Release Filament 1.51.8

- a field was added, which broke the layout of the structure. We fix it
by adding constructors which will handle the old and new way of
initializing this structure.

- one of the test needed a hash update

- OpenGLContext wrongly asserted when trying to unbind texture 0

CMakeLists.txt

@@ -71,6 +71,10 @@ 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)
 
@@ -169,15 +173,6 @@ 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 ${CMAKE_SHARED_LIBRARY_CREATE_CXX_FLAGS})
-endif()
-
 if (WIN32)
     # Link statically against c/c++ lib to avoid missing redistriburable such as
     # "VCRUNTIME140.dll not found. Try reinstalling the app.", but give users
@@ -389,8 +384,9 @@ endif()
 if (NOT MSVC AND NOT IOS)
     # Omitting stack frame pointers prevents the generation of readable stack traces in crash reports on iOS
     set(CMAKE_CXX_FLAGS_RELEASE "${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()
 
@@ -548,6 +544,12 @@ 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=${FILAMENT_BACKEND_DEBUG_FLAG})
+    unset(FILAMENT_BACKEND_DEBUG_FLAG)
+endif()
+
 # ==================================================================================================
 # Material compilation flags
 # ==================================================================================================
@@ -819,6 +821,9 @@ if (IS_HOST_PLATFORM)
     add_subdirectory(${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)

NEW_RELEASE_NOTES.md

@@ -7,5 +7,3 @@ for next branch cut* header.
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).
 
 ## Release notes for next branch cut
-
-- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.51.0'
+    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', '~> 1.51.0'
+pod 'Filament', '~> 1.51.8'
 ```
 
 ### Snapshots

RELEASE_NOTES.md

@@ -7,6 +7,41 @@ 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
 
 

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 +406,13 @@ 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) {
@@ -495,7 +517,9 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
         jlong textureUseAfterFreePoolSize, jboolean disableParallelShaderCompile,
         jint stereoscopicType, jlong stereoscopicEyeCount,
         jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge,
-        jboolean disableHandleUseAfterFreeCheck) {
+        jboolean disableHandleUseAfterFreeCheck,
+        jint preferredShaderLanguage,
+        jboolean forceGLES2Context) {
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
     Engine::Config config = {
             .commandBufferSizeMB = (uint32_t) commandBufferSizeMB,
@@ -511,6 +535,8 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
             .resourceAllocatorCacheSizeMB = (uint32_t) resourceAllocatorCacheSizeMB,
             .resourceAllocatorCacheMaxAge = (uint8_t) resourceAllocatorCacheMaxAge,
             .disableHandleUseAfterFreeCheck = (bool) disableHandleUseAfterFreeCheck,
+            .preferredShaderLanguage = (Engine::Config::ShaderLanguage) preferredShaderLanguage,
+            .forceGLES2Context = (bool) forceGLES2Context,
     };
     builder->config(&config);
 }

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

@@ -225,7 +225,9 @@ public class Engine {
                     config.textureUseAfterFreePoolSize, config.disableParallelShaderCompile,
                     config.stereoscopicType.ordinal(), config.stereoscopicEyeCount,
                     config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge,
-                    config.disableHandleUseAfterFreeCheck);
+                    config.disableHandleUseAfterFreeCheck,
+                    config.preferredShaderLanguage.ordinal(),
+                    config.forceGLES2Context);
             return this;
         }
 
@@ -428,6 +430,35 @@ public class Engine {
          * 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) {
@@ -849,6 +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
@@ -1209,6 +1258,17 @@ 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.
      *
@@ -1291,6 +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);
@@ -1299,6 +1361,7 @@ 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);
@@ -1321,7 +1384,9 @@ public class Engine {
             long textureUseAfterFreePoolSize, boolean disableParallelShaderCompile,
             int stereoscopicType, long stereoscopicEyeCount,
             long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge,
-            boolean disableHandleUseAfterFreeCheck);
+            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);

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/gltfio-android/CMakeLists.txt

@@ -86,7 +86,10 @@ set(GLTFIO_SRCS
         ${GLTFIO_DIR}/src/Wireframe.cpp
         ${GLTFIO_DIR}/src/Wireframe.h
         ${GLTFIO_DIR}/src/downcast.h
+        ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.cpp
         ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.h
+        ${GLTFIO_DIR}/src/extended/ResourceLoaderExtended.cpp
+        ${GLTFIO_DIR}/src/extended/ResourceLoaderExtended.h
         ${GLTFIO_DIR}/src/extended/TangentsJobExtended.cpp
         ${GLTFIO_DIR}/src/extended/TangentsJobExtended.h
         ${GLTFIO_DIR}/src/extended/TangentSpaceMeshWrapper.cpp

android/gradle.properties

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

@@ -61,6 +61,11 @@ 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"
@@ -172,6 +177,8 @@ MATOPT_GRADLE_OPTION=""
 
 ASAN_UBSAN_OPTION=""
 
+BACKEND_DEBUG_FLAG_OPTION=""
+
 IOS_BUILD_SIMULATOR=false
 BUILD_UNIVERSAL_LIBRARIES=false
 
@@ -231,6 +238,7 @@ function build_desktop_target {
             ${MATDBG_OPTION} \
             ${MATOPT_OPTION} \
             ${ASAN_UBSAN_OPTION} \
+            ${BACKEND_DEBUG_FLAG_OPTION} \
             ${architectures} \
             ../..
         ln -sf "out/cmake-${lc_target}/compile_commands.json" \
@@ -289,6 +297,7 @@ 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 +372,7 @@ 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
@@ -597,6 +607,7 @@ 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}-${arch}/compile_commands.json" \
            ../../compile_commands.json
@@ -730,6 +741,13 @@ function validate_build_command {
             exit 1
         fi
     fi
+
+    # Make sure FILAMENT_BACKEND_DEBUG_FLAG is only meant for debug builds
+    if [[ "${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 +794,7 @@ function check_debug_release_build {
 
 pushd "$(dirname "$0")" > /dev/null
 
-while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
+while getopts ":hacCfgijmp:q:uvslwtedk:bx:" opt; do
     case ${opt} in
         h)
             print_help
@@ -840,7 +858,7 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
                         echo "Platform must be one of [desktop|android|ios|webgl|all]"
                         echo ""
                         exit 1
-                    ;;    
+                    ;;
                 esac
             done
             ;;
@@ -918,6 +936,8 @@ while getopts ":hacCfgijmp:q:uvslwtedk:b" opt; do
         b)  ASAN_UBSAN_OPTION="-DFILAMENT_ENABLE_ASAN_UBSAN=ON"
             echo "Enabled ASAN/UBSAN"
             ;;
+        x)  BACKEND_DEBUG_FLAG_OPTION="-DFILAMENT_BACKEND_DEBUG_FLAG=${OPTARG}"
+            ;;
         \?)
             echo "Invalid option: -${OPTARG}" >&2
             echo ""

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

filament/backend/CMakeLists.txt

@@ -71,6 +71,8 @@ if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3 AND NOT FILAMEN
             include/backend/platforms/OpenGLPlatform.h
             src/opengl/gl_headers.cpp
             src/opengl/gl_headers.h
+            src/opengl/GLBufferObject.h
+            src/opengl/GLTexture.h
             src/opengl/GLUtils.cpp
             src/opengl/GLUtils.h
             src/opengl/OpenGLBlobCache.cpp
@@ -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
@@ -474,6 +478,10 @@ 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

@@ -65,6 +65,12 @@ public:
          * 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;

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.
@@ -114,6 +117,14 @@ 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;
@@ -136,6 +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;
     }
@@ -143,6 +156,15 @@ 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; }
@@ -155,10 +177,12 @@ 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;

filament/backend/include/backend/TargetBufferInfo.h

@@ -29,17 +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
+    // 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;
 };
 
@@ -64,7 +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 } {
     }
 
@@ -84,7 +103,7 @@ public:
 
     // this is here for backward compatibility
     MRT(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
-            : mInfos{{ handle, 0, level, layer }} {
+            : mInfos{{ handle, level, layer, 0 }} {
     }
 };
 

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

@@ -140,6 +140,23 @@ public:
      */
     virtual uint32_t getDefaultFramebufferObject() noexcept;
 
+    /**
+     * Called by the backend when a frame starts.
+     * @param steady_clock_ns vsync time point on the monotonic clock
+     * @param refreshIntervalNs refresh interval in nanosecond
+     * @param frameId a frame id
+     */
+    virtual void beginFrame(
+            int64_t monotonic_clock_ns,
+            int64_t refreshIntervalNs,
+            uint32_t frameId) noexcept;
+
+    /**
+     * Called by the backend when a frame ends.
+     * @param frameId the frame id used in beginFrame
+     */
+    virtual void endFrame(
+            uint32_t frameId) noexcept;
 
     /**
      * Type of contexts available
@@ -191,6 +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/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/private/backend/CommandBufferQueue.h

@@ -82,6 +82,7 @@ public:
     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

@@ -133,12 +133,13 @@ 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,
@@ -303,6 +304,7 @@ 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)
@@ -432,6 +434,11 @@ 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)

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

@@ -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);
@@ -163,7 +163,7 @@ public:
     inline typename std::enable_if_t<
             std::is_pointer_v<Dp> &&
             std::is_base_of_v<B, typename std::remove_pointer_t<Dp>>, Dp>
-    handle_cast(Handle<B>& handle) noexcept {
+    handle_cast(Handle<B>& handle) {
         assert_invariant(handle);
         auto [p, tag] = handleToPointer(handle.getId());
 
@@ -185,7 +185,7 @@ public:
     inline typename std::enable_if_t<
             std::is_pointer_v<Dp> &&
             std::is_base_of_v<B, typename std::remove_pointer_t<Dp>>, Dp>
-    handle_cast(Handle<B> const& handle) noexcept {
+    handle_cast(Handle<B> const& handle) {
         return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
     }
 
@@ -317,7 +317,7 @@ private:
         return (id & HANDLE_HEAP_FLAG) == 0u;
     }
 
-    HandleBase::HandleId allocateHandleSlow(size_t size) 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

filament/backend/src/CommandBufferQueue.cpp

@@ -57,6 +57,11 @@ 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) {

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

@@ -107,7 +107,7 @@ void* HandleAllocator<P0, P1, P2>::handleToPointerSlow(HandleBase::HandleId id)
 }
 
 template <size_t P0, size_t P1, size_t P2>
-HandleBase::HandleId HandleAllocator<P0, P1, P2>::allocateHandleSlow(size_t size) noexcept {
+HandleBase::HandleId HandleAllocator<P0, P1, P2>::allocateHandleSlow(size_t size) {
     void* p = ::malloc(size);
     std::unique_lock lock(mLock);
 

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) {
@@ -86,6 +91,12 @@ 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;

filament/backend/src/metal/MetalBuffer.h

@@ -18,6 +18,7 @@
 #define TNT_FILAMENT_DRIVER_METALBUFFER_H
 
 #include "MetalContext.h"
+#include "MetalPlatform.h"
 
 #include <backend/DriverEnums.h>
 
@@ -28,47 +29,122 @@
 #include <utility>
 #include <memory>
 #include <atomic>
+#include <chrono>
 
 namespace filament::backend {
 
-class TrackedMetalBuffer {
+class ScopedAllocationTimer {
 public:
-    TrackedMetalBuffer() noexcept : mBuffer(nil) {}
-    TrackedMetalBuffer(id<MTLBuffer> buffer) noexcept : mBuffer(buffer) {
-        if (buffer) {
-            aliveBuffers++;
-        }
-    }
-    ~TrackedMetalBuffer() {
-        if (mBuffer) {
-            aliveBuffers--;
+    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()));
+            }
         }
     }
 
-    TrackedMetalBuffer(TrackedMetalBuffer&&) = delete;
-    TrackedMetalBuffer(TrackedMetalBuffer const&) = delete;
-    TrackedMetalBuffer& operator=(TrackedMetalBuffer const&) = delete;
-
-    TrackedMetalBuffer& operator=(TrackedMetalBuffer&& rhs) noexcept {
-        swap(rhs);
-        return *this;
-    }
-
-    id<MTLBuffer> get() const noexcept { return mBuffer; }
-    operator bool() const noexcept { return bool(mBuffer); }
-
-    static uint64_t getAliveBuffers() { return aliveBuffers; }
+    static void setPlatform(MetalPlatform* p) { platform = p; }
 
 private:
-    void swap(TrackedMetalBuffer& other) noexcept {
-        id<MTLBuffer> temp = mBuffer;
-        mBuffer = other.mBuffer;
-        other.mBuffer = temp;
+    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;
+        }
     }
 
-    id<MTLBuffer> mBuffer;
+#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 std::atomic<uint64_t> aliveBuffers;
+    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 {
@@ -122,7 +198,7 @@ public:
 
 private:
 
-    TrackedMetalBuffer mBuffer;
+    id<MTLBuffer> mBuffer;
     size_t mBufferSize = 0;
     void* mCpuBuffer = nullptr;
     MetalContext& mContext;
@@ -171,7 +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);
     }
 
@@ -189,9 +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.get(), 0};
+            return { mAuxBuffer, 0 };
         }
         mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
         mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);
@@ -220,9 +304,9 @@ public:
      */
     std::pair<id<MTLBuffer>, NSUInteger> getCurrentAllocation() const {
         if (UTILS_UNLIKELY(mAuxBuffer)) {
-            return { mAuxBuffer.get(), 0 };
+            return { mAuxBuffer, 0 };
         }
-        return { mBuffer.get(), mCurrentSlot * mSlotSizeBytes };
+        return { mBuffer, mCurrentSlot * mSlotSizeBytes };
     }
 
     bool canAccomodateLayout(MTLSizeAndAlign layout) const {
@@ -231,8 +315,8 @@ public:
 
 private:
     id<MTLDevice> mDevice;
-    TrackedMetalBuffer mBuffer;
-    TrackedMetalBuffer mAuxBuffer;
+    id<MTLBuffer> mBuffer;
+    id<MTLBuffer> mAuxBuffer;
 
     MTLResourceOptions mBufferOptions;
 

filament/backend/src/metal/MetalBuffer.mm

@@ -22,7 +22,13 @@
 namespace filament {
 namespace backend {
 
-std::atomic<uint64_t> TrackedMetalBuffer::aliveBuffers = 0;
+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) {
@@ -37,7 +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);
 }
 
@@ -64,7 +74,7 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     // Acquire a staging buffer to hold the contents of this update.
     MetalBufferPool* bufferPool = mContext.bufferPool;
     const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
-    memcpy(staging->buffer.get().contents, src, size);
+    memcpy(staging->buffer.contents, src, size);
 
     // The blit below requires that byteOffset be a multiple of 4.
     ASSERT_PRECONDITION(!(byteOffset & 0x3u), "byteOffset must be a multiple of 4");
@@ -73,9 +83,9 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
     id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
     blitEncoder.label = @"Buffer upload blit";
-    [blitEncoder copyFromBuffer:staging->buffer.get()
+    [blitEncoder copyFromBuffer:staging->buffer
                    sourceOffset:0
-                       toBuffer:mBuffer.get()
+                       toBuffer:mBuffer
               destinationOffset:byteOffset
                            size:size];
     [blitEncoder endEncoding];
@@ -96,7 +106,7 @@ id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) n
         return nil;
     }
     assert_invariant(mBuffer);
-    return mBuffer.get();
+    return mBuffer;
 }
 
 void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncoder> encoder,

filament/backend/src/metal/MetalBufferPool.h

@@ -32,7 +32,7 @@ struct MetalContext;
 
 // Immutable POD representing a shared CPU-GPU buffer.
 struct MetalBufferPoolEntry {
-    TrackedMetalBuffer buffer;
+    id<MTLBuffer> buffer;
     size_t capacity;
     mutable uint64_t lastAccessed;
     mutable uint32_t referenceCount;

filament/backend/src/metal/MetalBufferPool.mm

@@ -42,8 +42,13 @@ 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 {
         .buffer = buffer,

filament/backend/src/metal/MetalDriver.mm

@@ -105,6 +105,10 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
                   driverConfig.disableHandleUseAfterFreeCheck) {
     mContext->driver = this;
 
+    ScopedAllocationTimer::setPlatform(platform);
+    MetalBufferTracking::initialize();
+    MetalBufferTracking::setPlatform(platform);
+
     mContext->device = mPlatform.createDevice();
     assert_invariant(mContext->device);
 
@@ -198,6 +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);
@@ -212,19 +218,29 @@ 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()) {
-        mPlatform.debugUpdateStat("filament.metal.alive_buffers", TrackedMetalBuffer::getAliveBuffers());
+#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,
@@ -805,6 +821,10 @@ bool MetalDriver::isDepthStencilResolveSupported() {
     return false;
 }
 
+bool MetalDriver::isDepthStencilBlitSupported(TextureFormat format) {
+    return true;
+}
+
 bool MetalDriver::isProtectedTexturesSupported() {
     return false;
 }
@@ -1236,6 +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) {
 
 }
@@ -1634,7 +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, CallbackHandler::Callback callback, void* user);
 
     // For CAMetalLayer-backed SwapChains, presents the drawable or schedules a
     // FrameScheduledCallback.
@@ -110,14 +112,15 @@ 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;
+        FrameScheduledCallback callback = {};
+    } frameScheduled;
 
     struct {
         CallbackHandler* handler = nullptr;

filament/backend/src/metal/MetalHandles.mm

@@ -221,9 +221,10 @@ 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::move(callback);
 }
 
 void MetalSwapChain::setFrameCompletedCallback(CallbackHandler* handler,
@@ -238,7 +239,7 @@ void MetalSwapChain::present() {
         scheduleFrameCompletedCallback();
     }
     if (drawable) {
-        if (frameScheduledCallback) {
+        if (frameScheduled.callback) {
             scheduleFrameScheduledCallback();
         } else  {
             [getPendingCommandBuffer(&context) presentDrawable:drawable];
@@ -296,21 +297,38 @@ 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(FrameScheduledCallback&& callback, id<CAMetalDrawable> drawable,
+                MetalDriver* driver)
+            : f(std::move(callback)), data(PresentDrawableData::create(drawable, driver)) {}
+        FrameScheduledCallback f;
+        // PresentDrawableData* is destroyed by maybePresentAndDestroyAsync() later.
