let v1 = load_src_image(base + vec2<u32>(0, 1), slice);
+    let v2 = load_src_image(base + vec2<u32>(1, 0), slice);
+    let v3 = load_src_image(base + vec2<u32>(1, 1), slice);
+    return spd_reduce_4(v0, v1, v2, v3);
+}
+
+fn spd_reduce_load_mid_mip_4(base: vec2<u32>, slice: u32) -> vec4<SPDScalar> {
+    let v0 = load_mid_mip_image(base + vec2<u32>(0, 0), slice);
+    let v1 = load_mid_mip_image(base + vec2<u32>(0, 1), slice);
+    let v2 = load_mid_mip_image(base + vec2<u32>(1, 0), slice);
+    let v3 = load_mid_mip_image(base + vec2<u32>(1, 1), slice);
+    return spd_reduce_4(v0, v1, v2, v3);
+}
+
+// Main logic ---------------------------------------------------------------------------------------------------------
+
+fn spd_downsample_mips_0_1(x: u32, y: u32, workgroup_id: vec2<u32>, local_invocation_index: u32, mip: u32, slice: u32) {
+    var v: array<vec4<SPDScalar>, 4>;
+
+    let workgroup64 = workgroup_id.xy * 64;
+    let workgroup32 = workgroup_id.xy * 32;
+    let workgroup16 = workgroup_id.xy * 16;
+
+    var tex = workgroup64 + vec2<u32>(x * 2, y * 2);
+    var pix = workgroup32 + vec2<u32>(x, y);
+    v[0] = spd_reduce_load_4(tex, slice);
+    spd_store(pix, v[0], 0, slice);
+
+    tex = workgroup64 + vec2<u32>(x * 2 + 32, y * 2);
+    pix = workgroup32 + vec2<u32>(x + 16, y);
+    v[1] = spd_reduce_load_4(tex, slice);
+    spd_store(pix, v[1], 0, slice);
+
+    tex = workgroup64 + vec2<u32>(x * 2, y * 2 + 32);
+    pix = workgroup32 + vec2<u32>(x, y + 16);
+    v[2] = spd_reduce_load_4(tex, slice);
+    spd_store(pix, v[2], 0, slice);
+
+    tex = workgroup64 + vec2<u32>(x * 2 + 32, y * 2 + 32);
+    pix = workgroup32 + vec2<u32>(x + 16, y + 16);
+    v[3] = spd_reduce_load_4(tex, slice);
+    spd_store(pix, v[3], 0, slice);
+
+    if mip <= 1 {
+        return;
+    }
+
+    for (var i = 0u; i < 4u; i++) {
+        spd_store_intermediate(x, y, v[i]);
+        spd_barrier();
+        if local_invocation_index < 64 {
+            v[i] = spd_reduce_intermediate(
+                vec2<u32>(x * 2 + 0, y * 2 + 0),
+                vec2<u32>(x * 2 + 1, y * 2 + 0),
+                vec2<u32>(x * 2 + 0, y * 2 + 1),
+                vec2<u32>(x * 2 + 1, y * 2 + 1)
+            );
+            spd_store(workgroup16 + vec2<u32>(x + (i % 2) * 8, y + (i / 2) * 8), v[i], 1, slice);
+        }
+        spd_barrier();
+    }
+
+    if local_invocation_index < 64 {
+        spd_store_intermediate(x + 0, y + 0, v[0]);
+        spd_store_intermediate(x + 8, y + 0, v[1]);
+        spd_store_intermediate(x + 0, y + 8, v[2]);
+        spd_store_intermediate(x + 8, y + 8, v[3]);
+    }
+}
+
+fn spd_downsample_mip_2(x: u32, y: u32, workgroup_id: vec2<u32>, local_invocation_index: u32, mip: u32, slice: u32) {
+    if local_invocation_index < 64u {
+        let v = spd_reduce_intermediate(
+            vec2<u32>(x * 2 + 0, y * 2 + 0),
+            vec2<u32>(x * 2 + 1, y * 2 + 0),
+            vec2<u32>(x * 2 + 0, y * 2 + 1),
+            vec2<u32>(x * 2 + 1, y * 2 + 1)
+        );
+        spd_store(workgroup_id.xy * 8 + vec2<u32>(x, y), v, mip, slice);
+        // store to LDS, try to reduce bank conflicts
+        // x 0 x 0 x 0 x 0 x 0 x 0 x 0 x 0
+        // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
+        // 0 x 0 x 0 x 0 x 0 x 0 x 0 x 0 x
+        // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
+        // x 0 x 0 x 0 x 0 x 0 x 0 x 0 x 0
+        // ...
