Throw an NSException on Apple platforms when panicking

2024-03-27 16:04:49 -07:00
106 changed files with 2560 additions and 4241 deletions
@@ -819,9 +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)
@@ -8,4 +8,4 @@ appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).

 ## Release notes for next branch cut

- Add new matedit tool
+- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.51.5'
+    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', '~> 1.51.5'
+pod 'Filament', '~> 1.51.0'
 ```

 ### Snapshots
@@ -7,19 +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).

-## v1.51.5
-
-
-## v1.51.4
-
-
-## v1.51.3
-
-
-## v1.51.2
-
- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`
-
 ## v1.51.1


@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.51.5
+VERSION_NAME=1.51.0

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

@@ -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
@@ -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
@@ -158,16 +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) {
@@ -180,8 +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";
    }
 }

@@ -1241,6 +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
@@ -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;
@@ -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,
@@ -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
@@ -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) {
@@ -1634,7 +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();

@@ -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) {
@@ -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), reserved(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 targetIndex = 0;    // optimization: index corresponding to target
-        bool imported           : 1;
-        uint8_t sidecarSamples  : 4;
-        uint8_t reserved        : 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") ||
@@ -947,37 +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
@@ -311,6 +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;
@@ -468,29 +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;
@@ -500,11 +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;

@@ -551,6 +522,9 @@ private:
            {   bugs.delay_fbo_destruction,
                    "delay_fbo_destruction",
                    ""},
+            {   bugs.disable_thread_affinity,
+                    "disable_thread_affinity",
+                    ""},
            {   bugs.force_feature_level0,
                    "force_feature_level0",
                    ""},
@@ -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();
 }
@@ -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;
@@ -806,6 +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);
@@ -2059,6 +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;
    }
@@ -2337,7 +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
            {
@@ -3062,10 +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);

@@ -3083,7 +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;
    }

@@ -3100,6 +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
@@ -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;

@@ -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

@@ -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>
@@ -19,7 +19,6 @@

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

 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.
@@ -19,6 +19,7 @@

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

 #include <utils/bitset.h>
@@ -41,7 +42,7 @@ struct VulkanTimerQuery;
 struct VulkanCommandBuffer;

 struct VulkanAttachment {
-    VulkanTexture* texture = nullptr;
+    VulkanTexture* texture;
    uint8_t level = 0;
    uint16_t layer = 0;
    VkImage getImage() const;
@@ -14,7 +14,7 @@
 * limitations under the License.
 */

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

 #include "CommandStreamDispatcher.h"
 #include "DataReshaper.h"
@@ -102,15 +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,
@@ -223,14 +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)
@@ -254,17 +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;
@@ -325,14 +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();
@@ -345,12 +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);

@@ -382,8 +360,6 @@ void VulkanDriver::collectGarbage() {
    mCommands.gc();
    mStagePool.gc();
    mFramebufferCache.gc();
-    mPipelineCache.gc();
-    mDescriptorSetManager.gc();

 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
    mResourceAllocator.print();
@@ -513,6 +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);
 }

@@ -563,7 +542,6 @@ void VulkanDriver::destroyProgram(Handle<HwProgram> ph) {
        return;
    }
    auto vkprogram = mResourceAllocator.handle_cast<VulkanProgram*>(ph);
-    mDescriptorSetManager.clearProgram(vkprogram);
    mResourceManager.release(vkprogram);
 }

@@ -1451,7 +1429,12 @@ void VulkanDriver::endRenderPass(int) {
                0, 1, &barrier, 0, nullptr, 0, nullptr);
    }

-    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();
@@ -1470,9 +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);
+        }
    }
 }

@@ -1512,24 +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) {
@@ -1777,14 +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;
        }

@@ -1793,14 +1791,18 @@ 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
@@ -1814,20 +1816,43 @@ void VulkanDriver::bindPipeline(PipelineState pipelineState) {
            texture = mEmptyTexture;
        }

+        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());
-        VulkanDriver::DebugUtils::setName(VK_OBJECT_TYPE_SAMPLER,
-                reinterpret_cast<uint64_t>(samplerInfo.sampler), bindingToName[binding].c_str());
 #endif

-        VkSampler const vksampler = mSamplerCache.getSampler(boundSampler->s);
+        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);
+        }

-        mDescriptorSetManager.updateSampler({}, binding, texture, vksampler);
+        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();
 }

@@ -1867,8 +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"
@@ -72,8 +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;

@@ -110,9 +106,7 @@ private:
    VulkanPlatform* mPlatform = nullptr;
    std::unique_ptr<VulkanTimestamps> mTimestamps;

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

    VulkanSwapChain* mCurrentSwapChain = nullptr;
    VulkanRenderTarget* mDefaultRenderTarget = nullptr;
@@ -129,17 +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;

    bool const mIsSRGBSwapChainSupported;
 };
@@ -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);
@@ -53,125 +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;
-}
-
-UsageFlags 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;
-}
-
-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];

@@ -184,12 +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,
@@ -217,25 +99,12 @@ 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];
@@ -243,9 +112,7 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
        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 = getUsageFlags(binding, group.stageFlags, usage);
+            usage = VulkanPipelineCache::getUsageFlags(binding, group.stageFlags, usage);

 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
            bindingToName[binding] = samplers[i].name.c_str();
@@ -259,6 +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);
@@ -520,35 +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 {
@@ -194,61 +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;
    }

-    inline UsageFlags getUsage() const {
-        return mInfo->usage;
-    }
-
-    // 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              
            {}

        // 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).
-        UsageFlags usage;
-
-        BindingList bindings;
+        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;
@@ -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
@@ -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)                                                                      \
@@ -17,9 +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"

@@ -23,7 +23,6 @@
 #include "VulkanResources.h"
 #include "VulkanImageUtility.h"

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

 #include <unordered_map>
@@ -655,6 +655,4 @@ uint8_t reduceSampleCount(uint8_t sampleCount, VkSampleCountFlags mask) {
    return mostSignificantBit((sampleCount - 1) & mask);
 }

-_BitCountHelper _BitCountHelper::BitCounter = {};
-
 } // namespace filament::backend
@@ -90,9 +90,6 @@ utils::FixedCapacityVector<OutType> enumerate(
 #undef EXPAND_ENUM_NO_ARGS
 #undef EXPAND_ENUM_ARGS

-// Used across pipeline related classes.
-using UsageFlags = utils::bitset128;
-
 // Useful shorthands
 using VkFormatList = utils::FixedCapacityVector<VkFormat>;

@@ -125,14 +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;
        }
@@ -173,26 +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);
@@ -200,127 +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
-
-// 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;
-    }
-
-    static _BitCountHelper BitCounter;
-private:
-    uint8_t data[256];
-};
-
-template<typename MaskType>
-inline uint8_t countBits(MaskType num) {
-    return _BitCountHelper::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
@@ -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;

 };

@@ -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));

@@ -34,7 +34,6 @@
 #include <utils/CString.h>

 #include <stdlib.h>
-#include <optional>

 using namespace utils;
 using namespace filament::backend;
@@ -43,30 +42,31 @@ using namespace filamat;

 namespace filament {

-constexpr std::pair<ChunkType, ChunkType> shaderLanguageToTags(ShaderLanguage language) {
-    switch (language) {
-        case ShaderLanguage::ESSL3:
-            return { ChunkType::MaterialGlsl, ChunkType::DictionaryText };
-        case ShaderLanguage::ESSL1:
-            return { ChunkType::MaterialEssl1, ChunkType::DictionaryText };
-        case ShaderLanguage::MSL:
-            return { ChunkType::MaterialMetal, ChunkType::DictionaryText };
-        case ShaderLanguage::SPIRV:
-            return { ChunkType::MaterialSpirv, ChunkType::DictionarySpirv };
-        case ShaderLanguage::METAL_LIBRARY:
-            return { ChunkType::MaterialMetalLibrary, ChunkType::DictionaryMetalLibrary };
-    }
-}
-
 // ------------------------------------------------------------------------------------------------

-MaterialParser::MaterialParserDetails::MaterialParserDetails(
-        const utils::FixedCapacityVector<ShaderLanguage>& preferredLanguages, const void* data,
-        size_t size)
-    : mManagedBuffer(data, size),
-      mChunkContainer(mManagedBuffer.data(), mManagedBuffer.size()),
-      mPreferredLanguages(preferredLanguages),
-      mMaterialChunk(mChunkContainer) {}
+MaterialParser::MaterialParserDetails::MaterialParserDetails(ShaderLanguage language, const void* data, size_t size)
+        : mManagedBuffer(data, size),
+          mChunkContainer(mManagedBuffer.data(), mManagedBuffer.size()),
+          mMaterialChunk(mChunkContainer) {
+    switch (language) {
+        case ShaderLanguage::ESSL3:
+            mMaterialTag = ChunkType::MaterialGlsl;
+            mDictionaryTag = ChunkType::DictionaryText;
+            break;
+        case ShaderLanguage::ESSL1:
+            mMaterialTag = ChunkType::MaterialEssl1;
+            mDictionaryTag = ChunkType::DictionaryText;
+            break;
+        case ShaderLanguage::MSL:
+            mMaterialTag = ChunkType::MaterialMetal;
+            mDictionaryTag = ChunkType::DictionaryText;
+            break;
+        case ShaderLanguage::SPIRV:
+            mMaterialTag = ChunkType::MaterialSpirv;
+            mDictionaryTag = ChunkType::DictionarySpirv;
+            break;
+    }
+}

 template<typename T>
 UTILS_NOINLINE
@@ -83,9 +83,9 @@ bool MaterialParser::MaterialParserDetails::getFromSimpleChunk(

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

-MaterialParser::MaterialParser(utils::FixedCapacityVector<ShaderLanguage> preferredLanguages,
-        const void* data, size_t size)
-    : mImpl(preferredLanguages, data, size) {}
+MaterialParser::MaterialParser(ShaderLanguage language, const void* data, size_t size)
+        : mImpl(language, data, size) {
+}

 ChunkContainer& MaterialParser::getChunkContainer() noexcept {
    return mImpl.mChunkContainer;
@@ -100,40 +100,20 @@ MaterialParser::ParseResult MaterialParser::parse() noexcept {
    if (UTILS_UNLIKELY(!cc.parse())) {
        return ParseResult::ERROR_OTHER;
    }
-
-    using MaybeShaderLanguageAndChunks =
-            std::optional<std::tuple<ShaderLanguage, ChunkType, ChunkType>>;
-    auto chooseLanguage = [this, &cc]() -> MaybeShaderLanguageAndChunks {
-        for (auto language : mImpl.mPreferredLanguages) {
-            const auto [matTag, dictTag] = shaderLanguageToTags(language);
-            if (cc.hasChunk(matTag) && cc.hasChunk(dictTag)) {
-                return std::make_tuple(language, matTag, dictTag);
-            }
-        }
-        return {};
-    };
-    const auto result = chooseLanguage();
-
-    if (!result.has_value()) {
+    const ChunkType matTag = mImpl.mMaterialTag;
+    const ChunkType dictTag = mImpl.mDictionaryTag;
+    if (UTILS_UNLIKELY(!cc.hasChunk(matTag) || !cc.hasChunk(dictTag))) {
        return ParseResult::ERROR_MISSING_BACKEND;
    }
-
-    const auto [chosenLanguage, matTag, dictTag] = result.value();
    if (UTILS_UNLIKELY(!DictionaryReader::unflatten(cc, dictTag, mImpl.mBlobDictionary))) {
        return ParseResult::ERROR_OTHER;
    }
    if (UTILS_UNLIKELY(!mImpl.mMaterialChunk.initialize(matTag))) {
        return ParseResult::ERROR_OTHER;
    }
-
-    mImpl.mChosenLanguage = chosenLanguage;
    return ParseResult::SUCCESS;
 }

-backend::ShaderLanguage MaterialParser::getShaderLanguage() const noexcept {
-    return mImpl.mChosenLanguage;
-}
-
 // Accessors
 bool MaterialParser::getMaterialVersion(uint32_t* value) const noexcept {
    return mImpl.getFromSimpleChunk(ChunkType::MaterialVersion, value);
@@ -297,16 +277,6 @@ bool MaterialParser::getBlendingMode(BlendingMode* value) const noexcept {
    return mImpl.getFromSimpleChunk(ChunkType::MaterialBlendingMode, reinterpret_cast<uint8_t*>(value));
 }

-bool MaterialParser::getCustomBlendFunction(std::array<BlendFunction, 4>* value) const noexcept {
-    uint32_t blendFunctions = 0;
-    bool const result =  mImpl.getFromSimpleChunk(ChunkType::MaterialBlendFunction, &blendFunctions);
-    (*value)[0] = BlendFunction((blendFunctions >> 24) & 0xFF);
-    (*value)[1] = BlendFunction((blendFunctions >> 16) & 0xFF);
-    (*value)[2] = BlendFunction((blendFunctions >>  8) & 0xFF);
-    (*value)[3] = BlendFunction((blendFunctions >>  0) & 0xFF);
-    return result;
-}
-
 bool MaterialParser::getMaskThreshold(float* value) const noexcept {
    return mImpl.getFromSimpleChunk(ChunkType::MaterialMaskThreshold, value);
 }
@@ -29,12 +29,10 @@
 #include <backend/DriverEnums.h>
 #include <backend/Program.h>

-#include <utils/CString.h>
-#include <utils/FixedCapacityVector.h>
 #include <utils/compiler.h>
+#include <utils/CString.h>

 #include <inttypes.h>
-#include <utility>

 namespace filaflat {
 class ChunkContainer;
@@ -50,8 +48,7 @@ struct MaterialConstant;

 class MaterialParser {
 public:
-    MaterialParser(utils::FixedCapacityVector<backend::ShaderLanguage> preferredLanguages,
-            const void* data, size_t size);
+    MaterialParser(backend::ShaderLanguage language, const void* data, size_t size);

    MaterialParser(MaterialParser const& rhs) noexcept = delete;
    MaterialParser& operator=(MaterialParser const& rhs) noexcept = delete;
@@ -63,7 +60,6 @@ public:
    };

    ParseResult parse() noexcept;
-    backend::ShaderLanguage getShaderLanguage() const noexcept;

    // Accessors
    bool getMaterialVersion(uint32_t* value) const noexcept;
@@ -106,7 +102,6 @@ public:

    bool getShading(Shading*) const noexcept;
    bool getBlendingMode(BlendingMode*) const noexcept;
-    bool getCustomBlendFunction(std::array<backend::BlendFunction, 4>*) const noexcept;
    bool getMaskThreshold(float*) const noexcept;
    bool getAlphaToCoverageSet(bool*) const noexcept;
    bool getAlphaToCoverage(bool*) const noexcept;
@@ -134,9 +129,7 @@ public:

 private:
    struct MaterialParserDetails {
-        MaterialParserDetails(
-                const utils::FixedCapacityVector<backend::ShaderLanguage>& preferredLanguages,
-                const void* data, size_t size);
+        MaterialParserDetails(backend::ShaderLanguage language, const void* data, size_t size);

        template<typename T>
        bool getFromSimpleChunk(filamat::ChunkType type, T* value) const noexcept;
@@ -163,12 +156,12 @@ private:

        ManagedBuffer mManagedBuffer;
        filaflat::ChunkContainer mChunkContainer;
-        utils::FixedCapacityVector<backend::ShaderLanguage> mPreferredLanguages;
-        backend::ShaderLanguage mChosenLanguage;

        // Keep MaterialChunk alive between calls to getShader to avoid reload the shader index.
        filaflat::MaterialChunk mMaterialChunk;
        filaflat::BlobDictionary mBlobDictionary;
+        filamat::ChunkType mMaterialTag = filamat::ChunkType::Unknown;
+        filamat::ChunkType mDictionaryTag = filamat::ChunkType::Unknown;
    };

    filaflat::ChunkContainer& getChunkContainer() noexcept;
@@ -450,8 +450,11 @@ PostProcessManager::StructurePassOutput PostProcessManager::structure(FrameGraph
                        .levels = uint8_t(levelCount),
                        .format = isES2 ? TextureFormat::DEPTH24 : TextureFormat::DEPTH32F });

+                // workaround: since we have levels, this implies SAMPLEABLE (because of the gl
+                // backend, which implements non-sampleables with renderbuffers, which don't have levels).