+        std::unique_ptr<PresentDrawableData> data;
+        static void func(void* user) {
+            auto* const c = reinterpret_cast<Callback*>(user);
+            PresentDrawableData* presentDrawableData = c->data.release();
+            PresentCallable presentCallable(presentDrawable, presentDrawableData);
+            c->f(presentCallable);
+            delete c;
+        }
+    };
 
-    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>(std::move(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);
     }];
 }
 
@@ -771,13 +789,13 @@ void MetalTexture::loadWithCopyBuffer(uint32_t level, uint32_t slice, MTLRegion
         PixelBufferDescriptor const& data, const PixelBufferShape& shape) {
     const size_t stagingBufferSize = shape.totalBytes;
     auto entry = context.bufferPool->acquireBuffer(stagingBufferSize);
-    memcpy(entry->buffer.get().contents,
+    memcpy(entry->buffer.contents,
             static_cast<uint8_t*>(data.buffer) + shape.sourceOffset,
             stagingBufferSize);
     id<MTLCommandBuffer> blitCommandBuffer = getPendingCommandBuffer(&context);
     id<MTLBlitCommandEncoder> blitCommandEncoder = [blitCommandBuffer blitCommandEncoder];
     blitCommandEncoder.label = @"Texture upload buffer blit";
-    [blitCommandEncoder copyFromBuffer:entry->buffer.get()
+    [blitCommandEncoder copyFromBuffer:entry->buffer
                           sourceOffset:0
                      sourceBytesPerRow:shape.bytesPerRow
                    sourceBytesPerImage:shape.bytesPerSlice

filament/backend/src/metal/MetalShaderCompiler.mm

@@ -106,26 +106,41 @@ bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
             continue;
         }
 
-        assert_invariant(source[source.size() - 1] == '\0');
-
-        // the shader string is null terminated and the length includes the null character
-        NSString* objcSource = [[NSString alloc] initWithBytes:source.data()
-                                                        length:source.size() - 1
-                                                      encoding:NSUTF8StringEncoding];
-
-        // By default, Metal uses the most recent language version.
-        MTLCompileOptions* options = [MTLCompileOptions new];
-
-        // Disable Fast Math optimizations.
-        // This ensures that operations adhere to IEEE standards for floating-point arithmetic,
-        // which is crucial for half precision floats in scenarios where fast math optimizations
-        // lead to inaccuracies, such as in handling special values like NaN or Infinity.
-        options.fastMathEnabled = NO;
-
         NSError* error = nil;
-        id<MTLLibrary> library = [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) {
@@ -154,6 +169,23 @@ bool MetalShaderCompiler::isParallelShaderCompileSupported() const 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());
         }

filament/backend/src/noop/NoopDriver.cpp

@@ -49,11 +49,12 @@ 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) {
 
 }
 
@@ -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) {
 }
 

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

@@ -63,8 +63,10 @@ bool OpenGLContext::queryOpenGLVersion(GLint* major, GLint* minor) noexcept {
 #endif
 }
 
-OpenGLContext::OpenGLContext(OpenGLPlatform& platform) noexcept
-        : mPlatform(platform) {
+OpenGLContext::OpenGLContext(OpenGLPlatform& platform,
+        Platform::DriverConfig const& driverConfig) noexcept
+        : mPlatform(platform),
+          mSamplerMap(32) {
 
     state.vao.p = &mDefaultVAO;
 
@@ -84,6 +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);
@@ -252,9 +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) {
@@ -516,8 +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") ||
@@ -868,19 +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!!!
@@ -934,6 +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

@@ -34,6 +34,8 @@
 #include <math/vec2.h>
 #include <math/vec4.h>
 
+#include <tsl/robin_map.h>
+
 #include <array>
 #include <functional>
 #include <optional>
@@ -89,9 +91,13 @@ public:
 
     static bool queryOpenGLVersion(GLint* major, GLint* minor) noexcept;
 
-    explicit OpenGLContext(OpenGLPlatform& platform) noexcept;
+    explicit OpenGLContext(OpenGLPlatform& platform,
+            Platform::DriverConfig const& driverConfig) noexcept;
+
     ~OpenGLContext() noexcept final;
 
+    void terminate() noexcept;
+
     // TimerQueryInterface ------------------------------------------------------------------------
 
     // note: OpenGLContext being final ensures (clang) these are not called through the vtable
@@ -132,7 +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;
 
@@ -144,10 +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;
@@ -306,10 +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;
@@ -321,8 +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; }
 
@@ -425,9 +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;
 
@@ -467,6 +473,29 @@ 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;
@@ -476,6 +505,11 @@ private:
     std::vector<std::function<void(OpenGLContext&)>> mDestroyWithNormalContext;
     RenderPrimitive mDefaultVAO;
     std::optional<GLuint> mDefaultFbo[2];
+    std::array<
+            std::tuple<GLuint, void const*, uint16_t>,
+            CONFIG_UNIFORM_BINDING_COUNT> mUniformBindings = {};
+    mutable tsl::robin_map<SamplerParams, GLuint,
+            SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;
 
     void bindFramebufferResolved(GLenum target, GLuint buffer) noexcept;
 
@@ -522,9 +556,6 @@ private:
             {   bugs.delay_fbo_destruction,
                     "delay_fbo_destruction",
                     ""},
-            {   bugs.disable_thread_affinity,
-                    "disable_thread_affinity",
-                    ""},
             {   bugs.force_feature_level0,
                     "force_feature_level0",
                     ""},
@@ -572,31 +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) {
@@ -744,19 +754,17 @@ void OpenGLContext::bindBufferRange(GLenum target, GLuint index, GLuint buffer,
 #endif
 }
 
-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

@@ -173,6 +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))) {
@@ -203,13 +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,
                   driverConfig.disableHandleUseAfterFreeCheck),
-          mSamplerMap(32),
-          mDriverConfig(driverConfig) {
+          mDriverConfig(driverConfig),
+          mCurrentPushConstants(new(std::nothrow) PushConstantBundle{}) {
 
     std::fill(mSamplerBindings.begin(), mSamplerBindings.end(), nullptr);
 
@@ -233,6 +237,7 @@ OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfi
 }
 
 OpenGLDriver::~OpenGLDriver() noexcept { // NOLINT(modernize-use-equals-default)
+    // this is called from the main thread. Can't call GL.
 }
 
 Dispatcher OpenGLDriver::getDispatcher() const noexcept {
@@ -263,16 +268,13 @@ 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();
 }
 
@@ -291,9 +293,42 @@ 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 {
@@ -307,7 +342,7 @@ 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);
             }
@@ -724,10 +759,19 @@ 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;
@@ -746,7 +790,6 @@ 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;
@@ -758,30 +801,23 @@ 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 = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D);
                     break;
                 case SamplerType::SAMPLER_3D:
                     t->gl.target = GL_TEXTURE_3D;
-                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_3D);
                     break;
                 case SamplerType::SAMPLER_2D_ARRAY:
                     t->gl.target = GL_TEXTURE_2D_ARRAY;
-                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
                     break;
                 case SamplerType::SAMPLER_CUBEMAP:
                     t->gl.target = GL_TEXTURE_CUBE_MAP;
-                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
                     break;
                 case SamplerType::SAMPLER_CUBEMAP_ARRAY:
                     t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
-                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
                     break;
             }
 
@@ -792,8 +828,6 @@ 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)
-                            OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
                 } else {
                     // Turn off multi-sampling for that texture. It's just not supported.
                 }
@@ -803,13 +837,6 @@ 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);
@@ -859,32 +886,24 @@ 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 = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_EXTERNAL_OES);
             break;
         case SamplerType::SAMPLER_2D:
             t->gl.target = GL_TEXTURE_2D;
-            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D);
             break;
         case SamplerType::SAMPLER_3D:
             t->gl.target = GL_TEXTURE_3D;
-            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_3D);
             break;
         case SamplerType::SAMPLER_2D_ARRAY:
             t->gl.target = GL_TEXTURE_2D_ARRAY;
-            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
             break;
         case SamplerType::SAMPLER_CUBEMAP:
             t->gl.target = GL_TEXTURE_CUBE_MAP;
-            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
             break;
         case SamplerType::SAMPLER_CUBEMAP_ARRAY:
             t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
-            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
             break;
     }
 
@@ -895,7 +914,6 @@ 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 = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
         } else {
             // Turn off multi-sampling for that texture. It's just not supported.
         }
@@ -1785,7 +1803,6 @@ 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);
                     }
                 }
@@ -2047,6 +2064,10 @@ bool OpenGLDriver::isDepthStencilResolveSupported() {
     return true;
 }
 
+bool OpenGLDriver::isDepthStencilBlitSupported(TextureFormat format) {
+    return true;
+}
+
 bool OpenGLDriver::isProtectedTexturesSupported() {
     return getContext().ext.EXT_protected_textures;
 }
@@ -2063,8 +2084,6 @@ 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;
     }
@@ -2343,7 +2362,7 @@ 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
             {
@@ -2673,7 +2692,6 @@ 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);
     }
 }
@@ -3068,11 +3086,10 @@ 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);
 
@@ -3090,7 +3107,7 @@ 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;
     }
 
@@ -3107,36 +3124,6 @@ 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
@@ -3436,17 +3423,21 @@ 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) {
@@ -3455,7 +3446,7 @@ 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);
             }
         }
@@ -3463,7 +3454,7 @@ void OpenGLDriver::beginFrame(
 }
 
 void OpenGLDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
-        FrameScheduledCallback callback, void* user) {
+        CallbackHandler* handler, FrameScheduledCallback&& callback) {
     DEBUG_MARKER()
 }
 
@@ -3494,6 +3485,7 @@ void OpenGLDriver::endFrame(UTILS_UNUSED uint32_t frameId) {
 #endif
     //SYSTRACE_NAME("glFinish");
     //glFinish();
+    mPlatform.endFrame(frameId);
     insertEventMarker("endFrame");
 }
 
@@ -3705,7 +3697,7 @@ 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) {
@@ -3731,7 +3723,7 @@ 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(
@@ -3853,6 +3845,7 @@ 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) {

filament/backend/src/opengl/OpenGLDriver.h

@@ -20,6 +20,8 @@
 #include "DriverBase.h"
 #include "OpenGLContext.h"
 #include "OpenGLTimerQuery.h"
+#include "GLBufferObject.h"
+#include "GLTexture.h"
 #include "ShaderCompilerService.h"
 
 #include <backend/platforms/OpenGLPlatform.h>
@@ -64,9 +66,9 @@ 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,
@@ -92,25 +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,
@@ -140,7 +123,6 @@ public:
         } gl;
     };
 
-    struct GLTexture;
     struct GLSamplerGroup : public HwSamplerGroup {
         using HwSamplerGroup::HwSamplerGroup;
         struct Entry {
@@ -157,27 +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;
 
@@ -268,13 +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) ...);
     }
 
@@ -288,7 +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);
     }
 
@@ -296,7 +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);
     }
 
@@ -348,22 +312,6 @@ 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;
     }
@@ -390,14 +338,10 @@ private:
 
     // ES2 only. Uniform buffer emulation binding points
     GLuint mLastAssignedEmulatedUboId = 0;
-    std::array<std::tuple<GLuint, void const*, uint16_t>, Program::UNIFORM_BINDING_COUNT> mUniformBindings = {};
 
     // sampler buffer binding points (nullptr if not used)
     std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT> mSamplerBindings = {};   // 4 pointers
 
-    mutable tsl::robin_map<SamplerParams, GLuint,
-            SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;
-
     // this must be accessed from the driver thread only
     std::vector<GLTexture*> mTexturesWithStreamsAttached;
 
@@ -431,6 +375,8 @@ private:
     // for ES2 sRGB support
     GLSwapChain* mCurrentDrawSwapChain = nullptr;
     bool mRec709OutputColorspace = false;
+
+    PushConstantBundle* mCurrentPushConstants = nullptr;
 };
 
 // ------------------------------------------------------------------------------------------------

filament/backend/src/opengl/OpenGLPlatform.cpp

@@ -55,6 +55,16 @@ 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;
 }

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -46,6 +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;
 };
 
 
@@ -53,7 +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())) {
@@ -61,6 +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));
@@ -203,6 +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 +27,7 @@
 
 #include <utils/compiler.h>
 #include <utils/FixedCapacityVector.h>
+#include <utils/Slice.h>
 
 #include <array>
 #include <limits>
@@ -38,6 +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:
 
@@ -70,14 +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;
@@ -89,14 +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;
@@ -104,12 +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/platforms/PlatformEGL.cpp

@@ -174,7 +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) {

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));
 

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);

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>

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,

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.

filament/backend/src/vulkan/VulkanConstants.h

@@ -47,7 +47,7 @@
 // granualarity of a renderpass. You can enable this along with FVK_DEBUG_DEBUG_UTILS to take
 // advantage of vkCmdBegin/EndDebugUtilsLabelEXT. You can also just enable this with
 // FVK_DEBUG_PRINT_GROUP_MARKERS to print the current marker to stdout.
-#define FVK_DEBUG_GROUP_MARKERS 0x00000002
+#define FVK_DEBUG_GROUP_MARKERS           0x00000002
 
 #define FVK_DEBUG_TEXTURE                 0x00000004
 #define FVK_DEBUG_LAYOUT_TRANSITION       0x00000008
@@ -75,18 +75,14 @@
 #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
@@ -112,7 +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();
 }
@@ -214,12 +220,14 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
       mThreadSafeResourceManager(&mResourceAllocator),
       mCommands(mPlatform->getDevice(), mPlatform->getGraphicsQueue(),
               mPlatform->getGraphicsQueueFamilyIndex(), &mContext, &mResourceAllocator),
-      mPipelineCache(&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)
@@ -243,13 +251,17 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
 #endif
 
     mTimestamps = std::make_unique<VulkanTimestamps>(mPlatform->getDevice());
-    mCommands.setObserver(&mPipelineCache);
-    mPipelineCache.setDevice(mPlatform->getDevice(), mAllocator);
 
     mEmptyTexture = createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
             mContext, mAllocator, &mCommands, mStagePool);
+    mEmptyBufferObject = createEmptyBufferObject(mAllocator, mStagePool, &mCommands);
 
-    mPipelineCache.setDummyTexture(mEmptyTexture->getPrimaryImageView());
+    mDescriptorSetManager.setPlaceHolders(mSamplerCache.getSampler({}), mEmptyTexture,
+            mEmptyBufferObject);
+
+    mGetPipelineFunction = [this](VulkanDescriptorSetLayoutList const& layouts, VulkanProgram* program) {
+        return mPipelineLayoutCache.getLayout(layouts, program);
+    };
 }
 
 VulkanDriver::~VulkanDriver() noexcept = default;
@@ -310,13 +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.terminate();
 
-    delete mEmptyTexture;
     mResourceManager.clear();
-
     mTimestamps.reset();
 
     mBlitter.terminate();
@@ -329,6 +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);
 
@@ -360,6 +379,8 @@ void VulkanDriver::collectGarbage() {
     mCommands.gc();
     mStagePool.gc();
     mFramebufferCache.gc();
+    mPipelineCache.gc();
+    mDescriptorSetManager.gc();
 
 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
     mResourceAllocator.print();
@@ -367,12 +388,13 @@ 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,
@@ -489,9 +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);
 }
 
@@ -542,6 +561,7 @@ void VulkanDriver::destroyProgram(Handle<HwProgram> ph) {
         return;
     }
     auto vkprogram = mResourceAllocator.handle_cast<VulkanProgram*>(ph);
+    mDescriptorSetManager.clearProgram(vkprogram);
     mResourceManager.release(vkprogram);
 }
 
@@ -898,6 +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;
 }
@@ -1211,12 +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;
@@ -1225,16 +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,
@@ -1251,9 +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;
@@ -1271,27 +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)
@@ -1300,7 +1339,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         // - If the RT is MS then all SS attachments are auto resolved if not discarded.
         assert_invariant(!(rt->getSamples() > 1 &&
                 rt->getDepth().texture->samples == 1 &&
-                !any(rpkey.discardEnd & TargetBufferFlags::DEPTH)));
+                !depthDiscardEnd));
     }
     VkFramebuffer vkfb = mFramebufferCache.getFramebuffer(fbkey);
 
@@ -1313,16 +1352,10 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     }
 #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.
@@ -1408,33 +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();
@@ -1453,15 +1505,9 @@ void VulkanDriver::nextSubpass(int) {
     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);
     }
 }
 
@@ -1501,25 +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) {
@@ -1527,6 +1572,14 @@ 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);
@@ -1767,22 +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;
         }
 
@@ -1791,24 +1836,20 @@ 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;
@@ -1816,43 +1857,22 @@ void VulkanDriver::bindPipeline(PipelineState pipelineState) {
             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,
+    };
 
-    // 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;
-    }
+    mPipelineCache.bindLayout(pipelineLayout);
+    mPipelineCache.bindPipeline(commands);
     FVK_SYSTRACE_END();
 }
 
@@ -1892,12 +1912,8 @@ void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
     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;

filament/backend/src/vulkan/VulkanDriver.h

@@ -28,6 +28,8 @@
 #include "VulkanSamplerCache.h"
 #include "VulkanStagePool.h"
 #include "VulkanUtility.h"
+#include "caching/VulkanDescriptorSetManager.h"
+#include "caching/VulkanPipelineLayoutCache.h"
 
 #include "DriverBase.h"
 #include "private/backend/Driver.h"
@@ -40,6 +42,25 @@ namespace filament::backend {
 class VulkanPlatform;
 struct VulkanSamplerGroup;
 
+// The maximum number of attachments for any renderpass (color + resolve + depth)
+constexpr uint8_t MAX_RENDERTARGET_ATTACHMENT_TEXTURES =
+        MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT * 2 + 1;
+
+// We need to store information about a render pass to enable better barriers at the end of a
+// renderpass.
+struct RenderPassFboBundle {
+    using AttachmentArray =
+            CappedArray<VulkanAttachment, MAX_RENDERTARGET_ATTACHMENT_TEXTURES>;
+
+    AttachmentArray attachments;
+    bool hasColorResolve = false;
+
+    void clear() {
+        attachments.clear();
+        hasColorResolve = false;
+    }
+};
+
 class VulkanDriver final : public DriverBase {
 public:
     static Driver* create(VulkanPlatform* platform, VulkanContext const& context,
@@ -70,6 +91,8 @@ public:
 #endif // FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
 
 private:
+    static constexpr uint8_t MAX_SAMPLER_BINDING_COUNT = Program::SAMPLER_BINDING_COUNT;
+
     void debugCommandBegin(CommandStream* cmds, bool synchronous,
             const char* methodName) noexcept override;
 
@@ -106,7 +129,9 @@ private:
     VulkanPlatform* mPlatform = nullptr;
     std::unique_ptr<VulkanTimestamps> mTimestamps;
 
+    // Placeholder resources
     VulkanTexture* mEmptyTexture;
+    VulkanBufferObject* mEmptyBufferObject;
 
     VulkanSwapChain* mCurrentSwapChain = nullptr;
     VulkanRenderTarget* mDefaultRenderTarget = nullptr;
@@ -123,13 +148,26 @@ private:
     VulkanThreadSafeResourceManager mThreadSafeResourceManager;
 
     VulkanCommands mCommands;
+    VulkanPipelineLayoutCache mPipelineLayoutCache;
     VulkanPipelineCache mPipelineCache;
     VulkanStagePool mStagePool;
     VulkanFboCache mFramebufferCache;
     VulkanSamplerCache mSamplerCache;
     VulkanBlitter mBlitter;
-    VulkanSamplerGroup* mSamplerBindings[VulkanPipelineCache::SAMPLER_BINDING_COUNT] = {};
+    VulkanSamplerGroup* mSamplerBindings[MAX_SAMPLER_BINDING_COUNT] = {};
     VulkanReadPixels mReadPixels;
+    VulkanDescriptorSetManager mDescriptorSetManager;
+
+    VulkanDescriptorSetManager::GetPipelineLayoutFunction mGetPipelineFunction;
+
+    // This is necessary for us to write to push constants after binding a pipeline.
