Throw an NSException on Apple platforms when panicking

2024-03-27 16:04:49 -07:00
205 changed files with 3728 additions and 7735 deletions
@@ -71,10 +71,6 @@ set(FILAMENT_METAL_HANDLE_ARENA_SIZE_IN_MB "8" CACHE STRING
    "Size of the Metal handle arena, default 8."
 )

-set(FILAMENT_BACKEND_DEBUG_FLAG "" CACHE STRING
-    "A debug flag meant for enabling/disabling backend debugging paths"
-)
-
 # Enable exceptions by default in spirv-cross.
 set(SPIRV_CROSS_EXCEPTIONS_TO_ASSERTIONS OFF)

@@ -552,12 +548,6 @@ if (FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview")
    set(FILAMENT_ENABLE_MULTIVIEW ON)
 endif ()

-# Define backend flag for debug only
-if (CMAKE_BUILD_TYPE STREQUAL "Debug" AND NOT FILAMENT_BACKEND_DEBUG_FLAG STREQUAL "")
-    add_definitions(-DFILAMENT_BACKEND_DEBUG_FLAG=${FILAMENT_BACKEND_DEBUG_FLAG})
-    unset(FILAMENT_BACKEND_DEBUG_FLAG)
-endif()
-
 # ==================================================================================================
 # Material compilation flags
 # ==================================================================================================
@@ -829,9 +819,6 @@ if (IS_HOST_PLATFORM)
    add_subdirectory(${TOOLS}/glslminifier)
    add_subdirectory(${TOOLS}/matc)
    add_subdirectory(${TOOLS}/matinfo)
-    if (NOT WIN32)  # matedit not yet supported on Windows
-        add_subdirectory(${TOOLS}/matedit)
-    endif()
    add_subdirectory(${TOOLS}/mipgen)
    add_subdirectory(${TOOLS}/normal-blending)
    add_subdirectory(${TOOLS}/resgen)
@@ -8,5 +8,4 @@ appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).

 ## Release notes for next branch cut

- filagui: Fix regression which broke WebGL
- Add a new Engine::Config setting to control preferred shader language
+- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.51.7'
+    implementation 'com.google.android.filament:filament-android:1.51.0'
 }
 ```

@@ -51,7 +51,7 @@ Here are all the libraries available in the group `com.google.android.filament`:
 iOS projects can use CocoaPods to install the latest release:

 ```shell
-pod 'Filament', '~> 1.51.7'
+pod 'Filament', '~> 1.51.0'
 ```

 ### Snapshots
@@ -7,31 +7,6 @@ A new header is inserted each time a *tag* is created.
 Instead, if you are authoring a PR for the main branch, add your release note to
 [NEW_RELEASE_NOTES.md](./NEW_RELEASE_NOTES.md).

-## v1.51.7
-
- Add new matedit tool
- filagui: Support rendering `GL_TEXTURE_EXTERNAL_OES` textures.
- `setFrameScheduledCallback` now takes a `utils::Invocable`.
- engine: Add `isPaused()`
-
-## v1.51.6
-
- Add new matedit tool
- filagui: Support rendering `GL_TEXTURE_EXTERNAL_OES` textures.
-
-## v1.51.5
-
-
-## v1.51.4
-
-
-## v1.51.3
-
-
-## v1.51.2
-
- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`
-
 ## v1.51.1


@@ -342,21 +342,6 @@ Java_com_google_android_filament_Engine_nIsValidMaterial(JNIEnv*, jclass,
    return (jboolean)engine->isValid((Material*)nativeMaterial);
 }

-extern "C" JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_Engine_nIsValidMaterialInstance(JNIEnv*, jclass,
-        jlong nativeEngine, jlong nativeMaterial, jlong nativeMaterialInstance) {
-    Engine* engine = (Engine *)nativeEngine;
-    return (jboolean)engine->isValid((Material*)nativeMaterial,
-            (MaterialInstance*)nativeMaterialInstance);
-}
-
-extern "C" JNIEXPORT jboolean JNICALL
-        Java_com_google_android_filament_Engine_nIsValidExpensiveMaterialInstance(JNIEnv*, jclass,
-        jlong nativeEngine, jlong nativeMaterialInstance) {
-    Engine* engine = (Engine *)nativeEngine;
-    return (jboolean)engine->isValidExpensive((MaterialInstance*)nativeMaterialInstance);
-}
-
 extern "C" JNIEXPORT jboolean JNICALL
 Java_com_google_android_filament_Engine_nIsValidSkybox(JNIEnv*, jclass,
        jlong nativeEngine, jlong nativeSkybox) {
@@ -406,13 +391,6 @@ Java_com_google_android_filament_Engine_nFlush(JNIEnv*, jclass,
    engine->flush();
 }

-extern "C" JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_Engine_nIsPaused(JNIEnv*, jclass,
-        jlong nativeEngine) {
-    Engine* engine = (Engine*) nativeEngine;
-    return (jboolean)engine->isPaused();
-}
-
 extern "C" JNIEXPORT void JNICALL
 Java_com_google_android_filament_Engine_nSetPaused(JNIEnv*, jclass,
        jlong nativeEngine, jboolean paused) {
@@ -49,12 +49,6 @@ Java_com_google_android_filament_View_nSetCamera(JNIEnv*, jclass,
    view->setCamera(camera);
 }

-extern "C" JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_View_nHasCamera(JNIEnv*, jclass, jlong nativeView) {
-    View* view = (View*) nativeView;
-    return (jboolean)view->hasCamera();
-}
-
 extern "C" JNIEXPORT void JNICALL
 Java_com_google_android_filament_View_nSetColorGrading(JNIEnv*, jclass,
        jlong nativeView, jlong nativeColorGrading) {
@@ -849,24 +849,6 @@ public class Engine {
        return nIsValidMaterial(getNativeObject(), object.getNativeObject());
    }

-    /**
-     * Returns whether the object is valid.
-     * @param object Object to check for validity
-     * @return returns true if the specified object is valid.
-     */
-    public boolean isValidMaterialInstance(@NonNull Material ma, MaterialInstance mi) {
-        return nIsValidMaterialInstance(getNativeObject(), ma.getNativeObject(), mi.getNativeObject());
-    }
-
-    /**
-     * Returns whether the object is valid.
-     * @param object Object to check for validity
-     * @return returns true if the specified object is valid.
-     */
-    public boolean isValidExpensiveMaterialInstance(@NonNull MaterialInstance object) {
-        return nIsValidExpensiveMaterialInstance(getNativeObject(), object.getNativeObject());
-    }
-
    /**
     * Returns whether the object is valid.
     * @param object Object to check for validity
@@ -1227,17 +1209,6 @@ public class Engine {
        nFlush(getNativeObject());
    }

-    /**
-     * Get paused state of rendering thread.
-     *
-     * <p>Warning: This is an experimental API.
-     *
-     * @see #setPaused
-     */
-    public boolean isPaused() {
-        return nIsPaused(getNativeObject());
-    }
-
    /**
     * Pause or resume the rendering thread.
     *
@@ -1320,8 +1291,6 @@ public class Engine {
    private static native boolean nIsValidSkinningBuffer(long nativeEngine, long nativeSkinningBuffer);
    private static native boolean nIsValidIndirectLight(long nativeEngine, long nativeIndirectLight);
    private static native boolean nIsValidMaterial(long nativeEngine, long nativeMaterial);
-    private static native boolean nIsValidMaterialInstance(long nativeEngine, long nativeMaterial, long nativeMaterialInstance);
-    private static native boolean nIsValidExpensiveMaterialInstance(long nativeEngine, long nativeMaterialInstance);
    private static native boolean nIsValidSkybox(long nativeEngine, long nativeSkybox);
    private static native boolean nIsValidColorGrading(long nativeEngine, long nativeColorGrading);
    private static native boolean nIsValidTexture(long nativeEngine, long nativeTexture);
@@ -1330,7 +1299,6 @@ public class Engine {
    private static native void nDestroyEntity(long nativeEngine, int entity);
    private static native void nFlushAndWait(long nativeEngine);
    private static native void nFlush(long nativeEngine);
-    private static native boolean nIsPaused(long nativeEngine);
    private static native void nSetPaused(long nativeEngine, boolean paused);
    private static native long nGetTransformManager(long nativeEngine);
    private static native long nGetLightManager(long nativeEngine);
@@ -241,15 +241,6 @@ public class View {
        nSetCamera(getNativeObject(), camera == null ? 0 : camera.getNativeObject());
    }

-    /**
-     * Query whether a camera is set.
-     * @return true if a camera is set, false otherwise
-     * @see #setCamera
-     */
-    public boolean hasCamera() {
-        return nHasCamera(getNativeObject());
-    }
-
    /**
     * Gets this View's associated Camera, or null if none has been assigned.
     *
@@ -1247,7 +1238,6 @@ public class View {
    private static native void nSetName(long nativeView, String name);
    private static native void nSetScene(long nativeView, long nativeScene);
    private static native void nSetCamera(long nativeView, long nativeCamera);
-    private static native boolean nHasCamera(long nativeView);
    private static native void nSetViewport(long nativeView, int left, int bottom, int width, int height);
    private static native void nSetVisibleLayers(long nativeView, int select, int value);
    private static native void nSetShadowingEnabled(long nativeView, boolean enabled);
@@ -86,10 +86,7 @@ set(GLTFIO_SRCS
        ${GLTFIO_DIR}/src/Wireframe.cpp
        ${GLTFIO_DIR}/src/Wireframe.h
        ${GLTFIO_DIR}/src/downcast.h
-        ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.cpp
        ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.h
-        ${GLTFIO_DIR}/src/extended/ResourceLoaderExtended.cpp
-        ${GLTFIO_DIR}/src/extended/ResourceLoaderExtended.h
        ${GLTFIO_DIR}/src/extended/TangentsJobExtended.cpp
        ${GLTFIO_DIR}/src/extended/TangentsJobExtended.h
        ${GLTFIO_DIR}/src/extended/TangentSpaceMeshWrapper.cpp
@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.51.7
+VERSION_NAME=1.51.0

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

@@ -321,7 +321,7 @@ class MainActivity : Activity() {
            var flags = uiHelper.swapChainFlags
            if (engine.activeFeatureLevel == Engine.FeatureLevel.FEATURE_LEVEL_0) {
                if (SwapChain.isSRGBSwapChainSupported(engine)) {
-                    flags = flags or SwapChainFlags.CONFIG_SRGB_COLORSPACE
+                    flags = flags or SwapChain.CONFIG_SRGB_COLORSPACE
                }
            }

@@ -61,11 +61,6 @@ function print_help {
    echo "    -b"
    echo "        Enable Address and Undefined Behavior Sanitizers (asan/ubsan) for debugging."
    echo "        This is only for the desktop build."
-    echo "    -x value"
-    echo "        Define a preprocessor flag FILAMENT_BACKEND_DEBUG_FLAG with [value]. This is useful for"
-    echo "        enabling debug paths in the backend from the build script. For example, make a"
-    echo "        systrace-enabled build without directly changing #defines. Remember to add -f when"
-    echo "        changing this option."
    echo ""
    echo "Build types:"
    echo "    release"
@@ -177,8 +172,6 @@ MATOPT_GRADLE_OPTION=""

 ASAN_UBSAN_OPTION=""

-BACKEND_DEBUG_FLAG_OPTION=""
-
 IOS_BUILD_SIMULATOR=false
 BUILD_UNIVERSAL_LIBRARIES=false

@@ -238,7 +231,6 @@ function build_desktop_target {
            ${MATDBG_OPTION} \
            ${MATOPT_OPTION} \
            ${ASAN_UBSAN_OPTION} \
-            ${BACKEND_DEBUG_FLAG_OPTION} \
            ${architectures} \
            ../..
        ln -sf "out/cmake-${lc_target}/compile_commands.json" \
@@ -297,7 +289,6 @@ function build_webgl_with_target {
            -DCMAKE_BUILD_TYPE="$1" \
            -DCMAKE_INSTALL_PREFIX="../webgl-${lc_target}/filament" \
            -DWEBGL=1 \
-            ${BACKEND_DEBUG_FLAG_OPTION} \
            ../..
        ln -sf "out/cmake-webgl-${lc_target}/compile_commands.json" \
           ../../compile_commands.json
@@ -372,7 +363,6 @@ function build_android_target {
            ${MATDBG_OPTION} \
            ${MATOPT_OPTION} \
            ${VULKAN_ANDROID_OPTION} \
-            ${BACKEND_DEBUG_FLAG_OPTION} \
            ../..
        ln -sf "out/cmake-android-${lc_target}-${arch}/compile_commands.json" \
           ../../compile_commands.json
@@ -607,7 +597,6 @@ function build_ios_target {
            -DCMAKE_TOOLCHAIN_FILE=../../third_party/clang/iOS.cmake \
            ${MATDBG_OPTION} \
            ${MATOPT_OPTION} \
-            ${BACKEND_DEBUG_FLAG_OPTION} \
            ../..
        ln -sf "out/cmake-ios-${lc_target}-${arch}/compile_commands.json" \
           ../../compile_commands.json
@@ -741,13 +730,6 @@ function validate_build_command {
            exit 1
        fi
    fi
-
-    # Make sure FILAMENT_BACKEND_DEBUG_FLAG is only meant for debug builds
-    if [[ "${ISSUE_DEBUG_BUILD}" != "true" ]] && [[ ! -z "${BACKEND_DEBUG_FLAG_OPTION}" ]]; then
-        echo "Error: cannot specify FILAMENT_BACKEND_DEBUG_FLAG in non-debug build"
-        exit 1
-    fi
-
    set -e
 }

@@ -794,7 +776,7 @@ function check_debug_release_build {

 pushd "$(dirname "$0")" > /dev/null

-while getopts ":hacCfgijmp:q:uvslwtedk:bx:" opt; do
+while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
    case ${opt} in
        h)
            print_help
@@ -858,7 +840,7 @@ while getopts ":hacCfgijmp:q:uvslwtedk:bx:" opt; do
                        echo "Platform must be one of [desktop|android|ios|webgl|all]"
                        echo ""
                        exit 1
-                    ;;
+                    ;;    
                esac
            done
            ;;
@@ -936,8 +918,6 @@ while getopts ":hacCfgijmp:q:uvslwtedk:bx:" opt; do
        b)  ASAN_UBSAN_OPTION="-DFILAMENT_ENABLE_ASAN_UBSAN=ON"
            echo "Enabled ASAN/UBSAN"
            ;;
-        x)  BACKEND_DEBUG_FLAG_OPTION="-DFILAMENT_BACKEND_DEBUG_FLAG=${OPTARG}"
-            ;;
        \?)
            echo "Invalid option: -${OPTARG}" >&2
            echo ""
@@ -1397,7 +1397,7 @@ Type
 :    `string`

 Value
-:     Any of `opaque`, `transparent`, `fade`, `add`, `masked`, `multiply`, `screen`, `custom`. Defaults to `opaque`.
+:     Any of `opaque`, `transparent`, `fade`, `add`, `masked`, `multiply`, `screen`. Defaults to `opaque`.

 Description
 :     Defines how/if the rendered object is blended with the content of the render target.
@@ -1420,7 +1420,6 @@ Description
      of the material's output defines whether a fragment is discarded or not. Additionally,
      ALPHA_TO_COVERAGE is enabled for non-translucent views. See the maskThreshold section for more
      information.
-    - **Custom**: blending is enabled. But the blending function is user specified. See `blendFunction`.

 !!! Note
    When `blending` is set to `masked`, alpha to coverage is automatically enabled for the material.
@@ -1433,36 +1432,6 @@ material {
 }
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-### Blending and transparency: blendFunction
-
-Type
-:    `object`
-
-Fields
-:     `srcRGB`, `srcA`, `dstRGB`, `dstA`
-
-Description
-:       - *srcRGB*: source function applied to the RGB channels
-        - *srcA*: source function applied to the alpha channel
-        - *srcRGB*: destination function applied to the RGB channels
-        - *srcRGB*: destination function applied to the alpha channel
-        The values possible for each functions are one of `zero`, `one`, `srcColor`, `oneMinusSrcColor`,
-               `dstColor`, `oneMinusDstColor`, `srcAlpha`, `oneMinusSrcAlpha`, `dstAlpha`,
-               `oneMinusDstAlpha`, `srcAlphaSaturate`
-
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ JSON
-material {
-    blending : custom,
-    blendFunction :
-    {
-        srcRGB: one,
-        srcA: one,
-        dstRGB: oneMinusSrcColor,
-        dstA: oneMinusSrcAlpha
-    }
- }
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
 ### Blending and transparency: postLightingBlending

 Type
@@ -163,7 +163,6 @@ set(PRIVATE_HDRS
        src/ResourceList.h
        src/ShadowMap.h
        src/ShadowMapManager.h
-        src/SharedHandle.h
        src/TypedUniformBuffer.h
        src/UniformBuffer.h
        src/components/CameraManager.h
@@ -215,7 +214,6 @@ set(PRIVATE_HDRS
 set(MATERIAL_SRCS
        src/materials/antiAliasing/fxaa.mat
        src/materials/antiAliasing/taa.mat
-        src/materials/blitDepth.mat
        src/materials/blitLow.mat
        src/materials/blitArray.mat
        src/materials/bloom/bloomDownsample.mat
@@ -71,8 +71,6 @@ if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3 AND NOT FILAMEN
            include/backend/platforms/OpenGLPlatform.h
            src/opengl/gl_headers.cpp
            src/opengl/gl_headers.h
-            src/opengl/GLBufferObject.h
-            src/opengl/GLTexture.h
            src/opengl/GLUtils.cpp
            src/opengl/GLUtils.h
            src/opengl/OpenGLBlobCache.cpp
@@ -168,10 +166,8 @@ endif()
 if (FILAMENT_SUPPORTS_VULKAN)
    list(APPEND SRCS
            include/backend/platforms/VulkanPlatform.h
-            src/vulkan/caching/VulkanDescriptorSetManager.cpp
-            src/vulkan/caching/VulkanDescriptorSetManager.h
-            src/vulkan/caching/VulkanPipelineLayoutCache.cpp
-            src/vulkan/caching/VulkanPipelineLayoutCache.h
+            src/vulkan/caching/VulkanDescriptorSet.cpp
+            src/vulkan/caching/VulkanDescriptorSet.h
            src/vulkan/platform/VulkanPlatform.cpp
            src/vulkan/platform/VulkanPlatformSwapChainImpl.cpp
            src/vulkan/platform/VulkanPlatformSwapChainImpl.h
@@ -24,7 +24,6 @@

 #include <backend/PresentCallable.h>

-#include <utils/Invocable.h>
 #include <utils/ostream.h>

 #include <math/vec4.h>
@@ -159,16 +158,14 @@ static constexpr const char* backendToString(Backend backend) {
 }

 /**
- * Defines the shader language. Similar to the above backend enum, with some differences:
- * - The OpenGL backend can select between two shader languages: ESSL 1.0 and ESSL 3.0.
- * - The Metal backend can prefer precompiled Metal libraries, while falling back to MSL.
+ * Defines the shader language. Similar to the above backend enum, but the OpenGL backend can select
+ * between two shader languages: ESSL 1.0 and ESSL 3.0.
 */
 enum class ShaderLanguage {
    ESSL1 = 0,
    ESSL3 = 1,
    SPIRV = 2,
    MSL = 3,
-    METAL_LIBRARY = 4,
 };

 static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguage) {
@@ -181,8 +178,6 @@ static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguag
            return "SPIR-V";
        case ShaderLanguage::MSL:
            return "MSL";
-        case ShaderLanguage::METAL_LIBRARY:
-            return "Metal precompiled library";
    }
 }

@@ -687,7 +682,7 @@ enum class TextureUsage : uint16_t {
    SUBPASS_INPUT       = 0x0020,            //!< Texture can be used as a subpass input
    BLIT_SRC            = 0x0040,            //!< Texture can be used the source of a blit()
    BLIT_DST            = 0x0080,            //!< Texture can be used the destination of a blit()
-    PROTECTED           = 0x0100,            //!< Texture can be used for protected content
+    PROTECTED           = 0x0100,            //!< Texture can be used the destination of a blit()
    DEFAULT             = UPLOADABLE | SAMPLEABLE   //!< Default texture usage
 };

@@ -1225,7 +1220,7 @@ static_assert(sizeof(StencilState::StencilOperations) == 5u,
 static_assert(sizeof(StencilState) == 12u,
        "StencilState size not what was intended");

-using FrameScheduledCallback = utils::Invocable<void(backend::PresentCallable)>;
+using FrameScheduledCallback = void(*)(PresentCallable callable, void* user);

 enum class Workaround : uint16_t {
    // The EASU pass must split because shader compiler flattens early-exit branch
@@ -1242,6 +1237,9 @@ enum class Workaround : uint16_t {
    DISABLE_BLIT_INTO_TEXTURE_ARRAY,
    // Multiple workarounds needed for PowerVR GPUs
    POWER_VR_SHADER_WORKAROUNDS,
+    // The driver has some threads pinned, and we can't easily know on which core, it can hurt
+    // performance more if we end-up pinned on the same one.
+    DISABLE_THREAD_AFFINITY
 };

 //! The type of technique for stereoscopic rendering
@@ -75,19 +75,6 @@ protected:
    HandleBase(HandleBase const& rhs) noexcept = default;
    HandleBase& operator=(HandleBase const& rhs) noexcept = default;

-    HandleBase(HandleBase&& rhs) noexcept
-            : object(rhs.object) {
-        rhs.object = nullid;
-    }
-
-    HandleBase& operator=(HandleBase&& rhs) noexcept {
-        if (this != &rhs) {
-            object = rhs.object;
-            rhs.object = nullid;
-        }
-        return *this;
-    }
-
 private:
    HandleId object;
 };
@@ -102,10 +89,8 @@ struct Handle : public HandleBase {
    Handle() noexcept = default;

    Handle(Handle const& rhs) noexcept = default;
-    Handle(Handle&& rhs) noexcept = default;

    Handle& operator=(Handle const& rhs) noexcept = default;
-    Handle& operator=(Handle&& rhs) noexcept = default;

    explicit Handle(HandleId id) noexcept : HandleBase(id) { }

@@ -48,7 +48,7 @@ namespace filament::backend {
 * and optional user data:
 *
 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- * swapChain->setFrameScheduledCallback(nullptr, myFrameScheduledCallback);
+ * swapChain->setFrameScheduledCallback(myFrameScheduledCallback, nullptr);
 * if (renderer->beginFrame(swapChain)) {
 *     renderer->render(view);
 *     renderer->endFrame();
@@ -58,6 +58,8 @@ namespace filament::backend {
 * @remark Only Filament's Metal backend supports PresentCallables and frame callbacks. Other
 * backends ignore the callback (which will never be called) and proceed normally.
 *
+ * @remark The SwapChain::FrameScheduledCallback is called on an arbitrary thread.
+ *
 * Applications *must* call each PresentCallable they receive. Each PresentCallable represents a
 * frame that is waiting to be presented. If an application fails to call a PresentCallable, a
 * memory leak could occur. To "cancel" the presentation of a frame, pass false to the
@@ -84,9 +84,6 @@ public:
    // null terminating character.
    Program& shader(ShaderStage shader, void const* data, size_t size);

-    // sets the language of the shader sources provided with shader() (defaults to ESSL3)
-    Program& shaderLanguage(ShaderLanguage shaderLanguage);
-
    // Note: This is only needed for GLES3.0 backends, because the layout(binding=) syntax is
    //       not permitted in glsl. The backend needs a way to associate a uniform block
    //       to a binding point.
@@ -139,8 +136,6 @@ public:
    utils::CString const& getName() const noexcept { return mName; }
    utils::CString& getName() noexcept { return mName; }

-    auto const& getShaderLanguage() const { return mShaderLanguage; }
-
    utils::FixedCapacityVector<SpecializationConstant> const& getSpecializationConstants() const noexcept {
        return mSpecializationConstants;
    }
@@ -160,7 +155,6 @@ private:
    UniformBlockInfo mUniformBlocks = {};
    SamplerGroupInfo mSamplerGroups = {};
    ShaderSource mShadersSource;
-    ShaderLanguage mShaderLanguage = ShaderLanguage::ESSL3;
    utils::CString mName;
    uint64_t mCacheId{};
    utils::Invocable<utils::io::ostream&(utils::io::ostream& out)> mLogger;
@@ -140,23 +140,6 @@ public:
     */
    virtual uint32_t getDefaultFramebufferObject() noexcept;

-    /**
-     * Called by the backend when a frame starts.
-     * @param steady_clock_ns vsync time point on the monotonic clock
-     * @param refreshIntervalNs refresh interval in nanosecond
-     * @param frameId a frame id
-     */
-    virtual void beginFrame(
-            int64_t monotonic_clock_ns,
-            int64_t refreshIntervalNs,
-            uint32_t frameId) noexcept;
-
-    /**
-     * Called by the backend when a frame ends.
-     * @param frameId the frame id used in beginFrame
-     */
-    virtual void endFrame(
-            uint32_t frameId) noexcept;

    /**
     * Type of contexts available
@@ -208,12 +191,6 @@ public:
            utils::Invocable<void()> preContextChange,
            utils::Invocable<void(size_t index)> postContextChange) noexcept;

-    /**
-     * Called by the backend just before calling commit()
-     * @see commit()
-     */
-    virtual void preCommit() noexcept;
-
    /**
     * Called by the driver once the current frame finishes drawing. Typically, this should present
     * the drawSwapChain. This is for example where `eglMakeCurrent()` would be called.
@@ -22,10 +22,6 @@
 #include <backend/platforms/OpenGLPlatform.h>
 #include <backend/platforms/PlatformEGL.h>

-#include <utils/android/PerformanceHintManager.h>
-
-#include <chrono>
-
 #include <stddef.h>
 #include <stdint.h>

@@ -62,13 +58,6 @@ protected:

    void terminate() noexcept override;

-    void beginFrame(
-            int64_t monotonic_clock_ns,
-            int64_t refreshIntervalNs,
-            uint32_t frameId) noexcept override;
-
-    void preCommit() noexcept override;
-
    /**
     * Set the presentation time using `eglPresentationTimeANDROID`
     * @param presentationTimeInNanosecond
@@ -92,11 +81,6 @@ protected:
 private:
    int mOSVersion;
    ExternalStreamManagerAndroid& mExternalStreamManager;
-    utils::PerformanceHintManager mPerformanceHintManager;
-    utils::PerformanceHintManager::Session mPerformanceHintSession;
-
-    using clock = std::chrono::high_resolution_clock;
-    clock::time_point mStartTimeOfActualWork;
 };

 } // namespace filament::backend
@@ -82,7 +82,6 @@ public:
    void requestExit();

    // suspend or unsuspend the queue.
-    bool isPaused() const noexcept;
    void setPaused(bool paused);

    bool isExitRequested() const;
@@ -134,7 +134,7 @@ struct CommandType<void (Driver::*)(ARGS...)> {

    public:
        template<typename M, typename D>
-        static inline void execute(M&& method, D&& driver, CommandBase* base, intptr_t* next) {
+        static inline void execute(M&& method, D&& driver, CommandBase* base, intptr_t* next) noexcept {
            Command* self = static_cast<Command*>(base);
            *next = align(sizeof(Command));
 #if DEBUG_COMMAND_STREAM
@@ -168,7 +168,7 @@ struct CommandType<void (Driver::*)(ARGS...)> {

 class CustomCommand : public CommandBase {
    std::function<void()> mCommand;
-    static void execute(Driver&, CommandBase* base, intptr_t* next);
+    static void execute(Driver&, CommandBase* base, intptr_t* next) noexcept;
 public:
    inline CustomCommand(CustomCommand&& rhs) = default;
    inline explicit CustomCommand(std::function<void()> cmd)
@@ -76,7 +76,7 @@ public:
    // the fn function will execute a batch of driver commands
    // this gives the driver a chance to wrap their execution in a meaningful manner
    // the default implementation simply calls fn
-    virtual void execute(std::function<void(void)> const& fn);
+    virtual void execute(std::function<void(void)> const& fn) noexcept;

    // This is called on debug build, or when enabled manually on the backend thread side.
    virtual void debugCommandBegin(CommandStream* cmds,
@@ -133,13 +133,12 @@ DECL_DRIVER_API_0(tick)

 DECL_DRIVER_API_N(beginFrame,
        int64_t, monotonic_clock_ns,
-        int64_t, refreshIntervalNs,
        uint32_t, frameId)

 DECL_DRIVER_API_N(setFrameScheduledCallback,
        backend::SwapChainHandle, sch,
-        backend::CallbackHandler*, handler,
-        backend::FrameScheduledCallback&&, callback)
+        backend::FrameScheduledCallback, callback,
+        void*, user)

 DECL_DRIVER_API_N(setFrameCompletedCallback,
        backend::SwapChainHandle, sch,
@@ -304,7 +303,6 @@ DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedContentSupported)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isStereoSupported, backend::StereoscopicType, stereoscopicType)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isParallelShaderCompileSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthStencilResolveSupported)
-DECL_DRIVER_API_SYNCHRONOUS_N(bool, isDepthStencilBlitSupported, backend::TextureFormat, format)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedTexturesSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(uint8_t, getMaxDrawBuffers)
 DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getMaxUniformBufferSize)
@@ -65,7 +65,7 @@ public:
     *
     */
    template<typename D, typename ... ARGS>
-    Handle<D> allocateAndConstruct(ARGS&& ... args) {
+    Handle<D> allocateAndConstruct(ARGS&& ... args) noexcept {
        Handle<D> h{ allocateHandle<D>() };
        D* addr = handle_cast<D*>(h);
        new(addr) D(std::forward<ARGS>(args)...);
@@ -97,7 +97,7 @@ public:
     */
    template<typename D, typename B, typename ... ARGS>
    typename std::enable_if_t<std::is_base_of_v<B, D>, D>*
-    destroyAndConstruct(Handle<B> const& handle, ARGS&& ... args) {
+    destroyAndConstruct(Handle<B> const& handle, ARGS&& ... args) noexcept {
        assert_invariant(handle);
        D* addr = handle_cast<D*>(const_cast<Handle<B>&>(handle));
        assert_invariant(addr);
@@ -163,7 +163,7 @@ public:
    inline typename std::enable_if_t<
            std::is_pointer_v<Dp> &&
            std::is_base_of_v<B, typename std::remove_pointer_t<Dp>>, Dp>
-    handle_cast(Handle<B>& handle) {
+    handle_cast(Handle<B>& handle) noexcept {
        assert_invariant(handle);
        auto [p, tag] = handleToPointer(handle.getId());

@@ -185,7 +185,7 @@ public:
    inline typename std::enable_if_t<
            std::is_pointer_v<Dp> &&
            std::is_base_of_v<B, typename std::remove_pointer_t<Dp>>, Dp>
-    handle_cast(Handle<B> const& handle) {
+    handle_cast(Handle<B> const& handle) noexcept {
        return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
    }

@@ -317,7 +317,7 @@ private:
        return (id & HANDLE_HEAP_FLAG) == 0u;
    }

-    HandleBase::HandleId allocateHandleSlow(size_t size);
+    HandleBase::HandleId allocateHandleSlow(size_t size) noexcept;
    void deallocateHandleSlow(HandleBase::HandleId id, size_t size) noexcept;

    // We inline this because it's just 4 instructions in the fast case
@@ -57,11 +57,6 @@ void CommandBufferQueue::requestExit() {
    mCondition.notify_one();
 }

-bool CommandBufferQueue::isPaused() const noexcept {
-    std::lock_guard<utils::Mutex> const lock(mLock);
-    return mPaused;
-}
-
 void CommandBufferQueue::setPaused(bool paused) {
    std::lock_guard<utils::Mutex> const lock(mLock);
    if (paused) {
@@ -149,7 +149,7 @@ void CommandType<void (Driver::*)(ARGS...)>::Command<METHOD>::log() noexcept  {

 // ------------------------------------------------------------------------------------------------

-void CustomCommand::execute(Driver&, CommandBase* base, intptr_t* next) {
+void CustomCommand::execute(Driver&, CommandBase* base, intptr_t* next) noexcept {
    *next = CustomCommand::align(sizeof(CustomCommand));
    static_cast<CustomCommand*>(base)->mCommand();
    static_cast<CustomCommand*>(base)->~CustomCommand();
@@ -214,7 +214,7 @@ size_t Driver::getElementTypeSize(ElementType type) noexcept {

 Driver::~Driver() noexcept = default;

-void Driver::execute(std::function<void(void)> const& fn) {
+void Driver::execute(std::function<void(void)> const& fn) noexcept {
    fn();
 }

@@ -107,7 +107,7 @@ void* HandleAllocator<P0, P1, P2>::handleToPointerSlow(HandleBase::HandleId id)
 }

 template <size_t P0, size_t P1, size_t P2>
-HandleBase::HandleId HandleAllocator<P0, P1, P2>::allocateHandleSlow(size_t size) {
+HandleBase::HandleId HandleAllocator<P0, P1, P2>::allocateHandleSlow(size_t size) noexcept {
    void* p = ::malloc(size);
    std::unique_lock lock(mLock);

@@ -21,7 +21,7 @@ namespace filament::backend {
 using namespace utils;

 // We want these in the .cpp file, so they're not inlined (not worth it)
-Program::Program() noexcept {  // NOLINT(modernize-use-equals-default)
+Program::Program() noexcept { // NOLINT(modernize-use-equals-default)
 }

 Program::Program(Program&& rhs) noexcept = default;
@@ -47,11 +47,6 @@ Program& Program::shader(ShaderStage shader, void const* data, size_t size) {
    return *this;
 }

-Program& Program::shaderLanguage(ShaderLanguage shaderLanguage) {
-    mShaderLanguage = shaderLanguage;
-    return *this;
-}
-
 Program& Program::uniformBlockBindings(
        FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept {
    for (auto const& item : uniformBlockBindings) {
@@ -18,7 +18,6 @@
 #define TNT_FILAMENT_DRIVER_METALBUFFER_H

 #include "MetalContext.h"
-#include "MetalPlatform.h"

 #include <backend/DriverEnums.h>

@@ -34,50 +33,15 @@ namespace filament::backend {

 class TrackedMetalBuffer {
 public:
-
-    static constexpr size_t EXCESS_BUFFER_COUNT = 30000;
-
-    enum class Type {
-        NONE = 0,
-        GENERIC = 1,
-        RING = 2,
-        STAGING = 3,
-    };
-    static constexpr size_t TypeCount = 3;
-
-    static constexpr auto toIndex(Type t) {
-        assert_invariant(t != Type::NONE);
-        switch (t) {
-            case Type::NONE:
-            case Type::GENERIC:
-                return 0;
-            case Type::RING:
-                return 1;
-            case Type::STAGING:
-                return 2;
-        }
-    }
-
    TrackedMetalBuffer() noexcept : mBuffer(nil) {}
-    TrackedMetalBuffer(nullptr_t) noexcept : mBuffer(nil) {}
-    TrackedMetalBuffer(id<MTLBuffer> buffer, Type type) : mBuffer(buffer), mType(type) {
-        assert_invariant(type != Type::NONE);
+    TrackedMetalBuffer(id<MTLBuffer> buffer) noexcept : mBuffer(buffer) {
        if (buffer) {
-            aliveBuffers[toIndex(type)]++;
-            mType = type;
-            if (getAliveBuffers() >= EXCESS_BUFFER_COUNT) {
-                if (platform && platform->hasDebugUpdateStatFunc()) {
-                    platform->debugUpdateStat("filament.metal.excess_buffers_allocated",
-                            TrackedMetalBuffer::getAliveBuffers());
-                }
-            }
+            aliveBuffers++;
        }
    }
-
    ~TrackedMetalBuffer() {
        if (mBuffer) {
-            assert_invariant(mType != Type::NONE);
-            aliveBuffers[toIndex(mType)]--;
+            aliveBuffers--;
        }
    }

@@ -93,31 +57,18 @@ public:
    id<MTLBuffer> get() const noexcept { return mBuffer; }
    operator bool() const noexcept { return bool(mBuffer); }

-    static uint64_t getAliveBuffers() {
-        uint64_t sum = 0;
-        for (const auto& v : aliveBuffers) {
-            sum += v;
-        }
-        return sum;
-    }
-
-    static uint64_t getAliveBuffers(Type type) {
-        assert_invariant(type != Type::NONE);
-        return aliveBuffers[toIndex(type)];
-    }
-    static void setPlatform(MetalPlatform* p) { platform = p; }
+    static uint64_t getAliveBuffers() { return aliveBuffers; }

 private:
    void swap(TrackedMetalBuffer& other) noexcept {
-        std::swap(mBuffer, other.mBuffer);
-        std::swap(mType, other.mType);
+        id<MTLBuffer> temp = mBuffer;
+        mBuffer = other.mBuffer;
+        other.mBuffer = temp;
    }

    id<MTLBuffer> mBuffer;
-    Type mType = Type::NONE;

-    static MetalPlatform* platform;
-    static std::array<uint64_t, TypeCount> aliveBuffers;
+    static std::atomic<uint64_t> aliveBuffers;
 };

 class MetalBuffer {
@@ -220,8 +171,7 @@ public:
          mBufferOptions(options),
          mSlotSizeBytes(computeSlotSize(layout)),
          mSlotCount(slotCount) {
-        mBuffer = { [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions],
-            TrackedMetalBuffer::Type::RING };
+        mBuffer = [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions];
        assert_invariant(mBuffer);
    }

@@ -239,10 +189,9 @@ public:
            // If we already have an aux buffer, it will get freed here, unless it has been retained
            // by a MTLCommandBuffer. In that case, it will be freed when the command buffer
            // finishes executing.