+                // (should the gl driver revert to textures, in that case?)
                data.depth = builder.write(data.depth,
-                        FrameGraphTexture::Usage::DEPTH_ATTACHMENT);
+                        FrameGraphTexture::Usage::DEPTH_ATTACHMENT | FrameGraphTexture::Usage::SAMPLEABLE);

                if (config.picking) {
                    data.picking = builder.createTexture("Picking Buffer", {
@@ -16,35 +16,17 @@

 #include "RendererUtils.h"

-#include "PostProcessManager.h"
-
 #include "details/Engine.h"
 #include "details/View.h"

 #include "fg/FrameGraph.h"
 #include "fg/FrameGraphId.h"
 #include "fg/FrameGraphResources.h"
-#include "fg/FrameGraphTexture.h"

-#include <filament/Options.h>
-#include <filament/RenderableManager.h>
-#include <filament/Viewport.h>
-
-#include <backend/DriverEnums.h>
-#include <backend/Handle.h>
-#include <backend/PixelBufferDescriptor.h>
-
-#include <utils/BitmaskEnum.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/Panic.h>

-#include <algorithm>
-#include <utility>
-
-#include <stddef.h>
-#include <stdint.h>
-
 namespace filament {

 using namespace backend;
@@ -171,9 +153,16 @@ FrameGraphId<FrameGraphTexture> RendererUtils::colorPass(
                data.depth = builder.read(data.depth, FrameGraphTexture::Usage::DEPTH_ATTACHMENT);

                data.color = builder.write(data.color, FrameGraphTexture::Usage::COLOR_ATTACHMENT);
-                data.depth = builder.write(data.depth, FrameGraphTexture::Usage::DEPTH_ATTACHMENT);
                if (engine.getConfig().stereoscopicType == StereoscopicType::MULTIVIEW) {
+                    // Add sampleable usage flag for depth in multiview rendering, othewise it's
+                    // treated as renderbuffer in the backend and crashed.
+                    data.depth = builder.write(data.depth,
+                        FrameGraphTexture::Usage::DEPTH_ATTACHMENT |
+                        FrameGraphTexture::Usage::SAMPLEABLE);
                    layerCount = engine.getConfig().stereoscopicEyeCount;
+                } else {
+                    data.depth = builder.write(data.depth,
+                        FrameGraphTexture::Usage::DEPTH_ATTACHMENT);
                }

                /*
@@ -63,7 +63,8 @@ namespace filament {
 using namespace backend;
 using namespace math;

-ShadowMapManager::ShadowMapManager(FEngine& engine) {
+ShadowMapManager::ShadowMapManager(FEngine& engine)
+        : mEngine(engine) {
    FDebugRegistry& debugRegistry = engine.getDebugRegistry();
    debugRegistry.registerProperty("d.shadowmap.visualize_cascades",
            &engine.debug.shadowmap.visualize_cascades);
@@ -272,26 +273,9 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::render(FEngine& engine, FrameG
                if (!spotShadowCastersRange.empty()) {
                    for (auto& shadowMap : getSpotShadowMaps()) {
                        assert_invariant(!shadowMap.isDirectionalShadow());
-
-                        switch (shadowMap.getShadowType()) {
-                            case ShadowType::DIRECTIONAL:
-                                // we should never be here
-                                break;
-                            case ShadowType::SPOT:
-                                prepareSpotShadowMap(shadowMap, engine, view, mainCameraInfo,
-                                        scene->getLightData(), mSceneInfo);
-                                break;
-                            case ShadowType::POINT:
-                                preparePointShadowMap(shadowMap, engine, view, mainCameraInfo,
-                                        scene->getLightData());
-                                break;
-                        }
-
-                        if (shadowMap.hasVisibleShadows()) {
-                            passList.push_back({
-                                    {}, &shadowMap, spotShadowCastersRange,
-                                    VISIBLE_DYN_SHADOW_RENDERABLE });
-                        }
+                        passList.push_back({
+                                {}, &shadowMap, spotShadowCastersRange,
+                                VISIBLE_DYN_SHADOW_RENDERABLE });
                    }
                }

@@ -305,7 +289,6 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::render(FEngine& engine, FrameG
                scene, mainCameraInfo, userTime, passBuilder = passBuilder](
                    FrameGraphResources const&, auto const& data, DriverApi& driver) mutable {

-
                // Note: we could almost parallel_for the loop below, the problem currently is
                // that updatePrimitivesLod() updates temporary global state.
                // prepareSpotShadowMap() also update the visibility of renderable. These two
@@ -314,54 +297,53 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::render(FEngine& engine, FrameG

                // Generate a RenderPass for each shadow map
                for (auto const& entry : data.passList) {
-                    ShadowMap const& shadowMap = *entry.shadowMap;
-                    assert_invariant(shadowMap.hasVisibleShadows());
-
-                    // Note: this loop can generate a lot of commands that come out of the
-                    //       "per frame command arena". The allocation persists until the
-                    //       end of the frame.
-                    //       One way to possibly mitigate this, would be to always use the
-                    //       same command buffer for all shadow map, but then we'd generate
-                    //       a lot of unneeded draw calls.
-                    //       To do this efficiently, we'd need a way to cull draw calls already
-                    //       recorded in the command buffer, per shadow map.
-
-                    // Note: the output of culling below is stored in scene->getRenderableData()
+                    ShadowMap& shadowMap = *entry.shadowMap;

+                    // for spot shadow map, we need to do the culling
                    switch (shadowMap.getShadowType()) {
                        case ShadowType::DIRECTIONAL:
-                            // we should never be here
                            break;
                        case ShadowType::SPOT:
-                            ShadowMapManager::cullSpotShadowMap(shadowMap, engine, view,
+                            prepareSpotShadowMap(shadowMap, engine, view, mainCameraInfo,
                                    scene->getRenderableData(), entry.range,
-                                    scene->getLightData());
+                                    scene->getLightData(), mSceneInfo);
                            break;
                        case ShadowType::POINT:
-                            ShadowMapManager::cullPointShadowMap(shadowMap, view,
+                            preparePointShadowMap(shadowMap, engine, view, mainCameraInfo,
                                    scene->getRenderableData(), entry.range,
-                                    scene->getLightData());
+                                    scene->getLightData(), mSceneInfo);
                            break;
                    }

-                    // cameraInfo only valid after calling update
-                    const CameraInfo cameraInfo{ shadowMap.getCamera(), mainCameraInfo };
+                    if (shadowMap.hasVisibleShadows()) {
+                        // Note: this loop can generate a lot of commands that come out of the
+                        //       "per frame command arena". The allocation persists until the
+                        //       end of the frame.
+                        //       One way to possibly mitigate this, would be to always use the
+                        //       same command buffer for all shadow map, but then we'd generate
+                        //       a lot of unneeded draw calls.
+                        //       To do this efficiently, we'd need a way to cull draw calls already
+                        //       recorded in the command buffer, per shadow map.

-                    auto transaction = ShadowMap::open(driver);
-                    ShadowMap::prepareCamera(transaction, engine, cameraInfo);
-                    ShadowMap::prepareViewport(transaction, shadowMap.getViewport());
-                    ShadowMap::prepareTime(transaction, engine, userTime);
-                    ShadowMap::prepareShadowMapping(transaction,
-                            vsmShadowOptions.highPrecision);
-                    shadowMap.commit(transaction, driver);

-                    // updatePrimitivesLod must be run before RenderPass::appendCommands.
-                    view.updatePrimitivesLod(engine,
-                            cameraInfo, scene->getRenderableData(), entry.range);
+                        // cameraInfo only valid after calling update
+                        const CameraInfo cameraInfo{ shadowMap.getCamera(), mainCameraInfo };

-                    // generate and sort the commands for rendering the shadow map
+                        auto transaction = ShadowMap::open(driver);
+                        ShadowMap::prepareCamera(transaction, engine, cameraInfo);
+                        ShadowMap::prepareViewport(transaction, shadowMap.getViewport());
+                        ShadowMap::prepareTime(transaction, engine, userTime);
+                        ShadowMap::prepareShadowMapping(transaction,
+                                vsmShadowOptions.highPrecision);
+                        shadowMap.commit(transaction, driver);

-                    RenderPass const pass = passBuilder
+                        // updatePrimitivesLod must be run before RenderPass::appendCommands.
+                        view.updatePrimitivesLod(engine,
+                                cameraInfo, scene->getRenderableData(), entry.range);
+
+                        // generate and sort the commands for rendering the shadow map
+
+                        RenderPass const pass = passBuilder
                            .camera(cameraInfo)
                            .visibilityMask(entry.visibilityMask)
                            .geometry(scene->getRenderableData(),
@@ -369,15 +351,16 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::render(FEngine& engine, FrameG
                            .commandTypeFlags(RenderPass::CommandTypeFlags::SHADOW)
                            .build(engine);

-                    entry.executor = pass.getExecutor();
+                        entry.executor = pass.getExecutor();

-                    if (!view.hasVSM()) {
-                        auto const* options = shadowMap.getShadowOptions();
-                        PolygonOffset const polygonOffset = { // handle reversed Z
-                                .slope    = -options->polygonOffsetSlope,
-                                .constant = -options->polygonOffsetConstant
-                        };
-                        entry.executor.overridePolygonOffset(&polygonOffset);
+                        if (!view.hasVSM()) {
+                            auto const* options = shadowMap.getShadowOptions();
+                            PolygonOffset const polygonOffset = { // handle reversed Z
+                                    .slope    = -options->polygonOffsetSlope,
+                                    .constant = -options->polygonOffsetConstant
+                            };
+                            entry.executor.overridePolygonOffset(&polygonOffset);
+                        }
                    }
                }

@@ -402,8 +385,10 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::render(FEngine& engine, FrameG

    auto const& passList = prepareShadowPass.getData().passList;
    for (auto const& entry: passList) {
-        assert_invariant(entry.shadowMap->hasVisibleShadows());
-
+        if (!entry.shadowMap->hasVisibleShadows()) {
+            continue;
+        }
+        
        const uint8_t layer = entry.shadowMap->getLayer();
        const auto* options = entry.shadowMap->getShadowOptions();
        const auto msaaSamples = textureRequirements.msaaSamples;
@@ -630,7 +615,7 @@ ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEng
        // Compute scene-dependent values shared across all cascades
        ShadowMap::updateSceneInfoDirectional(MvAtOrigin, *scene, sceneInfo);

-        shadowMap.updateDirectional(engine,
+        shadowMap.updateDirectional(mEngine,
                lightData, 0, cameraInfo, shadowMapInfo, sceneInfo);

        hasVisibleShadows = shadowMap.hasVisibleShadows();
@@ -696,7 +681,7 @@ ShadowMapManager::ShadowTechnique ShadowMapManager::updateCascadeShadowMaps(FEng

            sceneInfo.csNearFar = { csSplitPosition[i], csSplitPosition[i + 1] };

-            auto shaderParameters = shadowMap.updateDirectional(engine,
+            auto shaderParameters = shadowMap.updateDirectional(mEngine,
                    lightData, 0, cameraInfo, shadowMapInfo, sceneInfo);

            if (shadowMap.hasVisibleShadows()) {
@@ -772,58 +757,15 @@ void ShadowMapManager::updateSpotVisibilityMasks(
    }
 }

-void ShadowMapManager::prepareSpotShadowMap(ShadowMap& shadowMap, FEngine& engine, FView& view,
-        CameraInfo const& mainCameraInfo,
-        FScene::LightSoa& lightData, ShadowMap::SceneInfo const& sceneInfo) noexcept {
-
-    const size_t lightIndex = shadowMap.getLightIndex();
-    FLightManager::ShadowOptions const* const options = shadowMap.getShadowOptions();
-
-    // update the shadow map frustum/camera
-    const ShadowMap::ShadowMapInfo shadowMapInfo{
-            .atlasDimension      = mTextureAtlasRequirements.size,
-            .textureDimension    = uint16_t(options->mapSize),
-            .shadowDimension     = uint16_t(options->mapSize - 2u),
-            .textureSpaceFlipped = engine.getBackend() == Backend::METAL ||
-                                   engine.getBackend() == Backend::VULKAN,
-            .vsm                 = view.hasVSM()
-    };
-
-    auto shaderParameters = shadowMap.updateSpot(engine,
-            lightData, lightIndex, mainCameraInfo, shadowMapInfo, *view.getScene(), sceneInfo);
-
-    // and if we need to generate it, update all the UBO data
-    if (shadowMap.hasVisibleShadows()) {
-        const size_t shadowIndex = shadowMap.getShadowIndex();
-        const float wsTexelSizeAtOneMeter = shaderParameters.texelSizeAtOneMeterWs;
-        // note: normalBias is set to zero for VSM
-        const float normalBias = shadowMapInfo.vsm ? 0.0f : options->normalBias;
-
-        auto& s = mShadowUb.edit();
-        const double n = shadowMap.getCamera().getNear();
-        const double f = shadowMap.getCamera().getCullingFar();
-        s.shadows[shadowIndex].layer = shadowMap.getLayer();
-        s.shadows[shadowIndex].lightFromWorldMatrix = shaderParameters.lightSpace;
-        s.shadows[shadowIndex].scissorNormalized = shaderParameters.scissorNormalized;
-        s.shadows[shadowIndex].normalBias = normalBias * wsTexelSizeAtOneMeter;
-        s.shadows[shadowIndex].lightFromWorldZ = shaderParameters.lightFromWorldZ;
-        s.shadows[shadowIndex].texelSizeAtOneMeter = wsTexelSizeAtOneMeter;
-        s.shadows[shadowIndex].nearOverFarMinusNear = float(n / (f - n));
-        s.shadows[shadowIndex].elvsm = options->vsm.elvsm;
-        s.shadows[shadowIndex].bulbRadiusLs =
-                mSoftShadowOptions.penumbraScale * options->shadowBulbRadius
-                        / wsTexelSizeAtOneMeter;
-
-    }
-}
-
-void ShadowMapManager::cullSpotShadowMap(ShadowMap const& shadowMap, FEngine& engine, FView& view,
+void ShadowMapManager::prepareSpotShadowMap(ShadowMap& shadowMap,
+        FEngine& engine, FView& view, CameraInfo const& mainCameraInfo,
        FScene::RenderableSoa& renderableData, utils::Range<uint32_t> range,
-        FScene::LightSoa& lightData) noexcept {
+        FScene::LightSoa& lightData, ShadowMap::SceneInfo const& sceneInfo) noexcept {
    auto& lcm = engine.getLightManager();

    const size_t lightIndex = shadowMap.getLightIndex();
    const FLightManager::Instance li = lightData.elementAt<FScene::LIGHT_INSTANCE>(lightIndex);
+    FLightManager::ShadowOptions const* const options = shadowMap.getShadowOptions();

    // compute the frustum for this light
    // for spotlights, we cull shadow casters first because we already know the frustum,
@@ -860,28 +802,19 @@ void ShadowMapManager::cullSpotShadowMap(ShadowMap const& shadowMap, FEngine& en
            visibility + range.first,
            visibleArray + range.first,
            range.size());
-}
-
-void ShadowMapManager::preparePointShadowMap(ShadowMap& shadowMap,
-        FEngine& engine, FView& view, CameraInfo const& mainCameraInfo,
-        FScene::LightSoa& lightData) noexcept {
-
-    const uint8_t face = shadowMap.getFace();
-    const size_t lightIndex = shadowMap.getLightIndex();