+        // x 0 x 0 x 0 x 0 x 0 x 0 x 0 x 0
+        spd_store_intermediate(x * 2 + y % 2, y * 2, v);
+    }
+}
+
+fn spd_downsample_mip_3(x: u32, y: u32, workgroup_id: vec2<u32>, local_invocation_index: u32, mip: u32, slice: u32) {
+    if local_invocation_index < 16u {
+        // x 0 x 0
+        // 0 0 0 0
+        // 0 x 0 x
+        // 0 0 0 0
+        let v = spd_reduce_intermediate(
+            vec2<u32>(x * 4 + 0 + 0, y * 4 + 0),
+            vec2<u32>(x * 4 + 2 + 0, y * 4 + 0),
+            vec2<u32>(x * 4 + 0 + 1, y * 4 + 2),
+            vec2<u32>(x * 4 + 2 + 1, y * 4 + 2)
+        );
+        spd_store(workgroup_id.xy * 4 + vec2<u32>(x, y), v, mip, slice);
+        // store to LDS
+        // x 0 0 0 x 0 0 0 x 0 0 0 x 0 0 0
+        // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
+        // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
+        // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
+        // 0 x 0 0 0 x 0 0 0 x 0 0 0 x 0 0
+        // ...
+        // 0 0 x 0 0 0 x 0 0 0 x 0 0 0 x 0
+        // ...
+        // 0 0 0 x 0 0 0 x 0 0 0 x 0 0 0 x
+        // ...
+        spd_store_intermediate(x * 4 + y, y * 4, v);
+    }
+}
+
+fn spd_downsample_mip_4(x: u32, y: u32, workgroup_id: vec2<u32>, local_invocation_index: u32, mip: u32, slice: u32) {
+    if local_invocation_index < 4u {
+        // x 0 0 0 x 0 0 0
+        // ...
+        // 0 x 0 0 0 x 0 0
+        let v = spd_reduce_intermediate(
+            vec2<u32>(x * 8 + 0 + 0 + y * 2, y * 8 + 0),
+            vec2<u32>(x * 8 + 4 + 0 + y * 2, y * 8 + 0),
+            vec2<u32>(x * 8 + 0 + 1 + y * 2, y * 8 + 4),
+            vec2<u32>(x * 8 + 4 + 1 + y * 2, y * 8 + 4)
+        );
+        spd_store(workgroup_id.xy * 2 + vec2<u32>(x, y), v, mip, slice);
+        // store to LDS
+        // x x x x 0 ...
+        // 0 ...
+        spd_store_intermediate(x + y * 2, 0, v);
+    }
+}
+
+fn spd_downsample_mip_5(workgroup_id: vec2<u32>, local_invocation_index: u32, mip: u32, slice: u32) {
+    if local_invocation_index < 1u {
+        // x x x x 0 ...
+        // 0 ...
