Add unfilterable

Bump MATERIAL_VERSION to 61
Update RELEASE_NOTES for the revert
2025-06-16 10:25:13 -07:00 · 2025-06-13 15:39:09 -07:00 · 2025-06-13 15:26:56 -07:00 · 2025-06-13 15:02:55 -07:00 · 2025-06-13 21:06:22 +00:00 · 2025-06-13 16:46:18 -04:00
254 changed files with 8364 additions and 3789 deletions
@@ -0,0 +1,45 @@
+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: Find commit message (PR)
+      shell: bash
+      id: checkout_code
+      if: github.event_name == 'pull_request'
+      run: |
+        echo "+++++ head commit message +++++"
+        echo "$(git log -1 --no-merges)"
+        echo "+++++++++++++++++++++++++++++++"
+        echo "hash=$(git rev-parse HEAD)" >> "$GITHUB_OUTPUT"
+        git checkout ${{ github.event.pull_request.head.sha }}
+        echo "$(git log -1 --no-merges)" >> /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: Cleanup Find commit message (PR)
+      shell: bash
+      if: github.event_name == 'pull_request'
+      run: |
+        git checkout ${{ steps.checkout_code.outputs.hash }}
@@ -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
@@ -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:
@@ -110,9 +109,9 @@ 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
@@ -121,20 +120,21 @@ jobs:
      - uses: actions/checkout@v4.1.6
        with:
          fetch-depth: 0
+      - id: get_commit_msg
+        uses: ./.github/actions/get-commit-msg
      - 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
@@ -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:
@@ -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
@@ -49,6 +49,8 @@ option(FILAMENT_SUPPORTS_OSMESA "Enable OSMesa (headless GL context) for Filamen

 option(FILAMENT_ENABLE_FGVIEWER "Enable the frame graph viewer" OFF)

+option(FILAMENT_USE_ABSEIL_LOGGING "Use Abseil to log, may increase binary size" 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)
@@ -591,6 +593,10 @@ if (CMAKE_BUILD_TYPE STREQUAL "Debug" AND NOT FILAMENT_BACKEND_DEBUG_FLAG STREQU
    unset(FILAMENT_BACKEND_DEBUG_FLAG)
 endif()

+if (FILAMENT_USE_ABSEIL_LOGGING)
+    add_definitions(-DFILAMENT_USE_ABSEIL_LOGGING)
+endif()
+
 # ==================================================================================================
 # Material compilation flags
 # ==================================================================================================
@@ -1,9 +1,9 @@
 # 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
+## Release notes for next branch cut
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.60.1'
+    implementation 'com.google.android.filament:filament-android:1.61.1'
 }
 ```

@@ -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', '~> 1.60.1'
+pod 'Filament', '~> 1.61.1'
 ```

 ## Documentation
@@ -7,6 +7,13 @@ 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).

+## v1.61.2
+
+- samples: samples now have a CLI to select backend api
+
+## v1.61.1
+
+
 ## v1.61.0

 - materials: sampler now export their type in the material binary [⚠️ **New Material Version**]
@@ -59,6 +59,10 @@ add_library(smol-v STATIC IMPORTED)
 set_target_properties(smol-v PROPERTIES IMPORTED_LOCATION
        ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libsmol-v.a)

+add_library(abseil STATIC IMPORTED)
+set_target_properties(abseil PROPERTIES IMPORTED_LOCATION
+        ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libabseil.a)
+
 if (FILAMENT_ENABLE_FGVIEWER)
    add_library(fgviewer STATIC IMPORTED)
    set_target_properties(fgviewer PROPERTIES IMPORTED_LOCATION
@@ -128,6 +132,7 @@ target_link_libraries(filament-jni
    PRIVATE jnigraphics
    PRIVATE utils
    PRIVATE perfetto
+    PRIVATE abseil

    # libgeometry is PUBLIC because gltfio uses it.
    PUBLIC geometry
@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.60.1
+VERSION_NAME=1.61.1

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

@@ -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
@@ -204,9 +204,10 @@
      "license": "Apache-2.0"
    },
    "node_modules/brace-expansion": {
-      "version": "1.1.11",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.11.tgz",
-      "integrity": "sha512-iCuPHDFgrHX7H2vEI/5xpz07zSHB00TpugqhmYtVmMO6518mCuRMoOYFldEBl0g187ufozdaHgWKcYFb61qGiA==",
+      "version": "1.1.12",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.12.tgz",
+      "integrity": "sha512-9T9UjW3r0UW5c1Q7GTwllptXwhvYmEzFhzMfZ9H7FQWt+uZePjZPjBP/W1ZEyZ1twGWom5/56TF4lPcqjnDHcg==",
+      "license": "MIT",
      "dependencies": {
        "balanced-match": "^1.0.0",
        "concat-map": "0.0.1"
@@ -409,9 +410,9 @@
      "integrity": "sha512-Nik3Sc0ncrMK4UUdXQmAnRtzmNQTAAXmXIopizwZ1W1t8QmfJj+zL4OA2I7XPTPW5z5TDqv4hRo/JzouDJnX3A=="
    },
    "brace-expansion": {
-      "version": "1.1.11",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.11.tgz",
-      "integrity": "sha512-iCuPHDFgrHX7H2vEI/5xpz07zSHB00TpugqhmYtVmMO6518mCuRMoOYFldEBl0g187ufozdaHgWKcYFb61qGiA==",
+      "version": "1.1.12",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.12.tgz",
+      "integrity": "sha512-9T9UjW3r0UW5c1Q7GTwllptXwhvYmEzFhzMfZ9H7FQWt+uZePjZPjBP/W1ZEyZ1twGWom5/56TF4lPcqjnDHcg==",
      "requires": {
        "balanced-match": "^1.0.0",
        "concat-map": "0.0.1"
@@ -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% ^
@@ -601,6 +601,10 @@ target_link_libraries(${TARGET} PUBLIC filaflat)
 target_link_libraries(${TARGET} PUBLIC filabridge)
 target_link_libraries(${TARGET} PUBLIC ibl-lite)

+if (FILAMENT_USE_ABSEIL_LOGGING)
+    target_link_libraries(${TARGET} PUBLIC absl::log)
+endif()
+
 if (FILAMENT_ENABLE_FGVIEWER)
    target_link_libraries(${TARGET} PUBLIC fgviewer)
    add_definitions(-DFILAMENT_ENABLE_FGVIEWER=1)
@@ -92,7 +92,7 @@ Copy your platform's Makefile below into a `Makefile` inside the same directory.
 ### Linux

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

 main: main.o
@@ -110,12 +110,13 @@ clean:
 ### macOS

 ```make
-FILAMENT_LIBS=-lfilament -lbackend -lbluegl -lbluevk -lfilabridge -lfilaflat -lutils -lgeometry -lsmol-v -lvkshaders -libl
+FILAMENT_LIBS=-lfilament -lbackend -lbluegl -lbluevk -lfilabridge -lfilaflat -lutils -lgeometry -lsmol-v -lvkshaders -libl -labseil
 FRAMEWORKS=-framework Cocoa -framework Metal -framework CoreVideo
 CC=clang++
+ARCH ?= $(shell uname -m)

 main: main.o
-	$(CC) -Llib/x86_64/ main.o $(FILAMENT_LIBS) $(FRAMEWORKS) -o main
+	$(CC) -Llib/$(ARCH)/ main.o $(FILAMENT_LIBS) $(FRAMEWORKS) -o main

 main.o: main.cpp
 	$(CC) -Iinclude/ -std=c++17 -c main.cpp
@@ -139,7 +140,7 @@ used to change the run-time library version.

 ```make
 FILAMENT_LIBS=filament.lib backend.lib bluegl.lib bluevk.lib filabridge.lib filaflat.lib \
-              utils.lib geometry.lib smol-v.lib ibl.lib vkshaders.lib
+              utils.lib geometry.lib smol-v.lib ibl.lib vkshaders.lib abseil.lib
 CC=cl.exe

 main.exe: main.obj
@@ -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
@@ -254,17 +257,36 @@ if (FILAMENT_SUPPORTS_WEBGPU)
    list(APPEND SRCS
            include/backend/platforms/WebGPUPlatform.h
            src/webgpu/platform/WebGPUPlatform.cpp
+            src/webgpu/SpdMipmapGenerator/SpdMipmapGenerator.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
    )
    if (WIN32)
        list(APPEND SRCS src/webgpu/platform/WebGPUPlatformWindows.cpp)
@@ -382,6 +404,10 @@ endif()
 target_link_libraries(${TARGET} PUBLIC math)
 target_link_libraries(${TARGET} PUBLIC utils)

+if (FILAMENT_USE_ABSEIL_LOGGING)
+    target_link_libraries(${TARGET} PRIVATE absl::log)
+endif()
+
 # Android, iOS, and WebGL do not use bluegl.
 if(FILAMENT_SUPPORTS_OPENGL AND NOT IOS AND NOT ANDROID AND NOT WEBGL)
    target_link_libraries(${TARGET} PRIVATE bluegl)
@@ -474,7 +474,12 @@ constexpr std::string_view to_string(DescriptorType type) noexcept {

 enum class DescriptorFlags : uint8_t {
    NONE = 0x00,
-    DYNAMIC_OFFSET = 0x01
+
+    // Indicate a UNIFORM_BUFFER will have dynamic offsets.
+    DYNAMIC_OFFSET = 0x01,
+
+    // To indicate a texture/sampler type should be unfiltered.
+    UNFILTERABLE = 0x02,
 };

 using descriptor_set_t = uint8_t;
@@ -149,6 +149,13 @@ public:
         *      - PlatformEGLAndroid
         */
        bool assertNativeWindowIsValid = false;
+
+        /**
+         * The action to take if a Drawable cannot be acquired. If true, the
+         * frame is aborted instead of panic. This is only supported for:
+         *      - PlatformMetal
+         */
+        bool metalDisablePanicOnDrawableFailure = false;
    };

    Platform() noexcept;
@@ -128,7 +128,7 @@ protected:
    bool setExternalImage(ExternalImageHandleRef externalImage, ExternalTexture* texture) noexcept override;

    /**
-     * Logs glGetError() to slog.e
+     * Logs glGetError() to LOG(ERROR)
     * @param name a string giving some context on the error. Typically __func__.
     */
    static void logEglError(const char* name) noexcept;
@@ -39,7 +39,7 @@
 #define FILAMENT_DEBUG_COMMANDS_NONE         0x0
 // Command debugging enabled. No logging by default.
 #define FILAMENT_DEBUG_COMMANDS_ENABLE       0x1
-// Command debugging enabled. Every command logged to slog.d
+// Command debugging enabled. Every command logged to DLOG(INFO)
 #define FILAMENT_DEBUG_COMMANDS_LOG          0x2
 // Command debugging enabled. Every command logged to systrace
 #define FILAMENT_DEBUG_COMMANDS_SYSTRACE     0x4
@@ -222,56 +222,6 @@ public:
        return static_cast<Dp>(p);
    }

-    template<typename Dp, typename B>
-    inline std::enable_if_t<
-            std::is_pointer_v<Dp> && std::is_base_of_v<B, std::remove_pointer_t<Dp>>, int16_t>
-    handle_set_value(Handle<B>& handle, uint8_t v) {
-        assert_invariant(handle);
-        auto [p, tag] = handleToPointer(handle.getId());
-
-        if (isPoolHandle(handle.getId())) {
-            // check for pool handle use-after-free
-            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
-                auto const pNode = static_cast<typename Allocator::Node*>(p);
-                pNode[-1].test_value = v;
-                return 1;
-            }
-            return 0;
-        } else {
-            // check for heap handle use-after-free
-            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
-                return -1;
-            }
-            return -2;
-        }
-    }
-
-    template<typename Dp, typename B>
-    inline std::enable_if_t<
-            std::is_pointer_v<Dp> && std::is_base_of_v<B, std::remove_pointer_t<Dp>>, int16_t>
-    handle_get_value(Handle<B>& handle) {
-        assert_invariant(handle);
-        auto [p, tag] = handleToPointer(handle.getId());
-
-        if (isPoolHandle(handle.getId())) {
-            // check for pool handle use-after-free
-            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
-                auto const pNode = static_cast<typename Allocator::Node*>(p);
-                uint8_t age = pNode[-1].age;
-                uint8_t test_value = pNode[-1].test_value;
-                int16_t ret = (int16_t) ((age << 8) | test_value);
-                return ret;
-            }
-            return 0;
-        } else {
-            // check for heap handle use-after-free
-            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
-                return -1;
-            }
-            return -2;
-        }
-    }
-
    utils::CString getHandleTag(HandleBase::HandleId key) const noexcept;

    template<typename B>
@@ -326,10 +276,7 @@ private:
    class Allocator {
        friend class HandleAllocator;
        static constexpr size_t MIN_ALIGNMENT = alignof(std::max_align_t);
-        struct Node {
-            uint8_t age;
-            uint8_t test_value = 0xff;
-        };
+        struct Node { uint8_t age; };
        // Note: using the `extra` parameter of PoolAllocator<>, even with a 1-byte structure,
        // generally increases all pool allocations by 8-bytes because of alignment restrictions.
        template<size_t SIZE>
@@ -16,7 +16,7 @@

 #include "private/backend/CircularBuffer.h"

-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
 #include <utils/architecture.h>
 #include <utils/ashmem.h>
@@ -127,7 +127,7 @@ void* CircularBuffer::alloc(size_t size) {
                "couldn't allocate " << (size * 2 / 1024) <<
                " KiB of virtual address space for the command buffer";

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

        // guard page at the end
        void* guard = (void*)(uintptr_t(data) + size * 2);
@@ -20,12 +20,12 @@

 #include <private/utils/Tracing.h>

-#include <utils/compiler.h>
-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Mutex.h>
-#include <utils/ostream.h>
 #include <utils/Panic.h>
+#include <utils/compiler.h>
 #include <utils/debug.h>
+#include <utils/ostream.h>

 #include <algorithm>
 #include <mutex>
@@ -121,11 +121,10 @@ void CommandBufferQueue::flush() {

 #ifndef NDEBUG
        size_t const totalUsed = circularBuffer.size() - mFreeSpace;
-        slog.d << "CommandStream used too much space (will block): "
-                << "needed space " << requiredSize << " out of " << mFreeSpace
-                << ", totalUsed=" << totalUsed << ", current=" << used
-                << ", queue size=" << mCommandBuffersToExecute.size() << " buffers"
-                << io::endl;
+        DLOG(INFO) << "CommandStream used too much space (will block): "
+                   << "needed space " << requiredSize << " out of " << mFreeSpace
+                   << ", totalUsed=" << totalUsed << ", current=" << used
+                   << ", queue size=" << mCommandBuffersToExecute.size() << " buffers";

        mHighWatermark = std::max(mHighWatermark, totalUsed);
 #endif
@@ -22,10 +22,13 @@
 #include <utils/CallStack.h>
 #endif

-#include <utils/compiler.h>
-#include <utils/Log.h>
-#include <utils/ostream.h>
+#include <private/utils/Tracing.h>
+
+#include <utils/Logger.h>
 #include <utils/Profiler.h>
+#include <utils/compiler.h>
+#include <utils/ostream.h>
+#include <utils/sstream.h>

 #include <cstddef>
 #include <functional>
@@ -129,9 +132,10 @@ void CommandType<void (Driver::*)(ARGS...)>::Command<METHOD>::log(std::index_seq
 #if DEBUG_COMMAND_STREAM
    static_assert(UTILS_HAS_RTTI, "DEBUG_COMMAND_STREAM can only be used with RTTI");
    std::string command = utils::CallStack::demangleTypeName(typeid(Command).name()).c_str();
-    slog.d << extractMethodName(command) << " : size=" << sizeof(Command) << "\n\t";
-    printParameterPack(slog.d, std::get<I>(mArgs)...);
-    slog.d << io::endl;
+    DLOG(INFO) << extractMethodName(command) << " : size=" << sizeof(Command);
+    utils::io::sstream parameterPack;
+    printParameterPack(parameterPack, std::get<I>(mArgs)...);
+    DLOG(INFO) << "\t" << parameterPack.c_str();
 #endif
 }

@@ -25,9 +25,9 @@

 #include <private/utils/Tracing.h>

+#include <utils/Logger.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
-#include <utils/Log.h>
 #include <utils/ostream.h>

 #include <math/half.h>
@@ -149,7 +149,7 @@ void DriverBase::scheduleRelease(AcquiredImage const& image) noexcept {
 void DriverBase::debugCommandBegin(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_LOG)) {
-            utils::slog.d << methodName << utils::io::endl;
+            DLOG(INFO) << methodName;
        }
        if constexpr (bool(FILAMENT_DEBUG_COMMANDS & FILAMENT_DEBUG_COMMANDS_SYSTRACE)) {
            FILAMENT_TRACING_CONTEXT(FILAMENT_TRACING_CATEGORY_FILAMENT);
@@ -20,7 +20,7 @@

 #include <utils/Allocator.h>
 #include <utils/CString.h>
-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -57,8 +57,8 @@ HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& ar
    size_t const maxHeapSize = std::min(area.size(), HANDLE_INDEX_MASK * getAlignment());

    if (UTILS_UNLIKELY(maxHeapSize != area.size())) {
-        slog.w << "HandleAllocator heap size reduced to "
-               << maxHeapSize << " from " << area.size() << io::endl;
+        LOG(WARNING) << "HandleAllocator heap size reduced to " << maxHeapSize << " from "
+                     << area.size();
    }

    // make sure we start with a clean arena. This is needed to ensure that all blocks start
@@ -19,8 +19,8 @@
 #include "MetalContext.h"
 #include "MetalUtils.h"

+#include <utils/Logger.h>
 #include <utils/Panic.h>
-#include <utils/Log.h>

 namespace filament::backend {

@@ -323,7 +323,7 @@ id<MTLFunction> MetalBlitter::compileFragmentFunction(BlitFunctionKey key) const
    if (!library || !function) {
        if (error) {
            auto description = [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
-            utils::slog.e << description << utils::io::endl;
+            LOG(ERROR) << description;
        }
    }
    FILAMENT_CHECK_POSTCONDITION(library && function)
@@ -349,7 +349,7 @@ id<MTLFunction> MetalBlitter::getBlitVertexFunction() {
    if (!library || !function) {
        if (error) {
            auto description = [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
-            utils::slog.e << description << utils::io::endl;
+            LOG(ERROR) << description;
        }
    }
    FILAMENT_CHECK_POSTCONDITION(library && function)
@@ -18,8 +18,8 @@

 #include "MetalContext.h"

+#include <utils/Logger.h>
 #include <utils/Panic.h>
-#include <utils/Log.h>
 #include <utils/trap.h>

 #include <thread>
@@ -77,8 +77,7 @@ void MetalBufferPool::releaseBuffer(MetalBufferPoolEntry const *stage) noexcept

    auto iter = mUsedStages.find(stage);
    if (iter == mUsedStages.end()) {
-        utils::slog.e << "Unknown Metal buffer: " << stage->capacity << " bytes"
-                << utils::io::endl;
+        LOG(ERROR) << "Unknown Metal buffer: " << stage->capacity << " bytes";
        return;
    }
    stage->lastAccessed = mCurrentFrame;
@@ -18,8 +18,9 @@

 #include "MetalHandles.h"

-#include <utils/debug.h>
 #include <utils/FixedCapacityVector.h>
+#include <utils/Logger.h>
+#include <utils/debug.h>

 #include <utility>

@@ -112,8 +113,8 @@ id<MTLCommandBuffer> getPendingCommandBuffer(MetalContext* context) {
        auto errorCode = (MTLCommandBufferError)buffer.error.code;
        if (@available(macOS 11.0, *)) {
            if (errorCode == MTLCommandBufferErrorMemoryless) {
-                utils::slog.w << "Metal: memoryless geometry limit reached. "
-                        "Continuing with private storage mode." << utils::io::endl;
+                LOG(WARNING) << "Metal: memoryless geometry limit reached. Continuing with private "
+                                "storage mode.";
                context->memorylessLimitsReached = true;
            }
        }
@@ -38,10 +38,10 @@
 #include <Metal/Metal.h>
 #include <QuartzCore/QuartzCore.h>

-#include <utils/Log.h>
+#include <utils/Invocable.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
 #include <utils/sstream.h>
-#include <utils/Invocable.h>

 #include <algorithm>

@@ -78,20 +78,18 @@ Driver* MetalDriverFactory::create(PlatformMetal* const platform, const Platform
    //    MetalVertexBufferInfo        : 552       moderate
    // -- less than or equal to 552 bytes

-    utils::slog.d
-           << "\nMetalSwapChain: " << sizeof(MetalSwapChain)
-           << "\nMetalBufferObject: " << sizeof(MetalBufferObject)
-           << "\nMetalVertexBuffer: " << sizeof(MetalVertexBuffer)
-           << "\nMetalVertexBufferInfo: " << sizeof(MetalVertexBufferInfo)
-           << "\nMetalIndexBuffer: " << sizeof(MetalIndexBuffer)
-           << "\nMetalRenderPrimitive: " << sizeof(MetalRenderPrimitive)
-           << "\nMetalTexture: " << sizeof(MetalTexture)
-           << "\nMetalTimerQuery: " << sizeof(MetalTimerQuery)
-           << "\nHwStream: " << sizeof(HwStream)
-           << "\nMetalRenderTarget: " << sizeof(MetalRenderTarget)
-           << "\nMetalFence: " << sizeof(MetalFence)
-           << "\nMetalProgram: " << sizeof(MetalProgram)
-           << utils::io::endl;
+    DLOG(INFO) << "MetalSwapChain: " << sizeof(MetalSwapChain);
+    DLOG(INFO) << "MetalBufferObject: " << sizeof(MetalBufferObject);
+    DLOG(INFO) << "MetalVertexBuffer: " << sizeof(MetalVertexBuffer);
+    DLOG(INFO) << "MetalVertexBufferInfo: " << sizeof(MetalVertexBufferInfo);
+    DLOG(INFO) << "MetalIndexBuffer: " << sizeof(MetalIndexBuffer);
+    DLOG(INFO) << "MetalRenderPrimitive: " << sizeof(MetalRenderPrimitive);
+    DLOG(INFO) << "MetalTexture: " << sizeof(MetalTexture);
+    DLOG(INFO) << "MetalTimerQuery: " << sizeof(MetalTimerQuery);
+    DLOG(INFO) << "HwStream: " << sizeof(HwStream);
+    DLOG(INFO) << "MetalRenderTarget: " << sizeof(MetalRenderTarget);
+    DLOG(INFO) << "MetalFence: " << sizeof(MetalFence);
+    DLOG(INFO) << "MetalProgram: " << sizeof(MetalProgram);
 #endif
    return MetalDriver::create(platform, driverConfig);
 }
@@ -135,19 +133,18 @@ MetalDriver::MetalDriver(

    initializeSupportedGpuFamilies(mContext);