-            mAuxBuffer = { [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions],
-                TrackedMetalBuffer::Type::RING };
+            mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
            assert_invariant(mAuxBuffer);
-            return { mAuxBuffer.get(), 0 };
+            return {mAuxBuffer.get(), 0};
        }
        mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
        mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);
@@ -22,8 +22,7 @@
 namespace filament {
 namespace backend {

-std::array<uint64_t, TrackedMetalBuffer::TypeCount> TrackedMetalBuffer::aliveBuffers = { 0 };
-MetalPlatform* TrackedMetalBuffer::platform = nullptr;
+std::atomic<uint64_t> TrackedMetalBuffer::aliveBuffers = 0;

 MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
        size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
@@ -38,8 +37,7 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
    }

    // Otherwise, we allocate a private GPU buffer.
-    mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
-        TrackedMetalBuffer::Type::GENERIC };
+    mBuffer = [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate];
    ASSERT_POSTCONDITION(mBuffer, "Could not allocate Metal buffer of size %zu.", size);
 }

@@ -46,7 +46,7 @@ MetalBufferPoolEntry const* MetalBufferPool::acquireBuffer(size_t numBytes) {
                                                        options:MTLResourceStorageModeShared];
    ASSERT_POSTCONDITION(buffer, "Could not allocate Metal staging buffer of size %zu.", numBytes);
    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry {
-        .buffer = { buffer, TrackedMetalBuffer::Type::STAGING },
+        .buffer = buffer,
        .capacity = numBytes,
        .lastAccessed = mCurrentFrame,
        .referenceCount = 1
@@ -105,8 +105,6 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
                  driverConfig.disableHandleUseAfterFreeCheck) {
    mContext->driver = this;

-    TrackedMetalBuffer::setPlatform(platform);
-
    mContext->device = mPlatform.createDevice();
    assert_invariant(mContext->device);

@@ -200,7 +198,6 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
 }

 MetalDriver::~MetalDriver() noexcept {
-    TrackedMetalBuffer::setPlatform(nullptr);
    mContext->device = nil;
    mContext->emptyTexture = nil;
    CFRelease(mContext->textureCache);
@@ -215,26 +212,19 @@ void MetalDriver::tick(int) {
    executeTickOps();
 }

-void MetalDriver::beginFrame(int64_t monotonic_clock_ns,
-        int64_t refreshIntervalNs, uint32_t frameId) {
+void MetalDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
 #if defined(FILAMENT_METAL_PROFILING)
    os_signpost_interval_begin(mContext->log, mContext->signpostId, "Frame encoding", "%{public}d", frameId);
 #endif
    if (mPlatform.hasDebugUpdateStatFunc()) {
        mPlatform.debugUpdateStat("filament.metal.alive_buffers", TrackedMetalBuffer::getAliveBuffers());
-        mPlatform.debugUpdateStat("filament.metal.alive_buffers.generic",
-                TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::GENERIC));
-        mPlatform.debugUpdateStat("filament.metal.alive_buffers.ring",
-                TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::RING));
-        mPlatform.debugUpdateStat("filament.metal.alive_buffers.staging",
-                TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::STAGING));
    }
 }

-void MetalDriver::setFrameScheduledCallback(
-        Handle<HwSwapChain> sch, CallbackHandler* handler, FrameScheduledCallback&& callback) {
+void MetalDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
+        FrameScheduledCallback callback, void* user) {
    auto* swapChain = handle_cast<MetalSwapChain>(sch);
-    swapChain->setFrameScheduledCallback(handler, std::move(callback));
+    swapChain->setFrameScheduledCallback(callback, user);
 }

 void MetalDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
@@ -815,10 +805,6 @@ bool MetalDriver::isDepthStencilResolveSupported() {
    return false;
 }

-bool MetalDriver::isDepthStencilBlitSupported(TextureFormat format) {
-    return true;
-}
-
 bool MetalDriver::isProtectedTexturesSupported() {
    return false;
 }
@@ -1648,7 +1634,7 @@ void MetalDriver::bindPipeline(PipelineState ps) {
        return;
    }

-    functions.validate();
+    ASSERT_PRECONDITION(bool(functions), "Attempting to bind an invalid Metal program.");

    auto [fragment, vertex] = functions.getRasterFunctions();

@@ -31,8 +31,6 @@

 #include "private/backend/SamplerGroup.h"

-#include <backend/DriverEnums.h>
-
 #include <utils/bitset.h>
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
@@ -73,9 +71,9 @@ public:

    void releaseDrawable();

-    void setFrameScheduledCallback(CallbackHandler* handler, FrameScheduledCallback&& callback);
-    void setFrameCompletedCallback(
-            CallbackHandler* handler, CallbackHandler::Callback callback, void* user);
+    void setFrameScheduledCallback(FrameScheduledCallback callback, void* user);
+    void setFrameCompletedCallback(CallbackHandler* handler,
+            CallbackHandler::Callback callback, void* user);

    // For CAMetalLayer-backed SwapChains, presents the drawable or schedules a
    // FrameScheduledCallback.
@@ -112,15 +110,14 @@ private:
    MetalExternalImage externalImage;
    SwapChainType type;

-    // These fields store a callback to notify the client that a frame is ready for presentation. If
-    // !frameScheduled.callback, then the Metal backend automatically calls presentDrawable when the
-    // frame is committed. Otherwise, the Metal backend will not automatically present the frame.
-    // Instead, clients bear the responsibility of presenting the frame by calling the
-    // PresentCallable object.
-    struct {
-        CallbackHandler* handler = nullptr;
-        FrameScheduledCallback callback = {};
-    } frameScheduled;
+    // These two fields store a callback and user data to notify the client that a frame is ready
+    // for presentation.
+    // If frameScheduledCallback is nullptr, then the Metal backend automatically calls
+    // presentDrawable when the frame is committed.
+    // Otherwise, the Metal backend will not automatically present the frame. Instead, clients bear
+    // the responsibility of presenting the frame by calling the PresentCallable object.
+    FrameScheduledCallback frameScheduledCallback = nullptr;
+    void* frameScheduledUserData = nullptr;

    struct {
        CallbackHandler* handler = nullptr;
@@ -221,10 +221,9 @@ void MetalSwapChain::ensureDepthStencilTexture() {
    depthStencilTexture = [context.device newTextureWithDescriptor:descriptor];
 }

-void MetalSwapChain::setFrameScheduledCallback(
-        CallbackHandler* handler, FrameScheduledCallback&& callback) {
-    frameScheduled.handler = handler;
-    frameScheduled.callback = std::move(callback);
+void MetalSwapChain::setFrameScheduledCallback(FrameScheduledCallback callback, void* user) {
+    frameScheduledCallback = callback;
+    frameScheduledUserData = user;
 }

 void MetalSwapChain::setFrameCompletedCallback(CallbackHandler* handler,
@@ -239,7 +238,7 @@ void MetalSwapChain::present() {
        scheduleFrameCompletedCallback();
    }
    if (drawable) {
-        if (frameScheduled.callback) {
+        if (frameScheduledCallback) {
            scheduleFrameScheduledCallback();
        } else  {
            [getPendingCommandBuffer(&context) presentDrawable:drawable];
@@ -297,38 +296,21 @@ void presentDrawable(bool presentFrame, void* user) {
 }

 void MetalSwapChain::scheduleFrameScheduledCallback() {
-    if (!frameScheduled.callback) {
+    if (!frameScheduledCallback) {
        return;
    }

    assert_invariant(drawable);

-    struct Callback {
-        Callback(FrameScheduledCallback&& callback, id<CAMetalDrawable> drawable,
-                MetalDriver* driver)
-            : f(std::move(callback)), data(PresentDrawableData::create(drawable, driver)) {}
-        FrameScheduledCallback f;
-        // PresentDrawableData* is destroyed by maybePresentAndDestroyAsync() later.
-        std::unique_ptr<PresentDrawableData> data;
-        static void func(void* user) {
-            auto* const c = reinterpret_cast<Callback*>(user);
-            PresentDrawableData* presentDrawableData = c->data.release();
-            PresentCallable presentCallable(presentDrawable, presentDrawableData);
-            c->f(presentCallable);
-            delete c;
-        }
-    };
+    // Destroy this by calling maybePresentAndDestroyAsync() later.
+    auto* presentData = PresentDrawableData::create(drawable, context.driver);

-    // This callback pointer will be captured by the block. Even if the scheduled handler is never
-    // called, the unique_ptr will still ensure we don't leak memory.
-    __block auto callback =
-            std::make_unique<Callback>(std::move(frameScheduled.callback), drawable, context.driver);
+    FrameScheduledCallback userCallback = frameScheduledCallback;
+    void* userData = frameScheduledUserData;

-    backend::CallbackHandler* handler = frameScheduled.handler;
-    MetalDriver* driver = context.driver;
    [getPendingCommandBuffer(&context) addScheduledHandler:^(id<MTLCommandBuffer> cb) {
-        Callback* user = callback.release();
-        driver->scheduleCallback(handler, user, &Callback::func);
+        PresentCallable callable(presentDrawable, static_cast<void*>(presentData));
+        userCallback(callable, userData);
    }];
 }

@@ -106,41 +106,26 @@ bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
            continue;
        }

+        assert_invariant(source[source.size() - 1] == '\0');
+
+        // the shader string is null terminated and the length includes the null character
+        NSString* objcSource = [[NSString alloc] initWithBytes:source.data()
+                                                        length:source.size() - 1
+                                                      encoding:NSUTF8StringEncoding];
+
+        // By default, Metal uses the most recent language version.
+        MTLCompileOptions* options = [MTLCompileOptions new];
+
+        // Disable Fast Math optimizations.
+        // This ensures that operations adhere to IEEE standards for floating-point arithmetic,
+        // which is crucial for half precision floats in scenarios where fast math optimizations
+        // lead to inaccuracies, such as in handling special values like NaN or Infinity.
+        options.fastMathEnabled = NO;
+
        NSError* error = nil;
-        id<MTLLibrary> library = nil;
-        switch (program.getShaderLanguage()) {
-            case ShaderLanguage::MSL: {
-                // By default, Metal uses the most recent language version.
-                MTLCompileOptions* options = [MTLCompileOptions new];
-
-                // Disable Fast Math optimizations.
-                // This ensures that operations adhere to IEEE standards for floating-point
-                // arithmetic, which is crucial for half precision floats in scenarios where fast
-                // math optimizations lead to inaccuracies, such as in handling special values like
-                // NaN or Infinity.
-                options.fastMathEnabled = NO;
-
-                assert_invariant(source[source.size() - 1] == '\0');
-                // the shader string is null terminated and the length includes the null character
-                NSString* objcSource = [[NSString alloc] initWithBytes:source.data()
-                                                                length:source.size() - 1
-                                                              encoding:NSUTF8StringEncoding];
-                library = [device newLibraryWithSource:objcSource options:options error:&error];
-                break;
-            }
-            case ShaderLanguage::METAL_LIBRARY: {
-                dispatch_data_t data = dispatch_data_create(source.data(), source.size(),
-                        dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0),
-                        DISPATCH_DATA_DESTRUCTOR_DEFAULT);
-                library = [device newLibraryWithData:data error:&error];
-                break;
-            }
-            case ShaderLanguage::ESSL1:
-            case ShaderLanguage::ESSL3:
-            case ShaderLanguage::SPIRV:
-                break;
-        }
-
+        id<MTLLibrary> library = [device newLibraryWithSource:objcSource
+                                                      options:options
+                                                        error:&error];
        if (library == nil) {
            NSString* errorMessage = @"unknown error";
            if (error) {
@@ -49,12 +49,11 @@ void NoopDriver::terminate() {
 void NoopDriver::tick(int) {
 }

-void NoopDriver::beginFrame(int64_t monotonic_clock_ns,
-        int64_t refreshIntervalNs, uint32_t frameId) {
+void NoopDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
 }

 void NoopDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, FrameScheduledCallback&& callback) {
+        FrameScheduledCallback callback, void* user) {

 }

@@ -194,10 +193,6 @@ bool NoopDriver::isDepthStencilResolveSupported() {
    return true;
 }

-bool NoopDriver::isDepthStencilBlitSupported(TextureFormat format) {
-    return true;
-}
-
 bool NoopDriver::isProtectedTexturesSupported() {
    return true;
 }
@@ -1,51 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_OPENGL_GLBUFFEROBJECT_H
-#define TNT_FILAMENT_BACKEND_OPENGL_GLBUFFEROBJECT_H
-
-#include "DriverBase.h"
-
-#include "gl_headers.h"
-
-#include <backend/DriverEnums.h>
-
-#include <stdint.h>
-
-namespace filament::backend {
-
-struct GLBufferObject : public HwBufferObject {
-    using HwBufferObject::HwBufferObject;
-    GLBufferObject(uint32_t size,
-            BufferObjectBinding bindingType, BufferUsage usage) noexcept
-            : HwBufferObject(size), usage(usage), bindingType(bindingType) {
-    }
-
-    struct {
-        GLuint id;
-        union {
-            GLenum binding;
-            void* buffer;
-        };
-    } gl;
-    BufferUsage usage;
-    BufferObjectBinding bindingType;
-    uint16_t age = 0;
-};
-
-} // namespace filament::backend
-
-#endif //TNT_FILAMENT_BACKEND_OPENGL_GLBUFFEROBJECT_H
@@ -1,55 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_OPENGL_GLTEXTURE_H
-#define TNT_FILAMENT_BACKEND_OPENGL_GLTEXTURE_H
-
-#include "DriverBase.h"
-
-#include "gl_headers.h"
-
-#include <backend/platforms/OpenGLPlatform.h>
-
-#include <stdint.h>
-
-namespace filament::backend {
-
-struct GLTexture : public HwTexture {
-    using HwTexture::HwTexture;
-    struct GL {
-        GL() noexcept : imported(false), sidecarSamples(1), reserved1(0) {}
-        GLuint id = 0;          // texture or renderbuffer id
-        GLenum target = 0;
-        GLenum internalFormat = 0;
-        GLuint sidecarRenderBufferMS = 0;  // multi-sample sidecar renderbuffer
-
-        // texture parameters go here too
-        GLfloat anisotropy = 1.0;
-        int8_t baseLevel = 127;
-        int8_t maxLevel = -1;
-        uint8_t reserved0 = 0;
-        bool imported           : 1;
-        uint8_t sidecarSamples  : 4;
-        uint8_t reserved1       : 3;
-    } gl;
-
-    OpenGLPlatform::ExternalTexture* externalTexture = nullptr;
-};
-
-
-} // namespace filament::backend
-
-#endif //TNT_FILAMENT_BACKEND_OPENGL_GLTEXTURE_H
@@ -64,8 +64,7 @@ bool OpenGLContext::queryOpenGLVersion(GLint* major, GLint* minor) noexcept {
 }

 OpenGLContext::OpenGLContext(OpenGLPlatform& platform) noexcept
-        : mPlatform(platform),
-          mSamplerMap(32) {
+        : mPlatform(platform) {

    state.vao.p = &mDefaultVAO;

@@ -253,23 +252,9 @@ OpenGLContext::OpenGLContext(OpenGLPlatform& platform) noexcept
 }

 OpenGLContext::~OpenGLContext() noexcept {
-    // note: this is called from the main thread. Can't do any GL calls.
    delete mTimerQueryFactory;
 }

-void OpenGLContext::terminate() noexcept {
-    // note: this is called from the backend thread
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-    if (!isES2()) {
-        for (auto& item: mSamplerMap) {
-            unbindSampler(item.second);
-            glDeleteSamplers(1, &item.second);
-        }
-        mSamplerMap.clear();
-    }
-#endif
-}
-
 void OpenGLContext::destroyWithContext(
        size_t index, std::function<void(OpenGLContext&)> const& closure) noexcept {
    if (index == 0) {
@@ -531,6 +516,8 @@ void OpenGLContext::initBugs(Bugs* bugs, Extensions const& exts,
            bugs->delay_fbo_destruction = true;
            // PowerVR seems to have no problem with this (which is good for us)
            bugs->allow_read_only_ancillary_feedback_loop = true;
+            // PowerVR has a shader compiler thread pinned on the last core
+            bugs->disable_thread_affinity = true;
        } else if (strstr(renderer, "Apple")) {
            // Apple GPU
        } else if (strstr(renderer, "Tegra") ||
@@ -881,28 +868,19 @@ default_case:
    }
 }

-void OpenGLContext::unbindTexture(
-        UTILS_UNUSED_IN_RELEASE GLenum target, GLuint texture_id) noexcept {
+void OpenGLContext::unbindTexture(GLenum target, GLuint texture_id) noexcept {
    // unbind this texture from all the units it might be bound to
    // no need unbind the texture from FBOs because we're not tracking that state (and there is
    // no need to).
+    const size_t index = getIndexForTextureTarget(target);
    UTILS_NOUNROLL
    for (GLuint unit = 0; unit < MAX_TEXTURE_UNIT_COUNT; unit++) {
-        if (state.textures.units[unit].id == texture_id) {
-            // if this texture is bound, it should be at the same target
-            assert_invariant(state.textures.units[unit].target == target);
-            unbindTextureUnit(unit);
+        if (state.textures.units[unit].targets[index].texture_id == texture_id) {
+            bindTexture(unit, target, (GLuint)0, index);
        }
    }
 }

-void OpenGLContext::unbindTextureUnit(GLuint unit) noexcept {
-    update_state(state.textures.units[unit].id, 0u, [&]() {
-        activeTexture(unit);
-        glBindTexture(state.textures.units[unit].target, 0u);
-    });
-}
-
 void OpenGLContext::unbindSampler(GLuint sampler) noexcept {
    // unbind this sampler from all the units it might be bound to
    UTILS_NOUNROLL  // clang generates >800B of code!!!
@@ -956,37 +934,6 @@ void OpenGLContext::deleteVertexArray(GLuint vao) noexcept {
    }
 }

-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-GLuint OpenGLContext::getSamplerSlow(SamplerParams params) const noexcept {
-    assert_invariant(mSamplerMap.find(params) == mSamplerMap.end());
-
-    using namespace GLUtils;
-
-    GLuint s;
-    glGenSamplers(1, &s);
-    glSamplerParameteri(s, GL_TEXTURE_MIN_FILTER,   (GLint)getTextureFilter(params.filterMin));
-    glSamplerParameteri(s, GL_TEXTURE_MAG_FILTER,   (GLint)getTextureFilter(params.filterMag));
-    glSamplerParameteri(s, GL_TEXTURE_WRAP_S,       (GLint)getWrapMode(params.wrapS));
-    glSamplerParameteri(s, GL_TEXTURE_WRAP_T,       (GLint)getWrapMode(params.wrapT));
-    glSamplerParameteri(s, GL_TEXTURE_WRAP_R,       (GLint)getWrapMode(params.wrapR));
-    glSamplerParameteri(s, GL_TEXTURE_COMPARE_MODE, (GLint)getTextureCompareMode(params.compareMode));
-    glSamplerParameteri(s, GL_TEXTURE_COMPARE_FUNC, (GLint)getTextureCompareFunc(params.compareFunc));
-
-#if defined(GL_EXT_texture_filter_anisotropic)
-    if (ext.EXT_texture_filter_anisotropic &&
-        !bugs.texture_filter_anisotropic_broken_on_sampler) {
-        GLfloat const anisotropy = float(1u << params.anisotropyLog2);
-        glSamplerParameterf(s, GL_TEXTURE_MAX_ANISOTROPY_EXT,
-                std::min(gets.max_anisotropy, anisotropy));
-    }
-#endif
-    CHECK_GL_ERROR(utils::slog.e)
-    mSamplerMap[params] = s;
-    return s;
-}
-#endif
-
-
 void OpenGLContext::resetState() noexcept {
    // Force GL state to match the Filament state

@@ -34,8 +34,6 @@
 #include <math/vec2.h>
 #include <math/vec4.h>

-#include <tsl/robin_map.h>
-
 #include <array>
 #include <functional>
 #include <optional>
@@ -92,11 +90,8 @@ public:
    static bool queryOpenGLVersion(GLint* major, GLint* minor) noexcept;

    explicit OpenGLContext(OpenGLPlatform& platform) noexcept;
-
    ~OpenGLContext() noexcept final;

-    void terminate() noexcept;
-
    // TimerQueryInterface ------------------------------------------------------------------------

    // note: OpenGLContext being final ensures (clang) these are not called through the vtable
@@ -137,6 +132,7 @@ public:
 #endif
    }

+    constexpr static inline size_t getIndexForTextureTarget(GLuint target) noexcept;
    constexpr        inline size_t getIndexForCap(GLenum cap) noexcept;
    constexpr static inline size_t getIndexForBufferTarget(GLenum target) noexcept;

@@ -148,10 +144,10 @@ public:

          void pixelStore(GLenum, GLint) noexcept;
    inline void activeTexture(GLuint unit) noexcept;
+    inline void bindTexture(GLuint unit, GLuint target, GLuint texId, size_t targetIndex) noexcept;
    inline void bindTexture(GLuint unit, GLuint target, GLuint texId) noexcept;

           void unbindTexture(GLenum target, GLuint id) noexcept;
-           void unbindTextureUnit(GLuint unit) noexcept;
    inline void bindVertexArray(RenderPrimitive const* p) noexcept;
    inline void bindSampler(GLuint unit, GLuint sampler) noexcept;
           void unbindSampler(GLuint sampler) noexcept;
@@ -310,6 +306,10 @@ public:
        // a glFinish. So we must delay the destruction until we know the GPU is finished.
        bool delay_fbo_destruction;

+        // The driver has some threads pinned, and we can't easily know on which core, it can hurt
+        // performance more if we end-up pinned on the same one.
+        bool disable_thread_affinity;
+
        // Force feature level 0. Typically used for low end ES3 devices with significant driver
        // bugs or performance issues.
        bool force_feature_level0;
@@ -321,14 +321,8 @@ public:

    // function to handle state changes we don't control
    void updateTexImage(GLenum target, GLuint id) noexcept {
-        assert_invariant(target == GL_TEXTURE_EXTERNAL_OES);
-        // if another target is bound to this texture unit, unbind that texture
-        if (UTILS_UNLIKELY(state.textures.units[state.textures.active].target != target)) {
-            glBindTexture(state.textures.units[state.textures.active].target, 0);
-            state.textures.units[state.textures.active].target = GL_TEXTURE_EXTERNAL_OES;
-        }
-        // the texture is already bound to `target`, we just update our internal state
-        state.textures.units[state.textures.active].id = id;
+        const size_t index = getIndexForTextureTarget(target);
+        state.textures.units[state.textures.active].targets[index].texture_id = id;
    }
    void resetProgram() noexcept { state.program.use = 0; }

@@ -431,8 +425,9 @@ public:
            GLuint active = 0;      // zero-based
            struct {
                GLuint sampler = 0;
-                GLuint target = 0;
-                GLuint id = 0;
+                struct {
+                    GLuint texture_id = 0;
+                } targets[7];  // this must match getIndexForTextureTarget()
            } units[MAX_TEXTURE_UNIT_COUNT];
        } textures;

@@ -472,29 +467,6 @@ public:

    void unbindEverything() noexcept;
    void synchronizeStateAndCache(size_t index) noexcept;
-    void setEs2UniformBinding(size_t index, GLuint id, void const* data, uint16_t age) noexcept {
-        mUniformBindings[index] = { id, data, age };
-    }
-    auto getEs2UniformBinding(size_t index) const noexcept {
-        return mUniformBindings[index];
-    }
-
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-    GLuint getSamplerSlow(SamplerParams sp) const noexcept;
-
-    inline GLuint getSampler(SamplerParams sp) const noexcept {
-        assert_invariant(!sp.padding0);
-        assert_invariant(!sp.padding1);
-        assert_invariant(!sp.padding2);
-        auto& samplerMap = mSamplerMap;
-        auto pos = samplerMap.find(sp);
-        if (UTILS_UNLIKELY(pos == samplerMap.end())) {
-            return getSamplerSlow(sp);
-        }
-        return pos->second;
-    }
-#endif
-

 private:
    OpenGLPlatform& mPlatform;
@@ -504,11 +476,6 @@ private:
    std::vector<std::function<void(OpenGLContext&)>> mDestroyWithNormalContext;
    RenderPrimitive mDefaultVAO;
    std::optional<GLuint> mDefaultFbo[2];
-    std::array<
-            std::tuple<GLuint, void const*, uint16_t>,
-            CONFIG_UNIFORM_BINDING_COUNT> mUniformBindings = {};
-    mutable tsl::robin_map<SamplerParams, GLuint,
-            SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;

    void bindFramebufferResolved(GLenum target, GLuint buffer) noexcept;

@@ -555,6 +522,9 @@ private:
            {   bugs.delay_fbo_destruction,
                    "delay_fbo_destruction",
                    ""},
+            {   bugs.disable_thread_affinity,
+                    "disable_thread_affinity",
+                    ""},
            {   bugs.force_feature_level0,
                    "force_feature_level0",
                    ""},
@@ -602,10 +572,31 @@ private:
    }

    void setDefaultState() noexcept;
+
+    static constexpr const size_t TEXTURE_TARGET_COUNT =
+            sizeof(state.textures.units[0].targets) / sizeof(state.textures.units[0].targets[0]);
+
 };

 // ------------------------------------------------------------------------------------------------

+constexpr size_t OpenGLContext::getIndexForTextureTarget(GLuint target) noexcept {
+    // this must match state.textures[].targets[]
+    switch (target) {
+        case GL_TEXTURE_2D:                     return 0;
+        case GL_TEXTURE_2D_ARRAY:               return 1;
+        case GL_TEXTURE_CUBE_MAP:               return 2;
+#if defined(BACKEND_OPENGL_LEVEL_GLES31)
+        case GL_TEXTURE_2D_MULTISAMPLE:         return 3;
+#endif
+        case GL_TEXTURE_EXTERNAL_OES:           return 4;
+        case GL_TEXTURE_3D:                     return 5;
+        case GL_TEXTURE_CUBE_MAP_ARRAY:         return 6;
+        default:
+            return 0;
+    }
+}
+
 constexpr size_t OpenGLContext::getIndexForCap(GLenum cap) noexcept { //NOLINT
    size_t index = 0;
    switch (cap) {
@@ -753,17 +744,19 @@ void OpenGLContext::bindBufferRange(GLenum target, GLuint index, GLuint buffer,
 #endif
 }

-void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId) noexcept {
-    update_state(state.textures.units[unit].target, target, [&]() {
-        activeTexture(unit);
-        glBindTexture(state.textures.units[unit].target, 0);
-    });
-    update_state(state.textures.units[unit].id, texId, [&]() {
+void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId, size_t targetIndex) noexcept {
+    assert_invariant(targetIndex == getIndexForTextureTarget(target));
+    assert_invariant(targetIndex < TEXTURE_TARGET_COUNT);
+    update_state(state.textures.units[unit].targets[targetIndex].texture_id, texId, [&]() {
        activeTexture(unit);
        glBindTexture(target, texId);
    }, target == GL_TEXTURE_EXTERNAL_OES);
 }

+void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId) noexcept {
+    bindTexture(unit, target, texId, getIndexForTextureTarget(target));
+}
+
 void OpenGLContext::useProgram(GLuint program) noexcept {
    update_state(state.program.use, program, [&]() {
        glUseProgram(program);
@@ -208,6 +208,7 @@ OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfi
          mHandleAllocator("Handles",
                  driverConfig.handleArenaSize,
                  driverConfig.disableHandleUseAfterFreeCheck),
+          mSamplerMap(32),
          mDriverConfig(driverConfig) {

    std::fill(mSamplerBindings.begin(), mSamplerBindings.end(), nullptr);
@@ -232,7 +233,6 @@ OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfi
 }

 OpenGLDriver::~OpenGLDriver() noexcept { // NOLINT(modernize-use-equals-default)
-    // this is called from the main thread. Can't call GL.
 }

 Dispatcher OpenGLDriver::getDispatcher() const noexcept {
@@ -263,9 +263,15 @@ void OpenGLDriver::terminate() {

    // because we called glFinish(), all callbacks should have been executed
    assert_invariant(mGpuCommandCompleteOps.empty());
-#endif

-    mContext.terminate();
+    if (!getContext().isES2()) {
+        for (auto& item: mSamplerMap) {
+            mContext.unbindSampler(item.second);
+            glDeleteSamplers(1, &item.second);
+        }
+        mSamplerMap.clear();
+    }
+#endif

    mPlatform.terminate();
 }
@@ -287,7 +293,7 @@ void OpenGLDriver::bindSampler(GLuint unit, GLuint sampler) noexcept {

 void OpenGLDriver::bindTexture(GLuint unit, GLTexture const* t) noexcept {
    assert_invariant(t != nullptr);
-    mContext.bindTexture(unit, t->gl.target, t->gl.id);
+    mContext.bindTexture(unit, t->gl.target, t->gl.id, t->gl.targetIndex);
 }

 bool OpenGLDriver::useProgram(OpenGLProgram* p) noexcept {
@@ -301,7 +307,7 @@ bool OpenGLDriver::useProgram(OpenGLProgram* p) noexcept {

    if (UTILS_UNLIKELY(mContext.isES2())) {
        for (uint32_t i = 0; i < Program::UNIFORM_BINDING_COUNT; i++) {
-            auto [id, buffer, age] = mContext.getEs2UniformBinding(i);
+            auto [id, buffer, age] = mUniformBindings[i];
            if (buffer) {
                p->updateUniforms(i, id, buffer, age);
            }
@@ -718,19 +724,10 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
    GLenum internalFormat = getInternalFormat(format);
    assert_invariant(internalFormat);

-    if (any(usage & TextureUsage::PROTECTED)) {
+    if (UTILS_UNLIKELY(usage & TextureUsage::PROTECTED)) {
        // renderbuffers don't have a protected mode, so we need to use a texture
        // because protected textures are only supported on GLES 3.2, MSAA will be available.
        usage |= TextureUsage::SAMPLEABLE;
-    } else if (any(usage & TextureUsage::UPLOADABLE)) {
-        // if we have the uploadable flag, we also need to use a texture
-        usage |= TextureUsage::SAMPLEABLE;
-    } else if (target != SamplerType::SAMPLER_2D) {
-        // renderbuffers can only be 2D
-        usage |= TextureUsage::SAMPLEABLE;
-    } else if (levels > 1) {
-        // renderbuffers can't have mip levels
-        usage |= TextureUsage::SAMPLEABLE;
    }

    auto& gl = mContext;
@@ -749,6 +746,7 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
            if (t->externalTexture) {
                t->gl.target = t->externalTexture->target;
                t->gl.id = t->externalTexture->id;
+                t->gl.targetIndex = (uint8_t)OpenGLContext::getIndexForTextureTarget(t->gl.target);
                // internalFormat actually depends on the external image, but it doesn't matter
                // because it's not used anywhere for anything important.
                t->gl.internalFormat = internalFormat;
@@ -760,23 +758,30 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint

            t->gl.internalFormat = internalFormat;

+            // We DO NOT update targetIndex at function exit to take advantage of the fact that
+            // getIndexForTextureTarget() is constexpr -- so all of this disappears at compile time.
            switch (target) {
                case SamplerType::SAMPLER_EXTERNAL:
                    // we can't be here -- doesn't matter what we do
                case SamplerType::SAMPLER_2D:
                    t->gl.target = GL_TEXTURE_2D;
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D);
                    break;
                case SamplerType::SAMPLER_3D:
                    t->gl.target = GL_TEXTURE_3D;
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_3D);
                    break;
                case SamplerType::SAMPLER_2D_ARRAY:
                    t->gl.target = GL_TEXTURE_2D_ARRAY;
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
                    break;
                case SamplerType::SAMPLER_CUBEMAP:
                    t->gl.target = GL_TEXTURE_CUBE_MAP;
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
                    break;
                case SamplerType::SAMPLER_CUBEMAP_ARRAY:
                    t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
                    break;
            }

@@ -787,6 +792,8 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
                if (gl.features.multisample_texture) {
                    // multi-sample texture on GL 3.2 / GLES 3.1 and above
                    t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
+                    t->gl.targetIndex = (uint8_t)
+                            OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
                } else {
                    // Turn off multi-sampling for that texture. It's just not supported.
                }
@@ -796,6 +803,13 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
            textureStorage(t, w, h, depth, bool(usage & TextureUsage::PROTECTED));
        }
    } else {
+        assert_invariant(any(usage & (
+                TextureUsage::COLOR_ATTACHMENT |
+                TextureUsage::DEPTH_ATTACHMENT |
+                TextureUsage::STENCIL_ATTACHMENT)));
+        assert_invariant(levels == 1);
+        assert_invariant(target == SamplerType::SAMPLER_2D);
+        assert_invariant(none(usage & TextureUsage::PROTECTED));
        t->gl.internalFormat = internalFormat;
        t->gl.target = GL_RENDERBUFFER;
        glGenRenderbuffers(1, &t->gl.id);
@@ -845,24 +859,32 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
    t->gl.internalFormat = getInternalFormat(format);
    assert_invariant(t->gl.internalFormat);

+    // We DO NOT update targetIndex at function exit to take advantage of the fact that
+    // getIndexForTextureTarget() is constexpr -- so all of this disappears at compile time.
    switch (target) {
        case SamplerType::SAMPLER_EXTERNAL:
            t->gl.target = GL_TEXTURE_EXTERNAL_OES;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_EXTERNAL_OES);
            break;
        case SamplerType::SAMPLER_2D:
            t->gl.target = GL_TEXTURE_2D;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D);
            break;
        case SamplerType::SAMPLER_3D:
            t->gl.target = GL_TEXTURE_3D;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_3D);
            break;
        case SamplerType::SAMPLER_2D_ARRAY:
            t->gl.target = GL_TEXTURE_2D_ARRAY;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
            break;
        case SamplerType::SAMPLER_CUBEMAP:
            t->gl.target = GL_TEXTURE_CUBE_MAP;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
            break;
        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
            t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
            break;
    }

@@ -873,6 +895,7 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
        if (gl.features.multisample_texture) {
            // multi-sample texture on GL 3.2 / GLES 3.1 and above
            t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
        } else {
            // Turn off multi-sampling for that texture. It's just not supported.