+        let v = spd_reduce_intermediate(vec2<u32>(0, 0), vec2<u32>(1, 0), vec2<u32>(2, 0), vec2<u32>(3, 0));
+        spd_store(workgroup_id.xy, v, mip, slice);
+    }
+}
+
+fn spd_downsample_next_four(x: u32, y: u32, workgroup_id: vec2<u32>, local_invocation_index: u32, base_mip: u32, mips: u32, slice: u32) {
+    if mips <= base_mip {
+        return;
+    }
+    spd_barrier();
+    spd_downsample_mip_2(x, y, workgroup_id, local_invocation_index, base_mip, slice);
+
+    if mips <= base_mip + 1 {
+        return;
+    }
+    spd_barrier();
+    spd_downsample_mip_3(x, y, workgroup_id, local_invocation_index, base_mip + 1, slice);
+
+    if mips <= base_mip + 2 {
+        return;
+    }
+    spd_barrier();
+    spd_downsample_mip_4(x, y, workgroup_id, local_invocation_index, base_mip + 2, slice);
+
+    if mips <= base_mip + 3 {
+        return;
+    }
+    spd_barrier();
+    spd_downsample_mip_5(workgroup_id, local_invocation_index, base_mip + 3, slice);
+}
+
+fn spd_downsample_last_four(x: u32, y: u32, workgroup_id: vec2<u32>, local_invocation_index: u32, base_mip: u32, mips: u32, slice: u32, exit: bool) {
+    if mips <= base_mip {
+        return;
+    }
+    spd_barrier();
+    if !exit {
+        spd_downsample_mip_2(x, y, workgroup_id, local_invocation_index, base_mip, slice);
+    }
+
+    if mips <= base_mip + 1 {
+        return;
+    }
+    spd_barrier();
+    if !exit {
+        spd_downsample_mip_3(x, y, workgroup_id, local_invocation_index, base_mip + 1, slice);
+    }
+
+    if mips <= base_mip + 2 {
+        return;
+    }
+    spd_barrier();
+    if !exit {
+        spd_downsample_mip_4(x, y, workgroup_id, local_invocation_index, base_mip + 2, slice);
+    }
+
+    if mips <= base_mip + 3 {
+        return;
+    }
+    spd_barrier();
+    if !exit {
+        spd_downsample_mip_5(workgroup_id, local_invocation_index, base_mip + 3, slice);
+    }
+}
+
+fn spd_downsample_mips_6_7(x: u32, y: u32, mips: u32, slice: u32) {
+    var tex = vec2<u32>(x * 4 + 0, y * 4 + 0);
+    var pix = vec2<u32>(x * 2 + 0, y * 2 + 0);
+    let v0 = spd_reduce_load_mid_mip_4(tex, slice);
+    spd_store(pix, v0, 6, slice);
+
+    tex = vec2<u32>(x * 4 + 2, y * 4 + 0);
+    pix = vec2<u32>(x * 2 + 1, y * 2 + 0);
+    let v1 = spd_reduce_load_mid_mip_4(tex, slice);
+    spd_store(pix, v1, 6, slice);
+
+    tex = vec2<u32>(x * 4 + 0, y * 4 + 2);
+    pix = vec2<u32>(x * 2 + 0, y * 2 + 1);
+    let v2 = spd_reduce_load_mid_mip_4(tex, slice);
+    spd_store(pix, v2, 6, slice);
+
+    tex = vec2<u32>(x * 4 + 2, y * 4 + 2);
+    pix = vec2<u32>(x * 2 + 1, y * 2 + 1);
+    let v3 = spd_reduce_load_mid_mip_4(tex, slice);
+    spd_store(pix, v3, 6, slice);
+
+    if mips <= 7 {
+        return;
+    }
+    // no barrier needed, working on values only from the same thread
+
+    let v = spd_reduce_4(v0, v1, v2, v3);
+    spd_store(vec2<u32>(x, y), v, 7, slice);
+    spd_store_intermediate(x, y, v);
+}
+
+fn spd_downsample_last_6(x: u32, y: u32, local_invocation_index: u32, mips: u32, num_work_groups: u32, slice: u32) {
+    if mips <= 6 {
+        return;
+    }
+
+    // increase the global atomic counter for the given slice and check if it's the last remaining thread group:
+    // terminate if not, continue if yes.