-    utils::slog.v << "Supported GPU families: " << utils::io::endl;
+    LOG(INFO) << "Supported GPU families: ";
    if (mContext->highestSupportedGpuFamily.common > 0) {
-        utils::slog.v << "  MTLGPUFamilyCommon" << (int) mContext->highestSupportedGpuFamily.common << utils::io::endl;
+        LOG(INFO) << "  MTLGPUFamilyCommon" << (int) mContext->highestSupportedGpuFamily.common;
    }
    if (mContext->highestSupportedGpuFamily.apple > 0) {
-        utils::slog.v << "  MTLGPUFamilyApple" << (int) mContext->highestSupportedGpuFamily.apple << utils::io::endl;
+        LOG(INFO) << "  MTLGPUFamilyApple" << (int) mContext->highestSupportedGpuFamily.apple;
    }
    if (mContext->highestSupportedGpuFamily.mac > 0) {
-        utils::slog.v << "  MTLGPUFamilyMac" << (int) mContext->highestSupportedGpuFamily.mac << utils::io::endl;
+        LOG(INFO) << "  MTLGPUFamilyMac" << (int) mContext->highestSupportedGpuFamily.mac;
    }
-    utils::slog.v << "Features:" << utils::io::endl;
-    utils::slog.v << "  readWriteTextureSupport: " <<
-            (bool) mContext->device.readWriteTextureSupport << utils::io::endl;
+    LOG(INFO) << "Features:";
+    LOG(INFO) << "  readWriteTextureSupport: " << (bool) mContext->device.readWriteTextureSupport;

    // In order to support texture swizzling, the GPU needs to support it and the system be running
    // iOS 13+.
@@ -588,7 +585,7 @@ void MetalDriver::createProgramR(Handle<HwProgram> rph, Program&& program) {
 #if FILAMENT_METAL_DEBUG_LOG
    auto handleId = rph.getId();
    DEBUG_LOG("createProgramR(rph = %d, program = ", handleId);
-    utils::slog.d << program << utils::io::endl;
+    DLOG(INFO) << program;
 #endif
    construct_handle<MetalProgram>(rph, *mContext, std::move(program));
 }
@@ -2077,7 +2074,7 @@ void MetalDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGro
                                                            error:&error];
    if (error) {
        auto description = [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
-        utils::slog.e << description << utils::io::endl;
+        LOG(ERROR) << description;
    }
    assert_invariant(!error);

@@ -20,14 +20,14 @@
 #include "MetalEnums.h"
 #include "MetalUtils.h"

+#include <utils/Logger.h>
 #include <utils/Panic.h>
-#include <utils/Log.h>
 #include <utils/trap.h>

 #define NSERROR_CHECK(message)                                                                     \
    if (error) {                                                                                   \
        auto description = [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding]; \
-        utils::slog.e << description << utils::io::endl;                                           \
+        LOG(ERROR) << description;                                                                 \
    }                                                                                              \
    FILAMENT_CHECK_POSTCONDITION(error == nil) << message;

@@ -27,10 +27,11 @@

 #include "private/backend/BackendUtils.h"

-#include <utils/compiler.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
-#include <utils/trap.h>
+#include <utils/compiler.h>
 #include <utils/debug.h>
+#include <utils/trap.h>

 #include <math/scalar.h>

@@ -115,14 +116,14 @@ MetalSwapChain::MetalSwapChain(
      type(SwapChainType::CAMETALLAYER) {

    if (!(flags & SwapChain::CONFIG_TRANSPARENT) && !nativeWindow.opaque) {
-        utils::slog.w << "Warning: Filament SwapChain has no CONFIG_TRANSPARENT flag, "
-                         "but the CAMetaLayer(" << (__bridge void*) nativeWindow << ")"
-                         " has .opaque set to NO." << utils::io::endl;
+        LOG(WARNING) << "Warning: Filament SwapChain has no CONFIG_TRANSPARENT flag, but the "
+                        "CAMetaLayer("
+                     << (__bridge void*) nativeWindow << ") has .opaque set to NO.";
    }
    if ((flags & SwapChain::CONFIG_TRANSPARENT) && nativeWindow.opaque) {
-        utils::slog.w << "Warning: Filament SwapChain has the CONFIG_TRANSPARENT flag, "
-                         "but the CAMetaLayer(" << (__bridge void*) nativeWindow << ")"
-                         " has .opaque set to YES." << utils::io::endl;
+        LOG(WARNING) << "Warning: Filament SwapChain has the CONFIG_TRANSPARENT flag, but the "
+                        "CAMetaLayer("
+                     << (__bridge void*) nativeWindow << ") has .opaque set to YES.";
    }

    // Needed so we can use the SwapChain as a blit source.
@@ -21,6 +21,7 @@
 #include <backend/Program.h>

 #include <utils/JobSystem.h>
+#include <utils/Logger.h>
 #include <utils/Mutex.h>

 #include <chrono>
@@ -147,7 +148,7 @@ bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
            if (error) {
                auto description =
                        [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
-                utils::slog.w << description << utils::io::endl;
+                LOG(WARNING) << description;
                errorMessage = error.localizedDescription;
            }
            PANIC_LOG("Failed to compile Metal program.");
@@ -178,7 +179,7 @@ bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
            if (error) {
                auto description =
                        [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
-                utils::slog.w << description << utils::io::endl;
+                LOG(WARNING) << description;
                errorMessage = error.localizedDescription;
            }
            PANIC_LOG("Failed to load main0 in Metal program.");
@@ -18,7 +18,7 @@

 #include "MetalEnums.h"

-#include <utils/Log.h>
+#include <utils/Logger.h>

 namespace filament {
 namespace backend {
@@ -95,7 +95,7 @@ id<MTLRenderPipelineState> PipelineStateCreator::operator()(id<MTLDevice> device
            [NSString stringWithFormat:@"Could not create Metal pipeline state: %@",
                error ? error.localizedDescription : @"unknown error"];
        auto description = [errorMessage cStringUsingEncoding:NSUTF8StringEncoding];
-        utils::slog.e << description << utils::io::endl;
+        LOG(ERROR) << description;
        [[NSException exceptionWithName:@"MetalRenderPipelineFailure"
                                 reason:errorMessage
                               userInfo:nil] raise];
@@ -155,7 +155,7 @@ id<MTLSamplerState> SamplerStateCreator::operator()(id<MTLDevice> device,
    // MTLSamplerDescriptor.
    // In practice, this means shadows are not supported when running in the simulator.
    if (![device supportsFeatureSet:MTLFeatureSet_iOS_GPUFamily3_v1]) {
-        utils::slog.w << "Warning: sample comparison not supported by this GPU" << utils::io::endl;
+        LOG(WARNING) << "Warning: sample comparison not supported by this GPU";
        samplerDescriptor.compareFunction = MTLCompareFunctionNever;
    }
 #endif
@@ -19,7 +19,7 @@

 #include "MetalDriverFactory.h"

-#include <utils/Log.h>
+#include <utils/Logger.h>

 #import <Foundation/Foundation.h>

@@ -53,6 +53,9 @@ PlatformMetal::~PlatformMetal() noexcept {
 }

 Driver* PlatformMetal::createDriver(void* /*sharedContext*/, const Platform::DriverConfig& driverConfig) noexcept {
+    pImpl->mDrawableFailureBehavior = driverConfig.metalDisablePanicOnDrawableFailure
+            ? DrawableFailureBehavior::ABORT_FRAME
+            : DrawableFailureBehavior::PANIC;
    return MetalDriverFactory::create(this, driverConfig);
 }

@@ -129,9 +132,8 @@ void PlatformMetalImpl::createDeviceImpl(MetalDevice& outDevice) {
        result = MTLCreateSystemDefaultDevice();
    }

-    utils::slog.i << "Selected physical device '"
-                  << [result.name cStringUsingEncoding:NSUTF8StringEncoding] << "'"
-                  << utils::io::endl;
+    LOG(INFO) << "Selected physical device '"
+              << [result.name cStringUsingEncoding:NSUTF8StringEncoding] << "'";

    outDevice.device = result;
    mDevice = result;
@@ -360,7 +360,7 @@ void GLDescriptorSet::bind(
            }
        }, entry.desc);
    });
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void GLDescriptorSet::validate(HandleAllocatorGL& allocator,
@@ -18,6 +18,7 @@

 #include "private/backend/Driver.h"

+#include <utils/Logger.h>
 #include <utils/compiler.h>
 #include <utils/ostream.h>
 #include <utils/trap.h>
@@ -25,6 +26,7 @@
 #include <string_view>

 #include <stddef.h>
+#include <cstdio>

 namespace filament::backend {

@@ -55,19 +57,21 @@ std::string_view getGLErrorString(GLenum error) noexcept {
 }

 UTILS_NOINLINE
-GLenum checkGLError(io::ostream& out, const char* function, size_t line) noexcept {
+GLenum checkGLError(const char* function, size_t line) noexcept {
    GLenum const error = glGetError();
    if (UTILS_VERY_UNLIKELY(error != GL_NO_ERROR)) {
        auto const string = getGLErrorString(error);
-        out << "OpenGL error " << io::hex << error << " (" << string << ") in \""
-            << function << "\" at line " << io::dec << line << io::endl;
+        char hexError[16];
+        snprintf(hexError, sizeof(hexError), "%#x", error);
+        LOG(ERROR) << "OpenGL error " << hexError << " (" << string << ") in \"" << function
+                   << "\" at line " << line;
    }
    return error;
 }

 UTILS_NOINLINE
-void assertGLError(io::ostream& out, const char* function, size_t line) noexcept {
-    GLenum const err = checkGLError(out, function, line);
+void assertGLError(const char* function, size_t line) noexcept {
+    GLenum const err = checkGLError(function, line);
    if (UTILS_VERY_UNLIKELY(err != GL_NO_ERROR)) {
        debug_trap();
    }
@@ -97,19 +101,21 @@ std::string_view getFramebufferStatusString(GLenum status) noexcept {
 }

 UTILS_NOINLINE
-GLenum checkFramebufferStatus(io::ostream& out, GLenum target, const char* function, size_t line) noexcept {
+GLenum checkFramebufferStatus(GLenum target, const char* function, size_t line) noexcept {
    GLenum const status = glCheckFramebufferStatus(target);
    if (UTILS_VERY_UNLIKELY(status != GL_FRAMEBUFFER_COMPLETE)) {
        auto const string = getFramebufferStatusString(status);
-        out << "OpenGL framebuffer error " << io::hex << status << " (" << string << ") in \""
-            << function << "\" at line " << io::dec << line << io::endl;
+        char hexStatus[16];
+        snprintf(hexStatus, sizeof(hexStatus), "%#x", status);
+        LOG(ERROR) << "OpenGL framebuffer error " << hexStatus << " (" << string << ") in \""
+                   << function << "\" at line " << line;
    }
    return status;
 }

 UTILS_NOINLINE
-void assertFramebufferStatus(io::ostream& out, GLenum target, const char* function, size_t line) noexcept {
-    GLenum const status = checkFramebufferStatus(out, target, function, line);
+void assertFramebufferStatus(GLenum target, const char* function, size_t line) noexcept {
+    GLenum const status = checkFramebufferStatus(target, function, line);
    if (UTILS_VERY_UNLIKELY(status != GL_FRAMEBUFFER_COMPLETE)) {
        debug_trap();
    }
@@ -33,21 +33,21 @@
 namespace filament::backend::GLUtils {

 std::string_view getGLErrorString(GLenum error) noexcept;
-GLenum checkGLError(utils::io::ostream& out, const char* function, size_t line) noexcept;
-void assertGLError(utils::io::ostream& out, const char* function, size_t line) noexcept;
+GLenum checkGLError(const char* function, size_t line) noexcept;
+void assertGLError(const char* function, size_t line) noexcept;

 std::string_view getFramebufferStatusString(GLenum err) noexcept;
-GLenum checkFramebufferStatus(utils::io::ostream& out, GLenum target, const char* function, size_t line) noexcept;
-void assertFramebufferStatus(utils::io::ostream& out, GLenum target, const char* function, size_t line) noexcept;
+GLenum checkFramebufferStatus(GLenum target, const char* function, size_t line) noexcept;
+void assertFramebufferStatus(GLenum target, const char* function, size_t line) noexcept;

 #ifdef NDEBUG
-#   define CHECK_GL_ERROR(out)
-#   define CHECK_GL_ERROR_NON_FATAL(out)
-#   define CHECK_GL_FRAMEBUFFER_STATUS(out, target)
+#   define CHECK_GL_ERROR()
+#   define CHECK_GL_ERROR_NON_FATAL()
+#   define CHECK_GL_FRAMEBUFFER_STATUS(target)
 #else
-#   define CHECK_GL_ERROR(out) { GLUtils::assertGLError(out, __func__, __LINE__); }
-#   define CHECK_GL_ERROR_NON_FATAL(out) { GLUtils::checkGLError(out, __func__, __LINE__); }
-#   define CHECK_GL_FRAMEBUFFER_STATUS(out, target) { GLUtils::checkFramebufferStatus(out, target, __func__, __LINE__); }
+#   define CHECK_GL_ERROR() { GLUtils::assertGLError(__func__, __LINE__); }
+#   define CHECK_GL_ERROR_NON_FATAL() { GLUtils::checkGLError(__func__, __LINE__); }
+#   define CHECK_GL_FRAMEBUFFER_STATUS(target) { GLUtils::checkFramebufferStatus( target, __func__, __LINE__); }
 #endif

 constexpr GLuint getComponentCount(ElementType const type) noexcept {
@@ -22,9 +22,9 @@
 #include <backend/platforms/OpenGLPlatform.h>
 #include <backend/DriverEnums.h>

+#include <utils/Logger.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
-#include <utils/Log.h>
 #include <utils/ostream.h>

 #include <functional>
@@ -77,8 +77,8 @@ OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
    state.version  = (char const*)glGetString(GL_VERSION);
    state.shader   = (char const*)glGetString(GL_SHADING_LANGUAGE_VERSION);

-    slog.v << "[" << state.vendor << "], [" << state.renderer << "], "
-              "[" << state.version << "], [" << state.shader << "]" << io::endl;
+    LOG(INFO) << "[" << state.vendor << "], [" << state.renderer << "], "
+                 "[" << state.version << "], [" << state.shader << "]";

    /*
     * Figure out GL / GLES version, extensions and capabilities we need to
@@ -164,51 +164,33 @@ OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
    }
 #endif

-    slog.v << "Feature level: " << +mFeatureLevel << '\n';
-    slog.v << "Active workarounds: " << '\n';
+    LOG(INFO) << "Feature level: " << +mFeatureLevel;
+    LOG(INFO) << "Active workarounds: ";
    UTILS_NOUNROLL
    for (auto [enabled, name, _] : mBugDatabase) {
        if (enabled) {
-            slog.v << name << '\n';
+            LOG(INFO) << name;
        }
    }
-    flush(slog.v);

 #ifndef NDEBUG
    // this is useful for development
-    slog.v
-            << "GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT = "
-                    << gets.max_anisotropy << '\n'
-            << "GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS = "
-                    << gets.max_combined_texture_image_units << '\n'
-            << "GL_MAX_TEXTURE_SIZE = "
-                    << gets.max_texture_size << '\n'
-            << "GL_MAX_CUBE_MAP_TEXTURE_SIZE = "
-                    << gets.max_cubemap_texture_size << '\n'
-            << "GL_MAX_3D_TEXTURE_SIZE = "
-                    << gets.max_3d_texture_size << '\n'
-            << "GL_MAX_ARRAY_TEXTURE_LAYERS = "
-                    << gets.max_array_texture_layers << '\n'
-            << "GL_MAX_DRAW_BUFFERS = "
-                    << gets.max_draw_buffers << '\n'
-            << "GL_MAX_RENDERBUFFER_SIZE = "
-                    << gets.max_renderbuffer_size << '\n'
-            << "GL_MAX_SAMPLES = "
-                    << gets.max_samples << '\n'
-            << "GL_MAX_TEXTURE_IMAGE_UNITS = "
-                    << gets.max_texture_image_units << '\n'
-            << "GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS = "
-                    << gets.max_transform_feedback_separate_attribs << '\n'
-            << "GL_MAX_UNIFORM_BLOCK_SIZE = "
-                    << gets.max_uniform_block_size << '\n'
-            << "GL_MAX_UNIFORM_BUFFER_BINDINGS = "
-                    << gets.max_uniform_buffer_bindings << '\n'
-            << "GL_NUM_PROGRAM_BINARY_FORMATS = "
-                    << gets.num_program_binary_formats << '\n'
-            << "GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT = "
-                    << gets.uniform_buffer_offset_alignment << '\n'
-            ;
-    flush(slog.v);
+    LOG(INFO) << "GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT = " << gets.max_anisotropy;
+    LOG(INFO) << "GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS = " << gets.max_combined_texture_image_units;
+    LOG(INFO) << "GL_MAX_TEXTURE_SIZE = " << gets.max_texture_size;
+    LOG(INFO) << "GL_MAX_CUBE_MAP_TEXTURE_SIZE = " << gets.max_cubemap_texture_size;
+    LOG(INFO) << "GL_MAX_3D_TEXTURE_SIZE = " << gets.max_3d_texture_size;
+    LOG(INFO) << "GL_MAX_ARRAY_TEXTURE_LAYERS = " << gets.max_array_texture_layers;
+    LOG(INFO) << "GL_MAX_DRAW_BUFFERS = " << gets.max_draw_buffers;
+    LOG(INFO) << "GL_MAX_RENDERBUFFER_SIZE = " << gets.max_renderbuffer_size;
+    LOG(INFO) << "GL_MAX_SAMPLES = " << gets.max_samples;
+    LOG(INFO) << "GL_MAX_TEXTURE_IMAGE_UNITS = " << gets.max_texture_image_units;
+    LOG(INFO) << "GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS = "
+              << gets.max_transform_feedback_separate_attribs;
+    LOG(INFO) << "GL_MAX_UNIFORM_BLOCK_SIZE = " << gets.max_uniform_block_size;
+    LOG(INFO) << "GL_MAX_UNIFORM_BUFFER_BINDINGS = " << gets.max_uniform_buffer_bindings;
+    LOG(INFO) << "GL_NUM_PROGRAM_BINARY_FORMATS = " << gets.num_program_binary_formats;
+    LOG(INFO) << "GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT = " << gets.uniform_buffer_offset_alignment;
 #endif

 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
@@ -242,15 +224,14 @@ OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
    if (ext.KHR_debug) {
        auto cb = +[](GLenum, GLenum type, GLuint, GLenum severity, GLsizei length,
                const GLchar* message, const void *) {
-            io::ostream* stream = &slog.i;
+            auto logSeverity = utils::LogSeverity::kInfo;
            switch (severity) {
-                case GL_DEBUG_SEVERITY_HIGH:    stream = &slog.e;   break;
-                case GL_DEBUG_SEVERITY_MEDIUM:  stream = &slog.w;   break;
-                case GL_DEBUG_SEVERITY_LOW:     stream = &slog.d;   break;
+                case GL_DEBUG_SEVERITY_HIGH:    logSeverity = utils::LogSeverity::kError;   break;
+                case GL_DEBUG_SEVERITY_MEDIUM:  logSeverity = utils::LogSeverity::kWarning; break;
+                case GL_DEBUG_SEVERITY_LOW:     logSeverity = utils::LogSeverity::kInfo;    break;
                case GL_DEBUG_SEVERITY_NOTIFICATION:
                default: break;
            }
-            io::ostream& out = *stream;
            const char* level = ": ";
            switch (type) {
                case GL_DEBUG_TYPE_ERROR:               level = "ERROR: ";               break;
@@ -262,7 +243,7 @@ OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
                case GL_DEBUG_TYPE_MARKER:              level = "MARKER: ";              break;
                default: break;
            }
-            out << "KHR_debug " << level << std::string_view{ message, size_t(length) } << io::endl;
+            LOG(LEVEL(logSeverity)) << "KHR_debug " << level << std::string_view{ message, size_t(length) };
        };
        glEnable(GL_DEBUG_OUTPUT);
        glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS);
@@ -707,9 +688,8 @@ void OpenGLContext::initExtensionsGLES(Extensions* ext, GLint major, GLint minor
    GLUtils::unordered_string_set const exts = GLUtils::split(extensions);
    if constexpr (DEBUG_PRINT_EXTENSIONS) {
        for (auto extension: exts) {
-            slog.d << "\"" << std::string_view(extension) << "\"\n";
+            DLOG(INFO) << "\"" << std::string_view(extension) << "\"";
        }
-        flush(slog.d);
    }

    // figure out and initialize the extensions we need
@@ -783,9 +763,8 @@ void OpenGLContext::initExtensionsGL(Extensions* ext, GLint major, GLint minor)
    }
    if constexpr (DEBUG_PRINT_EXTENSIONS) {
        for (auto extension: exts) {
-            slog.d << "\"" << std::string_view(extension) << "\"\n";
+            DLOG(INFO) << "\"" << std::string_view(extension) << "\"";
        }
-        flush(slog.d);
    }

    using namespace std::literals;
@@ -1045,7 +1024,7 @@ GLuint OpenGLContext::getSamplerSlow(SamplerParams params) const noexcept {
                std::min(gets.max_anisotropy, anisotropy));
    }
 #endif
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
    mSamplerMap[params] = s;
    return s;
 }
@@ -48,12 +48,13 @@

 #include <type_traits>
 #include <utils/BitmaskEnum.h>
-#include <utils/FixedCapacityVector.h>
 #include <utils/CString.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Invocable.h>
-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
 #include <utils/Slice.h>
+#include <utils/Systrace.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/ostream.h>
@@ -184,20 +185,18 @@ OpenGLDriver* OpenGLDriver::create(OpenGLPlatform* platform,
    //    GLVertexBufferInfo        : 132       moderate
    // -- less than or equal to 136 bytes

-    slog.d
-           << "\nGLSwapChain: " << sizeof(GLSwapChain)
-           << "\nGLBufferObject: " << sizeof(GLBufferObject)
-           << "\nGLVertexBuffer: " << sizeof(GLVertexBuffer)
-           << "\nGLVertexBufferInfo: " << sizeof(GLVertexBufferInfo)
-           << "\nGLIndexBuffer: " << sizeof(GLIndexBuffer)
-           << "\nGLRenderPrimitive: " << sizeof(GLRenderPrimitive)
-           << "\nGLTexture: " << sizeof(GLTexture)
-           << "\nGLTimerQuery: " << sizeof(GLTimerQuery)
-           << "\nGLStream: " << sizeof(GLStream)
-           << "\nGLRenderTarget: " << sizeof(GLRenderTarget)
-           << "\nGLFence: " << sizeof(GLFence)
-           << "\nOpenGLProgram: " << sizeof(OpenGLProgram)
-           << io::endl;
+    DLOG(INFO) << "GLSwapChain: " << sizeof(GLSwapChain);
+    DLOG(INFO) << "GLBufferObject: " << sizeof(GLBufferObject);
+    DLOG(INFO) << "GLVertexBuffer: " << sizeof(GLVertexBuffer);
+    DLOG(INFO) << "GLVertexBufferInfo: " << sizeof(GLVertexBufferInfo);
+    DLOG(INFO) << "GLIndexBuffer: " << sizeof(GLIndexBuffer);
+    DLOG(INFO) << "GLRenderPrimitive: " << sizeof(GLRenderPrimitive);
+    DLOG(INFO) << "GLTexture: " << sizeof(GLTexture);
+    DLOG(INFO) << "GLTimerQuery: " << sizeof(GLTimerQuery);
+    DLOG(INFO) << "GLStream: " << sizeof(GLStream);
+    DLOG(INFO) << "GLRenderTarget: " << sizeof(GLRenderTarget);
+    DLOG(INFO) << "GLFence: " << sizeof(GLFence);
+    DLOG(INFO) << "OpenGLProgram: " << sizeof(OpenGLProgram);
 #endif

