backend-test: centralize read pixel logic

we use a different hook for the draw() call when on an ES2 context,
this eliminates completely the overhead of supporting ES2 for the draw
call. draw calls are expected to be the most common calls.

BUILDING.md

@@ -73,7 +73,6 @@ The following CMake options are boolean options specific to Filament:
 - `FILAMENT_SUPPORTS_VULKAN`:      Include the Vulkan backend
 - `FILAMENT_INSTALL_BACKEND_TEST`: Install the backend test library so it can be consumed on iOS
 - `FILAMENT_USE_EXTERNAL_GLES3`:   Experimental: Compile Filament against OpenGL ES 3
-- `FILAMENT_USE_SWIFTSHADER`:      Compile Filament against SwiftShader
 - `FILAMENT_SKIP_SAMPLES`:         Don't build sample apps
 
 To turn an option on or off:
@@ -426,7 +425,7 @@ value is the desired roughness between 0 and 1.
 
 ## Generating C++ documentation
 
-To generate the documentation you must first install `doxygen` and `graphviz`, then run the 
+To generate the documentation you must first install `doxygen` and `graphviz`, then run the
 following commands:
 
 ```shell
@@ -436,32 +435,27 @@ doxygen docs/doxygen/filament.doxygen
 
 Finally simply open `docs/html/index.html` in your web browser.
 
-## SwiftShader
+## Software Rasterization
 
-To try out Filament's Vulkan support with SwiftShader, first build SwiftShader and set the
-`SWIFTSHADER_LD_LIBRARY_PATH` variable to the folder that contains `libvk_swiftshader.dylib`:
+We have tested swiftshader for running software rasterization on the Vulkan backend. To use this,
+please first make sure that the [Vulkan SDK](https://www.lunarg.com/vulkan-sdk/) is installed on
+your machine. If you are doing a manual installation of the SDK on Linux, you will have to source
+`setup-env.sh` in the SDK's root folder to make sure the Vulkan loader is the first lib loaded.
+
+### Swiftshader (Vulkan) [tested on macOS and Linux]
+
+First, build SwiftShader
 
 ```shell
 git clone https://github.com/google/swiftshader.git
 cd swiftshader/build
 cmake .. &&  make -j
-export SWIFTSHADER_LD_LIBRARY_PATH=`pwd`
 ```
 
-Next, go to your Filament repo and use the [easy build](#easy-build) script with `-t`.
-
-## SwiftShader for CI
-
-Continuous testing turnaround can be quite slow if you need to build SwiftShader from scratch, so we
-provide an Ubuntu-based Docker image that has it already built. The Docker image also includes
-everything necessary for building Filament. You can fetch and run the image as follows:
-
+and then set `VK_ICD_FILENAMES` to the ICD json produced in the build. For example,
 ```shell
-docker pull ghcr.io/filament-assets/swiftshader
-docker run -it ghcr.io/filament-assets/swiftshader
+export VK_ICD_FILENAMES=/Users/user/swiftshader/build/Darwin/vk_swiftshader_icd.json
 ```
 
-To do more with the container, see the helper script at `build/swiftshader/test.sh`.
+Build Filament as normal and use the vulkan backend.
 
-If you are a team member, you can update the public image to the latest SwiftShader by
-following the instructions at the top of `build/swiftshader/Dockerfile`.

CMakeLists.txt

@@ -21,8 +21,6 @@ project(TNT)
 # ==================================================================================================
 option(FILAMENT_USE_EXTERNAL_GLES3 "Experimental: Compile Filament against OpenGL ES 3" OFF)
 
-option(FILAMENT_USE_SWIFTSHADER "Compile Filament against SwiftShader" OFF)
-
 option(FILAMENT_ENABLE_LTO "Enable link-time optimizations if supported by the compiler" OFF)
 
 option(FILAMENT_SKIP_SAMPLES "Don't build samples" OFF)
@@ -45,6 +43,8 @@ option(FILAMENT_ENABLE_TSAN "Enable Thread Sanitizer" OFF)
 
 option(FILAMENT_ENABLE_FEATURE_LEVEL_0 "Enable Feature Level 0" ON)
 
+option(FILAMENT_ENABLE_MULTIVIEW "Enable multiview for Filament" OFF)
+
 set(FILAMENT_NDK_VERSION "" CACHE STRING
     "Android NDK version or version prefix to be used when building for Android."
 )
@@ -69,6 +69,10 @@ set(FILAMENT_METAL_HANDLE_ARENA_SIZE_IN_MB "8" CACHE STRING
     "Size of the Metal handle arena, default 8."
 )
 
+set(FILAMENT_BACKEND_DEBUG_FLAG "" CACHE STRING
+    "A debug flag meant for enabling/disabling backend debugging paths"
+)
+
 # Enable exceptions by default in spirv-cross.
 set(SPIRV_CROSS_EXCEPTIONS_TO_ASSERTIONS OFF)
 
@@ -139,11 +143,6 @@ if (LINUX)
             add_definitions(-DFILAMENT_SUPPORTS_XCB)
         endif()
 
-        # Default Swiftshader build does not enable the xlib extension
-        if (FILAMENT_SUPPORTS_XLIB AND FILAMENT_USE_SWIFTSHADER)
-            set(FILAMENT_SUPPORTS_XLIB OFF)
-        endif()
-
         if (FILAMENT_SUPPORTS_XLIB)
             add_definitions(-DFILAMENT_SUPPORTS_XLIB)
         endif()
@@ -167,15 +166,6 @@ if (NOT ANDROID AND NOT WEBGL AND NOT IOS AND NOT FILAMENT_LINUX_IS_MOBILE)
     set(IS_HOST_PLATFORM TRUE)
 endif()
 
-if (IOS)
-    # Remove the headerpad_max_install_names linker flag on iOS. It causes warnings when linking
-    # executables with bitcode.
-    string(REPLACE "-Wl,-headerpad_max_install_names" "" CMAKE_C_LINK_FLAGS ${CMAKE_C_LINK_FLAGS})
-    string(REPLACE "-Wl,-headerpad_max_install_names" "" CMAKE_CXX_LINK_FLAGS ${CMAKE_CXX_LINK_FLAGS})
-    string(REPLACE "-Wl,-headerpad_max_install_names" "" CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS ${CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS})
-    string(REPLACE "-Wl,-headerpad_max_install_names" "" CMAKE_SHARED_LIBRARY_CREATE_CXX_FLAGS ${CMAKE_SHARED_LIBRARY_CREATE_CXX_FLAGS})
-endif()
-
 if (WIN32)
     # Link statically against c/c++ lib to avoid missing redistriburable such as
     # "VCRUNTIME140.dll not found. Try reinstalling the app.", but give users
@@ -330,10 +320,6 @@ if (FILAMENT_SUPPORTS_EGL_ON_LINUX)
   set(EGL TRUE)
 endif()
 
-if (FILAMENT_USE_SWIFTSHADER)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DFILAMENT_USE_SWIFTSHADER")
-endif()
-
 if (WIN32)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_USE_MATH_DEFINES=1")
 endif()
@@ -387,8 +373,9 @@ endif()
 if (NOT MSVC AND NOT IOS)
     # Omitting stack frame pointers prevents the generation of readable stack traces in crash reports on iOS
     set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -fomit-frame-pointer")
+endif()
 
-    # These aren't compatible with -fembed-bitcode (and seem to have no effect on Apple platforms anyway)
+if (NOT MSVC)
     set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -ffunction-sections -fdata-sections")
 endif()
 
@@ -453,8 +440,13 @@ endif()
 if (NOT WEBGL)
     set(GC_SECTIONS "-Wl,--gc-sections")
 endif()
+
 set(B_SYMBOLIC_FUNCTIONS "-Wl,-Bsymbolic-functions")
 
+if (ANDROID)
+set(BINARY_ALIGNMENT "-Wl,-z,max-page-size=16384")
+endif()
+
 if (APPLE)
     set(GC_SECTIONS "-Wl,-dead_strip")
     set(B_SYMBOLIC_FUNCTIONS "")
@@ -468,7 +460,7 @@ if (APPLE)
 endif()
 
 set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${GC_SECTIONS}")
-set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${GC_SECTIONS} ${B_SYMBOLIC_FUNCTIONS}")
+set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${GC_SECTIONS} ${B_SYMBOLIC_FUNCTIONS} ${BINARY_ALIGNMENT}")
 
 if (WEBGL_PTHREADS)
     set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -pthread")
@@ -531,6 +523,27 @@ else()
     option(FILAMENT_DISABLE_MATOPT "Disable material optimizations" ON)
 endif()
 
+# This only affects the prebuilt shader files in gltfio and samples, not filament library.
+# The value can be either "instanced" or "multiview".
+set(FILAMENT_SAMPLES_STEREO_TYPE "instanced" CACHE STRING
+    "Stereoscopic type that shader files in gltfio and samples are built for."
+)
+string(TOLOWER "${FILAMENT_SAMPLES_STEREO_TYPE}" FILAMENT_SAMPLES_STEREO_TYPE)
+if (NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "instanced" AND NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview")
+    message(FATAL_ERROR "Invalid stereo type: \"${FILAMENT_SAMPLES_STEREO_TYPE}\" choose either \"instanced\" or \"multiview\" ")
+endif ()
+
+# Compiling samples for multiview implies enabling multiview feature as well.
+if (FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview")
+    set(FILAMENT_ENABLE_MULTIVIEW ON)
+endif ()
+
+# Define backend flag for debug only
+if (CMAKE_BUILD_TYPE STREQUAL "Debug" AND NOT FILAMENT_BACKEND_DEBUG_FLAG STREQUAL "")
+    add_definitions(-DFILAMENT_BACKEND_DEBUG_FLAG=${FILAMENT_BACKEND_DEBUG_FLAG})
+    unset(FILAMENT_BACKEND_DEBUG_FLAG)
+endif()
+
 # ==================================================================================================
 # Material compilation flags
 # ==================================================================================================
@@ -659,21 +672,6 @@ else()
     set(IMPORT_EXECUTABLES ${FILAMENT}/${IMPORT_EXECUTABLES_DIR}/ImportExecutables-${CMAKE_BUILD_TYPE}.cmake)
 endif()
 
-# ==================================================================================================
-# Try to find Vulkan if the SDK is installed, otherwise fall back to the bundled version.
-# This needs to stay in our top-level CMakeLists because it sets up variables that are used by the
-# "bluevk" and "samples" targets.
-# ==================================================================================================
-
-if (FILAMENT_USE_SWIFTSHADER)
-    if (NOT FILAMENT_SUPPORTS_VULKAN)
-        message(ERROR "SwiftShader is only useful when Vulkan is enabled.")
-    endif()
-    find_library(SWIFTSHADER_VK NAMES vk_swiftshader HINTS "$ENV{SWIFTSHADER_LD_LIBRARY_PATH}")
-    message(STATUS "Found SwiftShader VK library in: ${SWIFTSHADER_VK}.")
-    add_definitions(-DFILAMENT_VKLIBRARY_PATH=\"${SWIFTSHADER_VK}\")
-endif()
-
 # ==================================================================================================
 # Common Functions
 # ==================================================================================================
@@ -802,6 +800,9 @@ 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)

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.50.6'
+    implementation 'com.google.android.filament:filament-android:1.51.8'
 }
 ```
 
@@ -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', '~> 1.50.6'
+pod 'Filament', '~> 1.51.8'
 ```
 
 ### Snapshots

RELEASE_NOTES.md

@@ -7,6 +7,44 @@ 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).
 
+## v1.51.9
+
+
+## v1.51.8
+
+- filagui: Fix regression which broke WebGL
+- Add a new Engine::Config setting to control preferred shader language
+- Add `getEyeIndex` vertex API
+- ios: Remove bitcode from iOS builds
+
+## v1.51.7
+
+- Add new matedit tool
+- filagui: Support rendering `GL_TEXTURE_EXTERNAL_OES` textures.
+- `setFrameScheduledCallback` now takes a `utils::Invocable`.
+- engine: Add `isPaused()`
+
+## v1.51.6
+
+- Add new matedit tool
+- filagui: Support rendering `GL_TEXTURE_EXTERNAL_OES` textures.
+
+## v1.51.5
+
+
+## v1.51.4
+
+
+## v1.51.3
+
+
+## v1.51.2
+
+- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`
+
+## v1.51.1
+
+
 ## v1.51.0
 
 - materials: add support for post-lighting mix factor (b/328498606) [⚠️ **New Material Version**]

android/filamat-android/CMakeLists.txt

@@ -38,6 +38,7 @@ set(FILAMAT_INCLUDE_DIRS
 include_directories(${FILAMENT_DIR}/include)
 
 set(CMAKE_SHARED_LINKER_FLAGS_RELEASE "${CMAKE_SHARED_LINKER_FLAGS_RELEASE} -Wl,--version-script=${CMAKE_SOURCE_DIR}/libfilamat-jni.map")
+set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size=16384")
 
 add_library(filamat-jni SHARED src/main/cpp/MaterialBuilder.cpp)
 target_include_directories(filamat-jni PRIVATE ${FILAMAT_INCLUDE_DIRS})

android/filament-android/CMakeLists.txt

@@ -59,6 +59,7 @@ endif()
 
 set(VERSION_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/libfilament-jni.map")
 set(CMAKE_SHARED_LINKER_FLAGS_RELEASE "${CMAKE_SHARED_LINKER_FLAGS_RELEASE} -Wl,--version-script=${VERSION_SCRIPT}")
+set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size=16384")
 
 add_library(filament-jni SHARED
     src/main/cpp/BufferObject.cpp

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

@@ -342,6 +342,21 @@ Java_com_google_android_filament_Engine_nIsValidMaterial(JNIEnv*, jclass,
     return (jboolean)engine->isValid((Material*)nativeMaterial);
 }
 
+extern "C" JNIEXPORT jboolean JNICALL
+Java_com_google_android_filament_Engine_nIsValidMaterialInstance(JNIEnv*, jclass,
+        jlong nativeEngine, jlong nativeMaterial, jlong nativeMaterialInstance) {
+    Engine* engine = (Engine *)nativeEngine;
+    return (jboolean)engine->isValid((Material*)nativeMaterial,
+            (MaterialInstance*)nativeMaterialInstance);
+}
+
+extern "C" JNIEXPORT jboolean JNICALL
+        Java_com_google_android_filament_Engine_nIsValidExpensiveMaterialInstance(JNIEnv*, jclass,
+        jlong nativeEngine, jlong nativeMaterialInstance) {
+    Engine* engine = (Engine *)nativeEngine;
+    return (jboolean)engine->isValidExpensive((MaterialInstance*)nativeMaterialInstance);
+}
+
 extern "C" JNIEXPORT jboolean JNICALL
 Java_com_google_android_filament_Engine_nIsValidSkybox(JNIEnv*, jclass,
         jlong nativeEngine, jlong nativeSkybox) {
@@ -391,6 +406,20 @@ Java_com_google_android_filament_Engine_nFlush(JNIEnv*, jclass,
     engine->flush();
 }
 
+extern "C" JNIEXPORT jboolean JNICALL
+Java_com_google_android_filament_Engine_nIsPaused(JNIEnv*, jclass,
+        jlong nativeEngine) {
+    Engine* engine = (Engine*) nativeEngine;
+    return (jboolean)engine->isPaused();
+}
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Engine_nSetPaused(JNIEnv*, jclass,
+        jlong nativeEngine, jboolean paused) {
+    Engine* engine = (Engine*) nativeEngine;
+    engine->setPaused(paused);
+}
+
 // Managers...
 
 extern "C" JNIEXPORT jlong JNICALL
@@ -487,7 +516,10 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
         jlong jobSystemThreadCount,
         jlong textureUseAfterFreePoolSize, jboolean disableParallelShaderCompile,
         jint stereoscopicType, jlong stereoscopicEyeCount,
-        jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge) {
+        jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge,
+        jboolean disableHandleUseAfterFreeCheck,
+        jint preferredShaderLanguage,
+        jboolean forceGLES2Context) {
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
     Engine::Config config = {
             .commandBufferSizeMB = (uint32_t) commandBufferSizeMB,
@@ -502,6 +534,9 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
             .stereoscopicEyeCount = (uint8_t) stereoscopicEyeCount,
             .resourceAllocatorCacheSizeMB = (uint32_t) resourceAllocatorCacheSizeMB,
             .resourceAllocatorCacheMaxAge = (uint8_t) resourceAllocatorCacheMaxAge,
+            .disableHandleUseAfterFreeCheck = (bool) disableHandleUseAfterFreeCheck,
+            .preferredShaderLanguage = (Engine::Config::ShaderLanguage) preferredShaderLanguage,
+            .forceGLES2Context = (bool) forceGLES2Context,
     };
     builder->config(&config);
 }
@@ -518,6 +553,12 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
     builder->sharedContext((void*) sharedContext);
 }
 
+extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBuilderPaused(
+        JNIEnv*, jclass, jlong nativeBuilder, jboolean paused) {
+    Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
+    builder->paused((bool) paused);
+}
+
 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Engine_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder) {
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;

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

@@ -104,6 +104,14 @@ Java_com_google_android_filament_RenderableManager_nBuilderGeometry__JIIJJIIII(J
             (size_t) count);
 }
 
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_RenderableManager_nBuilderGeometryType(JNIEnv*, jclass,
+        jlong nativeBuilder, int type) {
+    RenderableManager::Builder *builder = (RenderableManager::Builder *) nativeBuilder;
+    builder->geometryType((RenderableManager::Builder::GeometryType)type);
+}
+
 extern "C"
 JNIEXPORT void JNICALL
 Java_com_google_android_filament_RenderableManager_nBuilderMaterial(JNIEnv*, jclass,

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

@@ -49,6 +49,12 @@ Java_com_google_android_filament_View_nSetCamera(JNIEnv*, jclass,
     view->setCamera(camera);
 }
 
+extern "C" JNIEXPORT jboolean JNICALL
+Java_com_google_android_filament_View_nHasCamera(JNIEnv*, jclass, jlong nativeView) {
+    View* view = (View*) nativeView;
+    return (jboolean)view->hasCamera();
+}
+
 extern "C" JNIEXPORT void JNICALL
 Java_com_google_android_filament_View_nSetColorGrading(JNIEnv*, jclass,
         jlong nativeView, jlong nativeColorGrading) {

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

@@ -224,7 +224,10 @@ public class Engine {
                     config.jobSystemThreadCount,
                     config.textureUseAfterFreePoolSize, config.disableParallelShaderCompile,
                     config.stereoscopicType.ordinal(), config.stereoscopicEyeCount,
-                    config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge);
+                    config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge,
+                    config.disableHandleUseAfterFreeCheck,
+                    config.preferredShaderLanguage.ordinal(),
+                    config.forceGLES2Context);
             return this;
         }
 
@@ -239,6 +242,20 @@ public class Engine {
             return this;
         }
 
+        /**
+         * Sets the initial paused state of the rendering thread.
+         *
+         * <p>Warning: This is an experimental API. See {@link Engine#setPaused(boolean)} for
+         * caveats.
+         *
+         * @param paused Whether to start the rendering thread paused.
+         * @return A reference to this Builder for chaining calls.
+         */
+        public Builder paused(boolean paused) {
+            nSetBuilderPaused(mNativeBuilder, paused);
+            return this;
+        }
+
         /**
          * Creates an instance of Engine
          *
@@ -408,6 +425,40 @@ public class Engine {
          * This value determines for how many frames are texture entries kept in the cache.
          */
         public long resourceAllocatorCacheMaxAge = 2;
+
+        /*
+         * Disable backend handles use-after-free checks.
+         */
+        public boolean disableHandleUseAfterFreeCheck = false;
+
+        /*
+         * Sets a preferred shader language for Filament to use.
+         *
+         * The Metal backend supports two shader languages: MSL (Metal Shading Language) and
+         * METAL_LIBRARY (precompiled .metallib). This option controls which shader language is
+         * used when materials contain both.
+         *
+         * By default, when preferredShaderLanguage is unset, Filament will prefer METAL_LIBRARY
+         * shaders if present within a material, falling back to MSL. Setting
+         * preferredShaderLanguage to ShaderLanguage::MSL will instead instruct Filament to check
+         * for the presence of MSL in a material first, falling back to METAL_LIBRARY if MSL is not
+         * present.
+         *
+         * When using a non-Metal backend, setting this has no effect.
+         */
+        public enum ShaderLanguage {
+            DEFAULT,
+            MSL,
+            METAL_LIBRARY,
+        };
+        public ShaderLanguage preferredShaderLanguage = ShaderLanguage.DEFAULT;
+
+        /*
+         * When the OpenGL ES backend is used, setting this value to true will force a GLES2.0
+         * context if supported by the Platform, or if not, will have the backend pretend
+         * it's a GLES2 context. Ignored on other backends.
+         */
+        public boolean forceGLES2Context = false;
     }
 
     private Engine(long nativeEngine, Config config) {
@@ -654,7 +705,7 @@ public class Engine {
      */
     @NonNull
     public SwapChain createSwapChain(@NonNull Object surface) {
-        return createSwapChain(surface, SwapChain.CONFIG_DEFAULT);
+        return createSwapChain(surface, SwapChainFlags.CONFIG_DEFAULT);
     }
 
     /**
@@ -662,15 +713,15 @@ public class Engine {
      *
      * @param surface on Android, <b>must be</b> an instance of {@link android.view.Surface}
      *
-     * @param flags configuration flags, see {@link SwapChain}
+     * @param flags configuration flags, see {@link SwapChainFlags}
      *
      * @return a newly created {@link SwapChain} object
      *
      * @exception IllegalStateException can be thrown if the SwapChain couldn't be created
      *
-     * @see SwapChain#CONFIG_DEFAULT
-     * @see SwapChain#CONFIG_TRANSPARENT
-     * @see SwapChain#CONFIG_READABLE
+     * @see SwapChainFlags#CONFIG_DEFAULT
+     * @see SwapChainFlags#CONFIG_TRANSPARENT
+     * @see SwapChainFlags#CONFIG_READABLE
      *
      */
     @NonNull
@@ -688,21 +739,22 @@ public class Engine {
      *
      * @param width  width of the rendering buffer
      * @param height height of the rendering buffer
-     * @param flags  configuration flags, see {@link SwapChain}
+     * @param flags  configuration flags, see {@link SwapChainFlags}
      *
      * @return a newly created {@link SwapChain} object
      *
      * @exception IllegalStateException can be thrown if the SwapChain couldn't be created
      *
-     * @see SwapChain#CONFIG_DEFAULT
-     * @see SwapChain#CONFIG_TRANSPARENT
-     * @see SwapChain#CONFIG_READABLE
+     * @see SwapChainFlags#CONFIG_DEFAULT
+     * @see SwapChainFlags#CONFIG_TRANSPARENT
+     * @see SwapChainFlags#CONFIG_READABLE
      *
      */
     @NonNull
     public SwapChain createSwapChain(int width, int height, long flags) {
         if (width >= 0 && height >= 0) {
-            long nativeSwapChain = nCreateSwapChainHeadless(getNativeObject(), width, height, flags);
+            long nativeSwapChain =
+                nCreateSwapChainHeadless(getNativeObject(), width, height, flags);
             if (nativeSwapChain == 0) throw new IllegalStateException("Couldn't create SwapChain");
             return new SwapChain(nativeSwapChain, null);
         }
@@ -714,11 +766,12 @@ public class Engine {
      *
      * @param surface a properly initialized {@link NativeSurface}
      *
-     * @param flags configuration flags, see {@link SwapChain}
+     * @param flags configuration flags, see {@link SwapChainFlags}
      *
      * @return a newly created {@link SwapChain} object
      *
-     * @exception IllegalStateException can be thrown if the {@link SwapChain} couldn't be created
+     * @exception IllegalStateException can be thrown if the {@link SwapChainFlags} couldn't be
+     *            created
      */
     @NonNull
     public SwapChain createSwapChainFromNativeSurface(@NonNull NativeSurface surface, long flags) {
@@ -827,6 +880,24 @@ public class Engine {
         return nIsValidMaterial(getNativeObject(), object.getNativeObject());
     }
 
+    /**
+     * Returns whether the object is valid.
+     * @param object Object to check for validity
+     * @return returns true if the specified object is valid.
+     */
+    public boolean isValidMaterialInstance(@NonNull Material ma, MaterialInstance mi) {
+        return nIsValidMaterialInstance(getNativeObject(), ma.getNativeObject(), mi.getNativeObject());
+    }
+
+    /**
+     * Returns whether the object is valid.
+     * @param object Object to check for validity
+     * @return returns true if the specified object is valid.
+     */
+    public boolean isValidExpensiveMaterialInstance(@NonNull MaterialInstance object) {
+        return nIsValidExpensiveMaterialInstance(getNativeObject(), object.getNativeObject());
+    }
+
     /**
      * Returns whether the object is valid.
      * @param object Object to check for validity
@@ -1187,6 +1258,33 @@ public class Engine {
         nFlush(getNativeObject());
     }
 
+    /**
+     * Get paused state of rendering thread.
+     *
+     * <p>Warning: This is an experimental API.
+     *
+     * @see #setPaused
+     */
+    public boolean isPaused() {
+        return nIsPaused(getNativeObject());
+    }
+
+    /**
+     * Pause or resume the rendering thread.
+     *
+     * <p>Warning: This is an experimental API. In particular, note the following caveats.
+     *
+     * <ul><li>
+     * Buffer callbacks will never be called as long as the rendering thread is paused.
+     * Do not rely on a buffer callback to unpause the thread.
+     * </li><li>
+     * While the rendering thread is paused, rendering commands will continue to be queued until the
+     * buffer limit is reached. When the limit is reached, the program will abort.
+     * </li></ul>
+     */
+    public void setPaused(boolean paused) {
+        nSetPaused(getNativeObject(), paused);
+    }
 
     @UsedByReflection("TextureHelper.java")
     public long getNativeObject() {
@@ -1253,6 +1351,8 @@ public class Engine {
     private static native boolean nIsValidSkinningBuffer(long nativeEngine, long nativeSkinningBuffer);
     private static native boolean nIsValidIndirectLight(long nativeEngine, long nativeIndirectLight);
     private static native boolean nIsValidMaterial(long nativeEngine, long nativeMaterial);
+    private static native boolean nIsValidMaterialInstance(long nativeEngine, long nativeMaterial, long nativeMaterialInstance);
+    private static native boolean nIsValidExpensiveMaterialInstance(long nativeEngine, long nativeMaterialInstance);
     private static native boolean nIsValidSkybox(long nativeEngine, long nativeSkybox);
     private static native boolean nIsValidColorGrading(long nativeEngine, long nativeColorGrading);
     private static native boolean nIsValidTexture(long nativeEngine, long nativeTexture);
@@ -1261,6 +1361,8 @@ public class Engine {
     private static native void nDestroyEntity(long nativeEngine, int entity);
     private static native void nFlushAndWait(long nativeEngine);
     private static native void nFlush(long nativeEngine);
+    private static native boolean nIsPaused(long nativeEngine);
+    private static native void nSetPaused(long nativeEngine, boolean paused);
     private static native long nGetTransformManager(long nativeEngine);
     private static native long nGetLightManager(long nativeEngine);
     private static native long nGetRenderableManager(long nativeEngine);
@@ -1281,8 +1383,12 @@ public class Engine {
             long minCommandBufferSizeMB, long perFrameCommandsSizeMB, long jobSystemThreadCount,
             long textureUseAfterFreePoolSize, boolean disableParallelShaderCompile,
             int stereoscopicType, long stereoscopicEyeCount,
-            long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge);
+            long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge,
+            boolean disableHandleUseAfterFreeCheck,
+            int preferredShaderLanguage,
+            boolean forceGLES2Context);
     private static native void nSetBuilderFeatureLevel(long nativeBuilder, int ordinal);
     private static native void nSetBuilderSharedContext(long nativeBuilder, long sharedContext);
+    private static native void nSetBuilderPaused(long nativeBuilder, boolean paused);
     private static native long nBuilderBuild(long nativeBuilder);
 }

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

@@ -175,6 +175,32 @@ public class RenderableManager {
             return this;
         }
 
+        /**
+         * Type of geometry for a Renderable
+         */
+        public enum GeometryType {
+            /** dynamic gemoetry has no restriction */
+            DYNAMIC,
+            /** bounds and world space transform are immutable */
+            STATIC_BOUNDS,
+            /** skinning/morphing not allowed and Vertex/IndexBuffer immutables */
+            STATIC
+        }
+
+        /**
+         * Specify whether this renderable has static bounds. In this context his means that
+         * the renderable's bounding box cannot change and that the renderable's transform is
+         * assumed immutable. Changing the renderable's transform via the TransformManager
+         * can lead to corrupted graphics. Note that skinning and morphing are not forbidden.
+         * Disabled by default.
+         * @param enable whether this renderable has static bounds. false by default.
+         */
+        @NonNull
+        public Builder geometryType(GeometryType type) {
+            nBuilderGeometryType(mNativeBuilder, type.ordinal());
+            return this;
+        }
+
         /**
          * Binds a material instance to the specified primitive.
          *
@@ -964,6 +990,7 @@ public class RenderableManager {
     private static native void nBuilderGeometry(long nativeBuilder, int index, int value, long nativeVertexBuffer, long nativeIndexBuffer);
     private static native void nBuilderGeometry(long nativeBuilder, int index, int value, long nativeVertexBuffer, long nativeIndexBuffer, int offset, int count);
     private static native void nBuilderGeometry(long nativeBuilder, int index, int value, long nativeVertexBuffer, long nativeIndexBuffer, int offset, int minIndex, int maxIndex, int count);
+    private static native void nBuilderGeometryType(long nativeBuilder, int type);
     private static native void nBuilderMaterial(long nativeBuilder, int index, long nativeMaterialInstance);
     private static native void nBuilderBlendOrder(long nativeBuilder, int index, int blendOrder);
     private static native void nBuilderGlobalBlendOrderEnabled(long nativeBuilder, int index, boolean enabled);

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

@@ -68,72 +68,6 @@ public class SwapChain {
     private final Object mSurface;
     private long mNativeObject;
 
-    public static final long CONFIG_DEFAULT = 0x0;
-
-    /**
-     * This flag indicates that the <code>SwapChain</code> must be allocated with an
-     * alpha-channel.
-     */
-    public static final long CONFIG_TRANSPARENT = 0x1;
-
-    /**
-     * This flag indicates that the <code>SwapChain</code> may be used as a source surface
-     * for reading back render results.  This config must be set when creating
-     * any <code>SwapChain</code>  that will be used as the source for a blit operation.
-     *
-     * @see Renderer#copyFrame
-     */
-    public static final long CONFIG_READABLE = 0x2;
-
-    /**
-     * Indicates that the native X11 window is an XCB window rather than an XLIB window.
-     * This is ignored on non-Linux platforms and in builds that support only one X11 API.
-     */
-    public static final long CONFIG_ENABLE_XCB = 0x4;
-
-    /**
-     * Indicates that the SwapChain must automatically perform linear to sRGB encoding.
-     *
-     * This flag is ignored if isSRGBSwapChainSupported() is false.
-     *
-     * When using this flag, post-processing should be disabled.
-     *
-     * @see SwapChain#isSRGBSwapChainSupported
-     * @see View#setPostProcessingEnabled
-     */
-    public static final long CONFIG_SRGB_COLORSPACE = 0x10;
-
-    /**
-     * Indicates that this SwapChain should allocate a stencil buffer in addition to a depth buffer.
-     *
-     * This flag is necessary when using View::setStencilBufferEnabled and rendering directly into
-     * the SwapChain (when post-processing is disabled).
-     *
-     * The specific format of the stencil buffer depends on platform support. The following pixel
-     * formats are tried, in order of preference:
-     *
-     * Depth only (without CONFIG_HAS_STENCIL_BUFFER):
-     * - DEPTH32F
-     * - DEPTH24
-     *
-     * Depth + stencil (with CONFIG_HAS_STENCIL_BUFFER):
-     * - DEPTH32F_STENCIL8
-     * - DEPTH24F_STENCIL8
-     *
-     * Note that enabling the stencil buffer may hinder depth precision and should only be used if
-     * necessary.
-     *
-     * @see View#setStencilBufferEnabled
-     * @see View#setPostProcessingEnabled
-     */
-    public static final long CONFIG_HAS_STENCIL_BUFFER = 0x20;
-
-    /**
-     * The SwapChain contains protected content. Only supported when isProtectedContentSupported()
-     * is true.
-     */
-    public static final long CONFIG_PROTECTED_CONTENT   = 0x40;
-
     SwapChain(long nativeSwapChain, Object surface) {
         mNativeObject = nativeSwapChain;
         mSurface = surface;
@@ -145,17 +79,19 @@ public class SwapChain {
      *
      * @param engine A reference to the filament Engine
      * @return true if CONFIG_PROTECTED_CONTENT is supported, false otherwise.
+     * @see SwapChainFlags#CONFIG_PROTECTED_CONTENT
      */
     public static boolean isProtectedContentSupported(@NonNull Engine engine) {
         return nIsProtectedContentSupported(engine.getNativeObject());
     }
 
     /**
-     * Return whether createSwapChain supports the SWAP_CHAIN_CONFIG_SRGB_COLORSPACE flag.
+     * Return whether createSwapChain supports the CONFIG_SRGB_COLORSPACE flag.
      * The default implementation returns false.
      *
      * @param engine A reference to the filament Engine
-     * @return true if SWAP_CHAIN_CONFIG_SRGB_COLORSPACE is supported, false otherwise.
+     * @return true if CONFIG_SRGB_COLORSPACE is supported, false otherwise.
+     * @see SwapChainFlags#CONFIG_SRGB_COLORSPACE
      */
     public static boolean isSRGBSwapChainSupported(@NonNull Engine engine) {
         return nIsSRGBSwapChainSupported(engine.getNativeObject());

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

@@ -0,0 +1,97 @@
+/*
+ * 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.
+ */
+
+package com.google.android.filament;
+
+// Note: SwapChainFlags is kept separate from SwapChain so that UiHelper does not need to depend
+// on SwapChain. This allows clients to use UiHelper without requiring all of Filament's Java
+// classes.
+
+/**
+ * Flags that a <code>SwapChain</code> can be created with to control behavior.
+ *
+ * @see Engine#createSwapChain
+ * @see Engine#createSwapChainFromNativeSurface
+ */
+public final class SwapChainFlags {
+
+    public static final long CONFIG_DEFAULT = 0x0;
+
+    /**
+     * This flag indicates that the <code>SwapChain</code> must be allocated with an
+     * alpha-channel.
+     */
+    public static final long CONFIG_TRANSPARENT = 0x1;
+
+    /**
+     * This flag indicates that the <code>SwapChain</code> may be used as a source surface
+     * for reading back render results.  This config must be set when creating
+     * any <code>SwapChain</code>  that will be used as the source for a blit operation.
+     *
+     * @see Renderer#copyFrame
+     */
+    public static final long CONFIG_READABLE = 0x2;
+
+    /**
+     * Indicates that the native X11 window is an XCB window rather than an XLIB window.
+     * This is ignored on non-Linux platforms and in builds that support only one X11 API.
+     */
+    public static final long CONFIG_ENABLE_XCB = 0x4;
+
+    /**
+     * Indicates that the SwapChain must automatically perform linear to sRGB encoding.