+    let exit = spd_exit_workgroup(num_work_groups, local_invocation_index, slice);
+
+    // can't exit directly because subsequent barrier calls break uniform control flow...
+    if !exit {
+        // reset the global atomic counter back to 0 for the next spd dispatch
+        spd_reset_atomic_counter(slice);
+
+        // After mip 5 there is only a single workgroup left that downsamples the remaining up to 64x64 texels.
+        // compute MIP level 6 and 7
+        spd_downsample_mips_6_7(x, y, mips, slice);
+    }
+
+    // compute MIP level 8, 9, 10, 11
+    spd_downsample_last_four(x, y, vec2<u32>(0, 0), local_invocation_index, 8, mips, slice, exit);
+}
+
+/// Downsamples a 64x64 tile based on the work group id.
+/// If after downsampling it's the last active thread group, computes the remaining MIP levels.
+///
+/// @param [in] workGroupID        index of the work group / thread group
+/// @param [in] localInvocationIndex   index of the thread within the thread group in 1D
+/// @param [in] mips             the number of total MIP levels to compute for the input texture
+/// @param [in] numWorkGroups        the total number of dispatched work groups / thread groups for this slice
+/// @param [in] slice             the slice of the input texture
+fn spd_downsample(workgroup_id: vec2<u32>, local_invocation_index: u32, mips: u32, num_work_groups: u32, slice: u32) {
+    let xy = map_to_xy(local_invocation_index);
+    spd_downsample_mips_0_1(xy.x, xy.y, workgroup_id, local_invocation_index, mips, slice);
+    spd_downsample_next_four(xy.x, xy.y, workgroup_id, local_invocation_index, 2, mips, slice);
+)";
+    if (numMips > 6) {
+        ss << "    spd_downsample_last_6(xy.x, xy.y, local_invocation_index, mips, num_work_groups, slice);\n";
+    }
+    ss << R"(}
+
+// Entry points -------------------------------------------------------------------------------------------------------
+
+@compute
+@workgroup_size(256, 1, 1)
+fn downsample(@builtin(local_invocation_index) local_invocation_index: u32, @builtin(workgroup_id) workgroup_id: vec3<u32>) {
+    spd_downsample(
+        workgroup_id.xy + get_work_group_offset(),
+        local_invocation_index,
+        get_mips(),
+        get_num_work_groups(),
+        workgroup_id.z
+    );
+}
+)";
+    return ss.str();
+}
+
+
+} // namespace spd

filament/backend/src/webgpu/SpdMipmapGenerator.h

@@ -0,0 +1,83 @@
+#pragma once
+
+#include <webgpu/webgpu_cpp.h>
+#include <string>
+#include <vector>
+#include <unordered_map>
+#include <optional>
+// C++ port of https://github.com/JolifantoBambla/webgpu-spd for early experiments- do not merge like this
+namespace spd {
+
+    // Enum for selecting the downsampling filter.
+    enum class SPDFilter {
+        Average,
+        Min,
+        Max,
+        MinMax
+    };
+
+    // Enum for shader scalar types.
+    enum class SPDScalarType {
+        F32,
+        F16,
+        I32,
+        U32
+    };
+
+    // Configuration for a single mipmap generation pass.
+struct SPDPassConfig {
+    SPDFilter filter = SPDFilter::Average;
+    wgpu::Texture targetTexture = nullptr;
+    uint32_t numMips = 0;
+    bool halfPrecision = false;
+    uint32_t sourceMipLevel = 0;
+};
+    
+    // Holds a pipeline and its corresponding bind group layout.
+    struct SPDPipeline {
+        wgpu::BindGroupLayout mipsBindGroupLayout = nullptr;
+        wgpu::ComputePipeline pipeline = nullptr;
+    };
+
+    // Manages pipeline creation, caching, and execution for mipmap generation.