    // here we check we're on a supported version of GL before initializing the driver
@@ -289,7 +288,7 @@ OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfi
    mStreamsWithPendingAcquiredImage.reserve(8);

 #ifndef NDEBUG
-    slog.i << "OS version: " << mPlatform.getOSVersion() << io::endl;
+    LOG(INFO) << "OS version: " << mPlatform.getOSVersion();
 #endif

    // Timer queries are core in GL 3.3, otherwise we need EXT_disjoint_timer_query
@@ -643,25 +642,6 @@ void OpenGLDriver::createVertexBufferInfoR(
        AttributeArray attributes) {
    DEBUG_MARKER()
    construct<GLVertexBufferInfo>(vbih, bufferCount, attributeCount, attributes);
-
-    // test ----
-    int16_t returned_value = handle_set_value<GLVertexBufferInfo*>(vbih, 90);
-    uint8_t age = (uint8_t) (returned_value >> 8);
-    uint8_t test_value = (returned_value & 0xff);
-    switch (test_value) {
-        case -2:
-        case -1:
-        case 0:
-            utils::slog.w << "@@@ createVertexBufferInfoR: unexpected result, age:" << (int) age
-                          << ", test_value:" << (int) test_value << utils::io::endl;
-            break;
-        default:
-            FILAMENT_CHECK_POSTCONDITION(test_value == 1)
-                    << "@@@ createVertexBufferInfoR: shouldn't happen, age:" << (int) age
-                    << ", test_value:" << (int) test_value << utils::io::endl;
-            break;
-    }
-    //----------
 }

 void OpenGLDriver::createVertexBufferR(
@@ -687,7 +667,7 @@ void OpenGLDriver::createIndexBufferR(
    gl.bindVertexArray(nullptr);
    gl.bindBuffer(GL_ELEMENT_ARRAY_BUFFER, ib->gl.buffer);
    glBufferData(GL_ELEMENT_ARRAY_BUFFER, size, nullptr, getBufferUsage(usage));
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createBufferObjectR(Handle<HwBufferObject> boh,
@@ -712,7 +692,7 @@ void OpenGLDriver::createBufferObjectR(Handle<HwBufferObject> boh,
        glBufferData(bo->gl.binding, byteCount, nullptr, getBufferUsage(usage));
    }

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
@@ -749,7 +729,7 @@ void OpenGLDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
    // this records the index buffer into the currently bound VAO
    gl.bindBuffer(GL_ELEMENT_ARRAY_BUFFER, ib->gl.buffer);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {
@@ -779,26 +759,7 @@ void OpenGLDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {
    }

    construct<OpenGLProgram>(ph, *this, std::move(program));
-    CHECK_GL_ERROR(utils::slog.e)
-
-    // test ----
-    int16_t returned_value = handle_set_value<OpenGLProgram*>(ph, 100);
-    uint8_t age = (uint8_t) (returned_value >> 8);
-    uint8_t test_value = (returned_value & 0xff);
-    switch (test_value) {
-        case -2:
-        case -1:
-        case 0:
-            utils::slog.w << "@@@ createProgramR: unexpected result, age:" << (int)age << ", test_value:" << (int)test_value
-                          << utils::io::endl;
-            break;
-        default:
-            FILAMENT_CHECK_POSTCONDITION(test_value == 1)
-                    << "@@@ createProgramR: shouldn't happen, age:" << (int)age
-                    << ", test_value:" << (int)test_value << utils::io::endl;
-            break;
-    }
-    //----------
+    CHECK_GL_ERROR()
 }

 UTILS_NOINLINE
@@ -991,7 +952,7 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
        renderBufferStorage(t->gl.id, internalFormat, w, h, samples);
    }

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createTextureViewR(Handle<HwTexture> th,
@@ -1037,7 +998,7 @@ void OpenGLDriver::createTextureViewR(Handle<HwTexture> th,
    assert_invariant(ref);
    ref->count++;

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createTextureViewSwizzleR(Handle<HwTexture> th, Handle<HwTexture> srch,
@@ -1100,7 +1061,7 @@ void OpenGLDriver::createTextureViewSwizzleR(Handle<HwTexture> th, Handle<HwText
    assert_invariant(ref);
    ref->count++;

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createTextureExternalImage2R(Handle<HwTexture> th, SamplerType target,
@@ -1263,7 +1224,7 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
 #endif
    }

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer const* vb) {
@@ -1523,7 +1484,7 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
                // we shouldn't be here
                break;
        }
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    } else
 #ifndef __EMSCRIPTEN__
 #ifdef GL_EXT_multisampled_render_to_texture
@@ -1544,7 +1505,7 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
            glFramebufferRenderbuffer(GL_FRAMEBUFFER, attachment,
                    GL_RENDERBUFFER, t->gl.id);
        }
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    } else
 #endif // GL_EXT_multisampled_render_to_texture
 #endif // __EMSCRIPTEN__
@@ -1626,13 +1587,13 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
                break;
        }

-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    }

    rt->gl.resolve |= resolveFlags;

-    CHECK_GL_ERROR(utils::slog.e)
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)
+    CHECK_GL_ERROR()
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_FRAMEBUFFER)
 }

 void OpenGLDriver::renderBufferStorage(GLuint rbo, GLenum internalformat, uint32_t width, // NOLINT(readability-convert-member-functions-to-static)
@@ -1661,7 +1622,7 @@ void OpenGLDriver::renderBufferStorage(GLuint rbo, GLenum internalformat, uint32
    // unbind the renderbuffer, to avoid any later confusion
    glBindRenderbuffer(GL_RENDERBUFFER, 0);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createDefaultRenderTargetR(
@@ -1758,7 +1719,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
        if (UTILS_LIKELY(!getContext().isES2())) {
            glDrawBuffers((GLsizei)maxDrawBuffers, bufs);
        }
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    }
 #endif

@@ -1798,7 +1759,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
    assert_invariant(any(targets & TargetBufferFlags::ALL));
    assert_invariant(tmin == tmax);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::createFenceR(Handle<HwFence> fh, int) {
@@ -1956,25 +1917,6 @@ void OpenGLDriver::destroyRenderPrimitive(Handle<HwRenderPrimitive> rph) {
 void OpenGLDriver::destroyProgram(Handle<HwProgram> ph) {
    DEBUG_MARKER()
    if (ph) {
-        // test ----
-        int16_t returned_value = handle_set_value<OpenGLProgram*>(ph, 101);
-        uint8_t age = (uint8_t) (returned_value >> 8);
-        uint8_t test_value = (returned_value & 0xff);
-        switch (test_value) {
-            case -2:
-            case -1:
-            case 0:
-                utils::slog.w << "@@@ destroyProgram: unexpected result, age:" << (int) age
-                              << ", test_value:" << (int) test_value << utils::io::endl;
-                break;
-            default:
-                FILAMENT_CHECK_POSTCONDITION(test_value == 1)
-                        << "@@@ destroyProgram: shouldn't happen, age:" << (int) age
-                        << ", test_value:" << (int) test_value << utils::io::endl;
-                break;
-        }
-        //----------
-
        OpenGLProgram* p = handle_cast<OpenGLProgram*>(ph);
        destruct(ph, p);
    }
@@ -2168,7 +2110,7 @@ void OpenGLDriver::setAcquiredImage(Handle<HwStream> sh, void* hwbuffer, const m

    if (UTILS_UNLIKELY(glstream->user_thread.pending.image)) {
        scheduleRelease(glstream->user_thread.pending);
-        slog.w << "Acquired image is set more than once per frame." << io::endl;
+        LOG(WARNING) << "Acquired image is set more than once per frame.";
    }

    glstream->user_thread.pending = mPlatform.transformAcquiredImage({
@@ -2599,7 +2541,12 @@ void OpenGLDriver::makeCurrent(Handle<HwSwapChain> schDraw, Handle<HwSwapChain>
            [this](size_t index) {
                for (auto t: mTexturesWithStreamsAttached) {
                    if (t->hwStream->streamType == StreamType::NATIVE) {
-                        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);
+                        }
                        mPlatform.attach(t->hwStream->stream, t->gl.id);
                        mContext.updateTexImage(GL_TEXTURE_EXTERNAL_OES, t->gl.id);
                    }
@@ -2612,7 +2559,7 @@ void OpenGLDriver::makeCurrent(Handle<HwSwapChain> schDraw, Handle<HwSwapChain>

                // OpenGL context has changed, resynchronize the state with the cache
                mContext.synchronizeStateAndCache(index);
-                slog.d << "*** OpenGL context change : " << (index ? "protected" : "default") << io::endl;
+                DLOG(INFO) << "*** OpenGL context change : " << (index ? "protected" : "default");
            });

    mCurrentDrawSwapChain = scDraw;
@@ -2653,7 +2600,7 @@ void OpenGLDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh,
        vb->bufferObjectsVersion = (version + 1) % kMaxVersion;
    }

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::updateIndexBuffer(
@@ -2670,7 +2617,7 @@ void OpenGLDriver::updateIndexBuffer(

    scheduleDestroy(std::move(p));

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh, BufferObjectStreamDescriptor&& streams) {
@@ -2750,7 +2697,7 @@ void OpenGLDriver::updateBufferObject(

    scheduleDestroy(std::move(bd));

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::updateBufferObjectUnsynchronized(
@@ -2797,7 +2744,7 @@ retry:
            scheduleDestroy(std::move(bd));
        }
    }
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 #endif
 }

@@ -2847,7 +2794,7 @@ void OpenGLDriver::generateMipmaps(Handle<HwTexture> th) {

    glGenerateMipmap(t->gl.target);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::setTextureData(GLTexture const* t, uint32_t level,
@@ -2956,7 +2903,7 @@ void OpenGLDriver::setTextureData(GLTexture const* t, uint32_t level,

    scheduleDestroy(std::move(p));

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::setCompressedTextureData(GLTexture const* t, uint32_t level,
@@ -3042,7 +2989,7 @@ void OpenGLDriver::setCompressedTextureData(GLTexture const* t, uint32_t level,

    scheduleDestroy(std::move(p));

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::setupExternalImage2(Platform::ExternalImageHandleRef image) {
@@ -3084,6 +3031,7 @@ 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;
@@ -3112,7 +3060,12 @@ 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;
 }
@@ -3136,8 +3089,14 @@ 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.
@@ -3193,7 +3152,7 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
    TargetBufferFlags discardFlags = params.flags.discardStart & rt->targets;

    GLuint const fbo = gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_FRAMEBUFFER)

    // each render-pass starts with a disabled scissor
    gl.disable(GL_SCISSOR_TEST);
@@ -3205,7 +3164,7 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
        if (attachmentCount) {
            gl.procs.invalidateFramebuffer(GL_FRAMEBUFFER, attachmentCount, attachments.data());
        }
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    } else {
        // It's important to clear the framebuffer before drawing, as it resets
        // the fb to a known state (resets fb compression and possibly other things).
@@ -3292,7 +3251,7 @@ void OpenGLDriver::endRenderPass(int) {
            if (attachmentCount) {
                gl.procs.invalidateFramebuffer(GL_FRAMEBUFFER, attachmentCount, attachments.data());
            }
-           CHECK_GL_ERROR(utils::slog.e)
+           CHECK_GL_ERROR()
        }
    }

@@ -3338,13 +3297,13 @@ void OpenGLDriver::resolvePass(ResolveAction action, GLRenderTarget const* rt,
        gl.bindFramebuffer(GL_READ_FRAMEBUFFER, read);
        gl.bindFramebuffer(GL_DRAW_FRAMEBUFFER, draw);

-        CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_READ_FRAMEBUFFER)
-        CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_DRAW_FRAMEBUFFER)
+        CHECK_GL_FRAMEBUFFER_STATUS(GL_READ_FRAMEBUFFER)
+        CHECK_GL_FRAMEBUFFER_STATUS(GL_DRAW_FRAMEBUFFER)

        gl.disable(GL_SCISSOR_TEST);
        glBlitFramebuffer(0, 0, (GLint)rt->width, (GLint)rt->height,
                0, 0, (GLint)rt->width, (GLint)rt->height, mask, GL_NEAREST);
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    }
 #endif
 }
@@ -3523,7 +3482,7 @@ void OpenGLDriver::readPixels(Handle<HwRenderTarget> src,
        if (buffer) {
            gl.bindFramebuffer(GL_FRAMEBUFFER, s->gl.fbo_read ? s->gl.fbo_read : s->gl.fbo);
            glReadPixels(GLint(x), GLint(y), GLint(width), GLint(height), glFormat, glType, buffer);
-            CHECK_GL_ERROR(utils::slog.e)
+            CHECK_GL_ERROR()

            // now we need to flip the buffer vertically to match our API
            size_t const stride = p.stride ? p.stride : width;
@@ -3555,7 +3514,7 @@ void OpenGLDriver::readPixels(Handle<HwRenderTarget> src,
    glBufferData(GL_PIXEL_PACK_BUFFER, pboSize, nullptr, GL_STATIC_DRAW);
    glReadPixels(GLint(x), GLint(y), GLint(width), GLint(height), glFormat, glType, nullptr);
    gl.bindBuffer(GL_PIXEL_PACK_BUFFER, 0);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    // we're forced to make a copy on the heap because otherwise it deletes std::function<> copy
    // constructor.
@@ -3595,7 +3554,7 @@ void OpenGLDriver::readPixels(Handle<HwRenderTarget> src,
        glDeleteBuffers(1, &pbo);
        scheduleDestroy(std::move(p));
        delete pUserBuffer;
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    });
 #endif
 }
@@ -3620,7 +3579,7 @@ void OpenGLDriver::readBufferSubData(BufferObjectHandle boh,
        glCopyBufferSubData(bo->gl.binding, GL_PIXEL_PACK_BUFFER, offset, 0, size);
        gl.bindBuffer(bo->gl.binding, 0);
        gl.bindBuffer(GL_PIXEL_PACK_BUFFER, 0);
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()

        // then, we schedule a mapBuffer of the PBO later, once the fence has signaled
        auto* pUserBuffer = new BufferDescriptor(std::move(p));
@@ -3637,7 +3596,7 @@ void OpenGLDriver::readBufferSubData(BufferObjectHandle boh,
            glDeleteBuffers(1, &pbo);
            scheduleDestroy(std::move(p));
            delete pUserBuffer;
-            CHECK_GL_ERROR(utils::slog.e)
+            CHECK_GL_ERROR()
        });
    } else {
        gl.bindBuffer(bo->gl.binding, bo->gl.id);
@@ -3650,7 +3609,7 @@ void OpenGLDriver::readBufferSubData(BufferObjectHandle boh,
        }
        gl.bindBuffer(bo->gl.binding, 0);
        scheduleDestroy(std::move(p));
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
    }
 #endif
 }
@@ -3680,7 +3639,7 @@ void OpenGLDriver::whenFrameComplete(const std::function<void()>& fn) noexcept {
 void OpenGLDriver::whenGpuCommandsComplete(const std::function<void()>& fn) noexcept {
    GLsync sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
    mGpuCommandCompleteOps.emplace_back(sync, fn);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::executeGpuCommandsCompleteOps() noexcept {
@@ -3898,7 +3857,7 @@ void OpenGLDriver::clearWithRasterPipe(TargetBufferFlags clearFlags,
        }
    }

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void OpenGLDriver::resolve(
@@ -4009,7 +3968,7 @@ void OpenGLDriver::blit(
        case SamplerType::SAMPLER_EXTERNAL:
            break;
    }
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_DRAW_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_DRAW_FRAMEBUFFER)

    gl.bindFramebuffer(GL_READ_FRAMEBUFFER, fbo[1]);
    switch (s->target) {
@@ -4035,14 +3994,14 @@ void OpenGLDriver::blit(
        case SamplerType::SAMPLER_EXTERNAL:
            break;
    }
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_READ_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_READ_FRAMEBUFFER)

    gl.disable(GL_SCISSOR_TEST);
    glBlitFramebuffer(
            srcOrigin.x, srcOrigin.y, srcOrigin.x + size.x, srcOrigin.y + size.y,
            dstOrigin.x, dstOrigin.y, dstOrigin.x + size.x, dstOrigin.y + size.y,
            mask, GL_NEAREST);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    gl.unbindFramebuffer(GL_DRAW_FRAMEBUFFER);
    gl.unbindFramebuffer(GL_READ_FRAMEBUFFER);
@@ -4105,15 +4064,15 @@ void OpenGLDriver::blitDEPRECATED(TargetBufferFlags buffers,
    gl.bindFramebuffer(GL_READ_FRAMEBUFFER, s->gl.fbo);
    gl.bindFramebuffer(GL_DRAW_FRAMEBUFFER, d->gl.fbo);

-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_READ_FRAMEBUFFER)
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_DRAW_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_READ_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_DRAW_FRAMEBUFFER)

    gl.disable(GL_SCISSOR_TEST);
    glBlitFramebuffer(
            srcRect.left, srcRect.bottom, srcRect.right(), srcRect.top(),
            dstRect.left, dstRect.bottom, dstRect.right(), dstRect.top(),
            GL_COLOR_BUFFER_BIT, glFilterMode);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 #endif
 }

@@ -4123,24 +4082,6 @@ void OpenGLDriver::bindPipeline(PipelineState const& state) {
    setRasterState(state.rasterState);
    setStencilState(state.stencilState);
    gl.polygonOffset(state.polygonOffset.slope, state.polygonOffset.constant);
-    // test ----
-    Handle<HwProgram> pro = state.program;
-    int16_t returned_value = handle_get_value<OpenGLProgram*>(pro);
-    uint8_t age = (uint8_t)(returned_value >> 8);
-    uint8_t test_value = (returned_value & 0xff);
-    switch (test_value) {
-    case -2:
-    case -1:
-    case 0:
-        utils::slog.w << "@@@ bindPipeline: unexpected result, age:" << (int)age << ", test_value:" << (int)test_value << utils::io::endl;
-        break;
-    default:
-        if (test_value != 100) {
-            utils::slog.w << "@@@ bindPipeline: program is not valid OpenGLProgram, age:" << (int)age << ", test_value:" << (int)test_value << utils::io::endl;
-        }
-        break;
-    }
-    //----------
    OpenGLProgram* const p = handle_cast<OpenGLProgram*>(state.program);
    mValidProgram = useProgram(p);
    (*mCurrentPushConstants) = p->getPushConstants();
@@ -4258,9 +4199,9 @@ void OpenGLDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
 #endif

 #if FILAMENT_ENABLE_MATDBG
-    CHECK_GL_ERROR_NON_FATAL(utils::slog.e)
+    CHECK_GL_ERROR_NON_FATAL()
 #else
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 #endif
 }

@@ -4290,9 +4231,9 @@ void OpenGLDriver::draw2GLES2(uint32_t indexOffset, uint32_t indexCount, uint32_
            reinterpret_cast<const void*>(indexOffset << rp->gl.indicesShift));

 #if FILAMENT_ENABLE_MATDBG
-    CHECK_GL_ERROR_NON_FATAL(utils::slog.e)
+    CHECK_GL_ERROR_NON_FATAL()
 #else
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 #endif
 }

@@ -4342,9 +4283,9 @@ void OpenGLDriver::dispatchCompute(Handle<HwProgram> program, uint3 workGroupCou
 #endif // BACKEND_OPENGL_LEVEL_GLES31

 #if FILAMENT_ENABLE_MATDBG
-    CHECK_GL_ERROR_NON_FATAL(utils::slog.e)
+    CHECK_GL_ERROR_NON_FATAL()
 #else
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 #endif
 }

@@ -282,20 +282,6 @@ private:
        return mHandleAllocator.handle_cast<Dp, B>(handle);
    }

-    template<typename Dp, typename B>
-    std::enable_if_t<std::is_pointer_v<Dp> && std::is_base_of_v<B, std::remove_pointer_t<Dp>>,
-            int16_t>
-    handle_set_value(Handle<B>& handle, uint8_t v) {
-        return mHandleAllocator.handle_set_value<Dp, B>(handle, v);
-    }
-
-    template<typename Dp, typename B>
-    std::enable_if_t<std::is_pointer_v<Dp> && std::is_base_of_v<B, std::remove_pointer_t<Dp>>,
-            int16_t>
-    handle_get_value(Handle<B>& handle) {
-        return mHandleAllocator.handle_get_value<Dp, B>(handle);
-    }
-
    friend class OpenGLProgram;
    friend class ShaderCompilerService;

@@ -214,7 +214,7 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
                case DescriptorType::INPUT_ATTACHMENT:
                    break;
            }
-            CHECK_GL_ERROR(utils::slog.e)
+            CHECK_GL_ERROR()
        }
    }

@@ -114,14 +114,14 @@ void TimerQueryNativeFactory::createTimerQuery(GLTimerQuery* tq) {

    tq->state = std::make_shared<GLTimerQuery::State>();
    mContext.procs.genQueries(1u, &tq->state->gl.query);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void TimerQueryNativeFactory::destroyTimerQuery(GLTimerQuery* tq) {
    assert_invariant(tq->state);

    mContext.procs.deleteQueries(1u, &tq->state->gl.query);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    tq->state.reset();
 }
@@ -131,14 +131,14 @@ void TimerQueryNativeFactory::beginTimeElapsedQuery(GLTimerQuery* tq) {

    tq->state->elapsed.store(int64_t(TimerQueryResult::NOT_READY), std::memory_order_relaxed);
    mContext.procs.beginQuery(GL_TIME_ELAPSED, tq->state->gl.query);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 void TimerQueryNativeFactory::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* tq) {
    assert_invariant(tq->state);

    mContext.procs.endQuery(GL_TIME_ELAPSED);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    std::weak_ptr<GLTimerQuery::State> const weak = tq->state;

@@ -153,7 +153,7 @@ void TimerQueryNativeFactory::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQ

        GLuint available = 0;
        context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
-        CHECK_GL_ERROR(utils::slog.e)
+        CHECK_GL_ERROR()
        if (!available) {
            // we need to try this one again later
            return false;
@@ -23,6 +23,7 @@
 #include "OpenGLDriver.h"

 #include <iterator>
+#include <optional>
 #include <private/backend/BackendUtils.h>

 #include <backend/DriverEnums.h>
@@ -30,14 +31,14 @@

 #include <private/utils/Tracing.h>

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

 #include <algorithm>
 #include <array>
@@ -64,9 +65,9 @@ static std::string to_string(bool const b) { return b ? "true" : "false"; }
 static std::string to_string(int const i) { return std::to_string(i); }
 static std::string to_string(float const f) { return "float(" + std::to_string(f) + ")"; }

-static void logCompilationError(io::ostream& out, ShaderStage shaderType, const char* name,
-        GLuint shaderId, CString const& sourceCode) noexcept;
-static void logProgramLinkError(io::ostream& out, char const* name, GLuint program) noexcept;
+static void logCompilationError(ShaderStage shaderType, const char* name, GLuint shaderId,
+        CString const& sourceCode) noexcept;
+static void logProgramLinkError(char const* name, GLuint program) noexcept;

 static void process_GOOGLE_cpp_style_line_directive(OpenGLContext const& context, char* source,
        size_t len) noexcept;
@@ -110,17 +111,33 @@ 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{};
+
+    // Only valid when the blob functions are provided by users. The validity of this variable
+    // doesn't guarantee that the program was created from the cache blob.
    BlobCacheKey key;

    // Used for the `THREAD_POOL` mode.
    mutable Mutex lock;
    mutable Condition cond;
    bool signaled = false;
+
+    // Indicate this program was created from the cache blob.
+    bool retrievedFromBlobCache = false;
 };

 ShaderCompilerService::OpenGLProgramToken::~OpenGLProgramToken() {
-    compiler.submitCallbackHandle(handle);
+    trySubmittingCallback();
 }

 /* static */ void ShaderCompilerService::setUserData(const program_token_t& token,
@@ -248,6 +265,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
    // Try retrieving the cached program blob if available.
    token->gl.program = mBlobCache.retrieve(&token->key, mDriver.mPlatform, program);
    if (token->gl.program) {
+        token->retrievedFromBlobCache = true;
        return token;
    }

@@ -326,12 +344,15 @@ GLuint ShaderCompilerService::getProgram(program_token_t& token) {

    assert_invariant(token);// This function should be called when the token is still alive.