-    FLightManager::ShadowOptions const* const options = shadowMap.getShadowOptions();

    // update the shadow map frustum/camera
    const ShadowMap::ShadowMapInfo shadowMapInfo{
            .atlasDimension      = mTextureAtlasRequirements.size,
            .textureDimension    = uint16_t(options->mapSize),
-            .shadowDimension     = uint16_t(options->mapSize), // point-lights don't have a border
+            .shadowDimension     = uint16_t(options->mapSize - 2u),
            .textureSpaceFlipped = engine.getBackend() == Backend::METAL ||
                                   engine.getBackend() == Backend::VULKAN,
            .vsm                 = view.hasVSM()
    };

-    auto shaderParameters = shadowMap.updatePoint(engine, lightData, lightIndex,
-            mainCameraInfo, shadowMapInfo, *view.getScene(), face);
+    auto shaderParameters = shadowMap.updateSpot(mEngine,
+            lightData, lightIndex, mainCameraInfo, shadowMapInfo, *view.getScene(), sceneInfo);

    // and if we need to generate it, update all the UBO data
    if (shadowMap.hasVisibleShadows()) {
@@ -904,15 +837,19 @@ void ShadowMapManager::preparePointShadowMap(ShadowMap& shadowMap,
        s.shadows[shadowIndex].bulbRadiusLs =
                mSoftShadowOptions.penumbraScale * options->shadowBulbRadius
                        / wsTexelSizeAtOneMeter;
+
    }
 }

-void ShadowMapManager::cullPointShadowMap(ShadowMap const& shadowMap, FView& view,
+void ShadowMapManager::preparePointShadowMap(ShadowMap& shadowMap,
+        FEngine& engine, FView& view, CameraInfo const& mainCameraInfo,
        FScene::RenderableSoa& renderableData, utils::Range<uint32_t> range,
-        FScene::LightSoa& lightData) noexcept {
+        FScene::LightSoa& lightData,
+        ShadowMap::SceneInfo const&) noexcept {

    const uint8_t face = shadowMap.getFace();
    const size_t lightIndex = shadowMap.getLightIndex();
+    FLightManager::ShadowOptions const* const options = shadowMap.getShadowOptions();

    // compute the frustum for this light
    // for spotlights, we cull shadow casters first because we already know the frustum,
@@ -946,6 +883,43 @@ void ShadowMapManager::cullPointShadowMap(ShadowMap const& shadowMap, FView& vie
            visibility + range.first,
            visibleArray + range.first,
            range.size());
+
+    // update the shadow map frustum/camera
+    const ShadowMap::ShadowMapInfo shadowMapInfo{
+            .atlasDimension      = mTextureAtlasRequirements.size,
+            .textureDimension    = uint16_t(options->mapSize),
+            .shadowDimension     = uint16_t(options->mapSize), // point-lights don't have a border
+            .textureSpaceFlipped = engine.getBackend() == Backend::METAL ||
+                                   engine.getBackend() == Backend::VULKAN,
+            .vsm                 = view.hasVSM()
+    };
+
+    auto shaderParameters = shadowMap.updatePoint(mEngine, lightData, lightIndex,
+            mainCameraInfo, shadowMapInfo, *view.getScene(), face);
+
+
+    // and if we need to generate it, update all the UBO data
+    if (shadowMap.hasVisibleShadows()) {
+        const size_t shadowIndex = shadowMap.getShadowIndex();
+        const float wsTexelSizeAtOneMeter = shaderParameters.texelSizeAtOneMeterWs;
+        // note: normalBias is set to zero for VSM
+        const float normalBias = shadowMapInfo.vsm ? 0.0f : options->normalBias;
+
+        auto& s = mShadowUb.edit();
+        const double n = shadowMap.getCamera().getNear();
+        const double f = shadowMap.getCamera().getCullingFar();
+        s.shadows[shadowIndex].layer = shadowMap.getLayer();
+        s.shadows[shadowIndex].lightFromWorldMatrix = shaderParameters.lightSpace;
+        s.shadows[shadowIndex].scissorNormalized = shaderParameters.scissorNormalized;
+        s.shadows[shadowIndex].normalBias = normalBias * wsTexelSizeAtOneMeter;
+        s.shadows[shadowIndex].lightFromWorldZ = shaderParameters.lightFromWorldZ;
+        s.shadows[shadowIndex].texelSizeAtOneMeter = wsTexelSizeAtOneMeter;
+        s.shadows[shadowIndex].nearOverFarMinusNear = float(n / (f - n));
+        s.shadows[shadowIndex].elvsm = options->vsm.elvsm;
+        s.shadows[shadowIndex].bulbRadiusLs =
+                mSoftShadowOptions.penumbraScale * options->shadowBulbRadius
+                        / wsTexelSizeAtOneMeter;
+    }
 }

 ShadowMapManager::ShadowTechnique ShadowMapManager::updateSpotShadowMaps(FEngine& engine,
@@ -160,20 +160,14 @@ private:

    void prepareSpotShadowMap(ShadowMap& shadowMap,
            FEngine& engine, FView& view, CameraInfo const& mainCameraInfo,
-            FScene::LightSoa& lightData, ShadowMap::SceneInfo const& sceneInfo) noexcept;
-
-    static void cullSpotShadowMap(ShadowMap const& map,
-            FEngine& engine, FView& view,
            FScene::RenderableSoa& renderableData, utils::Range<uint32_t> range,
-            FScene::LightSoa& lightData) noexcept;
+            FScene::LightSoa& lightData, ShadowMap::SceneInfo const& sceneInfo) noexcept;

    void preparePointShadowMap(ShadowMap& map,
            FEngine& engine, FView& view, CameraInfo const& mainCameraInfo,
-            FScene::LightSoa& lightData) noexcept;
-
-    static void cullPointShadowMap(ShadowMap const& shadowMap, FView& view,
            FScene::RenderableSoa& renderableData, utils::Range<uint32_t> range,
-            FScene::LightSoa& lightData) noexcept;
+            FScene::LightSoa& lightData,
+            ShadowMap::SceneInfo const& sceneInfo) noexcept;

    static void updateSpotVisibilityMasks(
            uint8_t visibleLayers,
@@ -209,6 +203,8 @@ private:
        size_t mSplitCount;
    };

+    FEngine& mEngine;
+
    // Atlas requirements, updated in ShadowMapManager::update(),
    // consumed in ShadowMapManager::render()
    struct TextureAtlasRequirements {
@@ -682,7 +682,18 @@ int FEngine::loop() {
        return 0;
    }

+    // Set thread affinity for the backend thread.
+    //  see https://developer.android.com/agi/sys-trace/threads-scheduling#cpu_core_affinity
+    // Certain backends already have some threads pinned, and we can't easily know on which core.
+    const bool disableThreadAffinity
+            = mDriver->isWorkaroundNeeded(Workaround::DISABLE_THREAD_AFFINITY);
+
+    uint32_t const id = std::thread::hardware_concurrency() - 1;
    while (true) {
+        // looks like thread affinity needs to be reset regularly (on Android)
+        if (!disableThreadAffinity) {
+            JobSystem::setThreadAffinityById(id);
+        }
        if (!execute()) {
            break;
        }
@@ -715,10 +726,9 @@ const FMaterial* FEngine::getSkyboxMaterial() const noexcept {
 * Object created from a Builder
 */

-template<typename T, typename ... ARGS>
-inline T* FEngine::create(ResourceList<T>& list,
-        typename T::Builder const& builder, ARGS&& ... args) noexcept {
-    T* p = mHeapAllocator.make<T>(*this, builder, std::forward<ARGS>(args)...);
+template<typename T>
+inline T* FEngine::create(ResourceList<T>& list, typename T::Builder const& builder) noexcept {
+    T* p = mHeapAllocator.make<T>(*this, builder);
    if (UTILS_UNLIKELY(p)) { // this should never happen
        list.insert(p);
    }
@@ -757,9 +767,8 @@ FIndirectLight* FEngine::createIndirectLight(const IndirectLight::Builder& build
    return create(mIndirectLights, builder);
 }

-FMaterial* FEngine::createMaterial(const Material::Builder& builder,
-        std::unique_ptr<MaterialParser> materialParser) noexcept {
-    return create(mMaterials, builder, std::move(materialParser));
+FMaterial* FEngine::createMaterial(const Material::Builder& builder) noexcept {
+    return create(mMaterials, builder);
 }

 FSkybox* FEngine::createSkybox(const Skybox::Builder& builder) noexcept {
@@ -59,11 +59,10 @@
 #include <filament/Texture.h>
 #include <filament/VertexBuffer.h>

-#include <utils/Allocator.h>
-#include <utils/CountDownLatch.h>
-#include <utils/FixedCapacityVector.h>
-#include <utils/JobSystem.h>
 #include <utils/compiler.h>
+#include <utils/Allocator.h>
+#include <utils/JobSystem.h>
+#include <utils/CountDownLatch.h>

 #include <chrono>
 #include <memory>
@@ -232,21 +231,19 @@ public:
        return mPlatform;
    }

-    // Return a vector of shader languages, in order of preference.
-    utils::FixedCapacityVector<backend::ShaderLanguage> getShaderLanguage() const noexcept {
+    backend::ShaderLanguage getShaderLanguage() const noexcept {
        switch (mBackend) {
            case Backend::DEFAULT:
            case Backend::NOOP:
            default:
-                return { backend::ShaderLanguage::ESSL3 };
+                return backend::ShaderLanguage::ESSL3;
            case Backend::OPENGL:
-                return { getDriver().getFeatureLevel() == FeatureLevel::FEATURE_LEVEL_0
-                                ? backend::ShaderLanguage::ESSL1
-                                : backend::ShaderLanguage::ESSL3 };
+                return getDriver().getFeatureLevel() == FeatureLevel::FEATURE_LEVEL_0
+                        ? backend::ShaderLanguage::ESSL1 : backend::ShaderLanguage::ESSL3;
            case Backend::VULKAN:
-                return { backend::ShaderLanguage::SPIRV };
+                return backend::ShaderLanguage::SPIRV;
            case Backend::METAL:
-                return { backend::ShaderLanguage::METAL_LIBRARY, backend::ShaderLanguage::MSL };
+                return backend::ShaderLanguage::MSL;
        }
    }

@@ -266,8 +263,8 @@ public:
        return mDefaultRenderTarget;
    }

-    template <typename T, typename ... ARGS>
-    T* create(ResourceList<T>& list, typename T::Builder const& builder, ARGS&& ... args) noexcept;
+    template <typename T>
+    T* create(ResourceList<T>& list, typename T::Builder const& builder) noexcept;

    FBufferObject* createBufferObject(const BufferObject::Builder& builder) noexcept;
    FVertexBuffer* createVertexBuffer(const VertexBuffer::Builder& builder) noexcept;
@@ -276,7 +273,7 @@ public:
    FMorphTargetBuffer* createMorphTargetBuffer(const MorphTargetBuffer::Builder& builder) noexcept;
    FInstanceBuffer* createInstanceBuffer(const InstanceBuffer::Builder& builder) noexcept;
    FIndirectLight* createIndirectLight(const IndirectLight::Builder& builder) noexcept;
-    FMaterial* createMaterial(const Material::Builder& builder, std::unique_ptr<MaterialParser> materialParser) noexcept;
+    FMaterial* createMaterial(const Material::Builder& builder) noexcept;
    FTexture* createTexture(const Texture::Builder& builder) noexcept;
    FSkybox* createSkybox(const Skybox::Builder& builder) noexcept;
    FColorGrading* createColorGrading(const ColorGrading::Builder& builder) noexcept;
@@ -15,59 +15,29 @@
 */

 #include "details/Material.h"
-#include "details/Engine.h"

 #include "Froxelizer.h"
 #include "MaterialParser.h"

+#include "details/Engine.h"
+
 #include "FilamentAPI-impl.h"

+#include <private/filament/Variant.h>
+
 #include <private/filament/EngineEnums.h>
 #include <private/filament/SamplerInterfaceBlock.h>
 #include <private/filament/BufferInterfaceBlock.h>
-#include <private/filament/Variant.h>
-
-#include <filament/Material.h>
-#include <filament/MaterialEnums.h>
-
-#if FILAMENT_ENABLE_MATDBG
-#include <matdbg/DebugServer.h>
-#endif
-
-#include <filaflat/ChunkContainer.h>
-#include <filaflat/MaterialChunk.h>

 #include <backend/DriverEnums.h>
-#include <backend/CallbackHandler.h>
 #include <backend/Program.h>

-#include <utils/BitmaskEnum.h>
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
-#include <utils/Hash.h>
-#include <utils/Invocable.h>
-#include <utils/Log.h>
 #include <utils/Panic.h>
-#include <utils/bitset.h>
-#include <utils/compiler.h>
-#include <utils/debug.h>
-#include <utils/ostream.h>
+#include <utils/Hash.h>

-#include <algorithm>
-#include <array>
-#include <iterator>
-#include <memory>
-#include <mutex>
-#include <new>
-#include <optional>
-#include <sstream>
-#include <string_view>
 #include <unordered_map>
-#include <utility>
-#include <variant>
-
-#include <stddef.h>
-#include <stdint.h>

 namespace filament {

@@ -75,32 +45,22 @@ using namespace backend;
 using namespace filaflat;
 using namespace utils;

-static std::unique_ptr<MaterialParser> createParser(Backend backend,
-        utils::FixedCapacityVector<ShaderLanguage> languages, const void* data, size_t size) {
+static MaterialParser* createParser(Backend backend, ShaderLanguage language,
+                                    const void* data, size_t size) {
    // unique_ptr so we don't leak MaterialParser on failures below
-    auto materialParser = std::make_unique<MaterialParser>(languages, data, size);
+    auto materialParser = std::make_unique<MaterialParser>(language, data, size);

    MaterialParser::ParseResult const materialResult = materialParser->parse();

-    if (UTILS_UNLIKELY(materialResult == MaterialParser::ParseResult::ERROR_MISSING_BACKEND)) {
-        std::stringstream languageNames;
-        for (auto it = languages.begin(); it != languages.end(); ++it) {
-            languageNames << shaderLanguageToString(*it);
-            if (std::next(it) != languages.end()) {
-                languageNames << ", ";
-            }
-        }
-
-        ASSERT_PRECONDITION(materialResult != MaterialParser::ParseResult::ERROR_MISSING_BACKEND,
-                "the material was not built for any of the %s backend's supported shader "
-                "languages (%s)\n",
-                backendToString(backend), languageNames.str().c_str());
-    }
-
    if (backend == Backend::NOOP) {
-        return materialParser;
+        return materialParser.release();
    }

+    ASSERT_PRECONDITION(materialResult != MaterialParser::ParseResult::ERROR_MISSING_BACKEND,
+                "the material was not built for the %s backend and %s shader language\n",
+                backendToString(backend),
+                shaderLanguageToString(language));
+
    ASSERT_PRECONDITION(materialResult == MaterialParser::ParseResult::SUCCESS,
                "could not parse the material package");

@@ -111,12 +71,13 @@ static std::unique_ptr<MaterialParser> createParser(Backend backend,

    assert_invariant(backend != Backend::DEFAULT && "Default backend has not been resolved.");

-    return materialParser;
+    return materialParser.release();
 }

 struct Material::BuilderDetails {
    const void* mPayload = nullptr;
    size_t mSize = 0;
+    MaterialParser* mMaterialParser = nullptr;
    bool mDefaultMaterial = false;
    std::unordered_map<
        utils::CString,
@@ -154,11 +115,11 @@ template Material::Builder& Material::Builder::constant<float>(const char*, size
 template Material::Builder& Material::Builder::constant<bool>(const char*, size_t, bool);

 Material* Material::Builder::build(Engine& engine) {
-    std::unique_ptr<MaterialParser> materialParser = createParser(
+    std::unique_ptr<MaterialParser> materialParser{ createParser(
        downcast(engine).getBackend(), downcast(engine).getShaderLanguage(),
-        mImpl->mPayload, mImpl->mSize);
+        mImpl->mPayload, mImpl->mSize) };

-    if (!materialParser) {
+    if (materialParser == nullptr) {
        return nullptr;
    }

@@ -185,17 +146,17 @@ Material* Material::Builder::build(Engine& engine) {
        return nullptr;
    }

-    return downcast(engine).createMaterial(*this, std::move(materialParser));
+    mImpl->mMaterialParser = materialParser.release();
+
+    return downcast(engine).createMaterial(*this);
 }

-FMaterial::FMaterial(FEngine& engine, const Material::Builder& builder,
-        std::unique_ptr<MaterialParser> materialParser)
-        : mIsDefaultMaterial(builder->mDefaultMaterial),
-          mEngine(engine),
-          mMaterialId(engine.getMaterialId()),
-          mMaterialParser(std::move(materialParser))
+FMaterial::FMaterial(FEngine& engine, const Material::Builder& builder)
+        : mEngine(engine),
+          mMaterialId(engine.getMaterialId())
 {
-    MaterialParser* const parser = mMaterialParser.get();
+    MaterialParser* parser = builder->mMaterialParser;
+    mMaterialParser = parser;

    UTILS_UNUSED_IN_RELEASE bool const nameOk = parser->getName(&mName);
    assert_invariant(nameOk);
@@ -228,7 +189,7 @@ FMaterial::FMaterial(FEngine& engine, const Material::Builder& builder,
    success = parser->getUIB(&mUniformInterfaceBlock);
    assert_invariant(success);

-    if (UTILS_UNLIKELY(parser->getShaderLanguage() == ShaderLanguage::ESSL1)) {
+    if (UTILS_UNLIKELY(engine.getShaderLanguage() == ShaderLanguage::ESSL1)) {
        success = parser->getBindingUniformInfo(&mBindingUniformInfo);
        assert_invariant(success);

@@ -244,66 +205,6 @@ FMaterial::FMaterial(FEngine& engine, const Material::Builder& builder,
            &mSamplerGroupBindingInfoList, &mSamplerBindingToNameMap);
    assert_invariant(success);

-    // Older materials will not have a subpass chunk; this should not be an error.