+    class MipmapGenerator {
+    public:
+        MipmapGenerator(const wgpu::Device& device);
+
+        // Pre-creates pipelines for specified formats and filters.
+        void PreparePipelines(wgpu::TextureFormat format, SPDFilter filter, bool halfPrecision = false);
+
+        // Generates a compute pass for creating mipmaps.
+        void Generate(
+            wgpu::CommandEncoder& commandEncoder,
+            wgpu::Texture srcTexture,
+            const SPDPassConfig& config
+        );
+
+    private:
+        wgpu::Device m_device;
+        wgpu::BindGroupLayout m_internalResourcesBindGroupLayout;
+
+        // Cached pipelines: Map<TextureFormat, Map<SPDScalarType, Map<Filter, Map<NumMips, Pipeline>>>>
+        std::unordered_map<wgpu::TextureFormat,
+            std::unordered_map<SPDScalarType,
+                std::unordered_map<SPDFilter,
+                    std::unordered_map<uint32_t, SPDPipeline>>>> m_pipelines;
+        
+        // Helper methods
+        SPDPipeline& GetOrCreatePipeline(wgpu::TextureFormat format, SPDFilter filter, uint32_t numMips, SPDScalarType scalarType);
+        SPDScalarType SanitizeScalarType(wgpu::TextureFormat format, bool halfPrecision);
+        std::string GetFilterCode(SPDFilter filter);
+    };
+    // Assuming SPD_FILTER_AVERAGE is a string constant
+    const std::string SPD_FILTER_AVERAGE = "value * 0.25"; // Original filter operation
+
+    // Generates the WGSL shader code dynamically.
+    std::string MakeShaderCode(
+        wgpu::TextureFormat outputFormat,
+        const std::string& filterOp = SPD_FILTER_AVERAGE,
+        uint32_t numMips = 0,
+        SPDScalarType scalarType = SPDScalarType::F32
+    );
+
+} // namespace spd

filament/backend/src/webgpu/WGPUProgram.cpp

@@ -1,203 +0,0 @@
-/*
- * Copyright (C) 2025 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 "WebGPUHandles.h"
-
-#include "WebGPUConstants.h"
-
-#include "DriverBase.h"
-#include <backend/DriverEnums.h>
-#include <backend/Program.h>
-
-#include <utils/Panic.h>
-#include <utils/ostream.h>
-
-#include <webgpu/webgpu_cpp.h>
-
-#include <algorithm>
-#include <sstream>
-#include <string_view>
-#include <vector>
-
-namespace filament::backend {
-
-namespace {
-
-[[nodiscard]] constexpr std::string_view toString(ShaderStage stage) {
-    switch (stage) {
-        case ShaderStage::VERTEX:
-            return "vertex";
-        case ShaderStage::FRAGMENT:
-            return "fragment";
-        case ShaderStage::COMPUTE:
-            return "compute";
-    }
-}
-
-[[nodiscard]] wgpu::ShaderModule createShaderModule(wgpu::Device& device, const char* programName,
-        std::array<utils::FixedCapacityVector<uint8_t>, Program::SHADER_TYPE_COUNT> const&
-                shaderSource,
-        ShaderStage stage) {
-    utils::FixedCapacityVector<uint8_t> const& sourceBytes =
-            shaderSource[static_cast<size_t>(stage)];
-    if (sourceBytes.empty()) {
-        return nullptr;// nothing to compile, the shader was not provided
-    }
-    wgpu::ShaderModuleWGSLDescriptor wgslDescriptor{};
-    wgslDescriptor.code = wgpu::StringView(reinterpret_cast<const char*>(sourceBytes.data()));
-    std::stringstream labelStream;
-    labelStream << programName << " " << toString(stage) << " shader";
-    auto label = labelStream.str();
-    wgpu::ShaderModuleDescriptor descriptor{
-        .nextInChain = &wgslDescriptor,
-        .label = label.data()
-    };
-    wgpu::ShaderModule module = device.CreateShaderModule(&descriptor);
-    FILAMENT_CHECK_POSTCONDITION(module != nullptr) << "Failed to create " << descriptor.label;
-
-    wgpu::Instance instance = device.GetAdapter().GetInstance();
-    instance.WaitAny(
-            module.GetCompilationInfo(wgpu::CallbackMode::WaitAnyOnly,
-                    [&descriptor](auto const& status,
-                            wgpu::CompilationInfo const* info) {