-    if (token->compiler.mMode == Mode::THREAD_POOL) {
-        auto const job = token->compiler.mCompilerThreadPool.dequeue(token);
-        if (!job) {
-            // It's likely that the job was already completed. But it may be still being
-            // executed at this moment. Just try waiting for it to avoid a race.
-            token->wait();
+    // Finalize any pending shader compilation tasks only when the token was created without cache.
+    if (!token->retrievedFromBlobCache) {
+        if (token->compiler.mMode == Mode::THREAD_POOL) {
+            auto const job = token->compiler.mCompilerThreadPool.dequeue(token);
+            if (!job) {
+                // It's likely that the job was already completed. But it may be still being
+                // executed at this moment. Just try waiting for it to avoid a race.
+                token->wait();
+            }
        }
    }

@@ -339,7 +360,7 @@ 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() {
@@ -392,7 +413,7 @@ 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;
 }
@@ -422,8 +443,9 @@ void ShaderCompilerService::ensureTokenIsReady(program_token_t const& token) {
            // just log warnings here instead of repeatedly checking compile status. If this turns
            // out to be a real issue later, we would need to consider doing the canonical way.
            if (!isCompileCompleted(token)) {
-                slog.w << "Shader compilation for OpenGL program " << token->name.c_str_safe()
-                       << " is not completed yet. The following program link may not succeed.";
+                LOG(WARNING)
+                        << "Shader compilation for OpenGL program " << token->name.c_str_safe()
+                        << " is not completed yet. The following program link may not succeed.";
            }

            linkProgram(mDriver.getContext(), token);
@@ -633,8 +655,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
        }
        // Something went wrong. Log the error message.
        const ShaderStage type = static_cast<ShaderStage>(i);
-        logCompilationError(slog.e, type, token->name.c_str_safe(), shader,
-                token->shaderSourceCode[i]);
+        logCompilationError(type, token->name.c_str_safe(), shader, token->shaderSourceCode[i]);
    }
 }

@@ -659,6 +680,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
    }
    glLinkProgram(program);
    token->gl.program = program;
+    token->trySubmittingCallback();
 }

 /* static */ bool ShaderCompilerService::isLinkCompleted(program_token_t const& token) noexcept {
@@ -685,7 +707,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
    glGetProgramiv(token->gl.program, GL_LINK_STATUS, &status);
    if (UTILS_UNLIKELY(status != GL_TRUE)) {
        // Something went wrong. Log the error message.
-        logProgramLinkError(slog.e, token->name.c_str_safe(), token->gl.program);
+        logProgramLinkError(token->name.c_str_safe(), token->gl.program);
        linked = false;
    }
    // No need to keep the shaders around regardless of the result of the program linking.
@@ -735,7 +757,7 @@ void ShaderCompilerService::executeTickOps() noexcept {
 // ------------------------------------------------------------------------------------------------

 UTILS_NOINLINE
-/* static */ void logCompilationError(io::ostream& out, ShaderStage shaderType, const char* name,
+/* static */ void logCompilationError(ShaderStage shaderType, const char* name,
        GLuint const shaderId, UTILS_UNUSED_IN_RELEASE CString const& sourceCode) noexcept {

    { // scope for the temporary string storage
@@ -757,8 +779,9 @@ UTILS_NOINLINE
        CString infoLog(length);
        glGetShaderInfoLog(shaderId, length, nullptr, infoLog.data());

-        out << "Compilation error in " << to_string(shaderType) << " shader \"" << name << "\":\n"
-            << "\"" << infoLog.c_str() << "\"" << io::endl;
+        LOG(ERROR) << "Compilation error in " << to_string(shaderType) << " shader \"" << name
+                   << "\":";
+        LOG(ERROR) << "\"" << infoLog.c_str() << "\"";
    }

 #ifndef NDEBUG
@@ -773,26 +796,25 @@ UTILS_NOINLINE
        } else {
            line = shader.substr(start, end - start);
        }
-        out << lc++ << ":   " << line.c_str() << '\n';
+        LOG(ERROR) << lc++ << ":   " << line.c_str();
        if (end == std::string::npos) {
            break;
        }
        start = end + 1;
    }
-    out << io::endl;
+    LOG(ERROR) << "";
 #endif
 }

 UTILS_NOINLINE
-/* static */ void logProgramLinkError(io::ostream& out, char const* name, GLuint program) noexcept {
+/* static */ void logProgramLinkError(char const* name, GLuint program) noexcept {
    GLint length = 0;
    glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);

    CString infoLog(length);
    glGetProgramInfoLog(program, length, nullptr, infoLog.data());

-    out << "Link error in \"" << name << "\":\n"
-        << "\"" << infoLog.c_str() << "\"" << io::endl;
+    LOG(ERROR) << "Link error in \"" << name << "\":\n" << "\"" << infoLog.c_str() << "\"";
 }

 // If usages of the Google-style line directive are present, remove them, as some
@@ -96,7 +96,7 @@ CocoaExternalImage::SharedGl::~SharedGl() noexcept {
 CocoaExternalImage::CocoaExternalImage(const CVOpenGLTextureCacheRef textureCache,
        const SharedGl &sharedGl) noexcept : mSharedGl(sharedGl), mTextureCache(textureCache) {
    glGenFramebuffers(1, &mFBO);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 CocoaExternalImage::~CocoaExternalImage() noexcept {
@@ -125,7 +125,7 @@ bool CocoaExternalImage::set(CVPixelBufferRef image) noexcept {
    mTexture = createTextureFromImage(image);
    mRgbaTexture = encodeCopyRectangleToTexture2D(CVOpenGLTextureGetName(mTexture),
            CVPixelBufferGetWidth(image), CVPixelBufferGetHeight(image));
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    return true;
 }
@@ -182,33 +182,33 @@ GLuint CocoaExternalImage::encodeCopyRectangleToTexture2D(GLuint rectangle,
    // Create a texture to hold the result of the blit image.
    glBindTexture(GL_TEXTURE_2D, texture);
    glTexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, width, height);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    // source textures
    glBindSampler(0, mSharedGl.sampler);
    glActiveTexture(GL_TEXTURE0);
    glBindTexture(GL_TEXTURE_RECTANGLE, rectangle);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    // destination texture
    glBindFramebuffer(GL_FRAMEBUFFER, mFBO);
    glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_FRAMEBUFFER)
+    CHECK_GL_ERROR()

    // draw
    glViewport(0, 0, width, height);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
    glUseProgram(mSharedGl.program);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
    glDisableVertexAttribArray(0);
    glDrawArrays(GL_TRIANGLES, 0, 3);
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    mState.restore();
-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    return texture;
 }
@@ -119,7 +119,7 @@ CocoaTouchExternalImage::CocoaTouchExternalImage(const CVOpenGLESTextureCacheRef

    glGenFramebuffers(1, &mFBO);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()
 }

 CocoaTouchExternalImage::~CocoaTouchExternalImage() noexcept {
@@ -247,7 +247,7 @@ GLuint CocoaTouchExternalImage::encodeColorConversionPass(GLuint yPlaneTexture,
    glBindTexture(GL_TEXTURE_2D, texture);
    glTexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, width, height);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    // source textures
    glBindSampler(0, mSharedGl.sampler);
@@ -261,8 +261,8 @@ GLuint CocoaTouchExternalImage::encodeColorConversionPass(GLuint yPlaneTexture,
    glBindFramebuffer(GL_FRAMEBUFFER, mFBO);
    glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);

-    CHECK_GL_ERROR(utils::slog.e)
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)
+    CHECK_GL_ERROR()
+    CHECK_GL_FRAMEBUFFER_STATUS(GL_FRAMEBUFFER)

    // geometry
    glBindVertexArray(0);
@@ -275,7 +275,7 @@ GLuint CocoaTouchExternalImage::encodeColorConversionPass(GLuint yPlaneTexture,
    glUseProgram(mSharedGl.program);
    glDrawArrays(GL_TRIANGLES, 0, 3);

-    CHECK_GL_ERROR(utils::slog.e)
+    CHECK_GL_ERROR()

    mState.restore();

@@ -20,9 +20,9 @@

 #include <backend/Platform.h>

+#include <utils/Logger.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
-#include <utils/Log.h>
 #include <utils/ostream.h>

 #if __has_include(<android/surface_texture.h>)
@@ -59,7 +59,7 @@ void ExternalStreamManagerAndroid::destroy(ExternalStreamManagerAndroid* pExtern
 ExternalStreamManagerAndroid::ExternalStreamManagerAndroid() noexcept
        : mVm(VirtualMachineEnv::get()) {
    if (__builtin_available(android 28, *)) {
-        slog.d << "Using ASurfaceTexture" << io::endl;
+        DLOG(INFO) << "Using ASurfaceTexture";
    }
 }

@@ -32,9 +32,9 @@
 #endif
 #include <utils/compiler.h>

-#include <utils/debug.h>
 #include <utils/Invocable.h>
-#include <utils/Log.h>
+#include <utils/Logger.h>
+#include <utils/debug.h>
 #include <utils/ostream.h>

 #include <algorithm>
@@ -74,7 +74,7 @@ void PlatformEGL::logEglError(const char* name) noexcept {
 }

 void PlatformEGL::logEglError(const char* name, EGLint error) noexcept {
-    slog.e << name << " failed with " << getEglErrorName(error) << io::endl;
+    LOG(ERROR) << name << " failed with " << getEglErrorName(error);
 }

 const char* PlatformEGL::getEglErrorName(EGLint error) noexcept {
@@ -101,7 +101,7 @@ void PlatformEGL::clearGlError() noexcept {
    // clear GL error that may have been set by previous calls
    GLenum const error = glGetError();
    if (error != GL_NO_ERROR) {
-        slog.w << "Ignoring pending GL error " << io::hex << error << io::endl;
+        LOG(WARNING) << "Ignoring pending GL error " << io::hex << error;
    }
 }

@@ -142,7 +142,7 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const DriverConfig& drive
    }

    if (UTILS_UNLIKELY(!initialized)) {
-        slog.e << "eglInitialize failed" << io::endl;
+        LOG(ERROR) << "eglInitialize failed";
        return nullptr;
    }

@@ -29,13 +29,12 @@
 #include <android/native_window.h>
 #include <android/hardware_buffer.h>

+#include <utils/Logger.h>
+#include <utils/Panic.h>
 #include <utils/android/PerformanceHintManager.h>
-
+#include <utils/compiler.h>
 #include <utils/compiler.h>
 #include <utils/ostream.h>
-#include <utils/Panic.h>
-#include <utils/Log.h>
-#include <utils/compiler.h>
 #include <utils/ostream.h>

 #include <EGL/egl.h>
@@ -327,7 +326,7 @@ bool PlatformEGLAndroid::setImage(ExternalImageEGLAndroid const* eglExternalImag
            EGL_NATIVE_BUFFER_ANDROID, clientBuffer, imageAttrs);
    if (eglImage == EGL_NO_IMAGE_KHR) {
        // Handle error
-        slog.e << "Failed to create EGL image" << io::endl;
+        LOG(ERROR) << "Failed to create EGL image";
        glDeleteTextures(1, &texture->id);
        return false;
    }
@@ -340,7 +339,7 @@ bool PlatformEGLAndroid::setImage(ExternalImageEGLAndroid const* eglExternalImag
    glBindTexture(texture->target, texture->id);
    GLenum error = glGetError();
    if (UTILS_UNLIKELY(error != GL_NO_ERROR)) {
-        slog.e << "Error after glBindTexture: " << error << io::endl;
+        LOG(ERROR) << "Error after glBindTexture: " << error;
        glDeleteTextures(1, &texture->id);
        eglDestroyImageKHR(eglGetCurrentDisplay(), eglImage);
        glActiveTexture(prevActiveTexture);
@@ -350,7 +349,7 @@ bool PlatformEGLAndroid::setImage(ExternalImageEGLAndroid const* eglExternalImag
    glEGLImageTargetTexture2DOES(texture->target, static_cast<GLeglImageOES>(eglImage));
    error = glGetError();
    if (UTILS_UNLIKELY(error != GL_NO_ERROR)) {
-        slog.e << "Error after glEGLImageTargetTexture2DOES: " << error << io::endl;
+        LOG(ERROR) << "Error after glEGLImageTargetTexture2DOES: " << error;
        glDeleteTextures(1, &texture->id);
        eglDestroyImageKHR(eglGetCurrentDisplay(), eglImage);
        glActiveTexture(prevActiveTexture);
@@ -408,7 +407,7 @@ AcquiredImage PlatformEGLAndroid::transformAcquiredImage(AcquiredImage source) n

    EGLClientBuffer clientBuffer = eglGetNativeClientBufferANDROID(pHardwareBuffer);
    if (!clientBuffer) {
-        slog.e << "Unable to get EGLClientBuffer from AHardwareBuffer." << io::endl;
+        LOG(ERROR) << "Unable to get EGLClientBuffer from AHardwareBuffer.";
        return {};
    }

@@ -427,7 +426,7 @@ AcquiredImage PlatformEGLAndroid::transformAcquiredImage(AcquiredImage source) n
    EGLImageKHR eglImage = eglCreateImageKHR(mEGLDisplay,
            EGL_NO_CONTEXT, EGL_NATIVE_BUFFER_ANDROID, clientBuffer, attributes.data());
    if (eglImage == EGL_NO_IMAGE_KHR) {
-        slog.e << "eglCreateImageKHR returned no image." << io::endl;
+        LOG(ERROR) << "eglCreateImageKHR returned no image.";
        return {};
    }

@@ -442,7 +441,7 @@ AcquiredImage PlatformEGLAndroid::transformAcquiredImage(AcquiredImage source) n
    auto patchedCallback = [](void* image, void* userdata) {
        Closure* closure = (Closure*)userdata;
        if (eglDestroyImageKHR(closure->display, (EGLImageKHR) image) == EGL_FALSE) {
-            slog.e << "eglDestroyImageKHR failed." << io::endl;
+            LOG(ERROR) << "eglDestroyImageKHR failed.";
        }
        closure->acquiredImage.callback(closure->acquiredImage.image, closure->acquiredImage.userData);
        delete closure;
@@ -21,9 +21,9 @@
 #include <EGL/egl.h>
 #include <EGL/eglext.h>

-#include <utils/compiler.h>
-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
+#include <utils/compiler.h>

 using namespace utils;

@@ -42,12 +42,12 @@ backend::Driver* PlatformEGLHeadless::createDriver(void* sharedContext,
        const Platform::DriverConfig& driverConfig) noexcept {
    EGLBoolean bindAPI = eglBindAPI(EGL_OPENGL_API);
    if (UTILS_UNLIKELY(!bindAPI)) {
-        slog.e << "eglBindAPI EGL_OPENGL_API failed" << io::endl;
+        LOG(ERROR) << "eglBindAPI EGL_OPENGL_API failed";
        return nullptr;
    }
    int bindBlueGL = bluegl::bind();
    if (UTILS_UNLIKELY(bindBlueGL != 0)) {
-        slog.e << "bluegl bind failed" << io::endl;
+        LOG(ERROR) << "bluegl bind failed";
        return nullptr;
    }

@@ -16,7 +16,7 @@

 #include <backend/platforms/PlatformGLX.h>

-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>

 #include <X11/Xlib.h>
@@ -84,7 +84,7 @@ static PFNGLXGETPROCADDRESSPROC getProcAddress;
 static bool loadLibraries() {
    g_glx.library = dlopen(LIBRARY_GLX, RTLD_LOCAL | RTLD_NOW);
    if (!g_glx.library) {
-        utils::slog.e << "Could not find library " << LIBRARY_GLX << utils::io::endl;
+        LOG(ERROR) << "Could not find library " << LIBRARY_GLX;
        return false;
    }

@@ -115,7 +115,7 @@ static bool loadLibraries() {

    g_x11.library = dlopen(LIBRARY_X11, RTLD_LOCAL | RTLD_NOW);
    if (!g_x11.library) {
-        utils::slog.e << "Could not find library " << LIBRARY_X11 << utils::io::endl;
+        LOG(ERROR) << "Could not find library " << LIBRARY_X11;
        return false;
    }

@@ -134,7 +134,7 @@ Driver* PlatformGLX::createDriver(void* sharedGLContext,
    // Get the display device
    mGLXDisplay = g_x11.openDisplay(NULL);
    if (mGLXDisplay == nullptr) {
-        utils::slog.e << "Failed to open X display. (exiting)." << utils::io::endl;
+        LOG(ERROR) << "Failed to open X display. (exiting).";
        exit(1);
    }

@@ -145,8 +145,7 @@ Driver* PlatformGLX::createDriver(void* sharedGLContext,

        r = g_glx.queryContext(mGLXDisplay, sharedCtx, GLX_FBCONFIG_ID, &usedFbId);
        if (r != 0) {
-            utils::slog.e << "Failed to get GLX_FBCONFIG_ID from shared GL context."
-                          << utils::io::endl;
+            LOG(ERROR) << "Failed to get GLX_FBCONFIG_ID from shared GL context.";
            return nullptr;
        }

@@ -154,7 +153,7 @@ Driver* PlatformGLX::createDriver(void* sharedGLContext,
        GLXFBConfig* fbConfigs = g_glx.getFbConfigs(mGLXDisplay, 0, &numConfigs);

        if (fbConfigs == nullptr) {
-            utils::slog.e << "Failed to get the available GLXFBConfigs." << utils::io::endl;
+            LOG(ERROR) << "Failed to get the available GLXFBConfigs.";
            return nullptr;
        }

@@ -164,8 +163,7 @@ Driver* PlatformGLX::createDriver(void* sharedGLContext,
        for (int i = 0; i < numConfigs; ++i) {
            r = g_glx.getFbConfigAttrib(mGLXDisplay, fbConfigs[i], GLX_FBCONFIG_ID, &fbId);
            if (r != 0) {
-                utils::slog.e << "Failed to get GLX_FBCONFIG_ID for entry " << i << "."
-                              << utils::io::endl;
+                LOG(ERROR) << "Failed to get GLX_FBCONFIG_ID for entry " << i << ".";
                continue;
            }

@@ -176,8 +174,7 @@ Driver* PlatformGLX::createDriver(void* sharedGLContext,
        }

        if (fbIndex < 0) {
-            utils::slog.e << "Failed to find an `GLXFBConfig` with the requested ID."
-                          << utils::io::endl;
+            LOG(ERROR) << "Failed to find an `GLXFBConfig` with the requested ID.";
            return nullptr;
        }

@@ -202,8 +199,7 @@ Driver* PlatformGLX::createDriver(void* sharedGLContext,
            getProcAddress((GLubyte*)"glXCreateContextAttribsARB");

    if (glXCreateContextAttribs == nullptr) {
-        utils::slog.i << "Unable to retrieve function pointer for `glXCreateContextAttribs()`."
-                      << utils::io::endl;
+        LOG(INFO) << "Unable to retrieve function pointer for `glXCreateContextAttribs()`.";
        return nullptr;
    }

@@ -32,7 +32,7 @@
 #include "GL/glext.h"
 #include "GL/wglext.h"

-#include <utils/Log.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>

 namespace {
@@ -55,8 +55,7 @@ void reportWindowsError(DWORD dwError) {
        0, nullptr
 	);

-    utils::slog.e << "Windows error code: " << dwError << ". " << lpMessageBuffer
-            << utils::io::endl;
+    LOG(ERROR) << "Windows error code: " << dwError << ". " << lpMessageBuffer;

    LocalFree(lpMessageBuffer);
 }
@@ -106,7 +105,7 @@ Driver* PlatformWGL::createDriver(void* sharedGLContext,
    HDC whdc = mWhdc = GetDC(mHWnd);
    if (whdc == NULL) {
        dwError = GetLastError();
-        utils::slog.e << "CreateWindowA() failed" << utils::io::endl;
+        LOG(ERROR) << "CreateWindowA() failed";
        goto error;
    }

@@ -117,8 +116,7 @@ Driver* PlatformWGL::createDriver(void* sharedGLContext,
    tempContext = wglCreateContext(whdc);
    if (!wglMakeCurrent(whdc, tempContext)) {
        dwError = GetLastError();
-        utils::slog.e << "wglMakeCurrent() failed, whdc=" << whdc << ", tempContext=" <<
-                tempContext << utils::io::endl;
+        LOG(ERROR) << "wglMakeCurrent() failed, whdc=" << whdc << ", tempContext=" << tempContext;
        goto error;
    }

@@ -142,7 +140,7 @@ Driver* PlatformWGL::createDriver(void* sharedGLContext,
    }

    if (!mContext) {
-        utils::slog.e << "wglCreateContextAttribs() failed, whdc=" << whdc << utils::io::endl;
+        LOG(ERROR) << "wglCreateContextAttribs() failed, whdc=" << whdc;
        goto error;
    }

@@ -152,8 +150,7 @@ Driver* PlatformWGL::createDriver(void* sharedGLContext,

    if (!wglMakeCurrent(whdc, mContext)) {
        dwError = GetLastError();
-        utils::slog.e << "wglMakeCurrent() failed, whdc=" << whdc << ", mContext=" <<
-                mContext << utils::io::endl;
+        LOG(ERROR) << "wglMakeCurrent() failed, whdc=" << whdc << ", mContext=" << mContext;
        goto error;
    }

@@ -40,7 +40,7 @@ inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, V
        FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                      << " layout=" << src.getLayout()
                      << " to=" << dst.texture->getVkImage() << ",level=" << (int) dst.level
-                      << " layout=" << dst.getLayout() << utils::io::endl;
+                      << " layout=" << dst.getLayout();
    }

    VkImageSubresourceRange const srcRange = src.getSubresourceRange();
@@ -80,7 +80,7 @@ inline void resolveFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect
        FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                      << " layout=" << src.getLayout()
                      << " to=" << dst.texture->getVkImage() << ",level=" << (int) dst.level
-                      << " layout=" << dst.getLayout() << utils::io::endl;
+                      << " layout=" << dst.getLayout();
    }