-    if (!parser->getSubpasses(&mSubpassInfo)) {
-        mSubpassInfo.isValid = false;
-    }
-
-    parser->getShading(&mShading);
-    parser->getMaterialProperties(&mMaterialProperties);
-    parser->getInterpolation(&mInterpolation);
-    parser->getVertexDomain(&mVertexDomain);
-    parser->getMaterialDomain(&mMaterialDomain);
-    parser->getMaterialVariantFilterMask(&mVariantFilterMask);
-    parser->getRequiredAttributes(&mRequiredAttributes);
-    parser->getRefractionMode(&mRefractionMode);
-    parser->getRefractionType(&mRefractionType);
-    parser->getReflectionMode(&mReflectionMode);
-    parser->getTransparencyMode(&mTransparencyMode);
-    parser->getDoubleSided(&mDoubleSided);
-    parser->getCullingMode(&mCullingMode);
-
-    if (mShading == Shading::UNLIT) {
-        parser->hasShadowMultiplier(&mHasShadowMultiplier);
-    }
-
-    mIsVariantLit = mShading != Shading::UNLIT || mHasShadowMultiplier;
-
-    // color write
-    bool colorWrite = false;
-    parser->getColorWrite(&colorWrite);
-    mRasterState.colorWrite = colorWrite;
-
-    // depth test
-    bool depthTest = false;
-    parser->getDepthTest(&depthTest);
-    mRasterState.depthFunc = depthTest ? RasterState::DepthFunc::GE : RasterState::DepthFunc::A;
-
-    // if doubleSided() was called we override culling()
-    bool doubleSideSet = false;
-    parser->getDoubleSidedSet(&doubleSideSet);
-    if (doubleSideSet) {
-        mDoubleSidedCapability = true;
-        mRasterState.culling = mDoubleSided ? CullingMode::NONE : mCullingMode;
-    } else {
-        mRasterState.culling = mCullingMode;
-    }
-
-    // specular anti-aliasing
-    parser->hasSpecularAntiAliasing(&mSpecularAntiAliasing);
-    if (mSpecularAntiAliasing) {
-        parser->getSpecularAntiAliasingVariance(&mSpecularAntiAliasingVariance);
-        parser->getSpecularAntiAliasingThreshold(&mSpecularAntiAliasingThreshold);
-    }
-
-    processBlendingMode(parser);
-    processSpecializationConstants(engine, builder, parser);
-    processDepthVariants(engine, parser);
-
-    // we can only initialize the default instance once we're initialized ourselves
-    mDefaultInstance.initDefaultInstance(engine, this);
-
-
 #if FILAMENT_ENABLE_MATDBG
    // Register the material with matdbg.
    matdbg::DebugServer* server = downcast(engine).debug.server;
@@ -312,9 +213,259 @@ FMaterial::FMaterial(FEngine& engine, const Material::Builder& builder,
        mDebuggerId = server->addMaterial(mName, details->mPayload, details->mSize, this);
    }
 #endif
+
+    // Older materials will not have a subpass chunk; this should not be an error.
+    if (!parser->getSubpasses(&mSubpassInfo)) {
+        mSubpassInfo.isValid = false;
+    }
+
+    // Older materials won't have a constants chunk, but that's okay.
+    parser->getConstants(&mMaterialConstants);
+    for (size_t i = 0, c = mMaterialConstants.size(); i < c; i++) {
+        auto& item = mMaterialConstants[i];
+        // the key can be a string_view because mMaterialConstant owns the CString
+        std::string_view const key{ item.name.data(), item.name.size() };
+        mSpecializationConstantsNameToIndex[key] = i;
+    }
+
+    // Verify that all the constant specializations exist in the material and that their types match.
+    // The first specialization constants are defined internally by Filament.
+    // The subsequent constants are user-defined in the material.
+
+    // Feature level 0 doesn't support instancing
+    int const maxInstanceCount = (engine.getActiveFeatureLevel() == FeatureLevel::FEATURE_LEVEL_0)
+            ? 1 : CONFIG_MAX_INSTANCES;
+
+    int const maxFroxelBufferHeight = std::min(
+            FROXEL_BUFFER_MAX_ENTRY_COUNT / 4,
+            engine.getDriverApi().getMaxUniformBufferSize() / 16u);
+
+    bool const staticTextureWorkaround =
+            engine.getDriverApi().isWorkaroundNeeded(Workaround::A8X_STATIC_TEXTURE_TARGET_ERROR);
+
+    bool const powerVrShaderWorkarounds =
+            engine.getDriverApi().isWorkaroundNeeded(Workaround::POWER_VR_SHADER_WORKAROUNDS);
+
+    mSpecializationConstants.reserve(mMaterialConstants.size() + CONFIG_MAX_RESERVED_SPEC_CONSTANTS);
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::BACKEND_FEATURE_LEVEL,
+                    (int)engine.getSupportedFeatureLevel() });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_MAX_INSTANCES,
+                    (int)maxInstanceCount });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_FROXEL_BUFFER_HEIGHT,
+                    (int)maxFroxelBufferHeight });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_DEBUG_DIRECTIONAL_SHADOWMAP,
+                    engine.debug.shadowmap.debug_directional_shadowmap });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_DEBUG_FROXEL_VISUALIZATION,
+                    engine.debug.lighting.debug_froxel_visualization });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_STATIC_TEXTURE_TARGET_WORKAROUND,
+                    staticTextureWorkaround });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_POWER_VR_SHADER_WORKAROUNDS,
+                    powerVrShaderWorkarounds });
+    mSpecializationConstants.push_back({
+                    +ReservedSpecializationConstants::CONFIG_STEREO_EYE_COUNT,
+                    (int)engine.getConfig().stereoscopicEyeCount });
+    if (UTILS_UNLIKELY(engine.getShaderLanguage() == ShaderLanguage::ESSL1)) {
+        // The actual value of this spec-constant is set in the OpenGLDriver backend.
+        mSpecializationConstants.push_back({
+            +ReservedSpecializationConstants::CONFIG_SRGB_SWAPCHAIN_EMULATION,
+            false});
+    }
+
+    for (auto const& [name, value] : builder->mConstantSpecializations) {
+        std::string_view const key{ name.data(), name.size() };
+        auto pos = mSpecializationConstantsNameToIndex.find(key);
+        ASSERT_PRECONDITION(pos != mSpecializationConstantsNameToIndex.end(),
+                "The material %s does not have a constant parameter named %s.",
+                mName.c_str_safe(), name.c_str());
+        const char* const types[3] = {"an int", "a float", "a bool"};
+        const char* const errorMessage =
+                "The constant parameter %s on material %s is of type %s, but %s was "
+                "provided.";
+        auto& constant = mMaterialConstants[pos->second];
+        switch (constant.type) {
+            case ConstantType::INT:
+                ASSERT_PRECONDITION(std::holds_alternative<int32_t>(value), errorMessage,
+                        name.c_str(), mName.c_str_safe(), "int", types[value.index()]);
+                break;
+            case ConstantType::FLOAT:
+                ASSERT_PRECONDITION(std::holds_alternative<float>(value), errorMessage,
+                        name.c_str(), mName.c_str_safe(), "float", types[value.index()]);
+                break;
+            case ConstantType::BOOL:
+                ASSERT_PRECONDITION(std::holds_alternative<bool>(value), errorMessage,
+                        name.c_str(), mName.c_str_safe(), "bool", types[value.index()]);
+                break;
+        }
+        uint32_t const index = pos->second + CONFIG_MAX_RESERVED_SPEC_CONSTANTS;
+        mSpecializationConstants.push_back({ index, value });
+    }
+
+    parser->getShading(&mShading);
+    parser->getMaterialProperties(&mMaterialProperties);
+    parser->getBlendingMode(&mBlendingMode);
+    parser->getInterpolation(&mInterpolation);
+    parser->getVertexDomain(&mVertexDomain);
+    parser->getMaterialDomain(&mMaterialDomain);
+    parser->getMaterialVariantFilterMask(&mVariantFilterMask);
+    parser->getRequiredAttributes(&mRequiredAttributes);
+    parser->getRefractionMode(&mRefractionMode);
+    parser->getRefractionType(&mRefractionType);
+    parser->getReflectionMode(&mReflectionMode);
+
+    if (mBlendingMode == BlendingMode::MASKED) {
+        parser->getMaskThreshold(&mMaskThreshold);
+    }
+
+    // The fade blending mode only affects shading. For proper sorting we need to
+    // treat this blending mode as a regular transparent blending operation.
+    if (UTILS_UNLIKELY(mBlendingMode == BlendingMode::FADE)) {
+        mRenderBlendingMode = BlendingMode::TRANSPARENT;
+    } else {
+        mRenderBlendingMode = mBlendingMode;
+    }
+
+    if (mShading == Shading::UNLIT) {
+        parser->hasShadowMultiplier(&mHasShadowMultiplier);
+    }
+
+    mIsVariantLit = mShading != Shading::UNLIT || mHasShadowMultiplier;
+
+    // create raster state
+    using BlendFunction = RasterState::BlendFunction;
+    using DepthFunc = RasterState::DepthFunc;
+    switch (mBlendingMode) {
+        // Do not change the MASKED behavior without checking for regressions with
+        // AlphaBlendModeTest and TextureLinearInterpolationTest, with and without
+        // View::BlendMode::TRANSLUCENT.