+    struct BoundPipeline {
+        VulkanProgram* program;
+        VkPipelineLayout pipelineLayout;
+    };
+    BoundPipeline mBoundPipeline = {};
+
+    RenderPassFboBundle mRenderPassFboInfo;
 
     bool const mIsSRGBSwapChainSupported;
 };

filament/backend/src/vulkan/VulkanFboCache.cpp

@@ -29,14 +29,10 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-using ImgUtil = VulkanImageUtility;
-
 bool VulkanFboCache::RenderPassEq::operator()(const RenderPassKey& k1,
         const RenderPassKey& k2) const {
     if (k1.initialColorLayoutMask != k2.initialColorLayoutMask) return false;
     if (k1.initialDepthLayout != k2.initialDepthLayout) return false;
-    if (k1.renderPassDepthLayout != k2.renderPassDepthLayout) return false;
-    if (k1.finalDepthLayout != k2.finalDepthLayout) return false;
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
         if (k1.colorFormat[i] != k2.colorFormat[i]) return false;
     }
@@ -197,7 +193,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
         if (config.colorFormat[i] == VK_FORMAT_UNDEFINED) {
             continue;
         }
-        const VkImageLayout subpassLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT);
+        const VkImageLayout subpassLayout = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT);
         uint32_t index;
 
         if (!hasSubpasses) {
@@ -243,9 +239,9 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
             .stencilLoadOp = kDontCare,
             .stencilStoreOp = kDisableStore,
             .initialLayout = ((!discard && config.initialColorLayoutMask & (1 << i)) || clear)
-                                     ? ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT)
-                                     : ImgUtil::getVkLayout(VulkanLayout::UNDEFINED),
-            .finalLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
+                                     ? imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT)
+                                     : imgutil::getVkLayout(VulkanLayout::UNDEFINED),
+            .finalLayout = imgutil::getVkLayout(FINAL_COLOR_ATTACHMENT_LAYOUT),
         };
     }
 
@@ -272,7 +268,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
 
         pResolveAttachment->attachment = attachmentIndex;
         pResolveAttachment->layout
-                = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT_RESOLVE);
+                = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT_RESOLVE);
         ++pResolveAttachment;
 
         attachments[attachmentIndex++] = {
@@ -282,8 +278,8 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
             .storeOp = kEnableStore,
             .stencilLoadOp = kDontCare,
             .stencilStoreOp = kDisableStore,
-            .initialLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
-            .finalLayout = ImgUtil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
+            .initialLayout = imgutil::getVkLayout(VulkanLayout::COLOR_ATTACHMENT),
+            .finalLayout = imgutil::getVkLayout(FINAL_COLOR_ATTACHMENT_LAYOUT),
         };
     }
 
@@ -292,7 +288,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
         const bool clear = any(config.clear & TargetBufferFlags::DEPTH);
         const bool discardStart = any(config.discardStart & TargetBufferFlags::DEPTH);
         const bool discardEnd = any(config.discardEnd & TargetBufferFlags::DEPTH);
-        depthAttachmentRef.layout = ImgUtil::getVkLayout(config.renderPassDepthLayout);
+        depthAttachmentRef.layout = imgutil::getVkLayout(VulkanLayout::DEPTH_ATTACHMENT);
         depthAttachmentRef.attachment = attachmentIndex;
         attachments[attachmentIndex++] = {
             .format = config.depthFormat,
@@ -301,8 +297,8 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
             .storeOp = discardEnd ? kDisableStore : kEnableStore,
             .stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE,
             .stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE,
-            .initialLayout = ImgUtil::getVkLayout(config.initialDepthLayout),
-            .finalLayout = ImgUtil::getVkLayout(config.finalDepthLayout),
+            .initialLayout = imgutil::getVkLayout(config.initialDepthLayout),
+            .finalLayout = imgutil::getVkLayout(FINAL_DEPTH_ATTACHMENT_LAYOUT),
         };
     }
     renderPassInfo.attachmentCount = attachmentIndex;

filament/backend/src/vulkan/VulkanFboCache.h

@@ -35,24 +35,25 @@ namespace filament::backend {
 //
 class VulkanFboCache {
 public:
+    constexpr static VulkanLayout FINAL_COLOR_ATTACHMENT_LAYOUT = VulkanLayout::COLOR_ATTACHMENT;
+    constexpr static VulkanLayout FINAL_RESOLVE_ATTACHMENT_LAYOUT = VulkanLayout::COLOR_ATTACHMENT;
+    constexpr static VulkanLayout FINAL_DEPTH_ATTACHMENT_LAYOUT = VulkanLayout::DEPTH_ATTACHMENT;
+
     // RenderPassKey is a small POD representing the immutable state that is used to construct
     // a VkRenderPass. It is hashed and used as a lookup key.
-    // TODO: This struct can be reduced in size by using a subset of formats instead of VkFormat
-    //       and removing the "finalDepthLayout" field.
     struct alignas(8) RenderPassKey {
         // For each target, we need to know three image layouts: the layout BEFORE the pass, the
         // layout DURING the pass, and the layout AFTER the pass. Here are the rules:
         // - For depth, we explicitly specify all three layouts.
         // - Color targets have their initial image layout specified with a bitmask.
         // - For each color target, the pre-existing layout is either UNDEFINED (0) or GENERAL (1).
-        // - The render pass and final images layout for color buffers is always GENERAL.
+        // - The render pass and final images layout for color buffers is always
+        //   VulkanLayout::COLOR_ATTACHMENT.
         uint8_t initialColorLayoutMask;
 
         // Note that if VulkanLayout grows beyond 16, we'd need to up this.
-        VulkanLayout initialDepthLayout : 4;
-        VulkanLayout renderPassDepthLayout : 4;
-        VulkanLayout finalDepthLayout : 4;
-        uint8_t padding0 : 4;
+        VulkanLayout initialDepthLayout : 8;
+        uint8_t padding0;
         uint8_t padding1;
 
         VkFormat colorFormat[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT]; // 32 bytes
@@ -63,7 +64,7 @@ public:
         uint8_t samples; // 1 byte
         uint8_t needsResolveMask; // 1 byte
         uint8_t subpassMask; // 1 byte
-        bool padding2; // 1 byte
+        uint8_t padding2; // 1 byte
     };
     struct RenderPassVal {
         VkRenderPass handle;

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -20,7 +20,9 @@
 #include "VulkanConstants.h"
 #include "VulkanDriver.h"
 #include "VulkanMemory.h"
+#include "VulkanUtility.h"
 #include "spirv/VulkanSpirvUtils.h"
+#include "utils/Log.h"
 
 #include <backend/platforms/VulkanPlatform.h>
 
@@ -30,15 +32,17 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-static void flipVertically(VkRect2D* rect, uint32_t framebufferHeight) {
+namespace {
+
+void flipVertically(VkRect2D* rect, uint32_t framebufferHeight) {
     rect->offset.y = framebufferHeight - rect->offset.y - rect->extent.height;
 }
 
-static void flipVertically(VkViewport* rect, uint32_t framebufferHeight) {
+void flipVertically(VkViewport* rect, uint32_t framebufferHeight) {
     rect->y = framebufferHeight - rect->y - rect->height;
 }
 
-static void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
+void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
     int32_t x = std::max(rect->offset.x, 0);
     int32_t y = std::max(rect->offset.y, 0);
     int32_t right = std::min(rect->offset.x + (int32_t) rect->extent.width, (int32_t) fbWidth);
@@ -49,17 +53,165 @@ static void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeig
     rect->extent.height = std::max(top - y, 0);
 }
 
+template<typename Bitmask>
+static constexpr Bitmask fromStageFlags(ShaderStageFlags2 flags, uint8_t binding) {
+    Bitmask ret = 0;
+    if (flags & ShaderStageFlags2::VERTEX) {
+        ret |= (getVertexStage<Bitmask>() << binding);
+    }
+    if (flags & ShaderStageFlags2::FRAGMENT) {
+        ret |= (getFragmentStage<Bitmask>() << binding);
+    }
+    return ret;
+}
+
+constexpr decltype(VulkanProgram::MAX_SHADER_MODULES) MAX_SHADER_MODULES =
+        VulkanProgram::MAX_SHADER_MODULES;
+
+using LayoutDescriptionList = VulkanProgram::LayoutDescriptionList;
+
+template<typename Bitmask>
+void addDescriptors(Bitmask mask,
+        utils::FixedCapacityVector<DescriptorSetLayoutBinding>& outputList) {
+    constexpr uint8_t MODULE_OFFSET = (sizeof(Bitmask) * 8) / MAX_SHADER_MODULES;
+    for (uint8_t i = 0; i < MODULE_OFFSET; ++i) {
+        bool const hasVertex = (mask & (1ULL << i)) != 0;
+        bool const hasFragment = (mask & (1ULL << (MODULE_OFFSET + i))) != 0;
+        if (!hasVertex && !hasFragment) {
+            continue;
+        }
+
+        DescriptorSetLayoutBinding binding{
+            .binding = i,
+            .flags = DescriptorFlags::NONE,
+            .count = 0,// This is always 0 for now as we pass the size of the UBOs in the Driver API
+                       // instead.
+        };
+        if (hasVertex) {
+            binding.stageFlags = ShaderStageFlags2::VERTEX;
+        }
+        if (hasFragment) {
+            binding.stageFlags = static_cast<ShaderStageFlags2>(
+                    binding.stageFlags | ShaderStageFlags2::FRAGMENT);
+        }
+        if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
+            binding.type = DescriptorType::UNIFORM_BUFFER;
+        } else if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
+            binding.type = DescriptorType::SAMPLER;
+        } else if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
+            binding.type = DescriptorType::INPUT_ATTACHMENT;
+        }
+        outputList.push_back(binding);
+    }
+}
+
+inline VkDescriptorSetLayout createDescriptorSetLayout(VkDevice device,
+        VkDescriptorSetLayoutCreateInfo const& info) {
+    VkDescriptorSetLayout layout;
+    vkCreateDescriptorSetLayout(device, &info, VKALLOC, &layout);
+    return layout;
+}
+
+inline VkShaderStageFlags getVkStage(backend::ShaderStage stage) {
+    switch(stage) {
+        case backend::ShaderStage::VERTEX:
+            return VK_SHADER_STAGE_VERTEX_BIT;
+        case backend::ShaderStage::FRAGMENT:
+            return VK_SHADER_STAGE_FRAGMENT_BIT;
+        case backend::ShaderStage::COMPUTE:
+            PANIC_POSTCONDITION("Unsupported stage");
+    }
+}
+
+} // anonymous namespace
+
+
+VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(VkDevice device,
+        VkDescriptorSetLayoutCreateInfo const& info, Bitmask const& bitmask)
+    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET_LAYOUT),
+      mDevice(device),
+      vklayout(createDescriptorSetLayout(device, info)),
+      bitmask(bitmask),
+      bindings(getBindings(bitmask)),
+      count(Count::fromLayoutBitmask(bitmask)) {}
+
+VulkanDescriptorSetLayout::~VulkanDescriptorSetLayout() {
+    vkDestroyDescriptorSetLayout(mDevice, vklayout, VKALLOC);
+}
+
+PushConstantDescription::PushConstantDescription(backend::Program const& program) noexcept {
+    mRangeCount = 0;
+    for (auto stage : { ShaderStage::VERTEX, ShaderStage::FRAGMENT, ShaderStage::COMPUTE }) {
+        auto const& constants = program.getPushConstants(stage);
+        if (constants.empty()) {
+            continue;
+        }
+
+        // We store the type of the constant for type-checking when writing.
+        auto& types = mTypes[(uint8_t) stage];
+        types.reserve(constants.size());
+        std::for_each(constants.cbegin(), constants.cend(), [&types] (Program::PushConstant t) {
+            types.push_back(t.type);
+        });
+
+        mRanges[mRangeCount++] = {
+            .stageFlags = getVkStage(stage),
+            .offset = 0,
+            .size = (uint32_t) constants.size() * ENTRY_SIZE,
+        };
+    }
+}
+
+void PushConstantDescription::write(VulkanCommands* commands, VkPipelineLayout layout,
+        backend::ShaderStage stage, uint8_t index, backend::PushConstantVariant const& value) {
+    VulkanCommandBuffer* cmdbuf = &(commands->get());
+    uint32_t binaryValue = 0;
+    UTILS_UNUSED_IN_RELEASE auto const& types = mTypes[(uint8_t) stage];
+    if (std::holds_alternative<bool>(value)) {
+        assert_invariant(types[index] == ConstantType::BOOL);
+        bool const bval = std::get<bool>(value);
+        binaryValue = static_cast<uint32_t const>(bval ? VK_TRUE : VK_FALSE);
+    } else if (std::holds_alternative<float>(value)) {
+        assert_invariant(types[index] == ConstantType::FLOAT);
+        float const fval = std::get<float>(value);
+        binaryValue = *reinterpret_cast<uint32_t const*>(&fval);
+    } else {
+        assert_invariant(types[index] == ConstantType::INT);
+        int const ival = std::get<int>(value);
+        binaryValue = *reinterpret_cast<uint32_t const*>(&ival);
+    }
+    vkCmdPushConstants(cmdbuf->buffer(), layout, getVkStage(stage), index * ENTRY_SIZE, ENTRY_SIZE,
+            &binaryValue);
+}
+
 VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
     : HwProgram(builder.getName()),
       VulkanResource(VulkanResourceType::PROGRAM),
-      mInfo(new PipelineInfo()),
+      mInfo(new(std::nothrow) PipelineInfo(builder)),
       mDevice(device) {
+
+    constexpr uint8_t UBO_MODULE_OFFSET = (sizeof(UniformBufferBitmask) * 8) / MAX_SHADER_MODULES;
+    constexpr uint8_t SAMPLER_MODULE_OFFSET = (sizeof(SamplerBitmask) * 8) / MAX_SHADER_MODULES;
+    constexpr uint8_t INPUT_ATTACHMENT_MODULE_OFFSET =
+            (sizeof(InputAttachmentBitmask) * 8) / MAX_SHADER_MODULES;
+
     Program::ShaderSource const& blobs = builder.getShadersSource();
     auto& modules = mInfo->shaders;
 
     auto const& specializationConstants = builder.getSpecializationConstants();
 
     std::vector<uint32_t> shader;
+
+    // TODO: this will be moved out of the shader as the descriptor set layout will be provided by
+    // Filament instead of parsed from the shaders. See [GDSR] in VulkanDescriptorSetManager.h
+    UniformBufferBitmask uboMask = 0;
+    SamplerBitmask samplerMask = 0;
+    InputAttachmentBitmask inputAttachmentMask = 0;
+
+    static_assert(static_cast<ShaderStage>(0) == ShaderStage::VERTEX &&
+            static_cast<ShaderStage>(1) == ShaderStage::FRAGMENT &&
+            MAX_SHADER_MODULES == 2);
+
     for (size_t i = 0; i < MAX_SHADER_MODULES; i++) {
         Program::ShaderBlob const& blob = blobs[i];
 
@@ -72,6 +224,12 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
             dataSize = shader.size() * 4;
         }
 
+        auto const [ubo, sampler, inputAttachment] = getProgramBindings(blob);
+        uboMask |= (static_cast<UniformBufferBitmask>(ubo) << (UBO_MODULE_OFFSET * i));
+        samplerMask |= (static_cast<SamplerBitmask>(sampler) << (SAMPLER_MODULE_OFFSET * i));
+        inputAttachmentMask |= (static_cast<InputAttachmentBitmask>(inputAttachment)
+                                << (INPUT_ATTACHMENT_MODULE_OFFSET * i));
+
         VkShaderModule& module = modules[i];
         VkShaderModuleCreateInfo moduleInfo = {
             .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
@@ -99,20 +257,33 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
     }
 
+    LayoutDescriptionList& layouts = mInfo->layouts;
+    layouts[0].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
+            countBits(collapseStages(uboMask)));
+    layouts[1].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
+            countBits(collapseStages(samplerMask)));
+    layouts[2].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
+            countBits(collapseStages(inputAttachmentMask)));
+
+    addDescriptors(uboMask, layouts[0].bindings);
+    addDescriptors(samplerMask, layouts[1].bindings);
+    addDescriptors(inputAttachmentMask, layouts[2].bindings);
+
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
     auto& bindingToName = mInfo->bindingToName;
 #endif
 
     auto& groupInfo = builder.getSamplerGroupInfo();
     auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
-    auto& usage = mInfo->usage;
+    auto& bindings = mInfo->bindings;
     for (uint8_t groupInd = 0; groupInd < Program::SAMPLER_BINDING_COUNT; groupInd++) {
         auto const& group = groupInfo[groupInd];
         auto const& samplers = group.samplers;
         for (size_t i = 0; i < samplers.size(); ++i) {
             uint32_t const binding = samplers[i].binding;
             bindingToSamplerIndex[binding] = (groupInd << 8) | (0xff & i);
-            usage = VulkanPipelineCache::getUsageFlags(binding, group.stageFlags, usage);
+            assert_invariant(bindings.find(binding) == bindings.end());
+            bindings.insert(binding);
 
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
             bindingToName[binding] = samplers[i].name.c_str();
@@ -126,24 +297,6 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
 }
 
-VulkanProgram::VulkanProgram(VkDevice device, VkShaderModule vs, VkShaderModule fs,
-        CustomSamplerInfoList const& samplerInfo) noexcept
-    : VulkanResource(VulkanResourceType::PROGRAM),
-      mInfo(new PipelineInfo()),
-      mDevice(device) {
-    mInfo->shaders[0] = vs;
-    mInfo->shaders[1] = fs;
-    auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
-    auto& usage = mInfo->usage;
-    bindingToSamplerIndex.resize(samplerInfo.size());
-    for (uint16_t binding = 0; binding < samplerInfo.size(); ++binding) {
-        auto const& sampler = samplerInfo[binding];
-        bindingToSamplerIndex[binding]
-                = (sampler.groupIndex << 8) | (0xff & sampler.samplerIndex);
-        usage = VulkanPipelineCache::getUsageFlags(binding, sampler.flags, usage);
-    }
-}
-
 VulkanProgram::~VulkanProgram() {
     for (auto shader: mInfo->shaders) {
         vkDestroyShaderModule(mDevice, shader, VKALLOC);
@@ -187,13 +340,13 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica
 
     // Constrain the sample count according to both kinds of sample count masks obtained from
     // VkPhysicalDeviceProperties. This is consistent with the VulkanTexture constructor.