+     *
+     * This flag is ignored if isSRGBSwapChainSupported() is false.
+     *
+     * When using this flag, post-processing should be disabled.
+     *
+     * @see SwapChain#isSRGBSwapChainSupported
+     * @see View#setPostProcessingEnabled
+     */
+    public static final long CONFIG_SRGB_COLORSPACE = 0x10;
+
+    /**
+     * Indicates that this SwapChain should allocate a stencil buffer in addition to a depth buffer.
+     *
+     * This flag is necessary when using View::setStencilBufferEnabled and rendering directly into
+     * the SwapChain (when post-processing is disabled).
+     *
+     * The specific format of the stencil buffer depends on platform support. The following pixel
+     * formats are tried, in order of preference:
+     *
+     * Depth only (without CONFIG_HAS_STENCIL_BUFFER):
+     * - DEPTH32F
+     * - DEPTH24
+     *
+     * Depth + stencil (with CONFIG_HAS_STENCIL_BUFFER):
+     * - DEPTH32F_STENCIL8
+     * - DEPTH24F_STENCIL8
+     *
+     * Note that enabling the stencil buffer may hinder depth precision and should only be used if
+     * necessary.
+     *
+     * @see View#setStencilBufferEnabled
+     * @see View#setPostProcessingEnabled
+     */
+    public static final long CONFIG_HAS_STENCIL_BUFFER = 0x20;
+
+    /**
+     * The SwapChain contains protected content. Only supported when isProtectedContentSupported()
+     * is true.
+     */
+    public static final long CONFIG_PROTECTED_CONTENT   = 0x40;
+}
+

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

@@ -241,6 +241,15 @@ public class View {
         nSetCamera(getNativeObject(), camera == null ? 0 : camera.getNativeObject());
     }
 
+    /**
+     * Query whether a camera is set.
+     * @return true if a camera is set, false otherwise
+     * @see #setCamera
+     */
+    public boolean hasCamera() {
+        return nHasCamera(getNativeObject());
+    }
+
     /**
      * Gets this View's associated Camera, or null if none has been assigned.
      *
@@ -1238,6 +1247,7 @@ public class View {
     private static native void nSetName(long nativeView, String name);
     private static native void nSetScene(long nativeView, long nativeScene);
     private static native void nSetCamera(long nativeView, long nativeCamera);
+    private static native boolean nHasCamera(long nativeView);
     private static native void nSetViewport(long nativeView, int left, int bottom, int width, int height);
     private static native void nSetVisibleLayers(long nativeView, int select, int value);
     private static native void nSetShadowingEnabled(long nativeView, boolean enabled);

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

@@ -27,7 +27,7 @@ import android.view.SurfaceHolder;
 import android.view.SurfaceView;
 import android.view.TextureView;
 
-import com.google.android.filament.SwapChain;
+import com.google.android.filament.SwapChainFlags;
 
 /**
  * UiHelper is a simple class that can manage either a SurfaceView, TextureView, or a SurfaceHolder
@@ -538,7 +538,7 @@ public class UiHelper {
      * the options set on this UiHelper.
      */
     public long getSwapChainFlags() {
-        return isOpaque() ? SwapChain.CONFIG_DEFAULT : SwapChain.CONFIG_TRANSPARENT;
+        return isOpaque() ? SwapChainFlags.CONFIG_DEFAULT : SwapChainFlags.CONFIG_TRANSPARENT;
     }
 
     /**

android/filament-utils-android/CMakeLists.txt

@@ -31,6 +31,7 @@ set_target_properties(iblprefilter PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libfilament-iblprefilter.a)
 
 set(CMAKE_SHARED_LINKER_FLAGS_RELEASE "${CMAKE_SHARED_LINKER_FLAGS_RELEASE} -Wl,--version-script=${CMAKE_CURRENT_SOURCE_DIR}/libfilament-utils-jni.map")
+set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size=16384")
 
 add_library(filament-utils-jni SHARED
         src/main/cpp/AutomationEngine.cpp

android/gltfio-android/CMakeLists.txt

@@ -44,6 +44,7 @@ set_target_properties(uberarchive PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libuberarchive.a)
 
 set(CMAKE_SHARED_LINKER_FLAGS_RELEASE "${CMAKE_SHARED_LINKER_FLAGS_RELEASE} -Wl,--version-script=${CMAKE_CURRENT_SOURCE_DIR}/libgltfio-jni.map")
+set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size=16384")
 
 set(GLTFIO_SRCS
         ${GLTFIO_DIR}/include/gltfio/Animator.h
@@ -81,9 +82,20 @@ set(GLTFIO_SRCS
         ${GLTFIO_DIR}/src/TangentsJob.cpp
         ${GLTFIO_DIR}/src/TangentsJob.h
         ${GLTFIO_DIR}/src/UbershaderProvider.cpp
+        ${GLTFIO_DIR}/src/Utility.cpp
+        ${GLTFIO_DIR}/src/Utility.h
         ${GLTFIO_DIR}/src/Wireframe.cpp
         ${GLTFIO_DIR}/src/Wireframe.h
         ${GLTFIO_DIR}/src/downcast.h
+        ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.cpp
+        ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.h
+        ${GLTFIO_DIR}/src/extended/ResourceLoaderExtended.cpp
+        ${GLTFIO_DIR}/src/extended/ResourceLoaderExtended.h
+        ${GLTFIO_DIR}/src/extended/TangentsJobExtended.cpp
+        ${GLTFIO_DIR}/src/extended/TangentsJobExtended.h
+        ${GLTFIO_DIR}/src/extended/TangentSpaceMeshWrapper.cpp
+        ${GLTFIO_DIR}/src/extended/TangentSpaceMeshWrapper.h
+
 
         src/main/cpp/Animator.cpp
         src/main/cpp/AssetLoader.cpp

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.50.6
+VERSION_NAME=1.51.8
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

android/samples/sample-hello-triangle/src/main/java/com/google/android/filament/hellotriangle/MainActivity.kt

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

build.sh

@@ -44,8 +44,6 @@ function print_help {
     echo "        Exclude Vulkan support from the Android build."
     echo "    -s"
     echo "        Add iOS simulator support to the iOS build."
-    echo "    -t"
-    echo "        Enable SwiftShader support for Vulkan in desktop builds."
     echo "    -e"
     echo "        Enable EGL on Linux support for desktop builds."
     echo "    -l"
@@ -61,6 +59,11 @@ function print_help {
     echo "    -b"
     echo "        Enable Address and Undefined Behavior Sanitizers (asan/ubsan) for debugging."
     echo "        This is only for the desktop build."
+    echo "    -x value"
+    echo "        Define a preprocessor flag FILAMENT_BACKEND_DEBUG_FLAG with [value]. This is useful for"
+    echo "        enabling debug paths in the backend from the build script. For example, make a"
+    echo "        systrace-enabled build without directly changing #defines. Remember to add -f when"
+    echo "        changing this option."
     echo ""
     echo "Build types:"
     echo "    release"
@@ -160,8 +163,6 @@ INSTALL_COMMAND=
 VULKAN_ANDROID_OPTION="-DFILAMENT_SUPPORTS_VULKAN=ON"
 VULKAN_ANDROID_GRADLE_OPTION=""
 
-SWIFTSHADER_OPTION="-DFILAMENT_USE_SWIFTSHADER=OFF"
-
 EGL_ON_LINUX_OPTION="-DFILAMENT_SUPPORTS_EGL_ON_LINUX=OFF"
 
 MATDBG_OPTION="-DFILAMENT_ENABLE_MATDBG=OFF"
@@ -172,6 +173,8 @@ MATOPT_GRADLE_OPTION=""
 
 ASAN_UBSAN_OPTION=""
 
+BACKEND_DEBUG_FLAG_OPTION=""
+
 IOS_BUILD_SIMULATOR=false
 BUILD_UNIVERSAL_LIBRARIES=false
 
@@ -226,11 +229,11 @@ function build_desktop_target {
             -DIMPORT_EXECUTABLES_DIR=out \
             -DCMAKE_BUILD_TYPE="$1" \
             -DCMAKE_INSTALL_PREFIX="../${lc_target}/filament" \
-            ${SWIFTSHADER_OPTION} \
             ${EGL_ON_LINUX_OPTION} \
             ${MATDBG_OPTION} \
             ${MATOPT_OPTION} \
             ${ASAN_UBSAN_OPTION} \
+            ${BACKEND_DEBUG_FLAG_OPTION} \
             ${architectures} \
             ../..
         ln -sf "out/cmake-${lc_target}/compile_commands.json" \
@@ -289,6 +292,7 @@ function build_webgl_with_target {
             -DCMAKE_BUILD_TYPE="$1" \
             -DCMAKE_INSTALL_PREFIX="../webgl-${lc_target}/filament" \
             -DWEBGL=1 \
+            ${BACKEND_DEBUG_FLAG_OPTION} \
             ../..
         ln -sf "out/cmake-webgl-${lc_target}/compile_commands.json" \
            ../../compile_commands.json
@@ -363,6 +367,7 @@ function build_android_target {
             ${MATDBG_OPTION} \
             ${MATOPT_OPTION} \
             ${VULKAN_ANDROID_OPTION} \
+            ${BACKEND_DEBUG_FLAG_OPTION} \
             ../..
         ln -sf "out/cmake-android-${lc_target}-${arch}/compile_commands.json" \
            ../../compile_commands.json
@@ -516,7 +521,7 @@ function build_android {
             if [[ "${BUILD_ANDROID_SAMPLES}" == "true" ]]; then
                 for sample in ${ANDROID_SAMPLES}; do
                     echo "Installing out/${sample}-debug.apk"
-                    cp samples/${sample}/build/outputs/apk/debug/${sample}-debug-unsigned.apk \
+                    cp samples/${sample}/build/outputs/apk/debug/${sample}-debug.apk \
                         ../out/${sample}-debug.apk
                 done
             fi
@@ -597,6 +602,7 @@ function build_ios_target {
             -DCMAKE_TOOLCHAIN_FILE=../../third_party/clang/iOS.cmake \
             ${MATDBG_OPTION} \
             ${MATOPT_OPTION} \
+            ${BACKEND_DEBUG_FLAG_OPTION} \
             ../..
         ln -sf "out/cmake-ios-${lc_target}-${arch}/compile_commands.json" \
            ../../compile_commands.json
@@ -730,6 +736,13 @@ function validate_build_command {
             exit 1
         fi
     fi
+
+    # Make sure FILAMENT_BACKEND_DEBUG_FLAG is only meant for debug builds
+    if [[ "${ISSUE_DEBUG_BUILD}" != "true" ]] && [[ ! -z "${BACKEND_DEBUG_FLAG_OPTION}" ]]; then
+        echo "Error: cannot specify FILAMENT_BACKEND_DEBUG_FLAG in non-debug build"
+        exit 1
+    fi
+
     set -e
 }
 
@@ -776,7 +789,7 @@ function check_debug_release_build {
 
 pushd "$(dirname "$0")" > /dev/null
 
-while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
+while getopts ":hacCfgijmp:q:uvslwedk:bx:" opt; do
     case ${opt} in
         h)
             print_help
@@ -840,7 +853,7 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
                         echo "Platform must be one of [desktop|android|ios|webgl|all]"
                         echo ""
                         exit 1
-                    ;;    
+                    ;;
                 esac
             done
             ;;
@@ -895,10 +908,6 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
             IOS_BUILD_SIMULATOR=true
             echo "iOS simulator support enabled."
             ;;
-        t)
-            SWIFTSHADER_OPTION="-DFILAMENT_USE_SWIFTSHADER=ON"
-            echo "SwiftShader support enabled."
-            ;;
         e)
             EGL_ON_LINUX_OPTION="-DFILAMENT_SUPPORTS_EGL_ON_LINUX=ON -DFILAMENT_SKIP_SDL2=ON -DFILAMENT_SKIP_SAMPLES=ON"
             echo "EGL on Linux support enabled; skipping SDL2."
@@ -918,6 +927,8 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
         b)  ASAN_UBSAN_OPTION="-DFILAMENT_ENABLE_ASAN_UBSAN=ON"
             echo "Enabled ASAN/UBSAN"
             ;;
+        x)  BACKEND_DEBUG_FLAG_OPTION="-DFILAMENT_BACKEND_DEBUG_FLAG=${OPTARG}"
+            ;;
         \?)
             echo "Invalid option: -${OPTARG}" >&2
             echo ""

build/swiftshader/Dockerfile

@@ -1,53 +0,0 @@
-# Build the image:
-#   docker build --no-cache --tag ssfilament -f build/swiftshader/Dockerfile .
-#   docker tag ssfilament ghcr.io/filament-assets/swiftshader
-#
-# Publish the image:
-#   docker login ghcr.io --username <user> --password <token>
-#   docker push ghcr.io/filament-assets/swiftshader
-#
-# Run the image and mount the current directory:
-#   docker run -it -v `pwd`:/trees/filament -t ssfilament
-
-FROM ubuntu:focal
-WORKDIR /trees
-ARG DEBIAN_FRONTEND=noninteractive
-ENV SWIFTSHADER_LD_LIBRARY_PATH=/trees/swiftshader/build
-ENV CXXFLAGS='-fno-builtin -Wno-pass-failed'
-
-RUN apt-get update && \
-    apt-get --no-install-recommends install -y \
-	apt-transport-https \
-	apt-utils \
-	build-essential \
-	cmake \
-	ca-certificates \
-	git \
-	ninja-build \
-	python \
-	python3 \
-	xorg-dev \
-	clang-7 \
-	libc++-7-dev \
-	libc++abi-7-dev \
-	lldb
-
-# Ensure that clang is used instead of gcc.
-RUN set -eux ;\
-    update-alternatives --install /usr/bin/clang clang /usr/bin/clang-7 100 ;\
-    update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-7 100 ;\
-    update-alternatives --install /usr/bin/cc cc /usr/bin/clang 100 ;\
-    update-alternatives --install /usr/bin/c++ c++ /usr/bin/clang++ 100
-
-# Get patch files from the local Filament tree.
-COPY build/swiftshader/*.diff .
-
-# Clone SwiftShader, apply patches, and build it.
-RUN set -eux ;\
-	git clone https://swiftshader.googlesource.com/SwiftShader swiftshader ;\
-    cd swiftshader ;\
-	git checkout 139f5c3 ;\
-	git apply /trees/*.diff ;\
-    cd build ;\
-    cmake .. -GNinja -DCMAKE_BUILD_TYPE=Release ;\
-    ninja

build/swiftshader/gallery.py

@@ -1,56 +0,0 @@
-#!/usr/bin/env python3
-
-from pathlib import Path
-import os
-
-spath = os.path.dirname(os.path.realpath(__file__))
-
-path = Path(spath)
-
-folder = "../../results/"
-
-images = list(path.glob(folder + '*.png'))
-
-images.sort()
-
-gallery = open(path.absolute().joinpath(folder + 'index.html'), 'w')
-
-gallery.write("""<html>
-<head>
-<script type="module" src="https://unpkg.com/img-comparison-slider@latest/dist/component/component.esm.js"></script>
-<script nomodule="" src="https://unpkg.com/img-comparison-slider@latest/dist/component/component.js"></script>
-<link rel="stylesheet" href="https://unpkg.com/img-comparison-slider@latest/dist/collection/styles/initial.css"/>
-<style>
-    h2 {
-        font-weight: normal;
-        margin-top: 150px;
-        margin-bottom: 20px;
-    }
-    a {
-        text-decoration: none;
-        font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
-        color: blue;
-    }
-    a:hover {
-        font-weight: bold;
-    }
-</style>
-</head>
-<body>
-""")
-
-tag = ''
-
-for image in images:
-    group = image.stem.rstrip('0123456789')
-    before = f'https://filament-assets.github.io/golden/{group}/{image.name}'
-    after = image.name
-    gallery.write('\n')
-    gallery.write(f'<h2><a href="{image.stem}.json">{image.stem}.json</a></h2>\n')
-    gallery.write('<img-comparison-slider>\n')
-    gallery.write(f'<img slot="before" src="{before}" /> <img slot="after" src="{after}" />\n')
-    gallery.write('</img-comparison-slider>\n')
-
-gallery.write("""</body>
-</html>
-""")

build/swiftshader/patch_00.diff

@@ -1,62 +0,0 @@
-diff --git a/src/Vulkan/VkPipeline.cpp b/src/Vulkan/VkPipeline.cpp
-index 86913ec72..3b35345af 100644
---- a/src/Vulkan/VkPipeline.cpp
-+++ b/src/Vulkan/VkPipeline.cpp
-@@ -71,7 +71,56 @@ std::vector<uint32_t> preprocessSpirv(
- 	if(optimize)
- 	{
- 		// Full optimization list taken from spirv-opt.
--		opt.RegisterPerformancePasses();
-+
-+		// We have removed CreateRedundancyEliminationPass because it segfaults when encountering:
-+		//     %389 = OpCompositeConstruct %7 %386 %387 %388 %86
-+		// When inserting an entry into instruction_to_value_ (which is an unordered_map)
-+		// This could perhaps be investigated further with help from asan.
-+
-+		using namespace spvtools;
-+		opt.RegisterPass(CreateWrapOpKillPass())
-+			.RegisterPass(CreateDeadBranchElimPass())
-+			.RegisterPass(CreateMergeReturnPass())
-+			.RegisterPass(CreateInlineExhaustivePass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreatePrivateToLocalPass())
-+			.RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
-+			.RegisterPass(CreateLocalSingleStoreElimPass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateScalarReplacementPass())
-+			.RegisterPass(CreateLocalAccessChainConvertPass())
-+			.RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
-+			.RegisterPass(CreateLocalSingleStoreElimPass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateLocalMultiStoreElimPass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateCCPPass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateLoopUnrollPass(true))
-+			.RegisterPass(CreateDeadBranchElimPass())
-+			.RegisterPass(CreateRedundancyEliminationPass()) // workaround for SEGFAULT
-+			.RegisterPass(CreateCombineAccessChainsPass())
-+			.RegisterPass(CreateSimplificationPass())
-+			.RegisterPass(CreateScalarReplacementPass())
-+			.RegisterPass(CreateLocalAccessChainConvertPass())
-+			.RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
-+			.RegisterPass(CreateLocalSingleStoreElimPass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateSSARewritePass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateVectorDCEPass())
-+			.RegisterPass(CreateDeadInsertElimPass())
-+			.RegisterPass(CreateDeadBranchElimPass())
-+			.RegisterPass(CreateSimplificationPass())
-+			.RegisterPass(CreateIfConversionPass())
-+			.RegisterPass(CreateCopyPropagateArraysPass())
-+			.RegisterPass(CreateReduceLoadSizePass())
-+			.RegisterPass(CreateAggressiveDCEPass())
-+			.RegisterPass(CreateBlockMergePass())
-+			.RegisterPass(CreateRedundancyEliminationPass()) // workaround for SEGFAULT
-+			.RegisterPass(CreateDeadBranchElimPass())
-+			.RegisterPass(CreateBlockMergePass())
-+			.RegisterPass(CreateSimplificationPass());
- 	}
- 
- 	std::vector<uint32_t> optimized;

build/swiftshader/test.sh

@@ -1,127 +0,0 @@
-#!/bin/bash
-set -e
-
-function print_help {
-    local self_name=$(basename "$0")
-    echo "This script issues docker commands for testing Filament with SwiftShader."
-    echo "The usual sequence of commands is: fetch, start, build filament release, and run."
-    echo ""
-    echo "Usage:"
-    echo "    $self_name [command]"
-    echo ""
-    echo "Commands:"
-    echo "    build filament [debug | release]"
-    echo "        Use the container to build Filament."
-    echo "    build swiftshader [debug | release]"
-    echo "        Use the container to do a clean rebuild of SwiftShader."
-    echo "        (Note that the container already has SwiftShader built.)"
-    echo "    fetch"
-    echo "        Download the docker image from the central repository."
-    echo "    help"
-    echo "        Print this help message."
-    echo "    logs"
-    echo "        Print messages from the container's kernel ring buffer."
-    echo "        This is useful for diagnosing OOM issues."
-    echo "    run [lldb]"
-    echo "        Launch a test inside the container, optionally via lldb."
-    echo "    shell"
-    echo "        Interact with a bash prompt in the container."
-    echo "    start"
-    echo "        Start a container from the image."
-    echo "    stop"
-    echo "        Stop the container."
-    echo ""
-}
-
-# Change the current working directory to the Filament root.
-pushd "$(dirname "$0")/../.." > /dev/null
-
-if [[ "$1" == "build" ]] && [[ "$2" == "filament" ]]; then
-    docker exec runner filament/build.sh -t $3 gltf_viewer
-    exit $?
-fi
-
-if [[ "$1" == "build" ]] && [[ "$2" == "swiftshader" ]]; then
-    BUILD_TYPE="$3"
-    BUILD_TYPE="$(tr '[:lower:]' '[:upper:]' <<< ${BUILD_TYPE:0:1})${BUILD_TYPE:1}"
-    docker exec --workdir /trees/swiftshader runner rm -rf build
-    docker exec --workdir /trees/swiftshader runner mkdir build
-    docker exec --workdir /trees/swiftshader/build runner cmake -GNinja -DCMAKE_BUILD_TYPE="$BUILD_TYPE" ..
-    docker exec --workdir /trees/swiftshader/build runner ninja
-    exit $?
-fi
-
-if [[ "$1" == "fetch" ]]; then
-    docker pull ghcr.io/filament-assets/swiftshader:latest
-    docker tag ghcr.io/filament-assets/swiftshader:latest ssfilament
-    exit $?
-fi
-
-if [[ "$1" == "help" ]]; then
-    print_help
-    exit 0
-fi
-
-if [[ "$1" == "logs" ]]; then
-    docker exec runner dmesg --human --read-clear
-    exit $?
-fi
-
-if [[ "$1" == "run" ]] && [[ "$2" == "lldb" ]]; then
-    docker exec -i --workdir /trees/filament/results runner \
-          lldb --batch -o run -o bt -- \
-          ../out/cmake-release/samples/gltf_viewer \
-          --headless \
-          --batch ../libs/viewer/tests/basic.json \
-          --api vulkan
-    docker exec runner /trees/filament/build/swiftshader/gallery.py
-    exit $?
-fi
-
-if [[ "$1" == "run" ]]; then
-    docker exec --tty  --workdir /trees/filament/results runner \
-          /usr/bin/catchsegv \
-          ../out/cmake-release/samples/gltf_viewer \
-          --headless \
-          --batch ../libs/viewer/tests/basic.json \
-          --api vulkan
-    docker exec runner /trees/filament/build/swiftshader/gallery.py
-    exit $?
-fi
-
-if [[ "$1" == "shell" ]]; then
-    docker exec --interactive --tty runner /bin/bash
-    exit $?
-fi
-
-# Notes on options being passed to docker's run command:
-#
-# - The memory constraint seems to prevent an OOM signal in GitHub Actions.
-# - The cap / security args allow use of lldb and creation of core dumps.
-# - The privileged arg allows use of dmesg for examining OOM logs.
-#
-# Currently, a GitHub Actions VM has 2 CPUs, 7 GB RAM, and 14 GB of SSD disk space.
-#
-# Please be aware that Docker Desktop might impose additional resource constraints, and that those
-# settings can only be controlled with its GUI. We recommend at least 7 GB of memory and 2 GB swap.
-if [[ "$1" == "start" ]]; then
-    mkdir -p results
-    docker run --tty --rm --detach --privileged \
-          --memory 6.5g \
-          --name runner \
-          --cap-add=SYS_PTRACE \
-          --security-opt seccomp=unconfined \
-          --security-opt apparmor=unconfined \
-          --volume `pwd`:/trees/filament \
-          --workdir /trees \
-          ssfilament
-    exit $?
-fi
-
-if [[ "$1" == "stop" ]]; then
-    docker container rm runner --force
-    exit $?
-fi
-
-print_help
-exit 1

docs/Filament.html

@@ -476,7 +476,7 @@ Energy conservation is one of the key components of a good BRDF for physically b
 <p>
 
 
-For the specular term, \(f_m\) is a mirror BRDF that can be modeled with the Fresnel law, noted \(F\) in the Cook-Torrance approximation of the microfacet model integration:
+For the specular term, \(f_r\) is a mirror BRDF that can be modeled with the Fresnel law, noted \(F\) in the Cook-Torrance approximation of the microfacet model integration:
 
 </p><p>
 
@@ -857,7 +857,7 @@ The full GLSL implementation of the standard model is shown in <a href="#listing
 <span class="line">    <span class="hljs-comment">// perceptually linear roughness to roughness (see parameterization)</span></span>
 <span class="line">    <span class="hljs-type">float</span> roughness = perceptualRoughness * perceptualRoughness;</span>
 <span class="line"></span>
-<span class="line">    <span class="hljs-type">float</span> D = D_GGX(NoH, a);</span>
+<span class="line">    <span class="hljs-type">float</span> D = D_GGX(NoH, roughness);</span>
 <span class="line">    <span class="hljs-type">vec3</span>  F = F_Schlick(LoH, f0);</span>
 <span class="line">    <span class="hljs-type">float</span> V = V_SmithGGXCorrelated(NoV, NoL, roughness);</span>
 <span class="line"></span>

docs/Materials.md.html

@@ -1397,7 +1397,7 @@ Type
 :    `string`
 
 Value
-:     Any of `opaque`, `transparent`, `fade`, `add`, `masked`, `multiply`, `screen`. Defaults to `opaque`.
+:     Any of `opaque`, `transparent`, `fade`, `add`, `masked`, `multiply`, `screen`, `custom`. Defaults to `opaque`.
 
 Description
 :     Defines how/if the rendered object is blended with the content of the render target.
@@ -1420,6 +1420,7 @@ 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.
@@ -1432,6 +1433,36 @@ 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
@@ -2394,6 +2425,7 @@ The following APIs are only available from the vertex block:
 **getWorldFromModelMatrix()**        | float4x4 |  Matrix that converts from model (object) space to world space
 **getWorldFromModelNormalMatrix()**  | float3x3 |  Matrix that converts normals from model (object) space to world space
 **getVertexIndex()**                 | int      |  Index of the current vertex
+**getEyeIndex()**                    | int      |  Index of the eye being rendered, starting at 0
 
 ### Fragment only
 

docs/viewer/index.html

@@ -43,7 +43,7 @@ filament-viewer::part(canvas) {
 </p>
 
 </main>
-<script src="https://unpkg.com/filament@1.44.0/filament.js"></script>
+<script src="https://unpkg.com/filament@1.51.6/filament.js"></script>
 <script src="https://unpkg.com/gltumble"></script>
 <script src="filament-viewer.js" type="module"></script>
 </body>

filament/CMakeLists.txt

@@ -163,6 +163,7 @@ set(PRIVATE_HDRS
         src/ResourceList.h
         src/ShadowMap.h
         src/ShadowMapManager.h
+        src/SharedHandle.h
         src/TypedUniformBuffer.h
         src/UniformBuffer.h
         src/components/CameraManager.h
@@ -214,6 +215,7 @@ set(PRIVATE_HDRS
 set(MATERIAL_SRCS
         src/materials/antiAliasing/fxaa.mat
         src/materials/antiAliasing/taa.mat
+        src/materials/blitDepth.mat
         src/materials/blitLow.mat
         src/materials/blitArray.mat
         src/materials/bloom/bloomDownsample.mat
@@ -256,6 +258,11 @@ set(MATERIAL_FL0_SRCS
         src/materials/skybox.mat
 )
 
+set(MATERIAL_MULTIVIEW_SRCS
+        src/materials/defaultMaterial.mat
+        src/materials/skybox.mat
+)
+
 # Embed the binary resource blob for materials.
 get_resgen_vars(${RESOURCE_DIR} materials)
 list(APPEND PRIVATE_HDRS ${RESGEN_HEADER})
@@ -285,6 +292,11 @@ if (FILAMENT_ENABLE_FEATURE_LEVEL_0)
     add_definitions(-DFILAMENT_ENABLE_FEATURE_LEVEL_0)
 endif()
 
+# Whether to include MULTIVIEW materials.
+if (FILAMENT_ENABLE_MULTIVIEW)
+    add_definitions(-DFILAMENT_ENABLE_MULTIVIEW)
+endif()
+
 # ==================================================================================================
 # Definitions
 # ==================================================================================================
@@ -315,30 +327,41 @@ foreach (mat_src ${MATERIAL_SRCS})
     get_filename_component(localname "${mat_src}" NAME_WE)
     get_filename_component(fullname "${mat_src}" ABSOLUTE)
     set(output_path "${MATERIAL_DIR}/${localname}.filamat")
+    add_custom_command(
+            OUTPUT ${output_path}
+            COMMAND matc ${MATC_BASE_FLAGS} -o ${output_path} ${fullname}
+            MAIN_DEPENDENCY ${fullname}
+            DEPENDS matc
+            COMMENT "Compiling material ${fullname} to ${output_path}"
+    )
+    list(APPEND MATERIAL_BINS ${output_path})
 
     list(FIND MATERIAL_FL0_SRCS ${mat_src} index)
     if (${index} GREATER -1 AND FILAMENT_ENABLE_FEATURE_LEVEL_0)
-        string(REGEX REPLACE "[.]filamat$" "0.filamat" output_path0 ${output_path})
+        string(REGEX REPLACE "[.]filamat$" "_fl0.filamat" output_path_fl0 ${output_path})
         add_custom_command(
-                OUTPUT ${output_path} ${output_path0}
-                COMMAND matc ${MATC_BASE_FLAGS} -o ${output_path} ${fullname}
-                COMMAND matc ${MATC_BASE_FLAGS} -PfeatureLevel=0 -o ${output_path0} ${fullname}
+                OUTPUT ${output_path_fl0}
+                COMMAND matc ${MATC_BASE_FLAGS} -PfeatureLevel=0 -o ${output_path_fl0} ${fullname}
                 MAIN_DEPENDENCY ${fullname}
                 DEPENDS matc
-                COMMENT "Compiling material ${fullname} to ${output_path} and ${output_path0}"
-        )
-        list(APPEND MATERIAL_BINS ${output_path0})
-    else ()
-        add_custom_command(
-                OUTPUT ${output_path}
-                COMMAND matc ${MATC_BASE_FLAGS} -o ${output_path} ${fullname}
-                MAIN_DEPENDENCY ${fullname}
-                DEPENDS matc
-                COMMENT "Compiling material ${fullname} to ${output_path}"
+                COMMENT "Compiling material ${fullname} to ${output_path_fl0}"
         )
+        list(APPEND MATERIAL_BINS ${output_path_fl0})
+    endif ()
+
+    list(FIND MATERIAL_MULTIVIEW_SRCS ${mat_src} index)
+    if (${index} GREATER -1 AND FILAMENT_ENABLE_MULTIVIEW)
+        string(REGEX REPLACE "[.]filamat$" "_multiview.filamat" output_path_multiview ${output_path})
+        add_custom_command(
+                OUTPUT ${output_path_multiview}
+                COMMAND matc ${MATC_BASE_FLAGS} -PstereoscopicType=multiview -o ${output_path_multiview} ${fullname}
+                MAIN_DEPENDENCY ${fullname}
+                DEPENDS matc
+                COMMENT "Compiling material ${fullname} to ${output_path_multiview}"
+        )
+        list(APPEND MATERIAL_BINS ${output_path_multiview})
     endif ()
 
-    list(APPEND MATERIAL_BINS ${output_path})
 endforeach()
 
 # Additional dependencies on included files for materials

filament/backend/CMakeLists.txt

@@ -66,11 +66,13 @@ set(PRIVATE_HDRS
 # OpenGL / OpenGL ES Sources
 # ==================================================================================================
 
-if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3 AND NOT FILAMENT_USE_SWIFTSHADER)
+if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3)
     list(APPEND SRCS
             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
@@ -166,8 +168,10 @@ endif()
 if (FILAMENT_SUPPORTS_VULKAN)
     list(APPEND SRCS
             include/backend/platforms/VulkanPlatform.h
-            src/vulkan/caching/VulkanDescriptorSet.cpp
-            src/vulkan/caching/VulkanDescriptorSet.h
+            src/vulkan/caching/VulkanDescriptorSetManager.cpp
+            src/vulkan/caching/VulkanDescriptorSetManager.h
+            src/vulkan/caching/VulkanPipelineLayoutCache.cpp
+            src/vulkan/caching/VulkanPipelineLayoutCache.h
             src/vulkan/platform/VulkanPlatform.cpp
             src/vulkan/platform/VulkanPlatformSwapChainImpl.cpp
             src/vulkan/platform/VulkanPlatformSwapChainImpl.h
@@ -413,6 +417,7 @@ if (APPLE OR LINUX)
         test/test_MissingRequiredAttributes.cpp
         test/test_ReadPixels.cpp
         test/test_BufferUpdates.cpp
+        test/test_Callbacks.cpp
         test/test_MRT.cpp
         test/test_LoadImage.cpp
         test/test_StencilBuffer.cpp
@@ -474,6 +479,10 @@ if (APPLE)
        # linker from removing "unused" symbols.
        target_link_libraries(backend_test_mac PRIVATE -force_load backend_test)
        set_target_properties(backend_test_mac PROPERTIES FOLDER Tests)
+
+       # This is needed after XCode 15.3
+       set_target_properties(backend_test_mac PROPERTIES BUILD_WITH_INSTALL_RPATH TRUE)
+       set_target_properties(backend_test_mac PROPERTIES INSTALL_RPATH /usr/local/lib)
     endif()
 endif()
 

filament/backend/include/backend/DriverEnums.h

@@ -24,12 +24,14 @@
 
 #include <backend/PresentCallable.h>
 
+#include <utils/Invocable.h>
 #include <utils/ostream.h>
 
 #include <math/vec4.h>
 
 #include <array>        // FIXME: STL headers are not allowed in public headers
 #include <type_traits>  // FIXME: STL headers are not allowed in public headers
+#include <variant>      // FIXME: STL headers are not allowed in public headers
 
 #include <stddef.h>
 #include <stdint.h>
@@ -90,12 +92,15 @@ static constexpr uint64_t SWAP_CHAIN_HAS_STENCIL_BUFFER         = SWAP_CHAIN_CON
  */
 static constexpr uint64_t SWAP_CHAIN_CONFIG_PROTECTED_CONTENT   = 0x40;
 
-
 static constexpr size_t MAX_VERTEX_ATTRIBUTE_COUNT  = 16;   // This is guaranteed by OpenGL ES.
 static constexpr size_t MAX_SAMPLER_COUNT           = 62;   // Maximum needed at feature level 3.
 static constexpr size_t MAX_VERTEX_BUFFER_COUNT     = 16;   // Max number of bound buffer objects.
 static constexpr size_t MAX_SSBO_COUNT              = 4;    // This is guaranteed by OpenGL ES.
 
+static constexpr size_t MAX_PUSH_CONSTANT_COUNT     = 32;   // Vulkan 1.1 spec allows for 128-byte
+                                                            // of push constant (we assume 4-byte
+                                                            // types).
+
 // Per feature level caps
 // Use (int)FeatureLevel to index this array
 static constexpr struct {
@@ -158,14 +163,16 @@ static constexpr const char* backendToString(Backend backend) {
 }
 
 /**
- * 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.
+ * 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.