+        case BlendingMode::MASKED:
+        case BlendingMode::OPAQUE:
+            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
+            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
+            mRasterState.blendFunctionDstRGB   = BlendFunction::ZERO;
+            mRasterState.blendFunctionDstAlpha = BlendFunction::ZERO;
+            mRasterState.depthWrite = true;
+            break;
+        case BlendingMode::TRANSPARENT:
+        case BlendingMode::FADE:
+            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
+            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
+            mRasterState.blendFunctionDstRGB   = BlendFunction::ONE_MINUS_SRC_ALPHA;
+            mRasterState.blendFunctionDstAlpha = BlendFunction::ONE_MINUS_SRC_ALPHA;
+            mRasterState.depthWrite = false;
+            break;
+        case BlendingMode::ADD:
+            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
+            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
+            mRasterState.blendFunctionDstRGB   = BlendFunction::ONE;
+            mRasterState.blendFunctionDstAlpha = BlendFunction::ONE;
+            mRasterState.depthWrite = false;
+            break;
+        case BlendingMode::MULTIPLY:
+            mRasterState.blendFunctionSrcRGB   = BlendFunction::ZERO;
+            mRasterState.blendFunctionSrcAlpha = BlendFunction::ZERO;
+            mRasterState.blendFunctionDstRGB   = BlendFunction::SRC_COLOR;
+            mRasterState.blendFunctionDstAlpha = BlendFunction::SRC_COLOR;
+            mRasterState.depthWrite = false;
+            break;
+        case BlendingMode::SCREEN:
+            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
+            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
+            mRasterState.blendFunctionDstRGB   = BlendFunction::ONE_MINUS_SRC_COLOR;
+            mRasterState.blendFunctionDstAlpha = BlendFunction::ONE_MINUS_SRC_COLOR;
+            mRasterState.depthWrite = false;
+            break;
+    }
+
+    bool depthWriteSet = false;
+    parser->getDepthWriteSet(&depthWriteSet);
+    if (depthWriteSet) {
+        bool depthWrite = false;
+        parser->getDepthWrite(&depthWrite);
+        mRasterState.depthWrite = depthWrite;
+    }
+
+    // if doubleSided() was called we override culling()
+    bool doubleSideSet = false;
+    parser->getDoubleSidedSet(&doubleSideSet);
+    parser->getDoubleSided(&mDoubleSided);
+    parser->getCullingMode(&mCullingMode);
+    bool depthTest = false;
+    parser->getDepthTest(&depthTest);
+
+    if (doubleSideSet) {
+        mDoubleSidedCapability = true;
+        mRasterState.culling = mDoubleSided ? CullingMode::NONE : mCullingMode;
+    } else {
+        mRasterState.culling = mCullingMode;
+    }
+
+    parser->getTransparencyMode(&mTransparencyMode);
+    parser->hasCustomDepthShader(&mHasCustomDepthShader);
+    mIsDefaultMaterial = builder->mDefaultMaterial;
+
+    if (UTILS_UNLIKELY(mIsDefaultMaterial)) {
+        assert_invariant(mMaterialDomain == MaterialDomain::SURFACE);
+        filaflat::MaterialChunk const& materialChunk{ mMaterialParser->getMaterialChunk() };
+        auto variants = FixedCapacityVector<Variant>::with_capacity(materialChunk.getShaderCount());
+        materialChunk.visitShaders([&variants](
+                ShaderModel, Variant variant, ShaderStage) {
+            if (Variant::isValidDepthVariant(variant)) {
+                variants.push_back(variant);
+            }
+        });
+        std::sort(variants.begin(), variants.end(),
+                [](Variant lhs, Variant rhs) { return lhs.key < rhs.key; });
+        auto pos = std::unique(variants.begin(), variants.end());
+        variants.resize(std::distance(variants.begin(), pos));
+        std::swap(mDepthVariants, variants);
+    }
+
+    if (mMaterialDomain == MaterialDomain::SURFACE) {
+        if (UTILS_UNLIKELY(!mIsDefaultMaterial && !mHasCustomDepthShader)) {
+            FMaterial const* const pDefaultMaterial = engine.getDefaultMaterial();
+            auto& cachedPrograms = mCachedPrograms;
+            for (Variant const variant: pDefaultMaterial->mDepthVariants) {
+                pDefaultMaterial->prepareProgram(variant);
+                cachedPrograms[variant.key] = pDefaultMaterial->getProgram(variant);
+            }
+        }
+    }
+
+    bool colorWrite = false;
+    parser->getColorWrite(&colorWrite);
+    mRasterState.colorWrite = colorWrite;
+    mRasterState.depthFunc = depthTest ? DepthFunc::GE : DepthFunc::A;
+
+    bool alphaToCoverageSet = false;
+    parser->getAlphaToCoverageSet(&alphaToCoverageSet);
+    if (alphaToCoverageSet) {
+        bool alphaToCoverage = false;
+        parser->getAlphaToCoverage(&alphaToCoverage);
+        mRasterState.alphaToCoverage = alphaToCoverage;
+    } else {
+        mRasterState.alphaToCoverage = mBlendingMode == BlendingMode::MASKED;
+    }
+
+    parser->hasSpecularAntiAliasing(&mSpecularAntiAliasing);
+    if (mSpecularAntiAliasing) {
+        parser->getSpecularAntiAliasingVariance(&mSpecularAntiAliasingVariance);
+        parser->getSpecularAntiAliasingThreshold(&mSpecularAntiAliasingThreshold);
+    }
+
+    // we can only initialize the default instance once we're initialized ourselves
+    mDefaultInstance.initDefaultInstance(engine, this);
 }

-FMaterial::~FMaterial() noexcept = default;
+FMaterial::~FMaterial() noexcept {
+    delete mMaterialParser;
+}

 void FMaterial::invalidate(Variant::type_t variantMask, Variant::type_t variantValue) noexcept {
    if (mMaterialDomain == MaterialDomain::SURFACE) {
@@ -541,7 +692,6 @@ Program FMaterial::getProgramWithVariants(
    Program program;
    program.shader(ShaderStage::VERTEX, vsBuilder.data(), vsBuilder.size())
           .shader(ShaderStage::FRAGMENT, fsBuilder.data(), fsBuilder.size())
-           .shaderLanguage(mMaterialParser->getShaderLanguage())
           .uniformBlockBindings(mUniformBlockBindings)
           .diagnostics(mName,
                    [this, variant](io::ostream& out) -> io::ostream& {
@@ -563,7 +713,8 @@ Program FMaterial::getProgramWithVariants(
                    samplers.data(), info.count);
        }
    }
-    if (UTILS_UNLIKELY(mMaterialParser->getShaderLanguage() == ShaderLanguage::ESSL1)) {
+
+    if (UTILS_UNLIKELY(mEngine.getShaderLanguage() == ShaderLanguage::ESSL1)) {
        assert_invariant(!mBindingUniformInfo.empty());
        for (auto const& [index, uniforms] : mBindingUniformInfo) {
            program.uniforms(uint32_t(index), uniforms);
@@ -637,7 +788,9 @@ void FMaterial::applyPendingEdits() noexcept {
    const char* name = mName.c_str();
    slog.d << "Applying edits to " << (name ? name : "(untitled)") << io::endl;
    destroyPrograms(mEngine); // FIXME: this will not destroy the shared variants
-    latchPendingEdits();
+    delete mMaterialParser;
+    mMaterialParser = mPendingEdits;
+    mPendingEdits = nullptr;
 }

 /**
@@ -654,9 +807,8 @@ void FMaterial::onEditCallback(void* userdata, const utils::CString&, const void

    // This is called on a web server thread, so we defer clearing the program cache
    // and swapping out the MaterialParser until the next getProgram call.
-    std::unique_ptr<MaterialParser> pending = createParser(
-            engine.getBackend(), engine.getShaderLanguage(), packageData, packageSize);
-    material->setPendingEdits(std::move(pending));
+    material->mPendingEdits = createParser(engine.getBackend(), engine.getShaderLanguage(),
+            packageData, packageSize);
 }

 void FMaterial::onQueryCallback(void* userdata, VariantList* pVariants) {
@@ -736,211 +888,6 @@ bool FMaterial::setConstant(uint32_t id, T value) noexcept {
    return false;
 }

-void FMaterial::processBlendingMode(MaterialParser const* const parser) {
-    parser->getBlendingMode(&mBlendingMode);
-
-    if (mBlendingMode == BlendingMode::MASKED) {
-        parser->getMaskThreshold(&mMaskThreshold);
-    }
-
-    if (mBlendingMode == BlendingMode::CUSTOM) {
-        parser->getCustomBlendFunction(&mCustomBlendFunctions);
-    }
-
-    // blending mode
-    switch (mBlendingMode) {
-        // Do not change the MASKED behavior without checking for regressions with
-        // AlphaBlendModeTest and TextureLinearInterpolationTest, with and without
-        // View::BlendMode::TRANSLUCENT.
-        case BlendingMode::MASKED:
-        case BlendingMode::OPAQUE:
-            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
-            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
-            mRasterState.blendFunctionDstRGB   = BlendFunction::ZERO;
-            mRasterState.blendFunctionDstAlpha = BlendFunction::ZERO;
-            mRasterState.depthWrite = true;
-            break;
-        case BlendingMode::TRANSPARENT:
-        case BlendingMode::FADE:
-            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
-            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
-            mRasterState.blendFunctionDstRGB   = BlendFunction::ONE_MINUS_SRC_ALPHA;
-            mRasterState.blendFunctionDstAlpha = BlendFunction::ONE_MINUS_SRC_ALPHA;
-            mRasterState.depthWrite = false;
-            break;
-        case BlendingMode::ADD:
-            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
-            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
-            mRasterState.blendFunctionDstRGB   = BlendFunction::ONE;
-            mRasterState.blendFunctionDstAlpha = BlendFunction::ONE;
-            mRasterState.depthWrite = false;
-            break;
-        case BlendingMode::MULTIPLY:
-            mRasterState.blendFunctionSrcRGB   = BlendFunction::ZERO;
-            mRasterState.blendFunctionSrcAlpha = BlendFunction::ZERO;
-            mRasterState.blendFunctionDstRGB   = BlendFunction::SRC_COLOR;
-            mRasterState.blendFunctionDstAlpha = BlendFunction::SRC_COLOR;
-            mRasterState.depthWrite = false;
-            break;
-        case BlendingMode::SCREEN:
-            mRasterState.blendFunctionSrcRGB   = BlendFunction::ONE;
-            mRasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
-            mRasterState.blendFunctionDstRGB   = BlendFunction::ONE_MINUS_SRC_COLOR;
-            mRasterState.blendFunctionDstAlpha = BlendFunction::ONE_MINUS_SRC_COLOR;
-            mRasterState.depthWrite = false;
-            break;
-        case BlendingMode::CUSTOM:
-            mRasterState.blendFunctionSrcRGB   = mCustomBlendFunctions[0];
-            mRasterState.blendFunctionSrcAlpha = mCustomBlendFunctions[1];
-            mRasterState.blendFunctionDstRGB   = mCustomBlendFunctions[2];
-            mRasterState.blendFunctionDstAlpha = mCustomBlendFunctions[3];
-            mRasterState.depthWrite = false;
-    }
-
-    // depth write
-    bool depthWriteSet = false;
-    parser->getDepthWriteSet(&depthWriteSet);
-    if (depthWriteSet) {
-        bool depthWrite = false;
-        parser->getDepthWrite(&depthWrite);
-        mRasterState.depthWrite = depthWrite;
-    }
-
-    // alpha to coverage
-    bool alphaToCoverageSet = false;
-    parser->getAlphaToCoverageSet(&alphaToCoverageSet);
-    if (alphaToCoverageSet) {
-        bool alphaToCoverage = false;
-        parser->getAlphaToCoverage(&alphaToCoverage);
-        mRasterState.alphaToCoverage = alphaToCoverage;
-    } else {
-        mRasterState.alphaToCoverage = mBlendingMode == BlendingMode::MASKED;
-    }
-}
-
-void FMaterial::processSpecializationConstants(FEngine& engine, Material::Builder const& builder,
-        MaterialParser const* const parser) {
-    // Older materials won't have a constants chunk, but that's okay.
-    parser->getConstants(&mMaterialConstants);
-    for (size_t i = 0, c = mMaterialConstants.size(); i < c; i++) {
-        auto& item = mMaterialConstants[i];
-        // the key can be a string_view because mMaterialConstant owns the CString
-        std::string_view const key{ item.name.data(), item.name.size() };
-        mSpecializationConstantsNameToIndex[key] = i;
-    }
-
-    // Verify that all the constant specializations exist in the material and that their types match.
-    // The first specialization constants are defined internally by Filament.
-    // The subsequent constants are user-defined in the material.
-
-    // Feature level 0 doesn't support instancing
-    int const maxInstanceCount = (engine.getActiveFeatureLevel() == FeatureLevel::FEATURE_LEVEL_0)
-                                 ? 1 : CONFIG_MAX_INSTANCES;
-
-    int const maxFroxelBufferHeight = (int)std::min(
-            FROXEL_BUFFER_MAX_ENTRY_COUNT / 4,
-            engine.getDriverApi().getMaxUniformBufferSize() / 16u);
-
-    bool const staticTextureWorkaround =
-            engine.getDriverApi().isWorkaroundNeeded(Workaround::A8X_STATIC_TEXTURE_TARGET_ERROR);
-
-    bool const powerVrShaderWorkarounds =
-            engine.getDriverApi().isWorkaroundNeeded(Workaround::POWER_VR_SHADER_WORKAROUNDS);
-
-    mSpecializationConstants.reserve(mMaterialConstants.size() + CONFIG_MAX_RESERVED_SPEC_CONSTANTS);
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::BACKEND_FEATURE_LEVEL,
-            (int)engine.getSupportedFeatureLevel() });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_MAX_INSTANCES,
-            (int)maxInstanceCount });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_FROXEL_BUFFER_HEIGHT,
-            (int)maxFroxelBufferHeight });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_DEBUG_DIRECTIONAL_SHADOWMAP,
-            engine.debug.shadowmap.debug_directional_shadowmap });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_DEBUG_FROXEL_VISUALIZATION,
-            engine.debug.lighting.debug_froxel_visualization });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_STATIC_TEXTURE_TARGET_WORKAROUND,
-            staticTextureWorkaround });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_POWER_VR_SHADER_WORKAROUNDS,
-            powerVrShaderWorkarounds });
-    mSpecializationConstants.push_back({
-            +ReservedSpecializationConstants::CONFIG_STEREO_EYE_COUNT,
-            (int)engine.getConfig().stereoscopicEyeCount });
-    if (UTILS_UNLIKELY(parser->getShaderLanguage() == ShaderLanguage::ESSL1)) {
-        // The actual value of this spec-constant is set in the OpenGLDriver backend.
-        mSpecializationConstants.push_back({
-                +ReservedSpecializationConstants::CONFIG_SRGB_SWAPCHAIN_EMULATION,
-                false});
-    }
-
-    for (auto const& [name, value] : builder->mConstantSpecializations) {
-        std::string_view const key{ name.data(), name.size() };
-        auto pos = mSpecializationConstantsNameToIndex.find(key);
-        ASSERT_PRECONDITION(pos != mSpecializationConstantsNameToIndex.end(),
-                "The material %s does not have a constant parameter named %s.",
-                mName.c_str_safe(), name.c_str());
-        const char* const types[3] = {"an int", "a float", "a bool"};
-        const char* const errorMessage =
-                "The constant parameter %s on material %s is of type %s, but %s was "
-                "provided.";
-        auto& constant = mMaterialConstants[pos->second];
-        switch (constant.type) {
-            case ConstantType::INT:
-                ASSERT_PRECONDITION(std::holds_alternative<int32_t>(value), errorMessage,
-                        name.c_str(), mName.c_str_safe(), "int", types[value.index()]);
-                break;
-            case ConstantType::FLOAT:
-                ASSERT_PRECONDITION(std::holds_alternative<float>(value), errorMessage,
-                        name.c_str(), mName.c_str_safe(), "float", types[value.index()]);
-                break;
-            case ConstantType::BOOL:
-                ASSERT_PRECONDITION(std::holds_alternative<bool>(value), errorMessage,
-                        name.c_str(), mName.c_str_safe(), "bool", types[value.index()]);
-                break;
-        }
-        uint32_t const index = pos->second + CONFIG_MAX_RESERVED_SPEC_CONSTANTS;
-        mSpecializationConstants.push_back({ index, value });
-    }
-}
-
-void FMaterial::processDepthVariants(FEngine& engine, MaterialParser const* const parser) {
-    parser->hasCustomDepthShader(&mHasCustomDepthShader);
-
-    if (UTILS_UNLIKELY(mIsDefaultMaterial)) {
-        assert_invariant(mMaterialDomain == MaterialDomain::SURFACE);
-        filaflat::MaterialChunk const& materialChunk{ parser->getMaterialChunk() };
-        auto variants = FixedCapacityVector<Variant>::with_capacity(materialChunk.getShaderCount());
-        materialChunk.visitShaders([&variants](
-                ShaderModel, Variant variant, ShaderStage) {
-            if (Variant::isValidDepthVariant(variant)) {
-                variants.push_back(variant);
-            }
-        });
-        std::sort(variants.begin(), variants.end(),
-                [](Variant lhs, Variant rhs) { return lhs.key < rhs.key; });
-        auto pos = std::unique(variants.begin(), variants.end());
-        variants.resize(std::distance(variants.begin(), pos));
-        std::swap(mDepthVariants, variants);
-    }
-
-    if (mMaterialDomain == MaterialDomain::SURFACE) {
-        if (UTILS_UNLIKELY(!mIsDefaultMaterial && !mHasCustomDepthShader)) {
-            FMaterial const* const pDefaultMaterial = engine.getDefaultMaterial();
-            auto& cachedPrograms = mCachedPrograms;
-            for (Variant const variant: pDefaultMaterial->mDepthVariants) {
-                pDefaultMaterial->prepareProgram(variant);
-                cachedPrograms[variant.key] = pDefaultMaterial->getProgram(variant);
-            }
-        }
-    }
-}
-
 template bool FMaterial::setConstant<int32_t>(uint32_t id, int32_t value) noexcept;
 template bool FMaterial::setConstant<float>(uint32_t id, float value) noexcept;
 template bool FMaterial::setConstant<bool>(uint32_t id, bool value) noexcept;
@@ -22,39 +22,18 @@
 #include "details/MaterialInstance.h"

 #include <filament/Material.h>
-#include <filament/MaterialEnums.h>

-#include <private/filament/EngineEnums.h>
-#include <private/filament/BufferInterfaceBlock.h>
 #include <private/filament/SamplerBindingsInfo.h>
 #include <private/filament/SamplerInterfaceBlock.h>
 #include <private/filament/SubpassInfo.h>
 #include <private/filament/Variant.h>
 #include <private/filament/ConstantInfo.h>

-#include <backend/CallbackHandler.h>
-#include <backend/DriverEnums.h>
-#include <backend/Handle.h>
-#include <backend/Program.h>
-
 #include <utils/compiler.h>
-#include <utils/CString.h>
-#include <utils/debug.h>
-#include <utils/FixedCapacityVector.h>
-#include <utils/Invocable.h>
 #include <utils/Mutex.h>

-#include <array>
 #include <atomic>
-#include <memory>
-#include <mutex>
 #include <optional>
-#include <string_view>
-#include <unordered_map>
-#include <utility>
-
-#include <stddef.h>
-#include <stdint.h>

 namespace filament {

@@ -64,8 +43,7 @@ class  FEngine;

 class FMaterial : public Material {
 public:
-    FMaterial(FEngine& engine, const Material::Builder& builder,
-            std::unique_ptr<MaterialParser> materialParser);
+    FMaterial(FEngine& engine, const Material::Builder& builder);
    ~FMaterial() noexcept;

    class DefaultMaterialBuilder : public Material::Builder {
@@ -149,6 +127,7 @@ public:
    Shading getShading() const noexcept { return mShading; }
    Interpolation getInterpolation() const noexcept { return mInterpolation; }
    BlendingMode getBlendingMode() const noexcept { return mBlendingMode; }
+    BlendingMode getRenderBlendingMode() const noexcept { return mRenderBlendingMode; }
    VertexDomain getVertexDomain() const noexcept { return mVertexDomain; }
    MaterialDomain getMaterialDomain() const noexcept { return mMaterialDomain; }
    CullingMode getCullingMode() const noexcept { return mCullingMode; }
@@ -219,7 +198,7 @@ public:
    static void onQueryCallback(void* userdata, VariantList* pActiveVariants);

    void checkProgramEdits() noexcept {
-        if (UTILS_UNLIKELY(hasPendingEdits())) {
+        if (UTILS_UNLIKELY(mPendingEdits.load())) {
            applyPendingEdits();
        }
    }
@@ -238,26 +217,19 @@ private:
    backend::Program getProgramWithVariants(Variant variant,
            Variant vertexVariant, Variant fragmentVariant) const noexcept;

-    void processBlendingMode(MaterialParser const* const parser);
-
-    void processSpecializationConstants(FEngine& engine, Material::Builder const& builder,
-            MaterialParser const* const parser);
-
-    void processDepthVariants(FEngine& engine, MaterialParser const* const parser);
-
    void createAndCacheProgram(backend::Program&& p, Variant variant) const noexcept;

    // try to order by frequency of use
    mutable std::array<backend::Handle<backend::HwProgram>, VARIANT_COUNT> mCachedPrograms;

    backend::RasterState mRasterState;
+    BlendingMode mRenderBlendingMode = BlendingMode::OPAQUE;
    TransparencyMode mTransparencyMode = TransparencyMode::DEFAULT;
    bool mIsVariantLit = false;
    backend::FeatureLevel mFeatureLevel = backend::FeatureLevel::FEATURE_LEVEL_1;
    Shading mShading = Shading::UNLIT;

    BlendingMode mBlendingMode = BlendingMode::OPAQUE;
-    std::array<backend::BlendFunction, 4> mCustomBlendFunctions = {};
    Interpolation mInterpolation = Interpolation::SMOOTH;
    VertexDomain mVertexDomain = VertexDomain::OBJECT;
    MaterialDomain mMaterialDomain = MaterialDomain::SURFACE;
@@ -309,21 +281,7 @@ private:
    matdbg::MaterialKey mDebuggerId;
    mutable utils::Mutex mActiveProgramsLock;
    mutable VariantList mActivePrograms;
-    mutable utils::Mutex mPendingEditsLock;
-    std::unique_ptr<MaterialParser> mPendingEdits;
-    void setPendingEdits(std::unique_ptr<MaterialParser> pendingEdits) noexcept {
-        std::lock_guard lock(mPendingEditsLock);
-        std::swap(pendingEdits, mPendingEdits);
-    }
-    bool hasPendingEdits() noexcept {
-        std::lock_guard lock(mPendingEditsLock);
-        return (bool)mPendingEdits;
-    }
-    void latchPendingEdits() noexcept {
-        std::lock_guard lock(mPendingEditsLock);
-        std::swap(mPendingEdits, mMaterialParser);
-    }
-
+    std::atomic<MaterialParser*> mPendingEdits = {};
 #endif

    utils::CString mName;
@@ -331,7 +289,7 @@ private:
    const uint32_t mMaterialId;
    uint64_t mCacheId = 0;
    mutable uint32_t mMaterialInstanceId = 0;
-    std::unique_ptr<MaterialParser> mMaterialParser;
+    MaterialParser* mMaterialParser = nullptr;
 };


@@ -31,7 +31,7 @@ using namespace filament;
 // This will re-compile the test material with the current version of matc.