        }
@@ -1762,6 +1785,7 @@ void OpenGLDriver::updateStreams(DriverApi* driver) {
                        // the target and id can be reset each time
                        t->gl.target = t->externalTexture->target;
                        t->gl.id = t->externalTexture->id;
+                        t->gl.targetIndex = (uint8_t)OpenGLContext::getIndexForTextureTarget(t->gl.target);
                        bindTexture(OpenGLContext::DUMMY_TEXTURE_BINDING, t);
                    }
                }
@@ -2023,10 +2047,6 @@ bool OpenGLDriver::isDepthStencilResolveSupported() {
    return true;
 }

-bool OpenGLDriver::isDepthStencilBlitSupported(TextureFormat format) {
-    return true;
-}
-
 bool OpenGLDriver::isProtectedTexturesSupported() {
    return getContext().ext.EXT_protected_textures;
 }
@@ -2043,6 +2063,8 @@ bool OpenGLDriver::isWorkaroundNeeded(Workaround workaround) {
            return mContext.bugs.disable_blit_into_texture_array;
        case Workaround::POWER_VR_SHADER_WORKAROUNDS:
            return mContext.bugs.powervr_shader_workarounds;
+        case Workaround::DISABLE_THREAD_AFFINITY:
+            return mContext.bugs.disable_thread_affinity;
        default:
            return false;
    }
@@ -2321,7 +2343,7 @@ void OpenGLDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
 #endif
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
            if (UTILS_LIKELY(!es2)) {
-                samplerId = mContext.getSampler(params);
+                samplerId = getSampler(params);
            } else
 #endif
            {
@@ -2651,6 +2673,7 @@ void OpenGLDriver::setExternalImage(Handle<HwTexture> th, void* image) {
        // the target and id can be reset each time
        t->gl.target = t->externalTexture->target;
        t->gl.id = t->externalTexture->id;
+        t->gl.targetIndex = (uint8_t)OpenGLContext::getIndexForTextureTarget(t->gl.target);
        bindTexture(OpenGLContext::DUMMY_TEXTURE_BINDING, t);
    }
 }
@@ -3045,10 +3068,11 @@ void OpenGLDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t ind
    assert_invariant(offset + size <= ub->byteCount);

    if (UTILS_UNLIKELY(ub->bindingType == BufferObjectBinding::UNIFORM && gl.isES2())) {
-        gl.setEs2UniformBinding(index,
+        mUniformBindings[index] = {
                ub->gl.id,
                static_cast<uint8_t const*>(ub->gl.buffer) + offset,
-                ub->age);
+                ub->age,
+        };
    } else {
        GLenum const target = GLUtils::getBufferBindingType(bindingType);

@@ -3066,7 +3090,7 @@ void OpenGLDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index)
    auto& gl = mContext;

    if (UTILS_UNLIKELY(bindingType == BufferObjectBinding::UNIFORM && gl.isES2())) {
-        gl.setEs2UniformBinding(index, 0, nullptr, 0);
+        mUniformBindings[index] = {};
        return;
    }

@@ -3083,6 +3107,36 @@ void OpenGLDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
    CHECK_GL_ERROR(utils::slog.e)
 }

+
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+GLuint OpenGLDriver::getSamplerSlow(SamplerParams params) const noexcept {
+    assert_invariant(mSamplerMap.find(params) == mSamplerMap.end());
+
+    GLuint s;
+    glGenSamplers(1, &s);
+    glSamplerParameteri(s, GL_TEXTURE_MIN_FILTER,   (GLint)getTextureFilter(params.filterMin));
+    glSamplerParameteri(s, GL_TEXTURE_MAG_FILTER,   (GLint)getTextureFilter(params.filterMag));
+    glSamplerParameteri(s, GL_TEXTURE_WRAP_S,       (GLint)getWrapMode(params.wrapS));
+    glSamplerParameteri(s, GL_TEXTURE_WRAP_T,       (GLint)getWrapMode(params.wrapT));
+    glSamplerParameteri(s, GL_TEXTURE_WRAP_R,       (GLint)getWrapMode(params.wrapR));
+    glSamplerParameteri(s, GL_TEXTURE_COMPARE_MODE, (GLint)getTextureCompareMode(params.compareMode));
+    glSamplerParameteri(s, GL_TEXTURE_COMPARE_FUNC, (GLint)getTextureCompareFunc(params.compareFunc));
+
+#if defined(GL_EXT_texture_filter_anisotropic)
+    auto& gl = mContext;
+    if (gl.ext.EXT_texture_filter_anisotropic &&
+            !gl.bugs.texture_filter_anisotropic_broken_on_sampler) {
+        GLfloat const anisotropy = float(1u << params.anisotropyLog2);
+        glSamplerParameterf(s, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                std::min(gl.gets.max_anisotropy, anisotropy));
+    }
+#endif
+    CHECK_GL_ERROR(utils::slog.e)
+    mSamplerMap[params] = s;
+    return s;
+}
+#endif
+
 void OpenGLDriver::insertEventMarker(char const* string, uint32_t len) {
 #ifndef __EMSCRIPTEN__
 #ifdef GL_EXT_debug_marker
@@ -3389,12 +3443,10 @@ void OpenGLDriver::tick(int) {

 void OpenGLDriver::beginFrame(
        UTILS_UNUSED int64_t monotonic_clock_ns,
-        UTILS_UNUSED int64_t refreshIntervalNs,
        UTILS_UNUSED uint32_t frameId) {
    DEBUG_MARKER()
    auto& gl = mContext;
    insertEventMarker("beginFrame");
-    mPlatform.beginFrame(monotonic_clock_ns, refreshIntervalNs, frameId);
    if (UTILS_UNLIKELY(!mTexturesWithStreamsAttached.empty())) {
        OpenGLPlatform& platform = mPlatform;
        for (GLTexture const* t : mTexturesWithStreamsAttached) {
@@ -3403,7 +3455,7 @@ void OpenGLDriver::beginFrame(
                assert_invariant(t->hwStream->stream);
                platform.updateTexImage(t->hwStream->stream,
                        &static_cast<GLStream*>(t->hwStream)->user_thread.timestamp); // NOLINT(cppcoreguidelines-pro-type-static-cast-downcast)
-                // NOTE: We assume that OpenGLPlatform::updateTexImage() binds the texture on our behalf
+                // NOTE: We assume that updateTexImage() binds the texture on our behalf
                gl.updateTexImage(GL_TEXTURE_EXTERNAL_OES, t->gl.id);
            }
        }
@@ -3411,7 +3463,7 @@ void OpenGLDriver::beginFrame(
 }

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

@@ -3442,7 +3494,6 @@ void OpenGLDriver::endFrame(UTILS_UNUSED uint32_t frameId) {
 #endif
    //SYSTRACE_NAME("glFinish");
    //glFinish();
-    mPlatform.endFrame(frameId);
    insertEventMarker("endFrame");
 }

@@ -20,8 +20,6 @@
 #include "DriverBase.h"
 #include "OpenGLContext.h"
 #include "OpenGLTimerQuery.h"
-#include "GLBufferObject.h"
-#include "GLTexture.h"
 #include "ShaderCompilerService.h"

 #include <backend/platforms/OpenGLPlatform.h>
@@ -94,6 +92,25 @@ public:
        bool rec709 = false;
    };

+    struct GLBufferObject : public HwBufferObject {
+        using HwBufferObject::HwBufferObject;
+        GLBufferObject(uint32_t size,
+                BufferObjectBinding bindingType, BufferUsage usage) noexcept
+                : HwBufferObject(size), usage(usage), bindingType(bindingType) {
+        }
+
+        struct {
+            GLuint id;
+            union {
+                GLenum binding;
+                void* buffer;
+            };
+        } gl;
+        BufferUsage usage;
+        BufferObjectBinding bindingType;
+        uint16_t age = 0;
+    };
+
    struct GLVertexBufferInfo : public HwVertexBufferInfo {
        GLVertexBufferInfo() noexcept = default;
        GLVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
@@ -123,6 +140,7 @@ public:
        } gl;
    };

+    struct GLTexture;
    struct GLSamplerGroup : public HwSamplerGroup {
        using HwSamplerGroup::HwSamplerGroup;
        struct Entry {
@@ -139,9 +157,27 @@ public:
        Handle<HwVertexBufferInfo> vbih;
    };

-    using GLBufferObject = filament::backend::GLBufferObject;
+    struct GLTexture : public HwTexture {
+        using HwTexture::HwTexture;
+        struct GL {
+            GL() noexcept : imported(false), sidecarSamples(1), reserved(0) {}
+            GLuint id = 0;          // texture or renderbuffer id
+            GLenum target = 0;
+            GLenum internalFormat = 0;
+            GLuint sidecarRenderBufferMS = 0;  // multi-sample sidecar renderbuffer

-    using GLTexture = filament::backend::GLTexture;
+            // texture parameters go here too
+            GLfloat anisotropy = 1.0;
+            int8_t baseLevel = 127;
+            int8_t maxLevel = -1;
+            uint8_t targetIndex = 0;    // optimization: index corresponding to target
+            bool imported           : 1;
+            uint8_t sidecarSamples  : 4;
+            uint8_t reserved        : 3;
+        } gl;
+
+        OpenGLPlatform::ExternalTexture* externalTexture = nullptr;
+    };

    using GLTimerQuery = filament::backend::GLTimerQuery;

@@ -232,13 +268,13 @@ private:
    HandleAllocatorGL mHandleAllocator;

    template<typename D, typename ... ARGS>
-    Handle<D> initHandle(ARGS&& ... args) {
+    Handle<D> initHandle(ARGS&& ... args) noexcept {
        return mHandleAllocator.allocateAndConstruct<D>(std::forward<ARGS>(args) ...);
    }

    template<typename D, typename B, typename ... ARGS>
    typename std::enable_if<std::is_base_of<B, D>::value, D>::type*
-    construct(Handle<B> const& handle, ARGS&& ... args) {
+    construct(Handle<B> const& handle, ARGS&& ... args) noexcept {
        return mHandleAllocator.destroyAndConstruct<D, B>(handle, std::forward<ARGS>(args) ...);
    }

@@ -252,7 +288,7 @@ private:
    typename std::enable_if_t<
            std::is_pointer_v<Dp> &&
            std::is_base_of_v<B, typename std::remove_pointer_t<Dp>>, Dp>
-    handle_cast(Handle<B>& handle) {
+    handle_cast(Handle<B>& handle) noexcept {
        return mHandleAllocator.handle_cast<Dp, B>(handle);
    }

@@ -260,7 +296,7 @@ private:
    inline typename std::enable_if_t<
            std::is_pointer_v<Dp> &&
            std::is_base_of_v<B, typename std::remove_pointer_t<Dp>>, Dp>
-    handle_cast(Handle<B> const& handle) {
+    handle_cast(Handle<B> const& handle) noexcept {
        return mHandleAllocator.handle_cast<Dp, B>(handle);
    }

@@ -312,6 +348,22 @@ private:
    void resolvePass(ResolveAction action, GLRenderTarget const* rt,
            TargetBufferFlags discardFlags) noexcept;

+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+    GLuint getSamplerSlow(SamplerParams sp) const noexcept;
+
+    inline GLuint getSampler(SamplerParams sp) const noexcept {
+        assert_invariant(!sp.padding0);
+        assert_invariant(!sp.padding1);
+        assert_invariant(!sp.padding2);
+        auto& samplerMap = mSamplerMap;
+        auto pos = samplerMap.find(sp);
+        if (UTILS_UNLIKELY(pos == samplerMap.end())) {
+            return getSamplerSlow(sp);
+        }
+        return pos->second;
+    }
+#endif
+
    const std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT>& getSamplerBindings() const {
        return mSamplerBindings;
    }
@@ -338,10 +390,14 @@ private:

    // ES2 only. Uniform buffer emulation binding points
    GLuint mLastAssignedEmulatedUboId = 0;
+    std::array<std::tuple<GLuint, void const*, uint16_t>, Program::UNIFORM_BINDING_COUNT> mUniformBindings = {};

    // sampler buffer binding points (nullptr if not used)
    std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT> mSamplerBindings = {};   // 4 pointers

+    mutable tsl::robin_map<SamplerParams, GLuint,
+            SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;
+
    // this must be accessed from the driver thread only
    std::vector<GLTexture*> mTexturesWithStreamsAttached;

@@ -55,16 +55,6 @@ uint32_t OpenGLPlatform::getDefaultFramebufferObject() noexcept {
    return 0;
 }

-void OpenGLPlatform::beginFrame(int64_t monotonic_clock_ns, int64_t refreshIntervalNs,
-        uint32_t frameId) noexcept {
-}
-
-void OpenGLPlatform::endFrame(uint32_t frameId) noexcept {
-}
-
-void OpenGLPlatform::preCommit() noexcept {
-}
-
 OpenGLPlatform::ContextType OpenGLPlatform::getCurrentContextType() const noexcept {
    return ContextType::UNPROTECTED;
 }
@@ -70,14 +70,14 @@ public:
        }
    }

+    struct {
+        GLuint program = 0;
+    } gl; // 12 bytes
+
    // For ES2 only
    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept;
    void setRec709ColorSpace(bool rec709) const noexcept;

-    struct {
-        GLuint program = 0;
-    } gl;                                               // 4 bytes
-
 private:
    // keep these away from of other class attributes
    struct LazyInitializationData;
@@ -89,17 +89,14 @@ private:

    void updateSamplers(OpenGLDriver* gld) const noexcept;

+    ShaderCompilerService::program_token_t mToken{};
+
    // number of bindings actually used by this program
+    uint8_t mUsedBindingsCount = 0u;
+    UTILS_UNUSED uint8_t padding[3] = {};
    std::array<uint8_t, Program::SAMPLER_BINDING_COUNT> mUsedSamplerBindingPoints;   // 4 bytes

-    ShaderCompilerService::program_token_t mToken{};    // 16 bytes
-
-    uint8_t mUsedBindingsCount = 0u;                    // 1 byte
-    UTILS_UNUSED uint8_t padding[3] = {};               // 3 bytes
-
-
    // only needed for ES2
-    GLint mRec709Location = -1; // 4 bytes
    using LocationInfo = utils::FixedCapacityVector<GLint>;
    struct UniformsRecord {
        Program::UniformInfo uniforms;
@@ -108,10 +105,11 @@ private:
        mutable uint16_t age = std::numeric_limits<uint16_t>::max();
    };
    UniformsRecord const* mUniformsRecords = nullptr;
+    GLint mRec709Location = -1;
 };

 // if OpenGLProgram is larger tha 64 bytes, it'll fall in a larger Handle bucket.
-static_assert(sizeof(OpenGLProgram) <= 64); // currently 48 bytes
+static_assert(sizeof(OpenGLProgram) <= 64);

 } // namespace filament::backend

@@ -25,8 +25,6 @@
 #include <android/api-level.h>
 #include <android/hardware_buffer.h>

-#include <utils/android/PerformanceHintManager.h>
-
 #include <utils/compiler.h>
 #include <utils/ostream.h>
 #include <utils/Log.h>
@@ -38,11 +36,8 @@

 #include <jni.h>

-#include <chrono>
 #include <new>

-#include <unistd.h>
-
 #include <stddef.h>
 #include <stdint.h>
 #include <stdlib.h>
@@ -117,38 +112,8 @@ void PlatformEGLAndroid::terminate() noexcept {
    PlatformEGL::terminate();
 }

-void PlatformEGLAndroid::beginFrame(
-        int64_t monotonic_clock_ns,
-        int64_t refreshIntervalNs,
-        uint32_t frameId) noexcept {
-    if (mPerformanceHintSession.isValid()) {
-        if (refreshIntervalNs <= 0) {
-            // we're not provided with a target time, assume 16.67ms
-            refreshIntervalNs = 16'666'667;
-        }
-        mStartTimeOfActualWork = clock::time_point(std::chrono::nanoseconds(monotonic_clock_ns));
-        mPerformanceHintSession.updateTargetWorkDuration(refreshIntervalNs);
-    }
-    PlatformEGL::beginFrame(monotonic_clock_ns, refreshIntervalNs, frameId);
-}
-
-void backend::PlatformEGLAndroid::preCommit() noexcept {
-    if (mPerformanceHintSession.isValid()) {
-        auto const actualWorkDuration = std::chrono::duration_cast<std::chrono::nanoseconds>(
-                clock::now() - mStartTimeOfActualWork);
-        mPerformanceHintSession.reportActualWorkDuration(actualWorkDuration.count());
-    }
-    PlatformEGL::preCommit();
-}
-
 Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
        const Platform::DriverConfig& driverConfig) noexcept {
-
-    // the refresh rate default value doesn't matter, we change it later
-    int32_t const tid = gettid();
-    mPerformanceHintSession = PerformanceHintManager::Session{
-            mPerformanceHintManager, &tid, 1, 16'666'667 };
-
    Driver* driver = PlatformEGL::createDriver(sharedContext, driverConfig);
    auto extensions = GLUtils::split(eglQueryString(mEGLDisplay, EGL_EXTENSIONS));

@@ -26,16 +26,21 @@

 #include <smolv.h>

+#include "generated/vkshaders/vkshaders.h"
+
 using namespace bluevk;
 using namespace utils;

 namespace filament::backend {

+using ImgUtil = VulkanImageUtility;
+
 namespace {

 inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect, VkFilter filter,
        VulkanAttachment src, VulkanAttachment dst,
        const VkOffset3D srcRect[2], const VkOffset3D dstRect[2]) {
+
    if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
        utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                      << " layout=" << src.getLayout()
@@ -43,8 +48,21 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
                      << " layout=" << dst.getLayout() << utils::io::endl;
    }

-    VkImageSubresourceRange const srcRange = src.getSubresourceRange();
-    VkImageSubresourceRange const dstRange = dst.getSubresourceRange();
+    const VkImageSubresourceRange srcRange = {
+            .aspectMask = aspect,
+            .baseMipLevel = src.level,
+            .levelCount = 1,
+            .baseArrayLayer = src.layer,
+            .layerCount = 1,
+    };
+
+    const VkImageSubresourceRange dstRange = {
+            .aspectMask = aspect,
+            .baseMipLevel = dst.level,
+            .levelCount = 1,
+            .baseArrayLayer = dst.layer,
+            .layerCount = 1,
+    };

    VulkanLayout oldSrcLayout = src.getLayout();
    VulkanLayout oldDstLayout = dst.getLayout();
@@ -59,15 +77,15 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
            .dstOffsets = { dstRect[0], dstRect[1] },
    }};
    vkCmdBlitImage(cmdbuffer,
-            src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
+            src.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC),
+            dst.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_DST),
            1, blitRegions, filter);

    if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
+        oldSrcLayout = ImgUtil::getDefaultLayout(src.texture->usage);
    }
    if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
+        oldDstLayout = ImgUtil::getDefaultLayout(dst.texture->usage);
    }
    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
@@ -75,6 +93,7 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,

 inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
        VulkanAttachment src, VulkanAttachment dst) {
+
    if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
        utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                      << " layout=" << src.getLayout()
@@ -82,8 +101,21 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
                      << " layout=" << dst.getLayout() << utils::io::endl;
    }

-    VkImageSubresourceRange const srcRange = src.getSubresourceRange();
-    VkImageSubresourceRange const dstRange = dst.getSubresourceRange();
+    const VkImageSubresourceRange srcRange = {
+            .aspectMask = aspect,
+            .baseMipLevel = src.level,
+            .levelCount = 1,
+            .baseArrayLayer = src.layer,
+            .layerCount = 1,
+    };
+
+    const VkImageSubresourceRange dstRange = {
+            .aspectMask = aspect,
+            .baseMipLevel = dst.level,
+            .levelCount = 1,
+            .baseArrayLayer = dst.layer,
+            .layerCount = 1,
+    };

    VulkanLayout oldSrcLayout = src.getLayout();
    VulkanLayout oldDstLayout = dst.getLayout();
@@ -101,15 +133,15 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
            .extent = { src.getExtent2D().width, src.getExtent2D().height, 1 },
    }};
    vkCmdResolveImage(cmdbuffer,
-            src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
+            src.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC),
+            dst.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_DST),
            1, resolveRegions);

    if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
+        oldSrcLayout = ImgUtil::getDefaultLayout(src.texture->usage);
    }
    if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
+        oldDstLayout = ImgUtil::getDefaultLayout(dst.texture->usage);
    }
    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
@@ -17,7 +17,6 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANBLITTER_H
 #define TNT_FILAMENT_BACKEND_VULKANBLITTER_H

-#include "VulkanCommands.h"
 #include "VulkanContext.h"

 #include <utils/compiler.h>
@@ -27,8 +27,8 @@ VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
        VkBufferUsageFlags usage, uint32_t numBytes)
    : mAllocator(allocator),
      mStagePool(stagePool),
-      mUsage(usage),
-      mUpdatedBytes(0) {
+      mUsage(usage) {
+
    // for now make sure that only 1 bit is set in usage
    // (because loadFromCpu() assumes that somewhat)
    assert_invariant(usage && !(usage & (usage - 1)));
@@ -49,7 +49,7 @@ VulkanBuffer::~VulkanBuffer() {
 }

 void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
-        uint32_t numBytes) {
+        uint32_t numBytes) const {
    assert_invariant(byteOffset == 0);
    VulkanStage const* stage = mStagePool.acquireStage(numBytes);
    void* mapped;
@@ -58,47 +58,15 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
    vmaUnmapMemory(mAllocator, stage->memory);
    vmaFlushAllocation(mAllocator, stage->memory, byteOffset, numBytes);

-    // If there was a previous update, then we need to make sure the following write is properly
-    // synced with the previous read.
-    if (mUpdatedBytes > 0) {
-        VkAccessFlags srcAccess = 0;
-        VkPipelineStageFlags srcStage = 0;
-        if (mUsage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT) {
-            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) {
-            srcAccess = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
-            srcStage = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
-        } else if (mUsage & VK_BUFFER_USAGE_INDEX_BUFFER_BIT) {
-            srcAccess = VK_ACCESS_INDEX_READ_BIT;
-            srcStage = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
-        }
-
-        VkBufferMemoryBarrier barrier{
-            .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-            .srcAccessMask = srcAccess,
-            .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .buffer = mGpuBuffer,
-            .size = mUpdatedBytes,
-        };
-        vkCmdPipelineBarrier(cmdbuf, srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 1,
-                &barrier, 0, nullptr);
-    }
-
    VkBufferCopy region{ .size = numBytes };
    vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);

-    mUpdatedBytes = numBytes;
-
    // Firstly, ensure that the copy finishes before the next draw call.
    // Secondly, in case the user decides to upload another chunk (without ever using the first one)
    // we need to ensure that this upload completes first (hence
    // dstStageMask=VK_PIPELINE_STAGE_TRANSFER_BIT).
    VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
    VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT;
-
    if (mUsage & VK_BUFFER_USAGE_VERTEX_BUFFER_BIT) {
        dstAccessMask |= VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
        dstStageMask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
@@ -107,24 +75,26 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
        dstStageMask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
    } else if (mUsage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT) {
        dstAccessMask |= VK_ACCESS_UNIFORM_READ_BIT;
-        dstStageMask |=
-                (VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
+        // NOTE: ideally dstStageMask would include VERTEX_SHADER_BIT | FRAGMENT_SHADER_BIT, but
+        // this seems to be insufficient on Mali devices. To work around this we are using a more
+        // aggressive ALL_GRAPHICS_BIT barrier.
+        dstStageMask |= VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
    } else if (mUsage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT) {
        // TODO: implement me
    }

    VkBufferMemoryBarrier barrier{
-        .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-        .dstAccessMask = dstAccessMask,
-        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .buffer = mGpuBuffer,
-        .size = VK_WHOLE_SIZE,
+	    .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+	    .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+	    .dstAccessMask = dstAccessMask,
+	    .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+	    .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+	    .buffer = mGpuBuffer,
+	    .size = VK_WHOLE_SIZE,
    };

    vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask, 0, 0, nullptr, 1,
-            &barrier, 0, nullptr);
+	    &barrier, 0, nullptr);
 }

 } // namespace filament::backend
@@ -19,7 +19,6 @@

 #include "VulkanContext.h"
 #include "VulkanStagePool.h"
-#include "VulkanMemory.h"

 namespace filament::backend {

@@ -30,7 +29,7 @@ public:
            uint32_t numBytes);
    ~VulkanBuffer();
    void loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
-            uint32_t numBytes);
+            uint32_t numBytes) const;
    VkBuffer getGpuBuffer() const {
        return mGpuBuffer;
    }
@@ -42,7 +41,6 @@ private:
    VmaAllocation mGpuMemory = VK_NULL_HANDLE;
    VkBuffer mGpuBuffer = VK_NULL_HANDLE;
    VkBufferUsageFlags mUsage = {};
-    uint32_t mUpdatedBytes = 0;
 };

 } // namespace filament::backend
@@ -47,8 +47,7 @@ VulkanCmdFence::VulkanCmdFence(VkFence ifence)

 VulkanCommandBuffer::VulkanCommandBuffer(VulkanResourceAllocator* allocator, VkDevice device,
        VkCommandPool pool)
-    : mResourceManager(allocator),
-      mPipeline(VK_NULL_HANDLE) {
+    : mResourceManager(allocator) {
    // Create the low-level command buffer.
    const VkCommandBufferAllocateInfo allocateInfo{
            .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
@@ -89,15 +89,6 @@ struct VulkanCommandBuffer {
    inline void reset() {
        fence.reset();
        mResourceManager.clear();
-        mPipeline = VK_NULL_HANDLE;
-    }
-
-    inline void setPipeline(VkPipeline pipeline) {
-        mPipeline = pipeline;
-    }
-
-    inline VkPipeline pipeline() const {
-        return mPipeline;
    }

    inline VkCommandBuffer buffer() const {
@@ -112,7 +103,6 @@ struct VulkanCommandBuffer {
 private:
    VulkanAcquireOnlyResourceManager mResourceManager;
    VkCommandBuffer mBuffer;
-    VkPipeline mPipeline;
 };

 // Allows classes to be notified after a new command buffer has been activated.
@@ -47,7 +47,7 @@
 // granualarity of a renderpass. You can enable this along with FVK_DEBUG_DEBUG_UTILS to take
 // advantage of vkCmdBegin/EndDebugUtilsLabelEXT. You can also just enable this with
 // FVK_DEBUG_PRINT_GROUP_MARKERS to print the current marker to stdout.
-#define FVK_DEBUG_GROUP_MARKERS           0x00000002
+#define FVK_DEBUG_GROUP_MARKERS 0x00000002

 #define FVK_DEBUG_TEXTURE                 0x00000004
 #define FVK_DEBUG_LAYOUT_TRANSITION       0x00000008
@@ -75,14 +75,18 @@
 #define FVK_DEBUG_PERFORMANCE     \
    FVK_DEBUG_SYSTRACE

-#if defined(FILAMENT_BACKEND_DEBUG_FLAG)
-#define FVK_DEBUG_FORWARDED_FLAG (FILAMENT_BACKEND_DEBUG_FLAG & FVK_DEBUG_EVERYTHING)
-#else
-#define FVK_DEBUG_FORWARDED_FLAG 0
-#endif
+#define FVK_DEBUG_CORRECTNESS     \
+    FVK_DEBUG_VALIDATION |        \
+    FVK_DEBUG_SHADER_MODULE |     \
+    FVK_DEBUG_TEXTURE |           \
+    FVK_DEBUG_LAYOUT_TRANSITION
+
+#define FVK_DEBUG_RENDER_PASSES   \
+    FVK_DEBUG_GROUP_MARKERS |     \
+    FVK_DEBUG_PRINT_GROUP_MARKERS

 #ifndef NDEBUG
-#define FVK_DEBUG_FLAGS (FVK_DEBUG_PERFORMANCE | FVK_DEBUG_FORWARDED_FLAG)
+#define FVK_DEBUG_FLAGS (FVK_DEBUG_PERFORMANCE)
 #else
 #define FVK_DEBUG_FLAGS 0
 #endif
@@ -108,7 +112,7 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 // end dependcy checks

 // Shorthand for combination of enabled debug flags
-#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS) && FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS) || FVK_ENABLED(FVK_DEBUG_TEXTURE)
 #define FVK_ENABLED_DEBUG_SAMPLER_NAME 1
 #else
 #define FVK_ENABLED_DEBUG_SAMPLER_NAME 0
@@ -57,19 +57,15 @@ VkExtent2D VulkanAttachment::getExtent2D() const {
    return { std::max(1u, texture->width >> level), std::max(1u, texture->height >> level) };
 }

-VkImageView VulkanAttachment::getImageView() {
+VkImageView VulkanAttachment::getImageView(VkImageAspectFlags aspect) {
    assert_invariant(texture);
-    return texture->getAttachmentView(getSubresourceRange());
+    return texture->getAttachmentView(getSubresourceRange(aspect));
 }

-bool VulkanAttachment::isDepth() const {
-    return texture->getImageAspect() & VK_IMAGE_ASPECT_DEPTH_BIT;
-}
-
-VkImageSubresourceRange VulkanAttachment::getSubresourceRange() const {
+VkImageSubresourceRange VulkanAttachment::getSubresourceRange(VkImageAspectFlags aspect) const {
    assert_invariant(texture);
    return {
-            .aspectMask = texture->getImageAspect(),
+            .aspectMask = aspect,
            .baseMipLevel = uint32_t(level),
            .levelCount = 1,
            .baseArrayLayer = uint32_t(layer),
@@ -19,6 +19,7 @@

 #include "VulkanConstants.h"
 #include "VulkanImageUtility.h"
+#include "VulkanPipelineCache.h"
 #include "VulkanUtility.h"

 #include <utils/bitset.h>
@@ -41,18 +42,16 @@ struct VulkanTimerQuery;
 struct VulkanCommandBuffer;

 struct VulkanAttachment {
-    VulkanTexture* texture = nullptr;
+    VulkanTexture* texture;
    uint8_t level = 0;
    uint16_t layer = 0;
-
-    bool isDepth() const;
    VkImage getImage() const;
    VkFormat getFormat() const;
    VulkanLayout getLayout() const;
    VkExtent2D getExtent2D() const;
-    VkImageView getImageView();
+    VkImageView getImageView(VkImageAspectFlags aspect);
    // TODO: maybe embed aspect into the attachment or texture itself.
-    VkImageSubresourceRange getSubresourceRange() const;
+    VkImageSubresourceRange getSubresourceRange(VkImageAspectFlags aspect) const;
 };

 class VulkanTimestamps {
@@ -103,12 +102,8 @@ public:
        return (uint32_t) VK_MAX_MEMORY_TYPES;
    }

-    inline VkFormatList const& getAttachmentDepthStencilFormats() const {
-        return mDepthStencilFormats;
-    }
-
-    inline VkFormatList const& getBlittableDepthStencilFormats() const {
-        return mBlittableDepthStencilFormats;
+    inline VkFormatList const& getAttachmentDepthFormats() const {
+        return mDepthFormats;
    }

    inline VkPhysicalDeviceLimits const& getPhysicalDeviceLimits() const noexcept {
@@ -137,8 +132,7 @@ private:
    bool mDebugMarkersSupported = false;
    bool mDebugUtilsSupported = false;

-    VkFormatList mDepthStencilFormats;
-    VkFormatList mBlittableDepthStencilFormats;
+    VkFormatList mDepthFormats;

    // For convenience so that VulkanPlatform can initialize the private fields.
    friend class VulkanPlatform;
@@ -14,7 +14,7 @@
 * limitations under the License.
 */

-#include "VulkanDriver.h"
+#include "vulkan/VulkanDriver.h"

 #include "CommandStreamDispatcher.h"
 #include "DataReshaper.h"
@@ -22,6 +22,7 @@
 #include "VulkanCommands.h"
 #include "VulkanDriverFactory.h"
 #include "VulkanHandles.h"
+#include "VulkanImageUtility.h"
 #include "VulkanMemory.h"
 #include "VulkanTexture.h"

@@ -101,15 +102,6 @@ VulkanTexture* createEmptyTexture(VkDevice device, VkPhysicalDevice physicalDevi
    return emptyTexture;
 }

-VulkanBufferObject* createEmptyBufferObject(VmaAllocator allocator, VulkanStagePool& stagePool,
-        VulkanCommands* commands) {
-    VulkanBufferObject* obj =
-            new VulkanBufferObject(allocator, stagePool, 1, BufferObjectBinding::UNIFORM);
-    uint8_t byte = 0;
-    obj->buffer.loadFromCpu(commands->get().buffer(), &byte, 0, 1);
-    return obj;
-}
-
 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
 VKAPI_ATTR VkBool32 VKAPI_CALL debugReportCallback(VkDebugReportFlagsEXT flags,
        VkDebugReportObjectTypeEXT objectType, uint64_t object, size_t location,
@@ -205,6 +197,8 @@ void DebugUtils::setName(VkObjectType type, uint64_t handle, char const* name) {
 }
 #endif // FVK_EANBLED(FVK_DEBUG_DEBUG_UTILS)

+using ImgUtil = VulkanImageUtility;
+
 Dispatcher VulkanDriver::getDispatcher() const noexcept {
    return ConcreteDispatcher<VulkanDriver>::make();
 }
@@ -220,14 +214,12 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
      mThreadSafeResourceManager(&mResourceAllocator),
      mCommands(mPlatform->getDevice(), mPlatform->getGraphicsQueue(),
              mPlatform->getGraphicsQueueFamilyIndex(), &mContext, &mResourceAllocator),
-      mPipelineLayoutCache(mPlatform->getDevice(), &mResourceAllocator),
-      mPipelineCache(mPlatform->getDevice(), mAllocator),
+      mPipelineCache(&mResourceAllocator),
      mStagePool(mAllocator, &mCommands),
      mFramebufferCache(mPlatform->getDevice()),
      mSamplerCache(mPlatform->getDevice()),
      mBlitter(mPlatform->getPhysicalDevice(), &mCommands),
      mReadPixels(mPlatform->getDevice()),
-      mDescriptorSetManager(mPlatform->getDevice(), &mResourceAllocator),
      mIsSRGBSwapChainSupported(mPlatform->getCustomization().isSRGBSwapChainSupported) {

 #if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
@@ -251,17 +243,13 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
 #endif

    mTimestamps = std::make_unique<VulkanTimestamps>(mPlatform->getDevice());
+    mCommands.setObserver(&mPipelineCache);
+    mPipelineCache.setDevice(mPlatform->getDevice(), mAllocator);

    mEmptyTexture = createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
            mContext, mAllocator, &mCommands, mStagePool);
-    mEmptyBufferObject = createEmptyBufferObject(mAllocator, mStagePool, &mCommands);

-    mDescriptorSetManager.setPlaceHolders(mSamplerCache.getSampler({}), mEmptyTexture,
-            mEmptyBufferObject);
-
-    mGetPipelineFunction = [this](VulkanDescriptorSetLayoutList const& layouts) {
-        return mPipelineLayoutCache.getLayout(layouts);
-    };
+    mPipelineCache.setDummyTexture(mEmptyTexture->getPrimaryImageView());
 }

 VulkanDriver::~VulkanDriver() noexcept = default;
@@ -322,14 +310,13 @@ ShaderModel VulkanDriver::getShaderModel() const noexcept {
 }

 void VulkanDriver::terminate() {
-    delete mEmptyBufferObject;
-    delete mEmptyTexture;
-
    // Command buffers should come first since it might have commands depending on resources that
    // are about to be destroyed.
    mCommands.terminate();

+    delete mEmptyTexture;
    mResourceManager.clear();
+
    mTimestamps.reset();

    mBlitter.terminate();
@@ -342,12 +329,6 @@ void VulkanDriver::terminate() {
    mPipelineCache.terminate();
    mFramebufferCache.reset();
    mSamplerCache.terminate();
-    mDescriptorSetManager.terminate();
-    mPipelineLayoutCache.terminate();
-
-#if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
-    mResourceAllocator.print();
-#endif

    vmaDestroyAllocator(mAllocator);

@@ -379,8 +360,6 @@ void VulkanDriver::collectGarbage() {
    mCommands.gc();
    mStagePool.gc();
    mFramebufferCache.gc();
-    mPipelineCache.gc();
-    mDescriptorSetManager.gc();

 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
    mResourceAllocator.print();
@@ -388,13 +367,12 @@ void VulkanDriver::collectGarbage() {

    FVK_SYSTRACE_END();
 }
-void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
-        int64_t refreshIntervalNs, uint32_t frameId) {
+void VulkanDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
    // Do nothing.