  */
 enum class ShaderLanguage {
     ESSL1 = 0,
     ESSL3 = 1,
     SPIRV = 2,
     MSL = 3,
+    METAL_LIBRARY = 4,
 };
 
 static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguage) {
@@ -178,6 +185,8 @@ static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguag
             return "SPIR-V";
         case ShaderLanguage::MSL:
             return "MSL";
+        case ShaderLanguage::METAL_LIBRARY:
+            return "Metal precompiled library";
     }
 }
 
@@ -327,7 +336,7 @@ enum class UniformType : uint8_t {
 /**
  * Supported constant parameter types
  */
- enum class ConstantType : uint8_t {
+enum class ConstantType : uint8_t {
   INT,
   FLOAT,
   BOOL
@@ -682,7 +691,7 @@ enum class TextureUsage : uint16_t {
     SUBPASS_INPUT       = 0x0020,            //!< Texture can be used as a subpass input
     BLIT_SRC            = 0x0040,            //!< Texture can be used the source of a blit()
     BLIT_DST            = 0x0080,            //!< Texture can be used the destination of a blit()
-    PROTECTED           = 0x0100,            //!< Texture can be used the destination of a blit()
+    PROTECTED           = 0x0100,            //!< Texture can be used for protected content
     DEFAULT             = UPLOADABLE | SAMPLEABLE   //!< Default texture usage
 };
 
@@ -1214,13 +1223,15 @@ struct StencilState {
     uint8_t padding = 0;
 };
 
+using PushConstantVariant = std::variant<int32_t, float, bool>;
+
 static_assert(sizeof(StencilState::StencilOperations) == 5u,
         "StencilOperations size not what was intended");
 
 static_assert(sizeof(StencilState) == 12u,
         "StencilState size not what was intended");
 
-using FrameScheduledCallback = void(*)(PresentCallable callable, void* user);
+using FrameScheduledCallback = utils::Invocable<void(backend::PresentCallable)>;
 
 enum class Workaround : uint16_t {
     // The EASU pass must split because shader compiler flattens early-exit branch
@@ -1237,9 +1248,6 @@ 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

filament/backend/include/backend/Handle.h

@@ -75,6 +75,19 @@ protected:
     HandleBase(HandleBase const& rhs) noexcept = default;
     HandleBase& operator=(HandleBase const& rhs) noexcept = default;
 
+    HandleBase(HandleBase&& rhs) noexcept
+            : object(rhs.object) {
+        rhs.object = nullid;
+    }
+
+    HandleBase& operator=(HandleBase&& rhs) noexcept {
+        if (this != &rhs) {
+            object = rhs.object;
+            rhs.object = nullid;
+        }
+        return *this;
+    }
+
 private:
     HandleId object;
 };
@@ -89,8 +102,10 @@ struct Handle : public HandleBase {
     Handle() noexcept = default;
 
     Handle(Handle const& rhs) noexcept = default;
+    Handle(Handle&& rhs) noexcept = default;
 
     Handle& operator=(Handle const& rhs) noexcept = default;
+    Handle& operator=(Handle&& rhs) noexcept = default;
 
     explicit Handle(HandleId id) noexcept : HandleBase(id) { }
 

filament/backend/include/backend/Platform.h

@@ -23,6 +23,7 @@
 #include <utils/Invocable.h>
 
 #include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
@@ -56,9 +57,20 @@ public:
 
         /**
          * Set to `true` to forcibly disable parallel shader compilation in the backend.
-         * Currently only honored by the GL backend.
+         * Currently only honored by the GL and Metal backends.
          */
         bool disableParallelShaderCompile = false;
+
+        /**
+         * Disable backend handles use-after-free checks.
+         */
+        bool disableHandleUseAfterFreeCheck = false;
+
+        /**
+         * Force GLES2 context if supported, or pretend the context is ES2. Only meaningful on
+         * GLES 3.x backends.
+         */
+        bool forceGLES2Context = false;
     };
 
     Platform() noexcept;
@@ -188,9 +200,40 @@ public:
     size_t retrieveBlob(const void* UTILS_NONNULL key, size_t keySize,
             void* UTILS_NONNULL value, size_t valueSize);
 
+    using DebugUpdateStatFunc = utils::Invocable<void(const char* UTILS_NONNULL key, uint64_t value)>;
+
+    /**
+     * Sets the callback function that the backend can use to update backend-specific statistics
+     * to aid with debugging. This callback is guaranteed to be called on the Filament driver
+     * thread.
+     *
+     * @param debugUpdateStat   an Invocable that updates debug statistics
+     */
+    void setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept;
+
+    /**
+     * @return true if debugUpdateStat is valid.
+     */
+    bool hasDebugUpdateStatFunc() const noexcept;
+
+    /**
+     * To track backend-specific statistics, the backend implementation can call the
+     * application-provided callback function debugUpdateStatFunc to associate or update a value
+     * with a given key. It is possible for this function to be called multiple times with the
+     * same key, in which case newer values should overwrite older values.
+     *
+     * This function is guaranteed to be called only on a single thread, the Filament driver
+     * thread.
+     *
+     * @param key          a null-terminated C-string with the key of the debug statistic
+     * @param value        the updated value of key
+     */
+    void debugUpdateStat(const char* UTILS_NONNULL key, uint64_t value);
+
 private:
     InsertBlobFunc mInsertBlob;
     RetrieveBlobFunc mRetrieveBlob;
+    DebugUpdateStatFunc mDebugUpdateStat;
 };
 
 } // namespace filament

filament/backend/include/backend/PresentCallable.h

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

filament/backend/include/backend/Program.h

@@ -84,6 +84,9 @@ 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.
@@ -114,8 +117,18 @@ public:
     Program& specializationConstants(
             utils::FixedCapacityVector<SpecializationConstant> specConstants) noexcept;
 
+    struct PushConstant {
+        utils::CString name;
+        ConstantType type;
+    };
+
+    Program& pushConstants(ShaderStage stage,
+            utils::FixedCapacityVector<PushConstant> constants) noexcept;
+
     Program& cacheId(uint64_t cacheId) noexcept;
 
+    Program& multiview(bool multiview) noexcept;
+
     ShaderSource const& getShadersSource() const noexcept { return mShadersSource; }
     ShaderSource& getShadersSource() noexcept { return mShadersSource; }
 
@@ -134,6 +147,8 @@ 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;
     }
@@ -141,8 +156,19 @@ public:
         return mSpecializationConstants;
     }
 
+    utils::FixedCapacityVector<PushConstant> const& getPushConstants(
+            ShaderStage stage) const noexcept {
+        return mPushConstants[static_cast<uint8_t>(stage)];
+    }
+
+    utils::FixedCapacityVector<PushConstant>& getPushConstants(ShaderStage stage) noexcept {
+        return mPushConstants[static_cast<uint8_t>(stage)];
+    }
+
     uint64_t getCacheId() const noexcept { return mCacheId; }
 
+    bool isMultiview() const noexcept { return mMultiview; }
+
     CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
 
 private:
@@ -151,13 +177,20 @@ 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;
     utils::FixedCapacityVector<SpecializationConstant> mSpecializationConstants;
+    std::array<utils::FixedCapacityVector<PushConstant>, SHADER_TYPE_COUNT> mPushConstants;
     utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> mAttributes;
     std::array<UniformInfo, Program::UNIFORM_BINDING_COUNT> mBindingUniformInfo;
     CompilerPriorityQueue mPriorityQueue = CompilerPriorityQueue::HIGH;
+    // Indicates the current engine was initialized with multiview stereo, and the variant for this
+    // program contains STE flag. This will be referred later for the OpenGL shader compiler to
+    // determine whether shader code replacement for the num_views should be performed.
+    // This variable could be promoted as a more generic variable later if other similar needs occur.
+    bool mMultiview = false;
 };
 
 } // namespace filament::backend

filament/backend/include/backend/TargetBufferInfo.h

@@ -29,13 +29,36 @@ namespace filament::backend {
 //! \privatesection
 
 struct TargetBufferInfo {
+    // note: the parameters of this constructor are not in the order of this structure's fields
+    TargetBufferInfo(Handle<HwTexture> handle, uint8_t level, uint16_t layer, uint8_t baseViewIndex) noexcept
+        : handle(handle), baseViewIndex(baseViewIndex), level(level), layer(layer) {
+    }
+
+    TargetBufferInfo(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
+            : handle(handle), level(level), layer(layer) {
+    }
+
+    TargetBufferInfo(Handle<HwTexture> handle, uint8_t level) noexcept
+            : handle(handle), level(level) {
+    }
+
+    TargetBufferInfo(Handle<HwTexture> handle) noexcept // NOLINT(*-explicit-constructor)
+            : handle(handle) {
+    }
+
+    TargetBufferInfo() noexcept = default;
+
     // texture to be used as render target
     Handle<HwTexture> handle;
 
+    // Starting layer index for multiview. This value is only used when the `layerCount` for the
+    // render target is greater than 1.
+    uint8_t baseViewIndex = 0;
+
     // level to be used
     uint8_t level = 0;
 
-    // for cubemaps and 3D textures. See TextureCubemapFace for the face->layer mapping
+    // For cubemaps and 3D textures. See TextureCubemapFace for the face->layer mapping
     uint16_t layer = 0;
 };
 
@@ -60,7 +83,7 @@ public:
 
     MRT() noexcept = default;
 
-    MRT(TargetBufferInfo const& color) noexcept // NOLINT(hicpp-explicit-conversions)
+    MRT(TargetBufferInfo const& color) noexcept // NOLINT(hicpp-explicit-conversions, *-explicit-constructor)
             : mInfos{ color } {
     }
 
@@ -80,7 +103,7 @@ public:
 
     // this is here for backward compatibility
     MRT(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
-            : mInfos{{ handle, level, layer }} {
+            : mInfos{{ handle, level, layer, 0 }} {
     }
 };
 

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

@@ -89,7 +89,7 @@ public:
      * @return              The driver's SwapChain object.
      *
      */
-    virtual SwapChain* UTILS_NONNULL createSwapChain(
+    virtual SwapChain* UTILS_NULLABLE createSwapChain(
             void* UTILS_NULLABLE nativeWindow, uint64_t flags) noexcept = 0;
 
     /**
@@ -126,29 +126,76 @@ public:
     virtual TargetBufferFlags getPreservedFlags(SwapChain* UTILS_NONNULL swapChain) noexcept;
 
     /**
-     * Called by the driver to establish the default FBO. The default implementation returns 0.
-      * @return a GLuint casted to a uint32_t that is an OpenGL framebuffer object.
+     * Returns true if the swapchain is protected
      */
-    virtual uint32_t createDefaultRenderTarget() noexcept;
+    virtual bool isSwapChainProtected(Platform::SwapChain* UTILS_NONNULL swapChain) noexcept;
 
     /**
-     * Called by the driver to make the OpenGL context active on the calling thread and bind
-     * the drawSwapChain to the default render target (FBO) created with createDefaultRenderTarget.
+     * Called by the driver to establish the default FBO. The default implementation returns 0.
+     *
+     * This method can be called either on the regular or protected OpenGL contexts and can return
+     * a different or identical name, since these names exist in different namespaces.
+     *
+     * @return a GLuint casted to a uint32_t that is an OpenGL framebuffer object.
+     */
+    virtual uint32_t getDefaultFramebufferObject() noexcept;
+
+    /**
+     * Called by the backend when a frame starts.
+     * @param steady_clock_ns vsync time point on the monotonic clock
+     * @param refreshIntervalNs refresh interval in nanosecond
+     * @param frameId a frame id
+     */
+    virtual void beginFrame(
+            int64_t monotonic_clock_ns,
+            int64_t refreshIntervalNs,
+            uint32_t frameId) noexcept;
+
+    /**
+     * Called by the backend when a frame ends.
+     * @param frameId the frame id used in beginFrame
+     */
+    virtual void endFrame(
+            uint32_t frameId) noexcept;
+
+    /**
+     * Type of contexts available
+     */
+    enum class ContextType {
+        NONE,           //!< No current context
+        UNPROTECTED,    //!< current context is unprotected
+        PROTECTED       //!< current context supports protected content
+    };
+
+    /**
+     * Returns the type of the context currently in use. This value is updated by makeCurrent()
+     * and therefore can be cached between calls. ContextType::PROTECTED can only be returned
+     * if isProtectedContextSupported() is true.
+     * @return ContextType
+     */
+    virtual ContextType getCurrentContextType() const noexcept;
+
+    /**
+     * Binds the requested context to the current thread and drawSwapChain to the default FBO
+     * returned by getDefaultFramebufferObject().
+     *
+     * @param type type of context to bind to the current thread.
      * @param drawSwapChain SwapChain to draw to. It must be bound to the default FBO.
      * @param readSwapChain SwapChain to read from (for operation like `glBlitFramebuffer`)
+     * @return true on success, false on error.
      */
-    virtual void makeCurrent(
+    virtual bool makeCurrent(ContextType type,
             SwapChain* UTILS_NONNULL drawSwapChain,
             SwapChain* UTILS_NONNULL readSwapChain) noexcept = 0;
 
     /**
      * Called by the driver to make the OpenGL context active on the calling thread and bind
-     * the drawSwapChain to the default render target (FBO) created with createDefaultRenderTarget.
+     * the drawSwapChain to the default FBO returned by getDefaultFramebufferObject().
      * The context used is either the default context or the protected context. When a context
      * change is necessary, the preContextChange and postContextChange callbacks are called,
      * before and after the context change respectively. postContextChange is given the index
      * of the new context (0 for default and 1 for protected).
-     * The default implementation just calls makeCurrent(SwapChain*, SwapChain*).
+     * The default implementation just calls makeCurrent(getCurrentContextType(), SwapChain*, SwapChain*).
      *
      * @param drawSwapChain SwapChain to draw to. It must be bound to the default FBO.
      * @param readSwapChain SwapChain to read from (for operation like `glBlitFramebuffer`)
@@ -161,6 +208,12 @@ public:
             utils::Invocable<void()> preContextChange,
             utils::Invocable<void(size_t index)> postContextChange) noexcept;
 
+    /**
+     * Called by the backend just before calling commit()
+     * @see commit()
+     */
+    virtual void preCommit() noexcept;
+
     /**
      * Called by the driver once the current frame finishes drawing. Typically, this should present
      * the drawSwapChain. This is for example where `eglMakeCurrent()` would be called.

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

@@ -57,7 +57,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
     OpenGLPlatform::ExternalTexture* createExternalImageTexture() noexcept override;
     void destroyExternalImage(ExternalTexture* texture) noexcept override;

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

@@ -45,7 +45,7 @@ public:
 
     void terminate() noexcept override;
 
-    uint32_t createDefaultRenderTarget() noexcept override;
+    uint32_t getDefaultFramebufferObject() noexcept override;
 
     bool isExtraContextSupported() const noexcept override;
     void createContext(bool shared) override;
@@ -53,7 +53,7 @@ public:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
 
     OpenGLPlatform::ExternalTexture* createExternalImageTexture() noexcept override;

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

@@ -99,11 +99,18 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
+    bool isSwapChainProtected(SwapChain* swapChain) noexcept override;
+
+    ContextType getCurrentContextType() const noexcept override;
+
+    bool makeCurrent(ContextType type,
+            SwapChain* drawSwapChain,
+            SwapChain* readSwapChain) noexcept override;
 
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain,
             utils::Invocable<void()> preContextChange,
             utils::Invocable<void(size_t index)> postContextChange) noexcept override;
+
     void commit(SwapChain* swapChain) noexcept override;
 
     bool canCreateFence() noexcept override;
@@ -132,21 +139,25 @@ protected:
     /**
      * Always use this instead of eglMakeCurrent(), as it tracks some state.
      */
-    EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept;
 
-    /**
-     * Returns true if the swapchain is protected
-     */
-    static bool isSwapChainProtected(Platform::SwapChain const* swapChain) noexcept;
+    EGLContext getContextForType(ContextType type) const noexcept;
+
+    // makes the draw and read surface current without changing the current context
+    EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
+        return egl.makeCurrent(drawSurface, readSurface);
+    }
+
+    // makes context current and set draw and read surfaces to EGL_NO_SURFACE
+    EGLBoolean makeCurrent(EGLContext context) noexcept {
+        return egl.makeCurrent(context, mEGLDummySurface, mEGLDummySurface);
+    }
 
     // TODO: this should probably use getters instead.
     EGLDisplay mEGLDisplay = EGL_NO_DISPLAY;
     EGLContext mEGLContext = EGL_NO_CONTEXT;
     EGLContext mEGLContextProtected = EGL_NO_CONTEXT;
-    EGLSurface mCurrentDrawSurface = EGL_NO_SURFACE;
-    EGLSurface mCurrentReadSurface = EGL_NO_SURFACE;
-    EGLContext mCurrentContext = EGL_NO_CONTEXT;
     EGLSurface mEGLDummySurface = EGL_NO_SURFACE;
+    ContextType mCurrentContextType = ContextType::NONE;
     // mEGLConfig is valid only if ext.egl.KHR_no_config_context is false
     EGLConfig mEGLConfig = EGL_NO_CONFIG_KHR;
     Config mContextAttribs;
@@ -181,8 +192,20 @@ protected:
     EGLConfig findSwapChainConfig(uint64_t flags, bool window, bool pbuffer) const;
 
 private:
-    EGLBoolean makeCurrent(EGLContext context,
-            EGLSurface drawSurface, EGLSurface readSurface) noexcept;
+    class EGL {
+        EGLDisplay& mEGLDisplay;
+        EGLSurface mCurrentDrawSurface = EGL_NO_SURFACE;
+        EGLSurface mCurrentReadSurface = EGL_NO_SURFACE;
+        EGLContext mCurrentContext = EGL_NO_CONTEXT;
+    public:
+        explicit EGL(EGLDisplay& dpy) : mEGLDisplay(dpy) {}
+        EGLBoolean makeCurrent(EGLContext context,
+                EGLSurface drawSurface, EGLSurface readSurface) noexcept;
+
+        EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
+            return makeCurrent(mCurrentContext, drawSurface, readSurface);
+        }
+    } egl{ mEGLDisplay };
 };
 
 } // namespace filament::backend

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

@@ -22,6 +22,10 @@
 #include <backend/platforms/OpenGLPlatform.h>
 #include <backend/platforms/PlatformEGL.h>
 
+#include <utils/android/PerformanceHintManager.h>
+
+#include <chrono>
+
 #include <stddef.h>
 #include <stdint.h>
 
@@ -58,6 +62,13 @@ protected:
 
     void terminate() noexcept override;
 
+    void beginFrame(
+            int64_t monotonic_clock_ns,
+            int64_t refreshIntervalNs,
+            uint32_t frameId) noexcept override;
+
+    void preCommit() noexcept override;
+
     /**
      * Set the presentation time using `eglPresentationTimeANDROID`
      * @param presentationTimeInNanosecond
@@ -81,6 +92,11 @@ protected:
 private:
     int mOSVersion;
     ExternalStreamManagerAndroid& mExternalStreamManager;
+    utils::PerformanceHintManager mPerformanceHintManager;
+    utils::PerformanceHintManager::Session mPerformanceHintSession;
+
+    using clock = std::chrono::high_resolution_clock;
+    clock::time_point mStartTimeOfActualWork;
 };
 
 } // namespace filament::backend

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

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

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

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

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

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

filament/backend/include/backend/platforms/VulkanPlatform.h

@@ -23,9 +23,9 @@
 
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
+#include <utils/Hash.h>
 #include <utils/PrivateImplementation.h>
 
-#include <string_view>
 #include <tuple>
 #include <unordered_set>
 
@@ -47,6 +47,14 @@ struct VulkanPlatformPrivate;
 class VulkanPlatform : public Platform, utils::PrivateImplementation<VulkanPlatformPrivate> {
 public:
 
+    struct ExtensionHashFn {
+        std::size_t operator()(utils::CString const& s) const noexcept {
+            return std::hash<std::string>{}(s.data());
+        }
+    };
+    // Utility for managing device or instance extensions during initialization.
+    using ExtensionSet = std::unordered_set<utils::CString, ExtensionHashFn>;
+
     /**
      * A collection of handles to objects and metadata that comprises a Vulkan context. The client
      * can instantiate this struct and pass to Engine::Builder::sharedContext if they wishes to
@@ -192,6 +200,13 @@ public:
     virtual SwapChainPtr createSwapChain(void* nativeWindow, uint64_t flags = 0,
             VkExtent2D extent = {0, 0});
 
+    /**
+     * Allows implementers to provide instance extensions that they'd like to include in the
+     * instance creation.
+     * @return          A set of extensions to enable for the instance.
+     */
+    virtual ExtensionSet getRequiredInstanceExtensions() { return {}; }
+
     /**
      * Destroy the swapchain.
      * @param handle    The handle returned by createSwapChain()
@@ -236,10 +251,9 @@ public:
     VkQueue getGraphicsQueue() const noexcept;
 
 private:
-    // Platform dependent helper methods
-    using ExtensionSet = std::unordered_set<std::string_view>;
-    static ExtensionSet getRequiredInstanceExtensions();
+    static ExtensionSet getSwapchainInstanceExtensions();
 
+    // Platform dependent helper methods
     using SurfaceBundle = std::tuple<VkSurfaceKHR, VkExtent2D>;
     static SurfaceBundle createVkSurfaceKHR(void* nativeWindow, VkInstance instance,
             uint64_t flags) noexcept;

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

@@ -50,12 +50,13 @@ class CommandBufferQueue {
     size_t mFreeSpace = 0;
     size_t mHighWatermark = 0;
     uint32_t mExitRequested = 0;
+    bool mPaused = false;
 
     static constexpr uint32_t EXIT_REQUESTED = 0x31415926;
 
 public:
     // requiredSize: guaranteed available space after flush()
-    CommandBufferQueue(size_t requiredSize, size_t bufferSize);
+    CommandBufferQueue(size_t requiredSize, size_t bufferSize, bool paused);
     ~CommandBufferQueue();
 
     CircularBuffer& getCircularBuffer() noexcept { return mCircularBuffer; }
@@ -80,6 +81,10 @@ public:
     // returns from waitForCommands() immediately.
     void requestExit();
 
+    // suspend or unsuspend the queue.
+    bool isPaused() const noexcept;
+    void setPaused(bool paused);
+
     bool isExitRequested() const;
 };
 

filament/backend/include/private/backend/CommandStream.h

@@ -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) noexcept {
+        static inline void execute(M&& method, D&& driver, CommandBase* base, intptr_t* next) {
             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) noexcept;
+    static void execute(Driver&, CommandBase* base, intptr_t* next);
 public:
     inline CustomCommand(CustomCommand&& rhs) = default;
     inline explicit CustomCommand(std::function<void()> cmd)

filament/backend/include/private/backend/Driver.h

@@ -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) noexcept;
+    virtual void execute(std::function<void(void)> const& fn);
 
     // This is called on debug build, or when enabled manually on the backend thread side.
     virtual void debugCommandBegin(CommandStream* cmds,

filament/backend/include/private/backend/DriverAPI.inc

@@ -133,18 +133,18 @@ DECL_DRIVER_API_0(tick)
 
 DECL_DRIVER_API_N(beginFrame,
         int64_t, monotonic_clock_ns,
+        int64_t, refreshIntervalNs,
         uint32_t, frameId)
 
 DECL_DRIVER_API_N(setFrameScheduledCallback,
         backend::SwapChainHandle, sch,
-        backend::FrameScheduledCallback, callback,
-        void*, user)
+        backend::CallbackHandler*, handler,
+        backend::FrameScheduledCallback&&, callback)
 
 DECL_DRIVER_API_N(setFrameCompletedCallback,
         backend::SwapChainHandle, sch,
         backend::CallbackHandler*, handler,
-        backend::CallbackHandler::Callback, callback,
-        void*, user)
+        utils::Invocable<void(void)>&&, callback)
 
 DECL_DRIVER_API_N(setPresentationTime,
         int64_t, monotonic_clock_ns)
@@ -239,6 +239,7 @@ DECL_DRIVER_API_R_N(backend::RenderTargetHandle, createRenderTarget,
         uint32_t, width,
         uint32_t, height,
         uint8_t, samples,
+        uint8_t, layerCount,
         backend::MRT, color,
         backend::TargetBufferInfo, depth,
         backend::TargetBufferInfo, stencil)
@@ -302,6 +303,7 @@ DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedContentSupported)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isStereoSupported, backend::StereoscopicType, stereoscopicType)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isParallelShaderCompileSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthStencilResolveSupported)
+DECL_DRIVER_API_SYNCHRONOUS_N(bool, isDepthStencilBlitSupported, backend::TextureFormat, format)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedTexturesSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(uint8_t, getMaxDrawBuffers)
 DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getMaxUniformBufferSize)
@@ -431,6 +433,11 @@ DECL_DRIVER_API_N(bindSamplers,
         uint32_t, index,
         backend::SamplerGroupHandle, sbh)
 
+DECL_DRIVER_API_N(setPushConstant,
+        backend::ShaderStage, stage,
+        uint8_t, index,
+        backend::PushConstantVariant, value)
+
 DECL_DRIVER_API_N(insertEventMarker,
         const char*, string,
         uint32_t, len = 0)
@@ -494,7 +501,7 @@ DECL_DRIVER_API_N(blit,
         math::uint2, size)
 
 DECL_DRIVER_API_N(bindPipeline,
-        backend::PipelineState, state)
+        backend::PipelineState const&, state)
 
 DECL_DRIVER_API_N(bindRenderPrimitive,
         backend::RenderPrimitiveHandle, rph)

filament/backend/include/private/backend/HandleAllocator.h

@@ -49,7 +49,7 @@ namespace filament::backend {
 template<size_t P0, size_t P1, size_t P2>
 class HandleAllocator {
 public:
-    HandleAllocator(const char* name, size_t size) noexcept;
+    HandleAllocator(const char* name, size_t size, bool disableUseAfterFreeCheck) noexcept;
     HandleAllocator(HandleAllocator const& rhs) = delete;
     HandleAllocator& operator=(HandleAllocator const& rhs) = delete;
     ~HandleAllocator();
@@ -65,7 +65,7 @@ public:
      *
      */
     template<typename D, typename ... ARGS>
-    Handle<D> allocateAndConstruct(ARGS&& ... args) noexcept {
+    Handle<D> allocateAndConstruct(ARGS&& ... args) {
         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) noexcept {
+    destroyAndConstruct(Handle<B> const& handle, ARGS&& ... args) {
         assert_invariant(handle);
         D* addr = handle_cast<D*>(const_cast<Handle<B>&>(handle));
         assert_invariant(addr);
@@ -163,17 +163,19 @@ 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) noexcept {
+    handle_cast(Handle<B>& handle) {
         assert_invariant(handle);
         auto [p, tag] = handleToPointer(handle.getId());
 
         if (isPoolHandle(handle.getId())) {
             // check for use after free
-            uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
-            auto const pNode = static_cast<typename Allocator::Node*>(p);
-            uint8_t const expectedAge = pNode[-1].age;
-            ASSERT_POSTCONDITION(expectedAge == age,
-                    "use-after-free of Handle with id=%d", handle.getId());
+            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
+                uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
+                auto const pNode = static_cast<typename Allocator::Node*>(p);
+                uint8_t const expectedAge = pNode[-1].age;
+                ASSERT_POSTCONDITION(expectedAge == age,
+                        "use-after-free of Handle with id=%d", handle.getId());
+            }
         }
 
         return static_cast<Dp>(p);
@@ -183,11 +185,10 @@ 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) noexcept {
+    handle_cast(Handle<B> const& handle) {
         return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
     }
 
-
 private:
 
     template<typename D>
@@ -210,8 +211,9 @@ private:
         Pool<P1> mPool1;
         Pool<P2> mPool2;
         UTILS_UNUSED_IN_RELEASE const utils::AreaPolicy::HeapArea& mArea;
+        bool mUseAfterFreeCheckDisabled;
     public:
-        explicit Allocator(const utils::AreaPolicy::HeapArea& area);
+        explicit Allocator(const utils::AreaPolicy::HeapArea& area, bool disableUseAfterFreeCheck);
 
         static constexpr size_t getAlignment() noexcept { return MIN_ALIGNMENT; }
 
@@ -237,8 +239,10 @@ private:
             // check for double-free
             Node* const pNode = static_cast<Node*>(p);
             uint8_t& expectedAge = pNode[-1].age;
-            ASSERT_POSTCONDITION(expectedAge == age,
-                    "double-free of Handle of size %d at %p", size, p);
+            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
+                ASSERT_POSTCONDITION(expectedAge == age,
+                        "double-free of Handle of size %d at %p", size, p);
+            }
             expectedAge = (expectedAge + 1) & 0xF; // fixme
 
             if (size <= mPool0.getSize()) { mPool0.free(p); return; }
@@ -313,7 +317,7 @@ private:
         return (id & HANDLE_HEAP_FLAG) == 0u;
     }
 
-    HandleBase::HandleId allocateHandleSlow(size_t size) noexcept;
+    HandleBase::HandleId allocateHandleSlow(size_t size);
     void deallocateHandleSlow(HandleBase::HandleId id, size_t size) noexcept;
 
     // We inline this because it's just 4 instructions in the fast case
@@ -348,6 +352,7 @@ private:
     mutable utils::Mutex mLock;
     tsl::robin_map<HandleBase::HandleId, void*> mOverflowMap;
     HandleBase::HandleId mId = 0;
+    bool mUseAfterFreeCheckDisabled = false;
 };
 
 } // namespace filament::backend

filament/backend/src/CommandBufferQueue.cpp

@@ -39,10 +39,11 @@ using namespace utils;
 
 namespace filament::backend {
 
-CommandBufferQueue::CommandBufferQueue(size_t requiredSize, size_t bufferSize)
+CommandBufferQueue::CommandBufferQueue(size_t requiredSize, size_t bufferSize, bool paused)
         : mRequiredSize((requiredSize + (CircularBuffer::getBlockSize() - 1u)) & ~(CircularBuffer::getBlockSize() -1u)),
           mCircularBuffer(bufferSize),
-          mFreeSpace(mCircularBuffer.size()) {
+          mFreeSpace(mCircularBuffer.size()),
+          mPaused(paused) {
     assert_invariant(mCircularBuffer.size() > requiredSize);
 }
 
@@ -56,6 +57,21 @@ void CommandBufferQueue::requestExit() {
     mCondition.notify_one();
 }
 
+bool CommandBufferQueue::isPaused() const noexcept {
+    std::lock_guard<utils::Mutex> const lock(mLock);
+    return mPaused;
+}
+
+void CommandBufferQueue::setPaused(bool paused) {
+    std::lock_guard<utils::Mutex> const lock(mLock);
+    if (paused) {
+        mPaused = true;
+    } else {
+        mPaused = false;
+        mCondition.notify_one();
+    }
+}
+
 bool CommandBufferQueue::isExitRequested() const {
     std::lock_guard<utils::Mutex> const lock(mLock);
     ASSERT_PRECONDITION( mExitRequested == 0 || mExitRequested == EXIT_REQUESTED,
@@ -115,6 +131,9 @@ void CommandBufferQueue::flush() noexcept {
 #endif
 
         SYSTRACE_NAME("waiting: CircularBuffer::flush()");
+        ASSERT_POSTCONDITION(!mPaused,
+                "CommandStream is full, but since the rendering thread is paused, "
+                "the buffer cannot flush and we will deadlock. Instead, abort.");
         mCondition.wait(lock, [this, requiredSize]() -> bool {
             // TODO: on macOS, we need to call pumpEvents from time to time
             return mFreeSpace >= requiredSize;
@@ -127,7 +146,7 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
         return std::move(mCommandBuffersToExecute);
     }
     std::unique_lock<utils::Mutex> lock(mLock);
-    while (mCommandBuffersToExecute.empty() && !mExitRequested) {
+    while ((mCommandBuffersToExecute.empty() || mPaused) && !mExitRequested) {
         mCondition.wait(lock);
     }
 

filament/backend/src/CommandStream.cpp

@@ -149,7 +149,7 @@ void CommandType<void (Driver::*)(ARGS...)>::Command<METHOD>::log() noexcept  {
 
 // ------------------------------------------------------------------------------------------------
 
-void CustomCommand::execute(Driver&, CommandBase* base, intptr_t* next) noexcept {
+void CustomCommand::execute(Driver&, CommandBase* base, intptr_t* next) {
     *next = CustomCommand::align(sizeof(CustomCommand));
     static_cast<CustomCommand*>(base)->mCommand();
     static_cast<CustomCommand*>(base)->~CustomCommand();

filament/backend/src/Driver.cpp

@@ -214,7 +214,7 @@ size_t Driver::getElementTypeSize(ElementType type) noexcept {
 
 Driver::~Driver() noexcept = default;
 
-void Driver::execute(std::function<void(void)> const& fn) noexcept {
+void Driver::execute(std::function<void(void)> const& fn) {
     fn();
 }
 

filament/backend/src/HandleAllocator.cpp

@@ -39,8 +39,10 @@ using namespace utils;
 
 template <size_t P0, size_t P1, size_t P2>
 UTILS_NOINLINE
-HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& area)
-        : mArea(area) {
+HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& area,
+        bool disableUseAfterFreeCheck)
+        : mArea(area),
+          mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
 
     // The largest handle this allocator can generate currently depends on the architecture's
     // min alignment, typically 8 or 16 bytes.