 // To verify, rebuild and re-run test_material_parser (this test suite).
 TEST(MaterialParser, Parse) {
-    MaterialParser parser({ backend::ShaderLanguage::ESSL3 },
+    MaterialParser parser(backend::ShaderLanguage::ESSL3,
            FILAMENT_TEST_RESOURCES_TEST_MATERIAL_DATA, FILAMENT_TEST_RESOURCES_TEST_MATERIAL_SIZE);
    MaterialParser::ParseResult materialOk = parser.parse();

@@ -1,12 +1,12 @@
 Pod::Spec.new do |spec|
  spec.name = "Filament"
-  spec.version = "1.51.5"
+  spec.version = "1.51.0"
  spec.license = { :type => "Apache 2.0", :file => "LICENSE" }
  spec.homepage = "https://google.github.io/filament"
  spec.authors = "Google LLC."
  spec.summary = "Filament is a real-time physically based rendering engine for Android, iOS, Windows, Linux, macOS, and WASM/WebGL."
  spec.platform = :ios, "11.0"
-  spec.source = { :http => "https://github.com/google/filament/releases/download/v1.51.5/filament-v1.51.5-ios.tgz" }
+  spec.source = { :http => "https://github.com/google/filament/releases/download/v1.51.0/filament-v1.51.0-ios.tgz" }

  # Fix linking error with Xcode 12; we do not yet support the simulator on Apple silicon.
  spec.pod_target_xcconfig = {
@@ -45,7 +45,6 @@ enum UTILS_PUBLIC ChunkType : uint64_t {
    MaterialEssl1 = charTo64bitNum("MAT_ESS1"),
    MaterialSpirv = charTo64bitNum("MAT_SPIR"),
    MaterialMetal = charTo64bitNum("MAT_METL"),
-    MaterialMetalLibrary = charTo64bitNum("MAT_MLIB"),
    MaterialShaderModels = charTo64bitNum("MAT_SMDL"),
    MaterialSamplerBindings = charTo64bitNum("MAT_SAMP"),
    MaterialUniformBindings = charTo64bitNum("MAT_UNIF"),
@@ -60,7 +59,6 @@ enum UTILS_PUBLIC ChunkType : uint64_t {
    MaterialFeatureLevel = charTo64bitNum("MAT_FEAT"),
    MaterialShading = charTo64bitNum("MAT_SHAD"),
    MaterialBlendingMode = charTo64bitNum("MAT_BLEN"),
-    MaterialBlendFunction = charTo64bitNum("MAT_BLFN"),
    MaterialTransparencyMode = charTo64bitNum("MAT_TRMD"),
    MaterialMaskThreshold = charTo64bitNum("MAT_THRS"),
    MaterialShadowMultiplier = charTo64bitNum("MAT_SHML"),
@@ -94,7 +92,6 @@ enum UTILS_PUBLIC ChunkType : uint64_t {

    DictionaryText = charTo64bitNum("DIC_TEXT"),
    DictionarySpirv = charTo64bitNum("DIC_SPIR"),
-    DictionaryMetalLibrary = charTo64bitNum("DIC_MLIB"),
 };

 } // namespace filamat
@@ -80,8 +80,6 @@ enum class BlendingMode : uint8_t {
    MULTIPLY,
    //! material brightens what's behind it
    SCREEN,
-    //! custom blending function
-    CUSTOM,
 };

 /**
@@ -74,7 +74,7 @@ private:
            BlobDictionary const& dictionary, ShaderContent& shaderContent,
            ShaderModel shaderModel, filament::Variant variant, ShaderStage shaderStage);

-    bool getBinaryShader(
+    bool getSpirvShader(
            BlobDictionary const& dictionary, ShaderContent& shaderContent,
            ShaderModel shaderModel, filament::Variant variant, ShaderStage shaderStage);
 };
@@ -78,25 +78,6 @@ bool DictionaryReader::unflatten(ChunkContainer const& container,

        }
        return true;
-    } else if (dictionaryTag == ChunkType::DictionaryMetalLibrary) {
-        uint32_t blobCount;
-        if (!unflattener.read(&blobCount)) {
-            return false;
-        }
-
-        dictionary.reserve(blobCount);
-        for (uint32_t i = 0; i < blobCount; i++) {
-            unflattener.skipAlignmentPadding();
-
-            const char* data;
-            size_t dataSize;
-            if (!unflattener.read(&data, &dataSize)) {
-                return false;
-            }
-            dictionary.emplace_back(dataSize);
-            memcpy(dictionary.back().data(), data, dictionary.back().size());
-        }
-        return true;
    } else if (dictionaryTag == ChunkType::DictionaryText) {
        uint32_t stringCount = 0;
        if (!unflattener.read(&stringCount)) {
@@ -153,7 +153,7 @@ bool MaterialChunk::getTextShader(Unflattener unflattener,
    return true;
 }

-bool MaterialChunk::getBinaryShader(BlobDictionary const& dictionary,
+bool MaterialChunk::getSpirvShader(BlobDictionary const& dictionary,
        ShaderContent& shaderContent, ShaderModel shaderModel, filament::Variant variant, ShaderStage shaderStage) {

    if (mBase == nullptr) {
@@ -186,8 +186,7 @@ bool MaterialChunk::getShader(ShaderContent& shaderContent, BlobDictionary const
        case filamat::ChunkType::MaterialMetal:
            return getTextShader(mUnflattener, dictionary, shaderContent, shaderModel, variant, stage);
        case filamat::ChunkType::MaterialSpirv:
-        case filamat::ChunkType::MaterialMetalLibrary:
-            return getBinaryShader(dictionary, shaderContent, shaderModel, variant, stage);
+            return getSpirvShader(dictionary, shaderContent, shaderModel, variant, stage);
        default:
            return false;
    }
@@ -50,8 +50,7 @@ set(PRIVATE_HDRS
        ${COMMON_PRIVATE_HDRS}
        src/eiff/BlobDictionary.h
        src/eiff/DictionarySpirvChunk.h
-        src/eiff/DictionaryMetalLibraryChunk.h
-        src/eiff/MaterialBinaryChunk.h
+        src/eiff/MaterialSpirvChunk.h
        src/GLSLPostProcessor.h
        src/MetalArgumentBuffer.h
        src/ShaderMinifier.h
@@ -64,8 +63,7 @@ set(SRCS
        ${COMMON_SRCS}
        src/eiff/BlobDictionary.cpp
        src/eiff/DictionarySpirvChunk.cpp
-        src/eiff/DictionaryMetalLibraryChunk.cpp
-        src/eiff/MaterialBinaryChunk.cpp
+        src/eiff/MaterialSpirvChunk.cpp
        src/MetalArgumentBuffer.cpp
        src/sca/ASTHelpers.cpp
        src/sca/GLSLTools.cpp
@@ -227,7 +227,6 @@ public:

    using ShaderQuality = filament::ShaderQuality;
    using BlendingMode = filament::BlendingMode;
-    using BlendFunction = filament::backend::BlendFunction;
    using Shading = filament::Shading;
    using Interpolation = filament::Interpolation;
    using VertexDomain = filament::VertexDomain;
@@ -408,15 +407,6 @@ public:
     */
    MaterialBuilder& blending(BlendingMode blending) noexcept;

-    /**
-     * Set the blend function  for this material. blending must be et to CUSTOM.
-     */
-    MaterialBuilder& customBlendFunctions(
-            BlendFunction srcRGB,
-            BlendFunction srcA,
-            BlendFunction dstRGB,
-            BlendFunction dstA) noexcept;
-
    /**
     * Set the blending mode of the post-lighting color for this material.
     * Only OPAQUE, TRANSPARENT and ADD are supported, the default is TRANSPARENT.
@@ -838,7 +828,6 @@ private:
    FeatureLevel mFeatureLevel = FeatureLevel::FEATURE_LEVEL_1;
    BlendingMode mBlendingMode = BlendingMode::OPAQUE;
    BlendingMode mPostLightingBlendingMode = BlendingMode::TRANSPARENT;
-    std::array<BlendFunction, 4> mCustomBlendFunctions = {};
    CullingMode mCullingMode = CullingMode::BACK;
    Shading mShading = Shading::LIT;
    MaterialDomain mMaterialDomain = MaterialDomain::SURFACE;
@@ -33,12 +33,11 @@
 #include "eiff/LineDictionary.h"
 #include "eiff/MaterialInterfaceBlockChunk.h"
 #include "eiff/MaterialTextChunk.h"
-#include "eiff/MaterialBinaryChunk.h"
+#include "eiff/MaterialSpirvChunk.h"
 #include "eiff/ChunkContainer.h"
 #include "eiff/SimpleFieldChunk.h"
 #include "eiff/DictionaryTextChunk.h"
 #include "eiff/DictionarySpirvChunk.h"
-#include "eiff/DictionaryMetalLibraryChunk.h"

 #include <private/filament/BufferInterfaceBlock.h>
 #include <private/filament/SamplerInterfaceBlock.h>
@@ -389,16 +388,6 @@ MaterialBuilder& MaterialBuilder::blending(BlendingMode blending) noexcept {
    return *this;
 }

-MaterialBuilder& MaterialBuilder::customBlendFunctions(
-        BlendFunction srcRGB, BlendFunction srcA,
-        BlendFunction dstRGB, BlendFunction dstA) noexcept {
-    mCustomBlendFunctions[0] = srcRGB;
-    mCustomBlendFunctions[1] = srcA;
-    mCustomBlendFunctions[2] = dstRGB;
-    mCustomBlendFunctions[3] = dstA;
-    return *this;
-}
-
 MaterialBuilder& MaterialBuilder::postLightingBlending(BlendingMode blending) noexcept {
    mPostLightingBlendingMode = blending;
    return *this;
@@ -822,7 +811,7 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
    Mutex entriesLock;
    std::vector<TextEntry> glslEntries;
    std::vector<TextEntry> essl1Entries;
-    std::vector<BinaryEntry> spirvEntries;
+    std::vector<SpirvEntry> spirvEntries;
    std::vector<TextEntry> metalEntries;
    LineDictionary textDictionary;
    BlobDictionary spirvDictionary;
@@ -873,7 +862,7 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
                std::string* pMsl = targetApiNeedsMsl ? &msl : nullptr;

                TextEntry glslEntry{};
-                BinaryEntry spirvEntry{};
+                SpirvEntry spirvEntry{};
                TextEntry metalEntry{};

                glslEntry.shaderModel  = params.shaderModel;
@@ -964,15 +953,12 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
                            glslEntries.push_back(glslEntry);
                        }
                        break;
-                    case TargetApi::VULKAN: {
+                    case TargetApi::VULKAN:
                        assert(!spirv.empty());
-                        const std::vector<uint8_t> d(reinterpret_cast<uint8_t*>(spirv.data()),
-                                reinterpret_cast<uint8_t*>(spirv.data() + spirv.size()));
                        spirvEntry.stage = v.stage;
-                        spirvEntry.data = std::move(d);
+                        spirvEntry.spirv = std::move(spirv);
                        spirvEntries.push_back(spirvEntry);
                        break;
-                    }
                    case TargetApi::METAL:
                        assert(!spirv.empty());
                        assert(msl.length() > 0);
@@ -1022,7 +1008,7 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
        textDictionary.addText(s.shader);
    }
    for (auto& s : spirvEntries) {
-        std::vector<uint8_t> spirv = std::move(s.data);
+        std::vector<uint32_t> spirv = std::move(s.spirv);
        s.dictionaryIndex = spirvDictionary.addBlob(spirv);
    }
    for (const auto& s : metalEntries) {
@@ -1045,11 +1031,11 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
                dictionaryChunk.getDictionary(), ChunkType::MaterialEssl1);
    }

-    // Emit SPIRV chunks (SpirvDictionaryReader and MaterialBinaryChunk).