 }

 void VulkanDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, FrameScheduledCallback&& callback) {
+        FrameScheduledCallback callback, void* user) {
 }

 void VulkanDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
@@ -511,6 +489,9 @@ void VulkanDriver::destroyBufferObject(Handle<HwBufferObject> boh) {
        return;
    }
    auto bufferObject = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
+    if (bufferObject->bindingType == BufferObjectBinding::UNIFORM) {
+        mPipelineCache.unbindUniformBuffer(bufferObject->buffer.getGpuBuffer());
+    }
    mResourceManager.release(bufferObject);
 }

@@ -561,7 +542,6 @@ void VulkanDriver::destroyProgram(Handle<HwProgram> ph) {
        return;
    }
    auto vkprogram = mResourceAllocator.handle_cast<VulkanProgram*>(ph);
-    mDescriptorSetManager.clearProgram(vkprogram);
    mResourceManager.release(vkprogram);
 }

@@ -918,11 +898,6 @@ bool VulkanDriver::isDepthStencilResolveSupported() {
    return false;
 }

-bool VulkanDriver::isDepthStencilBlitSupported(TextureFormat format) {
-    auto const& formats = mContext.getBlittableDepthStencilFormats();
-    return std::find(formats.begin(), formats.end(), getVkFormat(format)) != formats.end();
-}
-
 bool VulkanDriver::isProtectedTexturesSupported() {
    return false;
 }
@@ -1236,7 +1211,12 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
        }
    }

-    VulkanLayout currentDepthLayout = depth.getLayout();
+    VulkanLayout const currentDepthLayout = depth.getLayout();
+    VulkanLayout const renderPassDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
+    // We need to keep the final layout as an attachment because the implicit transition does not
+    // have any barrier guarrantees, meaning that if we want to sample from the output in the next
+    // pass, then we'd have a race-condition/validation error.
+    VulkanLayout const finalDepthLayout = renderPassDepthLayout;

    TargetBufferFlags clearVal = params.flags.clear;
    TargetBufferFlags discardEndVal = params.flags.discardEnd;
@@ -1245,20 +1225,16 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
            discardEndVal &= ~TargetBufferFlags::DEPTH;
            clearVal &= ~TargetBufferFlags::DEPTH;
        }
-        // If the depth attachment texture was previously sampled, then we need to manually
-        // transition it to an attachment. This is necessary to also set up a barrier between the
-        // previous read and the potentially coming write.
-        if (currentDepthLayout == VulkanLayout::DEPTH_SAMPLER) {
-            depth.texture->transitionLayout(cmdbuffer, depth.getSubresourceRange(),
-                    VulkanLayout::DEPTH_ATTACHMENT);
-            currentDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
-        }
+        auto const attachmentSubresourceRange = depth.getSubresourceRange(VK_IMAGE_ASPECT_DEPTH_BIT);
+        depth.texture->setLayout(attachmentSubresourceRange, VulkanLayout::DEPTH_ATTACHMENT);
    }

    // Create the VkRenderPass or fetch it from cache.
    VulkanFboCache::RenderPassKey rpkey = {
        .initialColorLayoutMask = 0,
        .initialDepthLayout = currentDepthLayout,
+        .renderPassDepthLayout = renderPassDepthLayout,
+        .finalDepthLayout = finalDepthLayout,
        .depthFormat = depth.getFormat(),
        .clear = clearVal,
        .discardStart = discardStart,
@@ -1275,9 +1251,9 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
                rpkey.needsResolveMask |= (1 << i);
            }
            if (info.texture->getPrimaryImageLayout() != VulkanLayout::COLOR_ATTACHMENT) {
-                ((VulkanTexture*) info.texture)
-                        ->transitionLayout(cmdbuffer, info.getSubresourceRange(),
-                                VulkanLayout::COLOR_ATTACHMENT);
+                ((VulkanTexture*) info.texture)->transitionLayout(cmdbuffer,
+                        info.getSubresourceRange(VK_IMAGE_ASPECT_COLOR_BIT),
+                        VulkanLayout::COLOR_ATTACHMENT);
            }
        } else {
            rpkey.colorFormat[i] = VK_FORMAT_UNDEFINED;
@@ -1295,42 +1271,27 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
        .layers = 1,
        .samples = rpkey.samples,
    };
-    auto& renderPassAttachments = mRenderPassFboInfo.attachments;
    for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
        if (!rt->getColor(i).texture) {
            fbkey.color[i] = VK_NULL_HANDLE;
            fbkey.resolve[i] = VK_NULL_HANDLE;
        } else if (fbkey.samples == 1) {
-            auto& colorAttachment = rt->getColor(i);
-            renderPassAttachments.insert(colorAttachment);
-            fbkey.color[i] = colorAttachment.getImageView();
+            fbkey.color[i] = rt->getColor(i).getImageView(VK_IMAGE_ASPECT_COLOR_BIT);
            fbkey.resolve[i] = VK_NULL_HANDLE;
            assert_invariant(fbkey.color[i]);
        } else {
-            auto& msaaColorAttachment = rt->getMsaaColor(i);
-            renderPassAttachments.insert(msaaColorAttachment);
-
-            auto& colorAttachment = rt->getColor(i);
-            fbkey.color[i] = msaaColorAttachment.getImageView();
-
-            VulkanTexture* texture = colorAttachment.texture;
+            fbkey.color[i] = rt->getMsaaColor(i).getImageView(VK_IMAGE_ASPECT_COLOR_BIT);
+            VulkanTexture* texture = (VulkanTexture*) rt->getColor(i).texture;
            if (texture->samples == 1) {
-                mRenderPassFboInfo.hasColorResolve = true;
-
-                renderPassAttachments.insert(colorAttachment);
-                fbkey.resolve[i] = colorAttachment.getImageView();
+                fbkey.resolve[i] = rt->getColor(i).getImageView(VK_IMAGE_ASPECT_COLOR_BIT);
                assert_invariant(fbkey.resolve[i]);
            }
            assert_invariant(fbkey.color[i]);
        }
    }
    if (depth.texture) {
-        fbkey.depth = depth.getImageView();
+        fbkey.depth = depth.getImageView(VK_IMAGE_ASPECT_DEPTH_BIT);
        assert_invariant(fbkey.depth);
-        renderPassAttachments.insert(depth);
-
-        UTILS_UNUSED_IN_RELEASE bool const depthDiscardEnd =
-                any(rpkey.discardEnd & TargetBufferFlags::DEPTH);

        // Vulkan 1.1 does not support multisampled depth resolve, so let's check here
        // and assert if this is requested. (c.f. isAutoDepthResolveSupported)
@@ -1339,7 +1300,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
        // - If the RT is MS then all SS attachments are auto resolved if not discarded.
        assert_invariant(!(rt->getSamples() > 1 &&
                rt->getDepth().texture->samples == 1 &&
-                !depthDiscardEnd));
+                !any(rpkey.discardEnd & TargetBufferFlags::DEPTH)));
    }
    VkFramebuffer vkfb = mFramebufferCache.getFramebuffer(fbkey);

@@ -1352,10 +1313,16 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
    }
 #endif

-    // The current command buffer now has references to the render target and its attachments.
+    // The current command buffer now owns a reference to the render target and its attachments.
+    // Note that we must acquire parent textures, not sidecars.
    commands.acquire(rt);
-    for (auto const& attachment: renderPassAttachments) {
-        commands.acquire(attachment.texture);
+    if (depth.texture) {
+        commands.acquire((VulkanTexture*) depth.texture);
+    }
+    for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
+        if (rt->getColor(i).texture) {
+            commands.acquire(rt->getColor(i).texture);
+        }
    }

    // Populate the structures required for vkCmdBeginRenderPass.
@@ -1441,52 +1408,33 @@ void VulkanDriver::endRenderPass(int) {
    // issue several of them when considering MRT. This would be very complex to set up and would
    // require more state tracking, so we've chosen to use a memory barrier for simplicity and
    // correctness.
+
+    // NOTE: ideally dstStageMask would merely be VERTEX_SHADER_BIT | FRAGMENT_SHADER_BIT, but this
+    // seems to be insufficient on Mali devices. To work around this we are adding a more aggressive
+    // TOP_OF_PIPE barrier.
    if (!rt->isSwapChain()) {
-        for (auto const& attachment: mRenderPassFboInfo.attachments) {
-            bool const isDepth = attachment.isDepth();
-            VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
-
-            // This is a workaround around a validation issue (might not be an actual driver issue).
-            if (mRenderPassFboInfo.hasColorResolve && !isDepth) {
-                srcStageMask = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
-            }
-
-            VkPipelineStageFlags dstStageMask =
-                    VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
-            VkAccessFlags srcAccess = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-            VkAccessFlags dstAccess = VK_ACCESS_SHADER_READ_BIT;
-            VulkanLayout layout = VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT;
-            if (isDepth) {
-                srcAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
-                dstAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
-                srcStageMask = VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
-                dstStageMask = VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
-                layout =  VulkanFboCache::FINAL_DEPTH_ATTACHMENT_LAYOUT;
-            }
-
-            auto const vkLayout = imgutil::getVkLayout(layout);
-            auto const& range = attachment.getSubresourceRange();
-            VkImageMemoryBarrier barrier = {
-                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = srcAccess,
-                .dstAccessMask = dstAccess,
-                .oldLayout = vkLayout,
-                .newLayout = vkLayout,
-                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-                .image = attachment.getImage(),
-                .subresourceRange = range,
-            };
-
-            attachment.texture->setLayout(range, layout);
-            vkCmdPipelineBarrier(cmdbuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr,
-                    1, &barrier);
+        VkMemoryBarrier barrier{
+                .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+                .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
+        };
+        VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+        if (rt->hasDepth()) {
+            barrier.srcAccessMask |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+            srcStageMask |= VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
        }
+        vkCmdPipelineBarrier(cmdbuffer, srcStageMask,
+                VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT | // <== For Mali
+                VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
+                0, 1, &barrier, 0, nullptr, 0, nullptr);
    }

-    mRenderPassFboInfo.clear();
-    mDescriptorSetManager.clearState();
+    if (mCurrentRenderPass.currentSubpass > 0) {
+        for (uint32_t i = 0; i < VulkanPipelineCache::INPUT_ATTACHMENT_COUNT; i++) {
+            mPipelineCache.bindInputAttachment(i, {});
+        }
+        mCurrentRenderPass.currentSubpass = 0;
+    }
    mCurrentRenderPass.renderTarget = nullptr;
    mCurrentRenderPass.renderPass = VK_NULL_HANDLE;
    FVK_SYSTRACE_END();
@@ -1505,9 +1453,15 @@ void VulkanDriver::nextSubpass(int) {
    mPipelineCache.bindRenderPass(mCurrentRenderPass.renderPass,
            ++mCurrentRenderPass.currentSubpass);

-    if (mCurrentRenderPass.params.subpassMask & 0x1) {
-        VulkanAttachment subpassInput = renderTarget->getColor(0);
-        mDescriptorSetManager.updateInputAttachment({}, subpassInput);
+    for (uint32_t i = 0; i < VulkanPipelineCache::INPUT_ATTACHMENT_COUNT; i++) {
+        if ((1 << i) & mCurrentRenderPass.params.subpassMask) {
+            VulkanAttachment subpassInput = renderTarget->getColor(i);
+            VkDescriptorImageInfo info = {
+                .imageView = subpassInput.getImageView(VK_IMAGE_ASPECT_COLOR_BIT),
+                .imageLayout = ImgUtil::getVkLayout(subpassInput.getLayout()),
+            };
+            mPipelineCache.bindInputAttachment(i, info);
+        }
    }
 }

@@ -1547,24 +1501,25 @@ void VulkanDriver::commit(Handle<HwSwapChain> sch) {

 void VulkanDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> boh) {
    auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    VkDeviceSize const offset = 0;
-    VkDeviceSize const size = VK_WHOLE_SIZE;
-    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
+    const VkDeviceSize offset = 0;
+    const VkDeviceSize size = VK_WHOLE_SIZE;
+    mPipelineCache.bindUniformBufferObject((uint32_t) index, bo, offset, size);
 }

 void VulkanDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t index,
        Handle<HwBufferObject> boh, uint32_t offset, uint32_t size) {

-    assert_invariant(bindingType == BufferObjectBinding::UNIFORM);
+    assert_invariant(bindingType == BufferObjectBinding::SHADER_STORAGE ||
+                     bindingType == BufferObjectBinding::UNIFORM);

    // TODO: implement BufferObjectBinding::SHADER_STORAGE case

    auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
+    mPipelineCache.bindUniformBufferObject((uint32_t)index, bo, offset, size);
 }

 void VulkanDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
-    mDescriptorSetManager.clearBuffer((uint32_t) index);
+    // TODO: implement unbindBuffer()
 }

 void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
@@ -1812,14 +1767,22 @@ void VulkanDriver::bindPipeline(PipelineState pipelineState) {
    // where "SamplerBinding" is the integer in the GLSL, and SamplerGroupBinding is the abstract
    // Filament concept used to form groups of samplers.

+    VkDescriptorImageInfo samplerInfo[VulkanPipelineCache::SAMPLER_BINDING_COUNT] = {};
+    VulkanTexture* samplerTextures[VulkanPipelineCache::SAMPLER_BINDING_COUNT] = {nullptr};
+
    auto const& bindingToSamplerIndex = program->getBindingToSamplerIndex();
+    VulkanPipelineCache::UsageFlags usage = program->getUsage();
+
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
    auto const& bindingToName = program->getBindingToName();
 #endif

-    for (auto binding: program->getBindings()) {
+    UTILS_NOUNROLL
+    for (uint8_t binding = 0; binding < VulkanPipelineCache::SAMPLER_BINDING_COUNT; binding++) {
        uint16_t const indexPair = bindingToSamplerIndex[binding];
+
        if (indexPair == 0xffff) {
+            usage = VulkanPipelineCache::disableUsageFlags(binding, usage);
            continue;
        }

@@ -1828,20 +1791,24 @@ void VulkanDriver::bindPipeline(PipelineState pipelineState) {

        VulkanSamplerGroup* vksb = mSamplerBindings[samplerGroupInd];
        if (!vksb) {
+            usage = VulkanPipelineCache::disableUsageFlags(binding, usage);
            continue;
        }
        SamplerDescriptor const* boundSampler = ((SamplerDescriptor*) vksb->sb->data()) + samplerInd;

        if (UTILS_UNLIKELY(!boundSampler->t)) {
+            usage = VulkanPipelineCache::disableUsageFlags(binding, usage);
            continue;
        }
+
        VulkanTexture* texture = mResourceAllocator.handle_cast<VulkanTexture*>(boundSampler->t);
+        VkImageViewType const expectedType = texture->getViewType();

        // TODO: can this uninitialized check be checked in a higher layer?
        // This fallback path is very flaky because the dummy texture might not have
        // matching characteristics. (e.g. if the missing texture is a 3D texture)
        if (UTILS_UNLIKELY(texture->getPrimaryImageLayout() == VulkanLayout::UNDEFINED)) {
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE) && FVK_ENABLED_DEBUG_SAMPLER_NAME
+#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
            utils::slog.w << "Uninitialized texture bound to '" << bindingToName[binding] << "'";
            utils::slog.w << " in material '" << program->name.c_str() << "'";
            utils::slog.w << " at binding point " << +binding << utils::io::endl;
@@ -1849,16 +1816,43 @@ void VulkanDriver::bindPipeline(PipelineState pipelineState) {
            texture = mEmptyTexture;
        }

-        VkSampler const vksampler = mSamplerCache.getSampler(boundSampler->s);
+        SamplerParams const& samplerParams = boundSampler->s;
+        VkSampler const vksampler = mSamplerCache.getSampler(samplerParams);
+
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
        VulkanDriver::DebugUtils::setName(VK_OBJECT_TYPE_SAMPLER,
                reinterpret_cast<uint64_t>(vksampler), bindingToName[binding].c_str());
 #endif
-        mDescriptorSetManager.updateSampler({}, binding, texture, vksampler);
+
+        VkImageView imageView = VK_NULL_HANDLE;
+        VkImageSubresourceRange const range = texture->getPrimaryViewRange();
+        if (any(texture->usage & TextureUsage::DEPTH_ATTACHMENT) &&
+                expectedType == VK_IMAGE_VIEW_TYPE_2D) {
+            // If the sampler is part of a mipmapped depth texture, where one of the level *can* be
+            // an attachment, then the sampler for this texture has the same view properties as a
+            // view for an attachment. Therefore, we can use getAttachmentView to get a
+            // corresponding VkImageView.
+            imageView = texture->getAttachmentView(range);
+        } else {
+            imageView = texture->getViewForType(range, expectedType);
+        }
+
+        samplerInfo[binding] = {
+            .sampler = vksampler,
+            .imageView = imageView,
+            .imageLayout = ImgUtil::getVkLayout(texture->getPrimaryImageLayout())
+        };
+        samplerTextures[binding] = texture;
    }

-    mPipelineCache.bindLayout(mDescriptorSetManager.bind(commands, program, mGetPipelineFunction));
-    mPipelineCache.bindPipeline(commands);
+    mPipelineCache.bindSamplers(samplerInfo, samplerTextures, usage);
+
+    // Bind a new pipeline if the pipeline state changed.
+    // If allocation failed, skip the draw call and bail. We do not emit an error since the
+    // validation layer will already do so.
+    if (!mPipelineCache.bindPipeline(commands)) {
+        return;
+    }
    FVK_SYSTRACE_END();
 }

@@ -1898,8 +1892,12 @@ void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
    VulkanCommandBuffer& commands = mCommands.get();
    VkCommandBuffer cmdbuffer = commands.buffer();

-    // Bind "dynamic" UBOs if they need to change.
-    mDescriptorSetManager.dynamicBind(&commands, {});
+    // Bind new descriptor sets if they need to change.
+    // If descriptor set allocation failed, skip the draw call and bail. No need to emit an error
+    // message since the validation layers already do so.
+    if (!mPipelineCache.bindDescriptors(cmdbuffer)) {
+        return;
+    }

    // Finally, make the actual draw call. TODO: support subranges
    const uint32_t firstIndex = indexOffset;
@@ -28,8 +28,6 @@
 #include "VulkanSamplerCache.h"
 #include "VulkanStagePool.h"
 #include "VulkanUtility.h"
-#include "caching/VulkanDescriptorSetManager.h"
-#include "caching/VulkanPipelineLayoutCache.h"

 #include "DriverBase.h"
 #include "private/backend/Driver.h"
@@ -42,25 +40,6 @@ namespace filament::backend {
 class VulkanPlatform;
 struct VulkanSamplerGroup;

-// The maximum number of attachments for any renderpass (color + resolve + depth)
-constexpr uint8_t MAX_RENDERTARGET_ATTACHMENT_TEXTURES =
-        MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT * 2 + 1;
-
-// We need to store information about a render pass to enable better barriers at the end of a
-// renderpass.
-struct RenderPassFboBundle {
-    using AttachmentArray =
-            CappedArray<VulkanAttachment, MAX_RENDERTARGET_ATTACHMENT_TEXTURES>;
-
-    AttachmentArray attachments;
-    bool hasColorResolve = false;
-
-    void clear() {
-        attachments.clear();
-        hasColorResolve = false;
-    }
-};
-
 class VulkanDriver final : public DriverBase {
 public:
    static Driver* create(VulkanPlatform* platform, VulkanContext const& context,
@@ -91,8 +70,6 @@ public:
 #endif // FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)

 private:
-    static constexpr uint8_t MAX_SAMPLER_BINDING_COUNT = Program::SAMPLER_BINDING_COUNT;
-
    void debugCommandBegin(CommandStream* cmds, bool synchronous,
            const char* methodName) noexcept override;

@@ -129,9 +106,7 @@ private:
    VulkanPlatform* mPlatform = nullptr;
    std::unique_ptr<VulkanTimestamps> mTimestamps;

-    // Placeholder resources
    VulkanTexture* mEmptyTexture;
-    VulkanBufferObject* mEmptyBufferObject;

    VulkanSwapChain* mCurrentSwapChain = nullptr;
    VulkanRenderTarget* mDefaultRenderTarget = nullptr;
@@ -148,19 +123,13 @@ private:
    VulkanThreadSafeResourceManager mThreadSafeResourceManager;

    VulkanCommands mCommands;
-    VulkanPipelineLayoutCache mPipelineLayoutCache;
    VulkanPipelineCache mPipelineCache;
    VulkanStagePool mStagePool;
    VulkanFboCache mFramebufferCache;
    VulkanSamplerCache mSamplerCache;
    VulkanBlitter mBlitter;
-    VulkanSamplerGroup* mSamplerBindings[MAX_SAMPLER_BINDING_COUNT] = {};
+    VulkanSamplerGroup* mSamplerBindings[VulkanPipelineCache::SAMPLER_BINDING_COUNT] = {};
    VulkanReadPixels mReadPixels;
-    VulkanDescriptorSetManager mDescriptorSetManager;
-
-    VulkanDescriptorSetManager::GetPipelineLayoutFunction mGetPipelineFunction;
-
-    RenderPassFboBundle mRenderPassFboInfo;

    bool const mIsSRGBSwapChainSupported;
 };
@@ -29,10 +29,14 @@ using namespace bluevk;

 namespace filament::backend {

+using ImgUtil = VulkanImageUtility;
+
 bool VulkanFboCache::RenderPassEq::operator()(const RenderPassKey& k1,
        const RenderPassKey& k2) const {
    if (k1.initialColorLayoutMask != k2.initialColorLayoutMask) return false;
    if (k1.initialDepthLayout != k2.initialDepthLayout) return false;
+    if (k1.renderPassDepthLayout != k2.renderPassDepthLayout) return false;
+    if (k1.finalDepthLayout != k2.finalDepthLayout) return false;
    for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
        if (k1.colorFormat[i] != k2.colorFormat[i]) return false;
    }
@@ -193,7 +197,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
        if (config.colorFormat[i] == VK_FORMAT_UNDEFINED) {
            continue;
        }
-        const VkImageLayout subpassLayout = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT);
+        const VkImageLayout subpassLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT);
        uint32_t index;

        if (!hasSubpasses) {
@@ -239,9 +243,9 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
            .stencilLoadOp = kDontCare,
            .stencilStoreOp = kDisableStore,
            .initialLayout = ((!discard && config.initialColorLayoutMask & (1 << i)) || clear)
-                                     ? imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT)
-                                     : imgutil::getVkLayout(VulkanLayout::UNDEFINED),
-            .finalLayout = imgutil::getVkLayout(FINAL_COLOR_ATTACHMENT_LAYOUT),
+                                     ? ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT)
+                                     : ImgUtil::getVkLayout(VulkanLayout::UNDEFINED),
+            .finalLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
        };
    }

@@ -268,7 +272,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {

        pResolveAttachment->attachment = attachmentIndex;
        pResolveAttachment->layout
-                = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT_RESOLVE);
+                = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT_RESOLVE);
        ++pResolveAttachment;

        attachments[attachmentIndex++] = {
@@ -278,8 +282,8 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
            .storeOp = kEnableStore,
            .stencilLoadOp = kDontCare,
            .stencilStoreOp = kDisableStore,
-            .initialLayout = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
-            .finalLayout = imgutil::getVkLayout(FINAL_COLOR_ATTACHMENT_LAYOUT),
+            .initialLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
+            .finalLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
        };
    }

@@ -288,7 +292,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
        const bool clear = any(config.clear & TargetBufferFlags::DEPTH);
        const bool discardStart = any(config.discardStart & TargetBufferFlags::DEPTH);
        const bool discardEnd = any(config.discardEnd & TargetBufferFlags::DEPTH);
-        depthAttachmentRef.layout = imgutil::getVkLayout(VulkanLayout::DEPTH_ATTACHMENT);
+        depthAttachmentRef.layout = ImgUtil::getVkLayout(config.renderPassDepthLayout);
        depthAttachmentRef.attachment = attachmentIndex;
        attachments[attachmentIndex++] = {
            .format = config.depthFormat,
@@ -297,8 +301,8 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
            .storeOp = discardEnd ? kDisableStore : kEnableStore,
            .stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE,
            .stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE,
-            .initialLayout = imgutil::getVkLayout(config.initialDepthLayout),
-            .finalLayout = imgutil::getVkLayout(FINAL_DEPTH_ATTACHMENT_LAYOUT),
+            .initialLayout = ImgUtil::getVkLayout(config.initialDepthLayout),
+            .finalLayout = ImgUtil::getVkLayout(config.finalDepthLayout),
        };
    }
    renderPassInfo.attachmentCount = attachmentIndex;
@@ -35,25 +35,24 @@ namespace filament::backend {
 //
 class VulkanFboCache {
 public:
-    constexpr static VulkanLayout FINAL_COLOR_ATTACHMENT_LAYOUT = VulkanLayout::COLOR_ATTACHMENT;
-    constexpr static VulkanLayout FINAL_RESOLVE_ATTACHMENT_LAYOUT = VulkanLayout::COLOR_ATTACHMENT;
-    constexpr static VulkanLayout FINAL_DEPTH_ATTACHMENT_LAYOUT = VulkanLayout::DEPTH_ATTACHMENT;
-
    // RenderPassKey is a small POD representing the immutable state that is used to construct
    // a VkRenderPass. It is hashed and used as a lookup key.
+    // TODO: This struct can be reduced in size by using a subset of formats instead of VkFormat
+    //       and removing the "finalDepthLayout" field.
    struct alignas(8) RenderPassKey {
        // For each target, we need to know three image layouts: the layout BEFORE the pass, the
        // layout DURING the pass, and the layout AFTER the pass. Here are the rules:
        // - For depth, we explicitly specify all three layouts.
        // - Color targets have their initial image layout specified with a bitmask.
        // - For each color target, the pre-existing layout is either UNDEFINED (0) or GENERAL (1).
-        // - The render pass and final images layout for color buffers is always
-        //   VulkanLayout::COLOR_ATTACHMENT.
+        // - The render pass and final images layout for color buffers is always GENERAL.
        uint8_t initialColorLayoutMask;

        // Note that if VulkanLayout grows beyond 16, we'd need to up this.
-        VulkanLayout initialDepthLayout : 8;
-        uint8_t padding0;
+        VulkanLayout initialDepthLayout : 4;
+        VulkanLayout renderPassDepthLayout : 4;
+        VulkanLayout finalDepthLayout : 4;
+        uint8_t padding0 : 4;
        uint8_t padding1;

        VkFormat colorFormat[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT]; // 32 bytes
@@ -64,7 +63,7 @@ public:
        uint8_t samples; // 1 byte
        uint8_t needsResolveMask; // 1 byte
        uint8_t subpassMask; // 1 byte
-        uint8_t padding2; // 1 byte
+        bool padding2; // 1 byte
    };
    struct RenderPassVal {
        VkRenderPass handle;
@@ -20,9 +20,7 @@
 #include "VulkanConstants.h"
 #include "VulkanDriver.h"
 #include "VulkanMemory.h"
-#include "VulkanUtility.h"
 #include "spirv/VulkanSpirvUtils.h"
-#include "utils/Log.h"

 #include <backend/platforms/VulkanPlatform.h>

@@ -32,17 +30,15 @@ using namespace bluevk;

 namespace filament::backend {

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

-void flipVertically(VkViewport* rect, uint32_t framebufferHeight) {
+static void flipVertically(VkViewport* rect, uint32_t framebufferHeight) {
    rect->y = framebufferHeight - rect->y - rect->height;
 }

-void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
+static void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
    int32_t x = std::max(rect->offset.x, 0);
    int32_t y = std::max(rect->offset.y, 0);
    int32_t right = std::min(rect->offset.x + (int32_t) rect->extent.width, (int32_t) fbWidth);
@@ -53,110 +49,17 @@ void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
    rect->extent.height = std::max(top - y, 0);
 }

-template<typename Bitmask>
-static constexpr Bitmask fromStageFlags(ShaderStageFlags2 flags, uint8_t binding) {
-    Bitmask ret = 0;
-    if (flags & ShaderStageFlags2::VERTEX) {
-        ret |= (getVertexStage<Bitmask>() << binding);
-    }
-    if (flags & ShaderStageFlags2::FRAGMENT) {
-        ret |= (getFragmentStage<Bitmask>() << binding);
-    }
-    return ret;
-}
-
-constexpr decltype(VulkanProgram::MAX_SHADER_MODULES) MAX_SHADER_MODULES =
-        VulkanProgram::MAX_SHADER_MODULES;
-
-using LayoutDescriptionList = VulkanProgram::LayoutDescriptionList;
-
-template<typename Bitmask>
-void addDescriptors(Bitmask mask,
-        utils::FixedCapacityVector<DescriptorSetLayoutBinding>& outputList) {
-    constexpr uint8_t MODULE_OFFSET = (sizeof(Bitmask) * 8) / MAX_SHADER_MODULES;
-    for (uint8_t i = 0; i < MODULE_OFFSET; ++i) {
-        bool const hasVertex = (mask & (1ULL << i)) != 0;
-        bool const hasFragment = (mask & (1ULL << (MODULE_OFFSET + i))) != 0;
-        if (!hasVertex && !hasFragment) {
-            continue;
-        }
-
-        DescriptorSetLayoutBinding binding{
-            .binding = i,
-            .flags = DescriptorFlags::NONE,
-            .count = 0,// This is always 0 for now as we pass the size of the UBOs in the Driver API
-                       // instead.
-        };
-        if (hasVertex) {
-            binding.stageFlags = ShaderStageFlags2::VERTEX;
-        }
-        if (hasFragment) {
-            binding.stageFlags = static_cast<ShaderStageFlags2>(
-                    binding.stageFlags | ShaderStageFlags2::FRAGMENT);
-        }
-        if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
-            binding.type = DescriptorType::UNIFORM_BUFFER;
-        } else if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
-            binding.type = DescriptorType::SAMPLER;
-        } else if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
-            binding.type = DescriptorType::INPUT_ATTACHMENT;
-        }
-        outputList.push_back(binding);
-    }
-}
-
-inline VkDescriptorSetLayout createDescriptorSetLayout(VkDevice device,
-        VkDescriptorSetLayoutCreateInfo const& info) {
-    VkDescriptorSetLayout layout;
-    vkCreateDescriptorSetLayout(device, &info, VKALLOC, &layout);
-    return layout;
-}
-
-} // anonymous namespace
-
-
-VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(VkDevice device, VkDescriptorSetLayoutCreateInfo const& info,
-        Bitmask const& bitmask)
-    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET_LAYOUT),
-      mDevice(device),
-      vklayout(createDescriptorSetLayout(device, info)),
-      bitmask(bitmask),
-      bindings(getBindings(bitmask)),
-      count(Count::fromLayoutBitmask(bitmask)) {
-}
-
-VulkanDescriptorSetLayout::~VulkanDescriptorSetLayout() {
-    vkDestroyDescriptorSetLayout(mDevice, vklayout, VKALLOC);
-}
-
 VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
    : HwProgram(builder.getName()),
      VulkanResource(VulkanResourceType::PROGRAM),
      mInfo(new PipelineInfo()),
      mDevice(device) {
-
-    constexpr uint8_t UBO_MODULE_OFFSET = (sizeof(UniformBufferBitmask) * 8) / MAX_SHADER_MODULES;
-    constexpr uint8_t SAMPLER_MODULE_OFFSET = (sizeof(SamplerBitmask) * 8) / MAX_SHADER_MODULES;
-    constexpr uint8_t INPUT_ATTACHMENT_MODULE_OFFSET =
-            (sizeof(InputAttachmentBitmask) * 8) / MAX_SHADER_MODULES;
-
    Program::ShaderSource const& blobs = builder.getShadersSource();
    auto& modules = mInfo->shaders;

    auto const& specializationConstants = builder.getSpecializationConstants();

    std::vector<uint32_t> shader;
-
-    // TODO: this will be moved out of the shader as the descriptor set layout will be provided by
-    // Filament instead of parsed from the shaders. See [GDSR] in VulkanDescriptorSetManager.h
-    UniformBufferBitmask uboMask = 0;
-    SamplerBitmask samplerMask = 0;
-    InputAttachmentBitmask inputAttachmentMask = 0;
-
-    static_assert(static_cast<ShaderStage>(0) == ShaderStage::VERTEX &&
-            static_cast<ShaderStage>(1) == ShaderStage::FRAGMENT &&
-            MAX_SHADER_MODULES == 2);
-
    for (size_t i = 0; i < MAX_SHADER_MODULES; i++) {
        Program::ShaderBlob const& blob = blobs[i];

@@ -169,12 +72,6 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
            dataSize = shader.size() * 4;
        }

-        auto const [ubo, sampler, inputAttachment] = getProgramBindings(blob);
-        uboMask |= (static_cast<UniformBufferBitmask>(ubo) << (UBO_MODULE_OFFSET * i));
-        samplerMask |= (static_cast<SamplerBitmask>(sampler) << (SAMPLER_MODULE_OFFSET * i));
-        inputAttachmentMask |= (static_cast<InputAttachmentBitmask>(inputAttachment)
-                                << (INPUT_ATTACHMENT_MODULE_OFFSET * i));
-
        VkShaderModule& module = modules[i];
        VkShaderModuleCreateInfo moduleInfo = {
            .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
@@ -202,33 +99,20 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
    }

-    LayoutDescriptionList& layouts = mInfo->layouts;
-    layouts[0].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
-            countBits(collapseStages(uboMask)));
-    layouts[1].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
-            countBits(collapseStages(samplerMask)));
-    layouts[2].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
-            countBits(collapseStages(inputAttachmentMask)));
-
-    addDescriptors(uboMask, layouts[0].bindings);
-    addDescriptors(samplerMask, layouts[1].bindings);
-    addDescriptors(inputAttachmentMask, layouts[2].bindings);
-
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
    auto& bindingToName = mInfo->bindingToName;
 #endif

    auto& groupInfo = builder.getSamplerGroupInfo();
    auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
-    auto& bindings = mInfo->bindings;
+    auto& usage = mInfo->usage;
    for (uint8_t groupInd = 0; groupInd < Program::SAMPLER_BINDING_COUNT; groupInd++) {
        auto const& group = groupInfo[groupInd];
        auto const& samplers = group.samplers;
        for (size_t i = 0; i < samplers.size(); ++i) {
            uint32_t const binding = samplers[i].binding;
            bindingToSamplerIndex[binding] = (groupInd << 8) | (0xff & i);
-            assert_invariant(bindings.find(binding) == bindings.end());
-            bindings.insert(binding);
+            usage = VulkanPipelineCache::getUsageFlags(binding, group.stageFlags, usage);

 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
            bindingToName[binding] = samplers[i].name.c_str();
@@ -242,6 +126,24 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
 }

+VulkanProgram::VulkanProgram(VkDevice device, VkShaderModule vs, VkShaderModule fs,
+        CustomSamplerInfoList const& samplerInfo) noexcept
+    : VulkanResource(VulkanResourceType::PROGRAM),
+      mInfo(new PipelineInfo()),
+      mDevice(device) {
+    mInfo->shaders[0] = vs;
+    mInfo->shaders[1] = fs;
+    auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
+    auto& usage = mInfo->usage;
+    bindingToSamplerIndex.resize(samplerInfo.size());
+    for (uint16_t binding = 0; binding < samplerInfo.size(); ++binding) {
+        auto const& sampler = samplerInfo[binding];
+        bindingToSamplerIndex[binding]
+                = (sampler.groupIndex << 8) | (0xff & sampler.samplerIndex);
+        usage = VulkanPipelineCache::getUsageFlags(binding, sampler.flags, usage);
+    }
+}
+
 VulkanProgram::~VulkanProgram() {
    for (auto shader: mInfo->shaders) {
        vkDestroyShaderModule(mDevice, shader, VKALLOC);
@@ -285,13 +187,13 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica

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

    // Create sidecar MSAA textures for color attachments if they don't already exist.
    for (int index = 0; index < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; index++) {
-        VulkanAttachment const& spec = color[index];
+        const VulkanAttachment& spec = color[index];
        VulkanTexture* texture = (VulkanTexture*) spec.texture;
        if (texture && texture->samples == 1) {
            auto msTexture = texture->getSidecar();
@@ -354,19 +256,19 @@ VkExtent2D VulkanRenderTarget::getExtent() const {
    return {width, height};
 }

-VulkanAttachment& VulkanRenderTarget::getColor(int target) {
+VulkanAttachment VulkanRenderTarget::getColor(int target) const {
    return mColor[target];
 }

-VulkanAttachment& VulkanRenderTarget::getMsaaColor(int target) {
+VulkanAttachment VulkanRenderTarget::getMsaaColor(int target) const {
    return mMsaaAttachments[target];
 }

-VulkanAttachment& VulkanRenderTarget::getDepth() {
+VulkanAttachment VulkanRenderTarget::getDepth() const {
    return mDepth;
 }

-VulkanAttachment& VulkanRenderTarget::getMsaaDepth() {
+VulkanAttachment VulkanRenderTarget::getMsaaDepth() const {
    return mMsaaDepthAttachment;
 }

@@ -503,35 +405,4 @@ VulkanRenderPrimitive::VulkanRenderPrimitive(VulkanResourceAllocator* resourceAl
    mResources.acquire(indexBuffer);
 }

-using Bitmask = VulkanDescriptorSetLayout::Bitmask;
-
-Bitmask Bitmask::fromBackendLayout(descset::DescriptorSetLayout const& layout) {
-    Bitmask mask;
-    for (auto const& binding: layout.bindings) {
-        switch (binding.type) {
-            case descset::DescriptorType::UNIFORM_BUFFER: {
-                if (binding.flags == descset::DescriptorFlags::DYNAMIC_OFFSET) {
-                    mask.dynamicUbo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
-                            binding.binding);
-                } else {
-                    mask.ubo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
-                            binding.binding);
-                }
-                break;
-            }
-            case descset::DescriptorType::SAMPLER: {
-                mask.sampler |= fromStageFlags<SamplerBitmask>(binding.stageFlags, binding.binding);
-                break;
-            }
-            case descset::DescriptorType::INPUT_ATTACHMENT: {
-                mask.inputAttachment |=
-                        fromStageFlags<InputAttachmentBitmask>(binding.stageFlags, binding.binding);
-                break;
-            }
-        }
-    }
-    return mask;
-}
-
-
 } // namespace filament::backend
@@ -21,171 +21,35 @@
 #include "DriverBase.h"

 #include "VulkanBuffer.h"
+#include "VulkanPipelineCache.h"
 #include "VulkanResources.h"
 #include "VulkanSwapChain.h"
 #include "VulkanTexture.h"
-#include "VulkanUtility.h"

 #include "private/backend/SamplerGroup.h"
-#include "utils/FixedCapacityVector.h"
-#include "vulkan/vulkan_core.h"

 #include <utils/Mutex.h>
 #include <utils/StructureOfArrays.h>

 namespace filament::backend {

-using namespace descset;
-
 class VulkanTimestamps;

-struct VulkanDescriptorSetLayout : public VulkanResource {
-    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 3;
-
-    // The bitmask representation of a set layout.