-    const auto& limits = context.getPhysicalDeviceLimits();
+    auto const& limits = context.getPhysicalDeviceLimits();
     mSamples = samples = reduceSampleCount(samples, limits.framebufferDepthSampleCounts &
             limits.framebufferColorSampleCounts);
 
     // Create sidecar MSAA textures for color attachments if they don't already exist.
     for (int index = 0; index < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; index++) {
-        const VulkanAttachment& spec = color[index];
+        VulkanAttachment const& spec = color[index];
         VulkanTexture* texture = (VulkanTexture*) spec.texture;
         if (texture && texture->samples == 1) {
             auto msTexture = texture->getSidecar();
@@ -256,19 +409,19 @@ VkExtent2D VulkanRenderTarget::getExtent() const {
     return {width, height};
 }
 
-VulkanAttachment VulkanRenderTarget::getColor(int target) const {
+VulkanAttachment& VulkanRenderTarget::getColor(int target) {
     return mColor[target];
 }
 
-VulkanAttachment VulkanRenderTarget::getMsaaColor(int target) const {
+VulkanAttachment& VulkanRenderTarget::getMsaaColor(int target) {
     return mMsaaAttachments[target];
 }
 
-VulkanAttachment VulkanRenderTarget::getDepth() const {
+VulkanAttachment& VulkanRenderTarget::getDepth() {
     return mDepth;
 }
 
-VulkanAttachment VulkanRenderTarget::getMsaaDepth() const {
+VulkanAttachment& VulkanRenderTarget::getMsaaDepth() {
     return mMsaaDepthAttachment;
 }
 
@@ -405,4 +558,35 @@ VulkanRenderPrimitive::VulkanRenderPrimitive(VulkanResourceAllocator* resourceAl
     mResources.acquire(indexBuffer);
 }
 
+using Bitmask = VulkanDescriptorSetLayout::Bitmask;
+
+Bitmask Bitmask::fromBackendLayout(descset::DescriptorSetLayout const& layout) {
+    Bitmask mask;
+    for (auto const& binding: layout.bindings) {
+        switch (binding.type) {
+            case descset::DescriptorType::UNIFORM_BUFFER: {
+                if (binding.flags == descset::DescriptorFlags::DYNAMIC_OFFSET) {
+                    mask.dynamicUbo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
+                            binding.binding);
+                } else {
+                    mask.ubo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
+                            binding.binding);
+                }
+                break;
+            }
+            case descset::DescriptorType::SAMPLER: {
+                mask.sampler |= fromStageFlags<SamplerBitmask>(binding.stageFlags, binding.binding);
+                break;
+            }
+            case descset::DescriptorType::INPUT_ATTACHMENT: {
+                mask.inputAttachment |=
+                        fromStageFlags<InputAttachmentBitmask>(binding.stageFlags, binding.binding);
+                break;
+            }
+        }
+    }
+    return mask;
+}
+
+
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanHandles.h

@@ -21,35 +21,193 @@
 #include "DriverBase.h"
 
 #include "VulkanBuffer.h"
-#include "VulkanPipelineCache.h"
 #include "VulkanResources.h"
 #include "VulkanSwapChain.h"
 #include "VulkanTexture.h"
+#include "VulkanUtility.h"
 
-#include "private/backend/SamplerGroup.h"
+#include <private/backend/SamplerGroup.h>
+#include <backend/Program.h>
 
+#include <utils/FixedCapacityVector.h>
 #include <utils/Mutex.h>
 #include <utils/StructureOfArrays.h>
 
 namespace filament::backend {
 
+using namespace descset;
+
 class VulkanTimestamps;
 
+struct VulkanDescriptorSetLayout : public VulkanResource {
+    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 3;
+
+    // The bitmask representation of a set layout.
+    struct Bitmask {
+        UniformBufferBitmask ubo = 0;         // 4 bytes
+        UniformBufferBitmask dynamicUbo = 0;  // 4 bytes
+        SamplerBitmask sampler = 0;           // 8 bytes
+        InputAttachmentBitmask inputAttachment = 0; // 1 bytes
+
+        // Because we're using this struct as hash key, must make it's 8-bytes aligned, with no
+        // unaccounted bytes.
+        uint8_t padding0 = 0; // 1 bytes
+        uint16_t padding1 = 0;// 2 bytes
+        uint32_t padding2 = 0;// 4 bytes
+
+        bool operator==(Bitmask const& right) const {
+            return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
+                   inputAttachment == right.inputAttachment;
+        }
+
+        static Bitmask fromBackendLayout(descset::DescriptorSetLayout const& layout);
+    };
+
+    // This is a convenience struct to quickly check layout compatibility in terms of descriptor set
+    // pools.
+    struct Count {
+        uint32_t ubo = 0;
+        uint32_t dynamicUbo = 0;
+        uint32_t sampler = 0;
+        uint32_t inputAttachment = 0;
+
+        bool operator==(Count const& right) const noexcept {
+            return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
+                   inputAttachment == right.inputAttachment;
+        }
+
+        static inline Count fromLayoutBitmask(Bitmask const& mask) {
+            return {
+                    .ubo = countBits(collapseStages(mask.ubo)),
+                    .dynamicUbo = countBits(collapseStages(mask.dynamicUbo)),
+                    .sampler = countBits(collapseStages(mask.sampler)),
+                    .inputAttachment = countBits(collapseStages(mask.inputAttachment)),
+            };
+        }
+
+        Count operator*(uint16_t mult) const noexcept {
+            // TODO: check for overflow.
+
+            Count ret;
+            ret.ubo = ubo * mult;
+            ret.dynamicUbo = dynamicUbo * mult;
+            ret.sampler = sampler * mult;
+            ret.inputAttachment = inputAttachment * mult;
+            return ret;
+        }
+    };
+
+    static_assert(sizeof(Bitmask) % 8 == 0);
+
+    explicit VulkanDescriptorSetLayout(VkDevice device, VkDescriptorSetLayoutCreateInfo const& info,
+            Bitmask const& bitmask);
+
+    ~VulkanDescriptorSetLayout();
+
+    VkDevice const mDevice;
+    VkDescriptorSetLayout const vklayout;
+    Bitmask const bitmask;
+
+    // This is a convenience struct so that we don't have to iterate through all the bits of the
+    // bitmask (which correspondings to binding indices).
+    struct _Bindings {
+        utils::FixedCapacityVector<uint8_t> const ubo;
+        utils::FixedCapacityVector<uint8_t> const dynamicUbo;
+        utils::FixedCapacityVector<uint8_t> const sampler;
+        utils::FixedCapacityVector<uint8_t> const inputAttachment;
+    } bindings;
+
+    Count const count;
+
+private:
+
+    template <typename MaskType>
+    utils::FixedCapacityVector<uint8_t> bits(MaskType mask) {
+        utils::FixedCapacityVector<uint8_t> ret =
+                utils::FixedCapacityVector<uint8_t>::with_capacity(countBits(mask));
+        for (uint8_t i = 0; i < sizeof(mask) * 4; ++i) {
+            if (mask & (1 << i)) {
+                ret.push_back(i);
+            }
+        }
+        return ret;
+    }
+
+    _Bindings getBindings(Bitmask const& bitmask) {
+        auto const uboCollapsed = collapseStages(bitmask.ubo);
+        auto const dynamicUboCollapsed = collapseStages(bitmask.dynamicUbo);
+        auto const samplerCollapsed = collapseStages(bitmask.sampler);
+        auto const inputAttachmentCollapsed = collapseStages(bitmask.inputAttachment);
+        return {
+            bits(uboCollapsed),
+            bits(dynamicUboCollapsed),
+            bits(samplerCollapsed),
+            bits(inputAttachmentCollapsed),
+        };
+    }
+};
+
+using VulkanDescriptorSetLayoutList = std::array<Handle<VulkanDescriptorSetLayout>,
+        VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
+struct VulkanDescriptorSet : public VulkanResource {
+public:
+    // Because we need to recycle descriptor sets not used, we allow for a callback that the "Pool"
+    // can use to repackage the vk handle.
+    using OnRecycle = std::function<void()>;
+
+    VulkanDescriptorSet(VulkanResourceAllocator* allocator,
+            VkDescriptorSet rawSet, OnRecycle&& onRecycleFn)
+        : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
+          resources(allocator),
+          vkSet(rawSet),
+          mOnRecycleFn(std::move(onRecycleFn)) {}
+
+    ~VulkanDescriptorSet() {
+        if (mOnRecycleFn) {
+            mOnRecycleFn();
+        }
+    }
+
+    // TODO: maybe change to fixed size for performance.
+    VulkanAcquireOnlyResourceManager resources;
+    VkDescriptorSet const vkSet;
+
+private:
+    OnRecycle mOnRecycleFn;
+};
+
+using VulkanDescriptorSetList = std::array<Handle<VulkanDescriptorSet>,
+        VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
+using PushConstantNameArray = utils::FixedCapacityVector<char const*>;
+using PushConstantNameByStage = std::array<PushConstantNameArray, Program::SHADER_TYPE_COUNT>;
+
+struct PushConstantDescription {
+
+    explicit PushConstantDescription(backend::Program const& program) noexcept;
+
+    VkPushConstantRange const* getVkRanges() const noexcept { return mRanges; }
+
+    uint32_t getVkRangeCount() const noexcept { return mRangeCount; }
+
+    void write(VulkanCommands* commands, VkPipelineLayout layout, backend::ShaderStage stage,
+            uint8_t index, backend::PushConstantVariant const& value);
+
+private:
+    static constexpr uint32_t ENTRY_SIZE = sizeof(uint32_t);
+
+    utils::FixedCapacityVector<backend::ConstantType> mTypes[Program::SHADER_TYPE_COUNT];
+    VkPushConstantRange mRanges[Program::SHADER_TYPE_COUNT];
+    uint32_t mRangeCount;
+};
+
 struct VulkanProgram : public HwProgram, VulkanResource {
 
+    using BindingList = CappedArray<uint16_t, MAX_SAMPLER_COUNT>;
+
     VulkanProgram(VkDevice device, Program const& builder) noexcept;
 
-    struct CustomSamplerInfo {
-        uint8_t groupIndex;
-        uint8_t samplerIndex;
-        ShaderStageFlags flags;
-    };
-    using CustomSamplerInfoList = utils::FixedCapacityVector<CustomSamplerInfo>;
-
-    // We allow custom descriptor of the samplers within shaders.  This is needed if we want to use
-    // a program that exists only in the backend - for example, for shader-based bliting.
-    VulkanProgram(VkDevice device, VkShaderModule vs, VkShaderModule fs,
-            CustomSamplerInfoList const& samplerInfo) noexcept;
     ~VulkanProgram();
 
     inline VkShaderModule getVertexShader() const {
@@ -58,39 +216,69 @@ struct VulkanProgram : public HwProgram, VulkanResource {
 
     inline VkShaderModule getFragmentShader() const { return mInfo->shaders[1]; }
 
-    inline VulkanPipelineCache::UsageFlags getUsage() const { return mInfo->usage; }
-
     inline utils::FixedCapacityVector<uint16_t> const& getBindingToSamplerIndex() const {
         return mInfo->bindingToSamplerIndex;
     }
 
+    // Get a list of the sampler binding indices so that we don't have to loop through all possible
+    // samplers.
+    inline BindingList const& getBindings() const { return mInfo->bindings; }
+
+    // TODO: this is currently not used. This will replace getLayoutDescriptionList below.
+    // inline descset::DescriptorSetLayout const& getLayoutDescription() const {
+    //    return mInfo->layout;
+    // }
+    // In the usual case, we would have just one layout per program. But in the current setup, we
+    // have a set/layout for each descriptor type. This will be changed in the future.
+    using LayoutDescriptionList = std::array<descset::DescriptorSetLayout,
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+    inline LayoutDescriptionList const& getLayoutDescriptionList() const { return mInfo->layouts; }
+
+    inline uint32_t getPushConstantRangeCount() const {
+        return mInfo->pushConstantDescription.getVkRangeCount();
+    }
+
+    inline VkPushConstantRange const* getPushConstantRanges() const {
+        return mInfo->pushConstantDescription.getVkRanges();
+    }
+
+    inline void writePushConstant(VulkanCommands* commands, VkPipelineLayout layout,
+            backend::ShaderStage stage, uint8_t index, backend::PushConstantVariant const& value) {
+        mInfo->pushConstantDescription.write(commands, layout, stage, index, value);
+    }
+
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
     inline utils::FixedCapacityVector<std::string> const& getBindingToName() const {
         return mInfo->bindingToName;
     }
 #endif
 
-private:
     // TODO: handle compute shaders.
     // The expected order of shaders - from frontend to backend - is vertex, fragment, compute.
     static constexpr uint8_t const MAX_SHADER_MODULES = 2;
 
+private:
     struct PipelineInfo {
-        PipelineInfo()
-            : bindingToSamplerIndex(MAX_SAMPLER_COUNT, 0xffff)
+        explicit PipelineInfo(backend::Program const& program) noexcept
+            : bindingToSamplerIndex(MAX_SAMPLER_COUNT, 0xffff),
+              pushConstantDescription(program)
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
             , bindingToName(MAX_SAMPLER_COUNT, "")
-#endif              
+#endif
             {}
 
-        // This bitset maps to each of the sampler in the sampler groups associated with this
-        // program, and whether each sampler is used in which shader (i.e. vert, frag, compute).
-        VulkanPipelineCache::UsageFlags usage;
+        BindingList bindings;
 
         // We store the samplerGroupIndex as the top 8-bit and the index within each group as the lower 8-bit.
         utils::FixedCapacityVector<uint16_t> bindingToSamplerIndex;
         VkShaderModule shaders[MAX_SHADER_MODULES] = { VK_NULL_HANDLE };
 
+        // TODO: Use this instead of `layouts` after Filament-side Descriptor Set API is in place.
+        // descset::DescriptorSetLayout layout;
+        LayoutDescriptionList layouts;
+
+        PushConstantDescription pushConstantDescription;
+
 #if FVK_ENABLED_DEBUG_SAMPLER_NAME
         // We store the sampler name mapped from binding index (only for debug purposes).
         utils::FixedCapacityVector<std::string> bindingToName;
@@ -124,10 +312,11 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
     void transformClientRectToPlatform(VkRect2D* bounds) const;
     void transformClientRectToPlatform(VkViewport* bounds) const;
     VkExtent2D getExtent() const;
-    VulkanAttachment getColor(int target) const;
-    VulkanAttachment getMsaaColor(int target) const;
-    VulkanAttachment getDepth() const;
-    VulkanAttachment getMsaaDepth() const;
+    // We return references in the following methods to avoid a copy.
+    VulkanAttachment& getColor(int target);
+    VulkanAttachment& getMsaaColor(int target);
+    VulkanAttachment& getDepth();
+    VulkanAttachment& getMsaaDepth();
     uint8_t getColorTargetCount(const VulkanRenderPass& pass) const;
     uint8_t getSamples() const { return mSamples; }
     bool hasDepth() const { return mDepth.texture; }

filament/backend/src/vulkan/VulkanImageUtility.cpp

@@ -26,38 +26,10 @@
 
 using namespace bluevk;
 
-namespace filament::backend {
+namespace filament::backend::imgutil {
 
 namespace {
 
-inline VkImageLayout getVkImageLayout(VulkanLayout layout) {
-    switch (layout) {
-        case VulkanLayout::UNDEFINED:
-            return VK_IMAGE_LAYOUT_UNDEFINED;
-        case VulkanLayout::READ_WRITE:
-            return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::READ_ONLY:
-            return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
-        case VulkanLayout::TRANSFER_SRC:
-            return VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
-        case VulkanLayout::TRANSFER_DST:
-            return VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
-        case VulkanLayout::DEPTH_ATTACHMENT:
-            return VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
-        case VulkanLayout::DEPTH_SAMPLER:
-            return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::PRESENT:
-            return VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
-        // Filament sometimes samples from one miplevel while writing to another level in the same
-        // texture (e.g. bloom does this). Moreover we'd like to avoid lots of expensive layout
-        // transitions. So, keep it simple and use GENERAL for all color-attachable textures.
-        case VulkanLayout::COLOR_ATTACHMENT:
-            return VK_IMAGE_LAYOUT_GENERAL;
-        case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:
-            return VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
-    }
-}
-
 inline std::tuple<VkAccessFlags, VkAccessFlags, VkPipelineStageFlags, VkPipelineStageFlags,
         VkImageLayout, VkImageLayout>
 getVkTransition(const VulkanLayoutTransition& transition) {
@@ -70,11 +42,11 @@ getVkTransition(const VulkanLayoutTransition& transition) {
             srcStage = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
             break;
         case VulkanLayout::COLOR_ATTACHMENT:
-            srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT
-                            | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
-                            | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT
-                       | VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+            srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT |
+                            VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+                            VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+            srcStage = VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
+                       VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
             break;
         case VulkanLayout::READ_WRITE:
             srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
@@ -154,27 +126,12 @@ getVkTransition(const VulkanLayoutTransition& transition) {
     }
 
     return std::make_tuple(srcAccessMask, dstAccessMask, srcStage, dstStage,
-            getVkImageLayout(transition.oldLayout), getVkImageLayout(transition.newLayout));
+            getVkLayout(transition.oldLayout), getVkLayout(transition.newLayout));
 }
 
 }// anonymous namespace
 
-VkImageViewType VulkanImageUtility::getViewType(SamplerType target) {
-    switch (target) {
-        case SamplerType::SAMPLER_CUBEMAP:
-            return VK_IMAGE_VIEW_TYPE_CUBE;
-        case SamplerType::SAMPLER_2D_ARRAY:
-            return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
-        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
-            return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
-        case SamplerType::SAMPLER_3D:
-            return VK_IMAGE_VIEW_TYPE_3D;
-        default:
-            return VK_IMAGE_VIEW_TYPE_2D;
-    }
-}
-
-void VulkanImageUtility::transitionLayout(VkCommandBuffer cmdbuffer,
+void transitionLayout(VkCommandBuffer cmdbuffer,
         VulkanLayoutTransition transition) {
     if (transition.oldLayout == transition.newLayout) {
         return;
@@ -197,10 +154,6 @@ void VulkanImageUtility::transitionLayout(VkCommandBuffer cmdbuffer,
     vkCmdPipelineBarrier(cmdbuffer, srcStage, dstStage, 0, 0, nullptr, 0, nullptr, 1, &barrier);
 }
 
-VkImageLayout VulkanImageUtility::getVkLayout(VulkanLayout layout) {
-    return getVkImageLayout(layout);
-}
-
 }// namespace filament::backend
 
 bool operator<(const VkImageSubresourceRange& a, const VkImageSubresourceRange& b) {
@@ -217,12 +170,12 @@ bool operator<(const VkImageSubresourceRange& a, const VkImageSubresourceRange&
     return false;
 }
 
-#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION | FVK_DEBUG_TEXTURE)
+#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION) || FVK_ENABLED(FVK_DEBUG_TEXTURE)
 #define CASE(VALUE)                                                                                \
     case filament::backend::VulkanLayout::VALUE: {                                                 \
         out << #VALUE;                                                                             \
         out << " ["                                                                                \
-            << filament::backend::VulkanImageUtility::getVkLayout(                                 \
+            << filament::backend::imgutil::getVkLayout(                                            \
                        filament::backend::VulkanLayout::VALUE)                                     \
             << "]";                                                                                \
         break;                                                                                     \

filament/backend/src/vulkan/VulkanImageUtility.h

@@ -59,44 +59,85 @@ struct VulkanLayoutTransition {
     VkImageSubresourceRange subresources;
 };
 
-class VulkanImageUtility {
-public:
-    static VkImageViewType getViewType(SamplerType target);
+namespace imgutil {
 
-    inline static VulkanLayout getDefaultLayout(TextureUsage usage) {
-        if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
-            if (any(usage & TextureUsage::SAMPLEABLE)) {
-                return VulkanLayout::DEPTH_SAMPLER;
-            } else {
-                return VulkanLayout::DEPTH_ATTACHMENT;
-            }
-        }
+inline VkImageViewType getViewType(SamplerType target) {
+    switch (target) {
+        case SamplerType::SAMPLER_CUBEMAP:
+            return VK_IMAGE_VIEW_TYPE_CUBE;
+        case SamplerType::SAMPLER_2D_ARRAY:
+            return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
+        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
+            return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
+        case SamplerType::SAMPLER_3D:
+            return VK_IMAGE_VIEW_TYPE_3D;
+        default:
+            return VK_IMAGE_VIEW_TYPE_2D;
+    }
+}
 
-        if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
-            return VulkanLayout::COLOR_ATTACHMENT;
+inline VulkanLayout getDefaultLayout(TextureUsage usage) {
+    if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
+        if (any(usage & TextureUsage::SAMPLEABLE)) {
+            return VulkanLayout::DEPTH_SAMPLER;
+        } else {
+            return VulkanLayout::DEPTH_ATTACHMENT;
         }
-        // Finally, the layout for an immutable texture is optimal read-only.        