@@ -74,8 +76,10 @@ HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& ar
 // ------------------------------------------------------------------------------------------------
 
 template <size_t P0, size_t P1, size_t P2>
-HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size) noexcept
-    : mHandleArena(name, size) {
+HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size,
+        bool disableUseAfterFreeCheck) noexcept
+    : mHandleArena(name, size, disableUseAfterFreeCheck),
+      mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
 }
 
 template <size_t P0, size_t P1, size_t P2>
@@ -103,7 +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) noexcept {
+HandleBase::HandleId HandleAllocator<P0, P1, P2>::allocateHandleSlow(size_t size) {
     void* p = ::malloc(size);
     std::unique_lock lock(mLock);
 

filament/backend/src/Platform.cpp

@@ -53,4 +53,18 @@ size_t Platform::retrieveBlob(void const* key, size_t keySize, void* value, size
     return 0;
 }
 
+void Platform::setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept {
+    mDebugUpdateStat = std::move(debugUpdateStat);
+}
+
+bool Platform::hasDebugUpdateStatFunc() const noexcept {
+    return bool(mDebugUpdateStat);
+}
+
+void Platform::debugUpdateStat(const char* key, uint64_t value) {
+    if (mDebugUpdateStat) {
+        mDebugUpdateStat(key, value);
+    }
+}
+
 } // namespace filament::backend

filament/backend/src/PlatformFactory.cpp

@@ -29,21 +29,21 @@
         #include "backend/platforms/PlatformCocoaTouchGL.h"
     #endif
 #elif defined(__APPLE__)
-    #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3) && !defined(FILAMENT_USE_SWIFTSHADER)
+    #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
         #include <backend/platforms/PlatformCocoaGL.h>
     #endif
 #elif defined(__linux__)
     #if defined(FILAMENT_SUPPORTS_X11)
-        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3) && !defined(FILAMENT_USE_SWIFTSHADER)
+        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
             #include "backend/platforms/PlatformGLX.h"
         #endif
     #elif defined(FILAMENT_SUPPORTS_EGL_ON_LINUX)
-        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3) && !defined(FILAMENT_USE_SWIFTSHADER)
+        #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
             #include "backend/platforms/PlatformEGLHeadless.h"
         #endif
     #endif
 #elif defined(WIN32)
-    #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3) && !defined(FILAMENT_USE_SWIFTSHADER)
+    #if defined(FILAMENT_SUPPORTS_OPENGL) && !defined(FILAMENT_USE_EXTERNAL_GLES3)
         #include "backend/platforms/PlatformWGL.h"
     #endif
 #elif defined(__EMSCRIPTEN__)
@@ -111,8 +111,7 @@ Platform* PlatformFactory::create(Backend* backend) noexcept {
     }
     assert_invariant(*backend == Backend::OPENGL);
     #if defined(FILAMENT_SUPPORTS_OPENGL)
-        #if defined(FILAMENT_USE_EXTERNAL_GLES3) || defined(FILAMENT_USE_SWIFTSHADER)
-            // Swiftshader OpenGLES support is deprecated and incomplete
+        #if defined(FILAMENT_USE_EXTERNAL_GLES3)
             return nullptr;
         #elif defined(__ANDROID__)
             return new PlatformEGLAndroid();

filament/backend/src/Program.cpp

@@ -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;
@@ -47,6 +47,11 @@ 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) {
@@ -86,11 +91,22 @@ Program& Program::specializationConstants(
     return *this;
 }
 
+Program& Program::pushConstants(ShaderStage stage,
+        utils::FixedCapacityVector<PushConstant> constants) noexcept {
+    mPushConstants[static_cast<uint8_t>(stage)] = std::move(constants);
+    return *this;
+}
+
 Program& Program::cacheId(uint64_t cacheId) noexcept {
     mCacheId = cacheId;
     return *this;
 }
 
+Program& Program::multiview(bool multiview) noexcept {
+    mMultiview = multiview;
+    return *this;
+}
+
 io::ostream& operator<<(io::ostream& out, const Program& builder) {
     out << "Program{";
     builder.mLogger(out);

filament/backend/src/metal/MetalBuffer.h

@@ -18,7 +18,7 @@
 #define TNT_FILAMENT_DRIVER_METALBUFFER_H
 
 #include "MetalContext.h"
-#include "MetalBufferPool.h"
+#include "MetalPlatform.h"
 
 #include <backend/DriverEnums.h>
 
@@ -28,9 +28,125 @@
 
 #include <utility>
 #include <memory>
+#include <atomic>
+#include <chrono>
 
 namespace filament::backend {
 
+class ScopedAllocationTimer {
+public:
+    ScopedAllocationTimer(const char* name) : mBeginning(clock_t::now()), mName(name) {}
+    ~ScopedAllocationTimer() {
+        using namespace std::literals::chrono_literals;
+        static constexpr std::chrono::seconds LONG_TIME_THRESHOLD = 10s;
+
+        auto end = clock_t::now();
+        std::chrono::duration<double, std::micro> allocationTimeMicroseconds = end - mBeginning;
+
+        if (UTILS_UNLIKELY(allocationTimeMicroseconds > LONG_TIME_THRESHOLD)) {
+            if (platform && platform->hasDebugUpdateStatFunc()) {
+                char buffer[64];
+                snprintf(buffer, sizeof(buffer), "filament.metal.long_buffer_allocation_time.%s",
+                        mName);
+                platform->debugUpdateStat(
+                        buffer, static_cast<uint64_t>(allocationTimeMicroseconds.count()));
+            }
+        }
+    }
+
+    static void setPlatform(MetalPlatform* p) { platform = p; }
+
+private:
+    typedef std::chrono::steady_clock clock_t;
+
+    static MetalPlatform* platform;
+
+    std::chrono::time_point<clock_t> mBeginning;
+    const char* mName;
+};
+
+#ifndef FILAMENT_METAL_BUFFER_TRACKING
+#define FILAMENT_METAL_BUFFER_TRACKING 0
+#endif
+
+class MetalBufferTracking {
+public:
+    static constexpr size_t EXCESS_BUFFER_COUNT = 30000;
+
+    enum class Type {
+        NONE = 0,
+        GENERIC = 1,
+        RING = 2,
+        STAGING = 3,
+    };
+    static constexpr size_t TypeCount = 3;
+
+    static constexpr auto toIndex(Type t) {
+        assert_invariant(t != Type::NONE);
+        switch (t) {
+            case Type::NONE:
+            case Type::GENERIC:
+                return 0;
+            case Type::RING:
+                return 1;
+            case Type::STAGING:
+                return 2;
+        }
+    }
+
+#if FILAMENT_METAL_BUFFER_TRACKING
+    static void initialize() {
+        static dispatch_once_t onceToken;
+        dispatch_once(&onceToken, ^{
+            for (size_t i = 0; i < TypeCount; i++) {
+                aliveBuffers[i] = [NSHashTable weakObjectsHashTable];
+            }
+        });
+    }
+
+    static void setPlatform(MetalPlatform* p) { platform = p; }
+
+    static void track(id<MTLBuffer> buffer, Type type) {
+        assert_invariant(type != Type::NONE);
+        if (UTILS_UNLIKELY(getAliveBuffers() >= EXCESS_BUFFER_COUNT)) {
+            if (platform && platform->hasDebugUpdateStatFunc()) {
+                platform->debugUpdateStat("filament.metal.excess_buffers_allocated",
+                        MetalBufferTracking::getAliveBuffers());
+            }
+        }
+        [aliveBuffers[toIndex(type)] addObject:buffer];
+    }
+
+    static uint64_t getAliveBuffers() {
+        uint64_t sum = 0;
+        for (size_t i = 1; i < TypeCount; i++) {
+            sum += getAliveBuffers(static_cast<Type>(i));
+        }
+        return sum;
+    }
+
+    static uint64_t getAliveBuffers(Type type) {
+        assert_invariant(type != Type::NONE);
+        NSHashTable* hashTable = aliveBuffers[toIndex(type)];
+        // Caution! We can't simply use hashTable.count here, which is inaccurate.
+        // See http://cocoamine.net/blog/2013/12/13/nsmaptable-and-zeroing-weak-references/
+        return hashTable.objectEnumerator.allObjects.count;
+    }
+#else
+    static void initialize() {}
+    static void setPlatform(MetalPlatform* p) {}
+    static id<MTLBuffer> track(id<MTLBuffer> buffer, Type type) { return buffer; }
+    static uint64_t getAliveBuffers() { return 0; }
+    static uint64_t getAliveBuffers(Type type) { return 0; }
+#endif
+
+private:
+#if FILAMENT_METAL_BUFFER_TRACKING
+    static std::array<NSHashTable<id<MTLBuffer>>*, TypeCount> aliveBuffers;
+    static MetalPlatform* platform;
+#endif
+};
+
 class MetalBuffer {
 public:
 
@@ -82,7 +198,7 @@ public:
 
 private:
 
-    id<MTLBuffer> mBuffer = nil;
+    id<MTLBuffer> mBuffer;
     size_t mBufferSize = 0;
     void* mCpuBuffer = nullptr;
     MetalContext& mContext;
@@ -131,7 +247,11 @@ public:
           mBufferOptions(options),
           mSlotSizeBytes(computeSlotSize(layout)),
           mSlotCount(slotCount) {
-        mBuffer = [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions];
+        {
+            ScopedAllocationTimer timer("ring");
+            mBuffer = [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions];
+        }
+        MetalBufferTracking::track(mBuffer, MetalBufferTracking::Type::RING);
         assert_invariant(mBuffer);
     }
 
@@ -149,9 +269,13 @@ public:
             // If we already have an aux buffer, it will get freed here, unless it has been retained
             // by a MTLCommandBuffer. In that case, it will be freed when the command buffer
             // finishes executing.
-            mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
+            {
+                ScopedAllocationTimer timer("ring");
+                mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
+            }
+            MetalBufferTracking::track(mAuxBuffer, MetalBufferTracking::Type::RING);
             assert_invariant(mAuxBuffer);
-            return {mAuxBuffer, 0};
+            return { mAuxBuffer, 0 };
         }
         mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
         mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);

filament/backend/src/metal/MetalBuffer.mm

@@ -15,12 +15,21 @@
  */
 
 #include "MetalBuffer.h"
+#include "MetalBufferPool.h"
 
 #include "MetalContext.h"
 
 namespace filament {
 namespace backend {
 
+MetalPlatform* ScopedAllocationTimer::platform = nullptr;
+
+#if FILAMENT_METAL_BUFFER_TRACKING
+std::array<NSHashTable<id<MTLBuffer>>*, MetalBufferTracking::TypeCount>
+        MetalBufferTracking::aliveBuffers;
+MetalPlatform* MetalBufferTracking::platform = nullptr;
+#endif
+
 MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
         size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
     // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
@@ -34,7 +43,11 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
     }
 
     // Otherwise, we allocate a private GPU buffer.
-    mBuffer = [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate];
+    {
+        ScopedAllocationTimer timer("generic");
+        mBuffer = [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate];
+    }
+    MetalBufferTracking::track(mBuffer, MetalBufferTracking::Type::GENERIC);
     ASSERT_POSTCONDITION(mBuffer, "Could not allocate Metal buffer of size %zu.", size);
 }
 

filament/backend/src/metal/MetalBufferPool.h

@@ -19,6 +19,8 @@
 
 #include <Metal/Metal.h>
 
+#include "MetalBuffer.h"
+
 #include <map>
 #include <mutex>
 #include <unordered_set>

filament/backend/src/metal/MetalBufferPool.mm

@@ -42,15 +42,20 @@ MetalBufferPoolEntry const* MetalBufferPool::acquireBuffer(size_t numBytes) {
     }
 
     // We were not able to find a sufficiently large stage, so create a new one.
-    id<MTLBuffer> buffer = [mContext.device newBufferWithLength:numBytes
-                                                        options:MTLResourceStorageModeShared];
+    id<MTLBuffer> buffer = nil;
+    {
+        ScopedAllocationTimer timer("staging");
+        buffer = [mContext.device newBufferWithLength:numBytes
+                                              options:MTLResourceStorageModeShared];
+    }
+    MetalBufferTracking::track(buffer, MetalBufferTracking::Type::STAGING);
     ASSERT_POSTCONDITION(buffer, "Could not allocate Metal staging buffer of size %zu.", numBytes);
-    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry({
+    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry {
         .buffer = buffer,
         .capacity = numBytes,
         .lastAccessed = mCurrentFrame,
         .referenceCount = 1
-    });
+    };
     mUsedStages.insert(stage);
 
     return stage;

filament/backend/src/metal/MetalDriver.mm

@@ -20,6 +20,7 @@
 #include "metal/MetalDriver.h"
 
 #include "MetalBlitter.h"
+#include "MetalBufferPool.h"
 #include "MetalContext.h"
 #include "MetalDriverFactory.h"
 #include "MetalEnums.h"
@@ -36,6 +37,7 @@
 
 #include <utils/Log.h>
 #include <utils/Panic.h>
+#include <utils/sstream.h>
 
 #include <algorithm>
 
@@ -98,9 +100,15 @@ Dispatcher MetalDriver::getDispatcher() const noexcept {
 MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept
         : mPlatform(*platform),
           mContext(new MetalContext(driverConfig.textureUseAfterFreePoolSize)),
-          mHandleAllocator("Handles", driverConfig.handleArenaSize) {
+          mHandleAllocator("Handles",
+                  driverConfig.handleArenaSize,
+                  driverConfig.disableHandleUseAfterFreeCheck) {
     mContext->driver = this;
 
+    ScopedAllocationTimer::setPlatform(platform);
+    MetalBufferTracking::initialize();
+    MetalBufferTracking::setPlatform(platform);
+
     mContext->device = mPlatform.createDevice();
     assert_invariant(mContext->device);
 
@@ -180,7 +188,10 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
         mContext->eventListener = [[MTLSharedEventListener alloc] initWithDispatchQueue:queue];
     }
 
-    mContext->shaderCompiler = new MetalShaderCompiler(mContext->device, *this);
+    const MetalShaderCompiler::Mode compilerMode = driverConfig.disableParallelShaderCompile
+            ? MetalShaderCompiler::Mode::SYNCHRONOUS
+            : MetalShaderCompiler::Mode::ASYNCHRONOUS;
+    mContext->shaderCompiler = new MetalShaderCompiler(mContext->device, *this, compilerMode);
     mContext->shaderCompiler->init();
 
 #if defined(FILAMENT_METAL_PROFILING)
@@ -191,6 +202,8 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
 }
 
 MetalDriver::~MetalDriver() noexcept {
+    MetalBufferTracking::setPlatform(nullptr);
+    ScopedAllocationTimer::setPlatform(nullptr);
     mContext->device = nil;
     mContext->emptyTexture = nil;
     CFRelease(mContext->textureCache);
@@ -205,22 +218,35 @@ void MetalDriver::tick(int) {
     executeTickOps();
 }
 
-void MetalDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
+void MetalDriver::beginFrame(int64_t monotonic_clock_ns,
+        int64_t refreshIntervalNs, uint32_t frameId) {
 #if defined(FILAMENT_METAL_PROFILING)
     os_signpost_interval_begin(mContext->log, mContext->signpostId, "Frame encoding", "%{public}d", frameId);
 #endif
+    if (mPlatform.hasDebugUpdateStatFunc()) {
+#if FILAMENT_METAL_BUFFER_TRACKING
+        const uint64_t generic = MetalBufferTracking::getAliveBuffers(MetalBufferTracking::Type::GENERIC);
+        const uint64_t ring = MetalBufferTracking::getAliveBuffers(MetalBufferTracking::Type::RING);
+        const uint64_t staging = MetalBufferTracking::getAliveBuffers(MetalBufferTracking::Type::STAGING);
+        const uint64_t total = generic + ring + staging;
+        mPlatform.debugUpdateStat("filament.metal.alive_buffers", total);
+        mPlatform.debugUpdateStat("filament.metal.alive_buffers.generic", generic);
+        mPlatform.debugUpdateStat("filament.metal.alive_buffers.ring", ring);
+        mPlatform.debugUpdateStat("filament.metal.alive_buffers.staging", staging);
+#endif
+    }
 }
 
-void MetalDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        FrameScheduledCallback callback, void* user) {
+void MetalDriver::setFrameScheduledCallback(
+        Handle<HwSwapChain> sch, CallbackHandler* handler, FrameScheduledCallback&& callback) {
     auto* swapChain = handle_cast<MetalSwapChain>(sch);
-    swapChain->setFrameScheduledCallback(callback, user);
+    swapChain->setFrameScheduledCallback(handler, std::move(callback));
 }
 
-void MetalDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
+void MetalDriver::setFrameCompletedCallback(
+        Handle<HwSwapChain> sch, CallbackHandler* handler, utils::Invocable<void(void)>&& callback) {
     auto* swapChain = handle_cast<MetalSwapChain>(sch);
-    swapChain->setFrameCompletedCallback(handler, callback, user);
+    swapChain->setFrameCompletedCallback(handler, std::move(callback));
 }
 
 void MetalDriver::execute(std::function<void(void)> const& fn) noexcept {
@@ -373,7 +399,7 @@ void MetalDriver::createDefaultRenderTargetR(Handle<HwRenderTarget> rth, int dum
 
 void MetalDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         TargetBufferFlags targetBufferFlags, uint32_t width, uint32_t height,
-        uint8_t samples, MRT color,
+        uint8_t samples, uint8_t layerCount, MRT color,
         TargetBufferInfo depth, TargetBufferInfo stencil) {
     ASSERT_PRECONDITION(!isInRenderPass(mContext),
             "createRenderTarget must be called outside of a render pass.");
@@ -788,13 +814,17 @@ bool MetalDriver::isStereoSupported(backend::StereoscopicType stereoscopicType)
 }
 
 bool MetalDriver::isParallelShaderCompileSupported() {
-    return true;
+    return mContext->shaderCompiler->isParallelShaderCompileSupported();
 }
 
 bool MetalDriver::isDepthStencilResolveSupported() {
     return false;
 }
 
+bool MetalDriver::isDepthStencilBlitSupported(TextureFormat format) {
+    return true;
+}
+
 bool MetalDriver::isProtectedTexturesSupported() {
     return false;
 }
@@ -1226,6 +1256,9 @@ void MetalDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
     mContext->samplerBindings[index] = sb;
 }
 
+void MetalDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
+        backend::PushConstantVariant value) {}
+
 void MetalDriver::insertEventMarker(const char* string, uint32_t len) {
 
 }
@@ -1604,7 +1637,7 @@ void MetalDriver::finalizeSamplerGroup(MetalSamplerGroup* samplerGroup) {
     }
 }
 
-void MetalDriver::bindPipeline(PipelineState ps) {
+void MetalDriver::bindPipeline(PipelineState const& ps) {
     ASSERT_PRECONDITION(mContext->currentRenderPassEncoder != nullptr,
             "bindPipeline() without a valid command encoder.");
 
@@ -1624,7 +1657,7 @@ void MetalDriver::bindPipeline(PipelineState ps) {
         return;
     }
 
-    ASSERT_PRECONDITION(bool(functions), "Attempting to bind an invalid Metal program.");
+    functions.validate();
 
     auto [fragment, vertex] = functions.getRasterFunctions();
 

filament/backend/src/metal/MetalHandles.h

@@ -31,6 +31,8 @@
 
 #include "private/backend/SamplerGroup.h"
 
+#include <backend/DriverEnums.h>
+
 #include <utils/bitset.h>
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
@@ -71,9 +73,9 @@ public:
 
     void releaseDrawable();
 
-    void setFrameScheduledCallback(FrameScheduledCallback callback, void* user);
-    void setFrameCompletedCallback(CallbackHandler* handler,
-            CallbackHandler::Callback callback, void* user);
+    void setFrameScheduledCallback(CallbackHandler* handler, FrameScheduledCallback&& callback);
+    void setFrameCompletedCallback(
+            CallbackHandler* handler, utils::Invocable<void(void)>&& callback);
 
     // For CAMetalLayer-backed SwapChains, presents the drawable or schedules a
     // FrameScheduledCallback.
@@ -110,19 +112,19 @@ private:
     MetalExternalImage externalImage;
     SwapChainType type;
 
-    // These two fields store a callback and user data to notify the client that a frame is ready
-    // for presentation.
-    // If frameScheduledCallback is nullptr, then the Metal backend automatically calls
-    // presentDrawable when the frame is committed.
-    // Otherwise, the Metal backend will not automatically present the frame. Instead, clients bear
-    // the responsibility of presenting the frame by calling the PresentCallable object.
-    FrameScheduledCallback frameScheduledCallback = nullptr;
-    void* frameScheduledUserData = nullptr;
+    // These fields store a callback to notify the client that a frame is ready for presentation. If
+    // !frameScheduled.callback, then the Metal backend automatically calls presentDrawable when the
+    // frame is committed. Otherwise, the Metal backend will not automatically present the frame.
+    // Instead, clients bear the responsibility of presenting the frame by calling the
+    // PresentCallable object.
+    struct {
+        CallbackHandler* handler = nullptr;
+        std::shared_ptr<FrameScheduledCallback> callback = nullptr;
+    } frameScheduled;
 
     struct {
         CallbackHandler* handler = nullptr;
-        CallbackHandler::Callback callback = {};
-        void* user = nullptr;
+        std::shared_ptr<utils::Invocable<void(void)>> callback = nullptr;
     } frameCompleted;
 };
 

filament/backend/src/metal/MetalHandles.mm

@@ -19,6 +19,7 @@
 #include "MetalBlitter.h"
 #include "MetalEnums.h"
 #include "MetalUtils.h"
+#include "MetalBufferPool.h"
 
 #include <filament/SwapChain.h>
 
@@ -220,16 +221,16 @@ void MetalSwapChain::ensureDepthStencilTexture() {
     depthStencilTexture = [context.device newTextureWithDescriptor:descriptor];
 }
 
-void MetalSwapChain::setFrameScheduledCallback(FrameScheduledCallback callback, void* user) {
-    frameScheduledCallback = callback;
-    frameScheduledUserData = user;
+void MetalSwapChain::setFrameScheduledCallback(
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
+    frameScheduled.handler = handler;
+    frameScheduled.callback = std::make_shared<FrameScheduledCallback>(std::move(callback));
 }
 
-void MetalSwapChain::setFrameCompletedCallback(CallbackHandler* handler,
-        CallbackHandler::Callback callback, void* user) {
+void MetalSwapChain::setFrameCompletedCallback(
+        CallbackHandler* handler, utils::Invocable<void(void)>&& callback) {
     frameCompleted.handler = handler;
-    frameCompleted.callback = callback;
-    frameCompleted.user = user;
+    frameCompleted.callback = std::make_shared<utils::Invocable<void(void)>>(std::move(callback));
 }
 
 void MetalSwapChain::present() {
@@ -237,7 +238,7 @@ void MetalSwapChain::present() {
         scheduleFrameCompletedCallback();
     }
     if (drawable) {
-        if (frameScheduledCallback) {
+        if (frameScheduled.callback) {
             scheduleFrameScheduledCallback();
         } else  {
             [getPendingCommandBuffer(&context) presentDrawable:drawable];
@@ -295,21 +296,38 @@ void presentDrawable(bool presentFrame, void* user) {
 }
 
 void MetalSwapChain::scheduleFrameScheduledCallback() {
-    if (!frameScheduledCallback) {
+    if (!frameScheduled.callback) {
         return;
     }
 
     assert_invariant(drawable);
 
-    // Destroy this by calling maybePresentAndDestroyAsync() later.
-    auto* presentData = PresentDrawableData::create(drawable, context.driver);
+    struct Callback {
+        Callback(std::shared_ptr<FrameScheduledCallback> callback, id<CAMetalDrawable> drawable,
+                MetalDriver* driver)
+            : f(callback), data(PresentDrawableData::create(drawable, driver)) {}
+        std::shared_ptr<FrameScheduledCallback> f;
+        // PresentDrawableData* is destroyed by maybePresentAndDestroyAsync() later.
+        std::unique_ptr<PresentDrawableData> data;
+        static void func(void* user) {
+            auto* const c = reinterpret_cast<Callback*>(user);
+            PresentDrawableData* presentDrawableData = c->data.release();
+            PresentCallable presentCallable(presentDrawable, presentDrawableData);
+            c->f->operator()(presentCallable);
+            delete c;
+        }
+    };
 
-    FrameScheduledCallback userCallback = frameScheduledCallback;
-    void* userData = frameScheduledUserData;
+    // This callback pointer will be captured by the block. Even if the scheduled handler is never
+    // called, the unique_ptr will still ensure we don't leak memory.
+    __block auto callback =
+            std::make_unique<Callback>(frameScheduled.callback, drawable, context.driver);
 
+    backend::CallbackHandler* handler = frameScheduled.handler;
+    MetalDriver* driver = context.driver;
     [getPendingCommandBuffer(&context) addScheduledHandler:^(id<MTLCommandBuffer> cb) {
-        PresentCallable callable(presentDrawable, static_cast<void*>(presentData));
-        userCallback(callable, userData);
+        Callback* user = callback.release();
+        driver->scheduleCallback(handler, user, &Callback::func);
     }];
 }
 
@@ -318,13 +336,25 @@ void MetalSwapChain::scheduleFrameCompletedCallback() {
         return;
     }
 
-    CallbackHandler* handler = frameCompleted.handler;
-    void* user = frameCompleted.user;
-    CallbackHandler::Callback callback = frameCompleted.callback;
+    struct Callback {
+        Callback(std::shared_ptr<utils::Invocable<void(void)>> callback) : f(callback) {}
+        std::shared_ptr<utils::Invocable<void(void)>> f;
+        static void func(void* user) {
+            auto* const c = reinterpret_cast<Callback*>(user);
+            c->f->operator()();
+            delete c;
+        }
+    };
 
+    // This callback pointer will be captured by the block. Even if the completed handler is never
+    // called, the unique_ptr will still ensure we don't leak memory.
+    __block auto callback = std::make_unique<Callback>(frameCompleted.callback);
+
+    CallbackHandler* handler = frameCompleted.handler;
     MetalDriver* driver = context.driver;
     [getPendingCommandBuffer(&context) addCompletedHandler:^(id<MTLCommandBuffer> cb) {
-        driver->scheduleCallback(handler, user, callback);
+        Callback* user = callback.release();
+        driver->scheduleCallback(handler, user, &Callback::func);
     }];
 }
 

filament/backend/src/metal/MetalShaderCompiler.h

@@ -41,6 +41,11 @@ class MetalShaderCompiler {
     struct MetalProgramToken;
 
 public:
+    enum class Mode {
+        SYNCHRONOUS,    // synchronous shader compilation
+        ASYNCHRONOUS    // asynchronous shader compilation
+    };
+
     class MetalFunctionBundle {
     public:
         using Raster = std::tuple<id<MTLFunction>, id<MTLFunction>>;
@@ -110,7 +115,7 @@ public:
 
     using program_token_t = std::shared_ptr<MetalProgramToken>;
 
-    explicit MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver);
+    explicit MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver, Mode mode);
 
     MetalShaderCompiler(MetalShaderCompiler const& rhs) = delete;
     MetalShaderCompiler(MetalShaderCompiler&& rhs) = delete;
@@ -120,15 +125,15 @@ public:
     void init() noexcept;
     void terminate() noexcept;
 
-    // Creates a program asynchronously
+    bool isParallelShaderCompileSupported() const noexcept;
+
+    // Creates a program, either synchronously or asynchronously, depending on the Mode
+    // MetalShaderCompiler was constructed with.
     program_token_t createProgram(utils::CString const& name, Program&& program);
 
     // Returns the functions, blocking if necessary. The Token is destroyed and becomes invalid.
     MetalFunctionBundle getProgram(program_token_t& token);
 
-    // Destroys a valid token and all associated resources. Used to "cancel" a program compilation.
-    static void terminate(program_token_t& token);
-
     void notifyWhenAllProgramsAreReady(
             CallbackHandler* handler, CallbackHandler::Callback callback, void* user);
 
@@ -138,6 +143,7 @@ private:
     CompilerThreadPool mCompilerThreadPool;
     id<MTLDevice> mDevice;
     CallbackManager mCallbackManager;
+    Mode mMode;
 };
 
 } // namespace filament::backend

filament/backend/src/metal/MetalShaderCompiler.mm

@@ -70,22 +70,31 @@ struct MetalShaderCompiler::MetalProgramToken : ProgramToken {
 
 MetalShaderCompiler::MetalProgramToken::~MetalProgramToken() = default;
 
-MetalShaderCompiler::MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver)
+MetalShaderCompiler::MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver, Mode mode)
         : mDevice(device),
-          mCallbackManager(driver) {
+          mCallbackManager(driver),
+          mMode(mode) {
 
 }
 
 void MetalShaderCompiler::init() noexcept {
     const uint32_t poolSize = 1;
-    mCompilerThreadPool.init(poolSize, []() {}, []() {});
+    if (mMode == Mode::ASYNCHRONOUS) {
+        mCompilerThreadPool.init(poolSize, []() {}, []() {});
+    }
 }
 
 void MetalShaderCompiler::terminate() noexcept {
-    mCompilerThreadPool.terminate();
+    if (mMode == Mode::ASYNCHRONOUS) {
+        mCompilerThreadPool.terminate();
+    }
     mCallbackManager.terminate();
 }
 
+bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
+    return mMode == Mode::ASYNCHRONOUS;
+}
+
 /* static */ MetalShaderCompiler::MetalFunctionBundle MetalShaderCompiler::compileProgram(
         const Program& program, id<MTLDevice> device) {
     std::array<id<MTLFunction>, Program::SHADER_TYPE_COUNT> functions = { nil };
@@ -97,26 +106,41 @@ void MetalShaderCompiler::terminate() 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 = [device newLibraryWithSource:objcSource
-                                                      options:options
-                                                        error:&error];
+        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;
+        }
+
         if (library == nil) {
             NSString* errorMessage = @"unknown error";
             if (error) {
@@ -145,6 +169,23 @@ void MetalShaderCompiler::terminate() noexcept {
         id<MTLFunction> function = [library newFunctionWithName:@"main0"
                                                  constantValues:constants
                                                           error:&error];
+        if (function == nil) {
+            // If the library loads but functions within it fail to load, it usually means the
+            // GPU backend crashed. (This can happen if it's a Metallib shader that was compiled
+            // with a minimum iOS version that's newer than this device.)
+            NSString* errorMessage = @"unknown error";
+            if (error) {
+                auto description =
+                        [error.localizedDescription cStringUsingEncoding:NSUTF8StringEncoding];
+                utils::slog.w << description << utils::io::endl;
+                errorMessage = error.localizedDescription;
+            }
+            PANIC_LOG("Failed to load main0 in Metal program.");
+            NSString* programName =
+                [NSString stringWithFormat:@"%s::main0", program.getName().c_str_safe()];
+            return MetalFunctionBundle::error(errorMessage, programName);
+        }
+
         if (!program.getName().empty()) {
             function.label = @(program.getName().c_str());
         }
@@ -180,14 +221,26 @@ MetalShaderCompiler::program_token_t MetalShaderCompiler::createProgram(
 
     token->handle = mCallbackManager.get();
 
-    CompilerPriorityQueue const priorityQueue = program.getPriorityQueue();
-    mCompilerThreadPool.queue(priorityQueue, token,
-            [this, name, device = mDevice, program = std::move(program), token]() {
-                MetalFunctionBundle compiledProgram = compileProgram(program, device);
+    switch (mMode) {
+        case Mode::ASYNCHRONOUS: {
+            CompilerPriorityQueue const priorityQueue = program.getPriorityQueue();
+            mCompilerThreadPool.queue(priorityQueue, token,
+                    [this, name, device = mDevice, program = std::move(program), token]() {
+                        MetalFunctionBundle compiledProgram = compileProgram(program, device);
+                        token->set(compiledProgram);
+                        mCallbackManager.put(token->handle);
+                    });
 
-                token->set(compiledProgram);
-                mCallbackManager.put(token->handle);
-            });
+            break;
+        }
+
+        case Mode::SYNCHRONOUS: {
+            MetalFunctionBundle compiledProgram = compileProgram(program, mDevice);
+            token->set(compiledProgram);
+            mCallbackManager.put(token->handle);
+            break;
+        }
+    }
 
     return token;
 }
@@ -195,35 +248,23 @@ MetalShaderCompiler::program_token_t MetalShaderCompiler::createProgram(
 MetalShaderCompiler::MetalFunctionBundle MetalShaderCompiler::getProgram(program_token_t& token) {
     assert_invariant(token);
 
-    if (!token->isReady()) {
-        auto job = mCompilerThreadPool.dequeue(token);
-        if (job) {
-            job();
+    if (mMode == Mode::ASYNCHRONOUS) {
+        if (!token->isReady()) {
+            auto job = mCompilerThreadPool.dequeue(token);
+            if (job) {
+                job();
+            }
         }
     }
 
+    // The job isn't guaranteed to have finished yet. We may have failed to dequeue it above,
+    // which means it's currently running. In that case get() will block until it finishes.
+
     MetalShaderCompiler::MetalFunctionBundle program = token->get();
-
     token = nullptr;
-
     return program;
 }
 
-/* static */ void MetalShaderCompiler::terminate(program_token_t& token) {
-    assert_invariant(token);
-
-    auto job = token->compiler.mCompilerThreadPool.dequeue(token);
-    if (!job) {
-        // The job is being executed right now (or has already executed).
-        token->wait();
-    } else {
-        // The job has not executed yet.
-        token->compiler.mCallbackManager.put(token->handle);
-    }
-
-    token.reset();
-}
-
 void MetalShaderCompiler::notifyWhenAllProgramsAreReady(
         CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
     mCallbackManager.setCallback(handler, callback, user);

filament/backend/src/noop/NoopDriver.cpp

@@ -49,16 +49,17 @@ void NoopDriver::terminate() {
 void NoopDriver::tick(int) {
 }
 
-void NoopDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
+void NoopDriver::beginFrame(int64_t monotonic_clock_ns,
+        int64_t refreshIntervalNs, uint32_t frameId) {
 }
 
 void NoopDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        FrameScheduledCallback callback, void* user) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
 
 }
 
 void NoopDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
+        CallbackHandler* handler, utils::Invocable<void(void)>&& callback) {
 
 }
 
@@ -193,6 +194,10 @@ bool NoopDriver::isDepthStencilResolveSupported() {
     return true;
 }
 
+bool NoopDriver::isDepthStencilBlitSupported(TextureFormat format) {
+    return true;
+}
+
 bool NoopDriver::isProtectedTexturesSupported() {
     return true;
 }
@@ -307,6 +312,10 @@ void NoopDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
 void NoopDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
 }
 
+void NoopDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
+        backend::PushConstantVariant value) {
+}
+
 void NoopDriver::insertEventMarker(char const* string, uint32_t len) {
 }
 
@@ -350,7 +359,7 @@ void NoopDriver::blit(
         math::uint2 size) {
 }
 
-void NoopDriver::bindPipeline(PipelineState pipelineState) {
+void NoopDriver::bindPipeline(PipelineState const& pipelineState) {
 }
 
 void NoopDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {

filament/backend/src/opengl/GLBufferObject.h

@@ -0,0 +1,51 @@
+/*
+ * 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

filament/backend/src/opengl/GLTexture.h

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

filament/backend/src/opengl/OpenGLContext.cpp

@@ -16,11 +16,25 @@
 
 #include "OpenGLContext.h"
 
+#include "GLUtils.h"
+#include "OpenGLTimerQuery.h"
+
 #include <backend/platforms/OpenGLPlatform.h>
+#include <backend/DriverEnums.h>
+
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/Log.h>
+#include <utils/ostream.h>
 
 #include <functional>
+#include <string_view>
 #include <utility>
 
+#include <stddef.h>
+#include <stdio.h>
+#include <string.h>
+
 // change to true to display all GL extensions in the console on start-up
 #define DEBUG_PRINT_EXTENSIONS false
 
@@ -49,7 +63,10 @@ bool OpenGLContext::queryOpenGLVersion(GLint* major, GLint* minor) noexcept {
 #endif
 }
 
-OpenGLContext::OpenGLContext(OpenGLPlatform& platform) noexcept {
+OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
+        Platform::DriverConfig const& driverConfig) noexcept
+        : mPlatform(platform),
+          mSamplerMap(32) {
 
     state.vao.p = &mDefaultVAO;
 
@@ -69,6 +86,13 @@ OpenGLContext::OpenGLContext(OpenGLPlatform& platform) noexcept {
 
     queryOpenGLVersion(&state.major, &state.minor);
 
+    #if defined(BACKEND_OPENGL_VERSION_GLES)
+    if (UTILS_UNLIKELY(driverConfig.forceGLES2Context)) {
+        state.major = 2;
+        state.minor = 0;
+    }
+    #endif
+
     OpenGLContext::initExtensions(&ext, state.major, state.minor);
 
     OpenGLContext::initProcs(&procs, ext, state.major, state.minor);
@@ -237,9 +261,23 @@ 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) {
@@ -270,6 +308,9 @@ void OpenGLContext::synchronizeStateAndCache(size_t index) noexcept {
         }
     }
 
+    // the default FBO could be invalid
+    mDefaultFbo[index].reset();
+
     contextIndex = index;
     resetState();
 }
@@ -498,8 +539,6 @@ 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") ||
@@ -751,6 +790,51 @@ void OpenGLContext::initExtensionsGL(Extensions* ext, GLint major, GLint minor)
 
 #endif // BACKEND_OPENGL_VERSION_GL
 
+
+GLuint OpenGLContext::bindFramebuffer(GLenum target, GLuint buffer) noexcept {
+    if (UTILS_UNLIKELY(buffer == 0)) {
+        // we're binding the default frame buffer, resolve its actual name
+        auto& defaultFboForThisContext = mDefaultFbo[contextIndex];
+        if (UTILS_UNLIKELY(!defaultFboForThisContext.has_value())) {
+            defaultFboForThisContext = GLuint(mPlatform.getDefaultFramebufferObject());
+        }
+        buffer = defaultFboForThisContext.value();
+    }
+    bindFramebufferResolved(target, buffer);
+    return buffer;
+}
+
+void OpenGLContext::unbindFramebuffer(GLenum target) noexcept {
+    bindFramebufferResolved(target, 0);
+}
+
+void OpenGLContext::bindFramebufferResolved(GLenum target, GLuint buffer) noexcept {
+    switch (target) {
+        case GL_FRAMEBUFFER:
+            if (state.draw_fbo != buffer || state.read_fbo != buffer) {
+                state.draw_fbo = state.read_fbo = buffer;
+                glBindFramebuffer(target, buffer);
+            }
+            break;
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+        case GL_DRAW_FRAMEBUFFER:
+            if (state.draw_fbo != buffer) {
+                state.draw_fbo = buffer;
+                glBindFramebuffer(target, buffer);
+            }
+            break;
+        case GL_READ_FRAMEBUFFER:
+            if (state.read_fbo != buffer) {
+                state.read_fbo = buffer;
+                glBindFramebuffer(target, buffer);
+            }
+            break;
+#endif
+        default:
+            break;
+    }
+}
+
 void OpenGLContext::bindBuffer(GLenum target, GLuint buffer) noexcept {
     if (target == GL_ELEMENT_ARRAY_BUFFER) {
         constexpr size_t targetIndex = getIndexForBufferTarget(GL_ELEMENT_ARRAY_BUFFER);
@@ -805,19 +889,32 @@ default_case:
     }
 }
 
-void OpenGLContext::unbindTexture(GLenum target, GLuint texture_id) noexcept {
+void OpenGLContext::unbindTexture(
+        UTILS_UNUSED_IN_RELEASE GLenum target, GLuint texture_id) noexcept {
     // unbind this texture from all the units it might be bound to
     // no need unbind the texture from FBOs because we're not tracking that state (and there is
     // no need to).