    VkImageSubresourceRange const srcRange = src.getSubresourceRange();
@@ -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
@@ -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
@@ -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
@@ -14,49 +14,35 @@
 * 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),
+      mBufferCache(bufferCache),
+      mBuffer(mBufferCache.acquire(usage, numBytes)),
      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)));
+      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, 0, 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.
@@ -64,13 +50,13 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
            (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;
        }
@@ -81,20 +67,20 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
            .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .buffer = mGpuBuffer,
+            .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,
+        .srcOffset = stage->offset(),
        .dstOffset = byteOffset,
        .size = numBytes,
    };
-    vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);
+    vkCmdCopyBuffer(commands.buffer(), stage->buffer(), getVkBuffer(), 1, &region);

    mUpdatedOffset = byteOffset;
    mUpdatedBytes = numBytes;
@@ -106,16 +92,16 @@ 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) {
+    } 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 (mUsage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT) {
+    } else if (getUsage() == VulkanBufferUsage::SHADER_STORAGE) {
        // TODO: implement me
    }

@@ -125,13 +111,21 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
        .dstAccessMask = dstAccessMask,
        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .buffer = mGpuBuffer,
+        .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
@@ -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
@@ -215,7 +215,7 @@ VkSemaphore VulkanCommandBuffer::submit() {
    }
    FVK_LOGI << ") "
             << " signal=" << mSubmission
-             << " fence=" << mFence << utils::io::endl;
+             << " fence=" << mFence;
 #endif

    mFenceStatus->setStatus(VK_NOT_READY);
@@ -224,7 +224,7 @@ VkSemaphore VulkanCommandBuffer::submit() {

 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
    if (result != VK_SUCCESS) {
-        FVK_LOGD << "Failed command buffer submission result: " << result << utils::io::endl;
+        FVK_LOGD << "Failed command buffer submission result: " << result;
    }
 #endif
    assert_invariant(result == VK_SUCCESS);
@@ -490,7 +490,7 @@ void VulkanCommands::pushGroupMarker(char const* str, VulkanGroupMarkers::Timest
        mProtectedPool->pushMarker(str, timestamp);
    }
 #if FVK_ENABLED(FVK_DEBUG_PRINT_GROUP_MARKERS)
-    FVK_LOGD << "----> " << str << utils::io::endl;
+    FVK_LOGD << "----> " << str;
 #endif
 }

@@ -502,8 +502,7 @@ void VulkanCommands::popGroupMarker() {
    auto const& startTime = ret.second;
    auto const endTime = std::chrono::high_resolution_clock::now();
    std::chrono::duration<double> diff = endTime - startTime;
-    FVK_LOGD << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms"
-             << utils::io::endl;
+    FVK_LOGD << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms";
 #else
    mPool->popMarker();
 #endif // FVK_DEBUG_PRINT_GROUP_MARKERS
@@ -17,7 +17,7 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANCONSTANTS_H
 #define TNT_FILAMENT_BACKEND_VULKANCONSTANTS_H

-#include <utils/Log.h>
+#include <utils/Logger.h>

 #include <stdint.h>

@@ -73,14 +73,16 @@
 #define FVK_DEBUG_RESOURCE_LEAK           0x00010000

 // Set this to enable logging "only" to one output stream. This is useful in the case where we want
-// to debug with print statements and want ordered logging (e.g slog.i and slog.e will not appear in
-// order of calls).
+// to debug with print statements and want ordered logging (e.g LOG(INFO) and LOG(ERROR) will not
+// appear in order of calls).
 #define FVK_DEBUG_FORCE_LOG_TO_I          0x00020000

 // Enable a minimal set of traces to assess the performance of the backend.
 // All other debug features must be disabled.
 #define FVK_DEBUG_PROFILING               0x00040000

+#define FVK_DEBUG_VULKAN_BUFFER_CACHE     0x00080000
+
 // Useful default combinations
 #define FVK_DEBUG_EVERYTHING              (0xFFFFFFFF & ~FVK_DEBUG_PROFILING)
 #define FVK_DEBUG_PERFORMANCE     \
@@ -172,15 +174,15 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 #endif

 #if FVK_ENABLED(FVK_DEBUG_FORCE_LOG_TO_I)
-    #define FVK_LOGI (utils::slog.i)
+    #define FVK_LOGI LOG(INFO)
    #define FVK_LOGD FVK_LOGI
    #define FVK_LOGE FVK_LOGI
    #define FVK_LOGW FVK_LOGI
 #else
-    #define FVK_LOGE (utils::slog.e)
-    #define FVK_LOGW (utils::slog.w)
-    #define FVK_LOGD (utils::slog.d)
-    #define FVK_LOGI (utils::slog.i)
+    #define FVK_LOGE LOG(ERROR)
+    #define FVK_LOGW LOG(WARNING)
+    #define FVK_LOGD DLOG(INFO)
+    #define FVK_LOGI LOG(INFO)
 #endif

 // All vkCreate* functions take an optional allocator. For now we select the default allocator by
@@ -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 {
@@ -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,
    };
@@ -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,
@@ -99,15 +104,15 @@ VKAPI_ATTR VkBool32 VKAPI_CALL debugReportCallback(VkDebugReportFlagsEXT flags,
        VkDebugReportObjectTypeEXT objectType, uint64_t object, size_t location,
        int32_t messageCode, const char* pLayerPrefix, const char* pMessage, void* pUserData) {
    if (flags & VK_DEBUG_REPORT_ERROR_BIT_EXT) {
-        FVK_LOGE << "VULKAN ERROR: (" << pLayerPrefix << ") " << pMessage << utils::io::endl;
+        FVK_LOGE << "VULKAN ERROR: (" << pLayerPrefix << ") " << pMessage;
    } else {
        // TODO: emit best practices warnings about aggressive pipeline barriers.
        if (strstr(pMessage, "ALL_GRAPHICS_BIT") || strstr(pMessage, "ALL_COMMANDS_BIT")) {
            return VK_FALSE;
        }
-        FVK_LOGW << "VULKAN WARNING: (" << pLayerPrefix << ") " << pMessage << utils::io::endl;
+        FVK_LOGW << "VULKAN WARNING: (" << pLayerPrefix << ") " << pMessage;
    }
-    FVK_LOGE << utils::io::endl;
+    FVK_LOGE;
    return VK_FALSE;
 }
 #endif // FVK_ENABLED(FVK_DEBUG_VALIDATION)
@@ -117,18 +122,16 @@ VKAPI_ATTR VkBool32 VKAPI_CALL debugUtilsCallback(VkDebugUtilsMessageSeverityFla
        VkDebugUtilsMessageTypeFlagsEXT types, const VkDebugUtilsMessengerCallbackDataEXT* cbdata,
        void* pUserData) {
    if (severity & VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT) {
-        FVK_LOGE << "VULKAN ERROR: (" << cbdata->pMessageIdName << ") " << cbdata->pMessage
-                 << utils::io::endl;
+        FVK_LOGE << "VULKAN ERROR: (" << cbdata->pMessageIdName << ") " << cbdata->pMessage;
    } else {
        // TODO: emit best practices warnings about aggressive pipeline barriers.
        if (strstr(cbdata->pMessage, "ALL_GRAPHICS_BIT")
                || strstr(cbdata->pMessage, "ALL_COMMANDS_BIT")) {
            return VK_FALSE;
        }
-        FVK_LOGW << "VULKAN WARNING: (" << cbdata->pMessageIdName << ") " << cbdata->pMessage
-                 << utils::io::endl;
+        FVK_LOGW << "VULKAN WARNING: (" << cbdata->pMessageIdName << ") " << cbdata->pMessage;
    }
-    FVK_LOGE << utils::io::endl;
+    FVK_LOGE << "";
    return VK_FALSE;
 }
 #endif // FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
@@ -209,7 +212,8 @@ 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()),
@@ -270,20 +274,18 @@ Driver* VulkanDriver::create(VulkanPlatform* platform, VulkanContext const& cont
    //    VulkanRenderTarget            : 312       few
    // -- less than or equal to 312 bytes

-    FVK_LOGD
-           << "\nVulkanSwapChain: " << sizeof(VulkanSwapChain)
-           << "\nVulkanBufferObject: " << sizeof(VulkanBufferObject)
-           << "\nVulkanVertexBuffer: " << sizeof(VulkanVertexBuffer)
-           << "\nVulkanVertexBufferInfo: " << sizeof(VulkanVertexBufferInfo)
-           << "\nVulkanIndexBuffer: " << sizeof(VulkanIndexBuffer)
-           << "\nVulkanRenderPrimitive: " << sizeof(VulkanRenderPrimitive)
-           << "\nVulkanTexture: " << sizeof(VulkanTexture)
-           << "\nVulkanTimerQuery: " << sizeof(VulkanTimerQuery)
-           << "\nHwStream: " << sizeof(HwStream)
-           << "\nVulkanRenderTarget: " << sizeof(VulkanRenderTarget)
-           << "\nVulkanFence: " << sizeof(VulkanFence)
-           << "\nVulkanProgram: " << sizeof(VulkanProgram)
-           << utils::io::endl;
+    FVK_LOGD << "VulkanSwapChain: " << sizeof(VulkanSwapChain);
+    FVK_LOGD << "VulkanBufferObject: " << sizeof(VulkanBufferObject);
+    FVK_LOGD << "VulkanVertexBuffer: " << sizeof(VulkanVertexBuffer);
+    FVK_LOGD << "VulkanVertexBufferInfo: " << sizeof(VulkanVertexBufferInfo);
+    FVK_LOGD << "VulkanIndexBuffer: " << sizeof(VulkanIndexBuffer);
+    FVK_LOGD << "VulkanRenderPrimitive: " << sizeof(VulkanRenderPrimitive);
+    FVK_LOGD << "VulkanTexture: " << sizeof(VulkanTexture);
+    FVK_LOGD << "VulkanTimerQuery: " << sizeof(VulkanTimerQuery);
+    FVK_LOGD << "HwStream: " << sizeof(HwStream);
+    FVK_LOGD << "VulkanRenderTarget: " << sizeof(VulkanRenderTarget);
+    FVK_LOGD << "VulkanFence: " << sizeof(VulkanFence);
+    FVK_LOGD << "VulkanProgram: " << sizeof(VulkanProgram);
 #endif

    assert_invariant(platform);
@@ -328,7 +330,6 @@ void VulkanDriver::terminate() {
    // descriptorSetLayoutCache
    mExternalImageManager.terminate();

-    mStagePool.terminate();
    mPipelineCache.terminate();
    mFramebufferCache.terminate();
    mSamplerCache.terminate();
@@ -339,6 +340,15 @@ 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 +381,7 @@ void VulkanDriver::collectGarbage() {
    mCommands.gc();
    mDescriptorSetCache.gc();
    mStagePool.gc();
+    mBufferCache.gc();
    mFramebufferCache.gc();
    mPipelineCache.gc();

@@ -514,7 +525,7 @@ 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 +542,7 @@ 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();
 }

@@ -784,14 +795,12 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
    mResourceManager.gc();

    if ((flags & backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) != 0 && !isSRGBSwapChainSupported()) {
-        FVK_LOGW << "sRGB swapchain requested, but Platform does not support it"
-                 << utils::io::endl;
+        FVK_LOGW << "sRGB swapchain requested, but Platform does not support it";
        flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
    }
    if (flags & backend::SWAP_CHAIN_CONFIG_PROTECTED_CONTENT) {
        if (!isProtectedContentSupported()) {
-            FVK_LOGW << "protected swapchain requested, but Platform does not support it"
-                << utils::io::endl;
+            FVK_LOGW << "protected swapchain requested, but Platform does not support it";
        }
    }
    auto swapChain = resource_ptr<VulkanSwapChain>::make(&mResourceManager, sch, mPlatform,
@@ -802,8 +811,7 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
 void VulkanDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t width,
        uint32_t height, uint64_t flags) {
    if ((flags & backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) != 0 && !isSRGBSwapChainSupported()) {
-        FVK_LOGW << "sRGB swapchain requested, but Platform does not support it"
-                      << utils::io::endl;
+        FVK_LOGW << "sRGB swapchain requested, but Platform does not support it";
        flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
    }
    assert_invariant(width > 0 && height > 0 && "Vulkan requires non-zero swap chain dimensions.");
@@ -1223,7 +1231,7 @@ 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 +1246,7 @@ 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 +1257,7 @@ 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));
 }

@@ -1291,7 +1299,7 @@ TimerQueryResult VulkanDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint
    uint64_t const end = results.endTime;
    if (begin >= end) {
        // TODO: queries might have ran on different command buffers.
-        FVK_LOGW << "Timestamps are not monotonically increasing. " << utils::io::endl;
+        FVK_LOGW << "Timestamps are not monotonically increasing. ";
        *elapsedTime = 0;
        return TimerQueryResult::ERROR;
    }
@@ -1622,8 +1630,8 @@ 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 +1906,7 @@ 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);
 }

@@ -2029,7 +2037,7 @@ void VulkanDriver::debugCommandBegin(CommandStream* cmds, bool synchronous, cons
        assert_invariant(inRenderPass);
        inRenderPass = false;
    } else if (inRenderPass && OUTSIDE_COMMANDS.find(command) != OUTSIDE_COMMANDS.end()) {
-        FVK_LOGE << command.data() << " issued inside a render pass." << utils::io::endl;
+        FVK_LOGE << command.data() << " issued inside a render pass.";
    }
 #endif
 }
@@ -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;
@@ -100,8 +100,7 @@ VkFramebuffer VulkanFboCache::getFramebuffer(FboKey const& config) noexcept {
        << "for render pass " << config.renderPass << ", "
        << "samples = " << int(config.samples) << ", "
        << "depth = " << (config.depth ? 1 : 0) << ", "
-        << "attachmentCount = " << attachmentCount
-        << utils::io::endl;
+        << "attachmentCount = " << attachmentCount;
    #endif

    VkFramebufferCreateInfo info {
@@ -341,8 +340,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey const& config) noexcept
             << "needsResolveMask = " << int(config.needsResolveMask) << ", "
             << "usesLazilyAllocatedMemory = " << int(config.usesLazilyAllocatedMemory) << ", "
             << "viewCount = " << int(config.viewCount) << ", "
-             << "colorAttachmentCount[0] = " << subpasses[0].colorAttachmentCount
-             << utils::io::endl;
+             << "colorAttachmentCount[0] = " << subpasses[0].colorAttachmentCount;
 #endif

    return renderPass;
@@ -287,7 +287,7 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept

 #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
    FVK_LOGD << "Created VulkanProgram " << builder << ", shaders = (" << modules[0]
-             << ", " << modules[1] << ")" << utils::io::endl;
+             << ", " << modules[1] << ")";
 #endif
 }

@@ -583,22 +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) {}

@@ -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
@@ -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
@@ -218,8 +218,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n

    #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
        FVK_LOGD << "vkCreateGraphicsPipelines with shaders = ("
-                 << shaderStages[0].module << ", " << shaderStages[1].module << ")"
-                 << utils::io::endl;
+                 << shaderStages[0].module << ", " << shaderStages[1].module << ")";
    #endif
    PipelineCacheEntry cacheEntry = {
        .lastUsed = mCurrentTime,
@@ -228,7 +227,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
            VKALLOC, &cacheEntry.handle);
    assert_invariant(error == VK_SUCCESS);
    if (error != VK_SUCCESS) {
-        FVK_LOGE << "vkCreateGraphicsPipelines error " << error << utils::io::endl;
+        FVK_LOGE << "vkCreateGraphicsPipelines error " << error;
        return nullptr;
    }
    return &mPipelines.emplace(mPipelineRequirements, cacheEntry).first.value();
@@ -242,7 +241,7 @@ void VulkanPipelineCache::bindProgram(fvkmemory::resource_ptr<VulkanProgram> pro
 #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
    if (mPipelineRequirements.shaders[0] == VK_NULL_HANDLE ||
            mPipelineRequirements.shaders[1] == VK_NULL_HANDLE) {
-        FVK_LOGE << "Binding missing shader: " << program->name.c_str() << utils::io::endl;
+        FVK_LOGE << "Binding missing shader: " << program->name.c_str();
    }
 #endif
 }
@@ -42,7 +42,7 @@ fvkmemory::resource_ptr<VulkanTimerQuery> VulkanQueryManager::getNextQuery(
        fvkmemory::ResourceManager* resourceManager) {
    auto unused = ~mUsed;
    if (unused.empty()) {
-        FVK_LOGE << "More than " << mUsed.size() << " timers are not supported." << utils::io::endl;
+        FVK_LOGE << "More than " << mUsed.size() << " timers are not supported.";
        return {};
    }

@@ -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;
@@ -171,7 +171,7 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
 #if FVK_ENABLED(FVK_DEBUG_READ_PIXELS)
    FVK_LOGD << "readPixels created image=" << stagingImage
             << " to copy from image=" << srcTexture->getVkImage()
-             << " src-layout=" << srcTexture->getLayout(0, 0) << utils::io::endl;
+             << " src-layout=" << srcTexture->getLayout(0, 0);
 #endif

    VkMemoryRequirements memReqs;
@@ -188,28 +188,27 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
        memoryTypeIndex = selectMemoryFunc(memReqs.memoryTypeBits,
                VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);
        FVK_LOGW
-                << "readPixels is slow because VK_MEMORY_PROPERTY_HOST_CACHED_BIT is not available"
-                << utils::io::endl;
+                << "readPixels is slow because VK_MEMORY_PROPERTY_HOST_CACHED_BIT is not available";
    }

    FILAMENT_CHECK_POSTCONDITION(memoryTypeIndex < VK_MAX_MEMORY_TYPES)
            << "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);

@@ -304,9 +303,8 @@ 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;
+            FVK_LOGE << "Failed to wait for readPixels fence";
            return;
        }

@@ -324,7 +322,7 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
                    fvkutils::getComponentCount(srcFormat), srcPixels,
                    static_cast<int>(subResourceLayout.rowPitch), static_cast<int>(width),
                    static_cast<int>(height), swizzle)) {
-            FVK_LOGE << "Unsupported PixelDataFormat or PixelDataType" << utils::io::endl;
+            FVK_LOGE << "Unsupported PixelDataFormat or PixelDataType";
        }

        vkUnmapMemory(device, stagingMemory);
@@ -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;
+        FVK_LOGE << "Allocation error: " << result;
+    } else {
+        FVK_LOGD << "Allocated stage with hndl " << buffer;
    }
 #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,
@@ -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);
@@ -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;
@@ -25,6 +25,7 @@
 #include <backend/DriverEnums.h>
 #include <private/backend/BackendUtils.h>

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

 using namespace bluevk;
@@ -187,8 +188,8 @@ 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 " << vkFormat << " is not "
-                        "sampleable with optimal tiling." << utils::io::endl;
+                FVK_LOGW << "Texture usage is SAMPLEABLE but format " << vkFormat
+                         << " is not sampleable with optimal tiling.";
            }
        }
 #endif
@@ -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,
@@ -326,7 +361,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
            imageInfo.flags = VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT;
        } else {
            FVK_LOGW << "Note: creating 2D views on 3D image is not available on this platform. "
-                     << "i.e. we cannot render to slices of a 3D image" << utils::io::endl;
+                     << "i.e. we cannot render to slices of a 3D image";
        }
    } else if (target == SamplerType::SAMPLER_CUBEMAP) {
        imageInfo.arrayLayers = 6;
@@ -385,7 +420,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
            << "type = " << imageInfo.imageType << ", "
            << "flags = " << imageInfo.flags << ", "
            << "target = " << static_cast<int>(target) <<", "
-            << "format = " << vkFormat << utils::io::endl;
+            << "format = " << vkFormat;
    }
    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS) << "Unable to create image."
                                                       << " error=" << static_cast<int32_t>(result);
@@ -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();
@@ -681,14 +731,14 @@ bool VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, VkImageSubresourceR
                 << range.levelCount << ")" << " from=" << oldLayout << " to=" << newLayout
                 << " format=" << mState->mVkFormat << " depth="
                 << fvkutils::isVkDepthFormat(mState->mVkFormat)
-                 << " slice-by-slice=" << transitionSliceBySlice << utils::io::endl;
+                 << " slice-by-slice=" << transitionSliceBySlice;
 #endif
    } else {
 #if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
        FVK_LOGD << "transition texture=" << mState->mTextureImage << " (" << range.baseArrayLayer
                 << "," << range.baseMipLevel << ")" << " count=(" << range.layerCount << ","
                 << range.levelCount << ")" << " to=" << newLayout
-                 << " is skipped because of no change in layout" << utils::io::endl;
+                 << " is skipped because of no change in layout";
 #endif
    }

@@ -799,8 +849,7 @@ void VulkanTexture::print() const {
                level < (mPrimaryViewRange.baseMipLevel + mPrimaryViewRange.levelCount);
            FVK_LOGD << "[" << mState->mTextureImage << "]: (" << layer << "," << level
                          << ")=" << getLayout(layer, level)
-                          << " primary=" << primary
-                          << utils::io::endl;
+                          << " primary=" << primary;
        }
    }

@@ -809,8 +858,7 @@ void VulkanTexture::print() const {
        FVK_LOGD << "[" << mState->mTextureImage << ", imageView=" << view.second << "]=>"
                      << " (" << range.baseArrayLayer << "," << range.baseMipLevel << ")"
                      << " count=(" << range.layerCount << "," << range.levelCount << ")"
-                      << " aspect=" << range.aspectMask << " viewType=" << view.first.type
-                      << utils::io::endl;
+                      << " aspect=" << range.aspectMask << " viewType=" << view.first.type;
    }
 }
 #endif
@@ -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 "";
    }
@@ -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>
@@ -17,6 +17,7 @@
 #include "vulkan/memory/ResourceManager.h"
 #include "vulkan/VulkanHandles.h"

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

 namespace filament::backend::fvkmemory {
@@ -77,6 +78,9 @@ 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 +108,9 @@ 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;
    }
@@ -121,11 +128,11 @@ void ResourceManager::traceConstruction(ResourceType type, HandleId id) {

 void ResourceManager::print() const noexcept {
 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
-    utils::slog.e << "-------------------" << utils::io::endl;
+    LOG(ERROR) << "-------------------";
    for (size_t i = 0; i < (size_t) ResourceType::UNDEFINED_TYPE; ++i) {
-        utils::slog.e <<"    " << getTypeStr((ResourceType) i) << "=" << COUNTER[i] << utils::io::endl;
+        LOG(ERROR) << "    " << getTypeStr((ResourceType) i) << "=" << COUNTER[i];
    }
-    utils::slog.e << "+++++++++++++++++++" << utils::io::endl;
+    LOG(ERROR) << "+++++++++++++++++++";
 #endif
 }

@@ -25,8 +25,9 @@
 #include "vulkan/utils/Helper.h"

 #include <bluevk/BlueVK.h>
-#include <utils/PrivateImplementation-impl.h>
+#include <utils/Logger.h>
 #include <utils/Panic.h>
+#include <utils/PrivateImplementation-impl.h>