-                        switch (status) {
-                            case wgpu::CompilationInfoRequestStatus::CallbackCancelled:
-                                FWGPU_LOGW << "Shader compilation info callback cancelled for "
-                                           << descriptor.label << "?" << utils::io::endl;
-                                return;
-                            case wgpu::CompilationInfoRequestStatus::Success:
-                                break;
-                        }
-                        if (info != nullptr) {
-                            std::stringstream errorStream;
-                            int errorCount = 0;
-                            for (size_t msgIndex = 0; msgIndex < info->messageCount; msgIndex++) {
-                                wgpu::CompilationMessage const& message = info->messages[msgIndex];
-                                switch (message.type) {
-                                    case wgpu::CompilationMessageType::Info:
-                                        FWGPU_LOGI << descriptor.label << ": " << message.message
-                                                   << " line#:" << message.lineNum
-                                                   << " linePos:" << message.linePos
-                                                   << " offset:" << message.offset
-                                                   << " length:" << message.length
-                                                   << utils::io::endl;
-                                        break;
-                                    case wgpu::CompilationMessageType::Warning:
-                                        FWGPU_LOGW
-                                                << "Warning compiling " << descriptor.label << ": "
-                                                << message.message << " line#:" << message.lineNum
-                                                << " linePos:" << message.linePos
-                                                << " offset:" << message.offset
-                                                << " length:" << message.length << utils::io::endl;
-                                        break;
-                                    case wgpu::CompilationMessageType::Error:
-                                        errorCount++;
-                                        errorStream << "Error " << errorCount << " : "
-                                                    << std::string_view(message.message)
-                                                    << " line#:" << message.lineNum
-                                                    << " linePos:" << message.linePos
-                                                    << " offset:" << message.offset
-                                                    << " length:" << message.length << "\n";
-                                        break;
-                                }
-                            }
-                            FILAMENT_CHECK_POSTCONDITION(errorCount < 1)
-                                    << errorCount << " error(s) compiling " << descriptor.label
-                                    << ":\n"
-                                    << errorStream.str();
-                        }
-                        FWGPU_LOGD << descriptor.label << " compiled successfully"
-                                   << utils::io::endl;
-                    }),
-            UINT16_MAX);
-    return module;
-}
-
-// This is a 1 to 1 mapping of the ReservedSpecializationConstants enum in EngineEnums.h
-// The _hack is a workaround until https://issues.chromium.org/issues/42250586 is resolved
-// This workaround is the same one being used on the generateSpecializationConstant() function
-wgpu::StringView getSpecConstantStringId(uint32_t id) {
-    switch (id) {
-        case 0:
-            return "0";// BACKEND_FEATURE_LEVEL_hack