-    const size_t index = getIndexForTextureTarget(target);
-    UTILS_NOUNROLL
-    for (GLuint unit = 0; unit < MAX_TEXTURE_UNIT_COUNT; unit++) {
-        if (state.textures.units[unit].targets[index].texture_id == texture_id) {
-            bindTexture(unit, target, (GLuint)0, index);
+    // Never attempt to unbind texture 0. This could happen with external textures w/ streaming if
+    // never populated.
+    if (texture_id) {
+        UTILS_NOUNROLL
+        for (GLuint unit = 0; unit < MAX_TEXTURE_UNIT_COUNT; unit++) {
+            if (state.textures.units[unit].id == texture_id) {
+                // if this texture is bound, it should be at the same target
+                assert_invariant(state.textures.units[unit].target == target);
+                unbindTextureUnit(unit);
+            }
         }
     }
 }
 
+void OpenGLContext::unbindTextureUnit(GLuint unit) noexcept {
+    update_state(state.textures.units[unit].id, 0u, [&]() {
+        activeTexture(unit);
+        glBindTexture(state.textures.units[unit].target, 0u);
+    });
+}
+
 void OpenGLContext::unbindSampler(GLuint sampler) noexcept {
     // unbind this sampler from all the units it might be bound to
     UTILS_NOUNROLL  // clang generates >800B of code!!!
@@ -871,6 +968,37 @@ 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
 

filament/backend/src/opengl/OpenGLContext.h

@@ -17,24 +17,34 @@
 #ifndef TNT_FILAMENT_BACKEND_OPENGLCONTEXT_H
 #define TNT_FILAMENT_BACKEND_OPENGLCONTEXT_H
 
-#include <math/vec4.h>
 
 #include "OpenGLTimerQuery.h"
 
-#include <utils/CString.h>
-#include <utils/debug.h>
+#include <backend/platforms/OpenGLPlatform.h>
 
+#include <backend/DriverEnums.h>
 #include <backend/Handle.h>
 
-#include "GLUtils.h"
+#include "gl_headers.h"
+
+#include <utils/compiler.h>
+#include <utils/bitset.h>
+#include <utils/debug.h>
+
+#include <math/vec2.h>
+#include <math/vec4.h>
+
+#include <tsl/robin_map.h>
 
 #include <array>
 #include <functional>
-#include <set>
+#include <optional>
 #include <tuple>
-#include <utility>
 #include <vector>
 
+#include <stddef.h>
+#include <stdint.h>
+
 namespace filament::backend {
 
 class OpenGLPlatform;
@@ -81,8 +91,12 @@ public:
 
     static bool queryOpenGLVersion(GLint* major, GLint* minor) noexcept;
 
-    explicit OpenGLContext(OpenGLPlatform& platform) noexcept;
-    ~OpenGLContext() noexcept;
+    explicit OpenGLContext(OpenGLPlatform& platform,
+            Platform::DriverConfig const& driverConfig) noexcept;
+
+    ~OpenGLContext() noexcept final;
+
+    void terminate() noexcept;
 
     // TimerQueryInterface ------------------------------------------------------------------------
 
@@ -124,7 +138,6 @@ public:
 #endif
     }
 
-    constexpr static inline size_t getIndexForTextureTarget(GLuint target) noexcept;
     constexpr        inline size_t getIndexForCap(GLenum cap) noexcept;
     constexpr static inline size_t getIndexForBufferTarget(GLenum target) noexcept;
 
@@ -136,10 +149,10 @@ public:
 
           void pixelStore(GLenum, GLint) noexcept;
     inline void activeTexture(GLuint unit) noexcept;
-    inline void bindTexture(GLuint unit, GLuint target, GLuint texId, size_t targetIndex) noexcept;
     inline void bindTexture(GLuint unit, GLuint target, GLuint texId) noexcept;
 
            void unbindTexture(GLenum target, GLuint id) noexcept;
+           void unbindTextureUnit(GLuint unit) noexcept;
     inline void bindVertexArray(RenderPrimitive const* p) noexcept;
     inline void bindSampler(GLuint unit, GLuint sampler) noexcept;
            void unbindSampler(GLuint sampler) noexcept;
@@ -148,7 +161,8 @@ public:
     inline void bindBufferRange(GLenum target, GLuint index, GLuint buffer,
             GLintptr offset, GLsizeiptr size) noexcept;
 
-    inline void bindFramebuffer(GLenum target, GLuint buffer) noexcept;
+    GLuint bindFramebuffer(GLenum target, GLuint buffer) noexcept;
+    void unbindFramebuffer(GLenum target) noexcept;
 
     inline void enableVertexAttribArray(RenderPrimitive const* rp, GLuint index) noexcept;
     inline void disableVertexAttribArray(RenderPrimitive const* rp, GLuint index) noexcept;
@@ -297,10 +311,6 @@ 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;
@@ -312,8 +322,14 @@ public:
 
     // function to handle state changes we don't control
     void updateTexImage(GLenum target, GLuint id) noexcept {
-        const size_t index = getIndexForTextureTarget(target);
-        state.textures.units[state.textures.active].targets[index].texture_id = id;
+        assert_invariant(target == GL_TEXTURE_EXTERNAL_OES);
+        // if another target is bound to this texture unit, unbind that texture
+        if (UTILS_UNLIKELY(state.textures.units[state.textures.active].target != target)) {
+            glBindTexture(state.textures.units[state.textures.active].target, 0);
+            state.textures.units[state.textures.active].target = GL_TEXTURE_EXTERNAL_OES;
+        }
+        // the texture is already bound to `target`, we just update our internal state
+        state.textures.units[state.textures.active].id = id;
     }
     void resetProgram() noexcept { state.program.use = 0; }
 
@@ -416,9 +432,8 @@ public:
             GLuint active = 0;      // zero-based
             struct {
                 GLuint sampler = 0;
-                struct {
-                    GLuint texture_id = 0;
-                } targets[7];  // this must match getIndexForTextureTarget()
+                GLuint target = 0;
+                GLuint id = 0;
             } units[MAX_TEXTURE_UNIT_COUNT];
         } textures;
 
@@ -458,12 +473,45 @@ 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;
     ShaderModel mShaderModel = ShaderModel::MOBILE;
     FeatureLevel mFeatureLevel = FeatureLevel::FEATURE_LEVEL_1;
     TimerQueryFactoryInterface* mTimerQueryFactory = nullptr;
     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;
 
     const std::array<std::tuple<bool const&, char const*, char const*>, sizeof(bugs)> mBugDatabase{{
             {   bugs.disable_glFlush,
@@ -508,16 +556,11 @@ private:
             {   bugs.delay_fbo_destruction,
                     "delay_fbo_destruction",
                     ""},
-            {   bugs.disable_thread_affinity,
-                    "disable_thread_affinity",
-                    ""},
             {   bugs.force_feature_level0,
                     "force_feature_level0",
                     ""},
     }};
 
-    RenderPrimitive mDefaultVAO;
-
     // this is chosen to minimize code size
 #if defined(BACKEND_OPENGL_VERSION_GLES)
     static void initExtensionsGLES(Extensions* ext, GLint major, GLint minor) noexcept;
@@ -560,31 +603,10 @@ private:
     }
 
     void setDefaultState() noexcept;
-
-    static constexpr const size_t TEXTURE_TARGET_COUNT =
-            sizeof(state.textures.units[0].targets) / sizeof(state.textures.units[0].targets[0]);
-
 };
 
 // ------------------------------------------------------------------------------------------------
 
-constexpr size_t OpenGLContext::getIndexForTextureTarget(GLuint target) noexcept {
-    // this must match state.textures[].targets[]
-    switch (target) {
-        case GL_TEXTURE_2D:                     return 0;
-        case GL_TEXTURE_2D_ARRAY:               return 1;
-        case GL_TEXTURE_CUBE_MAP:               return 2;
-#if defined(BACKEND_OPENGL_LEVEL_GLES31)
-        case GL_TEXTURE_2D_MULTISAMPLE:         return 3;
-#endif
-        case GL_TEXTURE_EXTERNAL_OES:           return 4;
-        case GL_TEXTURE_3D:                     return 5;
-        case GL_TEXTURE_CUBE_MAP_ARRAY:         return 6;
-        default:
-            return 0;
-    }
-}
-
 constexpr size_t OpenGLContext::getIndexForCap(GLenum cap) noexcept { //NOLINT
     size_t index = 0;
     switch (cap) {
@@ -732,46 +754,17 @@ void OpenGLContext::bindBufferRange(GLenum target, GLuint index, GLuint buffer,
 #endif
 }
 
-void OpenGLContext::bindFramebuffer(GLenum target, GLuint buffer) noexcept {
-    switch (target) {
-        case GL_FRAMEBUFFER:
-            if (state.draw_fbo != buffer || state.read_fbo != buffer) {
-                state.draw_fbo = state.read_fbo = buffer;
-                glBindFramebuffer(target, buffer);
-            }
-            break;
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-        case GL_DRAW_FRAMEBUFFER:
-            if (state.draw_fbo != buffer) {
-                state.draw_fbo = buffer;
-                glBindFramebuffer(target, buffer);
-            }
-            break;
-        case GL_READ_FRAMEBUFFER:
-            if (state.read_fbo != buffer) {
-                state.read_fbo = buffer;
-                glBindFramebuffer(target, buffer);
-            }
-            break;
-#endif
-        default:
-            break;
-    }
-}
-
-void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId, size_t targetIndex) noexcept {
-    assert_invariant(targetIndex == getIndexForTextureTarget(target));
-    assert_invariant(targetIndex < TEXTURE_TARGET_COUNT);
-    update_state(state.textures.units[unit].targets[targetIndex].texture_id, texId, [&]() {
+void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId) noexcept {
+    update_state(state.textures.units[unit].target, target, [&]() {
+        activeTexture(unit);
+        glBindTexture(state.textures.units[unit].target, 0);
+    });
+    update_state(state.textures.units[unit].id, texId, [&]() {
         activeTexture(unit);
         glBindTexture(target, texId);
     }, target == GL_TEXTURE_EXTERNAL_OES);
 }
 
-void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId) noexcept {
-    bindTexture(unit, target, texId, getIndexForTextureTarget(target));
-}
-
 void OpenGLContext::useProgram(GLuint program) noexcept {
     update_state(state.program.use, program, [&]() {
         glUseProgram(program);

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -16,21 +16,54 @@
 
 #include "OpenGLDriver.h"
 
-#include "private/backend/DriverApi.h"
-
 #include "CommandStreamDispatcher.h"
+#include "GLUtils.h"
 #include "OpenGLContext.h"
 #include "OpenGLDriverFactory.h"
 #include "OpenGLProgram.h"
 #include "OpenGLTimerQuery.h"
+#include "gl_headers.h"
 
 #include <backend/platforms/OpenGLPlatform.h>
-#include <backend/SamplerDescriptor.h>
 
-#include <utils/compiler.h>
+#include <backend/BufferDescriptor.h>
+#include <backend/CallbackHandler.h>
+#include <backend/DriverApiForward.h>
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+#include <backend/PipelineState.h>
+#include <backend/Platform.h>
+#include <backend/Program.h>
+#include <backend/SamplerDescriptor.h>
+#include <backend/TargetBufferInfo.h>
+
+#include "private/backend/Dispatcher.h"
+#include "private/backend/DriverApi.h"
+
+#include <utils/BitmaskEnum.h>
+#include <utils/CString.h>
 #include <utils/Log.h>
 #include <utils/Panic.h>
 #include <utils/Systrace.h>
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/ostream.h>
+
+#include <math/vec2.h>
+#include <math/vec3.h>
+
+#include <algorithm>
+#include <chrono>
+#include <functional>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <utility>
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
 
 #if defined(__EMSCRIPTEN__)
 #include <emscripten.h>
@@ -140,6 +173,10 @@ Driver* OpenGLDriver::create(OpenGLPlatform* const platform,
         PANIC_LOG("OpenGL ES 2.0 minimum needed (current %d.%d)", major, minor);
         goto cleanup;
     }
+    if (UTILS_UNLIKELY(driverConfig.forceGLES2Context)) {
+        major = 2;
+        minor = 0;
+    }
 #else
     // we require GL 4.1 headers and minimum version
     if (UTILS_UNLIKELY(!((major == 4 && minor >= 1) || major > 4))) {
@@ -170,11 +207,13 @@ OpenGLDriver::DebugMarker::~DebugMarker() noexcept {
 
 OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept
         : mPlatform(*platform),
-          mContext(mPlatform),
+          mContext(mPlatform, driverConfig),
           mShaderCompilerService(*this),
-          mHandleAllocator("Handles", driverConfig.handleArenaSize),
-          mSamplerMap(32),
-          mDriverConfig(driverConfig) {
+          mHandleAllocator("Handles",
+                  driverConfig.handleArenaSize,
+                  driverConfig.disableHandleUseAfterFreeCheck),
+          mDriverConfig(driverConfig),
+          mCurrentPushConstants(new(std::nothrow) PushConstantBundle{}) {
 
     std::fill(mSamplerBindings.begin(), mSamplerBindings.end(), nullptr);
 
@@ -198,10 +237,19 @@ 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 {
-    return ConcreteDispatcher<OpenGLDriver>::make();
+    auto dispatcher = ConcreteDispatcher<OpenGLDriver>::make();
+    if (mContext.isES2()) {
+        dispatcher.draw2_ = +[](Driver& driver, CommandBase* base, intptr_t* next){
+            using Cmd = COMMAND_TYPE(draw2);
+            OpenGLDriver& concreteDriver = static_cast<OpenGLDriver&>(driver);
+            Cmd::execute(&OpenGLDriver::draw2GLES2, concreteDriver, base, next);
+        };
+    }
+    return dispatcher;
 }
 
 // ------------------------------------------------------------------------------------------------
@@ -228,16 +276,13 @@ void OpenGLDriver::terminate() {
 
     // because we called glFinish(), all callbacks should have been executed
     assert_invariant(mGpuCommandCompleteOps.empty());
-
-    if (!getContext().isES2()) {
-        for (auto& item: mSamplerMap) {
-            mContext.unbindSampler(item.second);
-            glDeleteSamplers(1, &item.second);
-        }
-        mSamplerMap.clear();
-    }
 #endif
 
+    delete mCurrentPushConstants;
+    mCurrentPushConstants = nullptr;
+
+    mContext.terminate();
+
     mPlatform.terminate();
 }
 
@@ -256,9 +301,42 @@ void OpenGLDriver::bindSampler(GLuint unit, GLuint sampler) noexcept {
     mContext.bindSampler(unit, sampler);
 }
 
+void OpenGLDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
+        backend::PushConstantVariant value) {
+    assert_invariant(stage == ShaderStage::VERTEX || stage == ShaderStage::FRAGMENT);
+    utils::Slice<std::pair<GLint, ConstantType>> constants;
+    if (stage == ShaderStage::VERTEX) {
+        constants = mCurrentPushConstants->vertexConstants;
+    } else if (stage == ShaderStage::FRAGMENT) {
+        constants = mCurrentPushConstants->fragmentConstants;
+    }
+
+    assert_invariant(index < constants.size());
+    auto const& [location, type] = constants[index];
+
+    // This push constant wasn't found in the shader. It's ok to return without error-ing here.
+    if (location < 0) {
+        return;
+    }
+
+    if (std::holds_alternative<bool>(value)) {
+        assert_invariant(type == ConstantType::BOOL);
+        bool const bval = std::get<bool>(value);
+        glUniform1i(location, bval ? 1 : 0);
+    } else if (std::holds_alternative<float>(value)) {
+        assert_invariant(type == ConstantType::FLOAT);
+        float const fval = std::get<float>(value);
+        glUniform1f(location, fval);
+    } else {
+        assert_invariant(type == ConstantType::INT);
+        int const ival = std::get<int>(value);
+        glUniform1i(location, ival);
+    }
+}
+
 void OpenGLDriver::bindTexture(GLuint unit, GLTexture const* t) noexcept {
     assert_invariant(t != nullptr);
-    mContext.bindTexture(unit, t->gl.target, t->gl.id, t->gl.targetIndex);
+    mContext.bindTexture(unit, t->gl.target, t->gl.id);
 }
 
 bool OpenGLDriver::useProgram(OpenGLProgram* p) noexcept {
@@ -272,7 +350,7 @@ 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] = mUniformBindings[i];
+            auto [id, buffer, age] = mContext.getEs2UniformBinding(i);
             if (buffer) {
                 p->updateUniforms(i, id, buffer, age);
             }
@@ -689,10 +767,19 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
     GLenum internalFormat = getInternalFormat(format);
     assert_invariant(internalFormat);
 
-    if (UTILS_UNLIKELY(usage & TextureUsage::PROTECTED)) {
+    if (any(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;
@@ -711,7 +798,6 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
             if (t->externalTexture) {
                 t->gl.target = t->externalTexture->target;
                 t->gl.id = t->externalTexture->id;
-                t->gl.targetIndex = (uint8_t)OpenGLContext::getIndexForTextureTarget(t->gl.target);
                 // internalFormat actually depends on the external image, but it doesn't matter
                 // because it's not used anywhere for anything important.
                 t->gl.internalFormat = internalFormat;
@@ -723,30 +809,23 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
 
             t->gl.internalFormat = internalFormat;
 
-            // We DO NOT update targetIndex at function exit to take advantage of the fact that
-            // getIndexForTextureTarget() is constexpr -- so all of this disappears at compile time.
             switch (target) {
                 case SamplerType::SAMPLER_EXTERNAL:
                     // we can't be here -- doesn't matter what we do
                 case SamplerType::SAMPLER_2D:
                     t->gl.target = GL_TEXTURE_2D;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D);
                     break;
                 case SamplerType::SAMPLER_3D:
                     t->gl.target = GL_TEXTURE_3D;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_3D);
                     break;
                 case SamplerType::SAMPLER_2D_ARRAY:
                     t->gl.target = GL_TEXTURE_2D_ARRAY;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
                     break;
                 case SamplerType::SAMPLER_CUBEMAP:
                     t->gl.target = GL_TEXTURE_CUBE_MAP;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
                     break;
                 case SamplerType::SAMPLER_CUBEMAP_ARRAY:
                     t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
                     break;
             }
 
@@ -757,8 +836,6 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
                 if (gl.features.multisample_texture) {
                     // multi-sample texture on GL 3.2 / GLES 3.1 and above
                     t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
-                    t->gl.targetIndex = (uint8_t)
-                            gl.getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
                 } else {
                     // Turn off multi-sampling for that texture. It's just not supported.
                 }
@@ -768,13 +845,6 @@ 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);
@@ -824,32 +894,24 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
     t->gl.internalFormat = getInternalFormat(format);
     assert_invariant(t->gl.internalFormat);
 
-    // We DO NOT update targetIndex at function exit to take advantage of the fact that
-    // getIndexForTextureTarget() is constexpr -- so all of this disappears at compile time.
     switch (target) {
         case SamplerType::SAMPLER_EXTERNAL:
             t->gl.target = GL_TEXTURE_EXTERNAL_OES;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_EXTERNAL_OES);
             break;
         case SamplerType::SAMPLER_2D:
             t->gl.target = GL_TEXTURE_2D;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D);
             break;
         case SamplerType::SAMPLER_3D:
             t->gl.target = GL_TEXTURE_3D;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_3D);
             break;
         case SamplerType::SAMPLER_2D_ARRAY:
             t->gl.target = GL_TEXTURE_2D_ARRAY;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
             break;
         case SamplerType::SAMPLER_CUBEMAP:
             t->gl.target = GL_TEXTURE_CUBE_MAP;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
             break;
         case SamplerType::SAMPLER_CUBEMAP_ARRAY:
             t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
             break;
     }
 
@@ -860,7 +922,6 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
         if (gl.features.multisample_texture) {
             // multi-sample texture on GL 3.2 / GLES 3.1 and above
             t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
         } else {
             // Turn off multi-sampling for that texture. It's just not supported.
         }
@@ -871,16 +932,18 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
 }
 
 void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer const* vb) {
-    // NOTE: this is called from draw() and must be as efficient as possible.
+    // NOTE: this is called often and must be as efficient as possible.
 
     auto& gl = mContext;
 
+#ifndef NDEBUG
     if (UTILS_LIKELY(gl.ext.OES_vertex_array_object)) {
         // The VAO for the given render primitive must already be bound.
         GLint vaoBinding;
         glGetIntegerv(GL_VERTEX_ARRAY_BINDING, &vaoBinding);
         assert_invariant(vaoBinding == (GLint)rp->gl.vao[gl.contextIndex]);
     }
+#endif
 
     if (UTILS_LIKELY(rp->gl.vertexBufferVersion == vb->bufferObjectsVersion &&
                      rp->gl.stateVersion == gl.state.age)) {
@@ -896,12 +959,12 @@ void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer
             // if a buffer is defined it must not be invalid.
             assert_invariant(vb->gl.buffers[bi]);
 
-            // if w're on ES2, the user shouldn't use FLAG_INTEGER_TARGET
+            // if we're on ES2, the user shouldn't use FLAG_INTEGER_TARGET
             assert_invariant(!(gl.isES2() && (attribute.flags & Attribute::FLAG_INTEGER_TARGET)));
 
             gl.bindBuffer(GL_ARRAY_BUFFER, vb->gl.buffers[bi]);
             GLuint const index = i;
-            GLint const size = getComponentCount(attribute.type);
+            GLint const size = (GLint)getComponentCount(attribute.type);
             GLenum const type = getComponentType(attribute.type);
             GLboolean const normalized = getNormalization(attribute.flags & Attribute::FLAG_NORMALIZED);
             GLsizei const stride = attribute.stride;
@@ -956,7 +1019,7 @@ void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer
 }
 
 void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
-        GLRenderTarget const* rt, GLenum attachment) noexcept {
+        GLRenderTarget const* rt, GLenum attachment, uint8_t layerCount) noexcept {
 
 #if !defined(NDEBUG)
     // Only used by assert_invariant() checks below
@@ -1094,9 +1157,20 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
             case GL_TEXTURE_2D_ARRAY:
             case GL_TEXTURE_CUBE_MAP_ARRAY:
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-                // GL_TEXTURE_2D_MULTISAMPLE_ARRAY is not supported in GLES
-                glFramebufferTextureLayer(GL_FRAMEBUFFER, attachment,
+
+                // TODO: support multiview for iOS and WebGL
+#if !defined(__EMSCRIPTEN__) && !defined(IOS)
+                if (layerCount > 1) {
+                    // if layerCount > 1, it means we use the multiview extension.
+                    glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, attachment,
+                        t->gl.id, 0, binfo.baseViewIndex, layerCount);
+                } else
+#endif // !defined(__EMSCRIPTEN__) && !defined(IOS)
+                {
+                    // GL_TEXTURE_2D_MULTISAMPLE_ARRAY is not supported in GLES
+                    glFramebufferTextureLayer(GL_FRAMEBUFFER, attachment,
                         t->gl.id, binfo.level, binfo.layer);
+                }
 #endif
                 break;
             default:
@@ -1257,11 +1331,9 @@ void OpenGLDriver::createDefaultRenderTargetR(
 
     construct<GLRenderTarget>(rth, 0, 0);  // FIXME: we don't know the width/height
 
-    uint32_t const framebuffer = mPlatform.createDefaultRenderTarget();
-
     GLRenderTarget* rt = handle_cast<GLRenderTarget*>(rth);
     rt->gl.isDefault = true;
-    rt->gl.fbo = framebuffer;
+    rt->gl.fbo = 0; // the actual id is resolved at binding time
     rt->gl.samples = 1;
     // FIXME: these flags should reflect the actual attachments present
     rt->targets = TargetBufferFlags::COLOR0 | TargetBufferFlags::DEPTH;
@@ -1272,6 +1344,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         uint32_t width,
         uint32_t height,
         uint8_t samples,
+        uint8_t layerCount,
         MRT color,
         TargetBufferInfo depth,
         TargetBufferInfo stencil) {
@@ -1338,7 +1411,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
             if (any(targets & getTargetBufferFlagsAt(i))) {
                 assert_invariant(color[i].handle);
                 rt->gl.color[i] = handle_cast<GLTexture*>(color[i].handle);
-                framebufferTexture(color[i], rt, GL_COLOR_ATTACHMENT0 + i);
+                framebufferTexture(color[i], rt, GL_COLOR_ATTACHMENT0 + i, layerCount);
                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
                 checkDimensions(rt->gl.color[i], color[i].level);
             }
@@ -1360,7 +1433,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         // either we supplied only the depth handle or both depth/stencil are identical and not null
         if (depth.handle && (stencil.handle == depth.handle || !stencil.handle)) {
             rt->gl.depth = handle_cast<GLTexture*>(depth.handle);
-            framebufferTexture(depth, rt, GL_DEPTH_STENCIL_ATTACHMENT);
+            framebufferTexture(depth, rt, GL_DEPTH_STENCIL_ATTACHMENT, layerCount);
             specialCased = true;
             checkDimensions(rt->gl.depth, depth.level);
         }
@@ -1371,13 +1444,13 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         if (any(targets & TargetBufferFlags::DEPTH)) {
             assert_invariant(depth.handle);
             rt->gl.depth = handle_cast<GLTexture*>(depth.handle);
-            framebufferTexture(depth, rt, GL_DEPTH_ATTACHMENT);
+            framebufferTexture(depth, rt, GL_DEPTH_ATTACHMENT, layerCount);
             checkDimensions(rt->gl.depth, depth.level);
         }
         if (any(targets & TargetBufferFlags::STENCIL)) {
             assert_invariant(stencil.handle);
             rt->gl.stencil = handle_cast<GLTexture*>(stencil.handle);
-            framebufferTexture(stencil, rt, GL_STENCIL_ATTACHMENT);
+            framebufferTexture(stencil, rt, GL_STENCIL_ATTACHMENT, layerCount);
             checkDimensions(rt->gl.stencil, stencil.level);
         }
     }
@@ -1419,6 +1492,13 @@ void OpenGLDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
     GLSwapChain* sc = handle_cast<GLSwapChain*>(sch);
     sc->swapChain = mPlatform.createSwapChain(nativeWindow, flags);
 
+#if !defined(__EMSCRIPTEN__)
+    // note: in practice this should never happen on Android
+    ASSERT_POSTCONDITION(sc->swapChain,
+            "createSwapChain(%p, 0x%lx) failed. See logs for details.",
+            nativeWindow, flags);
+#endif
+
     // See if we need the emulated rec709 output conversion
     if (UTILS_UNLIKELY(mContext.isES2())) {
         sc->rec709 = (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE &&
@@ -1433,6 +1513,13 @@ void OpenGLDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch,
     GLSwapChain* sc = handle_cast<GLSwapChain*>(sch);
     sc->swapChain = mPlatform.createSwapChain(width, height, flags);
 
+#if !defined(__EMSCRIPTEN__)
+    // note: in practice this should never happen on Android
+    ASSERT_POSTCONDITION(sc->swapChain,
+            "createSwapChainHeadless(%u, %u, 0x%lx) failed. See logs for details.",
+            width, height, flags);
+#endif
+
     // See if we need the emulated rec709 output conversion
     if (UTILS_UNLIKELY(mContext.isES2())) {
         sc->rec709 = (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE &&
@@ -1575,11 +1662,11 @@ void OpenGLDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
         GLRenderTarget* rt = handle_cast<GLRenderTarget*>(rth);
         if (rt->gl.fbo) {
             // first unbind this framebuffer if needed
-            gl.bindFramebuffer(GL_FRAMEBUFFER, 0);
+            gl.unbindFramebuffer(GL_FRAMEBUFFER);
         }
         if (rt->gl.fbo_read) {
             // first unbind this framebuffer if needed
-            gl.bindFramebuffer(GL_FRAMEBUFFER, 0);
+            gl.unbindFramebuffer(GL_FRAMEBUFFER);
         }
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
@@ -1726,7 +1813,6 @@ void OpenGLDriver::updateStreams(DriverApi* driver) {
                         // the target and id can be reset each time
                         t->gl.target = t->externalTexture->target;
                         t->gl.id = t->externalTexture->id;
-                        t->gl.targetIndex = (uint8_t)OpenGLContext::getIndexForTextureTarget(t->gl.target);
                         bindTexture(OpenGLContext::DUMMY_TEXTURE_BINDING, t);
                     }
                 }
@@ -1988,6 +2074,10 @@ bool OpenGLDriver::isDepthStencilResolveSupported() {
     return true;
 }
 
+bool OpenGLDriver::isDepthStencilBlitSupported(TextureFormat format) {
+    return true;
+}
+
 bool OpenGLDriver::isProtectedTexturesSupported() {
     return getContext().ext.EXT_protected_textures;
 }
@@ -2004,8 +2094,6 @@ 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;
     }
@@ -2284,7 +2372,7 @@ void OpenGLDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
 #endif
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
             if (UTILS_LIKELY(!es2)) {
-                samplerId = getSampler(params);
+                samplerId = mContext.getSampler(params);
             } else
 #endif
             {
@@ -2413,7 +2501,9 @@ void OpenGLDriver::setTextureData(GLTexture* t, uint32_t level,
     size_t const stride = p.stride ? p.stride : width;
     size_t const bpp = PBD::computeDataSize(p.format, p.type, 1, 1, 1);
     size_t const bpr = PBD::computeDataSize(p.format, p.type, stride, 1, p.alignment);
-    void const* const buffer = static_cast<char const*>(p.buffer) + p.left * bpp + bpr * p.top;
+    size_t const bpl = bpr * height; // TODO: PBD should have a "layer stride"
+    void const* const buffer = static_cast<char const*>(p.buffer)
+            + bpp* p.left + bpr * p.top + bpl * 0; // TODO: PBD should have a p.depth
 
     switch (t->target) {
         case SamplerType::SAMPLER_EXTERNAL:
@@ -2478,7 +2568,7 @@ void OpenGLDriver::setTextureData(GLTexture* t, uint32_t level,
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     if (!gl.isES2()) {
-        // update the base/max LOD, so we don't access undefined LOD. this allows the app to
+        // Update the base/max LOD, so we don't access undefined LOD. this allows the app to
         // specify levels as they become available.
         if (int8_t(level) < t->gl.baseLevel) {
             t->gl.baseLevel = int8_t(level);
@@ -2579,7 +2669,7 @@ void OpenGLDriver::setCompressedTextureData(GLTexture* t, uint32_t level,
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     if (!gl.isES2()) {
-        // update the base/max LOD, so we don't access undefined LOD. this allows the app to
+        // Update the base/max LOD, so we don't access undefined LOD. this allows the app to
         // specify levels as they become available.
         if (int8_t(level) < t->gl.baseLevel) {
             t->gl.baseLevel = int8_t(level);
@@ -2612,7 +2702,6 @@ void OpenGLDriver::setExternalImage(Handle<HwTexture> th, void* image) {
         // the target and id can be reset each time
         t->gl.target = t->externalTexture->target;
         t->gl.id = t->externalTexture->id;
-        t->gl.targetIndex = (uint8_t)OpenGLContext::getIndexForTextureTarget(t->gl.target);
         bindTexture(OpenGLContext::DUMMY_TEXTURE_BINDING, t);
     }
 }
@@ -2732,7 +2821,7 @@ TimerQueryResult OpenGLDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint
     return TimerQueryFactoryInterface::getTimerQueryValue(tq, elapsedTime);
 }
 
-void OpenGLDriver::compilePrograms(CompilerPriorityQueue priority,
+void OpenGLDriver::compilePrograms(CompilerPriorityQueue,
         CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
     if (callback) {
         getShaderCompilerService().notifyWhenAllProgramsAreReady(handler, callback, user);
@@ -2760,13 +2849,13 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
     const TargetBufferFlags clearFlags = params.flags.clear & rt->targets;
     TargetBufferFlags discardFlags = params.flags.discardStart & rt->targets;
 
-    gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
+    GLuint const fbo = gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
     CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)
 
     if (gl.ext.EXT_discard_framebuffer
             && !gl.bugs.disable_invalidate_framebuffer) {
         AttachmentArray attachments; // NOLINT
-        GLsizei const attachmentCount = getAttachments(attachments, rt, discardFlags);
+        GLsizei const attachmentCount = getAttachments(attachments, discardFlags, !fbo);
         if (attachmentCount) {
             gl.procs.invalidateFramebuffer(GL_FRAMEBUFFER, attachmentCount, attachments.data());
         }
@@ -2854,9 +2943,9 @@ void OpenGLDriver::endRenderPass(int) {
         }
         if (!gl.bugs.disable_invalidate_framebuffer) {
             // we wouldn't have to bind the framebuffer if we had glInvalidateNamedFramebuffer()
-            gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
+            GLuint const fbo = gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
             AttachmentArray attachments; // NOLINT
-            GLsizei const attachmentCount = getAttachments(attachments, rt, effectiveDiscardFlags);
+            GLsizei const attachmentCount = getAttachments(attachments, effectiveDiscardFlags, !fbo);
             if (attachmentCount) {
                 gl.procs.invalidateFramebuffer(GL_FRAMEBUFFER, attachmentCount, attachments.data());
             }
@@ -2918,50 +3007,48 @@ void OpenGLDriver::resolvePass(ResolveAction action, GLRenderTarget const* rt,
 }
 
 GLsizei OpenGLDriver::getAttachments(AttachmentArray& attachments,
-        GLRenderTarget const* rt, TargetBufferFlags buffers) noexcept {
-    assert_invariant(buffers <= rt->targets);
-
+        TargetBufferFlags buffers, bool isDefaultFramebuffer) noexcept {
     GLsizei attachmentCount = 0;
     // the default framebuffer uses different constants!!!