 #define SWAPCHAIN_RET_FUNC(func, handle, ...)                                                      \
    if (mImpl->mSurfaceSwapChains.find(handle) != mImpl->mSurfaceSwapChains.end()) {               \
@@ -92,6 +93,16 @@ 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) {
@@ -104,7 +115,7 @@ void printDeviceInfo(VkInstance instance, VkPhysicalDevice device) {
        chainStruct(&physicalDeviceProperties2, &driverProperties);
        vkGetPhysicalDeviceProperties2(device, &physicalDeviceProperties2);
        FVK_LOGI << "Vulkan device driver: " << driverProperties.driverName << " "
-                      << driverProperties.driverInfo << utils::io::endl;
+                 << driverProperties.driverInfo;
    }

    VkPhysicalDeviceProperties deviceProperties;
@@ -141,7 +152,7 @@ void printDeviceInfo(VkInstance instance, VkPhysicalDevice device) {
                  << "(vendor " << utils::io::hex << vendorID << ", "
                  << "device " << deviceID << ", "
                  << "driver " << driverVersion << ", " << utils::io::dec << "api " << major << "."
-                  << minor << ")" << utils::io::endl;
+                  << minor << ")";
 }

 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
@@ -151,14 +162,20 @@ 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 << " ";
        }
    }
-    FVK_LOGI << utils::io::endl;
+    FVK_LOGI << "";
 }
 #endif

@@ -278,10 +295,10 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
    } else {
 #if defined(__ANDROID__)
        FVK_LOGD << "Validation layers are not available; did you set jniLibs in your "
-                 << "gradle file?" << utils::io::endl;
+                 << "gradle file?";
 #else
        FVK_LOGD << "Validation layer not available; did you install the Vulkan SDK?\n"
-                 << "Please ensure that VK_LAYER_PATH is set correctly." << utils::io::endl;
+                 << "Please ensure that VK_LAYER_PATH is set correctly.";
 #endif // __ANDROID__

    }
@@ -617,7 +634,13 @@ 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);
@@ -962,8 +985,7 @@ SwapChainPtr VulkanPlatform::createSwapChain(void* nativeWindow, uint64_t flags,

    if (flags & backend::SWAP_CHAIN_CONFIG_PROTECTED_CONTENT) {
        if (!mImpl->mContext.mProtectedMemorySupported) {
-            utils::slog.w << "protected swapchain requested, but VulkanPlatform does not support it"
-                << utils::io::endl;
+            LOG(WARNING) << "protected swapchain requested, but VulkanPlatform does not support it";
        }
    }

@@ -163,8 +163,7 @@ VkResult VulkanPlatformSurfaceSwapChain::create() {
    // the number of images, though there may be limits related to the total amount of memory used
    // by presentable images."
    if (maxImageCount != 0 && desiredImageCount > maxImageCount) {
-        FVK_LOGE << "Swap chain does not support " << desiredImageCount << " images."
-                      << utils::io::endl;
+        FVK_LOGE << "Swap chain does not support " << desiredImageCount << " images.";
        desiredImageCount = caps.minImageCount;
    }

@@ -266,8 +265,7 @@ VkResult VulkanPlatformSurfaceSwapChain::create() {
           << "swapchain-size=" << mSwapChainBundle.colors.size() << ", "
           << "identity-transform=" << (caps.currentTransform == 1) << ", "
           << "depth=" << mSwapChainBundle.depthFormat << ", "
-           << "protected=" << mSwapChainBundle.isProtected
-           << io::endl;
+           << "protected=" << mSwapChainBundle.isProtected;

    VkSemaphoreCreateInfo const semaphoreCreateInfo = {
            .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
@@ -293,7 +291,7 @@ VkResult VulkanPlatformSurfaceSwapChain::acquire(VulkanPlatform::ImageSyncData*
    // Users should be notified of a suboptimal surface, but it should not cause a cascade of
    // log messages or a loop of re-creations.
    if (result == VK_SUBOPTIMAL_KHR && !mSuboptimal) {
-        FVK_LOGW << "Vulkan Driver: Suboptimal swap chain." << io::endl;
+        FVK_LOGW << "Vulkan Driver: Suboptimal swap chain.";
        mSuboptimal = true;
    }
    return result;
@@ -315,7 +313,7 @@ VkResult VulkanPlatformSurfaceSwapChain::present(uint32_t index, VkSemaphore fin
    // On Android Q and above, a suboptimal surface is always reported after screen rotation:
    // https://android-developers.googleblog.com/2020/02/handling-device-orientation-efficiently.html
    if (result == VK_SUBOPTIMAL_KHR && !mSuboptimal) {
-        FVK_LOGW << "Vulkan Driver: Suboptimal swap chain." << utils::io::endl;
+        FVK_LOGW << "Vulkan Driver: Suboptimal swap chain.";
        mSuboptimal = true;
    }
    return result;
@@ -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;
@@ -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>
@@ -0,0 +1,5 @@
+WebGPU does not include a trivial way to create mipmaps, and thus we must provide a render pass or compute pass to generate them.
+
+This directory is the result of porting https://github.com/JolifantoBambla/webgpu-spd to C++, which itself is a WebGPU port of https://github.com/GPUOpen-LibrariesAndSDKs/FidelityFX-SDK/blob/main/docs/samples/single-pass-downsampler.md. 
+
+The first version of our port is primarily from a Gemini "Convert this to C++" request, which worked as an MVP with very few changes. Future commits will likely alter this version significantly, but it is being merged in this early state to facilitate rapid iteration. 
@@ -0,0 +1,956 @@
+#include "SpdMipmapGenerator.h"
+#include <sstream>
+#include <stdexcept>
+// C++ port of https://github.com/JolifantoBambla/webgpu-spd for early experiments
+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); // Basic BGL for <= 6 mips
+
+    wgpu::Limits deviceLimits;
+    m_device.GetLimits(&deviceLimits);
+    // Max number of *output* mips we can generate in a single pass.
+    // This is device.maxStorageTexturesPerShaderStage - 1 (for the source texture binding),
+    // further capped at 11 output mips (which means 12 total texture bindings: 1 source + 11
+    // outputs). WebGPU minspec for maxStorageTexturesPerShaderStage is 4, so
+    // deviceLimits.maxStorageTexturesPerShaderStage >= 4. Thus, m_maxMipsPerPass will be at least
+    // std::min(4u - 1, 11u) = 3.
+    m_maxMipsPerPass =
+            std::min(static_cast<uint32_t>(deviceLimits.maxStorageTexturesPerShaderStage) - 1, 11u);
+    m_maxArrayLayers = deviceLimits.maxStorageTexturesPerShaderStage;
+
+
+    // If we can generate more than 6 mips in a pass, we need an advanced BGL
+    // that includes atomic counter and mid-mip buffer for levels > 5.
+    if (m_maxMipsPerPass > 6) {
+        std::vector<wgpu::BindGroupLayoutEntry> entriesAdvanced(3);
+        // Entry 0: Uniform buffer (DownsamplePassMeta)
+        entriesAdvanced[0].binding = 0;
+        entriesAdvanced[0].visibility = wgpu::ShaderStage::Compute;
+        entriesAdvanced[0].buffer.type = wgpu::BufferBindingType::Uniform;
+        entriesAdvanced[0].buffer.minBindingSize = 16; // sizeof(DownsamplePassMeta) effectively
+
+        // Entry 1: Storage buffer (spd_global_counter)
+        entriesAdvanced[1].binding = 1;
+        entriesAdvanced[1].visibility = wgpu::ShaderStage::Compute;
+        entriesAdvanced[1].buffer.type = wgpu::BufferBindingType::Storage;
+        entriesAdvanced[1].buffer.minBindingSize = 4; // sizeof(atomic<u32>)
+
+        // Entry 2: Storage buffer (mip_dst_6_buffer)
+        entriesAdvanced[2].binding = 2;
+        entriesAdvanced[2].visibility = wgpu::ShaderStage::Compute;
+        entriesAdvanced[2].buffer.type = wgpu::BufferBindingType::Storage;
+        // Size for array<array<array<vec4<f32>, 64>, 64>> which is 16 (bytes for vec4f32) * 64 * 64
+        // * arrayLayers. The shader defines it as [slice][uv.y][uv.x], so 64x64 per slice.
+        // minBindingSize here is for one slice as a baseline. Actual buffer will be larger for
+        // multiple array layers.
+        entriesAdvanced[2].buffer.minBindingSize = 16 * 64 * 64;
+
+        wgpu::BindGroupLayoutDescriptor bglDescAdvanced{};
+        bglDescAdvanced.label = "SPD Internal BGL (Advanced)";
+        bglDescAdvanced.entryCount = static_cast<uint32_t>(entriesAdvanced.size());
+        bglDescAdvanced.entries = entriesAdvanced.data();
+        m_internalResourcesBindGroupLayout_Advanced =
+                m_device.CreateBindGroupLayout(&bglDescAdvanced);
+    }
+}
+
+void MipmapGenerator::PreparePipelines(wgpu::TextureFormat format, SPDFilter filter,
+        bool halfPrecision) {
+    SPDScalarType scalarType = SanitizeScalarType(format, halfPrecision);
+    // Prepare for a reasonable number of mips, up to m_maxMipsPerPass
+    // The loop should go from 1 up to and including m_maxMipsPerPass.
+    // If m_maxMipsPerPass is 0 (e.g. maxStorageTexturesPerShaderStage <=1 ), this loop won't run.
+    for (uint32_t i = 1; i <= m_maxMipsPerPass; ++i) {
+        GetOrCreatePipeline({ format, scalarType, filter, i });
+    }
+}
+
+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; // Default to F32
+    if (formatStr.find("sint") != std::string::npos) {
+        texelType = SPDScalarType::I32;
+    } else if (formatStr.find("uint") != std::string::npos) {
+        texelType = SPDScalarType::U32;
+    }
+
+    if (halfPrecision && texelType == SPDScalarType::F32) {
+        if (m_device.HasFeature(wgpu::FeatureName::ShaderF16)) {
+            return SPDScalarType::F16;
+        } else {
+            // Optional: Log a warning here that halfPrecision was requested for a float format
+            // but the device does not support ShaderF16, so F32 will be used.
+        }
+    } else if (halfPrecision && texelType != SPDScalarType::F32) {
+        // Optional: Log a warning here that halfPrecision was requested for a non-float format.
+    }
+    return texelType;
+}
+
+SPDPipeline& MipmapGenerator::GetOrCreatePipeline(const PipelineCacheKey& key) {
+    uint32_t numMipsForPipeline = std::min(key.numMips, m_maxMipsPerPass);
+    if (m_pipelines.find(key) == m_pipelines.end()) {
+        // Create the pipeline
+        SPDPipeline spdPipeline;
+
+        std::vector<wgpu::BindGroupLayoutEntry> mipsBglEntries;
+        // Bindings are 0 for source, 1 to numMipsForPipeline for destinations
+        for (uint32_t i = 0; i <= numMipsForPipeline; ++i) {
+            wgpu::BindGroupLayoutEntry entry{};
+            entry.binding = i;
+            entry.visibility = wgpu::ShaderStage::Compute;
+            if (i == 0) {
+                entry.texture.sampleType = (key.scalarType == SPDScalarType::I32)
+                                                   ? wgpu::TextureSampleType::Sint
+                                           : (key.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 = key.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(key.format, GetFilterCode(key.filter),
+                numMipsForPipeline, key.scalarType);
+        wgpu::ShaderModuleWGSLDescriptor wgslDesc{};
+        wgslDesc.code = shaderCode.c_str();
+
+        wgpu::ShaderModuleDescriptor shaderModuleDesc{};
+        shaderModuleDesc.nextInChain = &wgslDesc;
+        wgpu::ShaderModule shaderModule = m_device.CreateShaderModule(&shaderModuleDesc);
+
+        wgpu::BindGroupLayout chosenInternalBGL =
+                (numMipsForPipeline > 6 && m_internalResourcesBindGroupLayout_Advanced)
+                        ? m_internalResourcesBindGroupLayout_Advanced
+                        : m_internalResourcesBindGroupLayout;
+
+        wgpu::BindGroupLayout bgls[] = { spdPipeline.mipsBindGroupLayout, chosenInternalBGL };
+        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[key] = std::move(spdPipeline);
+    }
+    return m_pipelines[key];
+}
+
+
+void MipmapGenerator::Generate(wgpu::CommandEncoder& commandEncoder, wgpu::Texture srcTexture,
+        const SPDPassConfig& config) {
+    if (config.numMips <= 1 || config.numMips <= config.sourceMipLevel + 1) {
+        return;
+    }
+
+    if (m_maxMipsPerPass == 0) {
+        return;
+    }
+
+    PreparePipelines(srcTexture.GetFormat(), config.filter, config.halfPrecision);
+
+    uint32_t totalSrcArrayLayers = srcTexture.GetDepthOrArrayLayers();
+
+    // Loop over array layer chunks
+    for (uint32_t currentBaseArrayLayer = 0; currentBaseArrayLayer < totalSrcArrayLayers;
+            currentBaseArrayLayer += m_maxArrayLayers) {
+
+        uint32_t arrayLayersThisPass =
+                std::min(m_maxArrayLayers, totalSrcArrayLayers - currentBaseArrayLayer);
+
+        uint32_t mipsToGenerateCount = config.numMips - (config.sourceMipLevel + 1);
+        uint32_t currentSourceMipLevel = config.sourceMipLevel;
+
+        while (mipsToGenerateCount > 0) {
+            uint32_t mipsThisPass = std::min(mipsToGenerateCount, m_maxMipsPerPass);
+
+            SPDPassConfig passSpecificConfig = {};
+            passSpecificConfig.filter = config.filter;
+            passSpecificConfig.halfPrecision = config.halfPrecision;
+            passSpecificConfig.targetTexture =
+                    config.targetTexture ? config.targetTexture : srcTexture;
+            passSpecificConfig.numMips = mipsThisPass;
+            passSpecificConfig.sourceMipLevel = currentSourceMipLevel;
+
+            // Pass baseArrayLayer and numArrayLayers directly to generatePass
+            generatePass(commandEncoder, srcTexture, passSpecificConfig, currentBaseArrayLayer,
+                    arrayLayersThisPass);
+
+            mipsToGenerateCount -= mipsThisPass;
+            currentSourceMipLevel += mipsThisPass;
+        }
+    }
+}
+
+void MipmapGenerator::generatePass(wgpu::CommandEncoder& commandEncoder, wgpu::Texture srcTexture,
+        const SPDPassConfig& passConfig, uint32_t baseArrayLayer, uint32_t numArrayLayers) {
+    uint32_t width = srcTexture.GetWidth() >> passConfig.sourceMipLevel;
+    uint32_t height = srcTexture.GetHeight() >> passConfig.sourceMipLevel;
+
+    uint32_t arrayLayerCountForThisPass = numArrayLayers;
+
+    wgpu::Texture target = passConfig.targetTexture ? passConfig.targetTexture : srcTexture;
+    uint32_t numMipsThisPass = passConfig.numMips;
+
+    if (numMipsThisPass == 0) return;
+
+    SPDScalarType scalarType = SanitizeScalarType(srcTexture.GetFormat(), passConfig.halfPrecision);
+    SPDPipeline& spdPipeline = GetOrCreatePipeline(
+            { target.GetFormat(), scalarType, passConfig.filter, numMipsThisPass });
+
+    // --- Create Bind Group 0 (Mips) ---
+    std::vector<wgpu::BindGroupEntry> mipEntries;
+
+    wgpu::TextureViewDescriptor srcViewDesc{};
+    srcViewDesc.dimension = wgpu::TextureViewDimension::e2DArray;
+    srcViewDesc.baseMipLevel = passConfig.sourceMipLevel;
+    srcViewDesc.mipLevelCount = 1;
+    srcViewDesc.baseArrayLayer = baseArrayLayer;
+    srcViewDesc.arrayLayerCount = arrayLayerCountForThisPass;
+
+    wgpu::BindGroupEntry srcEntry{};
+    srcEntry.binding = 0;
+    srcEntry.textureView = srcTexture.CreateView(&srcViewDesc);
+    mipEntries.push_back(srcEntry);
+
+    for (uint32_t i = 0; i < numMipsThisPass; ++i) {
+        wgpu::TextureViewDescriptor dstViewDesc{};
+        dstViewDesc.dimension = wgpu::TextureViewDimension::e2DArray;
+        dstViewDesc.baseMipLevel = passConfig.sourceMipLevel + i + 1;
+        dstViewDesc.mipLevelCount = 1;
+        dstViewDesc.baseArrayLayer = baseArrayLayer;
+        dstViewDesc.arrayLayerCount = arrayLayerCountForThisPass;
+
+        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 widthDiv64 = (width + 63) / 64;
+    uint32_t heightDiv64 = (height + 63) / 64;
+    uint32_t numWorkGroups = widthDiv64 * heightDiv64;
+
+    struct DownsamplePassMeta {
+        uint32_t work_group_offset[2] = { 0, 0 };
+        uint32_t num_work_groups;
+        uint32_t mips;
+        uint32_t padding[12];
+    } meta;
+    meta.num_work_groups = numWorkGroups;
+    meta.mips = numMipsThisPass;
+
+    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::BindGroup internalBindGroup;
+    wgpu::BindGroupLayout chosenInternalBGL =
+            (numMipsThisPass > 6 && m_internalResourcesBindGroupLayout_Advanced)
+                    ? m_internalResourcesBindGroupLayout_Advanced
+                    : m_internalResourcesBindGroupLayout;
+
+    if (numMipsThisPass > 6 && m_internalResourcesBindGroupLayout_Advanced) {
+        wgpu::BufferDescriptor atomicCounterBufferDesc{};
+        atomicCounterBufferDesc.size = sizeof(uint32_t) * arrayLayerCountForThisPass;
+        atomicCounterBufferDesc.usage = wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopyDst;
+        wgpu::Buffer atomicCounterBuffer = m_device.CreateBuffer(&atomicCounterBufferDesc);
+
+        commandEncoder.ClearBuffer(atomicCounterBuffer, 0, WGPU_WHOLE_SIZE);
+
+        wgpu::BufferDescriptor midMipBufferDesc{};
+        midMipBufferDesc.size = 64 * 64 * 16 * arrayLayerCountForThisPass;
+        midMipBufferDesc.usage = wgpu::BufferUsage::Storage;
+        wgpu::Buffer midMipBuffer = m_device.CreateBuffer(&midMipBufferDesc);
+
+        std::vector<wgpu::BindGroupEntry> internalEntries(3);
+        internalEntries[0].binding = 0;
+        internalEntries[0].buffer = metaBuffer;
+        internalEntries[0].size = sizeof(DownsamplePassMeta);
+
+        internalEntries[1].binding = 1;
+        internalEntries[1].buffer = atomicCounterBuffer;
+        internalEntries[1].size = atomicCounterBuffer.GetSize();
+
+        internalEntries[2].binding = 2;
+        internalEntries[2].buffer = midMipBuffer;
+        internalEntries[2].size = midMipBuffer.GetSize();
+
+        wgpu::BindGroupDescriptor internalBindGroupDesc{};
+        internalBindGroupDesc.layout = chosenInternalBGL;
+        internalBindGroupDesc.entryCount = static_cast<uint32_t>(internalEntries.size());
+        internalBindGroupDesc.entries = internalEntries.data();
+        internalBindGroup = m_device.CreateBindGroup(&internalBindGroupDesc);
+    } else {
+        wgpu::BindGroupEntry metaEntry{};
+        metaEntry.binding = 0;
+        metaEntry.buffer = metaBuffer;
+        metaEntry.size = sizeof(DownsamplePassMeta);
+
+        wgpu::BindGroupDescriptor internalBindGroupDesc{};
+        internalBindGroupDesc.layout = chosenInternalBGL;
+        internalBindGroupDesc.entryCount = 1;
+        internalBindGroupDesc.entries = &metaEntry;
+        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(widthDiv64, heightDiv64, arrayLayerCountForThisPass);
+    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
@@ -0,0 +1,99 @@
+#pragma once
+
+#include <optional>
+#include <string>
+#include <tsl/robin_map.h>
+#include <vector>
+#include <webgpu/webgpu_cpp.h>
+// C++ port of https://github.com/JolifantoBambla/webgpu-spd for early experiments
+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; // For the public API, this is total desired mips. Internally, mips for
+                          // the current pass.
+    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:
+    void generatePass(wgpu::CommandEncoder& commandEncoder, wgpu::Texture srcTexture,
+            const SPDPassConfig& passConfig, uint32_t baseArrayLayer,
+            uint32_t numArrayLayers); // Internal pass config might differ slightly if needed
+
+    wgpu::Device m_device;
+    wgpu::BindGroupLayout m_internalResourcesBindGroupLayout;
+    wgpu::BindGroupLayout m_internalResourcesBindGroupLayout_Advanced = nullptr;
+
+    // Maximum number of mips that can be generated in a single pass.
+    uint32_t m_maxMipsPerPass;
+    uint32_t m_maxArrayLayers;
+
+    // Key for the pipeline cache.
+    struct PipelineCacheKey {
+        wgpu::TextureFormat format;
+        SPDScalarType scalarType;
+        SPDFilter filter;
+        uint32_t numMips;
+
+        bool operator==(const PipelineCacheKey& other) const {
+            return format == other.format && scalarType == other.scalarType &&
+                   filter == other.filter && numMips == other.numMips;
+        }
+    };
+
+    // Hash function for PipelineCacheKey.
+    struct PipelineCacheKeyHash {
+        std::size_t operator()(const PipelineCacheKey& key) const {
+            std::size_t h1 = std::hash<wgpu::TextureFormat>()(key.format);
+            std::size_t h2 = std::hash<SPDScalarType>()(key.scalarType);
+            std::size_t h3 = std::hash<SPDFilter>()(key.filter);
+            std::size_t h4 = std::hash<uint32_t>()(key.numMips);
+            // A simple way to combine hashes.
+            return h1 ^ (h2 << 1) ^ (h3 << 2) ^ (h4 << 3);
+        }
+    };
+
+    // Cached pipelines.
+    tsl::robin_map<PipelineCacheKey, SPDPipeline, PipelineCacheKeyHash> m_pipelines;
+
+    // Helper methods
+    SPDPipeline& GetOrCreatePipeline(const PipelineCacheKey& key);
+    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
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Lukas Herzberger
+
+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 OR 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.
@@ -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
@@ -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
@@ -0,0 +1,49 @@
+/*
+ * 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 "WebGPUBufferObject.h"
+
+#include "WebGPUBufferBase.h"
+
+#include "DriverBase.h"
+#include <backend/DriverEnums.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <cstdint>
+
+namespace filament::backend {
+
+namespace {
+
+[[nodiscard]] constexpr wgpu::BufferUsage getBufferObjectUsage(
+        const BufferObjectBinding bindingType) noexcept {
+    switch (bindingType) {
+        case BufferObjectBinding::VERTEX:         return wgpu::BufferUsage::Vertex;
+        case BufferObjectBinding::UNIFORM:        return wgpu::BufferUsage::Uniform;
+        case BufferObjectBinding::SHADER_STORAGE: return wgpu::BufferUsage::Storage;
+    }
+}
+
+} // namespace
+
+WebGPUBufferObject::WebGPUBufferObject(wgpu::Device const& device,
+        const BufferObjectBinding bindingType, const uint32_t byteCount)
+    : HwBufferObject{ byteCount },
+      WebGPUBufferBase{ device, wgpu::BufferUsage::CopyDst | getBufferObjectUsage(bindingType),
+          byteCount, "buffer_object" } {}
+
+} // namespace filament::backend
@@ -0,0 +1,41 @@
+/*
+ * 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_WEBGPUBUFFEROBJECT_H
+#define TNT_FILAMENT_BACKEND_WEBGPUBUFFEROBJECT_H
+
+#include "WebGPUBufferBase.h"
+
+#include "DriverBase.h"
+
+#include <cstdint>
+
+namespace wgpu {
+class Device;
+}
+
+namespace filament::backend {
+
+enum class BufferObjectBinding : uint8_t;
+
+class WebGPUBufferObject final : public HwBufferObject, public WebGPUBufferBase {
+public:
+    WebGPUBufferObject(wgpu::Device const&, BufferObjectBinding, uint32_t byteCount);
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_WEBGPUBUFFEROBJECT_H
@@ -17,10 +17,12 @@
 #ifndef TNT_FILAMENT_BACKEND_WEBGPUCONSTANTS_H
 #define TNT_FILAMENT_BACKEND_WEBGPUCONSTANTS_H