-        case 1:
-            return "1";// CONFIG_MAX_INSTANCES_hack
-        case 2:
-            return "2";// ONFIG_STATIC_TEXTURE_TARGET_WORKAROUND_hack
-        case 3:
-            return "3";// CONFIG_SRGB_SWAPCHAIN_EMULATION_hack
-        case 4:
-            return "4";// CONFIG_FROXEL_BUFFER_HEIGHT_hack
-        case 5:
-            return "5";// CONFIG_POWER_VR_SHADER_WORKAROUNDS_hack
-        case 6:
-            return "6";// CONFIG_DEBUG_DIRECTIONAL_SHADOWMAP_hack
-        case 7:
-            return "7";// CONFIG_DEBUG_FROXEL_VISUALIZATION_hack
-        case 8:
-            return "8";// CONFIG_STEREO_EYE_COUNT_hack
-        case 9:
-            return "9";// CONFIG_SH_BANDS_COUNT_hack
-        case 10:
-            return "10";// CONFIG_SHADOW_SAMPLING_METHOD_hack
-        default:
-            PANIC_POSTCONDITION("Unknown/unhandled spec constant key/id: %d", id);
-    }
-}
-
-std::vector<wgpu::ConstantEntry> convertConstants(
-        utils::FixedCapacityVector<filament::backend::Program::SpecializationConstant> const&
-                constantsInfo) {
-    std::vector<wgpu::ConstantEntry> constants;
-    constants.reserve(constantsInfo.size());
-    for (filament::backend::Program::SpecializationConstant const& constant: constantsInfo) {
-        // CONFIG_MAX_INSTANCES (1) and CONFIG_FROXEL_BUFFER_HEIGHT (4) will not be present
-        // as constant overrides in the generated WGSL, because WGSL doesn't support specialization
-        // constants as an array length
-        // More information at https://github.com/gpuweb/gpuweb/issues/572#issuecomment-649760005
-        // CONFIG_SRGB_SWAPCHAIN_EMULATION (3) is being skipped all together since it's only
-        // included for the case of mFeatureLevel == FeatureLevel::FEATURE_LEVEL_0, which should
-        // not be possible for WebGPU
-        if (constant.id == 1 || constant.id == 3 || constant.id == 4) {
-            continue;
-        }
-        double value = 0.0;
-        if (auto* v = std::get_if<int32_t>(&constant.value)) {
-            value = static_cast<double>(*v);
-        } else if (auto* f = std::get_if<float>(&constant.value)) {
-            value = static_cast<double>(*f);
-        } else if (auto* b = std::get_if<bool>(&constant.value)) {
-            value = *b ? 0.0 : 1.0;
-        }
-        constants.push_back(
-                wgpu::ConstantEntry{ .key = getSpecConstantStringId(constant.id), .value = value });
-    }
-    return constants;
-}
-
-}// namespace
-
-WGPUProgram::WGPUProgram(wgpu::Device& device, Program& program)
-    : HwProgram(program.getName()),
-      vertexShaderModule(createShaderModule(device, name.c_str_safe(), program.getShadersSource(),
-              ShaderStage::VERTEX)),
-      fragmentShaderModule(createShaderModule(device, name.c_str_safe(), program.getShadersSource(),
-              ShaderStage::FRAGMENT)),
-      computeShaderModule(createShaderModule(device, name.c_str_safe(), program.getShadersSource(),
-              ShaderStage::COMPUTE)),
-      constants(convertConstants(program.getSpecializationConstants())) {}
-
-}// namespace filament::backend

filament/backend/src/webgpu/WebGPUBufferBase.cpp

@@ -0,0 +1,92 @@
+/*
+ * Copyright (C) 2025 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 "WebGPUBufferBase.h"
+
+#include "WebGPUConstants.h"
+
+#include "DriverBase.h"
+#include <backend/BufferDescriptor.h>
+
+#include <utils/Panic.h>
+#include <utils/debug.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <cstdint>
+#include <cstring>
+
+namespace filament::backend {
+
+namespace {
+
+[[nodiscard]] wgpu::Buffer createBuffer(wgpu::Device const& device, const wgpu::BufferUsage usage,
+        uint32_t size, const char* const label) {
+    // Write size must be divisible by WEBGPU_BUFFER_SIZE_MODULUS (e.g. 4).
+    // If the whole buffer is written to as is common, so must the buffer size.