-    struct Bitmask {
-        UniformBufferBitmask ubo = 0;         // 4 bytes
-        UniformBufferBitmask dynamicUbo = 0;  // 4 bytes
-        SamplerBitmask sampler = 0;           // 8 bytes
-        InputAttachmentBitmask inputAttachment = 0; // 1 bytes
-
-        // Because we're using this struct as hash key, must make it's 8-bytes aligned, with no
-        // unaccounted bytes.
-        uint8_t padding0 = 0; // 1 bytes
-        uint16_t padding1 = 0;// 2 bytes
-        uint32_t padding2 = 0;// 4 bytes
-
-        bool operator==(Bitmask const& right) const {
-            return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
-                   inputAttachment == right.inputAttachment;
-        }
-
-        static Bitmask fromBackendLayout(descset::DescriptorSetLayout const& layout);
-    };
-
-    // This is a convenience struct to quickly check layout compatibility in terms of descriptor set
-    // pools.
-    struct Count {
-        uint32_t ubo = 0;
-        uint32_t dynamicUbo = 0;
-        uint32_t sampler = 0;
-        uint32_t inputAttachment = 0;
-
-        bool operator==(Count const& right) const noexcept {
-            return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
-                   inputAttachment == right.inputAttachment;
-        }
-
-        static inline Count fromLayoutBitmask(Bitmask const& mask) {
-            return {
-                    .ubo = countBits(collapseStages(mask.ubo)),
-                    .dynamicUbo = countBits(collapseStages(mask.dynamicUbo)),
-                    .sampler = countBits(collapseStages(mask.sampler)),
-                    .inputAttachment = countBits(collapseStages(mask.inputAttachment)),
-            };
-        }
-
-        Count operator*(uint16_t mult) const noexcept {
-            // TODO: check for overflow.
-
-            Count ret;
-            ret.ubo = ubo * mult;
-            ret.dynamicUbo = dynamicUbo * mult;
-            ret.sampler = sampler * mult;
-            ret.inputAttachment = inputAttachment * mult;
-            return ret;
-        }
-    };
-
-    static_assert(sizeof(Bitmask) % 8 == 0);
-
-    explicit VulkanDescriptorSetLayout(VkDevice device, VkDescriptorSetLayoutCreateInfo const& info,
-            Bitmask const& bitmask);
-
-    ~VulkanDescriptorSetLayout();
-
-    VkDevice const mDevice;
-    VkDescriptorSetLayout const vklayout;
-    Bitmask const bitmask;
-
-    // This is a convenience struct so that we don't have to iterate through all the bits of the
-    // bitmask (which correspondings to binding indices).
-    struct _Bindings {
-        utils::FixedCapacityVector<uint8_t> const ubo;
-        utils::FixedCapacityVector<uint8_t> const dynamicUbo;
-        utils::FixedCapacityVector<uint8_t> const sampler;
-        utils::FixedCapacityVector<uint8_t> const inputAttachment;
-    } bindings;
-
-    Count const count;
-
-private:
-
-    template <typename MaskType>
-    utils::FixedCapacityVector<uint8_t> bits(MaskType mask) {
-        utils::FixedCapacityVector<uint8_t> ret =
-                utils::FixedCapacityVector<uint8_t>::with_capacity(countBits(mask));
-        for (uint8_t i = 0; i < sizeof(mask) * 4; ++i) {
-            if (mask & (1 << i)) {
-                ret.push_back(i);
-            }
-        }
-        return ret;
-    }
-
-    _Bindings getBindings(Bitmask const& bitmask) {
-        auto const uboCollapsed = collapseStages(bitmask.ubo);
-        auto const dynamicUboCollapsed = collapseStages(bitmask.dynamicUbo);
-        auto const samplerCollapsed = collapseStages(bitmask.sampler);
-        auto const inputAttachmentCollapsed = collapseStages(bitmask.inputAttachment);
-        return {
-            bits(uboCollapsed),
-            bits(dynamicUboCollapsed),
-            bits(samplerCollapsed),
-            bits(inputAttachmentCollapsed),
-        };
-    }
-};
-
-using VulkanDescriptorSetLayoutList = std::array<Handle<VulkanDescriptorSetLayout>,
-        VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
-
-struct VulkanDescriptorSet : public VulkanResource {
-public:
-    // Because we need to recycle descriptor sets not used, we allow for a callback that the "Pool"
-    // can use to repackage the vk handle.
-    using OnRecycle = std::function<void()>;
-
-    VulkanDescriptorSet(VulkanResourceAllocator* allocator,
-            VkDescriptorSet rawSet, OnRecycle&& onRecycleFn)
-        : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
-          resources(allocator),
-          vkSet(rawSet),
-          mOnRecycleFn(std::move(onRecycleFn)) {}
-
-    ~VulkanDescriptorSet() {
-        if (mOnRecycleFn) {
-            mOnRecycleFn();
-        }
-    }
-
-    // TODO: maybe change to fixed size for performance.
-    VulkanAcquireOnlyResourceManager resources;
-    VkDescriptorSet const vkSet;
-
-private:
-    OnRecycle mOnRecycleFn;
-};
-
-using VulkanDescriptorSetList = std::array<Handle<VulkanDescriptorSet>,
-        VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
-
 struct VulkanProgram : public HwProgram, VulkanResource {

-    using BindingList = CappedArray<uint16_t, MAX_SAMPLER_COUNT>;
-
    VulkanProgram(VkDevice device, Program const& builder) noexcept;

+    struct CustomSamplerInfo {
+        uint8_t groupIndex;
+        uint8_t samplerIndex;
+        ShaderStageFlags flags;
+    };
+    using CustomSamplerInfoList = utils::FixedCapacityVector<CustomSamplerInfo>;
+
+    // We allow custom descriptor of the samplers within shaders.  This is needed if we want to use
+    // a program that exists only in the backend - for example, for shader-based bliting.
+    VulkanProgram(VkDevice device, VkShaderModule vs, VkShaderModule fs,
+            CustomSamplerInfoList const& samplerInfo) noexcept;
    ~VulkanProgram();

    inline VkShaderModule getVertexShader() const {
@@ -194,53 +58,39 @@ struct VulkanProgram : public HwProgram, VulkanResource {

    inline VkShaderModule getFragmentShader() const { return mInfo->shaders[1]; }

+    inline VulkanPipelineCache::UsageFlags getUsage() const { return mInfo->usage; }
+
    inline utils::FixedCapacityVector<uint16_t> const& getBindingToSamplerIndex() const {
        return mInfo->bindingToSamplerIndex;
    }

-    // Get a list of the sampler binding indices so that we don't have to loop through all possible
-    // samplers.
-    inline BindingList const& getBindings() const { return mInfo->bindings; }
-
-    // TODO: this is currently not used. This will replace getLayoutDescriptionList below.
-    // inline descset::DescriptorSetLayout const& getLayoutDescription() const {
-    //    return mInfo->layout;
-    // }
-    // In the usual case, we would have just one layout per program. But in the current setup, we
-    // have a set/layout for each descriptor type. This will be changed in the future.
-    using LayoutDescriptionList = std::array<descset::DescriptorSetLayout,
-            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
-    inline LayoutDescriptionList const& getLayoutDescriptionList() const { return mInfo->layouts; }
-
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
    inline utils::FixedCapacityVector<std::string> const& getBindingToName() const {
        return mInfo->bindingToName;
    }
 #endif

+private:
    // TODO: handle compute shaders.
    // The expected order of shaders - from frontend to backend - is vertex, fragment, compute.
    static constexpr uint8_t const MAX_SHADER_MODULES = 2;

-private:
    struct PipelineInfo {
        PipelineInfo()
            : bindingToSamplerIndex(MAX_SAMPLER_COUNT, 0xffff)
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
            , bindingToName(MAX_SAMPLER_COUNT, "")
-#endif
+#endif              
            {}

-        BindingList bindings;
+        // This bitset maps to each of the sampler in the sampler groups associated with this
+        // program, and whether each sampler is used in which shader (i.e. vert, frag, compute).
+        VulkanPipelineCache::UsageFlags usage;

        // We store the samplerGroupIndex as the top 8-bit and the index within each group as the lower 8-bit.
        utils::FixedCapacityVector<uint16_t> bindingToSamplerIndex;
        VkShaderModule shaders[MAX_SHADER_MODULES] = { VK_NULL_HANDLE };

-        // TODO: Use this instead of `layouts` after Filament-side Descriptor Set API is in place.
-        // descset::DescriptorSetLayout layout;
-        LayoutDescriptionList layouts;
-
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
        // We store the sampler name mapped from binding index (only for debug purposes).
        utils::FixedCapacityVector<std::string> bindingToName;
@@ -274,11 +124,10 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
    void transformClientRectToPlatform(VkRect2D* bounds) const;
    void transformClientRectToPlatform(VkViewport* bounds) const;
    VkExtent2D getExtent() const;
-    // We return references in the following methods to avoid a copy.
-    VulkanAttachment& getColor(int target);
-    VulkanAttachment& getMsaaColor(int target);
-    VulkanAttachment& getDepth();
-    VulkanAttachment& getMsaaDepth();
+    VulkanAttachment getColor(int target) const;
+    VulkanAttachment getMsaaColor(int target) const;
+    VulkanAttachment getDepth() const;
+    VulkanAttachment getMsaaDepth() const;
    uint8_t getColorTargetCount(const VulkanRenderPass& pass) const;
    uint8_t getSamples() const { return mSamples; }
    bool hasDepth() const { return mDepth.texture; }
@@ -26,10 +26,38 @@

 using namespace bluevk;

-namespace filament::backend::imgutil {
+namespace filament::backend {

 namespace {

+inline VkImageLayout getVkImageLayout(VulkanLayout layout) {
+    switch (layout) {
+        case VulkanLayout::UNDEFINED:
+            return VK_IMAGE_LAYOUT_UNDEFINED;
+        case VulkanLayout::READ_WRITE:
+            return VK_IMAGE_LAYOUT_GENERAL;
+        case VulkanLayout::READ_ONLY:
+            return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+        case VulkanLayout::TRANSFER_SRC:
+            return VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
+        case VulkanLayout::TRANSFER_DST:
+            return VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+        case VulkanLayout::DEPTH_ATTACHMENT:
+            return VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
+        case VulkanLayout::DEPTH_SAMPLER:
+            return VK_IMAGE_LAYOUT_GENERAL;
+        case VulkanLayout::PRESENT:
+            return VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
+        // Filament sometimes samples from one miplevel while writing to another level in the same
+        // texture (e.g. bloom does this). Moreover we'd like to avoid lots of expensive layout
+        // transitions. So, keep it simple and use GENERAL for all color-attachable textures.
+        case VulkanLayout::COLOR_ATTACHMENT:
+            return VK_IMAGE_LAYOUT_GENERAL;
+        case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:
+            return VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
+    }
+}
+
 inline std::tuple<VkAccessFlags, VkAccessFlags, VkPipelineStageFlags, VkPipelineStageFlags,
        VkImageLayout, VkImageLayout>
 getVkTransition(const VulkanLayoutTransition& transition) {
@@ -42,11 +70,11 @@ getVkTransition(const VulkanLayoutTransition& transition) {
            srcStage = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
            break;
        case VulkanLayout::COLOR_ATTACHMENT:
-            srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT |
-                            VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
-                            VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-                       VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+            srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT
+                            | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
+                            | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT
+                       | VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
            break;
        case VulkanLayout::READ_WRITE:
            srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
@@ -126,12 +154,27 @@ getVkTransition(const VulkanLayoutTransition& transition) {
    }

    return std::make_tuple(srcAccessMask, dstAccessMask, srcStage, dstStage,
-            getVkLayout(transition.oldLayout), getVkLayout(transition.newLayout));
+            getVkImageLayout(transition.oldLayout), getVkImageLayout(transition.newLayout));
 }

 }// anonymous namespace

-void transitionLayout(VkCommandBuffer cmdbuffer,
+VkImageViewType VulkanImageUtility::getViewType(SamplerType target) {
+    switch (target) {
+        case SamplerType::SAMPLER_CUBEMAP:
+            return VK_IMAGE_VIEW_TYPE_CUBE;
+        case SamplerType::SAMPLER_2D_ARRAY:
+            return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
+        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
+            return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
+        case SamplerType::SAMPLER_3D:
+            return VK_IMAGE_VIEW_TYPE_3D;
+        default:
+            return VK_IMAGE_VIEW_TYPE_2D;
+    }
+}
+
+void VulkanImageUtility::transitionLayout(VkCommandBuffer cmdbuffer,
        VulkanLayoutTransition transition) {
    if (transition.oldLayout == transition.newLayout) {
        return;
@@ -154,6 +197,10 @@ void transitionLayout(VkCommandBuffer cmdbuffer,
    vkCmdPipelineBarrier(cmdbuffer, srcStage, dstStage, 0, 0, nullptr, 0, nullptr, 1, &barrier);
 }

+VkImageLayout VulkanImageUtility::getVkLayout(VulkanLayout layout) {
+    return getVkImageLayout(layout);
+}
+
 }// namespace filament::backend

 bool operator<(const VkImageSubresourceRange& a, const VkImageSubresourceRange& b) {
@@ -170,12 +217,12 @@ bool operator<(const VkImageSubresourceRange& a, const VkImageSubresourceRange&
    return false;
 }

-#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION) || FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION | FVK_DEBUG_TEXTURE)
 #define CASE(VALUE)                                                                                \
    case filament::backend::VulkanLayout::VALUE: {                                                 \
        out << #VALUE;                                                                             \
        out << " ["                                                                                \
-            << filament::backend::imgutil::getVkLayout(                                            \
+            << filament::backend::VulkanImageUtility::getVkLayout(                                 \
                       filament::backend::VulkanLayout::VALUE)                                     \
            << "]";                                                                                \
        break;                                                                                     \
@@ -59,85 +59,44 @@ struct VulkanLayoutTransition {
    VkImageSubresourceRange subresources;
 };

-namespace imgutil {
+class VulkanImageUtility {
+public:
+    static VkImageViewType getViewType(SamplerType target);

-inline VkImageViewType getViewType(SamplerType target) {
-    switch (target) {
-        case SamplerType::SAMPLER_CUBEMAP:
-            return VK_IMAGE_VIEW_TYPE_CUBE;
-        case SamplerType::SAMPLER_2D_ARRAY:
-            return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
-        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
-            return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
-        case SamplerType::SAMPLER_3D:
-            return VK_IMAGE_VIEW_TYPE_3D;
-        default:
-            return VK_IMAGE_VIEW_TYPE_2D;
-    }
-}
-
-inline VulkanLayout getDefaultLayout(TextureUsage usage) {
-    if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
-        if (any(usage & TextureUsage::SAMPLEABLE)) {
-            return VulkanLayout::DEPTH_SAMPLER;
-        } else {
-            return VulkanLayout::DEPTH_ATTACHMENT;
+    inline static VulkanLayout getDefaultLayout(TextureUsage usage) {
+        if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
+            if (any(usage & TextureUsage::SAMPLEABLE)) {
+                return VulkanLayout::DEPTH_SAMPLER;
+            } else {
+                return VulkanLayout::DEPTH_ATTACHMENT;
+            }
        }
+
+        if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
+            return VulkanLayout::COLOR_ATTACHMENT;
+        }
+        // Finally, the layout for an immutable texture is optimal read-only.        
+        return VulkanLayout::READ_ONLY;
    }

-    if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
-        return VulkanLayout::COLOR_ATTACHMENT;
+    inline static VulkanLayout getDefaultLayout(VkImageUsageFlags vkusage) {
+        TextureUsage usage {};
+        if (vkusage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+            usage = usage | TextureUsage::DEPTH_ATTACHMENT;
+        }
+        if (vkusage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+            usage = usage | TextureUsage::COLOR_ATTACHMENT;
+        }
+        if (vkusage & VK_IMAGE_USAGE_SAMPLED_BIT) {
+            usage = usage | TextureUsage::SAMPLEABLE;
+        }
+        return getDefaultLayout(usage);
    }
-    // Finally, the layout for an immutable texture is optimal read-only.
-    return VulkanLayout::READ_ONLY;
-}

-inline VulkanLayout getDefaultLayout(VkImageUsageFlags vkusage) {
-    TextureUsage usage{};
-    if (vkusage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
-        usage = usage | TextureUsage::DEPTH_ATTACHMENT;
-    }
-    if (vkusage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
-        usage = usage | TextureUsage::COLOR_ATTACHMENT;
-    }
-    if (vkusage & VK_IMAGE_USAGE_SAMPLED_BIT) {
-        usage = usage | TextureUsage::SAMPLEABLE;
-    }
-    return getDefaultLayout(usage);
-}
-
-constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
-    switch (layout) {
-        case VulkanLayout::UNDEFINED:
-            return VK_IMAGE_LAYOUT_UNDEFINED;
-        case VulkanLayout::READ_WRITE:
-            return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::READ_ONLY:
-            return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
-        case VulkanLayout::TRANSFER_SRC:
-            return VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
-        case VulkanLayout::TRANSFER_DST:
-            return VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
-        case VulkanLayout::DEPTH_ATTACHMENT:
-            return VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
-        case VulkanLayout::DEPTH_SAMPLER:
-            return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::PRESENT:
-            return VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
-        // Filament sometimes samples from one miplevel while writing to another level in the
-        // same texture (e.g. bloom does this). Moreover we'd like to avoid lots of expensive
-        // layout transitions. So, keep it simple and use GENERAL for all color-attachable
-        // textures.
-        case VulkanLayout::COLOR_ATTACHMENT:
-            return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:
-            return VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
-    }
-}
-
-void transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
-
-} // namespace imgutil
+    static VkImageLayout getVkLayout(VulkanLayout layout);
+    
+    static void transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
+};

 } // namespace filament::backend

@@ -31,8 +31,4 @@

 #include "vk_mem_alloc.h"

-VK_DEFINE_HANDLE(VmaAllocator)
-VK_DEFINE_HANDLE(VmaAllocation)
-VK_DEFINE_HANDLE(VmaPool)
-
 #endif // TNT_FILAMENT_BACKEND_VULKANMEMORY_H
@@ -14,9 +14,8 @@
 * limitations under the License.
 */

-#include "VulkanPipelineCache.h"
-#include "VulkanMemory.h"
-#include "caching/VulkanDescriptorSetManager.h"
+#include "vulkan/VulkanMemory.h"
+#include "vulkan/VulkanPipelineCache.h"

 #include <utils/Log.h>
 #include <utils/Panic.h>
@@ -35,9 +34,63 @@ using namespace bluevk;

 namespace filament::backend {

-VulkanPipelineCache::VulkanPipelineCache(VkDevice device, VmaAllocator allocator)
-    : mDevice(device),
-      mAllocator(allocator) {
+static VkShaderStageFlags getShaderStageFlags(VulkanPipelineCache::UsageFlags key, uint16_t binding) {
+    // NOTE: if you modify this function, you also need to modify getUsageFlags.
+    assert_invariant(binding < MAX_SAMPLER_COUNT);
+    VkShaderStageFlags flags = 0;
+    if (key.test(binding)) {
+        flags |= VK_SHADER_STAGE_VERTEX_BIT;
+    }
+    if (key.test(MAX_SAMPLER_COUNT + binding)) {
+        flags |= VK_SHADER_STAGE_FRAGMENT_BIT;
+    }
+    return flags;
+}
+
+VulkanPipelineCache::UsageFlags
+VulkanPipelineCache::getUsageFlags(uint16_t binding, ShaderStageFlags flags, UsageFlags src) {
+    // NOTE: if you modify this function, you also need to modify getShaderStageFlags.
+    assert_invariant(binding < MAX_SAMPLER_COUNT);
+    if (any(flags & ShaderStageFlags::VERTEX)) {
+        src.set(binding);
+    }
+    if (any(flags & ShaderStageFlags::FRAGMENT)) {
+        src.set(MAX_SAMPLER_COUNT + binding);
+    }
+    // TODO: add support for compute by extending SHADER_MODULE_COUNT and ensuring UsageFlags
+    // has 186 bits (MAX_SAMPLER_COUNT * 3)
+    // assert_invariant(!any(flags & ~(ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT)));
+    return src;
+}
+
+VulkanPipelineCache::UsageFlags VulkanPipelineCache::disableUsageFlags(uint16_t binding,
+        UsageFlags src) {
+    src.unset(binding);
+    src.unset(MAX_SAMPLER_COUNT + binding);
+    return src;
+}
+
+VulkanPipelineCache::VulkanPipelineCache(VulkanResourceAllocator* allocator)
+    : mResourceAllocator(allocator),
+      mPipelineBoundResources(allocator) {
+    mDummyBufferWriteInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+    mDummyBufferWriteInfo.pNext = nullptr;
+    mDummyBufferWriteInfo.dstArrayElement = 0;
+    mDummyBufferWriteInfo.descriptorCount = 1;
+    mDummyBufferWriteInfo.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+    mDummyBufferWriteInfo.pImageInfo = nullptr;
+    mDummyBufferWriteInfo.pBufferInfo = &mDummyBufferInfo;
+    mDummyBufferWriteInfo.pTexelBufferView = nullptr;
+
+    mDummyTargetInfo.imageLayout = VulkanImageUtility::getVkLayout(VulkanLayout::READ_ONLY);
+    mDummyTargetWriteInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+    mDummyTargetWriteInfo.pNext = nullptr;
+    mDummyTargetWriteInfo.dstArrayElement = 0;
+    mDummyTargetWriteInfo.descriptorCount = 1;
+    mDummyTargetWriteInfo.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
+    mDummyTargetWriteInfo.pImageInfo = &mDummyTargetInfo;
+    mDummyTargetWriteInfo.pBufferInfo = nullptr;
+    mDummyTargetWriteInfo.pTexelBufferView = nullptr;
 }

 VulkanPipelineCache::~VulkanPipelineCache() {
@@ -45,47 +98,244 @@ VulkanPipelineCache::~VulkanPipelineCache() {
    // be explicit about teardown order of various components.
 }

-void VulkanPipelineCache::bindLayout(VkPipelineLayout layout) noexcept {
-    mPipelineRequirements.layout = layout;
+void VulkanPipelineCache::setDevice(VkDevice device, VmaAllocator allocator) {
+    assert_invariant(mDevice == VK_NULL_HANDLE);
+    mDevice = device;
+    mAllocator = allocator;
+    mDescriptorPool = createDescriptorPool(mDescriptorPoolSize);
+
+    // Formulate some dummy objects and dummy descriptor info used only for clearing out unused
+    // bindings. This is especially crucial after a texture has been destroyed. Since core Vulkan
+    // does not allow specifying VK_NULL_HANDLE without the robustness2 extension, we would need to
+    // change the pipeline layout more frequently if we wanted to get rid of these dummy objects.
+
+    VkBufferCreateInfo bufferInfo {
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .size = 16,
+        .usage = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT,
+    };
+    VmaAllocationCreateInfo allocInfo { .usage = VMA_MEMORY_USAGE_GPU_ONLY };
+    vmaCreateBuffer(mAllocator, &bufferInfo, &allocInfo, &mDummyBuffer, &mDummyMemory, nullptr);
+
+    mDummyBufferInfo.buffer = mDummyBuffer;
+    mDummyBufferInfo.range = bufferInfo.size;
 }

-VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::getOrCreatePipeline() noexcept {
-    // If a cached object exists, re-use it, otherwise create a new one.
-    if (PipelineMap::iterator pipelineIter = mPipelines.find(mPipelineRequirements);
-            pipelineIter != mPipelines.end()) {
-        auto& pipeline = pipelineIter.value();
-        pipeline.lastUsed = mCurrentTime;
-        return &pipeline;
+bool VulkanPipelineCache::bindDescriptors(VkCommandBuffer cmdbuffer) noexcept {
+    DescriptorMap::iterator descriptorIter = mDescriptorSets.find(mDescriptorRequirements);
+
+    // Check if the required descriptors are already bound. If so, there's no need to do anything.
+    if (DescEqual equals; UTILS_LIKELY(equals(mBoundDescriptor, mDescriptorRequirements))) {
+
+        // If the pipeline state during an app's first draw call happens to match the default state
+        // vector of the cache, then the cache is uninitialized and we should not return early.
+        if (UTILS_LIKELY(!mDescriptorSets.empty())) {
+
+            // Since the descriptors are already bound, they should be found in the cache.
+            assert_invariant(descriptorIter != mDescriptorSets.end());
+
+            // Update the LRU "time stamp" (really a count of cmd buf submissions) before returning.
+            descriptorIter.value().lastUsed = mCurrentTime;
+            return true;
+        }
    }
-    auto ret = createPipeline();
-    ret->lastUsed = mCurrentTime;
-    return ret;
+
+    // If a cached object exists, re-use it, otherwise create a new one.
+    DescriptorCacheEntry* cacheEntry = UTILS_LIKELY(descriptorIter != mDescriptorSets.end()) ?
+            &descriptorIter.value() : createDescriptorSets();
+
+    // If a descriptor set overflow occurred, allow higher levels to handle it gracefully.
+    assert_invariant(cacheEntry != nullptr);
+    if (UTILS_UNLIKELY(cacheEntry == nullptr)) {
+        return false;
+    }
+
+    cacheEntry->lastUsed = mCurrentTime;
+    mBoundDescriptor = mDescriptorRequirements;
+    // This passes the currently "bound" uniform buffer objects to pipeline that will be used in the
+    // draw call.
+    auto resourceEntry = mDescriptorResources.find(cacheEntry->id);
+    if (resourceEntry == mDescriptorResources.end()) {
+        mDescriptorResources[cacheEntry->id]
+                = std::make_unique<VulkanAcquireOnlyResourceManager>(mResourceAllocator);
+        resourceEntry = mDescriptorResources.find(cacheEntry->id);
+    }
+    resourceEntry->second->acquireAll(&mPipelineBoundResources);
+
+    vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
+            getOrCreatePipelineLayout()->handle, 0, VulkanPipelineCache::DESCRIPTOR_TYPE_COUNT,
+            cacheEntry->handles.data(), 0, nullptr);
+
+    return true;
 }

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

-    PipelineCacheEntry* cacheEntry = getOrCreatePipeline();
+    PipelineMap::iterator pipelineIter = mPipelines.find(mPipelineRequirements);
+
    // Check if the required pipeline is already bound.
-    if (cacheEntry->handle == commands->pipeline()) {
-        return;
+    if (PipelineEqual equals; UTILS_LIKELY(equals(mBoundPipeline, mPipelineRequirements))) {
+        assert_invariant(pipelineIter != mPipelines.end());
+        pipelineIter.value().lastUsed = mCurrentTime;
+        return true;
    }

+    // If a cached object exists, re-use it, otherwise create a new one.
+    PipelineCacheEntry* cacheEntry = UTILS_LIKELY(pipelineIter != mPipelines.end()) ?
+            &pipelineIter.value() : createPipeline();
+
    // If an error occurred, allow higher levels to handle it gracefully.
-    assert_invariant(cacheEntry != nullptr && "Failed to create/find pipeline");
+    assert_invariant(cacheEntry != nullptr);
+    if (UTILS_UNLIKELY(cacheEntry == nullptr)) {
+        return false;
+    }
+
+    cacheEntry->lastUsed = mCurrentTime;
+    getOrCreatePipelineLayout()->lastUsed = mCurrentTime;

    mBoundPipeline = mPipelineRequirements;
+
    vkCmdBindPipeline(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, cacheEntry->handle);
-    commands->setPipeline(cacheEntry->handle);
+    return true;
 }

 void VulkanPipelineCache::bindScissor(VkCommandBuffer cmdbuffer, VkRect2D scissor) noexcept {
-    vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
+    if (UTILS_UNLIKELY(!equivalent(mCurrentScissor, scissor))) {
+        mCurrentScissor = scissor;
+        vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
+    }
+}
+
+VulkanPipelineCache::DescriptorCacheEntry* VulkanPipelineCache::createDescriptorSets() noexcept {
+    PipelineLayoutCacheEntry* layoutCacheEntry = getOrCreatePipelineLayout();
+
+    DescriptorCacheEntry descriptorCacheEntry = {
+        .pipelineLayout = mPipelineRequirements.layout,
+        .id = mDescriptorCacheEntryCount++,
+    };
+
+    // Each of the arenas for this particular layout are guaranteed to have the same size. Check
+    // the first arena to see if any descriptor sets are available that can be re-claimed. If not,
+    // create brand new ones (one for each type). They will be added to the arena later, after they
+    // are no longer used. This occurs during the cleanup phase during command buffer submission.
+    auto& descriptorSetArenas = layoutCacheEntry->descriptorSetArenas;
+    if (descriptorSetArenas[0].empty()) {
+
+        // If allocating a new descriptor set from the pool would cause it to overflow, then
+        // recreate the pool. The number of descriptor sets that have already been allocated from
+        // the pool is the sum of the "active" descriptor sets (mDescriptorSets) and the "dormant"
+        // descriptor sets (mDescriptorArenasCount).
+        //
+        // NOTE: technically both sides of the inequality below should be multiplied by
+        // DESCRIPTOR_TYPE_COUNT to get the true number of descriptor sets.
+        if (mDescriptorSets.size() + mDescriptorArenasCount + 1 > mDescriptorPoolSize) {
+            growDescriptorPool();
+        }
+
+        VkDescriptorSetAllocateInfo allocInfo = {};
+        allocInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
+        allocInfo.descriptorPool = mDescriptorPool;
+        allocInfo.descriptorSetCount = DESCRIPTOR_TYPE_COUNT;
+        allocInfo.pSetLayouts = layoutCacheEntry->descriptorSetLayouts.data();
+        VkResult error = vkAllocateDescriptorSets(mDevice, &allocInfo,
+                descriptorCacheEntry.handles.data());
+        assert_invariant(error == VK_SUCCESS);
+        if (error != VK_SUCCESS) {
+            return nullptr;
+        }
+    } else {
+        for (uint32_t i = 0; i < DESCRIPTOR_TYPE_COUNT; ++i) {
+            descriptorCacheEntry.handles[i] = descriptorSetArenas[i].back();
+            descriptorSetArenas[i].pop_back();
+        }
+        assert_invariant(mDescriptorArenasCount > 0);
+        mDescriptorArenasCount--;
+    }
+
+    // Rewrite every binding in the new descriptor sets.
+    VkDescriptorBufferInfo descriptorBuffers[UBUFFER_BINDING_COUNT];
+    VkDescriptorImageInfo descriptorSamplers[SAMPLER_BINDING_COUNT];
+    VkDescriptorImageInfo descriptorInputAttachments[INPUT_ATTACHMENT_COUNT];
+    VkWriteDescriptorSet descriptorWrites[UBUFFER_BINDING_COUNT + SAMPLER_BINDING_COUNT +
+            INPUT_ATTACHMENT_COUNT];
+    uint32_t nwrites = 0;
+    VkWriteDescriptorSet* writes = descriptorWrites;
+    nwrites = 0;
+    for (uint32_t binding = 0; binding < UBUFFER_BINDING_COUNT; binding++) {
+        VkWriteDescriptorSet& writeInfo = writes[nwrites++];
+        if (mDescriptorRequirements.uniformBuffers[binding]) {
+            VkDescriptorBufferInfo& bufferInfo = descriptorBuffers[binding];
+            bufferInfo.buffer = mDescriptorRequirements.uniformBuffers[binding];
+            bufferInfo.offset = mDescriptorRequirements.uniformBufferOffsets[binding];
+            bufferInfo.range = mDescriptorRequirements.uniformBufferSizes[binding];
+
+            // We store size with 32 bits, so our "WHOLE" sentinel is different from Vk.