-#include <utils/Log.h>
+#include <utils/Logger.h>

 #include <cstdint>

+constexpr size_t WEBGPU_BUFFER_SIZE_MODULUS = 4;
+
 // FWGPU is short for Filament WebGPU

 // turn on runtime validation, namely for debugging, that would normally not run (for release)
@@ -30,8 +32,8 @@
 #define FWGPU_PRINT_SYSTEM         0x00000002

 // Set this to enable logging "only" to one output stream. This is useful in the case where we want
-// to debug with print statements and want ordered logging (e.g slog.i and slog.e will not appear in
-// order of calls).
+// to debug with print statements and want ordered logging (e.g LOG(INFO) and LOG(ERROR) will not
+// appear in order of calls).
 #define FWGPU_DEBUG_FORCE_LOG_TO_I 0x00000004

 // Useful default combinations
@@ -52,15 +54,15 @@
 #define FWGPU_ENABLED(flags) (((FWGPU_DEBUG_FLAGS) & (flags)) == (flags))

 #if FWGPU_ENABLED(FWGPU_DEBUG_FORCE_LOG_TO_I)
-    #define FWGPU_LOGI (utils::slog.i)
+    #define FWGPU_LOGI LOG(INFO)
    #define FWGPU_LOGD FWGPU_LOGI
    #define FWGPU_LOGE FWGPU_LOGI
    #define FWGPU_LOGW FWGPU_LOGI
 #else
-    #define FWGPU_LOGE (utils::slog.e)
-    #define FWGPU_LOGW (utils::slog.w)
-    #define FWGPU_LOGD (utils::slog.d)
-    #define FWGPU_LOGI (utils::slog.i)
+    #define FWGPU_LOGE LOG(ERROR)
+    #define FWGPU_LOGW LOG(WARNING)
+    #define FWGPU_LOGD DLOG(INFO)
+    #define FWGPU_LOGI LOG(INFO)
 #endif

 constexpr uint64_t REQUEST_ADAPTER_TIMEOUT_NANOSECONDS =
@@ -0,0 +1,106 @@
+/*
+ * 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 "WebGPUDescriptorSet.h"
+
+#include "WebGPUDescriptorSetLayout.h"
+
+#include <backend/DriverEnums.h>
+
+#include <utils/Panic.h>
+#include <utils/debug.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <algorithm>
+#include <cstdint>
+#include <utility>
+#include <vector>
+
+namespace filament::backend {
+
+namespace {
+
+constexpr uint8_t INVALID_INDEX = MAX_DESCRIPTOR_COUNT + 1;
+
+} // namespace
+
+WebGPUDescriptorSet::WebGPUDescriptorSet(wgpu::BindGroupLayout const& layout,
+        std::vector<WebGPUDescriptorSetLayout::BindGroupEntryInfo> const& bindGroupEntries)
+    : mLayout{ layout },
+      mEntriesWithDynamicOffsetsCount{ static_cast<size_t>(std::count_if(bindGroupEntries.begin(),
+              bindGroupEntries.end(), [](auto const& entry) { return entry.hasDynamicOffset; })) } {
+
+    mEntries.resize(bindGroupEntries.size());
+    for (size_t i = 0; i < bindGroupEntries.size(); ++i) {
+        mEntries[i].binding = bindGroupEntries[i].binding;
+    }
+    // Establish the size of entries based on the layout. This should be reliable and efficient.
+    assert_invariant(INVALID_INDEX > mEntryIndexByBinding.size());
+    for (size_t i = 0; i < mEntryIndexByBinding.size(); i++) {
+        mEntryIndexByBinding[i] = INVALID_INDEX;
+    }
+    for (size_t index = 0; index < mEntries.size(); index++) {
+        wgpu::BindGroupEntry const& entry = mEntries[index];
+        assert_invariant(entry.binding < mEntryIndexByBinding.size());
+        mEntryIndexByBinding[entry.binding] = static_cast<uint8_t>(index);
+    }
+}
+
+void WebGPUDescriptorSet::addEntry(const unsigned int index, wgpu::BindGroupEntry&& entry) {
+    if (mBindGroup) {
+        // We will keep getting hits from future updates, but shouldn't do anything
+        // Filament guarantees this won't change after things have locked.
+        return;
+    }
+    // TODO: Putting some level of trust that Filament is not going to reuse indexes or go past the
+    // layout index for efficiency. Add guards if wrong.
+    FILAMENT_CHECK_POSTCONDITION(index < mEntryIndexByBinding.size())
+            << "impossible/invalid index for a descriptor/binding (our of range or >= "
+               "MAX_DESCRIPTOR_COUNT) "
+            << index;
+    uint8_t entryIndex = mEntryIndexByBinding[index];
+    FILAMENT_CHECK_POSTCONDITION(entryIndex != INVALID_INDEX && entryIndex < mEntries.size())
+            << "Invalid binding " << index;
+    entry.binding = index;
+    mEntries[entryIndex] = std::move(entry);
+}
+
+wgpu::BindGroup WebGPUDescriptorSet::lockAndReturn(wgpu::Device const& device) {
+    if (mBindGroup) {
+        return mBindGroup;
+    }
+    // TODO label? Should we just copy layout label?
+    const wgpu::BindGroupDescriptor descriptor{
+        .layout = mLayout,
+        .entryCount = mEntries.size(),
+        .entries = mEntries.data()
+    };
+    mBindGroup = device.CreateBindGroup(&descriptor);
+    FILAMENT_CHECK_POSTCONDITION(mBindGroup) << "Failed to create bind group?";
+    // once we have created the bind group itself we should no longer need any other state
+    mLayout = nullptr;
+    mEntries.clear();
+    mEntries.shrink_to_fit();
+    return mBindGroup;
+}
+
+WebGPUDescriptorSet::~WebGPUDescriptorSet() {
+    mBindGroup = nullptr;
+    mLayout = nullptr;
+}
+
+} // namespace filament::backend
@@ -0,0 +1,62 @@
+/*
+ * 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_WEBGPUDESCRIPTORSET_H
+#define TNT_FILAMENT_BACKEND_WEBGPUDESCRIPTORSET_H
+
+#include "WebGPUDescriptorSetLayout.h"
+
+#include "DriverBase.h"
+#include <backend/DriverEnums.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <array>
+#include <cstdint>
+#include <vector>
+
+namespace filament::backend {
+
+class WebGPUDescriptorSet final : public HwDescriptorSet {
+public:
+    WebGPUDescriptorSet(wgpu::BindGroupLayout const&,
+            std::vector<WebGPUDescriptorSetLayout::BindGroupEntryInfo> const&);
+    ~WebGPUDescriptorSet();
+
+    void addEntry(unsigned int index, wgpu::BindGroupEntry&& entry);
+
+    [[nodiscard]] wgpu::BindGroup lockAndReturn(wgpu::Device const&);
+
+    [[nodiscard]] bool getIsLocked() const { return mBindGroup != nullptr; }
+
+    [[nodiscard]] size_t getEntitiesWithDynamicOffsetsCount() const {
+        return mEntriesWithDynamicOffsetsCount;
+    }
+
+    // May be nullptr. Use lockAndReturn to create the bind group when appropriate
+    [[nodiscard]] wgpu::BindGroup const& getBindGroup() const { return mBindGroup; }
+
+private:
+    wgpu::BindGroupLayout mLayout = nullptr;
+    std::array<uint8_t, MAX_DESCRIPTOR_COUNT> mEntryIndexByBinding{};
+    std::vector<wgpu::BindGroupEntry> mEntries;
+    const size_t mEntriesWithDynamicOffsetsCount;
+    wgpu::BindGroup mBindGroup = nullptr;
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_WEBGPUDESCRIPTORSET_H
@@ -0,0 +1,190 @@
+/*
+ * 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 "WebGPUDescriptorSetLayout.h"
+#include "WebGPUConstants.h"
+
+#include <backend/DriverEnums.h>
+
+#include <utils/BitmaskEnum.h>
+#include <utils/CString.h>
+#include <utils/Panic.h>
+#include <utils/StaticString.h>
+#include <utils/debug.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <algorithm>
+#include <string>
+#include <variant>
+#include <vector>
+
+namespace filament::backend {
+
+namespace {
+
+// Convert Filament Shader Stage Flags bitmask to webgpu equivalent
+[[nodiscard]] wgpu::ShaderStage filamentStageToWGPUStage(const ShaderStageFlags fFlags) {
+    wgpu::ShaderStage retStages = wgpu::ShaderStage::None;
+    if (any(ShaderStageFlags::VERTEX & fFlags)) {
+        retStages |= wgpu::ShaderStage::Vertex;
+    }
+    if (any(ShaderStageFlags::FRAGMENT & fFlags)) {
+        retStages |= wgpu::ShaderStage::Fragment;
+    }
+    if (any(ShaderStageFlags::COMPUTE & fFlags)) {
+        retStages |= wgpu::ShaderStage::Compute;
+    }
+    return retStages;
+}
+
+} // namespace
+
+WebGPUDescriptorSetLayout::WebGPUDescriptorSetLayout(DescriptorSetLayout const& layout,
+        wgpu::Device const& device) {
+    assert_invariant(device);
+
+    std::string baseLabel;
+    if (std::holds_alternative<utils::StaticString>(layout.label)) {
+        const auto& temp = std::get_if<utils::StaticString>(&layout.label);
+        baseLabel = temp->c_str();
+    } else if (std::holds_alternative<utils::CString>(layout.label)) {
+        const auto& temp = std::get_if<utils::CString>(&layout.label);
+        baseLabel = temp->c_str();
+    }
+
+    // TODO: layoutDescriptor has a "Label". Ideally we can get info on what this layout is for
+    // debugging. For now, hack an incrementing value.
+    static int layoutNum = 0;
+
+    const unsigned int samplerCount =
+            std::count_if(layout.bindings.begin(), layout.bindings.end(), [](auto& fEntry) {
+                return DescriptorSetLayoutBinding::isSampler(fEntry.type);
+            });
+
+    std::vector<wgpu::BindGroupLayoutEntry> wEntries;
+    wEntries.reserve(layout.bindings.size() + samplerCount);
+    mBindGroupEntries.reserve(wEntries.capacity());
+
+    for (auto fEntry: layout.bindings) {
+        auto& wEntry = wEntries.emplace_back();
+        auto& entryInfo = mBindGroupEntries.emplace_back();
+        wEntry.visibility = filamentStageToWGPUStage(fEntry.stageFlags);
+        wEntry.binding = fEntry.binding * 2;
+        entryInfo.binding = wEntry.binding;
+
+        switch (fEntry.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_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: {
+                auto& samplerEntry = wEntries.emplace_back();
+                auto& samplerEntryInfo = mBindGroupEntries.emplace_back();
+                samplerEntry.binding = fEntry.binding * 2 + 1;
+                samplerEntryInfo.binding = samplerEntry.binding;
+                samplerEntry.visibility = wEntry.visibility;
+                wEntry.texture.multisampled = isMultiSampledTypeDescriptor(fEntry.type);
+                if (isDepthDescriptor(fEntry.type)) {
+                    samplerEntry.sampler.type = wgpu::SamplerBindingType::Comparison;
+                } else if (isIntDescriptor(fEntry.type) ||
+                        any(fEntry.flags & DescriptorFlags::UNFILTERABLE)) {
+                    samplerEntry.sampler.type = wgpu::SamplerBindingType::NonFiltering;
+                } else {
+                    samplerEntry.sampler.type = wgpu::SamplerBindingType::Filtering;
+                }
+                break;
+            }
+            case DescriptorType::UNIFORM_BUFFER: {
+                wEntry.buffer.hasDynamicOffset =
+                        any(fEntry.flags & DescriptorFlags::DYNAMIC_OFFSET);
+                entryInfo.hasDynamicOffset = wEntry.buffer.hasDynamicOffset;
+                wEntry.buffer.type = wgpu::BufferBindingType::Uniform;
+                // TODO: Ideally we fill minBindingSize
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT: {
+                PANIC_POSTCONDITION("Input Attachment is not supported");
+                break;
+            }
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                PANIC_POSTCONDITION("Shader storage is not supported");
+                break;
+            }
+            case DescriptorType::SAMPLER_EXTERNAL: {
+                PANIC_POSTCONDITION("External Sampler is not supported");
+                break;
+            }
+        }
+        if (isDepthDescriptor(fEntry.type)) {
+            wEntry.texture.sampleType = wgpu::TextureSampleType::Depth;
+        } else if (isFloatDescriptor(fEntry.type)) {
+            if (any(fEntry.flags & DescriptorFlags::UNFILTERABLE)) {
+                wEntry.texture.sampleType = wgpu::TextureSampleType::UnfilterableFloat;
+            } else {
+                wEntry.texture.sampleType = wgpu::TextureSampleType::Float;
+            }
+        } else if (isIntDescriptor(fEntry.type)) {
+            wEntry.texture.sampleType = wgpu::TextureSampleType::Sint;
+        } else if (isUnsignedIntDescriptor(fEntry.type)) {
+            wEntry.texture.sampleType = wgpu::TextureSampleType::Uint;
+        }
+
+        if (is3dTypeDescriptor(fEntry.type)) {
+            wEntry.texture.viewDimension = wgpu::TextureViewDimension::e3D;
+        } else if (is2dTypeDescriptor(fEntry.type)) {
+            wEntry.texture.viewDimension = wgpu::TextureViewDimension::e2D;
+        } else if (is2dArrayTypeDescriptor(fEntry.type)) {
+            wEntry.texture.viewDimension = wgpu::TextureViewDimension::e2DArray;
+        } else if (isCubeTypeDescriptor(fEntry.type)) {
+            wEntry.texture.viewDimension = wgpu::TextureViewDimension::Cube;
+        } else if (isCubeArrayTypeDescriptor(fEntry.type)) {
+            wEntry.texture.viewDimension = wgpu::TextureViewDimension::CubeArray;
+        }
+        // fEntry.count is unused currently
+    }
+    std::string label =  "layout_" + baseLabel + std::to_string(++layoutNum) ;
+    const wgpu::BindGroupLayoutDescriptor layoutDescriptor{
+        .label{label.c_str()}, // Use .c_str() if label needs to be const char*
+        .entryCount = wEntries.size(),
+        .entries = wEntries.data()
+    };
+    mLayout = device.CreateBindGroupLayout(&layoutDescriptor);
+    FILAMENT_CHECK_POSTCONDITION(mLayout)
+            << "Failed to create bind group layout with label " << label;
+}
+
+} // namespace filament::backend
@@ -0,0 +1,53 @@
+/*
+ * 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_WEBGPUDESCRIPTORSETLAYOUT_H
+#define TNT_FILAMENT_BACKEND_WEBGPUDESCRIPTORSETLAYOUT_H
+
+#include "DriverBase.h"
+#include <backend/DriverEnums.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <cstdint>
+#include <vector>
+
+namespace filament::backend {
+
+class WebGPUDescriptorSetLayout final : public HwDescriptorSetLayout {
+public:
+    struct BindGroupEntryInfo final {
+        uint8_t binding = 0;
+        bool hasDynamicOffset = false;
+    };
+
+    WebGPUDescriptorSetLayout(DescriptorSetLayout const&, wgpu::Device const&);
+    ~WebGPUDescriptorSetLayout() = default;
+
+    [[nodiscard]] wgpu::BindGroupLayout const& getLayout() const { return mLayout; }
+
+    [[nodiscard]] std::vector<BindGroupEntryInfo> const& getBindGroupEntries() const {
+        return mBindGroupEntries;
+    }
+
+private:
+    std::vector<BindGroupEntryInfo> mBindGroupEntries;
+    wgpu::BindGroupLayout mLayout = nullptr;
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_WEBGPUDESCRIPTORSETLAYOUT_H
@@ -17,7 +17,7 @@
 #ifndef TNT_FILAMENT_BACKEND_WEBGPUDRIVER_H
 #define TNT_FILAMENT_BACKEND_WEBGPUDRIVER_H

-#include "WebGPUHandles.h"
+#include "WebGPURenderTarget.h"
 #include "webgpu/WebGPUConstants.h"
 #include <backend/platforms/WebGPUPlatform.h>

@@ -29,6 +29,8 @@

 #include <utils/compiler.h>

+#include "SpdMipmapGenerator/SpdMipmapGenerator.h"
+#include <tsl/robin_map.h>
 #include <webgpu/webgpu_cpp.h>

 #include <cstdint>
@@ -64,7 +66,7 @@ private:
    WebGPUPlatform& mPlatform;
    wgpu::Adapter mAdapter = nullptr;
    wgpu::Device mDevice = nullptr;
-    uint32_t mMinUniformBufferOffsetAlignment;
+    wgpu::Limits mDeviceLimits = {};
    wgpu::Queue mQueue = nullptr;
    void* mNativeWindow = nullptr;
    WebGPUSwapChain* mSwapChain = nullptr;
@@ -73,8 +75,9 @@ private:
    wgpu::TextureView mTextureView = nullptr;
    wgpu::RenderPassEncoder mRenderPassEncoder = nullptr;
    wgpu::CommandBuffer mCommandBuffer = nullptr;
-    WGPURenderTarget* mDefaultRenderTarget = nullptr;
-    WGPURenderTarget* mCurrentRenderTarget = nullptr;
+    WebGPURenderTarget* mDefaultRenderTarget = nullptr;
+    WebGPURenderTarget* mCurrentRenderTarget = nullptr;
+    spd::MipmapGenerator mMipMapGenerator;

    tsl::robin_map<uint32_t, wgpu::RenderPipeline> mPipelineMap;