+    size += (WEBGPU_BUFFER_SIZE_MODULUS - (size % WEBGPU_BUFFER_SIZE_MODULUS)) %
+            WEBGPU_BUFFER_SIZE_MODULUS;
+    wgpu::BufferDescriptor descriptor{
+        .label = label,
+        .usage = usage,
+        .size = size,
+        .mappedAtCreation = false };
+    wgpu::Buffer buffer = device.CreateBuffer(&descriptor);
+    FILAMENT_CHECK_POSTCONDITION(buffer) << "Failed to create buffer for " << label;
+    return buffer;
+}
+
+} // namespace
+
+WebGPUBufferBase::WebGPUBufferBase(wgpu::Device const& device, const wgpu::BufferUsage usage,
+        const uint32_t size, char const* const label)
+    : mBuffer{ createBuffer(device, usage, size, label) } {}
+
+void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
+        const uint32_t byteOffset, wgpu::Queue const& queue) {
+    FILAMENT_CHECK_PRECONDITION(bufferDescriptor.buffer)
+            << "updateGPUBuffer called with a null buffer";
+    FILAMENT_CHECK_PRECONDITION(bufferDescriptor.size + byteOffset <= mBuffer.GetSize())
+            << "Attempting to copy " << bufferDescriptor.size << " bytes into a buffer of size "
+            << mBuffer.GetSize() << " at offset " << byteOffset;
+    FILAMENT_CHECK_PRECONDITION(byteOffset % WEBGPU_BUFFER_SIZE_MODULUS == 0)
+            << "Byte offset must be a multiple of " << WEBGPU_BUFFER_SIZE_MODULUS << " but is "
+            << byteOffset;
+
+    // TODO: All buffer objects are created with CopyDst usage.
+    // This may have some performance implications. That should be investigated later.
+    assert_invariant(mBuffer.GetUsage() & wgpu::BufferUsage::CopyDst);
+
+    const size_t remainder = bufferDescriptor.size % WEBGPU_BUFFER_SIZE_MODULUS;
+
+    // WriteBuffer is an async call. But cpu buffer data is already written to the staging
+    // buffer on return from the WriteBuffer.
+    const size_t legalSize = bufferDescriptor.size - remainder;
+    queue.WriteBuffer(mBuffer, byteOffset, bufferDescriptor.buffer, legalSize);
+    if (remainder != 0) {
+        const uint8_t* remainderStart =
+                static_cast<const uint8_t*>(bufferDescriptor.buffer) + legalSize;
+        memcpy(mRemainderChunk.data(), remainderStart, remainder);
+        // Pad the remainder with zeros to ensure deterministic behavior, though GPU shouldn't
+        // access this
+        std::memset(mRemainderChunk.data() + remainder, 0, WEBGPU_BUFFER_SIZE_MODULUS - remainder);
+
+        queue.WriteBuffer(mBuffer, byteOffset + legalSize, &mRemainderChunk,
+                WEBGPU_BUFFER_SIZE_MODULUS);
+    }
+}
+
+}// namespace filament::backend

filament/backend/src/webgpu/WebGPUBufferBase.h

@@ -0,0 +1,48 @@
+/*
+ * Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_WEBGPUBUFFERBASE_H
+#define TNT_FILAMENT_BACKEND_WEBGPUBUFFERBASE_H
+
+#include "WebGPUConstants.h"
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <array>
+#include <cstdint>
+
+namespace filament::backend {
+
+class BufferDescriptor;
+
+class WebGPUBufferBase /* intended to be extended */ {
+public:
+    void updateGPUBuffer(BufferDescriptor const&, uint32_t byteOffset, wgpu::Queue const&);
+
+    [[nodiscard]] wgpu::Buffer const& getBuffer() const { return mBuffer; }
+
+protected:
+    WebGPUBufferBase(wgpu::Device const&, wgpu::BufferUsage, uint32_t size, char const* label);
+
+private:
+    const wgpu::Buffer mBuffer;
+    // WEBGPU_BUFFER_SIZE_MODULUS (e.g. 4) bytes to hold any extra chunk we need.
+    std::array<uint8_t, WEBGPU_BUFFER_SIZE_MODULUS> mRemainderChunk{};
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_WEBGPUBUFFERBASE_H