-        return VulkanLayout::READ_ONLY;
     }
 
-    inline static VulkanLayout getDefaultLayout(VkImageUsageFlags vkusage) {
-        TextureUsage usage {};
-        if (vkusage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
-            usage = usage | TextureUsage::DEPTH_ATTACHMENT;
-        }
-        if (vkusage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
-            usage = usage | TextureUsage::COLOR_ATTACHMENT;
-        }
-        if (vkusage & VK_IMAGE_USAGE_SAMPLED_BIT) {
-            usage = usage | TextureUsage::SAMPLEABLE;
-        }
-        return getDefaultLayout(usage);
+    if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
+        return VulkanLayout::COLOR_ATTACHMENT;
     }
+    // Finally, the layout for an immutable texture is optimal read-only.
+    return VulkanLayout::READ_ONLY;
+}
 
-    static VkImageLayout getVkLayout(VulkanLayout layout);
-    
-    static void transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
-};
+inline VulkanLayout getDefaultLayout(VkImageUsageFlags vkusage) {
+    TextureUsage usage{};
+    if (vkusage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+        usage = usage | TextureUsage::DEPTH_ATTACHMENT;
+    }
+    if (vkusage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+        usage = usage | TextureUsage::COLOR_ATTACHMENT;
+    }
+    if (vkusage & VK_IMAGE_USAGE_SAMPLED_BIT) {
+        usage = usage | TextureUsage::SAMPLEABLE;
+    }
+    return getDefaultLayout(usage);
+}
+
+constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
+    switch (layout) {
+        case VulkanLayout::UNDEFINED:
+            return VK_IMAGE_LAYOUT_UNDEFINED;
+        case VulkanLayout::READ_WRITE:
+            return VK_IMAGE_LAYOUT_GENERAL;
+        case VulkanLayout::READ_ONLY:
+            return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+        case VulkanLayout::TRANSFER_SRC:
+            return VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
+        case VulkanLayout::TRANSFER_DST:
+            return VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+        case VulkanLayout::DEPTH_ATTACHMENT:
+            return VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
+        case VulkanLayout::DEPTH_SAMPLER:
+            return VK_IMAGE_LAYOUT_GENERAL;
+        case VulkanLayout::PRESENT:
+            return VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
+        // Filament sometimes samples from one miplevel while writing to another level in the
+        // same texture (e.g. bloom does this). Moreover we'd like to avoid lots of expensive
+        // layout transitions. So, keep it simple and use GENERAL for all color-attachable
+        // textures.
+        case VulkanLayout::COLOR_ATTACHMENT:
+            return VK_IMAGE_LAYOUT_GENERAL;
+        case VulkanLayout::COLOR_ATTACHMENT_RESOLVE:
+            return VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
+    }
+}
+
+void transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
+
+} // namespace imgutil
 
 } // namespace filament::backend
 

filament/backend/src/vulkan/VulkanMemory.h

@@ -31,4 +31,8 @@
 
 #include "vk_mem_alloc.h"
 
+VK_DEFINE_HANDLE(VmaAllocator)
+VK_DEFINE_HANDLE(VmaAllocation)
+VK_DEFINE_HANDLE(VmaPool)
+
 #endif // TNT_FILAMENT_BACKEND_VULKANMEMORY_H

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -14,8 +14,9 @@
  * limitations under the License.
  */
 
-#include "vulkan/VulkanMemory.h"
-#include "vulkan/VulkanPipelineCache.h"
+#include "VulkanPipelineCache.h"
+#include "VulkanMemory.h"
+#include "caching/VulkanDescriptorSetManager.h"
 
 #include <utils/Log.h>
 #include <utils/Panic.h>
@@ -34,63 +35,9 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-static VkShaderStageFlags getShaderStageFlags(VulkanPipelineCache::UsageFlags key, uint16_t binding) {
-    // NOTE: if you modify this function, you also need to modify getUsageFlags.
-    assert_invariant(binding < MAX_SAMPLER_COUNT);
-    VkShaderStageFlags flags = 0;
-    if (key.test(binding)) {
-        flags |= VK_SHADER_STAGE_VERTEX_BIT;
-    }
-    if (key.test(MAX_SAMPLER_COUNT + binding)) {
-        flags |= VK_SHADER_STAGE_FRAGMENT_BIT;
-    }
-    return flags;
-}
-
-VulkanPipelineCache::UsageFlags
-VulkanPipelineCache::getUsageFlags(uint16_t binding, ShaderStageFlags flags, UsageFlags src) {
-    // NOTE: if you modify this function, you also need to modify getShaderStageFlags.
-    assert_invariant(binding < MAX_SAMPLER_COUNT);
-    if (any(flags & ShaderStageFlags::VERTEX)) {
-        src.set(binding);
-    }
-    if (any(flags & ShaderStageFlags::FRAGMENT)) {
-        src.set(MAX_SAMPLER_COUNT + binding);
-    }
-    // TODO: add support for compute by extending SHADER_MODULE_COUNT and ensuring UsageFlags
-    // has 186 bits (MAX_SAMPLER_COUNT * 3)
-    // assert_invariant(!any(flags & ~(ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT)));
-    return src;
-}
-
-VulkanPipelineCache::UsageFlags VulkanPipelineCache::disableUsageFlags(uint16_t binding,
-        UsageFlags src) {
-    src.unset(binding);
-    src.unset(MAX_SAMPLER_COUNT + binding);
-    return src;
-}
-
-VulkanPipelineCache::VulkanPipelineCache(VulkanResourceAllocator* allocator)
-    : mResourceAllocator(allocator),
-      mPipelineBoundResources(allocator) {
-    mDummyBufferWriteInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
-    mDummyBufferWriteInfo.pNext = nullptr;
-    mDummyBufferWriteInfo.dstArrayElement = 0;
-    mDummyBufferWriteInfo.descriptorCount = 1;
-    mDummyBufferWriteInfo.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
-    mDummyBufferWriteInfo.pImageInfo = nullptr;
-    mDummyBufferWriteInfo.pBufferInfo = &mDummyBufferInfo;
-    mDummyBufferWriteInfo.pTexelBufferView = nullptr;
-
-    mDummyTargetInfo.imageLayout = VulkanImageUtility::getVkLayout(VulkanLayout::READ_ONLY);
-    mDummyTargetWriteInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
-    mDummyTargetWriteInfo.pNext = nullptr;
-    mDummyTargetWriteInfo.dstArrayElement = 0;
-    mDummyTargetWriteInfo.descriptorCount = 1;
-    mDummyTargetWriteInfo.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
-    mDummyTargetWriteInfo.pImageInfo = &mDummyTargetInfo;
-    mDummyTargetWriteInfo.pBufferInfo = nullptr;
-    mDummyTargetWriteInfo.pTexelBufferView = nullptr;
+VulkanPipelineCache::VulkanPipelineCache(VkDevice device, VmaAllocator allocator)
+    : mDevice(device),
+      mAllocator(allocator) {
 }
 
 VulkanPipelineCache::~VulkanPipelineCache() {
@@ -98,244 +45,47 @@ VulkanPipelineCache::~VulkanPipelineCache() {
     // be explicit about teardown order of various components.
 }
 
-void VulkanPipelineCache::setDevice(VkDevice device, VmaAllocator allocator) {
-    assert_invariant(mDevice == VK_NULL_HANDLE);
-    mDevice = device;
-    mAllocator = allocator;
-    mDescriptorPool = createDescriptorPool(mDescriptorPoolSize);
-
-    // Formulate some dummy objects and dummy descriptor info used only for clearing out unused
-    // bindings. This is especially crucial after a texture has been destroyed. Since core Vulkan
-    // does not allow specifying VK_NULL_HANDLE without the robustness2 extension, we would need to
-    // change the pipeline layout more frequently if we wanted to get rid of these dummy objects.
-
-    VkBufferCreateInfo bufferInfo {
-        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
-        .size = 16,
-        .usage = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT,
-    };
-    VmaAllocationCreateInfo allocInfo { .usage = VMA_MEMORY_USAGE_GPU_ONLY };
-    vmaCreateBuffer(mAllocator, &bufferInfo, &allocInfo, &mDummyBuffer, &mDummyMemory, nullptr);
-
-    mDummyBufferInfo.buffer = mDummyBuffer;
-    mDummyBufferInfo.range = bufferInfo.size;
+void VulkanPipelineCache::bindLayout(VkPipelineLayout layout) noexcept {
+    mPipelineRequirements.layout = layout;
 }
 
-bool VulkanPipelineCache::bindDescriptors(VkCommandBuffer cmdbuffer) noexcept {
-    DescriptorMap::iterator descriptorIter = mDescriptorSets.find(mDescriptorRequirements);
-
-    // Check if the required descriptors are already bound. If so, there's no need to do anything.
-    if (DescEqual equals; UTILS_LIKELY(equals(mBoundDescriptor, mDescriptorRequirements))) {
-
-        // If the pipeline state during an app's first draw call happens to match the default state
-        // vector of the cache, then the cache is uninitialized and we should not return early.
-        if (UTILS_LIKELY(!mDescriptorSets.empty())) {
-
-            // Since the descriptors are already bound, they should be found in the cache.
-            assert_invariant(descriptorIter != mDescriptorSets.end());
-
-            // Update the LRU "time stamp" (really a count of cmd buf submissions) before returning.
-            descriptorIter.value().lastUsed = mCurrentTime;
-            return true;
-        }
-    }
-
+VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::getOrCreatePipeline() noexcept {
     // If a cached object exists, re-use it, otherwise create a new one.
-    DescriptorCacheEntry* cacheEntry = UTILS_LIKELY(descriptorIter != mDescriptorSets.end()) ?
-            &descriptorIter.value() : createDescriptorSets();
-
-    // If a descriptor set overflow occurred, allow higher levels to handle it gracefully.
-    assert_invariant(cacheEntry != nullptr);
-    if (UTILS_UNLIKELY(cacheEntry == nullptr)) {
-        return false;
+    if (PipelineMap::iterator pipelineIter = mPipelines.find(mPipelineRequirements);
+            pipelineIter != mPipelines.end()) {
+        auto& pipeline = pipelineIter.value();
+        pipeline.lastUsed = mCurrentTime;
+        return &pipeline;
     }
-
-    cacheEntry->lastUsed = mCurrentTime;
-    mBoundDescriptor = mDescriptorRequirements;
-    // This passes the currently "bound" uniform buffer objects to pipeline that will be used in the
-    // draw call.
-    auto resourceEntry = mDescriptorResources.find(cacheEntry->id);
-    if (resourceEntry == mDescriptorResources.end()) {
-        mDescriptorResources[cacheEntry->id]
-                = std::make_unique<VulkanAcquireOnlyResourceManager>(mResourceAllocator);
-        resourceEntry = mDescriptorResources.find(cacheEntry->id);
-    }
-    resourceEntry->second->acquireAll(&mPipelineBoundResources);
-
-    vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
-            getOrCreatePipelineLayout()->handle, 0, VulkanPipelineCache::DESCRIPTOR_TYPE_COUNT,
-            cacheEntry->handles.data(), 0, nullptr);
-
-    return true;
+    auto ret = createPipeline();
+    ret->lastUsed = mCurrentTime;
+    return ret;
 }
 
-bool VulkanPipelineCache::bindPipeline(VulkanCommandBuffer* commands) noexcept {
+void VulkanPipelineCache::bindPipeline(VulkanCommandBuffer* commands) {
     VkCommandBuffer const cmdbuffer = commands->buffer();
 
-    PipelineMap::iterator pipelineIter = mPipelines.find(mPipelineRequirements);
-
+    PipelineCacheEntry* cacheEntry = getOrCreatePipeline();
     // Check if the required pipeline is already bound.
-    if (PipelineEqual equals; UTILS_LIKELY(equals(mBoundPipeline, mPipelineRequirements))) {
-        assert_invariant(pipelineIter != mPipelines.end());
-        pipelineIter.value().lastUsed = mCurrentTime;
-        return true;
+    if (cacheEntry->handle == commands->pipeline()) {
+        return;
     }
 
-    // If a cached object exists, re-use it, otherwise create a new one.
-    PipelineCacheEntry* cacheEntry = UTILS_LIKELY(pipelineIter != mPipelines.end()) ?
-            &pipelineIter.value() : createPipeline();
-
     // If an error occurred, allow higher levels to handle it gracefully.
-    assert_invariant(cacheEntry != nullptr);
-    if (UTILS_UNLIKELY(cacheEntry == nullptr)) {
-        return false;
-    }
-
-    cacheEntry->lastUsed = mCurrentTime;
-    getOrCreatePipelineLayout()->lastUsed = mCurrentTime;
+    assert_invariant(cacheEntry != nullptr && "Failed to create/find pipeline");
 
     mBoundPipeline = mPipelineRequirements;
-
     vkCmdBindPipeline(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, cacheEntry->handle);
-    return true;
+    commands->setPipeline(cacheEntry->handle);
 }
 
 void VulkanPipelineCache::bindScissor(VkCommandBuffer cmdbuffer, VkRect2D scissor) noexcept {
-    if (UTILS_UNLIKELY(!equivalent(mCurrentScissor, scissor))) {
-        mCurrentScissor = scissor;
-        vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
-    }
-}
-
-VulkanPipelineCache::DescriptorCacheEntry* VulkanPipelineCache::createDescriptorSets() noexcept {
-    PipelineLayoutCacheEntry* layoutCacheEntry = getOrCreatePipelineLayout();
-
-    DescriptorCacheEntry descriptorCacheEntry = {
-        .pipelineLayout = mPipelineRequirements.layout,
-        .id = mDescriptorCacheEntryCount++,
-    };
-
-    // Each of the arenas for this particular layout are guaranteed to have the same size. Check
-    // the first arena to see if any descriptor sets are available that can be re-claimed. If not,
-    // create brand new ones (one for each type). They will be added to the arena later, after they
-    // are no longer used. This occurs during the cleanup phase during command buffer submission.
-    auto& descriptorSetArenas = layoutCacheEntry->descriptorSetArenas;
-    if (descriptorSetArenas[0].empty()) {
-
-        // If allocating a new descriptor set from the pool would cause it to overflow, then
-        // recreate the pool. The number of descriptor sets that have already been allocated from
-        // the pool is the sum of the "active" descriptor sets (mDescriptorSets) and the "dormant"
-        // descriptor sets (mDescriptorArenasCount).
-        //
-        // NOTE: technically both sides of the inequality below should be multiplied by
-        // DESCRIPTOR_TYPE_COUNT to get the true number of descriptor sets.
-        if (mDescriptorSets.size() + mDescriptorArenasCount + 1 > mDescriptorPoolSize) {
-            growDescriptorPool();
-        }
-
-        VkDescriptorSetAllocateInfo allocInfo = {};
-        allocInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
-        allocInfo.descriptorPool = mDescriptorPool;
-        allocInfo.descriptorSetCount = DESCRIPTOR_TYPE_COUNT;
-        allocInfo.pSetLayouts = layoutCacheEntry->descriptorSetLayouts.data();
-        VkResult error = vkAllocateDescriptorSets(mDevice, &allocInfo,
-                descriptorCacheEntry.handles.data());
-        assert_invariant(error == VK_SUCCESS);
-        if (error != VK_SUCCESS) {
-            return nullptr;
-        }
-    } else {
-        for (uint32_t i = 0; i < DESCRIPTOR_TYPE_COUNT; ++i) {
-            descriptorCacheEntry.handles[i] = descriptorSetArenas[i].back();
-            descriptorSetArenas[i].pop_back();
-        }
-        assert_invariant(mDescriptorArenasCount > 0);
-        mDescriptorArenasCount--;
-    }
-
-    // Rewrite every binding in the new descriptor sets.
-    VkDescriptorBufferInfo descriptorBuffers[UBUFFER_BINDING_COUNT];
-    VkDescriptorImageInfo descriptorSamplers[SAMPLER_BINDING_COUNT];
-    VkDescriptorImageInfo descriptorInputAttachments[INPUT_ATTACHMENT_COUNT];
-    VkWriteDescriptorSet descriptorWrites[UBUFFER_BINDING_COUNT + SAMPLER_BINDING_COUNT +
-            INPUT_ATTACHMENT_COUNT];
-    uint32_t nwrites = 0;
-    VkWriteDescriptorSet* writes = descriptorWrites;
-    nwrites = 0;
-    for (uint32_t binding = 0; binding < UBUFFER_BINDING_COUNT; binding++) {
-        VkWriteDescriptorSet& writeInfo = writes[nwrites++];
-        if (mDescriptorRequirements.uniformBuffers[binding]) {
-            VkDescriptorBufferInfo& bufferInfo = descriptorBuffers[binding];
-            bufferInfo.buffer = mDescriptorRequirements.uniformBuffers[binding];
-            bufferInfo.offset = mDescriptorRequirements.uniformBufferOffsets[binding];
-            bufferInfo.range = mDescriptorRequirements.uniformBufferSizes[binding];
-
-            // We store size with 32 bits, so our "WHOLE" sentinel is different from Vk.