@@ -1,272 +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.
- */
-
-
-#ifndef TNT_FILAMENT_BACKEND_WEBGPUHANDLES_H
-#define TNT_FILAMENT_BACKEND_WEBGPUHANDLES_H
-
-#include "DriverBase.h"
-
-#include <backend/DriverEnums.h>
-#include <backend/Handle.h>
-
-#include <utils/FixedCapacityVector.h>
-
-#include <webgpu/webgpu_cpp.h>
-
-#include <array>
-#include <cstdint>
-#include <vector>
-
-namespace filament::backend {
-
-class WGPUProgram final : public HwProgram {
-public:
-    WGPUProgram(wgpu::Device&, Program&);
-
-    wgpu::ShaderModule vertexShaderModule = nullptr;
-    wgpu::ShaderModule fragmentShaderModule = nullptr;
-    wgpu::ShaderModule computeShaderModule = nullptr;
-    std::vector<wgpu::ConstantEntry> constants;
-};
-
-
-// WGPUVertexBufferInfo maps Filament vertex attributes to WebGPU buffer binding model.
-class WGPUVertexBufferInfo : public HwVertexBufferInfo {
-public:
-    WGPUVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
-            AttributeArray const& attributes);
-
-    inline wgpu::VertexBufferLayout const* getVertexBufferLayouts() const {
-        return mVertexBufferLayouts.data();
-    }
-    inline uint32_t getVertexBufferLayoutCount() const {
-        return static_cast<uint32_t>(mVertexBufferLayouts.size());
-    }
-
-    inline wgpu::VertexAttribute const* getVertexAttributes(uint32_t i) const {
-        return mVertexAttributes[i].data();
-    }
-    inline uint32_t getVertexAttributeCount(uint32_t i) const {
-        return static_cast<uint32_t>(mVertexAttributes[i].size());
-    }
-
-    struct WebGPUSlotBindingInfo {
-        uint8_t sourceBuffer;
-        uint32_t slot;
-        uint64_t bufferOffset;
-        uint32_t stride;
-    };
-    inline const std::vector<WebGPUSlotBindingInfo>& getWebGPUSlotBindingInfos() const {
-        return mWebGPUSlotBindingInfos;
-    }
-
-private:
-    // This stores the final wgpu::VertexBufferLayout objects, one per WebGPU slot.
-    std::vector<wgpu::VertexBufferLayout> mVertexBufferLayouts;
-
-    // This stores all wgpu::VertexAttribute structs, indexed by their original
-    // Filament attributeIndex.
-    std::vector<std::vector<wgpu::VertexAttribute>> mVertexAttributes;
-
-    // Stores information for the driver to perform setVertexBuffer calls
-    std::vector<WebGPUSlotBindingInfo> mWebGPUSlotBindingInfos;
-};
-
-struct WGPUVertexBuffer : public HwVertexBuffer {
-    WGPUVertexBuffer(wgpu::Device const &device, uint32_t vertexCount, uint32_t bufferCount,
-                     Handle<HwVertexBufferInfo> vbih);
-
-    Handle<HwVertexBufferInfo> vbih;
-    utils::FixedCapacityVector<wgpu::Buffer> buffers;
-};
-
-class WGPUBufferBase {
-public:
-    void createBuffer(wgpu::Device const& device, wgpu::BufferUsage usage, uint32_t size,
-            char const* label);
-    void updateGPUBuffer(BufferDescriptor& bufferDescriptor, uint32_t byteOffset,
-            wgpu::Queue queue);
-    const wgpu::Buffer& getBuffer() const { return buffer; }
-protected:
-    wgpu::Buffer buffer;
-private:
-    // 4 bytes to hold any extra chunk we need.
-    std::array<uint8_t,4> mRemainderChunk;
-};
-
-class WGPUIndexBuffer : public HwIndexBuffer, public WGPUBufferBase {
-public:
-    WGPUIndexBuffer(wgpu::Device const &device, uint8_t elementSize,
-                    uint32_t indexCount);
-    wgpu::IndexFormat indexFormat;
-};
-
-class WGPUBufferObject : public HwBufferObject, public WGPUBufferBase {
-public:
-    WGPUBufferObject(wgpu::Device const &device, BufferObjectBinding bindingType, uint32_t byteCount);
-};
-
-class WebGPUDescriptorSetLayout final : public HwDescriptorSetLayout {
-public:
-
-    struct BindGroupEntryInfo final {
-        uint8_t binding = 0;
-        bool hasDynamicOffset = false;
-    };
-
-    WebGPUDescriptorSetLayout(DescriptorSetLayout const& layout, wgpu::Device const& device);
-    ~WebGPUDescriptorSetLayout();
-    [[nodiscard]] const wgpu::BindGroupLayout& getLayout() const { return mLayout; }
-    [[nodiscard]] std::vector<BindGroupEntryInfo> const& getBindGroupEntries() const {
-        return mBindGroupEntries;
-    }
-
-private:
-    // TODO: If this is useful elsewhere, remove it from this class
-    // Convert Filament Shader Stage Flags bitmask to webgpu equivalent
-    static wgpu::ShaderStage filamentStageToWGPUStage(ShaderStageFlags fFlags);
-    std::vector<BindGroupEntryInfo> mBindGroupEntries;
-    wgpu::BindGroupLayout mLayout;
-};
-
-class WebGPUDescriptorSet final : public HwDescriptorSet {
-public:
-
-    WebGPUDescriptorSet(wgpu::BindGroupLayout const& layout,
-            std::vector<WebGPUDescriptorSetLayout::BindGroupEntryInfo> const& bindGroupEntries);
-    ~WebGPUDescriptorSet();
-
-    wgpu::BindGroup lockAndReturn(wgpu::Device const&);
-    void addEntry(unsigned int index, wgpu::BindGroupEntry&& entry);
-    [[nodiscard]] bool getIsLocked() const { return mBindGroup != nullptr; }
-    [[nodiscard]] size_t countEntitiesWithDynamicOffsets() const;
-
-    // May be nullptr. Use lockAndReturn to create the bind group when appropriate
-    [[nodiscard]] const wgpu::BindGroup& getBindGroup() const { return mBindGroup; }
-
-private:
-    wgpu::BindGroupLayout mLayout = nullptr;
-    static constexpr uint8_t INVALID_INDEX = MAX_DESCRIPTOR_COUNT + 1;
-    std::array<uint8_t, MAX_DESCRIPTOR_COUNT> mEntryIndexByBinding{};
-    std::vector<wgpu::BindGroupEntry> mEntries;
-    const size_t mEntriesWithDynamicOffsetsCount;
-    wgpu::BindGroup mBindGroup = nullptr;
-};
-
-class WGPUTexture : public HwTexture {
-public:
-    WGPUTexture(SamplerType target, uint8_t levels, TextureFormat format, uint8_t samples,
-            uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
-            wgpu::Device const& device) noexcept;
-
-    WGPUTexture(WGPUTexture* src, uint8_t baseLevel, uint8_t levelCount) noexcept;
-    wgpu::TextureAspect getAspect() const { return mAspect; }
-    size_t getBlockWidth() const { return mBlockWidth; }
-    size_t getBlockHeight() const { return mBlockHeight; }
-
-    [[nodiscard]] const wgpu::Texture& getTexture() const { return mTexture; }
-    [[nodiscard]] const wgpu::TextureView& getTextureView() const { return mTexureView; }
-    [[nodiscard]] wgpu::TextureFormat getFormat() const { return mFormat; }
-
-    static wgpu::TextureFormat fToWGPUTextureFormat(
-            filament::backend::TextureFormat const& fFormat);
-    static wgpu::TextureAspect fToWGPUTextureViewAspect(
-            filament::backend::TextureUsage const& fUsage,
-            filament::backend::TextureFormat const& fFormat);
-
-private:
-    wgpu::TextureView makeTextureView(const uint8_t& baseLevel, const uint8_t& levelCount,
-            SamplerType target);
-    // CreateTextureR has info for a texture and sampler. Texture Views are needed for binding,
-    // along with a sampler Current plan: Inherit the sampler and Texture to always exist (It is a
-    // ref counted pointer) when making views. View is optional
-    wgpu::Texture mTexture = nullptr;
-    wgpu::TextureUsage mUsage = wgpu::TextureUsage::None;
-    wgpu::TextureFormat mFormat = wgpu::TextureFormat::Undefined;
-    wgpu::TextureAspect mAspect = wgpu::TextureAspect::Undefined;
-    uint32_t mArrayLayerCount = 1;
-    wgpu::TextureView mTexureView = nullptr;
-    wgpu::TextureUsage fToWGPUTextureUsage(filament::backend::TextureUsage const& fUsage);
-    size_t mBlockWidth;
-    size_t mBlockHeight;
-};
-
-struct WGPURenderPrimitive : public HwRenderPrimitive {
-
-    WGPUVertexBuffer* vertexBuffer = nullptr;
-    WGPUIndexBuffer* indexBuffer = nullptr;
-};
-
-class WGPURenderTarget : public HwRenderTarget {
-public:
-    using Attachment = TargetBufferInfo; // Using TargetBufferInfo directly for attachments
-
-    WGPURenderTarget(uint32_t width, uint32_t height, uint8_t samples,
-            const MRT& colorAttachments,
-            const Attachment& depthAttachment,
-            const Attachment& stencilAttachment);
-
-    // Default constructor for the default render target
-    WGPURenderTarget()
-        : HwRenderTarget(0, 0),
-          defaultRenderTarget(true),
-          samples(1) {}
-
-    // Updated signature: takes resolved views for custom RTs, and default views for default RT
-    void setUpRenderPassAttachments(
-            wgpu::RenderPassDescriptor& descriptor,
-            RenderPassParams const& params,
-            // For default render target:
-            wgpu::TextureView const& defaultColorTextureView,
-            wgpu::TextureView const& defaultDepthStencilTextureView,
-            // For custom render targets:
-            wgpu::TextureView const* customColorTextureViews, // Array of views
-            uint32_t customColorTextureViewCount,
-            wgpu::TextureView const& customDepthTextureView,
-            wgpu::TextureView const& customStencilTextureView,
-            wgpu::TextureFormat customDepthFormat,
-            wgpu::TextureFormat customStencilFormat);
-
-
-    bool isDefaultRenderTarget() const { return defaultRenderTarget; }
-    uint8_t getSamples() const { return samples; }
-
-    // Accessors for the driver to get stored attachment info
-    const MRT& getColorAttachmentInfos() const { return mColorAttachments; }
-    const Attachment& getDepthAttachmentInfo() const { return mDepthAttachment; }
-    const Attachment& getStencilAttachmentInfo() const { return mStencilAttachment; }
-
-    // Static helpers for load/store operations
-    static wgpu::LoadOp getLoadOperation(const RenderPassParams& params, TargetBufferFlags buffer);
-    static wgpu::StoreOp getStoreOperation(const RenderPassParams& params, TargetBufferFlags buffer);
-
-private:
-    bool defaultRenderTarget = false;
-    uint8_t samples = 1;
-
-    MRT mColorAttachments{};
-    // TODO WebGPU only supports a DepthStencil attachment, should this be just mDepthStencilAttachment?
-    Attachment mDepthAttachment{};
-    Attachment mStencilAttachment{};
-
-    // Cached descriptors for the render pass
-    std::vector<wgpu::RenderPassColorAttachment> mColorAttachmentDescriptors;
-    wgpu::RenderPassDepthStencilAttachment mDepthStencilAttachmentDescriptor{};
-    bool mHasDepthStencilAttachment = false;
-};
-}// namespace filament::backend
-#endif// TNT_FILAMENT_BACKEND_WEBGPUHANDLES_H
@@ -0,0 +1,36 @@
+/*
+ * 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 "WebGPUIndexBuffer.h"
+
+#include "WebGPUBufferBase.h"
+
+#include "DriverBase.h"
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <cstdint>
+
+namespace filament::backend {
+
+WebGPUIndexBuffer::WebGPUIndexBuffer(wgpu::Device const& device, const uint8_t elementSize,
+        const uint32_t indexCount)
+    : HwIndexBuffer{ elementSize, indexCount },
+      WebGPUBufferBase{ device, wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Index,
+          elementSize * indexCount, "index_buffer" },
+      mIndexFormat{ elementSize == 2 ? wgpu::IndexFormat::Uint16 : wgpu::IndexFormat::Uint32 } {}
+
+} // namespace filament::backend
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Powei Feng	4261577724	Add unfilterable	2025-06-16 10:25:13 -07:00
Sungun Park	a3922ac3a9	Bump MATERIAL_VERSION to 61	2025-06-13 15:39:09 -07:00
Sungun Park	5f1112d85e	Update RELEASE_NOTES for the revert	2025-06-13 15:26:56 -07:00
Sungun Park	09545690bf	Revert "materials: introduce mutable spec constants (#8795 )" This reverts commit `8a1a0b0fd2`.	2025-06-13 15:02:55 -07:00
Powei Feng	3ce752c47b	gltfio: enable passing in variant filters (#8851 ) This enables filtering variants when using the JitMaterialProvider.	2025-06-13 21:06:22 +00:00
bridgewaterrobbie	1f9f3ddb91	Use the mipmap generator to properly generate mipmaps.	2025-06-13 16:46:18 -04:00
bridgewaterrobbie	8e94688fbf	Request higher maxStorageTexturesPerShaderStage if the device supports it.	2025-06-13 16:46:18 -04:00
bridgewaterrobbie	438ee40f40	Account for array layers, utilizing maxStorageTexturesPerShaderStage to configure	2025-06-13 16:46:18 -04:00
bridgewaterrobbie	7f22d733d7	webgpu: Significant cleanup of mipmap generator interface and logic. Now properly queries and uses device limits of maxMipsPerPass. Still fairly hacky though	2025-06-13 16:46:18 -04:00
Andy Hovingh	77d77f0a5c	webgpu: fix logging	2025-06-13 07:23:56 -05:00
Sungun Park	0aabd3dd6a	Fix: Prevent deadlock when loading programs from cache (#8849 ) When a program was created directly from a cached blob in `ShaderCompilerService`, its associated token was not signaled as ready in the THREAD_POOL mode. This oversight caused a deadlock at the `token->wait()` call during program destruction. This commit resolves the issue by skipping the token's readiness check upon destruction if the program was created from the cache blob. BUGS=[423221474]	2025-06-12 21:05:59 +00:00
Benjamin Doherty	72a54b30f9	Temporary workaround for PlatformMetal	2025-06-12 13:23:49 -07:00
bridgewaterrobbie	00b2ab7ac4	webgpu: Use C++ converted mipmap generator. First pass, likely to be tweaked substantially in the future	2025-06-12 15:13:10 -04:00
dependabot[bot]andPowei Feng	39e15d67fc	Bump brace-expansion (#8855 ) Bumps the npm_and_yarn group with 1 update in the /build/common/upload-release-assets directory: [brace-expansion](https://github.com/juliangruber/brace-expansion). Updates `brace-expansion` from 1.1.11 to 1.1.12 - [Release notes](https://github.com/juliangruber/brace-expansion/releases) - [Commits](https://github.com/juliangruber/brace-expansion/compare/1.1.11...v1.1.12) --- updated-dependencies: - dependency-name: brace-expansion dependency-version: 1.1.12 dependency-type: indirect dependency-group: npm_and_yarn ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Powei Feng <powei@google.com>	2025-06-12 17:55:35 +00:00
Juan Caldas	921dbddad9	webgpu: add condition to sample (#8860 ) * Add option for enabling MSAA	2025-06-12 12:14:54 -04:00
Andy Hovingh	6406d4602c	webgpu: texture changes to better support mipmap generation later	2025-06-11 22:35:42 -05:00
Sungun Park	880cd66166	Release Filament 1.61.1	2025-06-12 00:31:33 +00:00
Juan Caldas	f27ff203cf	webgpu: Enable multisampling in the sample (#8854 )	2025-06-11 23:09:59 +00:00
Ben Doherty	3ea8e529cc	Switch filament and backend to Abseil logging (#8691 )	2025-06-11 10:08:52 -07:00
Jeremy Nelson	572dd233d7	adding check for Stencil8	2025-06-10 20:06:00 -07:00
Jeremy Nelson	c19070e477	better default cases for loadOp and storeOp	2025-06-10 20:06:00 -07:00
Jeremy Nelson	64328b7388	pass through defaultDepthStencilFormat	2025-06-10 20:06:00 -07:00
Powei Feng	1414fdcd38	github: fix commit message action (#8846 ) Checkout sha instead of the branch	2025-06-10 22:55:48 +00:00
Matthew Hoffman	e607a035ae	Add TODOs for backend test features that shouldn't be made yet. (#8829 ) Some edge case handling features are possibly needed, but should only be made once they have use cases that exercise that behavior. So for now just leave TODOs explaining where to add it in the future. BUGS=[422804941,422803382]	2025-06-10 21:43:47 +00:00
Ben Doherty	0496e089c1	Remove old screenshot output for readPixels backend test (#8840 )	2025-06-10 13:30:00 -07:00
Juan Caldas	b310e3d24a	webgpu: Add missing WEBGPU cases (#8839 )	2025-06-10 13:56:48 -04:00
Juan Caldas	b359d77669	webgpu: Adjust the wrap modes (#8844 ) * Change the WebGPU wrap modes	2025-06-10 11:09:37 -04:00
Sungun Park	e57a4061cc	Add missing header (#8842 )	2025-06-09 21:01:29 +00:00
Powei Feng	53f82e6b71	vk: rename selectMemory params (#8835 ) Use more accurate parameter names. 'types' indicate the types of memory we are considering. 'reqs' indicate the requirements each type must meet. BUGS=401579988	2025-06-09 20:33:08 +00:00
rafadevaiandSerge Metral	e88bba5940	VK: Disable VMA internal synchronization (#8836 ) The vulkan backend is currently single threads, so there's no need for VMA to use synchronization internally. This will improve the CPU performance when going through VMA. Co-authored-by: Serge Metral <sergemetral@google.com>	2025-06-09 20:05:02 +00:00
Matthew Hoffman	0b83454d08	Fix backend tests on macos vulkan. (#8815 )	2025-06-09 19:05:03 +00:00
Matthew Hoffman	1ef0d36f79	Improve and document the backend test helper script. (#8816 )	2025-06-09 18:19:30 +00:00
Powei Feng	e350737d1e	Remove texture pbd check for compressed textures (#8837 ) This will also fix the failing ktx test.	2025-06-09 17:47:55 +00:00
Andy Hovingh	cb203b13de	webgpu: refactor: sort CMake source files	2025-06-09 11:05:15 -05:00
Andy Hovingh	92a0d7bfc3	webgpu: refactor: move WebGPU render target class to its own source+header files.	2025-06-09 11:05:15 -05:00
Andy Hovingh	c1a7ed2799	webgpu: refactor: move WebGPU texture class to its own source+header files.	2025-06-09 11:05:15 -05:00
Andy Hovingh	692301bdf1	webgpu: refactor: move WebGPU descriptor set related classes to their own source+header files.	2025-06-09 11:05:15 -05:00
Andy Hovingh	05b87b3c42	webgpu: refactor: move WebGPU buffer classes to their own source+header files.	2025-06-09 11:05:15 -05:00
Andy Hovingh	9825297e70	webgpu: refactor: move WebGPUProgram to its own source+header files.	2025-06-09 11:05:15 -05:00
Powei Feng	6a59c887de	vk: implement proper offset when uploading to texture (#8830 ) We didn't take into account the top/left/stride parameters of the PixelBufferDescriptor.	2025-06-06 20:46:45 +00:00
bridgewaterrobbie	33b4b46220	Create simple gltf viewer based on gltf_instances, primarily to facilitate webgpu development	2025-06-06 14:51:17 -04:00
Powei Feng	06170a25f7	Fix PBD size check in Texture (#8826 ) Bug found when running ./samples/heightfield and press add pixel buffer padding.	2025-06-06 18:20:05 +00:00
Konrad Piascik	418dad883d	webgpu: Fix Android material compilation Instead of requiring GL_EXT_shader_non_constant_global_initializers we can just make the new variables const BUGS=[421457710]	2025-06-06 13:46:37 -04:00
Jeremy Nelson	447661efed	Use colorInfos.layer	2025-06-05 22:18:34 -07:00
Jeremy Nelson	bd61ab691c	add Default miplevel and array layer	2025-06-05 22:18:34 -07:00
Jeremy Nelson	2a04cee97f	rename member variables	2025-06-05 22:18:34 -07:00
Jeremy Nelson	5ae7760752	fix samplerType check	2025-06-05 22:18:34 -07:00
Jeremy Nelson	56b8bf5b4b	noexcept to makeTextureView	2025-06-05 22:18:34 -07:00
Jeremy Nelson	6af38fa47f	renaming param to samplerType	2025-06-05 22:18:34 -07:00
Jeremy Nelson	0c01799c50	overload getTextureView	2025-06-05 22:18:34 -07:00
Jeremy Nelson	e2519e0eed	Update WebGPUHandles.cpp	2025-06-05 22:18:34 -07:00
Jeremy Nelson	6178ae0459	Use colorInfos.layer # Conflicts: # filament/backend/src/webgpu/WebGPUHandles.cpp	2025-06-05 22:18:34 -07:00
Eliza	8a1a0b0fd2	materials: introduce mutable spec constants (#8795 ) * materials: introduce mutable spec constants Rationale & design of this feature has been discussed internally. The current implementation uses a `FixedCapacityVector` to store the new program handles, but I wouldn't object to replacing it with a hasmap as discussed offline. I have compiled but not tested this yet on Android, so I'm not certain that the API bindings are correctly wired up. * materials: mutable spec constant feedback * materials: address mutable spec constant comments	2025-06-06 03:52:20 +00:00
Powei Feng	a7c3cf4173	Ensure maximum number of vertex buffer does not exceed 8 (#8824 )	2025-06-05 21:16:38 +00:00
bridgewaterrobbie	57b68eeba3	webgpu: Add required feature RG11B10UfloatRenderable for gltf_viewer, with comment on why	2025-06-05 15:08:03 -04:00
Juan Caldas	ad8c9ce4e0	Add common CLI Args parser for the samples (#8819 )	2025-06-05 18:18:41 +00:00
Powei Feng	7dc0f2cc86	vk: fix two validation error (#8825 ) On starting gltf_viewer on mac, there were validation errors on 1) copying to a depth texture using the aspect COLOR. 2) querying formats that required extensions without enabling those extensions.	2025-06-05 17:50:40 +00:00
bridgewaterrobbie	37110799ef	webgpu: Correct stencil state handling for depth-only formats	2025-06-05 13:22:26 -04:00
Anish GoyalandSerge Metral	88a06ec8e7	Switch to block-based stage-pool for Vulkan (#8742 ) * Switch to block-based stage-pool for Vulkan Instead of allocating a staging buffer every time one is needed, allocate a large (8mb) block of memory, and divvy it up as needed. We will make this configurable in the future, to allow for tuning for different apps as needed. * Address PR comments: use fvkmemory::Resource Instead of having the child block be a unique_ptr that we create a separate container for within the command buffers, just have the stage block segments be fvkmemory::Resource instances. * Address PR comments for staging buff change - As per discussion with @poweifeng, change the name of a variable called "stage" to "stageSegment" for clarity - As per discussion with @rafadevai, change the order of terminate calls in VulkanDriver to better reflect cleanup order of some objects. * Align stage pool to nonCoherentAtomSize In order to prevent flushing more atoms than were modified when writing data to host-mapped memory in a staging buffer, ensure that all segments allocated are aligned to nonCoherentAtomSize. Also - fix merge conflict compile errors. --------- Co-authored-by: Serge Metral <sergemetral@google.com>	2025-06-05 09:15:26 -07:00
bridgewaterrobbie	38705d6226	Add todos for unimplemented funcs that we haven't yet started on	2025-06-05 10:48:59 -04:00
bridgewaterrobbie	d325bb43cf	Add destroy texture implementation	2025-06-05 10:48:59 -04:00
Sungun Park	0c52d3c9bf	Fix: Submit callback handle on completion (#8818 ) This reverts a behavioral regression introduced in commit `c3542b135e`, which deferred callback submission until the program was first used. This commit restores the correct behavior by submitting the callback handle as soon as the token's work is complete. This occurs either upon successful `gl.program` population or via cancellation, ensuring the caller is properly notified that the resource loading operation has concluded.	2025-06-04 16:44:48 +00:00
Konrad Piascik	770176a1e3	Fix Windows build error C2512	2025-06-04 11:17:05 -04:00
bridgewaterrobbie	672603f9b4	Generate mipmaps when requested, deferring as needed	2025-06-03 17:45:35 -04:00
Powei Feng	0ecf6c46e2	github: Fix commit message parsing (#8817 )	2025-06-03 14:24:29 -07:00
Andy Hovingh	8b87a54c1a	webgpu: fix vertex buffer info for hellopbr	2025-06-03 16:08:12 -05:00
Powei Feng	cd1d3e8749	github: update windows runner due to 2019 being "stuck" (#8814 )	2025-06-03 13:27:31 -07:00
rafadevai	e88072cec0	VK: Introduce VulkanBufferCache (#8757 ) This class will allow better tracking of memory allocations and recycling of buffers. Currently only the uniform buffers are recycled. It will eventually allow us to dynamically change the underlying GPU buffer of a VulkanBuffer when updating an UBO for uniforms and also keep track which of those buffers are still inflight and which ones are ready to be reuse for an UBO. Its the first step on moving towards by passing the staging buffer in UMA.	2025-06-03 12:36:50 -07:00
Powei Feng	1d2e165d99	Release Filament 1.61.0	2025-06-03 12:02:10 -07:00
Powei Feng	cfc4ac5511	Add missing include in JobSystem.cpp (#8812 )	2025-06-03 11:44:42 -07:00
Powei Feng	73a03d7af0	Flip conditional for fixing missing samplers (#8811 )	2025-06-03 11:19:29 -07:00
Powei Feng	56dc348cc8	utils: add additional guards for Tracing (#8810 ) The addition JobSystem.cpp allows for defining FILAMENT_TRACING_ENABLED across targets. Addingin FILAMENT_TRACING_ENABLED to the #if in Tracing.h prevents perfetto from being included.	2025-06-03 11:18:11 -07:00
Juan Caldas	1b46ddd8b6	Webgpu: Check for Shadows (#8807 ) BUGS=[397432947]	2025-06-03 11:57:34 +00:00
Powei Feng	a68aaa114e	renderdiff: fix breakage in parsing commit message (#8808 )	2025-06-02 15:55:35 -07:00
Powei Feng	3da7dabb2a	renderdiff: enable update goldens on commit merge (#8771 )	2025-06-02 14:12:26 -07:00
bridgewaterrobbie	1e2311da3d	Fix the fact that isSample might be combined with the other flags	2025-06-02 16:18:08 -04:00
bridgewaterrobbie	a38c55c82d	Fix use accidental use of uninitalized HwTexture::target	2025-06-02 16:18:08 -04:00
Powei Feng	cd22478e4f	renderdiff: add viewer for image differences (#8768 ) - Modify the compare script to output more details of a comparison. This will include the source/golden directory, the comparison directory (the new renderings), and a file path to difference images if the golden does not match the rendered image. - The image_diff script can now output a TIFF that is the difference of two input TIFFs. - Add a viewer for examining the differences between rendered output and golden images. - The viewer consists of a simple server of web API endpoints for querying difference results (along with rendered images in TIFF). - And a web-based (html + lit-element) UI for looking at the rendered images and differences.	2025-06-02 19:37:46 +00:00
bridgewaterrobbie	60036c75fe	Remove asserts that throw on shadowtest. Investigate them later	2025-06-02 14:17:39 -04:00
Powei Feng	044f2aeb7c	gl: keep external texture id in sync (#8803 ) (Attributed to @dsternfeld7)	2025-06-02 17:18:00 +00:00