+    // Emit SPIRV chunks (SpirvDictionaryReader and MaterialSpirvChunk).
    if (!spirvEntries.empty()) {
        const bool stripInfo = !mGenerateDebugInfo;
        container.push<filamat::DictionarySpirvChunk>(std::move(spirvDictionary), stripInfo);
-        container.push<MaterialBinaryChunk>(std::move(spirvEntries), ChunkType::MaterialSpirv);
+        container.push<MaterialSpirvChunk>(std::move(spirvEntries));
    }

    // Emit Metal chunk (MaterialTextChunk).
@@ -1524,16 +1510,6 @@ void MaterialBuilder::writeCommonChunks(ChunkContainer& container, MaterialInfo&
        container.emplace<bool>(ChunkType::MaterialDoubleSided, mDoubleSided);
        container.emplace<uint8_t>(ChunkType::MaterialBlendingMode,
                static_cast<uint8_t>(mBlendingMode));
-
-        if (mBlendingMode == BlendingMode::CUSTOM) {
-            uint32_t const blendFunctions =
-                    (uint32_t(mCustomBlendFunctions[0]) << 24) |
-                    (uint32_t(mCustomBlendFunctions[1]) << 16) |
-                    (uint32_t(mCustomBlendFunctions[2]) <<  8) |
-                    (uint32_t(mCustomBlendFunctions[3]) <<  0);
-            container.emplace< uint32_t >(ChunkType::MaterialBlendFunction, blendFunctions);
-        }
-
        container.emplace<uint8_t>(ChunkType::MaterialTransparencyMode,
                static_cast<uint8_t>(mTransparencyMode));
        container.emplace<uint8_t>(ChunkType::MaterialReflectionMode,
@@ -20,8 +20,8 @@

 namespace filamat {

-size_t BlobDictionary::addBlob(const std::vector<uint8_t>& vblob) noexcept {
-    std::string_view blob((char*) vblob.data(), vblob.size());
+size_t BlobDictionary::addBlob(const std::vector<uint32_t>& vblob) noexcept {
+    std::string_view blob((char*) vblob.data(), vblob.size() * 4);
    auto iter = mBlobIndices.find(blob);
    if (iter != mBlobIndices.end()) {
        return iter->second;
@@ -36,7 +36,7 @@ public:
    BlobDictionary(BlobDictionary&&) = default;

    // Adds a blob if it's not already a duplicate and returns its index.
-    size_t addBlob(const std::vector<uint8_t>& blob) noexcept;
+    size_t addBlob(const std::vector<uint32_t>& blob) noexcept;

    size_t getBlobCount() const noexcept {
        return mBlobs.size();
@@ -1,33 +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 "DictionaryMetalLibraryChunk.h"
-
-namespace filamat {
-
-DictionaryMetalLibraryChunk::DictionaryMetalLibraryChunk(BlobDictionary&& dictionary)
-    : Chunk(ChunkType::DictionaryMetalLibrary), mDictionary(std::move(dictionary)) {}
-
-void DictionaryMetalLibraryChunk::flatten(Flattener& f) {
-    f.writeUint32(mDictionary.getBlobCount());
-    for (size_t i = 0 ; i < mDictionary.getBlobCount() ; i++) {
-        std::string_view blob = mDictionary.getBlob(i);
-        f.writeAlignmentPadding();
-        f.writeBlob((const char*) blob.data(), blob.size());
-    }
-}
-
-} // namespace filamat
@@ -1,43 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMAT_DIC_METAL_LIBRARY_CHUNK_H
-#define TNT_FILAMAT_DIC_METAL_LIBRARY_CHUNK_H
-
-#include <stdint.h>
-#include <vector>
-
-#include "Chunk.h"
-#include "Flattener.h"
-#include "BlobDictionary.h"
-
-namespace filamat {
-
-class DictionaryMetalLibraryChunk final : public Chunk {
-public:
-    explicit DictionaryMetalLibraryChunk(BlobDictionary&& dictionary);
-    ~DictionaryMetalLibraryChunk() = default;
-
-private:
-    void flatten(Flattener& f) override;
-
-    BlobDictionary mDictionary;
-    bool mStripDebugInfo;
-};
-
-} // namespace filamat
-
-#endif // TNT_FILAMAT_DIC_METAL_LIBRARY_CHUNK_H
@@ -127,13 +127,6 @@ public:
        mCursor += nbytes;
    }

-    void writeRaw(const char* raw, size_t nbytes) {
-        if (mStart != nullptr) {
-            memcpy(reinterpret_cast<char*>(mCursor), raw, nbytes);
-        }
-        mCursor += nbytes;
-    }
-
    void writeSizePlaceholder() {
        mSizePlaceholders.push_back(mCursor);
        if (mStart != nullptr) {
@@ -1,6 +1,6 @@
 /*
 * Copyright (C) 2018 The Android Open Source Project
- *
+ *DictionaryGlsl
 * 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
@@ -14,17 +14,16 @@
 * limitations under the License.
 */

-#include "MaterialBinaryChunk.h"
+#include "MaterialSpirvChunk.h"

 namespace filamat {

-MaterialBinaryChunk::MaterialBinaryChunk(
-        const std::vector<BinaryEntry>&& entries, ChunkType chunkType)
-    : Chunk(chunkType), mEntries(entries) {}
+MaterialSpirvChunk::MaterialSpirvChunk(const std::vector<SpirvEntry>&& entries) :
+        Chunk(ChunkType::MaterialSpirv), mEntries(entries) {}

-void MaterialBinaryChunk::flatten(Flattener &f) {
+void MaterialSpirvChunk::flatten(Flattener &f) {
    f.writeUint64(mEntries.size());
-    for (const BinaryEntry& entry : mEntries) {
+    for (const SpirvEntry& entry : mEntries) {
        f.writeUint8(uint8_t(entry.shaderModel));
        f.writeUint8(entry.variant.key);
        f.writeUint8(uint8_t(entry.stage));
@@ -14,8 +14,8 @@
 * limitations under the License.
 */

-#ifndef TNT_FILAMAT_MATERIAL_BINARY_CHUNK_H
-#define TNT_FILAMAT_MATERIAL_BINARY_CHUNK_H
+#ifndef TNT_FILAMAT_MATERIAL_SPIRV_CHUNK_H
+#define TNT_FILAMAT_MATERIAL_SPIRV_CHUNK_H

 #include "Chunk.h"
 #include "ShaderEntry.h"
@@ -24,17 +24,17 @@

 namespace filamat {

-class MaterialBinaryChunk final : public Chunk {
+class MaterialSpirvChunk final : public Chunk {
 public:
-    explicit MaterialBinaryChunk(const std::vector<BinaryEntry>&& entries, ChunkType type);
-    ~MaterialBinaryChunk() = default;
+    explicit MaterialSpirvChunk(const std::vector<SpirvEntry>&& entries);
+    ~MaterialSpirvChunk() = default;

 private:
    void flatten(Flattener& f) override;

-    const std::vector<BinaryEntry> mEntries;
+    const std::vector<SpirvEntry> mEntries;
 };

 } // namespace filamat

-#endif // TNT_FILAMAT_MATERIAL_BINARY_CHUNK_H
+#endif // TNT_FILAMAT_MATERIAL_SPIRV_CHUNK_H
@@ -34,14 +34,14 @@ struct TextEntry {
    std::string shader;
 };

-struct BinaryEntry {
+struct SpirvEntry {
    filament::backend::ShaderModel shaderModel;
    filament::Variant variant;
    filament::backend::ShaderStage stage;
-    size_t dictionaryIndex;  // maps to an index in the blob dictionary
+    size_t dictionaryIndex;

-    // temporarily holds this entry's binary data until added to the dictionary
-    std::vector<uint8_t> data;
+    // temporarily holds this entry's spirv until added to the dictionary
+    std::vector<uint32_t> spirv;
 };

 }  // namespace filamat
@@ -157,9 +157,6 @@ void ShaderGenerator::generateSurfaceMaterialVariantDefines(utils::io::sstream&
            case BlendingMode::SCREEN:
                CodeGenerator::generateDefine(out, "BLEND_MODE_SCREEN", true);
                break;
-            case BlendingMode::CUSTOM:
-                CodeGenerator::generateDefine(out, "BLEND_MODE_CUSTOM", true);
-                break;
        }

        switch (material.postLightingBlendingMode) {
@@ -178,9 +175,6 @@ void ShaderGenerator::generateSurfaceMaterialVariantDefines(utils::io::sstream&
            case BlendingMode::SCREEN:
                CodeGenerator::generateDefine(out, "POST_LIGHTING_BLEND_MODE_SCREEN", true);
                break;
-            case BlendingMode::CUSTOM:
-                CodeGenerator::generateDefine(out, "POST_LIGHTING_BLEND_MODE_CUSTOM", true);
-                break;
            default:
                break;
        }
@@ -338,7 +338,6 @@ const char* toString(BlendingMode blendingMode) noexcept {
        case BlendingMode::FADE: return "fade";
        case BlendingMode::MULTIPLY: return "multiply";
        case BlendingMode::SCREEN: return "screen";
-        case BlendingMode::CUSTOM: return "custom";
    }
 }

@@ -122,9 +122,10 @@ void postProcess(inout PostProcessInputs postProcess) {
    float side = materialParams.side;

    // compute the view (and normal, since v = n) direction for each face
-    vec3 rx = normalize(vec3(      side,  -p.y, side * -p.x));
-    vec3 ry = normalize(vec3(       p.x,  side, side *  p.y));
-    vec3 rz = normalize(vec3(side * p.x,  -p.y, side));
+    float l = inversesqrt(p.x * p.x + p.y * p.y + 1.0);
+    vec3 rx = vec3(      side,  p.y,  side * -p.x) * l;
+    vec3 ry = vec3(       p.x,  side, side * -p.y) * l;
+    vec3 rz = vec3(side * p.x,  p.y,  side)        * l;

    // random rotation around r
    mediump float a = 2.0 * PI * random(gl_FragCoord.xy);
@@ -65,7 +65,6 @@ const char* toString(BlendingMode blendingMode) noexcept {
        case BlendingMode::FADE: return "fade";
        case BlendingMode::MULTIPLY: return "multiply";
        case BlendingMode::SCREEN: return "screen";
-        case BlendingMode::CUSTOM: return "custom";
    }
    return "--";
 }
@@ -53,10 +53,6 @@ ShaderExtractor::ShaderExtractor(backend::ShaderLanguage target, const void* dat
            mMaterialTag = ChunkType::MaterialMetal;
            mDictionaryTag = ChunkType::DictionaryText;
            break;
-        case backend::ShaderLanguage::METAL_LIBRARY:
-            mMaterialTag = ChunkType::MaterialMetalLibrary;
-            mDictionaryTag = ChunkType::DictionaryMetalLibrary;
-            break;
        case backend::ShaderLanguage::SPIRV:
            mMaterialTag = ChunkType::MaterialSpirv;
            mDictionaryTag = ChunkType::DictionarySpirv;
@@ -35,7 +35,7 @@
 #include "eiff/ChunkContainer.h"
 #include "eiff/DictionarySpirvChunk.h"
 #include "eiff/DictionaryTextChunk.h"
-#include "eiff/MaterialBinaryChunk.h"
+#include "eiff/MaterialSpirvChunk.h"
 #include "eiff/MaterialTextChunk.h"
 #include "eiff/LineDictionary.h"

@@ -73,7 +73,7 @@ private:
    vector<TextEntry> mShaderRecords;
 };

-// Tiny database of data blobs that can import / export MaterialBinaryChunk and DictionarySpirvChunk.
+// Tiny database of data blobs that can import / export MaterialSpirvChunk and DictionarySpirvChunk.
 // The blobs are stored *after* they have been compressed by SMOL-V.
 class BlobIndex {
 public:
@@ -90,7 +90,7 @@ public:
 private:
    const ChunkType mDictTag;
    const ChunkType mMatTag;
-    vector<BinaryEntry> mShaderRecords;
+    vector<SpirvEntry> mShaderRecords;
    filaflat::BlobDictionary mDataBlobs;
 };

@@ -364,7 +364,7 @@ BlobIndex::BlobIndex(ChunkType dictTag, ChunkType matTag, const filaflat::ChunkC
    const auto& offsets = matChunk.getOffsets();
    mShaderRecords.reserve(offsets.size());
    for (auto [key, offset] : offsets) {
-        BinaryEntry info;
+        SpirvEntry info;
        filaflat::MaterialChunk::decodeKey(key, &info.shaderModel, &info.variant, &info.stage);
        info.dictionaryIndex = offset;
        mShaderRecords.emplace_back(info);
@@ -378,7 +378,7 @@ void BlobIndex::writeChunks(ostream& stream) {
        const auto& src = mDataBlobs[record.dictionaryIndex];
        assert(src.size() % 4 == 0);
        const uint32_t* ptr = (const uint32_t*) src.data();
-        record.dictionaryIndex = blobs.addBlob(vector<uint8_t>(ptr, ptr + src.size()));
+        record.dictionaryIndex = blobs.addBlob(vector<uint32_t>(ptr, ptr + src.size() / 4));
    }

    // Adjust start cursor of flatteners to match alignment of output stream.
@@ -391,7 +391,7 @@ void BlobIndex::writeChunks(ostream& stream) {

    // Apply SMOL-V compression and write out the results.
    filamat::ChunkContainer cc;
-    cc.push<MaterialBinaryChunk>(std::move(mShaderRecords), ChunkType::MaterialSpirv);
+    cc.push<MaterialSpirvChunk>(std::move(mShaderRecords));
    cc.push<DictionarySpirvChunk>(std::move(blobs), false);

    Flattener prepass = Flattener::getDryRunner();
@@ -421,9 +421,6 @@ static bool printShaderInfo(ostream& text, const ChunkContainer& container, Chun
        case ChunkType::MaterialMetal:
            text << "Metal shaders:" << endl;
            break;
-        case ChunkType::MaterialMetalLibrary:
-            text << "Metal precompiled shader libraries:" << endl;
-            break;
        default:
            assert(false && "Invalid shader ChunkType");
            break;
@@ -458,9 +455,6 @@ bool TextWriter::writeMaterialInfo(const filaflat::ChunkContainer& container) {
    if (!printShaderInfo(text, container, ChunkType::MaterialMetal)) {
        return false;
    }
-    if (!printShaderInfo(text, container, ChunkType::MaterialMetalLibrary)) {
-        return false;
-    }

    printChunks(text, container);

@@ -76,8 +76,6 @@ static string_view readBlendingMode(string_view cursor, BlendingMode* blending)
        *blending = BlendingMode::MULTIPLY;
    } else if (sz = readPrefix("screen"sv, cursor); sz > 0) {
        *blending = BlendingMode::SCREEN;
-    } else if (sz = readPrefix("custom"sv, cursor); sz > 0) {
-        *blending = BlendingMode::CUSTOM;
    }
    return { cursor.data(), sz };
 }
@@ -38,7 +38,6 @@ set(DIST_HDRS
        ${PUBLIC_HDR_DIR}/${TARGET}/SingleInstanceComponentManager.h
        ${PUBLIC_HDR_DIR}/${TARGET}/Slice.h
        ${PUBLIC_HDR_DIR}/${TARGET}/StructureOfArrays.h
-        ${PUBLIC_HDR_DIR}/${TARGET}/Systrace.h
        ${PUBLIC_HDR_DIR}/${TARGET}/unwindows.h
 )

@@ -88,6 +87,7 @@ if (LINUX OR ANDROID)
    list(APPEND SRCS src/linux/Path.cpp)
 endif()
 if (APPLE)
+    list(APPEND SRCS src/darwin/Exception.mm)
    list(APPEND SRCS src/darwin/Path.mm)
    list(APPEND SRCS src/darwin/Systrace.cpp)
 endif()
@@ -116,11 +116,6 @@ if (WIN32)
    target_link_libraries(${TARGET} PUBLIC Shlwapi)
 endif()

-if (APPLE)
-    # Needed for NSTemporaryDirectory()
-    target_link_libraries(${TARGET} PRIVATE "-framework Foundation")
-endif()
-
 if (LINUX)
    set(THREADS_PREFER_PTHREAD_FLAG ON)
    find_package(Threads REQUIRED)
@@ -128,6 +123,10 @@ if (LINUX)
    target_link_libraries(${TARGET} PRIVATE dl)
 endif()

+if (APPLE AND NOT IOS)
+    target_link_libraries(${TARGET} PRIVATE "-framework Cocoa")
+endif()
+
 # ==================================================================================================
 # Installation
 # ==================================================================================================
@@ -250,71 +250,54 @@ namespace utils {
 */
 class UTILS_PUBLIC Panic {
 public:
-
-    using PanicHandlerCallback = void(*)(void* user, Panic const& panic);
-
-    /**
-     * Sets a user-defined handler for the Panic. If exceptions are enabled, the concrete Panic
-     * object will be thrown upon return; moreover it is acceptable to throw from the provided
-     * callback, but it is unsafe to throw the Panic object itself, since it's just an interface.