+            if (bufferInfo.range == WHOLE_SIZE) {
+                bufferInfo.range = VK_WHOLE_SIZE;
+            }
+
+            writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+            writeInfo.pNext = nullptr;
+            writeInfo.dstArrayElement = 0;
+            writeInfo.descriptorCount = 1;
+            writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+            writeInfo.pImageInfo = nullptr;
+            writeInfo.pBufferInfo = &bufferInfo;
+            writeInfo.pTexelBufferView = nullptr;
+        } else {
+            writeInfo = mDummyBufferWriteInfo;
+            assert_invariant(mDummyBufferWriteInfo.pBufferInfo->buffer);
+        }
+        assert_invariant(writeInfo.pBufferInfo->buffer);
+        writeInfo.dstSet = descriptorCacheEntry.handles[0];
+        writeInfo.dstBinding = binding;
+    }
+    for (uint32_t binding = 0; binding < SAMPLER_BINDING_COUNT; binding++) {
+        if (mDescriptorRequirements.samplers[binding].sampler) {
+            VkWriteDescriptorSet& writeInfo = writes[nwrites++];
+            VkDescriptorImageInfo& imageInfo = descriptorSamplers[binding];
+            imageInfo = mDescriptorRequirements.samplers[binding];
+            writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+            writeInfo.pNext = nullptr;
+            writeInfo.dstArrayElement = 0;
+            writeInfo.descriptorCount = 1;
+            writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+            writeInfo.pImageInfo = &imageInfo;
+            writeInfo.pBufferInfo = nullptr;
+            writeInfo.pTexelBufferView = nullptr;
+            writeInfo.dstSet = descriptorCacheEntry.handles[1];
+            writeInfo.dstBinding = binding;
+        }
+    }
+    for (uint32_t binding = 0; binding < INPUT_ATTACHMENT_COUNT; binding++) {
+        if (mDescriptorRequirements.inputAttachments[binding].imageView) {
+            VkWriteDescriptorSet& writeInfo = writes[nwrites++];
+            VkDescriptorImageInfo& imageInfo = descriptorInputAttachments[binding];
+            imageInfo = mDescriptorRequirements.inputAttachments[binding];
+            writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+            writeInfo.pNext = nullptr;
+            writeInfo.dstArrayElement = 0;
+            writeInfo.descriptorCount = 1;
+            writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
+            writeInfo.pImageInfo = &imageInfo;
+            writeInfo.pBufferInfo = nullptr;
+            writeInfo.pTexelBufferView = nullptr;
+            writeInfo.dstSet = descriptorCacheEntry.handles[2];
+            writeInfo.dstBinding = binding;
+        }
+    }
+
+    vkUpdateDescriptorSets(mDevice, nwrites, writes, 0, nullptr);
+
+    return &mDescriptorSets.emplace(mDescriptorRequirements, descriptorCacheEntry).first.value();
 }

 VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() noexcept {
    assert_invariant(mPipelineRequirements.shaders[0] && "Vertex shader is not bound.");
-    assert_invariant(mPipelineRequirements.layout && "No pipeline layout specified");
+
+    PipelineLayoutCacheEntry* layout = getOrCreatePipelineLayout();
+    assert_invariant(layout);

    VkPipelineShaderStageCreateInfo shaderStages[SHADER_MODULE_COUNT];
    shaderStages[0] = VkPipelineShaderStageCreateInfo{};
@@ -154,7 +404,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n

    VkGraphicsPipelineCreateInfo pipelineCreateInfo = {};
    pipelineCreateInfo.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO;
-    pipelineCreateInfo.layout = mPipelineRequirements.layout;
+    pipelineCreateInfo.layout = layout->handle;
    pipelineCreateInfo.renderPass = mPipelineRequirements.renderPass;
    pipelineCreateInfo.subpass = mPipelineRequirements.subpassIndex;
    pipelineCreateInfo.stageCount = hasFragmentShader ? SHADER_MODULE_COUNT : 1;
@@ -248,6 +498,68 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
    return &mPipelines.emplace(mPipelineRequirements, cacheEntry).first.value();
 }

+VulkanPipelineCache::PipelineLayoutCacheEntry* VulkanPipelineCache::getOrCreatePipelineLayout() noexcept {
+    auto iter = mPipelineLayouts.find(mPipelineRequirements.layout);
+    if (UTILS_LIKELY(iter != mPipelineLayouts.end())) {
+        return &iter.value();
+    }
+
+    PipelineLayoutCacheEntry cacheEntry = {};
+
+    VkDescriptorSetLayoutBinding binding = {};
+    binding.descriptorCount = 1; // NOTE: We never use arrays-of-blocks.
+    binding.stageFlags = VK_SHADER_STAGE_ALL_GRAPHICS; // NOTE: This is potentially non-optimal.
+
+    // First create the descriptor set layout for UBO's.
+    VkDescriptorSetLayoutBinding ubindings[UBUFFER_BINDING_COUNT];
+    binding.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+    for (uint32_t i = 0; i < UBUFFER_BINDING_COUNT; i++) {
+        binding.binding = i;
+        ubindings[i] = binding;
+    }
+    VkDescriptorSetLayoutCreateInfo dlinfo = {};
+    dlinfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
+    dlinfo.bindingCount = UBUFFER_BINDING_COUNT;
+    dlinfo.pBindings = ubindings;
+    vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &cacheEntry.descriptorSetLayouts[0]);
+
+    // Next create the descriptor set layout for samplers.
+    VkDescriptorSetLayoutBinding sbindings[SAMPLER_BINDING_COUNT];
+    binding.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+    for (uint32_t i = 0; i < SAMPLER_BINDING_COUNT; i++) {
+        binding.stageFlags = getShaderStageFlags(mPipelineRequirements.layout, i);
+        binding.binding = i;
+        sbindings[i] = binding;
+    }
+    dlinfo.bindingCount = SAMPLER_BINDING_COUNT;
+    dlinfo.pBindings = sbindings;
+    vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &cacheEntry.descriptorSetLayouts[1]);
+
+    // Next create the descriptor set layout for input attachments.
+    VkDescriptorSetLayoutBinding tbindings[INPUT_ATTACHMENT_COUNT];
+    binding.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
+    binding.stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT;
+    for (uint32_t i = 0; i < INPUT_ATTACHMENT_COUNT; i++) {
+        binding.binding = i;
+        tbindings[i] = binding;
+    }
+    dlinfo.bindingCount = INPUT_ATTACHMENT_COUNT;
+    dlinfo.pBindings = tbindings;
+    vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &cacheEntry.descriptorSetLayouts[2]);
+
+    // Create VkPipelineLayout based on how to resources are bounded.
+    VkPipelineLayoutCreateInfo pPipelineLayoutCreateInfo = {};
+    pPipelineLayoutCreateInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
+    pPipelineLayoutCreateInfo.setLayoutCount = cacheEntry.descriptorSetLayouts.size();
+    pPipelineLayoutCreateInfo.pSetLayouts = cacheEntry.descriptorSetLayouts.data();
+    VkResult result = vkCreatePipelineLayout(mDevice, &pPipelineLayoutCreateInfo, VKALLOC,
+            &cacheEntry.handle);
+    if (UTILS_UNLIKELY(result != VK_SUCCESS)) {
+        return nullptr;
+    }
+    return &mPipelineLayouts.emplace(mPipelineRequirements.layout, cacheEntry).first.value();
+}
+
 void VulkanPipelineCache::bindProgram(VulkanProgram* program) noexcept {
    mPipelineRequirements.shaders[0] = program->getVertexShader();
    mPipelineRequirements.shaders[1] = program->getFragmentShader();
@@ -288,15 +600,97 @@ void VulkanPipelineCache::bindVertexArray(VkVertexInputAttributeDescription cons
    }
 }

+VulkanPipelineCache::UniformBufferBinding VulkanPipelineCache::getUniformBufferBinding(
+        uint32_t bindingIndex) const noexcept {
+    auto& key = mDescriptorRequirements;
+    return {
+        key.uniformBuffers[bindingIndex],
+        key.uniformBufferOffsets[bindingIndex],
+        key.uniformBufferSizes[bindingIndex],
+    };
+}
+
+void VulkanPipelineCache::unbindUniformBuffer(VkBuffer uniformBuffer) noexcept {
+    auto& key = mDescriptorRequirements;
+    for (uint32_t bindingIndex = 0u; bindingIndex < UBUFFER_BINDING_COUNT; ++bindingIndex) {
+        if (key.uniformBuffers[bindingIndex] == uniformBuffer) {
+            key.uniformBuffers[bindingIndex] = {};
+            key.uniformBufferSizes[bindingIndex] = {};
+            key.uniformBufferOffsets[bindingIndex] = {};
+        }
+    }
+}
+
+void VulkanPipelineCache::unbindImageView(VkImageView imageView) noexcept {
+    for (auto& sampler : mDescriptorRequirements.samplers) {
+        if (sampler.imageView == imageView) {
+            sampler = {};
+        }
+    }
+    for (auto& target : mDescriptorRequirements.inputAttachments) {
+        if (target.imageView == imageView) {
+            target = {};
+        }
+    }
+}
+
+void VulkanPipelineCache::bindUniformBufferObject(uint32_t bindingIndex,
+        VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept {
+    VkBuffer buffer = bufferObject->buffer.getGpuBuffer();
+
+    ASSERT_POSTCONDITION(bindingIndex < UBUFFER_BINDING_COUNT,
+            "Uniform bindings overflow: index = %d, capacity = %d.", bindingIndex,
+            UBUFFER_BINDING_COUNT);
+    auto& key = mDescriptorRequirements;
+    key.uniformBuffers[bindingIndex] = buffer;
+
+    if (size == VK_WHOLE_SIZE) {
+        size = WHOLE_SIZE;
+    }
+
+    assert_invariant(offset <= 0xffffffffu);
+    assert_invariant(size <= 0xffffffffu);
+
+    key.uniformBufferOffsets[bindingIndex] = offset;
+    key.uniformBufferSizes[bindingIndex] = size;
+
+    mPipelineBoundResources.acquire(bufferObject);
+}
+
+void VulkanPipelineCache::bindSamplers(VkDescriptorImageInfo samplers[SAMPLER_BINDING_COUNT],
+        VulkanTexture* textures[SAMPLER_BINDING_COUNT], UsageFlags flags) noexcept {
+    for (uint32_t bindingIndex = 0; bindingIndex < SAMPLER_BINDING_COUNT; bindingIndex++) {
+        mDescriptorRequirements.samplers[bindingIndex] = samplers[bindingIndex];
+        if (textures[bindingIndex]) {
+            mPipelineBoundResources.acquire(textures[bindingIndex]);
+        }
+    }
+    mPipelineRequirements.layout = flags;
+}
+
+void VulkanPipelineCache::bindInputAttachment(uint32_t bindingIndex,
+        VkDescriptorImageInfo targetInfo) noexcept {
+    ASSERT_POSTCONDITION(bindingIndex < INPUT_ATTACHMENT_COUNT,
+            "Input attachment bindings overflow: index = %d, capacity = %d.",
+            bindingIndex, INPUT_ATTACHMENT_COUNT);
+    mDescriptorRequirements.inputAttachments[bindingIndex] = targetInfo;
+}
+
 void VulkanPipelineCache::terminate() noexcept {
+    // Symmetric to createLayoutsAndDescriptors.
+    destroyLayoutsAndDescriptors();
    for (auto& iter : mPipelines) {
        vkDestroyPipeline(mDevice, iter.second.handle, VKALLOC);
    }
+    mPipelineBoundResources.clear();
    mPipelines.clear();
    mBoundPipeline = {};
+    vmaDestroyBuffer(mAllocator, mDummyBuffer, mDummyMemory);
+    mDummyBuffer = VK_NULL_HANDLE;
+    mDummyMemory = VK_NULL_HANDLE;
 }

-void VulkanPipelineCache::gc() noexcept {
+void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
    // The timestamp associated with a given cache entry represents "time" as a count of flush
    // events since the cache was constructed. If any cache entry was most recently used over
    // FVK_MAX_PIPELINE_AGE flush events in the past, then we can be sure that it is no longer
@@ -306,22 +700,194 @@ void VulkanPipelineCache::gc() noexcept {
    // The Vulkan spec says: "When a command buffer begins recording, all state in that command
    // buffer is undefined." Therefore, we need to clear all bindings at this time.
    mBoundPipeline = {};
+    mBoundDescriptor = {};
    mCurrentScissor = {};

    // NOTE: Due to robin_map restrictions, we cannot use auto or range-based loops.

+    // Check if any bundles in the cache are no longer in use by any command buffer. Descriptors
+    // from unused bundles are moved back to their respective arenas.
+    using ConstDescIterator = decltype(mDescriptorSets)::const_iterator;
+    for (ConstDescIterator iter = mDescriptorSets.begin(); iter != mDescriptorSets.end();) {
+        const DescriptorCacheEntry& cacheEntry = iter.value();
+        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+            auto& arenas = mPipelineLayouts[cacheEntry.pipelineLayout].descriptorSetArenas;
+            for (uint32_t i = 0; i < DESCRIPTOR_TYPE_COUNT; ++i) {
+                arenas[i].push_back(cacheEntry.handles[i]);
+            }
+            ++mDescriptorArenasCount;
+            mDescriptorResources.erase(cacheEntry.id);
+            iter = mDescriptorSets.erase(iter);
+        } else {
+            ++iter;
+        }
+    }
+
    // Evict any pipelines that have not been used in a while.
    // Any pipeline older than FVK_MAX_COMMAND_BUFFERS can be safely destroyed.
-   using ConstPipeIterator = decltype(mPipelines)::const_iterator;
-   for (ConstPipeIterator iter = mPipelines.begin(); iter != mPipelines.end();) {
-       const PipelineCacheEntry& cacheEntry = iter.value();
-       if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
-           vkDestroyPipeline(mDevice, iter->second.handle, VKALLOC);
-           iter = mPipelines.erase(iter);
-       } else {
-           ++iter;
-       }
-   }
+    using ConstPipeIterator = decltype(mPipelines)::const_iterator;
+    for (ConstPipeIterator iter = mPipelines.begin(); iter != mPipelines.end();) {
+        const PipelineCacheEntry& cacheEntry = iter.value();
+        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+            vkDestroyPipeline(mDevice, iter->second.handle, VKALLOC);
+            iter = mPipelines.erase(iter);
+        } else {
+            ++iter;
+        }
+    }
+
+    // Evict any layouts that have not been used in a while.
+    using ConstLayoutIterator = decltype(mPipelineLayouts)::const_iterator;
+    for (ConstLayoutIterator iter = mPipelineLayouts.begin(); iter != mPipelineLayouts.end();) {
+        const PipelineLayoutCacheEntry& cacheEntry = iter.value();
+        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+            vkDestroyPipelineLayout(mDevice, iter->second.handle, VKALLOC);
+            for (auto setLayout : iter->second.descriptorSetLayouts) {
+                #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
+                PipelineLayoutKey key = iter.key();
+                for (auto& pair : mDescriptorSets) {
+                    assert_invariant(pair.second.pipelineLayout != key);
+                }
+                #endif
+                vkDestroyDescriptorSetLayout(mDevice, setLayout, VKALLOC);
+            }
+            auto& arenas = iter->second.descriptorSetArenas;
+            assert_invariant(mDescriptorArenasCount >= arenas[0].size());
+            mDescriptorArenasCount -= arenas[0].size();
+            for (auto& arena : arenas) {
+                vkFreeDescriptorSets(mDevice, mDescriptorPool, arena.size(), arena.data());
+            }
+            iter = mPipelineLayouts.erase(iter);
+        } else {
+            ++iter;
+        }
+    }
+
+    // If there are no descriptors from any extinct pool that are still in use, we can safely
+    // destroy the extinct pools, which implicitly frees their associated descriptor sets.
+    bool canPurgeExtinctPools = true;
+    for (auto& bundle : mExtinctDescriptorBundles) {
+        if (bundle.lastUsed + FVK_MAX_PIPELINE_AGE >= mCurrentTime) {
+            canPurgeExtinctPools = false;
+            break;
+        }
+    }
+    if (canPurgeExtinctPools) {
+        for (VkDescriptorPool pool : mExtinctDescriptorPools) {
+            vkDestroyDescriptorPool(mDevice, pool, VKALLOC);
+        }
+        mExtinctDescriptorPools.clear();
+
+        for (auto const& entry : mExtinctDescriptorBundles) {
+            mDescriptorResources.erase(entry.id);
+        }
+        mExtinctDescriptorBundles.clear();
+    }
+}
+
+VkDescriptorPool VulkanPipelineCache::createDescriptorPool(uint32_t size) const {
+    VkDescriptorPoolSize poolSizes[DESCRIPTOR_TYPE_COUNT] = {};
+    VkDescriptorPoolCreateInfo poolInfo {
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
+        .maxSets = size * DESCRIPTOR_TYPE_COUNT,
+        .poolSizeCount = DESCRIPTOR_TYPE_COUNT,
+        .pPoolSizes = poolSizes
+    };
+    poolSizes[0].type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+    poolSizes[0].descriptorCount = poolInfo.maxSets * UBUFFER_BINDING_COUNT;
+    poolSizes[1].type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+    poolSizes[1].descriptorCount = poolInfo.maxSets * SAMPLER_BINDING_COUNT;
+    poolSizes[2].type = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
+    poolSizes[2].descriptorCount = poolInfo.maxSets * INPUT_ATTACHMENT_COUNT;
+
+    VkDescriptorPool pool;
+    const UTILS_UNUSED VkResult result = vkCreateDescriptorPool(mDevice, &poolInfo, VKALLOC, &pool);
+    assert_invariant(result == VK_SUCCESS);
+    return pool;
+}
+
+void VulkanPipelineCache::destroyLayoutsAndDescriptors() noexcept {
+    // Our current descriptor set strategy can cause the # of descriptor sets to explode in certain
+    // situations, so it's interesting to report the number that get stuffed into the cache.
+    #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
+    utils::slog.d << "Destroying " << mDescriptorSets.size() << " bundles of descriptor sets."
+            << utils::io::endl;
+    #endif
+
+    mDescriptorSets.clear();
+
+    // Our current layout bundle strategy can cause the # of layout bundles to explode in certain
+    // situations, so it's interesting to report the number that get stuffed into the cache.
+    #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
+    utils::slog.d << "Destroying " << mPipelineLayouts.size() << " pipeline layouts."
+                  << utils::io::endl;
+    #endif
+
+    for (auto& iter : mPipelineLayouts) {
+        vkDestroyPipelineLayout(mDevice, iter.second.handle, VKALLOC);
+        for (auto setLayout : iter.second.descriptorSetLayouts) {
+            vkDestroyDescriptorSetLayout(mDevice, setLayout, VKALLOC);
+        }
+        // There is no need to free descriptor sets individually since destroying the VkDescriptorPool
+        // implicitly frees them.
+    }
+    mPipelineLayouts.clear();
+    vkDestroyDescriptorPool(mDevice, mDescriptorPool, VKALLOC);
+    mDescriptorPool = VK_NULL_HANDLE;
+
+    for (VkDescriptorPool pool : mExtinctDescriptorPools) {
+        vkDestroyDescriptorPool(mDevice, pool, VKALLOC);
+    }
+    mExtinctDescriptorPools.clear();
+    mExtinctDescriptorBundles.clear();
+
+    // Both mDescriptorSets and mExtinctDescriptorBundles have been cleared, so it's safe to call
+    // clear() on mDescriptorResources.
+    mDescriptorResources.clear();
+
+    mBoundDescriptor = {};
+}
+
+void VulkanPipelineCache::growDescriptorPool() noexcept {
+    // We need to destroy the old VkDescriptorPool, but we can't do so immediately because many
+    // of its descriptors are still in use. So, stash it in an "extinct" list.
+    mExtinctDescriptorPools.push_back(mDescriptorPool);
+
+    // Create the new VkDescriptorPool, twice as big as the old one.
+    mDescriptorPoolSize *= 2;
+    mDescriptorPool = createDescriptorPool(mDescriptorPoolSize);
+
+    // Clear out all unused descriptor sets in the arena so they don't get reclaimed. There is no
+    // need to free them individually since the old VkDescriptorPool will be destroyed.
+    for (auto iter = mPipelineLayouts.begin(); iter != mPipelineLayouts.end(); ++iter) {
+        for (auto& arena : iter.value().descriptorSetArenas) {
+            arena.clear();
+        }
+    }
+    mDescriptorArenasCount = 0;
+
+    // Move all in-use descriptors from the primary cache into an "extinct" list, so that they will
+    // later be destroyed rather than reclaimed.
+    using DescIterator = decltype(mDescriptorSets)::iterator;
+    for (DescIterator iter = mDescriptorSets.begin(); iter != mDescriptorSets.end(); ++iter) {
+        mExtinctDescriptorBundles.push_back(iter.value());
+    }
+    mDescriptorSets.clear();
+}
+
+size_t VulkanPipelineCache::PipelineLayoutKeyHashFn::operator()(
+        const PipelineLayoutKey& key) const {
+    std::hash<uint64_t> hasher;
+    auto h0 = hasher(key.getBitsAt(0));
+    auto h1 = hasher(key.getBitsAt(1));
+    return h0 ^ (h1 << 1);
+}
+
+bool VulkanPipelineCache::PipelineLayoutKeyEqual::operator()(const PipelineLayoutKey& k1,
+        const PipelineLayoutKey& k2) const {
+    return k1 == k2;
 }

 bool VulkanPipelineCache::PipelineEqual::operator()(const PipelineKey& k1,
@@ -329,6 +895,31 @@ bool VulkanPipelineCache::PipelineEqual::operator()(const PipelineKey& k1,
    return 0 == memcmp((const void*) &k1, (const void*) &k2, sizeof(k1));
 }

+bool VulkanPipelineCache::DescEqual::operator()(const DescriptorKey& k1,
+        const DescriptorKey& k2) const {
+    for (uint32_t i = 0; i < UBUFFER_BINDING_COUNT; i++) {
+        if (k1.uniformBuffers[i] != k2.uniformBuffers[i] ||
+            k1.uniformBufferOffsets[i] != k2.uniformBufferOffsets[i] ||
+            k1.uniformBufferSizes[i] != k2.uniformBufferSizes[i]) {
+            return false;
+        }
+    }
+    for (uint32_t i = 0; i < SAMPLER_BINDING_COUNT; i++) {
+        if (k1.samplers[i].sampler != k2.samplers[i].sampler ||
+            k1.samplers[i].imageView != k2.samplers[i].imageView ||
+            k1.samplers[i].imageLayout != k2.samplers[i].imageLayout) {
+            return false;
+        }
+    }
+    for (uint32_t i = 0; i < INPUT_ATTACHMENT_COUNT; i++) {
+        if (k1.inputAttachments[i].imageView != k2.inputAttachments[i].imageView ||
+            k1.inputAttachments[i].imageLayout != k2.inputAttachments[i].imageLayout) {
+            return false;
+        }
+    }
+    return true;
+}
+
 } // namespace filament::backend

 #pragma clang diagnostic pop
@@ -17,11 +17,6 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H
 #define TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H

-#include "VulkanCommands.h"
-#include "VulkanMemory.h"
-#include "VulkanResources.h"
-#include "VulkanUtility.h"
-
 #include <backend/DriverEnums.h>
 #include <backend/TargetBufferInfo.h>

@@ -39,6 +34,12 @@
 #include <vector>
 #include <unordered_map>

+#include "VulkanCommands.h"
+
+VK_DEFINE_HANDLE(VmaAllocator)
+VK_DEFINE_HANDLE(VmaAllocation)
+VK_DEFINE_HANDLE(VmaPool)
+
 namespace filament::backend {

 struct VulkanProgram;
@@ -55,14 +56,32 @@ class VulkanResourceAllocator;
 // - Assumes that viewport and scissor should be dynamic. (not baked into VkPipeline)
 // - Assumes that uniform buffers should be visible across all shader stages.
 //
-class VulkanPipelineCache {
+class VulkanPipelineCache : public CommandBufferObserver {
 public:
    VulkanPipelineCache(VulkanPipelineCache const&) = delete;
    VulkanPipelineCache& operator=(VulkanPipelineCache const&) = delete;

+    static constexpr uint32_t UBUFFER_BINDING_COUNT = Program::UNIFORM_BINDING_COUNT;
+    static constexpr uint32_t SAMPLER_BINDING_COUNT = MAX_SAMPLER_COUNT;
+
+    // We assume only one possible input attachment between two subpasses. See also the subpasses
+    // definition in VulkanFboCache.
+    static constexpr uint32_t INPUT_ATTACHMENT_COUNT = 1;
+
    static constexpr uint32_t SHADER_MODULE_COUNT = 2;
    static constexpr uint32_t VERTEX_ATTRIBUTE_COUNT = MAX_VERTEX_ATTRIBUTE_COUNT;

+    // Three descriptor set layouts: uniforms, combined image samplers, and input attachments.
+    static constexpr uint32_t DESCRIPTOR_TYPE_COUNT = 3;
+    static constexpr uint32_t INITIAL_DESCRIPTOR_SET_POOL_SIZE = 512;
+
+    // The VertexArray POD is an array of buffer targets and an array of attributes that refer to
+    // those targets. It does not include any references to actual buffers, so you can think of it
+    // as a vertex assembler configuration. For simplicity it contains fixed-size arrays and does
+    // not store sizes; all unused entries are simply zeroed out.
+    struct VertexArray {
+    };
+
    // The ProgramBundle contains weak references to the compiled vertex and fragment shaders.
    struct ProgramBundle {
        VkShaderModule vertex;
@@ -70,6 +89,10 @@ public:
        VkSpecializationInfo* specializationInfos = nullptr;
    };

+    using UsageFlags = utils::bitset128;
+    static UsageFlags getUsageFlags(uint16_t binding, ShaderStageFlags stages, UsageFlags src = {});
+    static UsageFlags disableUsageFlags(uint16_t binding, UsageFlags src);
+
    #pragma clang diagnostic push
    #pragma clang diagnostic warning "-Wpadded"

@@ -112,13 +135,17 @@ public:

    // Upon construction, the pipeCache initializes some internal state but does not make any Vulkan
    // calls. On destruction it will free any cached Vulkan objects that haven't already been freed.
-    VulkanPipelineCache(VkDevice device, VmaAllocator allocator);
+    VulkanPipelineCache(VulkanResourceAllocator* allocator);
    ~VulkanPipelineCache();
+    void setDevice(VkDevice device, VmaAllocator allocator);

-    void bindLayout(VkPipelineLayout layout) noexcept;
+    // Creates new descriptor sets if necessary and binds them using vkCmdBindDescriptorSets.
+    // Returns false if descriptor set allocation fails.
+    bool bindDescriptors(VkCommandBuffer cmdbuffer) noexcept;

    // Creates a new pipeline if necessary and binds it using vkCmdBindPipeline.
-    void bindPipeline(VulkanCommandBuffer* commands);
+    // Returns false if an error occurred.
+    bool bindPipeline(VulkanCommandBuffer* commands) noexcept;

    // Sets up a new scissor rectangle if it has been dirtied.
    void bindScissor(VkCommandBuffer cmdbuffer, VkRect2D scissor) noexcept;
@@ -128,13 +155,42 @@ public:
    void bindRasterState(const RasterState& rasterState) noexcept;
    void bindRenderPass(VkRenderPass renderPass, int subpassIndex) noexcept;
    void bindPrimitiveTopology(VkPrimitiveTopology topology) noexcept;
-
+    void bindUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
+            VkDeviceSize offset = 0, VkDeviceSize size = VK_WHOLE_SIZE) noexcept;
+    void bindSamplers(VkDescriptorImageInfo samplers[SAMPLER_BINDING_COUNT],
+            VulkanTexture* textures[SAMPLER_BINDING_COUNT], UsageFlags flags) noexcept;
+    void bindInputAttachment(uint32_t bindingIndex, VkDescriptorImageInfo imageInfo) noexcept;
    void bindVertexArray(VkVertexInputAttributeDescription const* attribDesc,
            VkVertexInputBindingDescription const* bufferDesc, uint8_t count);

+    // Gets the current UBO at the given slot, useful for push / pop.
+    UniformBufferBinding getUniformBufferBinding(uint32_t bindingIndex) const noexcept;
+
+    // Checks if the given uniform is bound to any slot, and if so binds "null" to that slot.
+    // Also invalidates all cached descriptors that refer to the given buffer.
+    // This is only necessary when the client knows that the UBO is about to be destroyed.
+    void unbindUniformBuffer(VkBuffer uniformBuffer) noexcept;
+
+    // Checks if an image view is bound to any sampler, and if so resets that particular slot.
+    // Also invalidates all cached descriptors that refer to the given image view.
+    // This is only necessary when the client knows that a texture is about to be destroyed.
+    void unbindImageView(VkImageView imageView) noexcept;
+
+    // NOTE: In theory we should proffer "unbindSampler" but in practice we never destroy samplers.
+
    // Destroys all managed Vulkan objects. This should be called before changing the VkDevice.
    void terminate() noexcept;

+    // vkCmdBindPipeline and vkCmdBindDescriptorSets establish bindings to a specific command
+    // buffer; they are not global to the device. Therefore we need to be notified when a
+    // new command buffer becomes active.
+    void onCommandBuffer(const VulkanCommandBuffer& cmdbuffer) override;
+
+    // Injects a dummy texture that can be used to clear out old descriptor sets.
+    void setDummyTexture(VkImageView imageView) {
+        mDummyTargetInfo.imageView = imageView;
+    }
+
    static VkPrimitiveTopology getPrimitiveTopology(PrimitiveType pt) noexcept {
        switch (pt) {
            case PrimitiveType::POINTS:
@@ -150,9 +206,22 @@ public:
        }
    }

-    void gc() noexcept;
-
 private:
+    // PIPELINE LAYOUT CACHE KEY
+    // -------------------------
+
+    using PipelineLayoutKey = utils::bitset128;
+
+    static_assert(PipelineLayoutKey::BIT_COUNT >= 2 * MAX_SAMPLER_COUNT);
+
+    struct PipelineLayoutKeyHashFn {
+        size_t operator()(const PipelineLayoutKey& key) const;
+    };
+
+    struct PipelineLayoutKeyEqual {
+        bool operator()(const PipelineLayoutKey& k1, const PipelineLayoutKey& k2) const;
+    };
+
    // PIPELINE CACHE KEY
    // ------------------

@@ -205,10 +274,10 @@ private:
        VertexInputBindingDescription vertexBuffers[VERTEX_ATTRIBUTE_COUNT];      //  128 : 156
        RasterState rasterState;                                                  //  16  : 284
        uint32_t padding;                                                         //  4   : 300
-        VkPipelineLayout layout;                                                  //  8   : 304
+        PipelineLayoutKey layout;                                                 // 16   : 304
    };

-    static_assert(sizeof(PipelineKey) == 312, "PipelineKey must not have implicit padding.");
+    static_assert(sizeof(PipelineKey) == 320, "PipelineKey must not have implicit padding.");

    using PipelineHashFn = utils::hash::MurmurHashFn<PipelineKey>;

@@ -216,6 +285,52 @@ private:
        bool operator()(const PipelineKey& k1, const PipelineKey& k2) const;
    };

+    // DESCRIPTOR SET CACHE KEY
+    // ------------------------
+
+    // Equivalent to VkDescriptorImageInfo but with explicit padding.
+    struct DescriptorImageInfo {
+        DescriptorImageInfo& operator=(const VkDescriptorImageInfo& that) {
+            sampler = that.sampler;
+            imageView = that.imageView;
+            imageLayout = that.imageLayout;
+            padding = 0;
+            return *this;
+        }
+        operator VkDescriptorImageInfo() const { return { sampler, imageView, imageLayout }; }
+
+        // TODO: replace the 64-bit sampler handle with `uint32_t samplerParams` and remove the
+        // padding field. This is possible if we have access to the VulkanSamplerCache.
+        VkSampler sampler;
+
+        VkImageView imageView;
+        VkImageLayout imageLayout;
+        uint32_t padding;
+    };
+
+    // We store size with 32 bits, so our "WHOLE" sentinel is different from Vk.
+    static const uint32_t WHOLE_SIZE = 0xffffffffu;
+
+    // Represents all the Vulkan state that comprises a bound descriptor set.
+    struct DescriptorKey {
+        VkBuffer uniformBuffers[UBUFFER_BINDING_COUNT];               //   80     0
+        DescriptorImageInfo samplers[SAMPLER_BINDING_COUNT];          // 1488    80
+        DescriptorImageInfo inputAttachments[INPUT_ATTACHMENT_COUNT]; //   24  1568
+        uint32_t uniformBufferOffsets[UBUFFER_BINDING_COUNT];         //   40  1592
+        uint32_t uniformBufferSizes[UBUFFER_BINDING_COUNT];           //   40  1632
+    };
+    static_assert(offsetof(DescriptorKey, samplers)              == 80);
+    static_assert(offsetof(DescriptorKey, inputAttachments)      == 1568);
+    static_assert(offsetof(DescriptorKey, uniformBufferOffsets)  == 1592);
+    static_assert(offsetof(DescriptorKey, uniformBufferSizes)    == 1632);
+    static_assert(sizeof(DescriptorKey) == 1672, "DescriptorKey must not have implicit padding.");
+
+    using DescHashFn = utils::hash::MurmurHashFn<DescriptorKey>;
+
+    struct DescEqual {
+        bool operator()(const DescriptorKey& k1, const DescriptorKey& k2) const;
+    };
+
    #pragma clang diagnostic pop

    // CACHE ENTRY STRUCTS
@@ -228,6 +343,16 @@ private:
    using Timestamp = uint64_t;
    Timestamp mCurrentTime = 0;

+    // The descriptor set cache entry is a group of descriptor sets that are bound simultaneously.
+    struct DescriptorCacheEntry {
+        std::array<VkDescriptorSet, DESCRIPTOR_TYPE_COUNT> handles;
+        Timestamp lastUsed;
+        PipelineLayoutKey pipelineLayout;
+        uint32_t id;
+    };
+    uint32_t mDescriptorCacheEntryCount = 0;
+
+
    struct PipelineCacheEntry {
        VkPipeline handle;
        Timestamp lastUsed;
@@ -236,36 +361,98 @@ private:
    struct PipelineLayoutCacheEntry {
        VkPipelineLayout handle;
        Timestamp lastUsed;
+
+        std::array<VkDescriptorSetLayout, DESCRIPTOR_TYPE_COUNT> descriptorSetLayouts;
+
+        // Each pipeline layout has 3 arenas of unused descriptors (one for each binding type).
+        //
+        // The difference between the "arenas" and the "pool" are as follows.
+        //
+        // - The "pool" is a single, centralized factory for all descriptors (VkDescriptorPool).
+        //
+        // - Each "arena" is a set of unused (but alive) descriptors that can only be used with a
+        //   specific pipeline layout and a specific binding type. We manually manage each arena.
+        //   The arenas are created in an empty state, and they are gradually populated as new
+        //   descriptors are reclaimed over time.  This is quite different from the pool, which is
+        //   given a fixed size when it is constructed.
+        //
+        std::array<std::vector<VkDescriptorSet>, DESCRIPTOR_TYPE_COUNT> descriptorSetArenas;
    };

    // CACHE CONTAINERS
    // ----------------

+    using PipelineLayoutMap = tsl::robin_map<PipelineLayoutKey , PipelineLayoutCacheEntry,
+            PipelineLayoutKeyHashFn, PipelineLayoutKeyEqual>;
    using PipelineMap = tsl::robin_map<PipelineKey, PipelineCacheEntry,
            PipelineHashFn, PipelineEqual>;
+    using DescriptorMap
+            = tsl::robin_map<DescriptorKey, DescriptorCacheEntry, DescHashFn, DescEqual>;
+    using DescriptorResourceMap
+            = std::unordered_map<uint32_t, std::unique_ptr<VulkanAcquireOnlyResourceManager>>;

-private:
-
-    PipelineCacheEntry* getOrCreatePipeline() noexcept;
-
+    PipelineLayoutMap mPipelineLayouts;
    PipelineMap mPipelines;
+    DescriptorMap mDescriptorSets;
+    DescriptorResourceMap mDescriptorResources;

    // These helpers all return unstable pointers that should not be stored.
+    DescriptorCacheEntry* createDescriptorSets() noexcept;
    PipelineCacheEntry* createPipeline() noexcept;
    PipelineLayoutCacheEntry* getOrCreatePipelineLayout() noexcept;

+    // Misc helper methods.
+    void destroyLayoutsAndDescriptors() noexcept;
+    VkDescriptorPool createDescriptorPool(uint32_t size) const;
+    void growDescriptorPool() noexcept;
+
    // Immutable state.
    VkDevice mDevice = VK_NULL_HANDLE;
    VmaAllocator mAllocator = VK_NULL_HANDLE;

    // Current requirements for the pipeline layout, pipeline, and descriptor sets.
    PipelineKey mPipelineRequirements = {};
+    DescriptorKey mDescriptorRequirements = {};

    // Current bindings for the pipeline and descriptor sets.
    PipelineKey mBoundPipeline = {};
+    DescriptorKey mBoundDescriptor = {};

    // Current state for scissoring.
    VkRect2D mCurrentScissor = {};
+
+    // The descriptor set pool starts out with a decent number of descriptor sets.  The cache can
+    // grow the pool by re-creating it with a larger size.  See growDescriptorPool().