-            if (bufferInfo.range == WHOLE_SIZE) {
-                bufferInfo.range = VK_WHOLE_SIZE;
-            }
-
-            writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
-            writeInfo.pNext = nullptr;
-            writeInfo.dstArrayElement = 0;
-            writeInfo.descriptorCount = 1;
-            writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
-            writeInfo.pImageInfo = nullptr;
-            writeInfo.pBufferInfo = &bufferInfo;
-            writeInfo.pTexelBufferView = nullptr;
-        } else {
-            writeInfo = mDummyBufferWriteInfo;
-            assert_invariant(mDummyBufferWriteInfo.pBufferInfo->buffer);
-        }
-        assert_invariant(writeInfo.pBufferInfo->buffer);
-        writeInfo.dstSet = descriptorCacheEntry.handles[0];
-        writeInfo.dstBinding = binding;
-    }
-    for (uint32_t binding = 0; binding < SAMPLER_BINDING_COUNT; binding++) {
-        if (mDescriptorRequirements.samplers[binding].sampler) {
-            VkWriteDescriptorSet& writeInfo = writes[nwrites++];
-            VkDescriptorImageInfo& imageInfo = descriptorSamplers[binding];
-            imageInfo = mDescriptorRequirements.samplers[binding];
-            writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
-            writeInfo.pNext = nullptr;
-            writeInfo.dstArrayElement = 0;
-            writeInfo.descriptorCount = 1;
-            writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
-            writeInfo.pImageInfo = &imageInfo;
-            writeInfo.pBufferInfo = nullptr;
-            writeInfo.pTexelBufferView = nullptr;
-            writeInfo.dstSet = descriptorCacheEntry.handles[1];
-            writeInfo.dstBinding = binding;
-        }
-    }
-    for (uint32_t binding = 0; binding < INPUT_ATTACHMENT_COUNT; binding++) {
-        if (mDescriptorRequirements.inputAttachments[binding].imageView) {
-            VkWriteDescriptorSet& writeInfo = writes[nwrites++];
-            VkDescriptorImageInfo& imageInfo = descriptorInputAttachments[binding];
-            imageInfo = mDescriptorRequirements.inputAttachments[binding];
-            writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
-            writeInfo.pNext = nullptr;
-            writeInfo.dstArrayElement = 0;
-            writeInfo.descriptorCount = 1;
-            writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
-            writeInfo.pImageInfo = &imageInfo;
-            writeInfo.pBufferInfo = nullptr;
-            writeInfo.pTexelBufferView = nullptr;
-            writeInfo.dstSet = descriptorCacheEntry.handles[2];
-            writeInfo.dstBinding = binding;
-        }
-    }
-
-    vkUpdateDescriptorSets(mDevice, nwrites, writes, 0, nullptr);
-
-    return &mDescriptorSets.emplace(mDescriptorRequirements, descriptorCacheEntry).first.value();
+    vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
 }
 
 VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() noexcept {
     assert_invariant(mPipelineRequirements.shaders[0] && "Vertex shader is not bound.");
-
-    PipelineLayoutCacheEntry* layout = getOrCreatePipelineLayout();
-    assert_invariant(layout);
+    assert_invariant(mPipelineRequirements.layout && "No pipeline layout specified");
 
     VkPipelineShaderStageCreateInfo shaderStages[SHADER_MODULE_COUNT];
     shaderStages[0] = VkPipelineShaderStageCreateInfo{};
@@ -404,7 +154,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
 
     VkGraphicsPipelineCreateInfo pipelineCreateInfo = {};
     pipelineCreateInfo.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO;
-    pipelineCreateInfo.layout = layout->handle;
+    pipelineCreateInfo.layout = mPipelineRequirements.layout;
     pipelineCreateInfo.renderPass = mPipelineRequirements.renderPass;
     pipelineCreateInfo.subpass = mPipelineRequirements.subpassIndex;
     pipelineCreateInfo.stageCount = hasFragmentShader ? SHADER_MODULE_COUNT : 1;
@@ -498,68 +248,6 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     return &mPipelines.emplace(mPipelineRequirements, cacheEntry).first.value();
 }
 
-VulkanPipelineCache::PipelineLayoutCacheEntry* VulkanPipelineCache::getOrCreatePipelineLayout() noexcept {
-    auto iter = mPipelineLayouts.find(mPipelineRequirements.layout);
-    if (UTILS_LIKELY(iter != mPipelineLayouts.end())) {
-        return &iter.value();
-    }
-
-    PipelineLayoutCacheEntry cacheEntry = {};
-
-    VkDescriptorSetLayoutBinding binding = {};
-    binding.descriptorCount = 1; // NOTE: We never use arrays-of-blocks.
-    binding.stageFlags = VK_SHADER_STAGE_ALL_GRAPHICS; // NOTE: This is potentially non-optimal.
-
-    // First create the descriptor set layout for UBO's.
-    VkDescriptorSetLayoutBinding ubindings[UBUFFER_BINDING_COUNT];
-    binding.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
-    for (uint32_t i = 0; i < UBUFFER_BINDING_COUNT; i++) {
-        binding.binding = i;
-        ubindings[i] = binding;
-    }
-    VkDescriptorSetLayoutCreateInfo dlinfo = {};
-    dlinfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
-    dlinfo.bindingCount = UBUFFER_BINDING_COUNT;
-    dlinfo.pBindings = ubindings;
-    vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &cacheEntry.descriptorSetLayouts[0]);
-
-    // Next create the descriptor set layout for samplers.
-    VkDescriptorSetLayoutBinding sbindings[SAMPLER_BINDING_COUNT];
-    binding.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
-    for (uint32_t i = 0; i < SAMPLER_BINDING_COUNT; i++) {
-        binding.stageFlags = getShaderStageFlags(mPipelineRequirements.layout, i);
-        binding.binding = i;
-        sbindings[i] = binding;
-    }
-    dlinfo.bindingCount = SAMPLER_BINDING_COUNT;
-    dlinfo.pBindings = sbindings;
-    vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &cacheEntry.descriptorSetLayouts[1]);
-
-    // Next create the descriptor set layout for input attachments.
-    VkDescriptorSetLayoutBinding tbindings[INPUT_ATTACHMENT_COUNT];
-    binding.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
-    binding.stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT;
-    for (uint32_t i = 0; i < INPUT_ATTACHMENT_COUNT; i++) {
-        binding.binding = i;
-        tbindings[i] = binding;
-    }
-    dlinfo.bindingCount = INPUT_ATTACHMENT_COUNT;
-    dlinfo.pBindings = tbindings;
-    vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &cacheEntry.descriptorSetLayouts[2]);
-
-    // Create VkPipelineLayout based on how to resources are bounded.
-    VkPipelineLayoutCreateInfo pPipelineLayoutCreateInfo = {};
-    pPipelineLayoutCreateInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
-    pPipelineLayoutCreateInfo.setLayoutCount = cacheEntry.descriptorSetLayouts.size();
-    pPipelineLayoutCreateInfo.pSetLayouts = cacheEntry.descriptorSetLayouts.data();
-    VkResult result = vkCreatePipelineLayout(mDevice, &pPipelineLayoutCreateInfo, VKALLOC,
-            &cacheEntry.handle);
-    if (UTILS_UNLIKELY(result != VK_SUCCESS)) {
-        return nullptr;
-    }
-    return &mPipelineLayouts.emplace(mPipelineRequirements.layout, cacheEntry).first.value();
-}
-
 void VulkanPipelineCache::bindProgram(VulkanProgram* program) noexcept {
     mPipelineRequirements.shaders[0] = program->getVertexShader();
     mPipelineRequirements.shaders[1] = program->getFragmentShader();
@@ -600,97 +288,15 @@ void VulkanPipelineCache::bindVertexArray(VkVertexInputAttributeDescription cons
     }
 }
 
-VulkanPipelineCache::UniformBufferBinding VulkanPipelineCache::getUniformBufferBinding(
-        uint32_t bindingIndex) const noexcept {
-    auto& key = mDescriptorRequirements;
-    return {
-        key.uniformBuffers[bindingIndex],
-        key.uniformBufferOffsets[bindingIndex],
-        key.uniformBufferSizes[bindingIndex],
-    };
-}
-
-void VulkanPipelineCache::unbindUniformBuffer(VkBuffer uniformBuffer) noexcept {
-    auto& key = mDescriptorRequirements;
-    for (uint32_t bindingIndex = 0u; bindingIndex < UBUFFER_BINDING_COUNT; ++bindingIndex) {
-        if (key.uniformBuffers[bindingIndex] == uniformBuffer) {
-            key.uniformBuffers[bindingIndex] = {};
-            key.uniformBufferSizes[bindingIndex] = {};
-            key.uniformBufferOffsets[bindingIndex] = {};
-        }
-    }
-}
-
-void VulkanPipelineCache::unbindImageView(VkImageView imageView) noexcept {
-    for (auto& sampler : mDescriptorRequirements.samplers) {
-        if (sampler.imageView == imageView) {
-            sampler = {};
-        }
-    }
-    for (auto& target : mDescriptorRequirements.inputAttachments) {
-        if (target.imageView == imageView) {
-            target = {};
-        }
-    }
-}
-
-void VulkanPipelineCache::bindUniformBufferObject(uint32_t bindingIndex,
-        VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept {
-    VkBuffer buffer = bufferObject->buffer.getGpuBuffer();
-
-    ASSERT_POSTCONDITION(bindingIndex < UBUFFER_BINDING_COUNT,
-            "Uniform bindings overflow: index = %d, capacity = %d.", bindingIndex,
-            UBUFFER_BINDING_COUNT);
-    auto& key = mDescriptorRequirements;
-    key.uniformBuffers[bindingIndex] = buffer;
-
-    if (size == VK_WHOLE_SIZE) {
-        size = WHOLE_SIZE;
-    }
-
-    assert_invariant(offset <= 0xffffffffu);
-    assert_invariant(size <= 0xffffffffu);
-
-    key.uniformBufferOffsets[bindingIndex] = offset;
-    key.uniformBufferSizes[bindingIndex] = size;
-
-    mPipelineBoundResources.acquire(bufferObject);
-}
-
-void VulkanPipelineCache::bindSamplers(VkDescriptorImageInfo samplers[SAMPLER_BINDING_COUNT],
-        VulkanTexture* textures[SAMPLER_BINDING_COUNT], UsageFlags flags) noexcept {
-    for (uint32_t bindingIndex = 0; bindingIndex < SAMPLER_BINDING_COUNT; bindingIndex++) {
-        mDescriptorRequirements.samplers[bindingIndex] = samplers[bindingIndex];
-        if (textures[bindingIndex]) {
-            mPipelineBoundResources.acquire(textures[bindingIndex]);
-        }
-    }
-    mPipelineRequirements.layout = flags;
-}
-
-void VulkanPipelineCache::bindInputAttachment(uint32_t bindingIndex,
-        VkDescriptorImageInfo targetInfo) noexcept {
-    ASSERT_POSTCONDITION(bindingIndex < INPUT_ATTACHMENT_COUNT,
-            "Input attachment bindings overflow: index = %d, capacity = %d.",
-            bindingIndex, INPUT_ATTACHMENT_COUNT);
-    mDescriptorRequirements.inputAttachments[bindingIndex] = targetInfo;
-}
-
 void VulkanPipelineCache::terminate() noexcept {
-    // Symmetric to createLayoutsAndDescriptors.
-    destroyLayoutsAndDescriptors();
     for (auto& iter : mPipelines) {
         vkDestroyPipeline(mDevice, iter.second.handle, VKALLOC);
     }
-    mPipelineBoundResources.clear();
     mPipelines.clear();
     mBoundPipeline = {};
-    vmaDestroyBuffer(mAllocator, mDummyBuffer, mDummyMemory);
-    mDummyBuffer = VK_NULL_HANDLE;
-    mDummyMemory = VK_NULL_HANDLE;
 }
 
-void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
+void VulkanPipelineCache::gc() noexcept {
     // The timestamp associated with a given cache entry represents "time" as a count of flush
     // events since the cache was constructed. If any cache entry was most recently used over
     // FVK_MAX_PIPELINE_AGE flush events in the past, then we can be sure that it is no longer
@@ -700,194 +306,22 @@ void VulkanPipelineCache::onCommandBuffer(const VulkanCommandBuffer& commands) {
     // The Vulkan spec says: "When a command buffer begins recording, all state in that command
     // buffer is undefined." Therefore, we need to clear all bindings at this time.
     mBoundPipeline = {};
-    mBoundDescriptor = {};
     mCurrentScissor = {};
 
     // NOTE: Due to robin_map restrictions, we cannot use auto or range-based loops.
 
-    // Check if any bundles in the cache are no longer in use by any command buffer. Descriptors
-    // from unused bundles are moved back to their respective arenas.
-    using ConstDescIterator = decltype(mDescriptorSets)::const_iterator;
-    for (ConstDescIterator iter = mDescriptorSets.begin(); iter != mDescriptorSets.end();) {
-        const DescriptorCacheEntry& cacheEntry = iter.value();
-        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
-            auto& arenas = mPipelineLayouts[cacheEntry.pipelineLayout].descriptorSetArenas;
-            for (uint32_t i = 0; i < DESCRIPTOR_TYPE_COUNT; ++i) {
-                arenas[i].push_back(cacheEntry.handles[i]);
-            }
-            ++mDescriptorArenasCount;
-            mDescriptorResources.erase(cacheEntry.id);
-            iter = mDescriptorSets.erase(iter);
-        } else {
-            ++iter;
-        }
-    }
-
     // Evict any pipelines that have not been used in a while.
     // Any pipeline older than FVK_MAX_COMMAND_BUFFERS can be safely destroyed.
-    using ConstPipeIterator = decltype(mPipelines)::const_iterator;
-    for (ConstPipeIterator iter = mPipelines.begin(); iter != mPipelines.end();) {
-        const PipelineCacheEntry& cacheEntry = iter.value();
-        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
-            vkDestroyPipeline(mDevice, iter->second.handle, VKALLOC);
-            iter = mPipelines.erase(iter);
-        } else {
-            ++iter;
-        }
-    }
-
-    // Evict any layouts that have not been used in a while.
-    using ConstLayoutIterator = decltype(mPipelineLayouts)::const_iterator;
-    for (ConstLayoutIterator iter = mPipelineLayouts.begin(); iter != mPipelineLayouts.end();) {
-        const PipelineLayoutCacheEntry& cacheEntry = iter.value();
-        if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
-            vkDestroyPipelineLayout(mDevice, iter->second.handle, VKALLOC);
-            for (auto setLayout : iter->second.descriptorSetLayouts) {
-                #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
-                PipelineLayoutKey key = iter.key();
-                for (auto& pair : mDescriptorSets) {
-                    assert_invariant(pair.second.pipelineLayout != key);
-                }
-                #endif
-                vkDestroyDescriptorSetLayout(mDevice, setLayout, VKALLOC);
-            }
-            auto& arenas = iter->second.descriptorSetArenas;
-            assert_invariant(mDescriptorArenasCount >= arenas[0].size());
-            mDescriptorArenasCount -= arenas[0].size();
-            for (auto& arena : arenas) {
-                vkFreeDescriptorSets(mDevice, mDescriptorPool, arena.size(), arena.data());
-            }
-            iter = mPipelineLayouts.erase(iter);
-        } else {
-            ++iter;
-        }
-    }
-
-    // If there are no descriptors from any extinct pool that are still in use, we can safely
-    // destroy the extinct pools, which implicitly frees their associated descriptor sets.
-    bool canPurgeExtinctPools = true;
-    for (auto& bundle : mExtinctDescriptorBundles) {
-        if (bundle.lastUsed + FVK_MAX_PIPELINE_AGE >= mCurrentTime) {
-            canPurgeExtinctPools = false;
-            break;
-        }
-    }
-    if (canPurgeExtinctPools) {
-        for (VkDescriptorPool pool : mExtinctDescriptorPools) {
-            vkDestroyDescriptorPool(mDevice, pool, VKALLOC);
-        }
-        mExtinctDescriptorPools.clear();
-
-        for (auto const& entry : mExtinctDescriptorBundles) {
-            mDescriptorResources.erase(entry.id);
-        }
-        mExtinctDescriptorBundles.clear();
-    }
-}
-
-VkDescriptorPool VulkanPipelineCache::createDescriptorPool(uint32_t size) const {
-    VkDescriptorPoolSize poolSizes[DESCRIPTOR_TYPE_COUNT] = {};
-    VkDescriptorPoolCreateInfo poolInfo {
-        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
-        .pNext = nullptr,
-        .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
-        .maxSets = size * DESCRIPTOR_TYPE_COUNT,
-        .poolSizeCount = DESCRIPTOR_TYPE_COUNT,
-        .pPoolSizes = poolSizes
-    };
-    poolSizes[0].type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
-    poolSizes[0].descriptorCount = poolInfo.maxSets * UBUFFER_BINDING_COUNT;
-    poolSizes[1].type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
-    poolSizes[1].descriptorCount = poolInfo.maxSets * SAMPLER_BINDING_COUNT;
-    poolSizes[2].type = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
-    poolSizes[2].descriptorCount = poolInfo.maxSets * INPUT_ATTACHMENT_COUNT;
-
-    VkDescriptorPool pool;
-    const UTILS_UNUSED VkResult result = vkCreateDescriptorPool(mDevice, &poolInfo, VKALLOC, &pool);
-    assert_invariant(result == VK_SUCCESS);
-    return pool;
-}
-
-void VulkanPipelineCache::destroyLayoutsAndDescriptors() noexcept {
-    // Our current descriptor set strategy can cause the # of descriptor sets to explode in certain
-    // situations, so it's interesting to report the number that get stuffed into the cache.
-    #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
-    utils::slog.d << "Destroying " << mDescriptorSets.size() << " bundles of descriptor sets."
-            << utils::io::endl;
-    #endif
-
-    mDescriptorSets.clear();
-
-    // Our current layout bundle strategy can cause the # of layout bundles to explode in certain
-    // situations, so it's interesting to report the number that get stuffed into the cache.
-    #if FVK_ENABLED(FVK_DEBUG_PIPELINE_CACHE)
-    utils::slog.d << "Destroying " << mPipelineLayouts.size() << " pipeline layouts."
-                  << utils::io::endl;
-    #endif
-
-    for (auto& iter : mPipelineLayouts) {
-        vkDestroyPipelineLayout(mDevice, iter.second.handle, VKALLOC);
-        for (auto setLayout : iter.second.descriptorSetLayouts) {
-            vkDestroyDescriptorSetLayout(mDevice, setLayout, VKALLOC);
-        }
-        // There is no need to free descriptor sets individually since destroying the VkDescriptorPool
-        // implicitly frees them.
-    }
-    mPipelineLayouts.clear();
-    vkDestroyDescriptorPool(mDevice, mDescriptorPool, VKALLOC);
-    mDescriptorPool = VK_NULL_HANDLE;
-
-    for (VkDescriptorPool pool : mExtinctDescriptorPools) {
-        vkDestroyDescriptorPool(mDevice, pool, VKALLOC);
-    }
-    mExtinctDescriptorPools.clear();
-    mExtinctDescriptorBundles.clear();
-
-    // Both mDescriptorSets and mExtinctDescriptorBundles have been cleared, so it's safe to call
-    // clear() on mDescriptorResources.
-    mDescriptorResources.clear();
-
-    mBoundDescriptor = {};
-}
-
-void VulkanPipelineCache::growDescriptorPool() noexcept {
-    // We need to destroy the old VkDescriptorPool, but we can't do so immediately because many
-    // of its descriptors are still in use. So, stash it in an "extinct" list.
-    mExtinctDescriptorPools.push_back(mDescriptorPool);
-
-    // Create the new VkDescriptorPool, twice as big as the old one.
-    mDescriptorPoolSize *= 2;
-    mDescriptorPool = createDescriptorPool(mDescriptorPoolSize);
-
-    // Clear out all unused descriptor sets in the arena so they don't get reclaimed. There is no
-    // need to free them individually since the old VkDescriptorPool will be destroyed.