-     * It is also acceptable to abort from the callback. If exceptions are not enabled, std::abort()
-     * will be automatically called upon return.
-     *
-     * The PanicHandlerCallback can be called from any thread.
-     *
-     * Caveat: this API can misbehave if is used as a static library in multiple translation units,
-     * some of these translation units might not see the callback.
-     *
-     * @param handler pointer to the user defined handler for the Panic
-     * @param user  user pointer given back to the callback
-     */
-    static void setPanicHandler(PanicHandlerCallback handler, void* user) noexcept;
-
-
    virtual ~Panic() noexcept;

    /**
-     * @return a formatted and detailed description of the error including all available
-     *         information.
+     * @return a detailed description of the error
     * @see std::exception
     */
    virtual const char* what() const noexcept = 0;

    /**
-     * Get the reason string for the panic
-     * @return a C string containing the reason for the panic
+     * Get the reason for the panic.
+     * @return a C string containing the reason for the panic.
     */
    virtual const char* getReason() const noexcept = 0;

    /**
-     * Get the function name where the panic was detected. On debug build the fully qualified
-     * function name is returned; on release builds only the function name is.
+     * Get the type of the panic. When compiled without RTTI, this returns a generic type name.
+     * @return a C string containing the type of the panic.
+     */
+    virtual const char* getType() const noexcept = 0;
+
+    /**
+     * Get the function name where the panic was detected
     * @return a C string containing the function name where the panic was detected
     */
    virtual const char* getFunction() const noexcept = 0;

    /**
-     * Get the file name where the panic was detected. Only available on debug builds.
+     * Get the file name where the panic was detected
     * @return a C string containing the file name where the panic was detected
     */
    virtual const char* getFile() const noexcept = 0;

    /**
-     * Get the line number in the file where the panic was detected. Only available on debug builds.
+     * Get the line number in the file where the panic was detected
     * @return an integer containing the line number in the file where the panic was detected
     */
    virtual int getLine() const noexcept = 0;

-    /**
-     * Get the CallStack when the panic was detected if available.
-     * @return the CallStack when the panic was detected
-     */
-    virtual const CallStack& getCallStack() const noexcept = 0;
-
    /**
     * Logs this exception to the system-log
     */
    virtual void log() const noexcept = 0;
+
+    /**
+     * Get the CallStack when the panic was detected
+     * @return the CallStack when the panic was detected
+     */
+    virtual const CallStack& getCallStack() const noexcept = 0;
 };

 // -----------------------------------------------------------------------------------------------
@@ -335,6 +318,7 @@ public:

    // Panic interface
    const char* getReason() const noexcept override;
+    const char* getType() const noexcept override;
    const char* getFunction() const noexcept override;
    const char* getFile() const noexcept override;
    int getLine() const noexcept override;
@@ -399,7 +383,8 @@ private:
    char const* const m_function = nullptr;
    char const* const m_file = nullptr;
    const int m_line = -1;
-    mutable std::string m_msg;
+    std::string m_msg;
+    std::string m_type;
 };

 namespace details {
@@ -97,7 +97,7 @@
 namespace utils {
 namespace details {

-class UTILS_PUBLIC Systrace {
+class Systrace {
   public:

    enum tags {
@@ -220,7 +220,7 @@ class UTILS_PUBLIC Systrace {

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

-class UTILS_PUBLIC ScopedTrace {
+class ScopedTrace {
 public:
    // we don't inline this because it's relatively heavy due to a global check
    ScopedTrace(uint32_t tag, const char* name) noexcept: mTrace(tag), mTag(tag) {
@@ -14,18 +14,16 @@
 * limitations under the License.
 */

-#ifndef TNT_MATEDIT_EXTERNALCOMPILE_H
-#define TNT_MATEDIT_EXTERNALCOMPILE_H
+#ifndef TNT_UTILS_DARWIN_EXCEPTION_H
+#define TNT_UTILS_DARWIN_EXCEPTION_H

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

-#include <string>
-#include <vector>
+namespace utils::details {

-namespace matedit {
+void throwDarwinException(const Panic& exception) UTILS_NORETURN;

-int externalCompile(utils::Path input, utils::Path output, std::vector<std::string> args);
+}

-} // namespace matedit
-
-#endif
+#endif // TNT_UTILS_DARWIN_EXCEPTION_H
@@ -20,7 +20,6 @@
 #include <utils/compiler.h>

 namespace utils {
-UTILS_PUBLIC
 void panic(const char *func, const char * file, int line, const char *assertion) noexcept;
 } // namespace filament

@@ -14,64 +14,21 @@
 * limitations under the License.
 */

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

+#if defined(__APPLE__)
+#include <utils/darwin/Exception.h>
+#endif
+
+#include <atomic>
+#include <cstdarg>
 #include <cstdlib>
 #include <cstring>
-#include <mutex>
-#include <new>
-#include <string>
-#include <utility>

-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
+#include <utils/Log.h>

 namespace utils {

-// ------------------------------------------------------------------------------------------------
-
-class UserPanicHandler {
-    struct CallBack {
-        Panic::PanicHandlerCallback handler = nullptr;
-        void* user = nullptr;
-        void call(Panic const& panic) const noexcept {
-            if (UTILS_UNLIKELY(handler)) {
-                handler(user, panic);
-            }
-        }
-    };
-
-    mutable std::mutex mLock{};
-    CallBack mCallBack{};
-
-    CallBack getCallback() const noexcept {
-        std::lock_guard const lock(mLock);
-        return mCallBack;
-    }
-
-public:
-    static UserPanicHandler& get() noexcept {
-        static UserPanicHandler data;
-        return data;
-    }
-
-    void call(Panic const& panic) const noexcept {
-        getCallback().call(panic);
-    }
-
-    void set(Panic::PanicHandlerCallback handler, void* user) noexcept {
-        std::lock_guard const lock(mLock);
-        mCallBack = { handler, user };
-    }
-};
-
-// ------------------------------------------------------------------------------------------------
-
 static std::string formatString(const char* format, va_list args) noexcept {
    std::string reason;

@@ -82,12 +39,10 @@ static std::string formatString(const char* format, va_list args) noexcept {

    if (n >= 0) {
        ++n; // for the nul-terminating char
-        char* const buf = new(std::nothrow) char[n];
-        if (buf) {
-            vsnprintf(buf, size_t(n), format, args);
-            reason.assign(buf);
-            delete [] buf;
-        }
+        char* buf = new char[n];
+        vsnprintf(buf, size_t(n), format, args);
+        reason.assign(buf);
+        delete [] buf;
    }
    return reason;
 }
@@ -112,16 +67,8 @@ static std::string panicString(
 #endif
 }

-// ------------------------------------------------------------------------------------------------
-
 Panic::~Panic() noexcept = default;

-void Panic::setPanicHandler(PanicHandlerCallback handler, void* user) noexcept {
-    UserPanicHandler::get().set(handler, user);
-}
-
-// ------------------------------------------------------------------------------------------------
-
 template<typename T>
 TPanic<T>::TPanic(std::string reason) :
    m_reason(std::move(reason)) {
@@ -150,6 +97,11 @@ const char* TPanic<T>::getReason() const noexcept {
    return m_reason.c_str();
 }

+template<typename T>
+const char* TPanic<T>::getType() const noexcept {
+    return m_type.c_str();
+}
+
 template<typename T>
 const char* TPanic<T>::getFunction() const noexcept {
    return m_function;
@@ -178,13 +130,12 @@ void TPanic<T>::log() const noexcept {

 template<typename T>
 void TPanic<T>::buildMessage() {
-    std::string type;
 #if UTILS_HAS_RTTI
-    type = CallStack::demangleTypeName(typeid(T).name()).c_str();
+    m_type = CallStack::demangleTypeName(typeid(T).name()).c_str();
 #else
-    type = "Panic";
+    m_type = "Panic";
 #endif
-    m_msg = panicString(type, m_function, m_line, m_file, m_reason.c_str());
+    m_msg = panicString(m_type, m_function, m_line, m_file, m_reason.c_str());
 }

 UTILS_ALWAYS_INLINE
@@ -197,22 +148,18 @@ template<typename T>
 void TPanic<T>::panic(char const* function, char const* file, int line, const char* format, ...) {
    va_list args;
    va_start(args, format);
-    std::string const reason(formatString(format, args));
+    std::string reason(formatString(format, args));
    va_end(args);
    T e(function, formatFile(file), line, reason);
-
-    // always log the Panic at the point it is detected
    e.log();

-    // Call the user provided handler
-    UserPanicHandler::get().call(e);
-
-    // if exceptions are enabled, throw now.
-#ifdef __EXCEPTIONS
-    throw e;
+#if defined(__APPLE__)
+    details::throwDarwinException(e);
 #endif

-    // and finally abort if we somehow get here
+#ifdef __EXCEPTIONS
+        throw e;
+#endif
    std::abort();
 }

@@ -223,7 +170,7 @@ namespace details {
 void panicLog(char const* function, char const* file, int line, const char* format, ...) noexcept {
    va_list args;
    va_start(args, format);
-    std::string const reason(formatString(format, args));
+    std::string reason(formatString(format, args));
    va_end(args);

    const std::string msg = panicString("" /* no extra message */,
@@ -0,0 +1,35 @@
+/*
+ * 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 <utils/Panic.h>
+
+#include <Foundation/Foundation.h>
+
+namespace utils::details {
+
+void throwDarwinException(const Panic& exception) {
+    // Turn this exception into an NSException and raise it.
+    // This is allowed even if we're built without Objective-C exceptions (-fno-objc-exceptions).
+    [[NSException exceptionWithName:@(exception.getType())
+                             reason:@(exception.getReason())
+                           userInfo:@{
+                               @"file" : @(exception.getFile()),
+                               @"line" : @(exception.getLine()),
+                               @"function" : @(exception.getFunction())
+                           }] raise];
+}
+
+}
@@ -671,7 +671,6 @@ static bool processBlending(MaterialBuilder& builder, const JsonishValue& value)
        { "fade", MaterialBuilder::BlendingMode::FADE },
        { "multiply", MaterialBuilder::BlendingMode::MULTIPLY },
        { "screen", MaterialBuilder::BlendingMode::SCREEN },
-        { "custom", MaterialBuilder::BlendingMode::CUSTOM },
    };
    auto jsonString = value.toJsonString();
    if (!isStringValidEnum(strToEnum, jsonString->getString())) {
@@ -682,52 +681,6 @@ static bool processBlending(MaterialBuilder& builder, const JsonishValue& value)
    return true;
 }

-static bool processBlendFunction(MaterialBuilder& builder, const JsonishValue& value) {
-    static const std::unordered_map<std::string, MaterialBuilder::BlendFunction> strToEnum{
-            { "zero",             MaterialBuilder::BlendFunction::ZERO },
-            { "one",              MaterialBuilder::BlendFunction::ONE },
-            { "srcColor",         MaterialBuilder::BlendFunction::SRC_COLOR },
-            { "oneMinusSrcColor", MaterialBuilder::BlendFunction::ONE_MINUS_SRC_COLOR },
-            { "dstColor",         MaterialBuilder::BlendFunction::DST_COLOR },
-            { "oneMinusDstColor", MaterialBuilder::BlendFunction::ONE_MINUS_DST_COLOR },
-            { "srcAlpha",         MaterialBuilder::BlendFunction::SRC_ALPHA },
-            { "oneMinusSrcAlpha", MaterialBuilder::BlendFunction::ONE_MINUS_SRC_ALPHA },
-            { "dstAlpha",         MaterialBuilder::BlendFunction::DST_ALPHA },
-            { "oneMinusDstAlpha", MaterialBuilder::BlendFunction::ONE_MINUS_DST_ALPHA },
-            { "srcAlphaSaturate", MaterialBuilder::BlendFunction::SRC_ALPHA_SATURATE },
-    };
-
-    if (value.getType() != JsonishValue::Type::OBJECT) {
-        std::cerr << "blendFunction must be an OBJECT." << std::endl;
-    }
-
-    JsonishObject const* const jsonObject = value.toJsonObject();
-
-    MaterialBuilder::BlendFunction srcRGB, srcA, dstRGB, dstA;
-    std::pair<const char*, MaterialBuilder::BlendFunction*> functions[] = {
-            { "srcRGB", &srcRGB },
-            { "srcA",   &srcA },
-            { "dstRGB", &dstRGB },
-            { "dstA",   &dstA },
-    };
-
-    for (auto&& entry : functions) {
-        const char* key = entry.first;
-        const JsonishValue* v = jsonObject->getValue(key);
-        if (!v) {
-            std::cerr << "blendFunction: entry without '" << key << "' key." << std::endl;
-            return false;
-        }
-        if (v->getType() != JsonishValue::STRING) {
-            std::cerr << "blendFunction: '" << key << "' value must be STRING." << std::endl;
-            return false;
-        }
-        *entry.second = stringToEnum(strToEnum, v->toJsonString()->getString());
-    }
-    builder.customBlendFunctions(srcRGB, srcA, dstRGB, dstA);
-    return true;
-}
-
 static bool processPostLightingBlending(MaterialBuilder& builder, const JsonishValue& value) {
    static const std::unordered_map<std::string, MaterialBuilder::BlendingMode> strToEnum {
        { "add", MaterialBuilder::BlendingMode::ADD },
@@ -1241,7 +1194,6 @@ ParametersProcessor::ParametersProcessor() {
    mParameters["variables"]                     = { &processVariables, Type::ARRAY };
    mParameters["requires"]                      = { &processRequires, Type::ARRAY };
    mParameters["blending"]                      = { &processBlending, Type::STRING };
-    mParameters["blendFunction"]                 = { &processBlendFunction, Type::OBJECT };
    mParameters["postLightingBlending"]          = { &processPostLightingBlending, Type::STRING };
    mParameters["vertexDomain"]                  = { &processVertexDomain, Type::STRING };
    mParameters["culling"]                       = { &processCulling, Type::STRING };
--- a/Show More
+++ b/Show More