+    VkDescriptorPool mDescriptorPool;
+
+    // This describes the number of descriptor sets in mDescriptorPool. Note that this needs to be
+    // multiplied by DESCRIPTOR_TYPE_COUNT to get the actual number of descriptor sets. Also note
+    // that the number of low-level "descriptors" (not descriptor *sets*) is actually much more than
+    // this size. It can be computed only by factoring in UBUFFER_BINDING_COUNT etc.
+    uint32_t mDescriptorPoolSize = INITIAL_DESCRIPTOR_SET_POOL_SIZE;
+
+    // To get the actual number of descriptor sets that have been allocated from the pool,
+    // take the sum of mDescriptorArenasCount (these are inactive descriptor sets) and the
+    // number of entries in the mDescriptorPool map (active descriptor sets). Multiply the result by
+    // DESCRIPTOR_TYPE_COUNT.
+    uint32_t mDescriptorArenasCount = 0;
+
+    // After a growth event (i.e. when the VkDescriptorPool is replaced with a bigger version), all
+    // currently used descriptors are moved into the "extinct" sets so that they can be safely
+    // destroyed a few frames later.
+    std::list<VkDescriptorPool> mExtinctDescriptorPools;
+    std::list<DescriptorCacheEntry> mExtinctDescriptorBundles;
+
+    VkDescriptorBufferInfo mDummyBufferInfo = {};
+    VkWriteDescriptorSet mDummyBufferWriteInfo = {};
+    VkDescriptorImageInfo mDummyTargetInfo = {};
+    VkWriteDescriptorSet mDummyTargetWriteInfo = {};
+
+    VkBuffer mDummyBuffer;
+    VmaAllocation mDummyMemory;
+
+    VulkanResourceAllocator* mResourceAllocator;
+    VulkanAcquireOnlyResourceManager mPipelineBoundResources;
 };

 } // namespace filament::backend
@@ -28,6 +28,7 @@ using namespace bluevk;

 namespace filament::backend {

+using ImgUtil = VulkanImageUtility;
 using TaskHandler = VulkanReadPixels::TaskHandler;
 using WorkloadFunc = TaskHandler::WorkloadFunc;
 using OnCompleteFunc = TaskHandler::OnCompleteFunc;
@@ -117,7 +118,7 @@ void VulkanReadPixels::terminate() noexcept {
 VulkanReadPixels::VulkanReadPixels(VkDevice device)
    : mDevice(device) {}

-void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint32_t const y,
+void VulkanReadPixels::run(VulkanRenderTarget const* srcTarget, uint32_t const x, uint32_t const y,
        uint32_t const width, uint32_t const height, uint32_t const graphicsQueueFamilyIndex,
        PixelBufferDescriptor&& pbd, SelecteMemoryFunction const& selectMemoryFunc,
        OnReadCompleteFunction const& readCompleteFunc) {
@@ -217,7 +218,7 @@ void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint
    };
    vkBeginCommandBuffer(cmdbuffer, &binfo);

-    imgutil::transitionLayout(cmdbuffer, {
+    ImgUtil::transitionLayout(cmdbuffer, {
        .image = stagingImage,
        .oldLayout = VulkanLayout::UNDEFINED,
        .newLayout = VulkanLayout::TRANSFER_DST,
@@ -231,7 +232,8 @@ void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint
    });

    VulkanAttachment const srcAttachment = srcTarget->getColor(0);
-    VkImageSubresourceRange const srcRange = srcAttachment.getSubresourceRange();
+    const VkImageSubresourceRange srcRange
+            = srcAttachment.getSubresourceRange(VK_IMAGE_ASPECT_COLOR_BIT);
    srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);

    VkImageCopy const imageCopyRegion = {
@@ -264,8 +266,8 @@ void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint
            imageCopyRegion.srcOffset.y + imageCopyRegion.extent.height <= srcExtent.height);

    vkCmdCopyImage(cmdbuffer, srcAttachment.getImage(),
-            imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC), stagingImage,
-            imgutil::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);
+            ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC), stagingImage,
+            ImgUtil::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);

    // Restore the source image layout.
    srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::COLOR_ATTACHMENT);
@@ -26,7 +26,9 @@
 #include <functional>
 #include <mutex>
 #include <queue>
+#include <set>
 #include <thread>
+#include <vector>

 namespace filament::backend {

@@ -72,8 +74,8 @@ public:

    void terminate() noexcept;

-    void run(VulkanRenderTarget* srcTarget, uint32_t x, uint32_t y, uint32_t width, uint32_t height,
-            uint32_t graphicsQueueFamilyIndex, PixelBufferDescriptor&& pbd,
+    void run(VulkanRenderTarget const* srcTarget, uint32_t x, uint32_t y, uint32_t width,
+            uint32_t height, uint32_t graphicsQueueFamilyIndex, PixelBufferDescriptor&& pbd,
            SelecteMemoryFunction const& selectMemoryFunc,
            OnReadCompleteFunction const& readCompleteFunc);

@@ -49,8 +49,8 @@ namespace filament::backend {
 #endif

 class VulkanResourceAllocator {
+
 public:
-    using AllocatorImpl = HandleAllocatorVK;
    VulkanResourceAllocator(size_t arenaSize, bool disableUseAfterFreeCheck)
        : mHandleAllocatorImpl("Handles", arenaSize, disableUseAfterFreeCheck)
 #if DEBUG_RESOURCE_LEAKS
@@ -106,7 +106,7 @@ public:
    }

 private:
-    AllocatorImpl mHandleAllocatorImpl;
+    HandleAllocatorVK mHandleAllocatorImpl;

 #if DEBUG_RESOURCE_LEAKS
 public:
@@ -17,7 +17,7 @@
 #include "VulkanResources.h"
 #include "VulkanHandles.h"
 #include "VulkanResourceAllocator.h"
-#include "VulkanPipelineCache.h"
+#include "caching/VulkanDescriptorSet.h"

 namespace filament::backend {

@@ -62,9 +62,6 @@ void deallocateResource(VulkanResourceAllocator* allocator, VulkanResourceType t
        case VulkanResourceType::RENDER_PRIMITIVE:
            allocator->destruct<VulkanRenderPrimitive>(Handle<VulkanRenderPrimitive>(id));
            break;
-        case VulkanResourceType::DESCRIPTOR_SET_LAYOUT:
-            allocator->destruct<VulkanDescriptorSetLayout>(Handle<VulkanDescriptorSetLayout>(id));
-            break;
        case VulkanResourceType::DESCRIPTOR_SET:
            allocator->destruct<VulkanDescriptorSet>(Handle<VulkanDescriptorSet>(id));
            break;
@@ -35,25 +35,24 @@ struct VulkanThreadSafeResource;

 // Subclasses of VulkanResource must provide this enum in their construction.
 enum class VulkanResourceType : uint8_t {
-    BUFFER_OBJECT = 0,
-    INDEX_BUFFER = 1,
-    PROGRAM = 2,
-    RENDER_TARGET = 3,
-    SAMPLER_GROUP = 4,
-    SWAP_CHAIN = 5,
-    RENDER_PRIMITIVE = 6,
-    TEXTURE = 7,
-    TIMER_QUERY = 8,
-    VERTEX_BUFFER = 9,
-    VERTEX_BUFFER_INFO = 10,
-    DESCRIPTOR_SET_LAYOUT = 11,
-    DESCRIPTOR_SET = 12,
+    BUFFER_OBJECT,
+    INDEX_BUFFER,
+    PROGRAM,
+    RENDER_TARGET,
+    SAMPLER_GROUP,
+    SWAP_CHAIN,
+    RENDER_PRIMITIVE,
+    TEXTURE,
+    TIMER_QUERY,
+    VERTEX_BUFFER,
+    VERTEX_BUFFER_INFO,
+    DESCRIPTOR_SET,

    // Below are resources that are managed manually (i.e. not ref counted).
-    FENCE = 13,
-    HEAP_ALLOCATED = 14,
+    FENCE,
+    HEAP_ALLOCATED,

-    END_TYPE = 15,  // A placeholder
+    END_TYPE,  // A placeholder
 };

 #define IS_HEAP_ALLOC_TYPE(f)                                                                      \
@@ -17,7 +17,6 @@
 #include "VulkanStagePool.h"

 #include "VulkanConstants.h"
-#include "VulkanImageUtility.h"
 #include "VulkanMemory.h"
 #include "VulkanUtility.h"

@@ -119,7 +118,7 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
    // VK_IMAGE_LAYOUT_PREINITIALIZED or VK_IMAGE_LAYOUT_GENERAL layout. Calling
    // vkGetImageSubresourceLayout for a linear image returns a subresource layout mapping that is
    // valid for either of those image layouts."
-    imgutil::transitionLayout(cmdbuffer, {
+    VulkanImageUtility::transitionLayout(cmdbuffer, {
            .image = image->image,
            .oldLayout = VulkanLayout::UNDEFINED,
            .newLayout = VulkanLayout::READ_WRITE, // (= VK_IMAGE_LAYOUT_GENERAL)
@@ -17,9 +17,7 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANSTAGEPOOL_H
 #define TNT_FILAMENT_BACKEND_VULKANSTAGEPOOL_H

-#include "VulkanCommands.h"
 #include "VulkanContext.h"
-#include "VulkanMemory.h"

 #include <map>
 #include <unordered_set>
@@ -19,7 +19,6 @@

 #include "DriverBase.h"

-#include "VulkanCommands.h"
 #include "VulkanContext.h"
 #include "VulkanResources.h"

@@ -28,6 +28,7 @@ using namespace bluevk;

 namespace filament::backend {

+using ImgUtil = VulkanImageUtility;
 VulkanTexture::VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
        VkImage image, VkFormat format, uint8_t samples, uint32_t width, uint32_t height,
        TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
@@ -36,7 +37,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanComm
      VulkanResource(
              heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
      mVkFormat(format),
-      mViewType(imgutil::getViewType(target)),
+      mViewType(ImgUtil::getViewType(target)),
      mSwizzle({}),
      mTextureImage(image),
      mFullViewRange{
@@ -61,7 +62,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
      VulkanResource(
              heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
      mVkFormat(backend::getVkFormat(tformat)),
-      mViewType(imgutil::getViewType(target)),
+      mViewType(ImgUtil::getViewType(target)),
      mSwizzle(swizzle),
      mStagePool(stagePool),
      mDevice(device),
@@ -235,7 +236,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
        VulkanCommandBuffer& commands = mCommands->get();
        VkCommandBuffer const cmdbuf = commands.buffer();
        commands.acquire(this);
-        transitionLayout(cmdbuf, mFullViewRange, imgutil::getDefaultLayout(imageInfo.usage));
+        transitionLayout(cmdbuf, mFullViewRange, ImgUtil::getDefaultLayout(imageInfo.usage));
    }
 }

@@ -326,10 +327,10 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt

    VulkanLayout const newLayout = VulkanLayout::TRANSFER_DST;
    VulkanLayout nextLayout = getLayout(transitionRange.baseArrayLayer, miplevel);
-    VkImageLayout const newVkLayout = imgutil::getVkLayout(newLayout);
+    VkImageLayout const newVkLayout = ImgUtil::getVkLayout(newLayout);

    if (nextLayout == VulkanLayout::UNDEFINED) {
-        nextLayout = imgutil::getDefaultLayout(this->usage);
+        nextLayout = ImgUtil::getDefaultLayout(this->usage);
    }

    transitionLayout(cmdbuf, transitionRange, newLayout);
@@ -373,8 +374,8 @@ void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, u
    VulkanLayout const oldLayout = getLayout(layer, miplevel);
    transitionLayout(cmdbuf, range, newLayout);

-    vkCmdBlitImage(cmdbuf, stage->image, imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            mTextureImage, imgutil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
+    vkCmdBlitImage(cmdbuf, stage->image, ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC),
+            mTextureImage, ImgUtil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);

    transitionLayout(cmdbuf, range, oldLayout);
 }
@@ -466,7 +467,7 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
        for (uint32_t i = firstLayer; i < lastLayer; ++i) {
            for (uint32_t j = firstLevel; j < lastLevel; ++j) {
                VulkanLayout const layout = getLayout(i, j);
-                imgutil::transitionLayout(cmdbuf, {
+                ImgUtil::transitionLayout(cmdbuf, {
                        .image = mTextureImage,
                        .oldLayout = layout,
                        .newLayout = newLayout,
@@ -481,7 +482,7 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
            }
        }
    } else {
-        imgutil::transitionLayout(cmdbuf, {
+        ImgUtil::transitionLayout(cmdbuf, {
            .image = mTextureImage,
            .oldLayout = oldLayout,
            .newLayout = newLayout,
@@ -23,7 +23,6 @@
 #include "VulkanResources.h"
 #include "VulkanImageUtility.h"

-#include <utils/Hash.h>
 #include <utils/RangeMap.h>

 #include <unordered_map>
@@ -93,318 +93,6 @@ utils::FixedCapacityVector<OutType> enumerate(
 // Useful shorthands
 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,
-};
-
 // An Array that will be fixed capacity, but the "size" (as in user added elements) is variable.
 // Note that this class is movable.
 template<typename T, uint16_t CAPACITY>
@@ -434,14 +122,14 @@ public:
        clear();
    }

-    inline const_iterator begin() const {
+    inline const_iterator begin() {
        if (mInd == 0) {
            return mArray.cend();
        }
        return mArray.cbegin();
    }

-    inline const_iterator end() const {
+    inline const_iterator end() {
        if (mInd > 0 && mInd < CAPACITY) {
            return mArray.begin() + mInd;
        }
@@ -482,26 +170,10 @@ public:
        return mArray[ind];
    }

-    inline T const& operator[](uint16_t ind) const {
-        return mArray[ind];
-    }
-
-    inline uint32_t size() const {
+    inline size_t size() {
        return mInd;
    }

-    T* data() {
-        return mArray.data();
-    }
-
-    T const* data() const {
-        return mArray.data();
-    }
-
-    bool operator==(CappedArray const& b) const {
-        return this->mArray == b.mArray;
-    }
-
 private:
    void swap(CappedArray& rhs) {
        std::swap(mArray, rhs.mArray);
@@ -509,129 +181,9 @@ private:
    }

    FixedSizeArray mArray;
-    uint32_t mInd = 0;
+    size_t mInd = 0;
 };

-// TODO: ok to remove once Filament-side API is complete
-namespace descset {
-
-// Used to describe the descriptor binding in shader stages. We assume that the binding index does
-// not exceed 31. We also assume that we have two shader stages - vertex and fragment.  The below
-// types and struct are used across VulkanDescriptorSet and VulkanProgram.
-using UniformBufferBitmask = uint32_t;
-using SamplerBitmask = uint64_t;
-
-// We only have at most one input attachment, so this bitmask exists only to make the code more
-// general.
-using InputAttachmentBitmask = uint8_t;
-
-constexpr UniformBufferBitmask UBO_VERTEX_STAGE = 0x1;
-constexpr UniformBufferBitmask UBO_FRAGMENT_STAGE = (0x1ULL << (sizeof(UniformBufferBitmask) * 4));
-constexpr SamplerBitmask SAMPLER_VERTEX_STAGE = 0x1;
-constexpr SamplerBitmask SAMPLER_FRAGMENT_STAGE = (0x1ULL << (sizeof(SamplerBitmask) * 4));
-constexpr InputAttachmentBitmask INPUT_ATTACHMENT_VERTEX_STAGE = 0x1;
-constexpr InputAttachmentBitmask INPUT_ATTACHMENT_FRAGMENT_STAGE =
-        (0x1ULL << (sizeof(InputAttachmentBitmask) * 4));
-
-template<typename Bitmask>
-static constexpr Bitmask getVertexStage() noexcept {
-    if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
-        return UBO_VERTEX_STAGE;
-    }
-    if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
-        return SAMPLER_VERTEX_STAGE;
-    }
-    if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
-        return INPUT_ATTACHMENT_VERTEX_STAGE;
-    }
-}
-
-template<typename Bitmask>
-static constexpr Bitmask getFragmentStage() noexcept {
-    if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
-        return UBO_FRAGMENT_STAGE;
-    }
-    if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
-        return SAMPLER_FRAGMENT_STAGE;
-    }
-    if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
-        return INPUT_ATTACHMENT_FRAGMENT_STAGE;
-    }
-}
-
-typedef enum ShaderStageFlags2 : uint8_t {
-    NONE        =    0,
-    VERTEX      =    0x1,
-    FRAGMENT    =    0x2,
-} ShaderStageFlags2;
-
-enum class DescriptorType : uint8_t {
-    UNIFORM_BUFFER,
-    SAMPLER,
-    INPUT_ATTACHMENT,
-};
-
-enum class DescriptorFlags : uint8_t {
-    NONE = 0x00,
-    DYNAMIC_OFFSET = 0x01
-};
-
-struct DescriptorSetLayoutBinding {
-    DescriptorType type;
-    ShaderStageFlags2 stageFlags;
-    uint8_t binding;
-    DescriptorFlags flags;
-    uint16_t count;
-};
-
-struct DescriptorSetLayout {
-    utils::FixedCapacityVector<DescriptorSetLayoutBinding> bindings;
-};
-
-} // namespace descset
-
-namespace {
-// Use constexpr to statically generate a bit count table for 8-bit numbers.
-struct _BitCountHelper {
-    constexpr _BitCountHelper() : data{} {
-        for (uint16_t i = 0; i < 256; ++i) {
-            data[i] = 0;
-            for (auto j = i; j > 0; j /= 2) {
-                if (j & 1) {
-                    data[i]++;
-                }
-            }
-        }
-    }
-
-    template<typename MaskType>
-    constexpr uint8_t count(MaskType num) {
-        uint8_t count = 0;
-        for (uint8_t i = 0; i < sizeof(MaskType) * 8; i+=8) {
-            count += data[(num >> i) & 0xFF];
-        }
-        return count;
-    }
-
-private:
-    uint8_t data[256];
-};
-} // namespace anonymous
-
-template<typename MaskType>
-inline uint8_t countBits(MaskType num) {
-    static _BitCountHelper BitCounter = {};
-    return BitCounter.count(num);
-}
-
-// This is useful for counting the total number of descriptors for both vertex and fragment stages.
-template<typename MaskType>
-inline MaskType collapseStages(MaskType mask) {
-    constexpr uint8_t NBITS_DIV_2 = sizeof(MaskType) * 4;
-    // First zero out the top-half and then or the bottom-half against the original top-half.
-    return ((mask << NBITS_DIV_2) >> NBITS_DIV_2) | (mask >> NBITS_DIV_2);
-}
-
 } // namespace filament::backend

 #endif // TNT_FILAMENT_BACKEND_VULKANUTILITY_H
@@ -0,0 +1,647 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "VulkanDescriptorSet.h"
+
+#include <vulkan/VulkanConstants.h>
+#include <vulkan/VulkanResources.h>
+
+#include <type_traits>
+
+namespace filament::backend {
+
+namespace {
+
+template<typename MaskType>
+using LayoutMap = tsl::robin_map<MaskType, VkDescriptorSetLayout>;
+using LayoutArray = VulkanDescriptorSetManager::LayoutArray;
+using UniformBufferBitmask = VulkanDescriptorSetManager::UniformBufferBitmask;
+using SamplerBitmask = VulkanDescriptorSetManager::SamplerBitmask;
+
+constexpr uint8_t EXPECTED_IN_FLIGHT_FRAMES = 3; // Asssume triple buffering
+
+#define PORT_CONSTANT(NameSpace, K) constexpr decltype(NameSpace::K) K = NameSpace::K;
+
+PORT_CONSTANT(VulkanDescriptorSetManager, MAX_SUPPORTED_SHADER_STAGE)
+PORT_CONSTANT(VulkanDescriptorSetManager, VERTEX_STAGE)
+PORT_CONSTANT(VulkanDescriptorSetManager, FRAGMENT_STAGE)
+
+PORT_CONSTANT(Program, UNIFORM_BINDING_COUNT)
+PORT_CONSTANT(Program, SAMPLER_BINDING_COUNT)
+
+#undef PORT_CONSTANT
+
+// Use constexpr to statically generate a bit count table for 8-bit numbers.
+struct BitCountHelper {
+    constexpr BitCountHelper() : data{} {
+        for (uint16_t i = 0; i < 256; ++i) {
+            data[i] = 0;
+            for (auto j = i; j > 0; j /= 2) {
+                if (j & 1) {
+                    data[i]++;
+                }
+            }
+        }
+    }
+    uint8_t data[256];
+} BitCountImpl;
+
+inline uint8_t countBits(uint32_t num) {
+    static constexpr uint8_t const* BIT_COUNT = BitCountImpl.data;
+    return BIT_COUNT[num & 0xFF] + BIT_COUNT[(num >> 8) & 0xFF] + BIT_COUNT[(num >> 16) & 0xFF] +
+           BIT_COUNT[(num >> 24) & 0xFF];
+}
+
+struct SamplerSet {
+    using Mask = UniformBufferBitmask;
+    struct Key {
+        Mask mask;
+        static_assert(sizeof(Mask) == 4);
+        uint32_t padding;// We need a padding to ensure 8-byte alignement.
+        VkDescriptorSetLayout layout;
+        VkSampler sampler[SAMPLER_BINDING_COUNT];
+        VkImageView imageView[SAMPLER_BINDING_COUNT];
+        VkImageLayout imageLayout[SAMPLER_BINDING_COUNT];
+    };
+
+    struct Equal {
+        bool operator()(Key const& k1, Key const& k2) const {
+            if (k1.mask != k2.mask) return false;
+            if (k1.layout != k2.layout) return false;
+
+            for (uint8_t i = 0, bitCount = countBits(k1.mask); i < bitCount; ++i) {
+                if (k1.sampler[i] != k2.sampler[i] || k1.imageView[i] != k2.imageView[i] ||
+                        k1.imageLayout[i] != k2.imageLayout[i]) {
+                    return false;
+                }
+            }
+            return true;
+        }
+    };
+    using HashFn = utils::hash::MurmurHashFn<Key>;
+    using Cache = std::unordered_map<Key, VulkanDescriptorSet*, HashFn, Equal>;
+};
+
+struct UBOSet {
+    using Mask = SamplerBitmask;
+
+    struct Key {
+        Mask mask;
+        static_assert(sizeof(Mask) == 4);
+        uint32_t padding;// We need a padding to ensure 8-byte alignement.
+        VkDescriptorSetLayout layout;
+        VkBuffer buffers[UNIFORM_BINDING_COUNT];    //   80     0
+        VkDeviceSize offsets[UNIFORM_BINDING_COUNT];//   40  1592
+        VkDeviceSize sizes[UNIFORM_BINDING_COUNT];  //   40  1632
+    };
+
+    struct Equal {
+        bool operator()(Key const& k1, Key const& k2) const {
+            if (k1.mask != k2.mask) return false;
+            if (k1.layout != k2.layout) return false;
+
+            for (uint8_t i = 0, bitCount = countBits(k1.mask); i < bitCount; ++i) {
+                if (k1.buffers[i] != k2.buffers[i] || k1.offsets[i] != k2.offsets[i] ||
+                        k1.sizes[i] != k2.sizes[i]) {
+                    return false;
+                }
+            }
+            return true;
+        }
+    };
+    using HashFn = utils::hash::MurmurHashFn<Key>;
+    using Cache = std::unordered_map<Key, VulkanDescriptorSet*, HashFn, Equal>;
+};
+
+constexpr uint8_t MAX_BINDINGS = UNIFORM_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE;
+static_assert(MAX_BINDINGS >= SAMPLER_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE);
+
+template<typename TYPE>
+class DescriptorPool {
+private:
+    static constexpr uint32_t COUNT_MULTIPLIER =
+            MAX_SUPPORTED_SHADER_STAGE * EXPECTED_IN_FLIGHT_FRAMES;
+    static constexpr uint32_t CAPACITY =
+            std::is_same_v<TYPE, UBOSet> ? Program::UNIFORM_BINDING_COUNT * COUNT_MULTIPLIER
+                                         : Program::SAMPLER_BINDING_COUNT * COUNT_MULTIPLIER;
+
+public:
+    DescriptorPool()
+        : mDevice(VK_NULL_HANDLE),
+          mResourceAllocator(nullptr) {}
+
+    DescriptorPool(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mResourceAllocator(allocator) {
+        VkDescriptorPoolSize sizes[1] = {
+            {
+                .type = std::is_same_v<TYPE, UBOSet> ? VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
+                                                     : VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                .descriptorCount = CAPACITY,
+            },
+        };
+        VkDescriptorPoolCreateInfo info{
+            .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
+            .maxSets = EXPECTED_IN_FLIGHT_FRAMES,
+            .poolSizeCount = 1,
+            .pPoolSizes = sizes,
+        };
+        vkCreateDescriptorPool(mDevice, &info, VKALLOC, &mPool);
+    }
+
+    DescriptorPool(DescriptorPool const&) = delete;
+    DescriptorPool& operator=(DescriptorPool const&) = delete;
+
+    ~DescriptorPool() {
+        vkDestroyDescriptorPool(mDevice, mPool, VKALLOC);
+    }
+
+    VulkanDescriptorSet* obtainSet(VkDescriptorSetLayout layout) {
+        if (mCount == CAPACITY) {
+            return nullptr;
+        }
+
+        if (auto unused = mUnused.find(layout);
+                unused != mUnused.end() && !unused->second.empty()) {
+            auto& unusedList = unused->second;
+            auto set = unusedList.back();
+            unusedList.pop_back();
+            mCount++;
+            return set;
+        }
+
+        // Creating a new set
+        VkDescriptorSetLayout layouts[1] = {layout};
+        VkDescriptorSetAllocateInfo allocInfo = {
+            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
+            .descriptorPool = mPool,
+            .descriptorSetCount = 1,
+            .pSetLayouts = layouts,
+        };
+        VkDescriptorSet vkSet;
+        UTILS_UNUSED VkResult result = vkAllocateDescriptorSets(mDevice, &allocInfo, &vkSet);
+        ASSERT_POSTCONDITION(result == VK_SUCCESS, "Cannot allocate descriptor set. error=%d",
+                (int) result);
+
+        mCount++;
+        return createSet(layout, vkSet);
+    }
+
+private:
+    inline VulkanDescriptorSet* createSet(VkDescriptorSetLayout layout, VkDescriptorSet vkSet) {
+        return VulkanDescriptorSet::create(mResourceAllocator, vkSet, layout,
+                [this](VulkanDescriptorSet* set) {
+                    auto const layout = set->layout;
+                    auto const vkSet = set->vkSet;
+                    auto listIter = mUnused.find(layout);
+                    assert_invariant(listIter != mUnused.end());
+                    auto& unusedList = listIter->second;
+
+                    // At this point, we know that the pointer will be stale, and we need to remove
+                    // any non-ref-counted references to it.
+                    // TODO: this will be inefficient
+                    auto iter = std::find(unusedList.begin(), unusedList.end(), set);
+                    unusedList.erase(iter);
+
+                    // We are recycling - release the vk resource back into the pool. Note that the
+                    // vk handle has not changed, but we need to change the backend handle to allow
+                    // for proper refcounting of resources referenced in this set.
+                    unusedList.push_back(createSet(layout, vkSet));
+                    mCount--;
+                });
+    }
+
+    VkDevice mDevice;
+    VulkanResourceAllocator* mResourceAllocator;
+    VkDescriptorPool mPool;
+
+    uint32_t mCount = 0;
+    std::unordered_map<VkDescriptorSetLayout, std::vector<VulkanDescriptorSet*>> mUnused;
+};
+
+template<typename TYPE>
+class DescriptorInfinitePool {
+public:
+    DescriptorInfinitePool(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mResourceAllocator(allocator) {}
+
+    ~DescriptorInfinitePool() {
+        for (auto pool: mPools) {
+            delete pool;
+        }
+    }
+
+    VulkanDescriptorSet* obtainSet(VkDescriptorSetLayout layout) {
+        for (auto pool: mPools) {
+            auto set = pool->obtainSet(layout);
+            if (set) {
+                return set;
+            }
+        }
+        // We need to increase the number of pools
+        mPools.push_back(new DescriptorPool<TYPE>{mDevice, mResourceAllocator});
+        auto pool = mPools.back();
+        return pool->obtainSet(layout);
+    }
+
+private:
+    VkDevice mDevice;
+    VulkanResourceAllocator* mResourceAllocator;
+    std::vector<DescriptorPool<TYPE>*> mPools;
+};
+
+// This holds a cache of descriptor sets of a TYPE (UBO or Sampler). The Key is defined as the
+// layout and the content (for example specific samplers).
+template<typename TYPE, typename Cache, typename Key>
+class DescriptorSetCache {
+public:
+    DescriptorSetCache(VkDevice device, VulkanResourceAllocator* allocator)
+        : mPool(device, allocator),
+          mResourceManager(allocator) {}
+
+    std::pair<VulkanDescriptorSet*, bool> obtainSet(Key const& key) {
+        mAge++;
+        if (auto result = mCache.find(key); result != mCache.end()) {
+            auto set = result->second;
+            auto const oldAge = mReverseHistory[set];
+            mHistory.erase(oldAge);
+            mHistory[mAge] = set;
+            mReverseHistory[set] = mAge;
+            return {set, true};
+        }
+
+        VulkanDescriptorSet* set = mPool.obtainSet(key.layout);
+        mCache[key] = set;
+        mReverseCache[set] = key;
+        mHistory[mAge] = set;
+        mReverseHistory[set] = mAge;
+        mResourceManager.acquire(set);
+        return {set, false};
+    }
+
+    // We need to periodically purge the descriptor sets so that we're not holding on to resources
+    // unnecessarily. The strategy for purging needs to be examined more.
+    void gc() noexcept {
+        constexpr uint32_t ALLOWED_ENTRIES = EXPECTED_IN_FLIGHT_FRAMES * 3;
+        int32_t const toCutCount = mHistory.size() - ALLOWED_ENTRIES;
+
+        if (toCutCount <= 0) {
+            return;
+        }
+
+        std::vector<uint64_t> remove;
+        for (auto entry = mHistory.begin(); entry != mHistory.end(); entry++) {
+            auto const& set = entry->second;
+            Key const& key = mReverseCache[set];
+            mCache.erase(key);
+            mReverseCache.erase(set);
+            mReverseHistory.erase(set);
+
+            remove.push_back(entry->first);
+            mResourceManager.release(set);
+        }
+        for (auto removeAge: remove) {
+            mHistory.erase(removeAge);
+        }
+    }
+
+private:
+    DescriptorInfinitePool<TYPE> mPool;
+
+    // TODO: combine some of these data structures
+    Cache mCache;
+    std::unordered_map<VulkanDescriptorSet*, Key> mReverseCache;
+
+    // Use the ordering for purging if needed;
+    std::map<uint64_t, VulkanDescriptorSet*> mHistory;
+    std::unordered_map<VulkanDescriptorSet*, uint64_t> mReverseHistory;
+    uint64_t mAge;
+
+    // Note that instead of owning the resources (i.e. descriptor set) in the pools, we keep them
+    // here since all the allocated sets have to pass through this class, and this class has better
+    // knowledge to make decisions about gc.
+    VulkanResourceManager mResourceManager;
+};
+
+using UBOSetCache = DescriptorSetCache<UBOSet, UBOSet::Cache, UBOSet::Key>;
+using SamplerSetCache = DescriptorSetCache<SamplerSet, SamplerSet::Cache, SamplerSet::Key>;
+
+template<typename MaskType>
+class LayoutCache {
+public:
+    explicit LayoutCache(VkDevice device)
+        : mDevice(device) {}
+
+    ~LayoutCache() {
+        for (auto [mask, layout]: mLayouts) {
+            vkDestroyDescriptorSetLayout(mDevice, layout, VKALLOC);
+        }
+        mLayouts.clear();
+    }
+
+    VkDescriptorSetLayout getLayout(MaskType stageMask) noexcept {
+        if (stageMask == 0) {
+            return VK_NULL_HANDLE;
+        }
+        if (auto iter = mLayouts.find(stageMask); iter != mLayouts.end()) {
+            return iter->second;
+        }
+        VkDescriptorType descriptorType;
+        if constexpr (std::is_same_v<MaskType, UBOSet::Mask>) {
+            descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+        } else if constexpr (std::is_same_v<MaskType, SamplerSet::Mask>) {
+            descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+        }
+
+        VkDescriptorSetLayoutBinding toBind[MAX_BINDINGS];
+        uint32_t count = 0;
+        for (uint8_t i = 0, maxBindings = sizeof(stageMask) * 4; i < maxBindings; i++) {
+            VkShaderStageFlags stages = 0;
+            if ((stageMask & (VERTEX_STAGE << i)) != 0) {
+                stages |= VK_SHADER_STAGE_VERTEX_BIT;
+            }
+            if ((stageMask & (FRAGMENT_STAGE << i)) == 0) {
+                stages |= VK_SHADER_STAGE_FRAGMENT_BIT;
+            }
+            if (stages == 0) {
+                continue;
+            }
+
+            auto& bindInfo = toBind[count++];
+            bindInfo = {
+                .binding = i,
+                .descriptorType = descriptorType,
+                .descriptorCount = 1,
+                .stageFlags = stages,
+            };
+        }
+        VkDescriptorSetLayoutCreateInfo dlinfo = {
+            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+            .bindingCount = count,
+            .pBindings = toBind,
+        };
+
+        VkDescriptorSetLayout handle;
+        vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &handle);
+        mLayouts[stageMask] = handle;
+
+        return handle;
+    }
+
+private:
+    VkDevice mDevice;
+    LayoutMap<MaskType> mLayouts;
+};
+
+using UBOSetLayoutCache = LayoutCache<UBOSet::Mask>;
+using SamplerSetLayoutCache = LayoutCache<SamplerSet::Mask>;
+
+} // anonymous namespace
+
+VulkanDescriptorSet::VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+        VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn)
+    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
+      resources(allocator),
+      vkSet(rawSet),
+      layout(layout),
+      mOnRecycleFn(std::move(onRecycleFn)) {}
+
+VulkanDescriptorSet::~VulkanDescriptorSet() {
+    if (mOnRecycleFn) {
+        mOnRecycleFn(this);
+    }
+}
+
+VulkanDescriptorSet* VulkanDescriptorSet::create(VulkanResourceAllocator* allocator,
+        VkDescriptorSet rawSet, VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn) {
+    auto handle = allocator->allocHandle<VulkanDescriptorSet>();
+    auto set = allocator->construct<VulkanDescriptorSet>(handle, allocator, rawSet, layout,
+            std::move(onRecycleFn));
+    return set;
+}
+
+class VulkanDescriptorSetManager::Impl {
+public:
+    Impl(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mUBOLayoutCache(device),
+          mUBOCache(device, allocator),
+          mSamplerLayoutCache(device),
+          mSamplerCache(device, allocator),
+          mResources(allocator) {}
+
+    void setUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
+            VkDeviceSize offset, VkDeviceSize size) noexcept {
+        mUbos.push_back({bindingIndex, bufferObject, offset, size});
+
+        // Between "set" and "commit", we need to ref the buffer object to avoid it being collected.
+        mResources.acquire(bufferObject);
+    }
+
+    void setSamplers(SamplerArray&& samplers) {
+        mSamplers = std::move(samplers);
+        for (auto const& sampler: mSamplers) {
+            mResources.acquire(sampler.texture);
+        }
+    }
+
+    void gc() noexcept {
+        mUBOCache.gc();
+        mSamplerCache.gc();
+    }
+
+    // This will write/update all of the descriptor set (and create a set if a one of the same
+    // layout is not available).