-    for (auto iter = mPipelineLayouts.begin(); iter != mPipelineLayouts.end(); ++iter) {
-        for (auto& arena : iter.value().descriptorSetArenas) {
-            arena.clear();
-        }
-    }
-    mDescriptorArenasCount = 0;
-
-    // Move all in-use descriptors from the primary cache into an "extinct" list, so that they will
-    // later be destroyed rather than reclaimed.
-    using DescIterator = decltype(mDescriptorSets)::iterator;
-    for (DescIterator iter = mDescriptorSets.begin(); iter != mDescriptorSets.end(); ++iter) {
-        mExtinctDescriptorBundles.push_back(iter.value());
-    }
-    mDescriptorSets.clear();
-}
-
-size_t VulkanPipelineCache::PipelineLayoutKeyHashFn::operator()(
-        const PipelineLayoutKey& key) const {
-    std::hash<uint64_t> hasher;
-    auto h0 = hasher(key.getBitsAt(0));
-    auto h1 = hasher(key.getBitsAt(1));
-    return h0 ^ (h1 << 1);
-}
-
-bool VulkanPipelineCache::PipelineLayoutKeyEqual::operator()(const PipelineLayoutKey& k1,
-        const PipelineLayoutKey& k2) const {
-    return k1 == k2;
+   using ConstPipeIterator = decltype(mPipelines)::const_iterator;
+   for (ConstPipeIterator iter = mPipelines.begin(); iter != mPipelines.end();) {
+       const PipelineCacheEntry& cacheEntry = iter.value();
+       if (cacheEntry.lastUsed + FVK_MAX_PIPELINE_AGE < mCurrentTime) {
+           vkDestroyPipeline(mDevice, iter->second.handle, VKALLOC);
+           iter = mPipelines.erase(iter);
+       } else {
+           ++iter;
+       }
+   }
 }
 
 bool VulkanPipelineCache::PipelineEqual::operator()(const PipelineKey& k1,
@@ -895,31 +329,6 @@ bool VulkanPipelineCache::PipelineEqual::operator()(const PipelineKey& k1,
     return 0 == memcmp((const void*) &k1, (const void*) &k2, sizeof(k1));
 }
 
-bool VulkanPipelineCache::DescEqual::operator()(const DescriptorKey& k1,
-        const DescriptorKey& k2) const {
-    for (uint32_t i = 0; i < UBUFFER_BINDING_COUNT; i++) {
-        if (k1.uniformBuffers[i] != k2.uniformBuffers[i] ||
-            k1.uniformBufferOffsets[i] != k2.uniformBufferOffsets[i] ||
-            k1.uniformBufferSizes[i] != k2.uniformBufferSizes[i]) {
-            return false;
-        }
-    }
-    for (uint32_t i = 0; i < SAMPLER_BINDING_COUNT; i++) {
-        if (k1.samplers[i].sampler != k2.samplers[i].sampler ||
-            k1.samplers[i].imageView != k2.samplers[i].imageView ||
-            k1.samplers[i].imageLayout != k2.samplers[i].imageLayout) {
-            return false;
-        }
-    }
-    for (uint32_t i = 0; i < INPUT_ATTACHMENT_COUNT; i++) {
-        if (k1.inputAttachments[i].imageView != k2.inputAttachments[i].imageView ||
-            k1.inputAttachments[i].imageLayout != k2.inputAttachments[i].imageLayout) {
-            return false;
-        }
-    }
-    return true;
-}
-
 } // namespace filament::backend
 
 #pragma clang diagnostic pop

filament/backend/src/vulkan/VulkanPipelineCache.h

@@ -17,6 +17,11 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H
 #define TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H
 
+#include "VulkanCommands.h"
+#include "VulkanMemory.h"
+#include "VulkanResources.h"
+#include "VulkanUtility.h"
+
 #include <backend/DriverEnums.h>
 #include <backend/TargetBufferInfo.h>
 
@@ -34,12 +39,6 @@
 #include <vector>
 #include <unordered_map>
 
-#include "VulkanCommands.h"
-
-VK_DEFINE_HANDLE(VmaAllocator)
-VK_DEFINE_HANDLE(VmaAllocation)
-VK_DEFINE_HANDLE(VmaPool)
-
 namespace filament::backend {
 
 struct VulkanProgram;
@@ -56,32 +55,14 @@ class VulkanResourceAllocator;
 // - Assumes that viewport and scissor should be dynamic. (not baked into VkPipeline)
 // - Assumes that uniform buffers should be visible across all shader stages.
 //
-class VulkanPipelineCache : public CommandBufferObserver {
+class VulkanPipelineCache {
 public:
     VulkanPipelineCache(VulkanPipelineCache const&) = delete;
     VulkanPipelineCache& operator=(VulkanPipelineCache const&) = delete;
 
-    static constexpr uint32_t UBUFFER_BINDING_COUNT = Program::UNIFORM_BINDING_COUNT;
-    static constexpr uint32_t SAMPLER_BINDING_COUNT = MAX_SAMPLER_COUNT;
-
-    // We assume only one possible input attachment between two subpasses. See also the subpasses
-    // definition in VulkanFboCache.
-    static constexpr uint32_t INPUT_ATTACHMENT_COUNT = 1;
-
     static constexpr uint32_t SHADER_MODULE_COUNT = 2;
     static constexpr uint32_t VERTEX_ATTRIBUTE_COUNT = MAX_VERTEX_ATTRIBUTE_COUNT;
 
-    // Three descriptor set layouts: uniforms, combined image samplers, and input attachments.
-    static constexpr uint32_t DESCRIPTOR_TYPE_COUNT = 3;
-    static constexpr uint32_t INITIAL_DESCRIPTOR_SET_POOL_SIZE = 512;
-
-    // The VertexArray POD is an array of buffer targets and an array of attributes that refer to
-    // those targets. It does not include any references to actual buffers, so you can think of it
-    // as a vertex assembler configuration. For simplicity it contains fixed-size arrays and does
-    // not store sizes; all unused entries are simply zeroed out.
-    struct VertexArray {
-    };
-
     // The ProgramBundle contains weak references to the compiled vertex and fragment shaders.
     struct ProgramBundle {
         VkShaderModule vertex;
@@ -89,10 +70,6 @@ public:
         VkSpecializationInfo* specializationInfos = nullptr;
     };
 
-    using UsageFlags = utils::bitset128;
-    static UsageFlags getUsageFlags(uint16_t binding, ShaderStageFlags stages, UsageFlags src = {});
-    static UsageFlags disableUsageFlags(uint16_t binding, UsageFlags src);
-
     #pragma clang diagnostic push
     #pragma clang diagnostic warning "-Wpadded"
 
@@ -135,17 +112,13 @@ public:
 
     // Upon construction, the pipeCache initializes some internal state but does not make any Vulkan
     // calls. On destruction it will free any cached Vulkan objects that haven't already been freed.
-    VulkanPipelineCache(VulkanResourceAllocator* allocator);
+    VulkanPipelineCache(VkDevice device, VmaAllocator allocator);
     ~VulkanPipelineCache();
-    void setDevice(VkDevice device, VmaAllocator allocator);
 
-    // Creates new descriptor sets if necessary and binds them using vkCmdBindDescriptorSets.
-    // Returns false if descriptor set allocation fails.
-    bool bindDescriptors(VkCommandBuffer cmdbuffer) noexcept;
+    void bindLayout(VkPipelineLayout layout) noexcept;
 
     // Creates a new pipeline if necessary and binds it using vkCmdBindPipeline.
-    // Returns false if an error occurred.
-    bool bindPipeline(VulkanCommandBuffer* commands) noexcept;
+    void bindPipeline(VulkanCommandBuffer* commands);
 
     // Sets up a new scissor rectangle if it has been dirtied.
     void bindScissor(VkCommandBuffer cmdbuffer, VkRect2D scissor) noexcept;
@@ -155,42 +128,13 @@ public:
     void bindRasterState(const RasterState& rasterState) noexcept;
     void bindRenderPass(VkRenderPass renderPass, int subpassIndex) noexcept;
     void bindPrimitiveTopology(VkPrimitiveTopology topology) noexcept;
-    void bindUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
-            VkDeviceSize offset = 0, VkDeviceSize size = VK_WHOLE_SIZE) noexcept;
-    void bindSamplers(VkDescriptorImageInfo samplers[SAMPLER_BINDING_COUNT],
-            VulkanTexture* textures[SAMPLER_BINDING_COUNT], UsageFlags flags) noexcept;
-    void bindInputAttachment(uint32_t bindingIndex, VkDescriptorImageInfo imageInfo) noexcept;
+
     void bindVertexArray(VkVertexInputAttributeDescription const* attribDesc,
             VkVertexInputBindingDescription const* bufferDesc, uint8_t count);
 
-    // Gets the current UBO at the given slot, useful for push / pop.
-    UniformBufferBinding getUniformBufferBinding(uint32_t bindingIndex) const noexcept;
-
-    // Checks if the given uniform is bound to any slot, and if so binds "null" to that slot.
-    // Also invalidates all cached descriptors that refer to the given buffer.
-    // This is only necessary when the client knows that the UBO is about to be destroyed.
-    void unbindUniformBuffer(VkBuffer uniformBuffer) noexcept;
-
-    // Checks if an image view is bound to any sampler, and if so resets that particular slot.
-    // Also invalidates all cached descriptors that refer to the given image view.
-    // This is only necessary when the client knows that a texture is about to be destroyed.
-    void unbindImageView(VkImageView imageView) noexcept;
-
-    // NOTE: In theory we should proffer "unbindSampler" but in practice we never destroy samplers.
-
     // Destroys all managed Vulkan objects. This should be called before changing the VkDevice.
     void terminate() noexcept;
 
-    // vkCmdBindPipeline and vkCmdBindDescriptorSets establish bindings to a specific command
-    // buffer; they are not global to the device. Therefore we need to be notified when a
-    // new command buffer becomes active.
-    void onCommandBuffer(const VulkanCommandBuffer& cmdbuffer) override;
-
-    // Injects a dummy texture that can be used to clear out old descriptor sets.
-    void setDummyTexture(VkImageView imageView) {
-        mDummyTargetInfo.imageView = imageView;
-    }
-
     static VkPrimitiveTopology getPrimitiveTopology(PrimitiveType pt) noexcept {
         switch (pt) {
             case PrimitiveType::POINTS:
@@ -206,22 +150,9 @@ public:
         }
     }
 
+    void gc() noexcept;
+
 private:
-    // PIPELINE LAYOUT CACHE KEY
-    // -------------------------
-
-    using PipelineLayoutKey = utils::bitset128;
-
-    static_assert(PipelineLayoutKey::BIT_COUNT >= 2 * MAX_SAMPLER_COUNT);
-
-    struct PipelineLayoutKeyHashFn {
-        size_t operator()(const PipelineLayoutKey& key) const;
-    };
-
-    struct PipelineLayoutKeyEqual {
-        bool operator()(const PipelineLayoutKey& k1, const PipelineLayoutKey& k2) const;
-    };
-
     // PIPELINE CACHE KEY
     // ------------------
 
@@ -274,10 +205,10 @@ private:
         VertexInputBindingDescription vertexBuffers[VERTEX_ATTRIBUTE_COUNT];      //  128 : 156
         RasterState rasterState;                                                  //  16  : 284
         uint32_t padding;                                                         //  4   : 300
-        PipelineLayoutKey layout;                                                 // 16   : 304
+        VkPipelineLayout layout;                                                  //  8   : 304
     };
 
-    static_assert(sizeof(PipelineKey) == 320, "PipelineKey must not have implicit padding.");
+    static_assert(sizeof(PipelineKey) == 312, "PipelineKey must not have implicit padding.");
 
     using PipelineHashFn = utils::hash::MurmurHashFn<PipelineKey>;
 
@@ -285,52 +216,6 @@ private:
         bool operator()(const PipelineKey& k1, const PipelineKey& k2) const;
     };
 
-    // DESCRIPTOR SET CACHE KEY
-    // ------------------------
-
-    // Equivalent to VkDescriptorImageInfo but with explicit padding.
-    struct DescriptorImageInfo {
-        DescriptorImageInfo& operator=(const VkDescriptorImageInfo& that) {
-            sampler = that.sampler;
-            imageView = that.imageView;
-            imageLayout = that.imageLayout;
-            padding = 0;
-            return *this;
-        }
-        operator VkDescriptorImageInfo() const { return { sampler, imageView, imageLayout }; }
-
-        // TODO: replace the 64-bit sampler handle with `uint32_t samplerParams` and remove the
-        // padding field. This is possible if we have access to the VulkanSamplerCache.
-        VkSampler sampler;
-
-        VkImageView imageView;
-        VkImageLayout imageLayout;
-        uint32_t padding;
-    };
-
-    // We store size with 32 bits, so our "WHOLE" sentinel is different from Vk.
-    static const uint32_t WHOLE_SIZE = 0xffffffffu;
-
-    // Represents all the Vulkan state that comprises a bound descriptor set.
-    struct DescriptorKey {
-        VkBuffer uniformBuffers[UBUFFER_BINDING_COUNT];               //   80     0
-        DescriptorImageInfo samplers[SAMPLER_BINDING_COUNT];          // 1488    80
-        DescriptorImageInfo inputAttachments[INPUT_ATTACHMENT_COUNT]; //   24  1568
-        uint32_t uniformBufferOffsets[UBUFFER_BINDING_COUNT];         //   40  1592
-        uint32_t uniformBufferSizes[UBUFFER_BINDING_COUNT];           //   40  1632
-    };
-    static_assert(offsetof(DescriptorKey, samplers)              == 80);
-    static_assert(offsetof(DescriptorKey, inputAttachments)      == 1568);
-    static_assert(offsetof(DescriptorKey, uniformBufferOffsets)  == 1592);
-    static_assert(offsetof(DescriptorKey, uniformBufferSizes)    == 1632);
-    static_assert(sizeof(DescriptorKey) == 1672, "DescriptorKey must not have implicit padding.");
-
-    using DescHashFn = utils::hash::MurmurHashFn<DescriptorKey>;
-
-    struct DescEqual {
-        bool operator()(const DescriptorKey& k1, const DescriptorKey& k2) const;
-    };
-
     #pragma clang diagnostic pop
 
     // CACHE ENTRY STRUCTS
@@ -343,16 +228,6 @@ private:
     using Timestamp = uint64_t;
     Timestamp mCurrentTime = 0;
 
-    // The descriptor set cache entry is a group of descriptor sets that are bound simultaneously.
-    struct DescriptorCacheEntry {
-        std::array<VkDescriptorSet, DESCRIPTOR_TYPE_COUNT> handles;
-        Timestamp lastUsed;
-        PipelineLayoutKey pipelineLayout;
-        uint32_t id;
-    };
-    uint32_t mDescriptorCacheEntryCount = 0;
-
-
     struct PipelineCacheEntry {
         VkPipeline handle;
         Timestamp lastUsed;
@@ -361,98 +236,36 @@ private:
     struct PipelineLayoutCacheEntry {
         VkPipelineLayout handle;
         Timestamp lastUsed;
-
-        std::array<VkDescriptorSetLayout, DESCRIPTOR_TYPE_COUNT> descriptorSetLayouts;
-
-        // Each pipeline layout has 3 arenas of unused descriptors (one for each binding type).
-        //
-        // The difference between the "arenas" and the "pool" are as follows.
-        //
-        // - The "pool" is a single, centralized factory for all descriptors (VkDescriptorPool).
-        //
-        // - Each "arena" is a set of unused (but alive) descriptors that can only be used with a
-        //   specific pipeline layout and a specific binding type. We manually manage each arena.
-        //   The arenas are created in an empty state, and they are gradually populated as new
-        //   descriptors are reclaimed over time.  This is quite different from the pool, which is
-        //   given a fixed size when it is constructed.
-        //
-        std::array<std::vector<VkDescriptorSet>, DESCRIPTOR_TYPE_COUNT> descriptorSetArenas;
     };
 
     // CACHE CONTAINERS
     // ----------------
 
-    using PipelineLayoutMap = tsl::robin_map<PipelineLayoutKey , PipelineLayoutCacheEntry,
-            PipelineLayoutKeyHashFn, PipelineLayoutKeyEqual>;
     using PipelineMap = tsl::robin_map<PipelineKey, PipelineCacheEntry,
             PipelineHashFn, PipelineEqual>;
-    using DescriptorMap
-            = tsl::robin_map<DescriptorKey, DescriptorCacheEntry, DescHashFn, DescEqual>;
-    using DescriptorResourceMap
-            = std::unordered_map<uint32_t, std::unique_ptr<VulkanAcquireOnlyResourceManager>>;
 
-    PipelineLayoutMap mPipelineLayouts;
+private:
+
+    PipelineCacheEntry* getOrCreatePipeline() noexcept;
+
     PipelineMap mPipelines;
-    DescriptorMap mDescriptorSets;
-    DescriptorResourceMap mDescriptorResources;
 
     // These helpers all return unstable pointers that should not be stored.
-    DescriptorCacheEntry* createDescriptorSets() noexcept;
     PipelineCacheEntry* createPipeline() noexcept;
     PipelineLayoutCacheEntry* getOrCreatePipelineLayout() noexcept;
 
-    // Misc helper methods.
-    void destroyLayoutsAndDescriptors() noexcept;
-    VkDescriptorPool createDescriptorPool(uint32_t size) const;
-    void growDescriptorPool() noexcept;
-
     // Immutable state.
     VkDevice mDevice = VK_NULL_HANDLE;
     VmaAllocator mAllocator = VK_NULL_HANDLE;
 
     // Current requirements for the pipeline layout, pipeline, and descriptor sets.
     PipelineKey mPipelineRequirements = {};
-    DescriptorKey mDescriptorRequirements = {};
 
     // Current bindings for the pipeline and descriptor sets.
     PipelineKey mBoundPipeline = {};
-    DescriptorKey mBoundDescriptor = {};
 
     // Current state for scissoring.
     VkRect2D mCurrentScissor = {};
-
-    // The descriptor set pool starts out with a decent number of descriptor sets.  The cache can
-    // grow the pool by re-creating it with a larger size.  See growDescriptorPool().
-    VkDescriptorPool mDescriptorPool;
-
-    // This describes the number of descriptor sets in mDescriptorPool. Note that this needs to be
-    // multiplied by DESCRIPTOR_TYPE_COUNT to get the actual number of descriptor sets. Also note
-    // that the number of low-level "descriptors" (not descriptor *sets*) is actually much more than
-    // this size. It can be computed only by factoring in UBUFFER_BINDING_COUNT etc.
-    uint32_t mDescriptorPoolSize = INITIAL_DESCRIPTOR_SET_POOL_SIZE;
-
-    // To get the actual number of descriptor sets that have been allocated from the pool,
-    // take the sum of mDescriptorArenasCount (these are inactive descriptor sets) and the
-    // number of entries in the mDescriptorPool map (active descriptor sets). Multiply the result by
-    // DESCRIPTOR_TYPE_COUNT.