-    const bool defaultFramebuffer = (rt->gl.fbo == 0);
+
     if (any(buffers & TargetBufferFlags::COLOR0)) {
-        attachments[attachmentCount++] = defaultFramebuffer ? GL_COLOR : GL_COLOR_ATTACHMENT0;
+        attachments[attachmentCount++] = isDefaultFramebuffer ? GL_COLOR : GL_COLOR_ATTACHMENT0;
     }
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     if (any(buffers & TargetBufferFlags::COLOR1)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT1;
     }
     if (any(buffers & TargetBufferFlags::COLOR2)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT2;
     }
     if (any(buffers & TargetBufferFlags::COLOR3)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT3;
     }
     if (any(buffers & TargetBufferFlags::COLOR4)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT4;
     }
     if (any(buffers & TargetBufferFlags::COLOR5)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT5;
     }
     if (any(buffers & TargetBufferFlags::COLOR6)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT6;
     }
     if (any(buffers & TargetBufferFlags::COLOR7)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT7;
     }
 #endif
     if (any(buffers & TargetBufferFlags::DEPTH)) {
-        attachments[attachmentCount++] = defaultFramebuffer ? GL_DEPTH : GL_DEPTH_ATTACHMENT;
+        attachments[attachmentCount++] = isDefaultFramebuffer ? GL_DEPTH : GL_DEPTH_ATTACHMENT;
     }
     if (any(buffers & TargetBufferFlags::STENCIL)) {
-        attachments[attachmentCount++] = defaultFramebuffer ? GL_STENCIL : GL_STENCIL_ATTACHMENT;
+        attachments[attachmentCount++] = isDefaultFramebuffer ? GL_STENCIL : GL_STENCIL_ATTACHMENT;
     }
     return attachmentCount;
 }
@@ -3009,11 +3096,10 @@ void OpenGLDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t ind
     assert_invariant(offset + size <= ub->byteCount);
 
     if (UTILS_UNLIKELY(ub->bindingType == BufferObjectBinding::UNIFORM && gl.isES2())) {
-        mUniformBindings[index] = {
+        gl.setEs2UniformBinding(index,
                 ub->gl.id,
                 static_cast<uint8_t const*>(ub->gl.buffer) + offset,
-                ub->age,
-        };
+                ub->age);
     } else {
         GLenum const target = GLUtils::getBufferBindingType(bindingType);
 
@@ -3031,7 +3117,7 @@ void OpenGLDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index)
     auto& gl = mContext;
 
     if (UTILS_UNLIKELY(bindingType == BufferObjectBinding::UNIFORM && gl.isES2())) {
-        mUniformBindings[index] = {};
+        gl.setEs2UniformBinding(index, 0, nullptr, 0);
         return;
     }
 
@@ -3048,36 +3134,6 @@ 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
@@ -3377,17 +3433,21 @@ void OpenGLDriver::executeGpuCommandsCompleteOps() noexcept {
 
 void OpenGLDriver::tick(int) {
     DEBUG_MARKER()
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     executeGpuCommandsCompleteOps();
+#endif
     executeEveryNowAndThenOps();
     getShaderCompilerService().tick();
 }
 
 void OpenGLDriver::beginFrame(
         UTILS_UNUSED int64_t monotonic_clock_ns,
+        UTILS_UNUSED int64_t refreshIntervalNs,
         UTILS_UNUSED uint32_t frameId) {
     DEBUG_MARKER()
     auto& gl = mContext;
     insertEventMarker("beginFrame");
+    mPlatform.beginFrame(monotonic_clock_ns, refreshIntervalNs, frameId);
     if (UTILS_UNLIKELY(!mTexturesWithStreamsAttached.empty())) {
         OpenGLPlatform& platform = mPlatform;
         for (GLTexture const* t : mTexturesWithStreamsAttached) {
@@ -3396,7 +3456,7 @@ void OpenGLDriver::beginFrame(
                 assert_invariant(t->hwStream->stream);
                 platform.updateTexImage(t->hwStream->stream,
                         &static_cast<GLStream*>(t->hwStream)->user_thread.timestamp); // NOLINT(cppcoreguidelines-pro-type-static-cast-downcast)
-                // NOTE: We assume that updateTexImage() binds the texture on our behalf
+                // NOTE: We assume that OpenGLPlatform::updateTexImage() binds the texture on our behalf
                 gl.updateTexImage(GL_TEXTURE_EXTERNAL_OES, t->gl.id);
             }
         }
@@ -3404,12 +3464,12 @@ void OpenGLDriver::beginFrame(
 }
 
 void OpenGLDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        FrameScheduledCallback callback, void* user) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
     DEBUG_MARKER()
 }
 
 void OpenGLDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
+        CallbackHandler* handler, utils::Invocable<void(void)>&& callback) {
     DEBUG_MARKER()
 }
 
@@ -3435,6 +3495,7 @@ void OpenGLDriver::endFrame(UTILS_UNUSED uint32_t frameId) {
 #endif
     //SYSTRACE_NAME("glFinish");
     //glFinish();
+    mPlatform.endFrame(frameId);
     insertEventMarker("endFrame");
 }
 
@@ -3646,7 +3707,7 @@ void OpenGLDriver::blit(
         case SamplerType::SAMPLER_EXTERNAL:
             break;
     }
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_READ_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_DRAW_FRAMEBUFFER)
 
     gl.bindFramebuffer(GL_READ_FRAMEBUFFER, fbo[1]);
     switch (s->target) {
@@ -3672,7 +3733,7 @@ void OpenGLDriver::blit(
         case SamplerType::SAMPLER_EXTERNAL:
             break;
     }
-    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_DRAW_FRAMEBUFFER)
+    CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_READ_FRAMEBUFFER)
 
     gl.disable(GL_SCISSOR_TEST);
     glBlitFramebuffer(
@@ -3681,8 +3742,8 @@ void OpenGLDriver::blit(
             mask, GL_NEAREST);
     CHECK_GL_ERROR(utils::slog.e)
 
-    gl.bindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
-    gl.bindFramebuffer(GL_READ_FRAMEBUFFER, 0);
+    gl.unbindFramebuffer(GL_DRAW_FRAMEBUFFER);
+    gl.unbindFramebuffer(GL_READ_FRAMEBUFFER);
     glDeleteFramebuffers(2, fbo);
 
     if (any(d->usage & TextureUsage::SAMPLEABLE)) {
@@ -3786,7 +3847,7 @@ void OpenGLDriver::updateTextureLodRange(GLTexture* texture, int8_t targetLevel)
 #endif
 }
 
-void OpenGLDriver::bindPipeline(PipelineState state) {
+void OpenGLDriver::bindPipeline(PipelineState const& state) {
     DEBUG_MARKER()
     auto& gl = mContext;
     setRasterState(state.rasterState);
@@ -3794,6 +3855,7 @@ void OpenGLDriver::bindPipeline(PipelineState state) {
     gl.polygonOffset(state.polygonOffset.slope, state.polygonOffset.constant);
     OpenGLProgram* const p = handle_cast<OpenGLProgram*>(state.program);
     mValidProgram = useProgram(p);
+    (*mCurrentPushConstants) = p->getPushConstants();
 }
 
 void OpenGLDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
@@ -3823,20 +3885,35 @@ void OpenGLDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
         return;
     }
 
-    if (UTILS_LIKELY(instanceCount <= 1)) {
-        glDrawElements(GLenum(rp->type), (GLsizei)indexCount, rp->gl.getIndicesType(),
-                reinterpret_cast<const void*>(indexOffset * rp->gl.indicesSize));
-    } else {
-        assert_invariant(!mContext.isES2());
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-        glDrawElementsInstanced(GLenum(rp->type), (GLsizei)indexCount,
-                rp->gl.getIndicesType(),
-                reinterpret_cast<const void*>(indexOffset * rp->gl.indicesSize),
-                (GLsizei)instanceCount);
+    assert_invariant(!mContext.isES2());
+    glDrawElementsInstanced(GLenum(rp->type), (GLsizei)indexCount,
+            rp->gl.getIndicesType(),
+            reinterpret_cast<const void*>(indexOffset * rp->gl.indicesSize),
+            (GLsizei)instanceCount);
 #endif
+
+#if FILAMENT_ENABLE_MATDBG
+    CHECK_GL_ERROR_NON_FATAL(utils::slog.e)
+#else
+    CHECK_GL_ERROR(utils::slog.e)
+#endif
+}
+
+void OpenGLDriver::draw2GLES2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+    GLRenderPrimitive const* const rp = mBoundRenderPrimitive;
+    if (UTILS_UNLIKELY(!rp || !mValidProgram)) {
+        return;
     }
 
-#ifdef FILAMENT_ENABLE_MATDBG
+    assert_invariant(mContext.isES2());
+    assert_invariant(instanceCount == 1);
+
+    glDrawElements(GLenum(rp->type), (GLsizei)indexCount, rp->gl.getIndicesType(),
+            reinterpret_cast<const void*>(indexOffset * rp->gl.indicesSize));
+
+
+#if FILAMENT_ENABLE_MATDBG
     CHECK_GL_ERROR_NON_FATAL(utils::slog.e)
 #else
     CHECK_GL_ERROR(utils::slog.e)
@@ -3855,7 +3932,11 @@ void OpenGLDriver::draw(PipelineState state, Handle<HwRenderPrimitive> rph,
     state.vertexBufferInfo = rp->vbih;
     bindPipeline(state);
     bindRenderPrimitive(rph);
-    draw2(indexOffset, indexCount, instanceCount);
+    if (UTILS_UNLIKELY(mContext.isES2())) {
+        draw2GLES2(indexOffset, indexCount, instanceCount);
+    } else {
+        draw2(indexOffset, indexCount, instanceCount);
+    }
 }
 
 void OpenGLDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGroupCount) {
@@ -3882,7 +3963,7 @@ void OpenGLDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGr
     glDispatchCompute(workGroupCount.x, workGroupCount.y, workGroupCount.z);
 #endif // BACKEND_OPENGL_LEVEL_GLES31
 
-#ifdef FILAMENT_ENABLE_MATDBG
+#if FILAMENT_ENABLE_MATDBG
     CHECK_GL_ERROR_NON_FATAL(utils::slog.e)
 #else
     CHECK_GL_ERROR(utils::slog.e)

filament/backend/src/opengl/OpenGLDriver.h

@@ -18,31 +18,43 @@
 #define TNT_FILAMENT_BACKEND_OPENGL_OPENGLDRIVER_H
 
 #include "DriverBase.h"
-#include "GLUtils.h"
 #include "OpenGLContext.h"
 #include "OpenGLTimerQuery.h"
+#include "GLBufferObject.h"
+#include "GLTexture.h"
 #include "ShaderCompilerService.h"
 
-#include "private/backend/Driver.h"
-#include "private/backend/HandleAllocator.h"
-
 #include <backend/platforms/OpenGLPlatform.h>
 
 #include <backend/AcquiredImage.h>
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+#include <backend/Platform.h>
 #include <backend/Program.h>
 #include <backend/TargetBufferInfo.h>
 
+#include "private/backend/Driver.h"
+#include "private/backend/HandleAllocator.h"
+
+#include <utils/FixedCapacityVector.h>
 #include <utils/compiler.h>
-#include <utils/Allocator.h>
+#include <utils/debug.h>
 
 #include <math/vec4.h>
 
 #include <tsl/robin_map.h>
 
-#include <atomic>
+#include <array>
+#include <condition_variable>
+#include <functional>
 #include <memory>
-#include <set>
+#include <mutex>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+#include <vector>
 
+#include <stddef.h>
 #include <stdint.h>
 
 #ifndef FILAMENT_OPENGL_HANDLE_ARENA_SIZE_IN_MB
@@ -54,17 +66,19 @@ namespace filament::backend {
 class OpenGLPlatform;
 class PixelBufferDescriptor;
 struct TargetBufferInfo;
-
 class OpenGLProgram;
 class TimerQueryFactoryInterface;
+struct PushConstantBundle;
 
 class OpenGLDriver final : public DriverBase {
-    inline explicit OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept;
+    inline explicit OpenGLDriver(OpenGLPlatform* platform,
+            const Platform::DriverConfig& driverConfig) noexcept;
     ~OpenGLDriver() noexcept final;
     Dispatcher getDispatcher() const noexcept final;
 
 public:
-    static Driver* create(OpenGLPlatform* platform, void* sharedGLContext, const Platform::DriverConfig& driverConfig) noexcept;
+    static Driver* create(OpenGLPlatform* platform, void* sharedGLContext,
+            const Platform::DriverConfig& driverConfig) noexcept;
 
     class DebugMarker {
         OpenGLDriver& driver;
@@ -80,25 +94,6 @@ 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,
@@ -128,7 +123,6 @@ public:
         } gl;
     };
 
-    struct GLTexture;
     struct GLSamplerGroup : public HwSamplerGroup {
         using HwSamplerGroup::HwSamplerGroup;
         struct Entry {
@@ -145,27 +139,9 @@ public:
         Handle<HwVertexBufferInfo> vbih;
     };
 
-    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 GLBufferObject = filament::backend::GLBufferObject;
 
-            // 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 GLTexture = filament::backend::GLTexture;
 
     using GLTimerQuery = filament::backend::GLTimerQuery;
 
@@ -256,13 +232,13 @@ private:
     HandleAllocatorGL mHandleAllocator;
 
     template<typename D, typename ... ARGS>
-    Handle<D> initHandle(ARGS&& ... args) noexcept {
+    Handle<D> initHandle(ARGS&& ... args) {
         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) noexcept {
+    construct(Handle<B> const& handle, ARGS&& ... args) {
         return mHandleAllocator.destroyAndConstruct<D, B>(handle, std::forward<ARGS>(args) ...);
     }
 
@@ -276,7 +252,7 @@ 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) noexcept {
+    handle_cast(Handle<B>& handle) {
         return mHandleAllocator.handle_cast<Dp, B>(handle);
     }
 
@@ -284,7 +260,7 @@ 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) noexcept {
+    handle_cast(Handle<B> const& handle) {
         return mHandleAllocator.handle_cast<Dp, B>(handle);
     }
 
@@ -302,7 +278,7 @@ private:
     void updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer const* vb);
 
     void framebufferTexture(TargetBufferInfo const& binfo,
-            GLRenderTarget const* rt, GLenum attachment) noexcept;
+            GLRenderTarget const* rt, GLenum attachment, uint8_t layerCount) noexcept;
 
     void setRasterState(RasterState rs) noexcept;
 
@@ -336,29 +312,13 @@ 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;
     }
 
     using AttachmentArray = std::array<GLenum, MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT + 2>;
-    static GLsizei getAttachments(AttachmentArray& attachments,
-            GLRenderTarget const* rt, TargetBufferFlags buffers) noexcept;
+    static GLsizei getAttachments(AttachmentArray& attachments, TargetBufferFlags buffers,
+            bool isDefaultFramebuffer) noexcept;
 
     // state required to represent the current render pass
     Handle<HwRenderTarget> mRenderPassTarget;
@@ -376,16 +336,14 @@ private:
 
     void setScissor(Viewport const& scissor) noexcept;
 
+    void draw2GLES2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount);
+
     // 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;
 
@@ -419,6 +377,8 @@ private:
     // for ES2 sRGB support
     GLSwapChain* mCurrentDrawSwapChain = nullptr;
     bool mRec709OutputColorspace = false;
+
+    PushConstantBundle* mCurrentPushConstants = nullptr;
 };
 
 // ------------------------------------------------------------------------------------------------

filament/backend/src/opengl/OpenGLPlatform.cpp

@@ -40,7 +40,7 @@ OpenGLPlatform::~OpenGLPlatform() noexcept = default;
 
 void OpenGLPlatform::makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain,
         utils::Invocable<void()>, utils::Invocable<void(size_t)>) noexcept {
-    makeCurrent(drawSwapChain, readSwapChain);
+    makeCurrent(getCurrentContextType(), drawSwapChain, readSwapChain);
 }
 
 bool OpenGLPlatform::isProtectedContextSupported() const noexcept {
@@ -51,10 +51,24 @@ bool OpenGLPlatform::isSRGBSwapChainSupported() const noexcept {
     return false;
 }
 
-uint32_t OpenGLPlatform::createDefaultRenderTarget() noexcept {
+uint32_t OpenGLPlatform::getDefaultFramebufferObject() noexcept {
     return 0;
 }
 
+void OpenGLPlatform::beginFrame(int64_t monotonic_clock_ns, int64_t refreshIntervalNs,
+        uint32_t frameId) noexcept {
+}
+
+void OpenGLPlatform::endFrame(uint32_t frameId) noexcept {
+}
+
+void OpenGLPlatform::preCommit() noexcept {
+}
+
+OpenGLPlatform::ContextType OpenGLPlatform::getCurrentContextType() const noexcept {
+    return ContextType::UNPROTECTED;
+}
+
 void OpenGLPlatform::setPresentationTime(
         UTILS_UNUSED int64_t presentationTimeInNanosecond) noexcept {
 }
@@ -125,10 +139,14 @@ AcquiredImage OpenGLPlatform::transformAcquiredImage(AcquiredImage source) noexc
     return source;
 }
 
-TargetBufferFlags OpenGLPlatform::getPreservedFlags(UTILS_UNUSED SwapChain* swapChain) noexcept {
+TargetBufferFlags OpenGLPlatform::getPreservedFlags(UTILS_UNUSED SwapChain*) noexcept {
     return TargetBufferFlags::NONE;
 }
 
+bool OpenGLPlatform::isSwapChainProtected(UTILS_UNUSED SwapChain*) noexcept {
+    return false;
+}
+
 bool OpenGLPlatform::isExtraContextSupported() const noexcept {
     return false;
 }

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -16,16 +16,25 @@
 
 #include "OpenGLProgram.h"
 
-#include "BlobCacheKey.h"
+#include "GLUtils.h"
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"
 
+#include <backend/Program.h>
+
+#include <private/backend/BackendUtils.h>
+
 #include <utils/debug.h>
 #include <utils/compiler.h>
 #include <utils/Log.h>
 #include <utils/Systrace.h>
 
-#include <private/backend/BackendUtils.h>
+#include <array>
+#include <string_view>
+#include <utility>
+#include <new>
+
+#include <stddef.h>
 
 namespace filament::backend {
 
@@ -37,6 +46,8 @@ struct OpenGLProgram::LazyInitializationData {
     Program::UniformBlockInfo uniformBlockInfo;
     Program::SamplerGroupInfo samplerGroupInfo;
     std::array<Program::UniformInfo, Program::UNIFORM_BINDING_COUNT> bindingUniformInfo;
+    utils::FixedCapacityVector<Program::PushConstant> vertexPushConstants;
+    utils::FixedCapacityVector<Program::PushConstant> fragmentPushConstants;
 };
 
 
@@ -44,7 +55,6 @@ OpenGLProgram::OpenGLProgram() noexcept = default;
 
 OpenGLProgram::OpenGLProgram(OpenGLDriver& gld, Program&& program) noexcept
         : HwProgram(std::move(program.getName())) {
-
     auto* const lazyInitializationData = new(std::nothrow) LazyInitializationData();
     lazyInitializationData->samplerGroupInfo = std::move(program.getSamplerGroupInfo());
     if (UTILS_UNLIKELY(gld.getContext().isES2())) {
@@ -52,6 +62,8 @@ OpenGLProgram::OpenGLProgram(OpenGLDriver& gld, Program&& program) noexcept
     } else {
         lazyInitializationData->uniformBlockInfo = std::move(program.getUniformBlockBindings());
     }
+    lazyInitializationData->vertexPushConstants = std::move(program.getPushConstants(ShaderStage::VERTEX));
+    lazyInitializationData->fragmentPushConstants = std::move(program.getPushConstants(ShaderStage::FRAGMENT));
 
     ShaderCompilerService& compiler = gld.getShaderCompilerService();
     mToken = compiler.createProgram(name, std::move(program));
@@ -194,6 +206,21 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
         }
     }
     mUsedBindingsCount = usedBindingCount;
+
+    auto& vertexConstants = lazyInitializationData.vertexPushConstants;
+    auto& fragmentConstants = lazyInitializationData.fragmentPushConstants;
+
+    size_t const totalConstantCount = vertexConstants.size() + fragmentConstants.size();
+    if (totalConstantCount > 0) {
+        mPushConstants.reserve(totalConstantCount);
+        mPushConstantFragmentStageOffset = vertexConstants.size();
+        auto const transformAndAdd = [&](Program::PushConstant const& constant) {
+            GLint const loc = glGetUniformLocation(program, constant.name.c_str());
+            mPushConstants.push_back({loc, constant.type});
+        };
+        std::for_each(vertexConstants.cbegin(), vertexConstants.cend(), transformAndAdd);
+        std::for_each(fragmentConstants.cbegin(), fragmentConstants.cend(), transformAndAdd);
+    }
 }
 
 void OpenGLProgram::updateSamplers(OpenGLDriver* const gld) const noexcept {

filament/backend/src/opengl/OpenGLProgram.h

@@ -27,6 +27,10 @@
 
 #include <utils/compiler.h>
 #include <utils/FixedCapacityVector.h>
+#include <utils/Slice.h>
+
+#include <array>
+#include <limits>
 
 #include <stddef.h>
 #include <stdint.h>
@@ -35,6 +39,11 @@ namespace filament::backend {
 
 class OpenGLDriver;
 
+struct PushConstantBundle {
+    utils::Slice<std::pair<GLint, ConstantType>> vertexConstants;
+    utils::Slice<std::pair<GLint, ConstantType>> fragmentConstants;
+};
+
 class OpenGLProgram : public HwProgram {
 public:
 
@@ -67,14 +76,22 @@ 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
+
+    PushConstantBundle getPushConstants() {
+        auto fragBegin = mPushConstants.begin() + mPushConstantFragmentStageOffset;
+        return {
+            .vertexConstants = utils::Slice(mPushConstants.begin(), fragBegin),
+            .fragmentConstants = utils::Slice(fragBegin, mPushConstants.end()),
+        };
+    }
+
 private:
     // keep these away from of other class attributes
     struct LazyInitializationData;
@@ -86,14 +103,20 @@ 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[2] = {};               // 2 byte
+
+    // Push constant array offset for fragment stage constants.
+    uint8_t mPushConstantFragmentStageOffset = 0u;      // 1 byte
+
     // only needed for ES2
+    GLint mRec709Location = -1;                         // 4 bytes
+
     using LocationInfo = utils::FixedCapacityVector<GLint>;
     struct UniformsRecord {
         Program::UniformInfo uniforms;
@@ -101,12 +124,15 @@ private:
         mutable GLuint id = 0;
         mutable uint16_t age = std::numeric_limits<uint16_t>::max();
     };
-    UniformsRecord const* mUniformsRecords = nullptr;
-    GLint mRec709Location = -1;
+    UniformsRecord const* mUniformsRecords = nullptr;               // 8 bytes
+
+    // Note that this can be replaced with a raw pointer and an uint8_t (for size) to reduce the
+    // size of the container to 9 bytes if there is a need in the future.
+    utils::FixedCapacityVector<std::pair<GLint, ConstantType>> mPushConstants;// 16 bytes
 };
 
 // if OpenGLProgram is larger tha 64 bytes, it'll fall in a larger Handle bucket.
-static_assert(sizeof(OpenGLProgram) <= 64);
+static_assert(sizeof(OpenGLProgram) <= 64); // currently 64 bytes
 
 } // namespace filament::backend
 

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -30,6 +30,7 @@
 #include <utils/Log.h>
 #include <utils/Systrace.h>
 
+#include <cctype>
 #include <chrono>
 #include <string>
 #include <string_view>
@@ -264,6 +265,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
                     compileShaders(gl,
                             std::move(program.getShadersSource()),
                             program.getSpecializationConstants(),
+                            program.isMultiview(),
                             shaders,
                             token->shaderSourceCode);
 
@@ -299,6 +301,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
         compileShaders(gl,
                 std::move(program.getShadersSource()),
                 program.getSpecializationConstants(),
+                program.isMultiview(),
                 token->gl.shaders,
                 token->shaderSourceCode);
 
@@ -501,6 +504,7 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
 void ShaderCompilerService::compileShaders(OpenGLContext& context,
         Program::ShaderSource shadersSource,
         utils::FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
+        bool multiview,
         std::array<GLuint, Program::SHADER_TYPE_COUNT>& outShaders,
         UTILS_UNUSED_IN_RELEASE std::array<CString, Program::SHADER_TYPE_COUNT>& outShaderSourceCode) noexcept {
 
@@ -514,8 +518,16 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
     };
 
     std::string specializationConstantString;
+    int32_t numViews = 2;
     for (auto const& sc : specializationConstants) {
         appendSpecConstantString(specializationConstantString, sc);
+        if (sc.id == 8) {
+            // This constant must match
+            // ReservedSpecializationConstants::CONFIG_STEREO_EYE_COUNT
+            // which we can't use here because it's defined in EngineEnums.h.
+            // (we're breaking layering here, but it's for the good cause).
+            numViews = std::get<int32_t>(sc.value);
+        }
     }
     if (!specializationConstantString.empty()) {
         specializationConstantString += '\n';
@@ -544,17 +556,23 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 
         if (UTILS_LIKELY(!shadersSource[i].empty())) {
             Program::ShaderBlob& shader = shadersSource[i];
+            char* shader_src = reinterpret_cast<char*>(shader.data());
+            size_t shader_len = shader.size();
 
             // remove GOOGLE_cpp_style_line_directive
-            std::string_view const source = process_GOOGLE_cpp_style_line_directive(context,
-                    reinterpret_cast<char*>(shader.data()), shader.size());
+            process_GOOGLE_cpp_style_line_directive(context, shader_src, shader_len);
+
+            // replace the value of layout(num_views = X) for multiview extension
+            if (multiview && stage == ShaderStage::VERTEX) {
+                process_OVR_multiview2(context, numViews, shader_src, shader_len);
+            }
 
             // add support for ARB_shading_language_packing if needed
             auto const packingFunctions = process_ARB_shading_language_packing(context);
 
             // split shader source, so we can insert the specialization constants and the packing
             // functions
-            auto const [prolog, body] = splitShaderSource(source);
+            auto const [prolog, body] = splitShaderSource({ shader_src, shader_len });
 
             const std::array<const char*, 4> sources = {
                     prolog.data(),
@@ -577,7 +595,7 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 #ifndef NDEBUG
             // for debugging we return the original shader source (without the modifications we
             // made here), otherwise the line numbers wouldn't match.
-            outShaderSourceCode[i] = { source.data(), source.length() };
+            outShaderSourceCode[i] = { shader_src, shader_len };
 #endif
 
             outShaders[i] = shaderId;
@@ -586,15 +604,59 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 }
 
 // If usages of the Google-style line directive are present, remove them, as some
-// drivers don't allow the quotation marks. This happens in-place.
-std::string_view ShaderCompilerService::process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
+// drivers don't allow the quotation marks. This source modification happens in-place.
+void ShaderCompilerService::process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
         char* source, size_t len) noexcept {
     if (!context.ext.GOOGLE_cpp_style_line_directive) {
         if (UTILS_UNLIKELY(requestsGoogleLineDirectivesExtension({ source, len }))) {
             removeGoogleLineDirectives(source, len); // length is unaffected
         }
     }
-    return { source, len };
+}
+
+// Look up the `source` to replace the number of eyes for multiview with the given number. This is
+// necessary for OpenGL because OpenGL relies on the number specified in shader files to determine
+// the number of views, which is assumed as a single digit, for multiview.
+// This source modification happens in-place.
+void ShaderCompilerService::process_OVR_multiview2(OpenGLContext& context,
+        int32_t eyeCount, char* source, size_t len) noexcept {
+    // We don't use regular expression in favor of performance.
+    if (context.ext.OVR_multiview2) {
+        const std::string_view shader{ source, len };
+        const std::string_view layout = "layout";
+        const std::string_view num_views = "num_views";
+        size_t found = 0;
+        while (true) {
+            found = shader.find(layout, found);
+            if (found == std::string_view::npos) {
+                break;
+            }
+            found = shader.find_first_not_of(' ', found + layout.size());
+            if (found == std::string_view::npos || shader[found] != '(') {
+                continue;
+            }
+            found = shader.find_first_not_of(' ', found + 1);
+            if (found == std::string_view::npos) {
+                continue;
+            }
+            if (shader.compare(found, num_views.size(), num_views) != 0) {
+                continue;
+            }
+            found = shader.find_first_not_of(' ', found + num_views.size());
+            if (found == std::string_view::npos || shader[found] != '=') {
+                continue;
+            }
+            found = shader.find_first_not_of(' ', found + 1);
+            if (found == std::string_view::npos) {
+                continue;
+            }
+            // We assume the value should be one-digit number.
+            assert_invariant(eyeCount < 10);
+            assert_invariant(!::isdigit(source[found + 1]));
+            source[found] = '0' + eyeCount;
+            break;
+        }
+    }
 }
 
 // Tragically, OpenGL 4.1 doesn't support unpackHalf2x16 (appeared in 4.2) and

filament/backend/src/opengl/ShaderCompilerService.h

@@ -134,10 +134,14 @@ private:
             OpenGLContext& context,
             Program::ShaderSource shadersSource,
             utils::FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
+            bool multiview,
             std::array<GLuint, Program::SHADER_TYPE_COUNT>& outShaders,
             std::array<utils::CString, Program::SHADER_TYPE_COUNT>& outShaderSourceCode) noexcept;
 
-    static std::string_view process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
+    static void process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
+            char* source, size_t len) noexcept;
+
+    static void process_OVR_multiview2(OpenGLContext& context, int32_t eyeCount,
             char* source, size_t len) noexcept;
 
     static std::string_view process_ARB_shading_language_packing(OpenGLContext& context) noexcept;

filament/backend/src/opengl/gl_headers.cpp

@@ -67,12 +67,14 @@ PFNGLDISCARDFRAMEBUFFEREXTPROC glDiscardFramebufferEXT;
 #ifdef GL_KHR_parallel_shader_compile
 PFNGLMAXSHADERCOMPILERTHREADSKHRPROC glMaxShaderCompilerThreadsKHR;
 #endif
+#ifdef GL_OVR_multiview
+PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
+#endif
 
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
 // On Android, If we want to support a build system less than ANDROID_API 21, we need to
 // use getProcAddress for ES3.1 and above entry points.