+    void bind(VulkanCommandBuffer* commands, GetPipelineLayoutFunction& getPipelineLayoutFn) {
+        LayoutArray layouts = {VK_NULL_HANDLE, VK_NULL_HANDLE};
+        std::array<VkDescriptorSet, 2> descSets = {VK_NULL_HANDLE, VK_NULL_HANDLE};
+        uint8_t descSetCount = 0;
+
+        VkWriteDescriptorSet descriptorWrites[UNIFORM_BINDING_COUNT + SAMPLER_BINDING_COUNT];
+        VkDescriptorBufferInfo uboWrite[UNIFORM_BINDING_COUNT];
+        VkDescriptorImageInfo samplerWrite[SAMPLER_BINDING_COUNT];
+        uint32_t nwrites = 0;
+
+        UBOSet::Key uboKey = {};
+        auto& uboMask = uboKey.mask;
+        uint8_t uboCount = 0;
+        for (auto const& ubo: mUbos) {
+            auto const& binding = std::get<0>(ubo);
+
+            uboKey.buffers[uboCount] = std::get<1>(ubo)->buffer.getGpuBuffer();
+            uboKey.offsets[uboCount] = std::get<2>(ubo);
+            uboKey.sizes[uboCount] = std::get<3>(ubo);
+            uboCount++;
+
+            // Currently we let ubo be visible in all stages.
+            uboMask |= (VERTEX_STAGE << binding);
+            uboMask |= (FRAGMENT_STAGE << binding);
+        }
+
+        if (uboMask) {
+            layouts[UBO_SET_INDEX] = uboKey.layout = mUBOLayoutCache.getLayout(uboMask);
+
+            auto [descriptorSet, cached] = mUBOCache.obtainSet(uboKey);
+            descSets[descSetCount++] = descriptorSet->vkSet;
+
+            // We need to write to the descriptor set since it wasn't cached.
+            if (!cached) {
+                // If it wasn't cached before, we need to ref the resources that it touches.
+                for (auto const& ubo: mUbos) {
+                    auto const buffer = std::get<1>(ubo);
+                    descriptorSet->resources.acquire(buffer);
+                }
+
+                for (uint8_t i = 0; i < uboCount; ++i) {
+                    auto const& binding = std::get<0>(mUbos[i]);
+                    VkWriteDescriptorSet& descriptorWrite = descriptorWrites[nwrites++];
+                    auto& writeInfo = uboWrite[i];
+                    writeInfo = {
+                        .buffer = uboKey.buffers[i],
+                        .offset = uboKey.offsets[i],
+                        .range = uboKey.sizes[i],
+                    };
+                    descriptorWrite = {
+                        .dstSet = descriptorSet->vkSet,
+                        .dstBinding = binding,
+                        .descriptorCount = 1,
+                        .descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
+                        .pBufferInfo = &writeInfo,
+                    };
+                }
+            }
+            commands->acquire(descriptorSet);
+        }
+
+        SamplerSet::Key samplerKey = {};
+        auto& samplerMask = samplerKey.mask;
+        uint8_t samplerCount = 0;
+        for (auto const sampler: mSamplers) {
+            samplerMask |= (sampler.stage << sampler.binding);
+
+            samplerKey.sampler[samplerCount] = sampler.info.sampler;
+            samplerKey.imageView[samplerCount] = sampler.info.imageView;
+            samplerKey.imageLayout[samplerCount] = sampler.info.imageLayout;
+            samplerCount++;
+        }
+
+        if (samplerMask) {
+            layouts[SAMPLER_SET_INDEX] = samplerKey.layout =
+                    mSamplerLayoutCache.getLayout(samplerMask);
+
+            auto [descriptorSet, cached] = mSamplerCache.obtainSet(samplerKey);
+            descSets[descSetCount++] = descriptorSet->vkSet;
+
+            if (!cached) {
+                for (auto const& sampler: mSamplers) {
+                    descriptorSet->resources.acquire(sampler.texture);
+                }
+
+                for (uint8_t i = 0; i < samplerCount; ++i) {
+                    auto const& binding = mSamplers[i].binding;
+                    VkWriteDescriptorSet& descriptorWrite = descriptorWrites[nwrites++];
+                    auto& writeInfo = samplerWrite[i];
+                    writeInfo = mSamplers[i].info;
+                    descriptorWrite = {
+                        .dstSet = descriptorSet->vkSet,
+                        .dstBinding = binding,
+                        .descriptorCount = 1,
+                        .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                        .pImageInfo = &writeInfo,
+                    };
+                }
+            }
+            commands->acquire(descriptorSet);
+        }
+        if (nwrites) {
+            vkUpdateDescriptorSets(mDevice, nwrites, descriptorWrites, 0, nullptr);
+        }
+
+        VkPipelineLayout const pipelineLayout = getPipelineLayoutFn(layouts);
+        VkCommandBuffer const cmdbuffer = commands->buffer();
+
+        BoundState state {
+            .cmdbuf = cmdbuffer,
+            .pipelineLayout = pipelineLayout,
+        };
+
+        if (state == mPreviousBoundState) {
+            return;
+        }
+
+        vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0,
+                descSetCount, descSets.data(), 0, nullptr);
+
+        mPreviousBoundState = state;
+
+        // Once bound, the resources are now ref'd in the descriptor set and the references in this
+        // class can be released and the descriptor set is ref'd by the command buffer.
+        mResources.clear();
+    }
+
+private:
+    struct BoundState {
+        VkCommandBuffer cmdbuf;
+        VkPipelineLayout pipelineLayout;
+
+        bool operator==(BoundState const& b) const {
+            return cmdbuf == b.cmdbuf && pipelineLayout == b.pipelineLayout;
+        }
+    };
+
+    VkDevice mDevice;
+
+    UBOSetLayoutCache mUBOLayoutCache;
+    UBOSetCache mUBOCache;
+
+    SamplerSetLayoutCache mSamplerLayoutCache;
+    SamplerSetCache mSamplerCache;
+
+    std::vector<std::tuple<uint32_t, VulkanBufferObject*, VkDeviceSize, VkDeviceSize>> mUbos;
+    SamplerArray mSamplers;
+    VulkanAcquireOnlyResourceManager mResources;
+
+    BoundState mPreviousBoundState;
+};
+
+VulkanDescriptorSetManager::VulkanDescriptorSetManager(VkDevice device,
+        VulkanResourceAllocator* resourceAllocator)
+    : mImpl(new Impl(device, resourceAllocator)) {}
+
+void VulkanDescriptorSetManager::terminate() noexcept {
+    assert_invariant(mImpl);
+    delete mImpl;
+    mImpl = nullptr;
+}
+
+void VulkanDescriptorSetManager::gc() noexcept { mImpl->gc(); }
+
+void VulkanDescriptorSetManager::bind(VulkanCommandBuffer* commands,
+        GetPipelineLayoutFunction& getPipelineLayoutFn) {
+    mImpl->bind(commands, getPipelineLayoutFn);
+}
+
+void VulkanDescriptorSetManager::setUniformBufferObject(uint32_t bindingIndex,
+        VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept {
+    mImpl->setUniformBufferObject(bindingIndex, bufferObject, offset, size);
+}
+
+void VulkanDescriptorSetManager::setSamplers(SamplerArray&& samplers) {
+    mImpl->setSamplers(std::move(samplers));
+}
+
+} // namespace filament::backend
@@ -0,0 +1,111 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
+#define TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
+
+#include <vulkan/VulkanResourceAllocator.h>
+#include <vulkan/VulkanTexture.h>
+#include <vulkan/VulkanUtility.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Program.h>
+#include <backend/TargetBufferInfo.h>
+
+#include <utils/bitset.h>
+
+#include <bluevk/BlueVK.h>
+#include <tsl/robin_map.h>
+
+namespace filament::backend {
+
+// We need to make this class public to enable allocation on the HandleAllocator.
+struct VulkanDescriptorSet : public VulkanResourceBase {
+public:
+    // Because we need to recycle descriptor set not used, we allow for a callback that the "Pool"
+    // can use to repackage the vk handle.
+    using OnRecycle = std::function<void(VulkanDescriptorSet*)>;
+
+    VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+            VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn);
+
+    ~VulkanDescriptorSet();
+
+    static VulkanDescriptorSet* create(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+            VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn);
+
+    // TODO: maybe change to fixed size for performance.
+    VulkanAcquireOnlyResourceManager resources;
+
+    VkDescriptorSet const vkSet;
+    VkDescriptorSetLayout const layout;
+
+private:
+    OnRecycle mOnRecycleFn;
+};
+
+// Abstraction over the pool and the layout cache.
+class VulkanDescriptorSetManager {
+public:
+    using StageBitMask = uint32_t;
+    using UniformBufferBitmask = StageBitMask;
+    using SamplerBitmask = StageBitMask;
+
+    static constexpr uint8_t CONCURRENT_DESCRIPTOR_SET_COUNT = 2;// UBO and samplers
+    static constexpr uint8_t MAX_SUPPORTED_SHADER_STAGE = 2;     // Vertex and fragment.
+
+    static_assert(sizeof(UniformBufferBitmask) * 8 >=
+                  (Program::UNIFORM_BINDING_COUNT) *MAX_SUPPORTED_SHADER_STAGE);
+    static_assert(sizeof(SamplerBitmask) * 8 >=
+                  Program::SAMPLER_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE);
+
+    static constexpr StageBitMask VERTEX_STAGE = 0x1;
+    static constexpr StageBitMask FRAGMENT_STAGE = (0x1 << (sizeof(StageBitMask) / 4));
+    static constexpr uint8_t UBO_SET_INDEX = 0;
+    static constexpr uint8_t SAMPLER_SET_INDEX = 1;
+
+    using LayoutArray = std::array<VkDescriptorSetLayout, CONCURRENT_DESCRIPTOR_SET_COUNT>;
+    struct SamplerBundle {
+        VkDescriptorImageInfo info = {};
+        VulkanTexture* texture = nullptr;
+        uint8_t binding = 0;
+        SamplerBitmask stage = 0;
+    };
+    using SamplerArray = CappedArray<SamplerBundle, Program::SAMPLER_BINDING_COUNT>;
+    using GetPipelineLayoutFunction = std::function<VkPipelineLayout(LayoutArray)>;
+
+    VulkanDescriptorSetManager(VkDevice device, VulkanResourceAllocator* resourceAllocator);
+
+    void terminate() noexcept;
+
+    void gc() noexcept;
+
+    // This will write/update all of the descriptor set.
+    void bind(VulkanCommandBuffer* commands, GetPipelineLayoutFunction& getPipelineFn);
+
+    void setUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
+            VkDeviceSize offset, VkDeviceSize size) noexcept;
+
+    void setSamplers(SamplerArray&& samplers);
+
+private:
+    class Impl;
+    Impl* mImpl;
+};
+
+} // namespace filament::backend
+
+#endif// TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
@@ -1,110 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
-#define TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
-
-#include <vulkan/VulkanResourceAllocator.h>
-#include <vulkan/VulkanTexture.h>
-#include <vulkan/VulkanUtility.h>
-
-#include <backend/DriverEnums.h>
-#include <backend/Program.h>
-#include <backend/TargetBufferInfo.h>
-
-#include <utils/bitset.h>
-
-#include <bluevk/BlueVK.h>
-#include <tsl/robin_map.h>
-
-#include <functional>
-
-namespace filament::backend {
-
-using namespace descset;
-
-// [GDSR]: Great-Descriptor-Set-Refactor: As of 03/20/24, the Filament frontend is planning to
-// introduce descriptor set. This PR will arrive before that change is complete. As such, some of
-// the methods introduced here will be obsolete, and certain logic will be generalized.
-
-// Abstraction over the pool and the layout cache.
-class VulkanDescriptorSetManager {
-public:
-    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT =
-            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT;
-    using GetPipelineLayoutFunction =
-            std::function<VkPipelineLayout(VulkanDescriptorSetLayoutList const&)>;
-
-    VulkanDescriptorSetManager(VkDevice device, VulkanResourceAllocator* resourceAllocator);
-
-    void terminate() noexcept;
-
-    void gc();
-
-    // TODO: Obsolete after [GDSR].
-    // This will write/update/bind all of the descriptor set. After [GDSR], the binding for
-    // descriptor sets will not depend on the layout described in the program.
-    VkPipelineLayout bind(VulkanCommandBuffer* commands, VulkanProgram* program,
-            GetPipelineLayoutFunction& getPipelineLayoutFn);
-
-    // TODO: Obsolete after [GDSR].
-    // This is to "dynamically" bind UBOs that might have offsets changed between pipeline binding
-    // and the draw call. We do this because UBOs for primitives that are part of the same
-    // renderable can be stored within one buffer. This can be a no-op if there were no range
-    // changes between the pipeline bind and the draw call. We will re-use applicable states
-    // provided within the bind() call, including the UBO descriptor set layout. TODO: make it a
-    // proper dynamic binding when Filament-side descriptor changes are completed.
-    void dynamicBind(VulkanCommandBuffer* commands, Handle<VulkanDescriptorSetLayout> uboLayout);
-
-    // TODO: Obsolete after [GDSR].
-    // Since we use program pointer as cache key, we need to clear the cache when it's freed.
-    void clearProgram(VulkanProgram* program) noexcept;
-
-    Handle<VulkanDescriptorSetLayout> createLayout(descset::DescriptorSetLayout const& layout);
-
-    void destroyLayout(Handle<VulkanDescriptorSetLayout> layout);
-
-    void updateBuffer(Handle<VulkanDescriptorSet> set, uint8_t binding,
-            VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept;
-
-    void updateSampler(Handle<VulkanDescriptorSet> set, uint8_t binding,
-            VulkanTexture* texture, VkSampler sampler) noexcept;
-
-    void updateInputAttachment(Handle<VulkanDescriptorSet> set, VulkanAttachment attachment) noexcept;
-
-    void clearBuffer(uint32_t bindingIndex);
-
-    void setPlaceHolders(VkSampler sampler, VulkanTexture* texture,
-            VulkanBufferObject* bufferObject) noexcept;
-
-    void clearState() noexcept;
-
-    // TODO: To be completed after [GDSR]
-    Handle<VulkanDescriptorSet> createSet(Handle<VulkanDescriptorSetLayout> layout) {
-        return Handle<VulkanDescriptorSet>();
-    }
-
-    // TODO: To be completed after [GDSR]
-    void destroySet(Handle<VulkanDescriptorSet> set) {}
-
-private:
-    class Impl;
-    Impl* mImpl;
-};
-
-}// namespace filament::backend
-
-#endif// TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
@@ -1,64 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "VulkanPipelineLayoutCache.h"
-
-#include <vulkan/VulkanResourceAllocator.h>
-
-namespace filament::backend {
-
-VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
-        VulkanDescriptorSetLayoutList const& descriptorSetLayouts) {
-    PipelineLayoutKey key = {VK_NULL_HANDLE};
-    uint8_t descSetLayoutCount = 0;
-    for (auto layoutHandle: descriptorSetLayouts) {
-        if (layoutHandle) {
-            auto layout = mAllocator->handle_cast<VulkanDescriptorSetLayout*>(layoutHandle);
-            key[descSetLayoutCount++] = layout->vklayout;
-        }
-    }
-
-    auto iter = mPipelineLayouts.find(key);
-    if (iter != mPipelineLayouts.end()) {
-        PipelineLayoutCacheEntry& entry = mPipelineLayouts[key];
-        entry.lastUsed = mTimestamp++;
-        return entry.handle;
-    }
-
-    VkPipelineLayoutCreateInfo info{
-        .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
-        .pNext = nullptr,
-        .setLayoutCount = (uint32_t) descSetLayoutCount,
-        .pSetLayouts = key.data(),
-        .pushConstantRangeCount = 0,
-    };
-    VkPipelineLayout layout;
-    vkCreatePipelineLayout(mDevice, &info, VKALLOC, &layout);
-
-    mPipelineLayouts[key] = {
-        .handle = layout,
-        .lastUsed = mTimestamp++,
-    };
-    return layout;
-}
-
-void VulkanPipelineLayoutCache::terminate() noexcept {
-    for (auto const& [key, entry]: mPipelineLayouts) {
-        vkDestroyPipelineLayout(mDevice, entry.handle, VKALLOC);
-    }
-}
-
-}// namespace filament::backend
@@ -1,71 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_VULKANPIPELINELAYOUTCACHE_H
-#define TNT_FILAMENT_BACKEND_VULKANPIPELINELAYOUTCACHE_H
-
-#include <vulkan/VulkanHandles.h>
-#include <bluevk/BlueVK.h>
-
-#include <utils/Hash.h>
-
-#include <tsl/robin_map.h>
-
-namespace filament::backend {
-
-class VulkanPipelineLayoutCache {
-public:
-    VulkanPipelineLayoutCache(VkDevice device, VulkanResourceAllocator* allocator)
-        : mDevice(device),
-          mAllocator(allocator),
-          mTimestamp(0) {}
-
-    void terminate() noexcept;
-
-    using PipelineLayoutKey = std::array<VkDescriptorSetLayout,
-            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
-
-    VulkanPipelineLayoutCache(VulkanPipelineLayoutCache const&) = delete;
-    VulkanPipelineLayoutCache& operator=(VulkanPipelineLayoutCache const&) = delete;
-
-    VkPipelineLayout getLayout(VulkanDescriptorSetLayoutList const& descriptorSetLayouts);
-
-private:
-    using Timestamp = uint64_t;
-    struct PipelineLayoutCacheEntry {
-        VkPipelineLayout handle;
-        Timestamp lastUsed;
-    };
-
-    using PipelineLayoutKeyHashFn = utils::hash::MurmurHashFn<PipelineLayoutKey>;
-    struct PipelineLayoutKeyEqual {
-        bool operator()(PipelineLayoutKey const& k1, PipelineLayoutKey const& k2) const {
-            return 0 == memcmp((const void*) &k1, (const void*) &k2, sizeof(PipelineLayoutKey));
-        }
-    };
-
-    using PipelineLayoutMap = tsl::robin_map<PipelineLayoutKey, PipelineLayoutCacheEntry,
-            PipelineLayoutKeyHashFn, PipelineLayoutKeyEqual>;
-
-    VkDevice mDevice;
-    VulkanResourceAllocator* mAllocator;
-    Timestamp mTimestamp;
-    PipelineLayoutMap mPipelineLayouts;
-};
-
-}
-
-#endif // TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H
@@ -137,14 +137,18 @@ void printDepthFormats(VkPhysicalDevice device) {
    const VkFormatFeatureFlags required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT
                                              | VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
    utils::slog.i << "Sampleable depth formats: ";
-    for (VkFormat format : ALL_VK_FORMATS) {
+    for (VkFormat format = (VkFormat) 1;;) {
        VkFormatProperties props;
        vkGetPhysicalDeviceFormatProperties(device, format, &props);
        if ((props.optimalTilingFeatures & required) == required) {
            utils::slog.i << format << " ";
        }
+        if (format == VK_FORMAT_ASTC_12x12_SRGB_BLOCK) {
+            utils::slog.i << utils::io::endl;
+            break;
+        }
+        format = (VkFormat) (1 + (int) format);
    }
-    utils::slog.i << utils::io::endl;
 }
 #endif

@@ -253,7 +257,6 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
    // Create the Vulkan instance.
    VkApplicationInfo appInfo = {};
    appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
-    appInfo.pEngineName = "Filament";
    appInfo.apiVersion
            = VK_MAKE_API_VERSION(0, FVK_REQUIRED_VERSION_MAJOR, FVK_REQUIRED_VERSION_MINOR, 0);
    instanceCreateInfo.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO;
@@ -340,13 +343,13 @@ std::tuple<ExtensionSet, ExtensionSet> pruneExtensions(VkPhysicalDevice device,
    ExtensionSet newInstExts = instExts;
    ExtensionSet newDeviceExts = deviceExts;

-#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
+#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)    
    // debugUtils and debugMarkers extensions are used mutually exclusively.
    if (newInstExts.find(VK_EXT_DEBUG_UTILS_EXTENSION_NAME) != newInstExts.end()
            && newDeviceExts.find(VK_EXT_DEBUG_MARKER_EXTENSION_NAME) != newDeviceExts.end()) {
        newDeviceExts.erase(VK_EXT_DEBUG_MARKER_EXTENSION_NAME);
    }
-#endif
+#endif    

 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
    // debugMarker must also request debugReport the instance extension. So check if that's present.
@@ -401,7 +404,7 @@ inline int deviceTypeOrder(VkPhysicalDeviceType deviceType) {
        case VK_PHYSICAL_DEVICE_TYPE_OTHER:
            return 1;
        default:
-            utils::slog.w << "deviceTypeOrder: Unexpected deviceType: " << deviceType
+            utils::slog.w << "devcieTypeOrder: Unexpected deviceType: " << deviceType
                          << utils::io::endl;
            return -1;
    }
@@ -496,7 +499,7 @@ VkPhysicalDevice selectPhysicalDevice(VkInstance instance,
    return device;
 }

-VkFormatList findAttachmentDepthStencilFormats(VkPhysicalDevice device) {
+VkFormatList findAttachmentDepthFormats(VkPhysicalDevice device) {
    VkFormatFeatureFlags const features = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;

    // The ordering here indicates the preference of choosing depth+stencil format.
@@ -520,24 +523,6 @@ VkFormatList findAttachmentDepthStencilFormats(VkPhysicalDevice device) {
    return ret;
 }

-VkFormatList findBlittableDepthStencilFormats(VkPhysicalDevice device) {
-    std::vector<VkFormat> selectedFormats;
-    VkFormatFeatureFlags const required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT |
-            VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
-    for (VkFormat format : ALL_VK_FORMATS) {
-        if (isVkDepthFormat(format)) {
-            VkFormatProperties props;
-            vkGetPhysicalDeviceFormatProperties(device, format, &props);
-            if ((props.optimalTilingFeatures & required) == required) {
-                selectedFormats.push_back(format);
-            }
-        }
-    }
-    VkFormatList ret(selectedFormats.size());
-    std::copy(selectedFormats.begin(), selectedFormats.end(), ret.begin());
-    return ret;
-}
-
 }// anonymous namespace

 using SwapChainPtr = VulkanPlatform::SwapChainPtr;
@@ -683,11 +668,9 @@ Driver* VulkanPlatform::createDriver(void* sharedContext,
            "Debug utils should not be enabled in release build.");
 #endif

-    context.mDepthStencilFormats = findAttachmentDepthStencilFormats(mImpl->mPhysicalDevice);
-    context.mBlittableDepthStencilFormats =
-            findBlittableDepthStencilFormats(mImpl->mPhysicalDevice);
+    context.mDepthFormats = findAttachmentDepthFormats(mImpl->mPhysicalDevice);

-    assert_invariant(context.mDepthStencilFormats.size() > 0);
+    assert_invariant(context.mDepthFormats.size() > 0);

 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
    printDepthFormats(mImpl->mPhysicalDevice);
@@ -243,7 +243,7 @@ VkResult VulkanPlatformSurfaceSwapChain::create() {
    mSwapChainBundle.colors = enumerate(vkGetSwapchainImagesKHR, mDevice, mSwapchain);
    mSwapChainBundle.colorFormat = surfaceFormat.format;
    mSwapChainBundle.depthFormat =
-            selectDepthFormat(mContext.getAttachmentDepthStencilFormats(), mHasStencil);
+            selectDepthFormat(mContext.getAttachmentDepthFormats(), mHasStencil);
    mSwapChainBundle.depth = createImage(mSwapChainBundle.extent, mSwapChainBundle.depthFormat);

    slog.i << "vkCreateSwapchain"
@@ -330,7 +330,7 @@ VulkanPlatformHeadlessSwapChain::VulkanPlatformHeadlessSwapChain(VulkanContext c

    bool const hasStencil = (flags & backend::SWAP_CHAIN_HAS_STENCIL_BUFFER) != 0;
    mSwapChainBundle.depthFormat =
-            selectDepthFormat(mContext.getAttachmentDepthStencilFormats(), hasStencil);
+            selectDepthFormat(mContext.getAttachmentDepthFormats(), hasStencil);
    mSwapChainBundle.depth = createImage(extent, mSwapChainBundle.depthFormat);
 }

@@ -24,7 +24,6 @@

 #include <unordered_map>
 #include <variant>
-#include <vector>

 namespace filament::backend {

@@ -73,7 +72,7 @@ void workaroundSpecConstant(Program::ShaderBlob const& blob,
        std::vector<uint32_t>& output) {
    using WordMap = std::unordered_map<uint32_t, uint32_t>;
    using SpecValueMap = std::unordered_map<uint32_t, SpecConstantValue>;
-    constexpr size_t HEADER_SIZE = 5;
+    constexpr size_t const HEADER_SIZE = 5;

    WordMap varToIdMap;
    SpecValueMap idToValue;
@@ -146,61 +145,4 @@ void workaroundSpecConstant(Program::ShaderBlob const& blob,
    output.resize(outputCursor);
 }

-std::tuple<uint32_t,uint32_t, uint32_t> getProgramBindings(Program::ShaderBlob const& blob) {
-    std::unordered_map<uint32_t, uint32_t> targetToSet, targetToBinding;
-
-    constexpr size_t HEADER_SIZE = 5;
-
-    size_t const dataSize = blob.size() / 4;
-    uint32_t const* data = (uint32_t*) blob.data();
-
-    for (uint32_t cursor = HEADER_SIZE, cursorEnd = dataSize; cursor < cursorEnd;) {
-        uint32_t const firstWord = data[cursor];
-        uint32_t const wordCount = firstWord >> 16;
-        uint32_t const op = firstWord & 0x0000FFFF;
-
-        switch(op) {
-            case spv::Op::OpDecorate: {
-                if (data[cursor + 2] == spv::Decoration::DecorationDescriptorSet) {
-                    uint32_t const targetVar = data[cursor + 1];
-                    uint32_t const setId = data[cursor + 3];
-                    targetToSet[targetVar] = setId;
-                } else if (data[cursor + 2] == spv::Decoration::DecorationBinding) {
-                    uint32_t const targetVar = data[cursor + 1];
-                    uint32_t const binding = data[cursor + 3];
-                    targetToBinding[targetVar] = binding;
-                }
-                break;
-            }
-            default:
-                break;
-        }
-        cursor += wordCount;
-    }
-
-    constexpr uint32_t UBO = 0;
-    constexpr uint32_t SAMPLER = 1;
-    constexpr uint32_t IATTACHMENT = 2;
-    uint32_t ubo = 0, sampler = 0, inputAttachment = 0;
-
-    for (auto const& [target, setId]: targetToSet) {
-        uint32_t const binding = targetToBinding[target];
-        assert_invariant(binding < 32);
-        switch(setId) {
-            case UBO:
-                ubo |= (1 << binding);
-                break;
-            case SAMPLER:
-                sampler |= (1 << binding);
-                break;
-            case IATTACHMENT:
-                inputAttachment |= (1 << binding);
-                break;
-            default:
-                PANIC_POSTCONDITION("unexpected %d", (int) setId);
-        }
-    }
-    return {ubo, sampler, inputAttachment};
-}
-
 } // namespace filament::backend
@@ -17,11 +17,10 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANSPIRVUTILS_H
 #define TNT_FILAMENT_BACKEND_VULKANSPIRVUTILS_H

-#include <backend/Program.h>
+#include "backend/Program.h"

 #include <utils/FixedCapacityVector.h>

-#include <tuple>
 #include <vector>

 namespace filament::backend {
@@ -43,9 +42,6 @@ void workaroundSpecConstant(Program::ShaderBlob const& blob,
        utils::FixedCapacityVector<Program::SpecializationConstant> const& specConstants,
        std::vector<uint32_t>& output);

-// bindings for UBO, samplers, input attachment
-std::tuple<uint32_t, uint32_t, uint32_t> getProgramBindings(Program::ShaderBlob const& blob);
-
 } // namespace filament::backend

 #endif // TNT_FILAMENT_BACKEND_VULKANSPIRVUTILS_H
@@ -79,20 +79,6 @@ ShaderGenerator::ShaderGenerator(std::string vertex, std::string fragment,
          mVertexBlob(transpileShader(ShaderStage::VERTEX, std::move(vertex), backend, isMobile, sib)),
          mFragmentBlob(transpileShader(ShaderStage::FRAGMENT, std::move(fragment), backend,
                  isMobile, sib)) {
-    switch (backend) {
-        case Backend::OPENGL:
-            mShaderLanguage = filament::backend::ShaderLanguage::ESSL3;
-            break;
-        case Backend::VULKAN:
-            mShaderLanguage = filament::backend::ShaderLanguage::SPIRV;
-            break;
-        case Backend::METAL:
-            mShaderLanguage = filament::backend::ShaderLanguage::MSL;
-            break;
-        case Backend::NOOP:
-            mShaderLanguage = filament::backend::ShaderLanguage::ESSL3;
-            break;
-    }
 }

 ShaderGenerator::Blob ShaderGenerator::transpileShader(
@@ -174,7 +160,6 @@ ShaderGenerator::Blob ShaderGenerator::transpileShader(

 Program ShaderGenerator::getProgram(filament::backend::DriverApi&) noexcept {
    Program program;
-    program.shaderLanguage(mShaderLanguage);
    program.shader(ShaderStage::VERTEX, mVertexBlob.data(), mVertexBlob.size());
    program.shader(ShaderStage::FRAGMENT, mFragmentBlob.data(), mFragmentBlob.size());
    return program;
@@ -60,7 +60,6 @@ private:
    Blob mFragmentBlob;
    std::string mCompiledVertexShader;
    std::string mCompiledFragmentShader;
-    filament::backend::ShaderLanguage mShaderLanguage;

 };

@@ -238,13 +238,13 @@ TEST_F(BackendTest, ColorMagnify) {
            {0, 0, kSrcTexWidth >> srcLevel, kSrcTexHeight >> srcLevel}, SamplerMagFilter::LINEAR);

    // Push through an empty frame to allow the texture to upload and the blit to execute.
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
    api.commit(swapChain);
    api.endFrame(0);

    // Grab a screenshot.
    ScreenshotParams params { kDstTexWidth, kDstTexHeight, "ColorMagnify.png" };
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
    dumpScreenshot(api, dstRenderTargets[0], &params);
    api.commit(swapChain);
    api.endFrame(0);
@@ -402,7 +402,7 @@ TEST_F(BackendTest, ColorResolve) {
    });

    // FIXME: on Metal this triangle is not drawn. Can't understand why.
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
        api.beginRenderPass(srcRenderTarget, params);
            api.bindUniformBuffer(0, ubuffer);
            api.draw(state, triangle.getRenderPrimitive(), 0, 3, 1);
@@ -484,13 +484,13 @@ TEST_F(BackendTest, Blit2DTextureArray) {
            {0, 0, kSrcTexWidth >> srcLevel, kSrcTexHeight >> srcLevel}, SamplerMagFilter::LINEAR);

    // Push through an empty frame to allow the texture to upload and the blit to execute.
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
    api.commit(swapChain);
    api.endFrame(0);

    // Grab a screenshot.
    ScreenshotParams params { kDstTexWidth, kDstTexHeight, "Blit2DTextureArray.png" };
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
    dumpScreenshot(api, dstRenderTarget, &params);
    api.commit(swapChain);
    api.endFrame(0);
@@ -574,13 +574,13 @@ TEST_F(BackendTest, BlitRegion) {
            SamplerMagFilter::LINEAR);

    // Push through an empty frame to allow the texture to upload and the blit to execute.
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
    api.commit(swapChain);
    api.endFrame(0);

    // Grab a screenshot.
    ScreenshotParams params { kDstTexWidth, kDstTexHeight, "BlitRegion.png" };
-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);
    dumpScreenshot(api, dstRenderTarget, &params);
    api.commit(swapChain);
    api.endFrame(0);
@@ -655,7 +655,7 @@ TEST_F(BackendTest, BlitRegionToSwapChain) {
        .height = kDstTexHeight - 10,
    };

-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);

    api.blitDEPRECATED(TargetBufferFlags::COLOR0, dstRenderTarget,
            dstRect, srcRenderTargets[srcLevel],
@@ -143,7 +143,7 @@ TEST_F(BackendTest, VertexBufferUpdate) {
        }

        getDriverApi().makeCurrent(swapChain, swapChain);
-        getDriverApi().beginFrame(0, 0, 0);
+        getDriverApi().beginFrame(0, 0);

        // Draw 10 triangles, updating the vertex buffer / index buffer each time.
        size_t triangleIndex = 0;
@@ -58,7 +58,6 @@ kernel void main0() {}
    }

    Program program;
-    program.shaderLanguage(ShaderLanguage::ESSL3);
    program.shader(ShaderStage::COMPUTE, shader.data(), shader.size() + 1);

    Handle<HwProgram> ph = driver.createProgram(std::move(program));
@@ -145,7 +144,6 @@ kernel void main0(device Output_data& output_data [[buffer(0)]],
    driver.updateBufferObject(input_data, { data.data(), size }, 0);

    Program program;
-    program.shaderLanguage(ShaderLanguage::ESSL3);
    program.shader(ShaderStage::COMPUTE, shader.data(), shader.size() + 1);
    Handle<HwProgram> ph = driver.createProgram(std::move(program));

@@ -200,7 +200,7 @@ TEST_F(BackendTest, FeedbackLoops) {
            auto ubuffer = api.createBufferObject(sizeof(MaterialParams),
                    BufferObjectBinding::UNIFORM, BufferUsage::STATIC);
            api.makeCurrent(swapChain, swapChain);
-            api.beginFrame(0, 0, 0);
+            api.beginFrame(0, 0);
            api.bindSamplers(0, sgroup);
            api.bindUniformBuffer(0, ubuffer);

@@ -414,7 +414,7 @@ TEST_F(BackendTest, UpdateImageSRGB) {

    api.update3DImage(texture, 0, 0, 0, 0, 512, 512, 1, std::move(descriptor));

-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);

    // Update samplers.
    SamplerGroup samplers(1);
@@ -486,7 +486,7 @@ TEST_F(BackendTest, UpdateImageMipLevel) {
    PixelBufferDescriptor descriptor = checkerboardPixelBuffer(pixelFormat, pixelType, 512);
    api.update3DImage(texture, /* level*/ 1, 0, 0, 0, 512, 512, 1, std::move(descriptor));

-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);

    // Update samplers.
    SamplerGroup samplers(1);
@@ -570,7 +570,7 @@ TEST_F(BackendTest, UpdateImage3D) {

    api.update3DImage(texture, 0, 0, 0, 0, 512, 512, 4, std::move(descriptor));

-    api.beginFrame(0, 0, 0);
+    api.beginFrame(0, 0);

    // Update samplers.
    SamplerGroup samplers(1);
@@ -127,7 +127,7 @@ TEST_F(BackendTest, MRT) {
        getDriverApi().startCapture(0);

        getDriverApi().makeCurrent(swapChain, swapChain);
-        getDriverApi().beginFrame(0, 0, 0);
+        getDriverApi().beginFrame(0, 0);

        // Draw a triangle.
        getDriverApi().beginRenderPass(renderTarget, params);
@@ -143,7 +143,7 @@ TEST_F(BackendTest, SetMinMaxLevel) {

        TrianglePrimitive triangle(api);

-        api.beginFrame(0, 0, 0);
+        api.beginFrame(0, 0);

        // We set the base mip to 1, and the max mip to 3
        // Level 0: 128x128 (red)
@@ -100,7 +100,7 @@ TEST_F(BackendTest, MissingRequiredAttributes) {
        getDriverApi().startCapture(0);

        getDriverApi().makeCurrent(swapChain, swapChain);
-        getDriverApi().beginFrame(0, 0, 0);
+        getDriverApi().beginFrame(0, 0);

        // Render a triangle.
        getDriverApi().beginRenderPass(defaultRenderTarget, params);
@@ -295,7 +295,7 @@ TEST_F(ReadPixelsTest, ReadPixels) {
        params.viewport.width = t.getRenderTargetSize();

        getDriverApi().makeCurrent(swapChain, swapChain);
-        getDriverApi().beginFrame(0, 0, 0);
+        getDriverApi().beginFrame(0, 0);

        // Render a white triangle over blue.
        getDriverApi().beginRenderPass(renderTarget, params);
@@ -434,7 +434,7 @@ TEST_F(ReadPixelsTest, ReadPixelsPerformance) {
        }

        getDriverApi().makeCurrent(swapChain, swapChain);
-        getDriverApi().beginFrame(0, 0, 0);
+        getDriverApi().beginFrame(0, 0);

        // Render some content, just so we don't read back uninitialized data.
        getDriverApi().beginRenderPass(renderTarget, params);
--- a/Show More
+++ b/Show More