 PFNGLDISPATCHCOMPUTEPROC glDispatchCompute;
-PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
 #endif
 static std::once_flag sGlExtInitialized;
 #endif // __EMSCRIPTEN__
@@ -118,9 +120,11 @@ void importGLESExtensionsEntryPoints() {
 #ifdef GL_KHR_parallel_shader_compile
         getProcAddress(glMaxShaderCompilerThreadsKHR, "glMaxShaderCompilerThreadsKHR");
 #endif
+#ifdef GL_OVR_multiview
+        getProcAddress(glFramebufferTextureMultiviewOVR, "glFramebufferTextureMultiviewOVR");
+#endif
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
         getProcAddress(glDispatchCompute, "glDispatchCompute");
-        getProcAddress(glFramebufferTextureMultiviewOVR, "glFramebufferTextureMultiviewOVR");
 #endif
     });
 #endif // __EMSCRIPTEN__

filament/backend/src/opengl/gl_headers.h

@@ -151,9 +151,11 @@ extern PFNGLDISCARDFRAMEBUFFEREXTPROC glDiscardFramebufferEXT;
 #ifdef GL_KHR_parallel_shader_compile
 extern PFNGLMAXSHADERCOMPILERTHREADSKHRPROC glMaxShaderCompilerThreadsKHR;
 #endif
+#ifdef GL_OVR_multiview
+extern PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
+#endif
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
 extern PFNGLDISPATCHCOMPUTEPROC glDispatchCompute;
-extern PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
 #endif
 #endif // __EMSCRIPTEN__
 } // namespace glext

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

@@ -247,8 +247,8 @@ void PlatformCocoaGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept
     delete cocoaSwapChain;
 }
 
-void PlatformCocoaGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-        Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformCocoaGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
             "ContextManagerCocoa does not support using distinct draw/read swap chains.");
     CocoaGLSwapChain* swapChain = (CocoaGLSwapChain*)drawSwapChain;
@@ -287,6 +287,7 @@ void PlatformCocoaGL::makeCurrent(Platform::SwapChain* drawSwapChain,
 
     swapChain->previousBounds = currentBounds;
     swapChain->previousWindowFrame = currentWindowFrame;
+    return true;
 }
 
 void PlatformCocoaGL::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -143,11 +143,12 @@ void PlatformCocoaTouchGL::destroySwapChain(Platform::SwapChain* swapChain) noex
     }
 }
 
-uint32_t PlatformCocoaTouchGL::createDefaultRenderTarget() noexcept {
+uint32_t PlatformCocoaTouchGL::getDefaultFramebufferObject() noexcept {
     return pImpl->mDefaultFramebuffer;
 }
 
-void PlatformCocoaTouchGL::makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept {
+bool PlatformCocoaTouchGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
             "PlatformCocoaTouchGL does not support using distinct draw/read swap chains.");
     CAEAGLLayer* const glLayer = (__bridge CAEAGLLayer*) drawSwapChain;
@@ -182,6 +183,7 @@ void PlatformCocoaTouchGL::makeCurrent(SwapChain* drawSwapChain, SwapChain* read
         ASSERT_POSTCONDITION(status == GL_FRAMEBUFFER_COMPLETE, "Incomplete framebuffer.");
         glBindFramebuffer(GL_FRAMEBUFFER, oldFramebuffer);
     }
+    return true;
 }
 
 void PlatformCocoaTouchGL::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -25,8 +25,9 @@
 
 #include <EGL/egl.h>
 #include <EGL/eglext.h>
-#if defined(__ANDROID__)
+#include <EGL/eglplatform.h>
 
+#if defined(__ANDROID__)
 #include <sys/system_properties.h>
 #endif
 #include <utils/compiler.h>
@@ -173,7 +174,7 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const Platform::DriverCon
     };
 
 #ifdef __ANDROID__
-    bool requestES2Context = false;
+    bool requestES2Context = driverConfig.forceGLES2Context;
     char property[PROP_VALUE_MAX];
     int const length = __system_property_get("debug.filament.es2", property);
     if (length > 0) {
@@ -282,12 +283,14 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const Platform::DriverCon
         }
     }
 
-    if (UTILS_UNLIKELY(makeCurrent(mEGLContext, mEGLDummySurface, mEGLDummySurface) == EGL_FALSE)) {
+    if (UTILS_UNLIKELY(
+            egl.makeCurrent(mEGLContext, mEGLDummySurface, mEGLDummySurface) == EGL_FALSE)) {
         // eglMakeCurrent failed
         logEglError("eglMakeCurrent");
         goto error;
     }
 
+    mCurrentContextType = ContextType::UNPROTECTED;
     mContextAttribs = std::move(contextAttribs);
 
     initializeGlExtensions();
@@ -362,26 +365,6 @@ void PlatformEGL::releaseContext() noexcept {
     eglReleaseThread();
 }
 
-EGLBoolean PlatformEGL::makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
-    return makeCurrent(mCurrentContext, drawSurface, readSurface);
-}
-
-EGLBoolean PlatformEGL::makeCurrent(EGLContext context,
-        EGLSurface drawSurface, EGLSurface readSurface) noexcept {
-    if (UTILS_UNLIKELY((mCurrentContext != context ||
-            drawSurface != mCurrentDrawSurface || readSurface != mCurrentReadSurface))) {
-        EGLBoolean const success = eglMakeCurrent(
-                mEGLDisplay, drawSurface, readSurface, context);
-        if (success) {
-            mCurrentDrawSurface = drawSurface;
-            mCurrentReadSurface = readSurface;
-            mCurrentContext = context;
-        }
-        return success;
-    }
-    return EGL_TRUE;
-}
-
 void PlatformEGL::terminate() noexcept {
     // it's always allowed to use EGL_NO_SURFACE, EGL_NO_CONTEXT
     eglMakeCurrent(mEGLDisplay, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
@@ -468,30 +451,10 @@ bool PlatformEGL::isSRGBSwapChainSupported() const noexcept {
     return ext.egl.KHR_gl_colorspace;
 }
 
-bool PlatformEGL::isSwapChainProtected(Platform::SwapChain const* swapChain) noexcept {
-    if (swapChain) {
-        SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
-        return bool(sc->flags & SWAP_CHAIN_CONFIG_PROTECTED_CONTENT);
-    }
-    return false;
-}
-
 Platform::SwapChain* PlatformEGL::createSwapChain(
         void* nativeWindow, uint64_t flags) noexcept {
 
-    EGLConfig config = EGL_NO_CONFIG_KHR;
-    if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
-        config = findSwapChainConfig(flags, true, false);
-    } else {
-        config = mEGLConfig;
-    }
-
-    if (UTILS_UNLIKELY(config == EGL_NO_CONFIG_KHR)) {
-        return nullptr;
-    }
-
     Config attribs;
-
     if (ext.egl.KHR_gl_colorspace) {
         if (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) {
             attribs[EGL_GL_COLORSPACE_KHR] = EGL_GL_COLORSPACE_SRGB_KHR;
@@ -508,16 +471,27 @@ Platform::SwapChain* PlatformEGL::createSwapChain(
         flags &= ~SWAP_CHAIN_CONFIG_PROTECTED_CONTENT;
     }
 
-    EGLSurface sur = eglCreateWindowSurface(mEGLDisplay, config,
-            (EGLNativeWindowType)nativeWindow, attribs.data());
-
-    if (UTILS_UNLIKELY(sur == EGL_NO_SURFACE)) {
-        logEglError("eglCreateWindowSurface");
-        return nullptr;
+    EGLConfig config = EGL_NO_CONFIG_KHR;
+    if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
+        config = findSwapChainConfig(flags, true, false);
+    } else {
+        config = mEGLConfig;
     }
 
-    // this is not fatal
-    eglSurfaceAttrib(mEGLDisplay, sur, EGL_SWAP_BEHAVIOR, EGL_BUFFER_DESTROYED);
+    EGLSurface sur = EGL_NO_SURFACE;
+    if (UTILS_LIKELY(config != EGL_NO_CONFIG_KHR)) {
+        sur = eglCreateWindowSurface(mEGLDisplay, config,
+                (EGLNativeWindowType)nativeWindow, attribs.data());
+
+        if (UTILS_LIKELY(sur != EGL_NO_SURFACE)) {
+            // this is not fatal
+            eglSurfaceAttrib(mEGLDisplay, sur, EGL_SWAP_BEHAVIOR, EGL_BUFFER_DESTROYED);
+        } else {
+            logEglError("PlatformEGL::createSwapChain: eglCreateWindowSurface");
+        }
+    } else {
+        // error already logged
+    }
 
     SwapChainEGL* const sc = new(std::nothrow) SwapChainEGL({
         .sur = sur,
@@ -532,17 +506,6 @@ Platform::SwapChain* PlatformEGL::createSwapChain(
 Platform::SwapChain* PlatformEGL::createSwapChain(
         uint32_t width, uint32_t height, uint64_t flags) noexcept {
 
-    EGLConfig config = EGL_NO_CONFIG_KHR;
-    if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
-        config = findSwapChainConfig(flags, false, true);
-    } else {
-        config = mEGLConfig;
-    }
-
-    if (UTILS_UNLIKELY(config == EGL_NO_CONFIG_KHR)) {
-        return nullptr;
-    }
-
     Config attribs = {
             { EGL_WIDTH,  EGLint(width) },
             { EGL_HEIGHT, EGLint(height) },
@@ -564,17 +527,28 @@ Platform::SwapChain* PlatformEGL::createSwapChain(
         flags &= ~SWAP_CHAIN_CONFIG_PROTECTED_CONTENT;
     }
 
-    EGLSurface sur = eglCreatePbufferSurface(mEGLDisplay, config, attribs.data());
+    EGLConfig config = EGL_NO_CONFIG_KHR;
+    if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
+        config = findSwapChainConfig(flags, true, false);
+    } else {
+        config = mEGLConfig;
+    }
 
-    if (UTILS_UNLIKELY(sur == EGL_NO_SURFACE)) {
-        logEglError("eglCreatePbufferSurface");
-        return nullptr;
+    EGLSurface sur = EGL_NO_SURFACE;
+    if (UTILS_LIKELY(config != EGL_NO_CONFIG_KHR)) {
+        sur = eglCreatePbufferSurface(mEGLDisplay, config, attribs.data());
+        if (UTILS_UNLIKELY(sur == EGL_NO_SURFACE)) {
+            logEglError("PlatformEGL::createSwapChain: eglCreatePbufferSurface");
+        }
+    } else {
+        // error already logged
     }
 
     SwapChainEGL* const sc = new(std::nothrow) SwapChainEGL({
             .sur = sur,
             .attribs = std::move(attribs),
-            .config = config
+            .config = config,
+            .flags = flags
     });
     return sc;
 }
@@ -583,29 +557,45 @@ void PlatformEGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     if (swapChain) {
         SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
         if (sc->sur != EGL_NO_SURFACE) {
-            makeCurrent(mCurrentContext, mEGLDummySurface, mEGLDummySurface);
+            egl.makeCurrent(mEGLDummySurface, mEGLDummySurface);
             eglDestroySurface(mEGLDisplay, sc->sur);
             delete sc;
         }
     }
 }
 
-void PlatformEGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-        Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformEGL::isSwapChainProtected(Platform::SwapChain* swapChain) noexcept {
+    if (swapChain) {
+        SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
+        return bool(sc->flags & SWAP_CHAIN_CONFIG_PROTECTED_CONTENT);
+    }
+    return false;
+}
+
+OpenGLPlatform::ContextType PlatformEGL::getCurrentContextType() const noexcept {
+    return mCurrentContextType;
+}
+
+bool PlatformEGL::makeCurrent(ContextType type,
+        SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept {
     SwapChainEGL const* const dsc = static_cast<SwapChainEGL const*>(drawSwapChain);
     SwapChainEGL const* const rsc = static_cast<SwapChainEGL const*>(readSwapChain);
-    makeCurrent(mEGLContext, dsc->sur, rsc->sur);
+    EGLContext context = getContextForType(type);
+    EGLBoolean const success = egl.makeCurrent(context, dsc->sur, rsc->sur);
+    return success == EGL_TRUE ? true : false;
 }
 
 void PlatformEGL::makeCurrent(Platform::SwapChain* drawSwapChain,
         Platform::SwapChain* readSwapChain,
         utils::Invocable<void()> preContextChange,
         utils::Invocable<void(size_t index)> postContextChange) noexcept {
-    SwapChainEGL const* const dsc = static_cast<SwapChainEGL const*>(drawSwapChain);
-    SwapChainEGL const* const rsc = static_cast<SwapChainEGL const*>(readSwapChain);
-    EGLContext context = mEGLContext;
+
+    assert_invariant(drawSwapChain);
+    assert_invariant(readSwapChain);
+
+    ContextType type = ContextType::UNPROTECTED;
     if (ext.egl.EXT_protected_content) {
-        bool const swapChainProtected = PlatformEGL::isSwapChainProtected(dsc);
+        bool const swapChainProtected = isSwapChainProtected(drawSwapChain);
         if (UTILS_UNLIKELY(swapChainProtected)) {
             // we need a protected context
             if (UTILS_UNLIKELY(mEGLContextProtected == EGL_NO_CONTEXT)) {
@@ -622,34 +612,36 @@ void PlatformEGL::makeCurrent(Platform::SwapChain* drawSwapChain,
                     goto error;
                 }
             }
-            context = mEGLContextProtected;
+            type = ContextType::PROTECTED;
             error: ;
         }
 
-        bool const contextChange = context != mCurrentContext;
+        bool const contextChange = type != mCurrentContextType;
+        mCurrentContextType = type;
+
         if (UTILS_UNLIKELY(contextChange)) {
             preContextChange();
-            EGLBoolean const success = makeCurrent(context, dsc->sur, rsc->sur);
+            bool const success = makeCurrent(mCurrentContextType, drawSwapChain, readSwapChain);
             if (UTILS_UNLIKELY(!success)) {
                 logEglError("PlatformEGL::makeCurrent");
                 if (mEGLContextProtected != EGL_NO_CONTEXT) {
                     eglDestroyContext(mEGLDisplay, mEGLContextProtected);
                     mEGLContextProtected = EGL_NO_CONTEXT;
                 }
-                context = mEGLContext;
+                mCurrentContextType = ContextType::UNPROTECTED;
             }
             if (UTILS_LIKELY(!swapChainProtected && mEGLContextProtected != EGL_NO_CONTEXT)) {
                 // We don't need the protected context anymore, unbind and destroy right away.
                 eglDestroyContext(mEGLDisplay, mEGLContextProtected);
                 mEGLContextProtected = EGL_NO_CONTEXT;
             }
-            size_t const contextIndex = (context == mEGLContext) ? 0 : 1;
+            size_t const contextIndex = (mCurrentContextType == ContextType::PROTECTED) ? 1 : 0;
             postContextChange(contextIndex);
             return;
         }
     }
 
-    EGLBoolean const success = makeCurrent(context, dsc->sur, rsc->sur);
+    bool const success = makeCurrent(mCurrentContextType, drawSwapChain, readSwapChain);
     if (UTILS_UNLIKELY(!success)) {
         logEglError("PlatformEGL::makeCurrent");
     }
@@ -746,6 +738,17 @@ void PlatformEGL::initializeGlExtensions() noexcept {
     ext.gl.OES_EGL_image_external_essl3 = glExtensions.has("GL_OES_EGL_image_external_essl3");
 }
 
+EGLContext PlatformEGL::getContextForType(OpenGLPlatform::ContextType type) const noexcept {
+    switch (type) {
+        case ContextType::NONE:
+            return EGL_NO_CONTEXT;
+        case ContextType::UNPROTECTED:
+            return mEGLContext;
+        case ContextType::PROTECTED:
+            return mEGLContextProtected;
+    }
+}
+
 // ---------------------------------------------------------------------------------------------
 
 PlatformEGL::Config::Config() = default;
@@ -782,4 +785,23 @@ void PlatformEGL::Config::erase(EGLint name) noexcept {
     }
 }
 
+// ------------------------------------------------------------------------------------------------
+
+EGLBoolean PlatformEGL::EGL::makeCurrent(EGLContext context, EGLSurface drawSurface,
+        EGLSurface readSurface) noexcept {
+    if (UTILS_UNLIKELY((
+            mCurrentContext != context ||
+            drawSurface != mCurrentDrawSurface || readSurface != mCurrentReadSurface))) {
+        EGLBoolean const success = eglMakeCurrent(
+                mEGLDisplay, drawSurface, readSurface, context);
+        if (success) {
+            mCurrentDrawSurface = drawSurface;
+            mCurrentReadSurface = readSurface;
+            mCurrentContext = context;
+        }
+        return success;
+    }
+    return EGL_TRUE;
+}
+
 } // namespace filament::backend

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

@@ -25,6 +25,8 @@
 #include <android/api-level.h>
 #include <android/hardware_buffer.h>
 
+#include <utils/android/PerformanceHintManager.h>
+
 #include <utils/compiler.h>
 #include <utils/ostream.h>
 #include <utils/Log.h>
@@ -36,8 +38,11 @@
 
 #include <jni.h>
 
+#include <chrono>
 #include <new>
 
+#include <unistd.h>
+
 #include <stddef.h>
 #include <stdint.h>
 #include <stdlib.h>
@@ -112,8 +117,38 @@ void PlatformEGLAndroid::terminate() noexcept {
     PlatformEGL::terminate();
 }
 
+void PlatformEGLAndroid::beginFrame(
+        int64_t monotonic_clock_ns,
+        int64_t refreshIntervalNs,
+        uint32_t frameId) noexcept {
+    if (mPerformanceHintSession.isValid()) {
+        if (refreshIntervalNs <= 0) {
+            // we're not provided with a target time, assume 16.67ms
+            refreshIntervalNs = 16'666'667;
+        }
+        mStartTimeOfActualWork = clock::time_point(std::chrono::nanoseconds(monotonic_clock_ns));
+        mPerformanceHintSession.updateTargetWorkDuration(refreshIntervalNs);
+    }
+    PlatformEGL::beginFrame(monotonic_clock_ns, refreshIntervalNs, frameId);
+}
+
+void backend::PlatformEGLAndroid::preCommit() noexcept {
+    if (mPerformanceHintSession.isValid()) {
+        auto const actualWorkDuration = std::chrono::duration_cast<std::chrono::nanoseconds>(
+                clock::now() - mStartTimeOfActualWork);
+        mPerformanceHintSession.reportActualWorkDuration(actualWorkDuration.count());
+    }
+    PlatformEGL::preCommit();
+}
+
 Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
         const Platform::DriverConfig& driverConfig) noexcept {
+
+    // the refresh rate default value doesn't matter, we change it later
+    int32_t const tid = gettid();
+    mPerformanceHintSession = PerformanceHintManager::Session{
+            mPerformanceHintManager, &tid, 1, 16'666'667 };
+
     Driver* driver = PlatformEGL::createDriver(sharedContext, driverConfig);
     auto extensions = GLUtils::split(eglQueryString(mEGLDisplay, EGL_EXTENSIONS));
 
@@ -142,11 +177,12 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
 }
 
 void PlatformEGLAndroid::setPresentationTime(int64_t presentationTimeInNanosecond) noexcept {
-    if (mCurrentDrawSurface != EGL_NO_SURFACE) {
+    EGLSurface currentDrawSurface = eglGetCurrentSurface(EGL_DRAW);
+    if (currentDrawSurface != EGL_NO_SURFACE) {
         if (eglPresentationTimeANDROID) {
             eglPresentationTimeANDROID(
                     mEGLDisplay,
-                    mCurrentDrawSurface,
+                    currentDrawSurface,
                     presentationTimeInNanosecond);
         }
     }

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

@@ -265,10 +265,11 @@ void PlatformGLX::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     }
 }
 
-void PlatformGLX::makeCurrent(
-        Platform::SwapChain* drawSwapChain, Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformGLX::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     g_glx.setCurrentContext(mGLXDisplay,
             (GLXDrawable)drawSwapChain, (GLXDrawable)readSwapChain, mGLXContext);
+    return true;
 }
 
 void PlatformGLX::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -255,8 +255,8 @@ void PlatformWGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     wglMakeCurrent(mWhdc, mContext);
 }
 
-void PlatformWGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-                              Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformWGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
                                   "PlatformWGL does not support distinct draw/read swap chains.");
 
@@ -269,6 +269,7 @@ void PlatformWGL::makeCurrent(Platform::SwapChain* drawSwapChain,
             wglMakeCurrent(0, NULL);
         }
     }
+    return true;
 }
 
 void PlatformWGL::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -46,8 +46,9 @@ Platform::SwapChain* PlatformWebGL::createSwapChain(
 void PlatformWebGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
 }
 
-void PlatformWebGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-        Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformWebGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
+    return true;
 }
 
 void PlatformWebGL::commit(Platform::SwapChain* swapChain) noexcept {

filament/backend/src/ostream.cpp

@@ -410,6 +410,7 @@ io::ostream& operator<<(io::ostream& out, const RasterState& rs) {
 io::ostream& operator<<(io::ostream& out, const TargetBufferInfo& tbi) {
     return out << "TargetBufferInfo{"
     << "handle=" << tbi.handle
+    << ", baseViewIndex=" << tbi.baseViewIndex
     << ", level=" << tbi.level
     << ", layer=" << tbi.layer << "}";
 }

filament/backend/src/vulkan/VulkanBlitter.cpp

@@ -26,21 +26,16 @@
 
 #include <smolv.h>
 
-#include "generated/vkshaders/vkshaders.h"
-
 using namespace bluevk;
 using namespace utils;
 
 namespace filament::backend {
 
-using ImgUtil = VulkanImageUtility;
-
 namespace {
 
 inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect, VkFilter filter,
         VulkanAttachment src, VulkanAttachment dst,
         const VkOffset3D srcRect[2], const VkOffset3D dstRect[2]) {
-
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
         utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
@@ -48,21 +43,8 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
                       << " layout=" << dst.getLayout() << utils::io::endl;
     }
 
-    const VkImageSubresourceRange srcRange = {
-            .aspectMask = aspect,
-            .baseMipLevel = src.level,
-            .levelCount = 1,
-            .baseArrayLayer = src.layer,
-            .layerCount = 1,
-    };
-
-    const VkImageSubresourceRange dstRange = {
-            .aspectMask = aspect,
-            .baseMipLevel = dst.level,
-            .levelCount = 1,
-            .baseArrayLayer = dst.layer,
-            .layerCount = 1,
-    };
+    VkImageSubresourceRange const srcRange = src.getSubresourceRange();
+    VkImageSubresourceRange const dstRange = dst.getSubresourceRange();
 
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
@@ -77,15 +59,15 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
             .dstOffsets = { dstRect[0], dstRect[1] },
     }};
     vkCmdBlitImage(cmdbuffer,
-            src.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            dst.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_DST),
+            src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
+            dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
             1, blitRegions, filter);
 
     if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = ImgUtil::getDefaultLayout(src.texture->usage);
+        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
     }
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = ImgUtil::getDefaultLayout(dst.texture->usage);
+        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
     }
     src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
     dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
@@ -93,7 +75,6 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
 
 inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
         VulkanAttachment src, VulkanAttachment dst) {
-
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
         utils::slog.d << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
@@ -101,21 +82,8 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
                       << " layout=" << dst.getLayout() << utils::io::endl;
     }
 
-    const VkImageSubresourceRange srcRange = {
-            .aspectMask = aspect,
-            .baseMipLevel = src.level,
-            .levelCount = 1,
-            .baseArrayLayer = src.layer,
-            .layerCount = 1,
-    };
-
-    const VkImageSubresourceRange dstRange = {
-            .aspectMask = aspect,
-            .baseMipLevel = dst.level,
-            .levelCount = 1,
-            .baseArrayLayer = dst.layer,
-            .layerCount = 1,
-    };
+    VkImageSubresourceRange const srcRange = src.getSubresourceRange();
+    VkImageSubresourceRange const dstRange = dst.getSubresourceRange();
 
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
@@ -133,15 +101,15 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
             .extent = { src.getExtent2D().width, src.getExtent2D().height, 1 },
     }};
     vkCmdResolveImage(cmdbuffer,
-            src.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            dst.getImage(), ImgUtil::getVkLayout(VulkanLayout::TRANSFER_DST),
+            src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
+            dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
             1, resolveRegions);
 
     if (oldSrcLayout == VulkanLayout::UNDEFINED) {
-        oldSrcLayout = ImgUtil::getDefaultLayout(src.texture->usage);
+        oldSrcLayout = imgutil::getDefaultLayout(src.texture->usage);
     }
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
-        oldDstLayout = ImgUtil::getDefaultLayout(dst.texture->usage);
+        oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
     }
     src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
     dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
@@ -154,15 +122,9 @@ struct BlitterUniforms {
 
 }// anonymous namespace
 
-VulkanBlitter::VulkanBlitter() noexcept = default;
-
-void VulkanBlitter::initialize(VkPhysicalDevice physicalDevice, VkDevice device,
-        VmaAllocator allocator, VulkanCommands* commands) noexcept {
-    mPhysicalDevice = physicalDevice;
-    mDevice = device;
-    mAllocator = allocator;
-    mCommands = commands;
-}
+VulkanBlitter::VulkanBlitter(VkPhysicalDevice physicalDevice, VulkanCommands* commands) noexcept
+    : mPhysicalDevice(physicalDevice),
+      mCommands(commands) {}
 
 void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {
 

filament/backend/src/vulkan/VulkanBlitter.h

@@ -17,6 +17,7 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANBLITTER_H
 #define TNT_FILAMENT_BACKEND_VULKANBLITTER_H
 
+#include "VulkanCommands.h"
 #include "VulkanContext.h"
 
 #include <utils/compiler.h>
@@ -32,10 +33,7 @@ struct VulkanProgram;
 
 class VulkanBlitter {
 public:
-    VulkanBlitter() noexcept;
-
-    void initialize(VkPhysicalDevice physicalDevice, VkDevice device,
-            VmaAllocator allocator, VulkanCommands* commands) noexcept;
+    VulkanBlitter(VkPhysicalDevice physicalDevice, VulkanCommands* commands) noexcept;
 
     void blit(VkFilter filter,
             VulkanAttachment dst, const VkOffset3D* dstRectPair,
@@ -47,8 +45,6 @@ public:
 
 private:
     UTILS_UNUSED VkPhysicalDevice mPhysicalDevice;
-    VkDevice mDevice;
-    VmaAllocator mAllocator;
     VulkanCommands* mCommands;
 };
 

filament/backend/src/vulkan/VulkanBuffer.cpp

@@ -27,8 +27,8 @@ VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
         VkBufferUsageFlags usage, uint32_t numBytes)
     : mAllocator(allocator),
       mStagePool(stagePool),
-      mUsage(usage) {
-
+      mUsage(usage),
+      mUpdatedBytes(0) {
     // for now make sure that only 1 bit is set in usage
     // (because loadFromCpu() assumes that somewhat)
     assert_invariant(usage && !(usage & (usage - 1)));
@@ -49,8 +49,7 @@ VulkanBuffer::~VulkanBuffer() {
 }
 
 void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
-        uint32_t numBytes) const {
-    assert_invariant(byteOffset == 0);
+        uint32_t numBytes) {
     VulkanStage const* stage = mStagePool.acquireStage(numBytes);
     void* mapped;
     vmaMapMemory(mAllocator, stage->memory, &mapped);
@@ -58,15 +57,51 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
     vmaUnmapMemory(mAllocator, stage->memory);
     vmaFlushAllocation(mAllocator, stage->memory, byteOffset, numBytes);
 
-    VkBufferCopy region{ .size = numBytes };
+    // If there was a previous update, then we need to make sure the following write is properly
+    // synced with the previous read.
+    if (mUpdatedBytes > 0) {
+        VkAccessFlags srcAccess = 0;
+        VkPipelineStageFlags srcStage = 0;
+        if (mUsage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT) {
+            srcAccess = VK_ACCESS_SHADER_READ_BIT;
+            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
+        } else if (mUsage & VK_BUFFER_USAGE_VERTEX_BUFFER_BIT) {
+            srcAccess = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
+            srcStage = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
+        } else if (mUsage & VK_BUFFER_USAGE_INDEX_BUFFER_BIT) {
+            srcAccess = VK_ACCESS_INDEX_READ_BIT;
+            srcStage = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
+        }
+
+        VkBufferMemoryBarrier barrier{
+            .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+            .srcAccessMask = srcAccess,
+            .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .buffer = mGpuBuffer,
+            .size = mUpdatedBytes,
+        };
+        vkCmdPipelineBarrier(cmdbuf, srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 1,
+                &barrier, 0, nullptr);
+    }
+
+    VkBufferCopy region {
+            .srcOffset = 0,
+            .dstOffset = byteOffset,
+            .size = numBytes,
+    };
     vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);
 
+    mUpdatedBytes = numBytes;
+
     // Firstly, ensure that the copy finishes before the next draw call.
     // Secondly, in case the user decides to upload another chunk (without ever using the first one)
     // we need to ensure that this upload completes first (hence
     // dstStageMask=VK_PIPELINE_STAGE_TRANSFER_BIT).
     VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
     VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT;
+
     if (mUsage & VK_BUFFER_USAGE_VERTEX_BUFFER_BIT) {
         dstAccessMask |= VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
         dstStageMask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
@@ -75,26 +110,24 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
         dstStageMask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
     } else if (mUsage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT) {
         dstAccessMask |= VK_ACCESS_UNIFORM_READ_BIT;
-        // NOTE: ideally dstStageMask would include VERTEX_SHADER_BIT | FRAGMENT_SHADER_BIT, but
-        // this seems to be insufficient on Mali devices. To work around this we are using a more
-        // aggressive ALL_GRAPHICS_BIT barrier.
-        dstStageMask |= VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
+        dstStageMask |=
+                (VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
     } else if (mUsage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT) {
         // TODO: implement me
     }
 
     VkBufferMemoryBarrier barrier{
-	    .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-	    .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-	    .dstAccessMask = dstAccessMask,
-	    .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-	    .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-	    .buffer = mGpuBuffer,
-	    .size = VK_WHOLE_SIZE,
+        .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+        .dstAccessMask = dstAccessMask,
+        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .buffer = mGpuBuffer,
+        .size = VK_WHOLE_SIZE,
     };
 
     vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask, 0, 0, nullptr, 1,
-	    &barrier, 0, nullptr);
+            &barrier, 0, nullptr);
 }
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanBuffer.h

@@ -19,6 +19,7 @@
 
 #include "VulkanContext.h"
 #include "VulkanStagePool.h"
+#include "VulkanMemory.h"
 
 namespace filament::backend {
 
@@ -29,7 +30,7 @@ public:
             uint32_t numBytes);
     ~VulkanBuffer();
     void loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint32_t byteOffset,
-            uint32_t numBytes) const;
+            uint32_t numBytes);
     VkBuffer getGpuBuffer() const {
         return mGpuBuffer;
     }
@@ -41,6 +42,7 @@ private:
     VmaAllocation mGpuMemory = VK_NULL_HANDLE;
     VkBuffer mGpuBuffer = VK_NULL_HANDLE;
     VkBufferUsageFlags mUsage = {};
+    uint32_t mUpdatedBytes = 0;
 };
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -47,7 +47,8 @@ VulkanCmdFence::VulkanCmdFence(VkFence ifence)
 
 VulkanCommandBuffer::VulkanCommandBuffer(VulkanResourceAllocator* allocator, VkDevice device,
         VkCommandPool pool)
-    : mResourceManager(allocator) {
+    : mResourceManager(allocator),
+      mPipeline(VK_NULL_HANDLE) {
     // Create the low-level command buffer.
     const VkCommandBufferAllocateInfo allocateInfo{
             .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
@@ -155,7 +156,7 @@ VulkanCommands::VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFam
 #endif
 }
 
-VulkanCommands::~VulkanCommands() {
+void VulkanCommands::terminate() {
     wait();
     gc();
     vkDestroyCommandPool(mDevice, mPool, VKALLOC);

filament/backend/src/vulkan/VulkanCommands.h

@@ -89,6 +89,15 @@ 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 {
@@ -103,6 +112,7 @@ struct VulkanCommandBuffer {
 private:
     VulkanAcquireOnlyResourceManager mResourceManager;
     VkCommandBuffer mBuffer;
+    VkPipeline mPipeline;
 };
 
 // Allows classes to be notified after a new command buffer has been activated.
@@ -139,71 +149,72 @@ public:
 //    - We do this because vkGetFenceStatus must be called from the rendering thread.
 //
 class VulkanCommands {
-    public:
-        VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
-                VulkanContext* context, VulkanResourceAllocator* allocator);
-        ~VulkanCommands();
+public:
+    VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
+            VulkanContext* context, VulkanResourceAllocator* allocator);
 
-        // Creates a "current" command buffer if none exists, otherwise returns the current one.
-        VulkanCommandBuffer& get();
+    void terminate();
 
-        // Submits the current command buffer if it exists, then sets "current" to null.
-        // If there are no outstanding commands then nothing happens and this returns false.
-        bool flush();
+    // Creates a "current" command buffer if none exists, otherwise returns the current one.
+    VulkanCommandBuffer& get();
 
-        // Returns the "rendering finished" semaphore for the most recent flush and removes
-        // it from the existing dependency chain. This is especially useful for setting up
-        // vkQueuePresentKHR.
-        VkSemaphore acquireFinishedSignal();
+    // Submits the current command buffer if it exists, then sets "current" to null.
+    // If there are no outstanding commands then nothing happens and this returns false.
+    bool flush();
 
-        // Takes a semaphore that signals when the next flush can occur. Only one injected
-        // semaphore is allowed per flush. Useful after calling vkAcquireNextImageKHR.
-        void injectDependency(VkSemaphore next);
+    // Returns the "rendering finished" semaphore for the most recent flush and removes
+    // it from the existing dependency chain. This is especially useful for setting up
+    // vkQueuePresentKHR.
+    VkSemaphore acquireFinishedSignal();
 
-        // Destroys all command buffers that are no longer in use.
-        void gc();
+    // Takes a semaphore that signals when the next flush can occur. Only one injected
+    // semaphore is allowed per flush. Useful after calling vkAcquireNextImageKHR.
+    void injectDependency(VkSemaphore next);
 
-        // Waits for all outstanding command buffers to finish.
-        void wait();
+    // Destroys all command buffers that are no longer in use.
+    void gc();
 
-        // Updates the atomic "status" variable in every extant fence.
-        void updateFences();
+    // Waits for all outstanding command buffers to finish.
+    void wait();
 
-        // Sets an observer who is notified every time a new command buffer has been made "current".
-        // The observer's event handler can only be called during get().
-        void setObserver(CommandBufferObserver* observer) { mObserver = observer; }
+    // Updates the atomic "status" variable in every extant fence.
+    void updateFences();
+
+    // Sets an observer who is notified every time a new command buffer has been made "current".
+    // The observer's event handler can only be called during get().
+    void setObserver(CommandBufferObserver* observer) { mObserver = observer; }
 
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-        void pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp = {});
+    void pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp = {});
 
-        void popGroupMarker();
+    void popGroupMarker();
 
-        void insertEventMarker(char const* string, uint32_t len);
+    void insertEventMarker(char const* string, uint32_t len);
 
-        std::string getTopGroupMarker() const;
+    std::string getTopGroupMarker() const;
 #endif
 
-    private:
-        static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
-        VkDevice const mDevice;
-        VkQueue const mQueue;
-        VkCommandPool const mPool;
-        VulkanContext const* mContext;
+private:
+    static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
+    VkDevice const mDevice;
+    VkQueue const mQueue;
+    VkCommandPool const mPool;
+    VulkanContext const* mContext;
 
-        // int8 only goes up to 127, therefore capacity must be less than that.
-        static_assert(CAPACITY < 128);
-        int8_t mCurrentCommandBufferIndex = -1;
-        VkSemaphore mSubmissionSignal = {};
-        VkSemaphore mInjectedSignal = {};
-        utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>> mStorage;
-        VkFence mFences[CAPACITY] = {};
-        VkSemaphore mSubmissionSignals[CAPACITY] = {};
-        uint8_t mAvailableBufferCount = CAPACITY;
-        CommandBufferObserver* mObserver = nullptr;
+    // int8 only goes up to 127, therefore capacity must be less than that.
+    static_assert(CAPACITY < 128);
+    int8_t mCurrentCommandBufferIndex = -1;
+    VkSemaphore mSubmissionSignal = {};
+    VkSemaphore mInjectedSignal = {};
+    utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>> mStorage;
+    VkFence mFences[CAPACITY] = {};
+    VkSemaphore mSubmissionSignals[CAPACITY] = {};
+    uint8_t mAvailableBufferCount = CAPACITY;
+    CommandBufferObserver* mObserver = nullptr;
 
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-        std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
-        std::unique_ptr<VulkanGroupMarkers> mCarriedOverMarkers;
+    std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
+    std::unique_ptr<VulkanGroupMarkers> mCarriedOverMarkers;
 #endif
 };
 

filament/backend/src/vulkan/VulkanConstants.h

@@ -43,9 +43,12 @@
 // Enables Android systrace
 #define FVK_DEBUG_SYSTRACE                0x00000001
 
-// Group markers are used to denote collection of GPU commands.  It is typically at the granualarity
-// of a renderpass.
+// Group markers are used to denote collections of GPU commands.  It is typically at the
+// granualarity of a renderpass. You can enable this along with FVK_DEBUG_DEBUG_UTILS to take
+// advantage of vkCmdBegin/EndDebugUtilsLabelEXT. You can also just enable this with
+// FVK_DEBUG_PRINT_GROUP_MARKERS to print the current marker to stdout.
 #define FVK_DEBUG_GROUP_MARKERS           0x00000002
+
 #define FVK_DEBUG_TEXTURE                 0x00000004
 #define FVK_DEBUG_LAYOUT_TRANSITION       0x00000008
 #define FVK_DEBUG_COMMAND_BUFFER          0x00000010
@@ -63,25 +66,23 @@
 // Enable the debug utils extension if it is available.
 #define FVK_DEBUG_DEBUG_UTILS             0x00008000
 
+// Use this to debug potential Handle/Resource leakage. It will print out reference counts for all
+// the currently active resources.
 #define FVK_DEBUG_RESOURCE_LEAK           0x00010000
 
-// Usefaul default combinations
+// Useful default combinations
 #define FVK_DEBUG_EVERYTHING              0xFFFFFFFF
 #define FVK_DEBUG_PERFORMANCE     \
     FVK_DEBUG_SYSTRACE
 
-#define FVK_DEBUG_CORRECTNESS     \
-    FVK_DEBUG_VALIDATION |        \
-    FVK_DEBUG_SHADER_MODULE |     \
-    FVK_DEBUG_TEXTURE |           \
-    FVK_DEBUG_LAYOUT_TRANSITION
-
-#define FVK_DEBUG_RENDER_PASSES   \
-    FVK_DEBUG_GROUP_MARKERS |     \
-    FVK_DEBUG_PRINT_GROUP_MARKERS
+#if defined(FILAMENT_BACKEND_DEBUG_FLAG)
+#define FVK_DEBUG_FORWARDED_FLAG (FILAMENT_BACKEND_DEBUG_FLAG & FVK_DEBUG_EVERYTHING)
+#else
+#define FVK_DEBUG_FORWARDED_FLAG 0
+#endif
 
 #ifndef NDEBUG
-#define FVK_DEBUG_FLAGS (FVK_DEBUG_PERFORMANCE)
+#define FVK_DEBUG_FLAGS (FVK_DEBUG_PERFORMANCE | FVK_DEBUG_FORWARDED_FLAG)
 #else
 #define FVK_DEBUG_FLAGS 0
 #endif
@@ -107,7 +108,7 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 // end dependcy checks
 
 // Shorthand for combination of enabled debug flags
-#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS) || FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS) && FVK_ENABLED(FVK_DEBUG_TEXTURE)
 #define FVK_ENABLED_DEBUG_SAMPLER_NAME 1
 #else
 #define FVK_ENABLED_DEBUG_SAMPLER_NAME 0

filament/backend/src/vulkan/VulkanContext.cpp

@@ -57,15 +57,19 @@ VkExtent2D VulkanAttachment::getExtent2D() const {
     return { std::max(1u, texture->width >> level), std::max(1u, texture->height >> level) };
 }
 
-VkImageView VulkanAttachment::getImageView(VkImageAspectFlags aspect) {
+VkImageView VulkanAttachment::getImageView() {
     assert_invariant(texture);
-    return texture->getAttachmentView(getSubresourceRange(aspect));
+    return texture->getAttachmentView(getSubresourceRange());
 }
 
-VkImageSubresourceRange VulkanAttachment::getSubresourceRange(VkImageAspectFlags aspect) const {
+bool VulkanAttachment::isDepth() const {
+    return texture->getImageAspect() & VK_IMAGE_ASPECT_DEPTH_BIT;
+}
+
+VkImageSubresourceRange VulkanAttachment::getSubresourceRange() const {
     assert_invariant(texture);
     return {
-            .aspectMask = aspect,
+            .aspectMask = texture->getImageAspect(),
             .baseMipLevel = uint32_t(level),
             .levelCount = 1,
             .baseArrayLayer = uint32_t(layer),

filament/backend/src/vulkan/VulkanContext.h

@@ -19,7 +19,6 @@
 
 #include "VulkanConstants.h"
 #include "VulkanImageUtility.h"
-#include "VulkanPipelineCache.h"
 #include "VulkanUtility.h"
 
 #include <utils/bitset.h>
@@ -42,16 +41,18 @@ struct VulkanTimerQuery;
 struct VulkanCommandBuffer;
 
 struct VulkanAttachment {
-    VulkanTexture* texture;
+    VulkanTexture* texture = nullptr;
     uint8_t level = 0;
     uint16_t layer = 0;
+
+    bool isDepth() const;
     VkImage getImage() const;
     VkFormat getFormat() const;
     VulkanLayout getLayout() const;
     VkExtent2D getExtent2D() const;
-    VkImageView getImageView(VkImageAspectFlags aspect);
+    VkImageView getImageView();
     // TODO: maybe embed aspect into the attachment or texture itself.
-    VkImageSubresourceRange getSubresourceRange(VkImageAspectFlags aspect) const;
+    VkImageSubresourceRange getSubresourceRange() const;
 };
 
 class VulkanTimestamps {
@@ -102,8 +103,12 @@ public:
         return (uint32_t) VK_MAX_MEMORY_TYPES;
     }
 
-    inline VkFormatList const& getAttachmentDepthFormats() const {
-        return mDepthFormats;
+    inline VkFormatList const& getAttachmentDepthStencilFormats() const {
+        return mDepthStencilFormats;
+    }
+
+    inline VkFormatList const& getBlittableDepthStencilFormats() const {
+        return mBlittableDepthStencilFormats;
     }
 
     inline VkPhysicalDeviceLimits const& getPhysicalDeviceLimits() const noexcept {
@@ -132,7 +137,8 @@ private:
     bool mDebugMarkersSupported = false;
     bool mDebugUtilsSupported = false;
 
-    VkFormatList mDepthFormats;
+    VkFormatList mDepthStencilFormats;
+    VkFormatList mBlittableDepthStencilFormats;
 
     // For convenience so that VulkanPlatform can initialize the private fields.
     friend class VulkanPlatform;

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "vulkan/VulkanDriver.h"
+#include "VulkanDriver.h"
 
 #include "CommandStreamDispatcher.h"
 #include "DataReshaper.h"
@@ -22,7 +22,6 @@
 #include "VulkanCommands.h"
 #include "VulkanDriverFactory.h"
 #include "VulkanHandles.h"
-#include "VulkanImageUtility.h"
 #include "VulkanMemory.h"
 #include "VulkanTexture.h"
 
@@ -102,6 +101,15 @@ 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,
@@ -197,8 +205,6 @@ void DebugUtils::setName(VkObjectType type, uint64_t handle, char const* name) {
 }
 #endif // FVK_EANBLED(FVK_DEBUG_DEBUG_UTILS)
 
-using ImgUtil = VulkanImageUtility;
-
 Dispatcher VulkanDriver::getDispatcher() const noexcept {
     return ConcreteDispatcher<VulkanDriver>::make();
 }
@@ -209,11 +215,19 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
       mAllocator(createAllocator(mPlatform->getInstance(), mPlatform->getPhysicalDevice(),
               mPlatform->getDevice())),
       mContext(context),
-      mResourceAllocator(driverConfig.handleArenaSize),
+      mResourceAllocator(driverConfig.handleArenaSize, driverConfig.disableHandleUseAfterFreeCheck),
       mResourceManager(&mResourceAllocator),
       mThreadSafeResourceManager(&mResourceAllocator),
-      mPipelineCache(&mResourceAllocator),
+      mCommands(mPlatform->getDevice(), mPlatform->getGraphicsQueue(),
+              mPlatform->getGraphicsQueueFamilyIndex(), &mContext, &mResourceAllocator),
+      mPipelineLayoutCache(mPlatform->getDevice(), &mResourceAllocator),
+      mPipelineCache(mPlatform->getDevice(), mAllocator),
+      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)
@@ -237,23 +251,17 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
 #endif
 
     mTimestamps = std::make_unique<VulkanTimestamps>(mPlatform->getDevice());
-    mCommands = std::make_unique<VulkanCommands>(mPlatform->getDevice(),
-            mPlatform->getGraphicsQueue(), mPlatform->getGraphicsQueueFamilyIndex(), &mContext,
-            &mResourceAllocator);
-    mCommands->setObserver(&mPipelineCache);
-    mPipelineCache.setDevice(mPlatform->getDevice(), mAllocator);
 
-    // TOOD: move them all to be initialized by constructor
-    mStagePool.initialize(mAllocator, mCommands.get());
-    mFramebufferCache.initialize(mPlatform->getDevice());
-    mSamplerCache.initialize(mPlatform->getDevice());
+    mEmptyTexture = createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
+            mContext, mAllocator, &mCommands, mStagePool);
+    mEmptyBufferObject = createEmptyBufferObject(mAllocator, mStagePool, &mCommands);
 
-    mEmptyTexture.reset(createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
-            mContext, mAllocator, mCommands.get(), mStagePool));
+    mDescriptorSetManager.setPlaceHolders(mSamplerCache.getSampler({}), mEmptyTexture,
+            mEmptyBufferObject);
 
-    mPipelineCache.setDummyTexture(mEmptyTexture->getPrimaryImageView());
-    mBlitter.initialize(mPlatform->getPhysicalDevice(), mPlatform->getDevice(), mAllocator,
-            mCommands.get());
+    mGetPipelineFunction = [this](VulkanDescriptorSetLayoutList const& layouts, VulkanProgram* program) {
+        return mPipelineLayoutCache.getLayout(layouts, program);
+    };
 }
 
 VulkanDriver::~VulkanDriver() noexcept = default;
@@ -314,10 +322,14 @@ 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.reset();
-    mEmptyTexture.reset();
+    mCommands.terminate();
+
+    mResourceManager.clear();
     mTimestamps.reset();
 
     mBlitter.terminate();
@@ -330,6 +342,12 @@ 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);
 
@@ -346,7 +364,7 @@ void VulkanDriver::terminate() {
 }
 
 void VulkanDriver::tick(int) {
-    mCommands->updateFences();
+    mCommands.updateFences();
 }
 
 // Garbage collection should not occur too frequently, only about once per frame. Internally, the
@@ -358,9 +376,11 @@ void VulkanDriver::collectGarbage() {
     FVK_SYSTRACE_START("gc");
     // Command buffers need to be submitted and completed before other resources can be gc'd. And
     // its gc() function carrys out the *wait*.
-    mCommands->gc();
+    mCommands.gc();
     mStagePool.gc();
     mFramebufferCache.gc();
+    mPipelineCache.gc();
+    mDescriptorSetManager.gc();
 
 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
     mResourceAllocator.print();
@@ -368,16 +388,17 @@ void VulkanDriver::collectGarbage() {
 
     FVK_SYSTRACE_END();
 }
-void VulkanDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
+void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
+        int64_t refreshIntervalNs, uint32_t frameId) {
     // Do nothing.
 }
 
 void VulkanDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        FrameScheduledCallback callback, void* user) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
 }
 
 void VulkanDriver::setFrameCompletedCallback(Handle<HwSwapChain> sch,
-        CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
+        CallbackHandler* handler, utils::Invocable<void(void)>&& callback) {
 }
 
 void VulkanDriver::setPresentationTime(int64_t monotonic_clock_ns) {
@@ -386,7 +407,7 @@ void VulkanDriver::setPresentationTime(int64_t monotonic_clock_ns) {
 void VulkanDriver::endFrame(uint32_t frameId) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("endframe");
-    mCommands->flush();
+    mCommands.flush();
     collectGarbage();
     FVK_SYSTRACE_END();
 }
@@ -394,7 +415,7 @@ void VulkanDriver::endFrame(uint32_t frameId) {
 void VulkanDriver::flush(int) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("flush");
-    mCommands->flush();
+    mCommands.flush();
     FVK_SYSTRACE_END();
 }
 
@@ -402,8 +423,8 @@ void VulkanDriver::finish(int dummy) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("finish");
 
-    mCommands->flush();
-    mCommands->wait();
+    mCommands.flush();
+    mCommands.wait();
 
     mReadPixels.runUntilComplete();
     FVK_SYSTRACE_END();
@@ -490,9 +511,6 @@ 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);
 }
 
@@ -500,7 +518,7 @@ void VulkanDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
         TextureUsage usage) {
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, mCommands.get(), target, levels,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
             format, samples, w, h, depth, usage, mStagePool);
     mResourceManager.acquire(vktexture);
 }
@@ -512,7 +530,7 @@ void VulkanDriver::createTextureSwizzledR(Handle<HwTexture> th, SamplerType targ
     TextureSwizzle swizzleArray[] = {r, g, b, a};
     const VkComponentMapping swizzleMap = getSwizzleMap(swizzleArray);
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, mCommands.get(), target, levels,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
             format, samples, w, h, depth, usage, mStagePool, false /*heap allocated */, swizzleMap);
     mResourceManager.acquire(vktexture);
 }
@@ -543,6 +561,7 @@ void VulkanDriver::destroyProgram(Handle<HwProgram> ph) {
         return;
     }
     auto vkprogram = mResourceAllocator.handle_cast<VulkanProgram*>(ph);
+    mDescriptorSetManager.clearProgram(vkprogram);
     mResourceManager.release(vkprogram);
 }
 
@@ -555,7 +574,7 @@ void VulkanDriver::createDefaultRenderTargetR(Handle<HwRenderTarget> rth, int) {
 
 void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         TargetBufferFlags targets, uint32_t width, uint32_t height, uint8_t samples,
-        MRT color, TargetBufferInfo depth, TargetBufferInfo stencil) {
+        uint8_t layerCount, MRT color, TargetBufferInfo depth, TargetBufferInfo stencil) {
     UTILS_UNUSED_IN_RELEASE math::vec2<uint32_t> tmin = {std::numeric_limits<uint32_t>::max()};
     UTILS_UNUSED_IN_RELEASE math::vec2<uint32_t> tmax = {0};
     UTILS_UNUSED_IN_RELEASE size_t attachmentCount = 0;
@@ -607,7 +626,7 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
     assert_invariant(tmin.x >= width && tmin.y >= height);
 
     auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, mCommands.get(), width, height,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, width, height,
             samples, colorTargets, depthStencil, mStagePool);
     mResourceManager.acquire(renderTarget);
 }
@@ -625,7 +644,7 @@ void VulkanDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
 }
 
 void VulkanDriver::createFenceR(Handle<HwFence> fh, int) {
-    VulkanCommandBuffer const& commandBuffer = mCommands->get();
+    VulkanCommandBuffer const& commandBuffer = mCommands.get();
     mResourceAllocator.construct<VulkanFence>(fh, commandBuffer.fence);
 }
 
@@ -636,7 +655,7 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
         flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
     }
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, mCommands.get(), mStagePool, nativeWindow, flags);
+            mAllocator, &mCommands, mStagePool, nativeWindow, flags);
     mResourceManager.acquire(swapChain);
 }
 
@@ -649,7 +668,7 @@ void VulkanDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t wi
     }
     assert_invariant(width > 0 && height > 0 && "Vulkan requires non-zero swap chain dimensions.");
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, mCommands.get(), mStagePool, nullptr, flags, VkExtent2D{width, height});
+            mAllocator, &mCommands, mStagePool, nullptr, flags, VkExtent2D{width, height});
     mResourceManager.acquire(swapChain);
 }
 
@@ -899,6 +918,11 @@ bool VulkanDriver::isDepthStencilResolveSupported() {
     return false;
 }
 
+bool VulkanDriver::isDepthStencilBlitSupported(TextureFormat format) {
+    auto const& formats = mContext.getBlittableDepthStencilFormats();
+    return std::find(formats.begin(), formats.end(), getVkFormat(format)) != formats.end();
+}
+
 bool VulkanDriver::isProtectedTexturesSupported() {
     return false;
 }
@@ -994,7 +1018,7 @@ void VulkanDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t in
 
 void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&& p,
         uint32_t byteOffset) {
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     auto ib = mResourceAllocator.handle_cast<VulkanIndexBuffer*>(ibh);
     commands.acquire(ib);
     ib->buffer.loadFromCpu(commands.buffer(), p.buffer, byteOffset, p.size);
@@ -1004,7 +1028,7 @@ void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor
 
 void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& bd,
         uint32_t byteOffset) {
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
 
     auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
     commands.acquire(bo);
@@ -1015,7 +1039,7 @@ void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescript
 
 void VulkanDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> boh,
         BufferDescriptor&& bd, uint32_t byteOffset) {
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
     commands.acquire(bo);
     // TODO: implement unsynchronized version
@@ -1180,7 +1204,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     // If that's the case, we need to change the layout of the texture to DEPTH_SAMPLER, which is a
     // more general layout. Otherwise, we prefer the DEPTH_ATTACHMENT layout, which is optimal for
     // the non-sampling case.
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer const cmdbuffer = commands.buffer();
 
     UTILS_NOUNROLL
@@ -1212,12 +1236,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         }
     }
 
-    VulkanLayout const currentDepthLayout = depth.getLayout();
-    VulkanLayout const renderPassDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
-    // We need to keep the final layout as an attachment because the implicit transition does not
-    // have any barrier guarrantees, meaning that if we want to sample from the output in the next
-    // pass, then we'd have a race-condition/validation error.
-    VulkanLayout const finalDepthLayout = renderPassDepthLayout;
+    VulkanLayout currentDepthLayout = depth.getLayout();
 
     TargetBufferFlags clearVal = params.flags.clear;
     TargetBufferFlags discardEndVal = params.flags.discardEnd;
@@ -1226,16 +1245,20 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
             discardEndVal &= ~TargetBufferFlags::DEPTH;
             clearVal &= ~TargetBufferFlags::DEPTH;
         }
-        auto const attachmentSubresourceRange = depth.getSubresourceRange(VK_IMAGE_ASPECT_DEPTH_BIT);
-        depth.texture->setLayout(attachmentSubresourceRange, VulkanLayout::DEPTH_ATTACHMENT);
+        // If the depth attachment texture was previously sampled, then we need to manually
+        // transition it to an attachment. This is necessary to also set up a barrier between the
+        // previous read and the potentially coming write.
+        if (currentDepthLayout == VulkanLayout::DEPTH_SAMPLER) {
+            depth.texture->transitionLayout(cmdbuffer, depth.getSubresourceRange(),
+                    VulkanLayout::DEPTH_ATTACHMENT);
+            currentDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
+        }
     }
 
     // Create the VkRenderPass or fetch it from cache.
     VulkanFboCache::RenderPassKey rpkey = {
         .initialColorLayoutMask = 0,
         .initialDepthLayout = currentDepthLayout,
-        .renderPassDepthLayout = renderPassDepthLayout,
-        .finalDepthLayout = finalDepthLayout,
         .depthFormat = depth.getFormat(),
         .clear = clearVal,
         .discardStart = discardStart,
@@ -1252,9 +1275,9 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
                 rpkey.needsResolveMask |= (1 << i);
             }
             if (info.texture->getPrimaryImageLayout() != VulkanLayout::COLOR_ATTACHMENT) {
-                ((VulkanTexture*) info.texture)->transitionLayout(cmdbuffer,
-                        info.getSubresourceRange(VK_IMAGE_ASPECT_COLOR_BIT),
-                        VulkanLayout::COLOR_ATTACHMENT);
+                ((VulkanTexture*) info.texture)
+                        ->transitionLayout(cmdbuffer, info.getSubresourceRange(),
+                                VulkanLayout::COLOR_ATTACHMENT);
             }
         } else {
             rpkey.colorFormat[i] = VK_FORMAT_UNDEFINED;
@@ -1272,27 +1295,42 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         .layers = 1,
         .samples = rpkey.samples,
     };
+    auto& renderPassAttachments = mRenderPassFboInfo.attachments;
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
         if (!rt->getColor(i).texture) {
             fbkey.color[i] = VK_NULL_HANDLE;
             fbkey.resolve[i] = VK_NULL_HANDLE;
         } else if (fbkey.samples == 1) {
-            fbkey.color[i] = rt->getColor(i).getImageView(VK_IMAGE_ASPECT_COLOR_BIT);
+            auto& colorAttachment = rt->getColor(i);
+            renderPassAttachments.insert(colorAttachment);
+            fbkey.color[i] = colorAttachment.getImageView();
             fbkey.resolve[i] = VK_NULL_HANDLE;
             assert_invariant(fbkey.color[i]);
         } else {
-            fbkey.color[i] = rt->getMsaaColor(i).getImageView(VK_IMAGE_ASPECT_COLOR_BIT);
-            VulkanTexture* texture = (VulkanTexture*) rt->getColor(i).texture;
+            auto& msaaColorAttachment = rt->getMsaaColor(i);
+            renderPassAttachments.insert(msaaColorAttachment);
+
+            auto& colorAttachment = rt->getColor(i);
+            fbkey.color[i] = msaaColorAttachment.getImageView();
+
+            VulkanTexture* texture = colorAttachment.texture;
             if (texture->samples == 1) {
-                fbkey.resolve[i] = rt->getColor(i).getImageView(VK_IMAGE_ASPECT_COLOR_BIT);
+                mRenderPassFboInfo.hasColorResolve = true;
+
+                renderPassAttachments.insert(colorAttachment);
+                fbkey.resolve[i] = colorAttachment.getImageView();
                 assert_invariant(fbkey.resolve[i]);
             }
             assert_invariant(fbkey.color[i]);
         }
     }
     if (depth.texture) {
-        fbkey.depth = depth.getImageView(VK_IMAGE_ASPECT_DEPTH_BIT);
+        fbkey.depth = depth.getImageView();
         assert_invariant(fbkey.depth);
+        renderPassAttachments.insert(depth);
+
+        UTILS_UNUSED_IN_RELEASE bool const depthDiscardEnd =
+                any(rpkey.discardEnd & TargetBufferFlags::DEPTH);
 
         // Vulkan 1.1 does not support multisampled depth resolve, so let's check here
         // and assert if this is requested. (c.f. isAutoDepthResolveSupported)
@@ -1301,29 +1339,23 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         // - If the RT is MS then all SS attachments are auto resolved if not discarded.
         assert_invariant(!(rt->getSamples() > 1 &&
                 rt->getDepth().texture->samples == 1 &&
-                !any(rpkey.discardEnd & TargetBufferFlags::DEPTH)));
+                !depthDiscardEnd));
     }
     VkFramebuffer vkfb = mFramebufferCache.getFramebuffer(fbkey);
 
 // Assign a label to the framebuffer for debugging purposes.
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS | FVK_DEBUG_DEBUG_UTILS)
-    auto const topMarker = mCommands->getTopGroupMarker();
+    auto const topMarker = mCommands.getTopGroupMarker();
     if (!topMarker.empty()) {
         DebugUtils::setName(VK_OBJECT_TYPE_FRAMEBUFFER, reinterpret_cast<uint64_t>(vkfb),
                 topMarker.c_str());
     }
 #endif
 
-    // The current command buffer now owns a reference to the render target and its attachments.
-    // Note that we must acquire parent textures, not sidecars.
+    // The current command buffer now has references to the render target and its attachments.
     commands.acquire(rt);
-    if (depth.texture) {
-        commands.acquire((VulkanTexture*) depth.texture);
-    }
-    for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
-        if (rt->getColor(i).texture) {
-            commands.acquire(rt->getColor(i).texture);
-        }
+    for (auto const& attachment: renderPassAttachments) {
+        commands.acquire(attachment.texture);
     }
 
     // Populate the structures required for vkCmdBeginRenderPass.
@@ -1396,7 +1428,7 @@ void VulkanDriver::endRenderPass(int) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("endRenderPass");
 
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer cmdbuffer = commands.buffer();
     vkCmdEndRenderPass(cmdbuffer);
 
@@ -1409,33 +1441,52 @@ void VulkanDriver::endRenderPass(int) {
     // issue several of them when considering MRT. This would be very complex to set up and would
     // require more state tracking, so we've chosen to use a memory barrier for simplicity and
     // correctness.
-
-    // NOTE: ideally dstStageMask would merely be VERTEX_SHADER_BIT | FRAGMENT_SHADER_BIT, but this
-    // seems to be insufficient on Mali devices. To work around this we are adding a more aggressive
-    // TOP_OF_PIPE barrier.
     if (!rt->isSwapChain()) {
-        VkMemoryBarrier barrier{
-                .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
-                .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-                .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
-        };
-        VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
-        if (rt->hasDepth()) {
-            barrier.srcAccessMask |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
-            srcStageMask |= VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
+        for (auto const& attachment: mRenderPassFboInfo.attachments) {
+            bool const isDepth = attachment.isDepth();
+            VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+
+            // This is a workaround around a validation issue (might not be an actual driver issue).
+            if (mRenderPassFboInfo.hasColorResolve && !isDepth) {
+                srcStageMask = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
+            }
+
+            VkPipelineStageFlags dstStageMask =
+                    VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
+            VkAccessFlags srcAccess = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+            VkAccessFlags dstAccess = VK_ACCESS_SHADER_READ_BIT;
+            VulkanLayout layout = VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT;
+            if (isDepth) {
+                srcAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+                dstAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
+                srcStageMask = VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
+                dstStageMask = VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
+                layout =  VulkanFboCache::FINAL_DEPTH_ATTACHMENT_LAYOUT;
+            }
+
+            auto const vkLayout = imgutil::getVkLayout(layout);
+            auto const& range = attachment.getSubresourceRange();
+            VkImageMemoryBarrier barrier = {
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = srcAccess,
+                .dstAccessMask = dstAccess,
+                .oldLayout = vkLayout,
+                .newLayout = vkLayout,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = attachment.getImage(),
+                .subresourceRange = range,
+            };
+
+            attachment.texture->setLayout(range, layout);
+            vkCmdPipelineBarrier(cmdbuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr,
+                    1, &barrier);
         }
-        vkCmdPipelineBarrier(cmdbuffer, srcStageMask,
-                VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT | // <== For Mali
-                VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
-                0, 1, &barrier, 0, nullptr, 0, nullptr);
     }
 
-    if (mCurrentRenderPass.currentSubpass > 0) {
-        for (uint32_t i = 0; i < VulkanPipelineCache::INPUT_ATTACHMENT_COUNT; i++) {
-            mPipelineCache.bindInputAttachment(i, {});
-        }
-        mCurrentRenderPass.currentSubpass = 0;
-    }
+    mRenderPassFboInfo.clear();
+    mDescriptorSetManager.clearState();
     mCurrentRenderPass.renderTarget = nullptr;
     mCurrentRenderPass.renderPass = VK_NULL_HANDLE;
     FVK_SYSTRACE_END();
@@ -1449,20 +1500,14 @@ void VulkanDriver::nextSubpass(int) {
     assert_invariant(renderTarget);
     assert_invariant(mCurrentRenderPass.params.subpassMask);
 
-    vkCmdNextSubpass(mCommands->get().buffer(), VK_SUBPASS_CONTENTS_INLINE);
+    vkCmdNextSubpass(mCommands.get().buffer(), VK_SUBPASS_CONTENTS_INLINE);
 
     mPipelineCache.bindRenderPass(mCurrentRenderPass.renderPass,
             ++mCurrentRenderPass.currentSubpass);
 
-    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);
-        }
+    if (mCurrentRenderPass.params.subpassMask & 0x1) {
+        VulkanAttachment subpassInput = renderTarget->getColor(0);
+        mDescriptorSetManager.updateInputAttachment({}, subpassInput);
     }
 }
 
@@ -1502,25 +1547,24 @@ void VulkanDriver::commit(Handle<HwSwapChain> sch) {
 
 void VulkanDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> boh) {
     auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    const VkDeviceSize offset = 0;
-    const VkDeviceSize size = VK_WHOLE_SIZE;
-    mPipelineCache.bindUniformBufferObject((uint32_t) index, bo, offset, size);
+    VkDeviceSize const offset = 0;
+    VkDeviceSize const size = VK_WHOLE_SIZE;
+    mDescriptorSetManager.updateBuffer({}, (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::SHADER_STORAGE ||
-                     bindingType == BufferObjectBinding::UNIFORM);
+    assert_invariant(bindingType == BufferObjectBinding::UNIFORM);
 
     // TODO: implement BufferObjectBinding::SHADER_STORAGE case
 
     auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    mPipelineCache.bindUniformBufferObject((uint32_t)index, bo, offset, size);
+    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
 }
 
 void VulkanDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
-    // TODO: implement unbindBuffer()
+    mDescriptorSetManager.clearBuffer((uint32_t) index);
 }
 
 void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
@@ -1528,16 +1572,24 @@ void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
     mSamplerBindings[index] = hwsb;
 }
 
+void VulkanDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
+        backend::PushConstantVariant value) {
+    assert_invariant(mBoundPipeline.program && "Expect a program when writing to push constants");
+    VulkanCommands* commands = &mCommands;
+    mBoundPipeline.program->writePushConstant(commands, mBoundPipeline.pipelineLayout, stage, index,
+            value);
+}
+
 void VulkanDriver::insertEventMarker(char const* string, uint32_t len) {
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands->insertEventMarker(string, len);
+    mCommands.insertEventMarker(string, len);
 #endif
 }
 
 void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
     // Turns out all the markers are 0-terminated, so we can just pass it without len.
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands->pushGroupMarker(string);
+    mCommands.pushGroupMarker(string);
 #endif
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START(string);
@@ -1545,7 +1597,7 @@ void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
 
 void VulkanDriver::popGroupMarker(int) {
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands->popGroupMarker();
+    mCommands.popGroupMarker();
 #endif
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_END();
@@ -1558,7 +1610,7 @@ void VulkanDriver::stopCapture(int) {}
 void VulkanDriver::readPixels(Handle<HwRenderTarget> src, uint32_t x, uint32_t y, uint32_t width,
         uint32_t height, PixelBufferDescriptor&& pbd) {
     VulkanRenderTarget* srcTarget = mResourceAllocator.handle_cast<VulkanRenderTarget*>(src);
-    mCommands->flush();
+    mCommands.flush();
     mReadPixels.run(
             srcTarget, x, y, width, height, mPlatform->getGraphicsQueueFamilyIndex(),
             std::move(pbd),
@@ -1711,11 +1763,11 @@ void VulkanDriver::blitDEPRECATED(TargetBufferFlags buffers,
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::bindPipeline(PipelineState pipelineState) {
+void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("draw");
 
-    VulkanCommandBuffer* commands = &mCommands->get();
+    VulkanCommandBuffer* commands = &mCommands.get();
     const VulkanVertexBufferInfo& vbi =
             *mResourceAllocator.handle_cast<VulkanVertexBufferInfo*>(pipelineState.vertexBufferInfo);
 
@@ -1768,22 +1820,14 @@ 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
 
-    UTILS_NOUNROLL
-    for (uint8_t binding = 0; binding < VulkanPipelineCache::SAMPLER_BINDING_COUNT; binding++) {
+    for (auto binding: program->getBindings()) {
         uint16_t const indexPair = bindingToSamplerIndex[binding];
-
         if (indexPair == 0xffff) {
-            usage = VulkanPipelineCache::disableUsageFlags(binding, usage);
             continue;
         }
 
@@ -1792,68 +1836,49 @@ void VulkanDriver::bindPipeline(PipelineState pipelineState) {
 
         VulkanSamplerGroup* vksb = mSamplerBindings[samplerGroupInd];
         if (!vksb) {
-            usage = VulkanPipelineCache::disableUsageFlags(binding, usage);
             continue;
         }
         SamplerDescriptor const* boundSampler = ((SamplerDescriptor*) vksb->sb->data()) + samplerInd;
 
         if (UTILS_UNLIKELY(!boundSampler->t)) {
-            usage = VulkanPipelineCache::disableUsageFlags(binding, usage);
             continue;
         }
-
         VulkanTexture* texture = mResourceAllocator.handle_cast<VulkanTexture*>(boundSampler->t);
-        VkImageViewType const expectedType = texture->getViewType();
 
         // TODO: can this uninitialized check be checked in a higher layer?
         // This fallback path is very flaky because the dummy texture might not have
         // matching characteristics. (e.g. if the missing texture is a 3D texture)
         if (UTILS_UNLIKELY(texture->getPrimaryImageLayout() == VulkanLayout::UNDEFINED)) {
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#if FVK_ENABLED(FVK_DEBUG_TEXTURE) && FVK_ENABLED_DEBUG_SAMPLER_NAME
             utils::slog.w << "Uninitialized texture bound to '" << bindingToName[binding] << "'";
             utils::slog.w << " in material '" << program->name.c_str() << "'";
             utils::slog.w << " at binding point " << +binding << utils::io::endl;
 #endif
-            texture = mEmptyTexture.get();
+            texture = mEmptyTexture;
         }
 
-        SamplerParams const& samplerParams = boundSampler->s;
-        VkSampler const vksampler = mSamplerCache.getSampler(samplerParams);
-
+        VkSampler const vksampler = mSamplerCache.getSampler(boundSampler->s);
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
         VulkanDriver::DebugUtils::setName(VK_OBJECT_TYPE_SAMPLER,
                 reinterpret_cast<uint64_t>(vksampler), bindingToName[binding].c_str());
 #endif
-
-        VkImageView imageView = VK_NULL_HANDLE;
-        VkImageSubresourceRange const range = texture->getPrimaryViewRange();
-        if (any(texture->usage & TextureUsage::DEPTH_ATTACHMENT) &&
-                expectedType == VK_IMAGE_VIEW_TYPE_2D) {
-            // If the sampler is part of a mipmapped depth texture, where one of the level *can* be
-            // an attachment, then the sampler for this texture has the same view properties as a
-            // view for an attachment. Therefore, we can use getAttachmentView to get a
-            // corresponding VkImageView.
-            imageView = texture->getAttachmentView(range);
-        } else {
-            imageView = texture->getViewForType(range, expectedType);
-        }
-
-        samplerInfo[binding] = {
-            .sampler = vksampler,
-            .imageView = imageView,
-            .imageLayout = ImgUtil::getVkLayout(texture->getPrimaryImageLayout())
-        };
-        samplerTextures[binding] = texture;
+        mDescriptorSetManager.updateSampler({}, binding, texture, vksampler);
     }
 
-    mPipelineCache.bindSamplers(samplerInfo, samplerTextures, usage);
+    auto const pipelineLayout = mDescriptorSetManager.bind(commands, program, mGetPipelineFunction);
+    mBoundPipeline = {
+        .program = program,
+        .pipelineLayout = pipelineLayout,
+    };
+
+    mPipelineCache.bindLayout(pipelineLayout);
+    mPipelineCache.bindPipeline(commands);
+
+    // Since we don't statically define scissor as part of the pipeline, we need to call scissor at
+    // least once. Context: VUID-vkCmdDrawIndexed-None-07832.
+    auto const& extent = rt->getExtent();
+    scissor({0, 0, extent.width, extent.height});
 
-    // 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();
 }
 
@@ -1861,7 +1886,7 @@ void VulkanDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("bindRenderPrimitive");
 
-    VulkanCommandBuffer* commands = &mCommands->get();
+    VulkanCommandBuffer* commands = &mCommands.get();
     VkCommandBuffer cmdbuffer = commands->buffer();
     const VulkanRenderPrimitive& prim = *mResourceAllocator.handle_cast<VulkanRenderPrimitive*>(rph);
     commands->acquire(prim.indexBuffer);
@@ -1890,15 +1915,11 @@ void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("draw2");
 
-    VulkanCommandBuffer* commands = &mCommands->get();
-    VkCommandBuffer cmdbuffer = commands->buffer();
+    VulkanCommandBuffer& commands = mCommands.get();
+    VkCommandBuffer cmdbuffer = commands.buffer();
 
-    // 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;
-    }
+    // Bind "dynamic" UBOs if they need to change.
+    mDescriptorSetManager.dynamicBind(&commands, {});
 
     // Finally, make the actual draw call. TODO: support subranges
     const uint32_t firstIndex = indexOffset;
@@ -1925,8 +1946,8 @@ void VulkanDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGr
 }
 
 void VulkanDriver::scissor(Viewport scissorBox) {
-    VulkanCommandBuffer* commands = &mCommands->get();
-    VkCommandBuffer cmdbuffer = commands->buffer();
+    VulkanCommandBuffer& commands = mCommands.get();
+    VkCommandBuffer cmdbuffer = commands.buffer();
 
     // Set scissoring.
     // clamp left-bottom to 0,0 and avoid overflows
@@ -1948,17 +1969,17 @@ void VulkanDriver::scissor(Viewport scissorBox) {
 
     const VulkanRenderTarget* rt = mCurrentRenderPass.renderTarget;
     rt->transformClientRectToPlatform(&scissor);
-    mPipelineCache.bindScissor(cmdbuffer, scissor);
+    vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
 }
 
 void VulkanDriver::beginTimerQuery(Handle<HwTimerQuery> tqh) {
     VulkanTimerQuery* vtq = mResourceAllocator.handle_cast<VulkanTimerQuery*>(tqh);
-    mTimestamps->beginQuery(&(mCommands->get()), vtq);
+    mTimestamps->beginQuery(&(mCommands.get()), vtq);
 }
 
 void VulkanDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
     VulkanTimerQuery* vtq = mResourceAllocator.handle_cast<VulkanTimerQuery*>(tqh);
-    mTimestamps->endQuery(&(mCommands->get()), vtq);
+    mTimestamps->endQuery(&(mCommands.get()), vtq);
 }
 
 void VulkanDriver::debugCommandBegin(CommandStream* cmds, bool synchronous, const char* methodName) noexcept {