-    uint32_t mDescriptorArenasCount = 0;
-
-    // After a growth event (i.e. when the VkDescriptorPool is replaced with a bigger version), all
-    // currently used descriptors are moved into the "extinct" sets so that they can be safely
-    // destroyed a few frames later.
-    std::list<VkDescriptorPool> mExtinctDescriptorPools;
-    std::list<DescriptorCacheEntry> mExtinctDescriptorBundles;
-
-    VkDescriptorBufferInfo mDummyBufferInfo = {};
-    VkWriteDescriptorSet mDummyBufferWriteInfo = {};
-    VkDescriptorImageInfo mDummyTargetInfo = {};
-    VkWriteDescriptorSet mDummyTargetWriteInfo = {};
-
-    VkBuffer mDummyBuffer;
-    VmaAllocation mDummyMemory;
-
-    VulkanResourceAllocator* mResourceAllocator;
-    VulkanAcquireOnlyResourceManager mPipelineBoundResources;
 };
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanReadPixels.cpp

@@ -28,7 +28,6 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-using ImgUtil = VulkanImageUtility;
 using TaskHandler = VulkanReadPixels::TaskHandler;
 using WorkloadFunc = TaskHandler::WorkloadFunc;
 using OnCompleteFunc = TaskHandler::OnCompleteFunc;
@@ -118,7 +117,7 @@ void VulkanReadPixels::terminate() noexcept {
 VulkanReadPixels::VulkanReadPixels(VkDevice device)
     : mDevice(device) {}
 
-void VulkanReadPixels::run(VulkanRenderTarget const* srcTarget, uint32_t const x, uint32_t const y,
+void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint32_t const y,
         uint32_t const width, uint32_t const height, uint32_t const graphicsQueueFamilyIndex,
         PixelBufferDescriptor&& pbd, SelecteMemoryFunction const& selectMemoryFunc,
         OnReadCompleteFunction const& readCompleteFunc) {
@@ -218,7 +217,7 @@ void VulkanReadPixels::run(VulkanRenderTarget const* srcTarget, uint32_t const x
     };
     vkBeginCommandBuffer(cmdbuffer, &binfo);
 
-    ImgUtil::transitionLayout(cmdbuffer, {
+    imgutil::transitionLayout(cmdbuffer, {
         .image = stagingImage,
         .oldLayout = VulkanLayout::UNDEFINED,
         .newLayout = VulkanLayout::TRANSFER_DST,
@@ -232,8 +231,7 @@ void VulkanReadPixels::run(VulkanRenderTarget const* srcTarget, uint32_t const x
     });
 
     VulkanAttachment const srcAttachment = srcTarget->getColor(0);
-    const VkImageSubresourceRange srcRange
-            = srcAttachment.getSubresourceRange(VK_IMAGE_ASPECT_COLOR_BIT);
+    VkImageSubresourceRange const srcRange = srcAttachment.getSubresourceRange();
     srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
 
     VkImageCopy const imageCopyRegion = {
@@ -266,8 +264,8 @@ void VulkanReadPixels::run(VulkanRenderTarget const* srcTarget, uint32_t const x
             imageCopyRegion.srcOffset.y + imageCopyRegion.extent.height <= srcExtent.height);
 
     vkCmdCopyImage(cmdbuffer, srcAttachment.getImage(),
-            ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC), stagingImage,
-            ImgUtil::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);
+            imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC), stagingImage,
+            imgutil::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);
 
     // Restore the source image layout.
     srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::COLOR_ATTACHMENT);

filament/backend/src/vulkan/VulkanReadPixels.h

@@ -26,9 +26,7 @@
 #include <functional>
 #include <mutex>
 #include <queue>
-#include <set>
 #include <thread>
-#include <vector>
 
 namespace filament::backend {
 
@@ -74,8 +72,8 @@ public:
 
     void terminate() noexcept;
 
-    void run(VulkanRenderTarget const* srcTarget, uint32_t x, uint32_t y, uint32_t width,
-            uint32_t height, uint32_t graphicsQueueFamilyIndex, PixelBufferDescriptor&& pbd,
+    void run(VulkanRenderTarget* srcTarget, uint32_t x, uint32_t y, uint32_t width, uint32_t height,
+            uint32_t graphicsQueueFamilyIndex, PixelBufferDescriptor&& pbd,
             SelecteMemoryFunction const& selectMemoryFunc,
             OnReadCompleteFunction const& readCompleteFunc);
 

filament/backend/src/vulkan/VulkanResourceAllocator.h

@@ -49,8 +49,8 @@ namespace filament::backend {
 #endif
 
 class VulkanResourceAllocator {
-
 public:
+    using AllocatorImpl = HandleAllocatorVK;
     VulkanResourceAllocator(size_t arenaSize, bool disableUseAfterFreeCheck)
         : mHandleAllocatorImpl("Handles", arenaSize, disableUseAfterFreeCheck)
 #if DEBUG_RESOURCE_LEAKS
@@ -106,7 +106,7 @@ public:
     }
 
 private:
-    HandleAllocatorVK mHandleAllocatorImpl;
+    AllocatorImpl mHandleAllocatorImpl;
 
 #if DEBUG_RESOURCE_LEAKS
 public:

filament/backend/src/vulkan/VulkanResources.cpp

@@ -17,7 +17,7 @@
 #include "VulkanResources.h"
 #include "VulkanHandles.h"
 #include "VulkanResourceAllocator.h"
-#include "caching/VulkanDescriptorSet.h"
+#include "VulkanPipelineCache.h"
 
 namespace filament::backend {
 
@@ -62,6 +62,9 @@ void deallocateResource(VulkanResourceAllocator* allocator, VulkanResourceType t
         case VulkanResourceType::RENDER_PRIMITIVE:
             allocator->destruct<VulkanRenderPrimitive>(Handle<VulkanRenderPrimitive>(id));
             break;
+        case VulkanResourceType::DESCRIPTOR_SET_LAYOUT:
+            allocator->destruct<VulkanDescriptorSetLayout>(Handle<VulkanDescriptorSetLayout>(id));
+            break;
         case VulkanResourceType::DESCRIPTOR_SET:
             allocator->destruct<VulkanDescriptorSet>(Handle<VulkanDescriptorSet>(id));
             break;

filament/backend/src/vulkan/VulkanResources.h

@@ -35,24 +35,25 @@ struct VulkanThreadSafeResource;
 
 // Subclasses of VulkanResource must provide this enum in their construction.
 enum class VulkanResourceType : uint8_t {
-    BUFFER_OBJECT,
-    INDEX_BUFFER,
-    PROGRAM,
-    RENDER_TARGET,
-    SAMPLER_GROUP,
-    SWAP_CHAIN,
-    RENDER_PRIMITIVE,
-    TEXTURE,
-    TIMER_QUERY,
-    VERTEX_BUFFER,
-    VERTEX_BUFFER_INFO,
-    DESCRIPTOR_SET,
+    BUFFER_OBJECT = 0,
+    INDEX_BUFFER = 1,
+    PROGRAM = 2,
+    RENDER_TARGET = 3,
+    SAMPLER_GROUP = 4,
+    SWAP_CHAIN = 5,
+    RENDER_PRIMITIVE = 6,
+    TEXTURE = 7,
+    TIMER_QUERY = 8,
+    VERTEX_BUFFER = 9,
+    VERTEX_BUFFER_INFO = 10,
+    DESCRIPTOR_SET_LAYOUT = 11,
+    DESCRIPTOR_SET = 12,
 
     // Below are resources that are managed manually (i.e. not ref counted).
-    FENCE,
-    HEAP_ALLOCATED,
+    FENCE = 13,
+    HEAP_ALLOCATED = 14,
 
-    END_TYPE,  // A placeholder
+    END_TYPE = 15,  // A placeholder
 };
 
 #define IS_HEAP_ALLOC_TYPE(f)                                                                      \

filament/backend/src/vulkan/VulkanStagePool.cpp

@@ -17,6 +17,7 @@
 #include "VulkanStagePool.h"
 
 #include "VulkanConstants.h"
+#include "VulkanImageUtility.h"
 #include "VulkanMemory.h"
 #include "VulkanUtility.h"
 
@@ -118,7 +119,7 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
     // VK_IMAGE_LAYOUT_PREINITIALIZED or VK_IMAGE_LAYOUT_GENERAL layout. Calling
     // vkGetImageSubresourceLayout for a linear image returns a subresource layout mapping that is
     // valid for either of those image layouts."
-    VulkanImageUtility::transitionLayout(cmdbuffer, {
+    imgutil::transitionLayout(cmdbuffer, {
             .image = image->image,
             .oldLayout = VulkanLayout::UNDEFINED,
             .newLayout = VulkanLayout::READ_WRITE, // (= VK_IMAGE_LAYOUT_GENERAL)

filament/backend/src/vulkan/VulkanStagePool.h

@@ -17,7 +17,9 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANSTAGEPOOL_H
 #define TNT_FILAMENT_BACKEND_VULKANSTAGEPOOL_H
 
+#include "VulkanCommands.h"
 #include "VulkanContext.h"
+#include "VulkanMemory.h"
 
 #include <map>
 #include <unordered_set>

filament/backend/src/vulkan/VulkanSwapChain.h

@@ -19,6 +19,7 @@
 
 #include "DriverBase.h"
 
+#include "VulkanCommands.h"
 #include "VulkanContext.h"
 #include "VulkanResources.h"
 

filament/backend/src/vulkan/VulkanTexture.cpp

@@ -28,7 +28,6 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-using ImgUtil = VulkanImageUtility;
 VulkanTexture::VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
         VkImage image, VkFormat format, uint8_t samples, uint32_t width, uint32_t height,
         TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
@@ -37,7 +36,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanComm
       VulkanResource(
               heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
       mVkFormat(format),
-      mViewType(ImgUtil::getViewType(target)),
+      mViewType(imgutil::getViewType(target)),
       mSwizzle({}),
       mTextureImage(image),
       mFullViewRange{
@@ -62,7 +61,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
       VulkanResource(
               heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
       mVkFormat(backend::getVkFormat(tformat)),
-      mViewType(ImgUtil::getViewType(target)),
+      mViewType(imgutil::getViewType(target)),
       mSwizzle(swizzle),
       mStagePool(stagePool),
       mDevice(device),
@@ -236,7 +235,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         VulkanCommandBuffer& commands = mCommands->get();
         VkCommandBuffer const cmdbuf = commands.buffer();
         commands.acquire(this);
-        transitionLayout(cmdbuf, mFullViewRange, ImgUtil::getDefaultLayout(imageInfo.usage));
+        transitionLayout(cmdbuf, mFullViewRange, imgutil::getDefaultLayout(imageInfo.usage));
     }
 }
 
@@ -327,10 +326,10 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
 
     VulkanLayout const newLayout = VulkanLayout::TRANSFER_DST;
     VulkanLayout nextLayout = getLayout(transitionRange.baseArrayLayer, miplevel);
-    VkImageLayout const newVkLayout = ImgUtil::getVkLayout(newLayout);
+    VkImageLayout const newVkLayout = imgutil::getVkLayout(newLayout);
 
     if (nextLayout == VulkanLayout::UNDEFINED) {
-        nextLayout = ImgUtil::getDefaultLayout(this->usage);
+        nextLayout = imgutil::getDefaultLayout(this->usage);
     }
 
     transitionLayout(cmdbuf, transitionRange, newLayout);
@@ -374,8 +373,8 @@ void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, u
     VulkanLayout const oldLayout = getLayout(layer, miplevel);
     transitionLayout(cmdbuf, range, newLayout);
 
-    vkCmdBlitImage(cmdbuf, stage->image, ImgUtil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            mTextureImage, ImgUtil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
+    vkCmdBlitImage(cmdbuf, stage->image, imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
+            mTextureImage, imgutil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
 
     transitionLayout(cmdbuf, range, oldLayout);
 }
@@ -467,7 +466,7 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
         for (uint32_t i = firstLayer; i < lastLayer; ++i) {
             for (uint32_t j = firstLevel; j < lastLevel; ++j) {
                 VulkanLayout const layout = getLayout(i, j);
-                ImgUtil::transitionLayout(cmdbuf, {
+                imgutil::transitionLayout(cmdbuf, {
                         .image = mTextureImage,
                         .oldLayout = layout,
                         .newLayout = newLayout,
@@ -482,7 +481,7 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
             }
         }
     } else {
-        ImgUtil::transitionLayout(cmdbuf, {
+        imgutil::transitionLayout(cmdbuf, {
             .image = mTextureImage,
             .oldLayout = oldLayout,
             .newLayout = newLayout,

filament/backend/src/vulkan/VulkanTexture.h

@@ -23,6 +23,7 @@
 #include "VulkanResources.h"
 #include "VulkanImageUtility.h"
 
+#include <utils/Hash.h>
 #include <utils/RangeMap.h>
 
 #include <unordered_map>

filament/backend/src/vulkan/VulkanUtility.h

@@ -93,6 +93,318 @@ utils::FixedCapacityVector<OutType> enumerate(
 // Useful shorthands
 using VkFormatList = utils::FixedCapacityVector<VkFormat>;
 
+// Copied from
+//   https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkFormat.html
+constexpr VkFormat ALL_VK_FORMATS[] = {
+        VK_FORMAT_UNDEFINED,
+        VK_FORMAT_R4G4_UNORM_PACK8,
+        VK_FORMAT_R4G4B4A4_UNORM_PACK16,
+        VK_FORMAT_B4G4R4A4_UNORM_PACK16,
+        VK_FORMAT_R5G6B5_UNORM_PACK16,
+        VK_FORMAT_B5G6R5_UNORM_PACK16,
+        VK_FORMAT_R5G5B5A1_UNORM_PACK16,
+        VK_FORMAT_B5G5R5A1_UNORM_PACK16,
+        VK_FORMAT_A1R5G5B5_UNORM_PACK16,
+        VK_FORMAT_R8_UNORM,
+        VK_FORMAT_R8_SNORM,
+        VK_FORMAT_R8_USCALED,
+        VK_FORMAT_R8_SSCALED,
+        VK_FORMAT_R8_UINT,
+        VK_FORMAT_R8_SINT,
+        VK_FORMAT_R8_SRGB,
+        VK_FORMAT_R8G8_UNORM,
+        VK_FORMAT_R8G8_SNORM,
+        VK_FORMAT_R8G8_USCALED,
+        VK_FORMAT_R8G8_SSCALED,
+        VK_FORMAT_R8G8_UINT,
+        VK_FORMAT_R8G8_SINT,
+        VK_FORMAT_R8G8_SRGB,
+        VK_FORMAT_R8G8B8_UNORM,
+        VK_FORMAT_R8G8B8_SNORM,
+        VK_FORMAT_R8G8B8_USCALED,
+        VK_FORMAT_R8G8B8_SSCALED,
+        VK_FORMAT_R8G8B8_UINT,
+        VK_FORMAT_R8G8B8_SINT,
+        VK_FORMAT_R8G8B8_SRGB,
+        VK_FORMAT_B8G8R8_UNORM,
+        VK_FORMAT_B8G8R8_SNORM,
+        VK_FORMAT_B8G8R8_USCALED,
+        VK_FORMAT_B8G8R8_SSCALED,
+        VK_FORMAT_B8G8R8_UINT,
+        VK_FORMAT_B8G8R8_SINT,
+        VK_FORMAT_B8G8R8_SRGB,
+        VK_FORMAT_R8G8B8A8_UNORM,
+        VK_FORMAT_R8G8B8A8_SNORM,
+        VK_FORMAT_R8G8B8A8_USCALED,
+        VK_FORMAT_R8G8B8A8_SSCALED,
+        VK_FORMAT_R8G8B8A8_UINT,
+        VK_FORMAT_R8G8B8A8_SINT,
+        VK_FORMAT_R8G8B8A8_SRGB,
+        VK_FORMAT_B8G8R8A8_UNORM,
+        VK_FORMAT_B8G8R8A8_SNORM,
+        VK_FORMAT_B8G8R8A8_USCALED,
+        VK_FORMAT_B8G8R8A8_SSCALED,
+        VK_FORMAT_B8G8R8A8_UINT,
+        VK_FORMAT_B8G8R8A8_SINT,
+        VK_FORMAT_B8G8R8A8_SRGB,
+        VK_FORMAT_A8B8G8R8_UNORM_PACK32,
+        VK_FORMAT_A8B8G8R8_SNORM_PACK32,
+        VK_FORMAT_A8B8G8R8_USCALED_PACK32,
+        VK_FORMAT_A8B8G8R8_SSCALED_PACK32,
+        VK_FORMAT_A8B8G8R8_UINT_PACK32,
+        VK_FORMAT_A8B8G8R8_SINT_PACK32,
+        VK_FORMAT_A8B8G8R8_SRGB_PACK32,
+        VK_FORMAT_A2R10G10B10_UNORM_PACK32,
+        VK_FORMAT_A2R10G10B10_SNORM_PACK32,
+        VK_FORMAT_A2R10G10B10_USCALED_PACK32,
+        VK_FORMAT_A2R10G10B10_SSCALED_PACK32,
+        VK_FORMAT_A2R10G10B10_UINT_PACK32,
+        VK_FORMAT_A2R10G10B10_SINT_PACK32,
+        VK_FORMAT_A2B10G10R10_UNORM_PACK32,
+        VK_FORMAT_A2B10G10R10_SNORM_PACK32,
+        VK_FORMAT_A2B10G10R10_USCALED_PACK32,
+        VK_FORMAT_A2B10G10R10_SSCALED_PACK32,
+        VK_FORMAT_A2B10G10R10_UINT_PACK32,
+        VK_FORMAT_A2B10G10R10_SINT_PACK32,
+        VK_FORMAT_R16_UNORM,
+        VK_FORMAT_R16_SNORM,
+        VK_FORMAT_R16_USCALED,
+        VK_FORMAT_R16_SSCALED,
+        VK_FORMAT_R16_UINT,
+        VK_FORMAT_R16_SINT,
+        VK_FORMAT_R16_SFLOAT,
+        VK_FORMAT_R16G16_UNORM,
+        VK_FORMAT_R16G16_SNORM,
+        VK_FORMAT_R16G16_USCALED,
+        VK_FORMAT_R16G16_SSCALED,
+        VK_FORMAT_R16G16_UINT,
+        VK_FORMAT_R16G16_SINT,
+        VK_FORMAT_R16G16_SFLOAT,
+        VK_FORMAT_R16G16B16_UNORM,
+        VK_FORMAT_R16G16B16_SNORM,
+        VK_FORMAT_R16G16B16_USCALED,
+        VK_FORMAT_R16G16B16_SSCALED,
+        VK_FORMAT_R16G16B16_UINT,
+        VK_FORMAT_R16G16B16_SINT,
+        VK_FORMAT_R16G16B16_SFLOAT,
+        VK_FORMAT_R16G16B16A16_UNORM,
+        VK_FORMAT_R16G16B16A16_SNORM,
+        VK_FORMAT_R16G16B16A16_USCALED,
+        VK_FORMAT_R16G16B16A16_SSCALED,
+        VK_FORMAT_R16G16B16A16_UINT,
+        VK_FORMAT_R16G16B16A16_SINT,
+        VK_FORMAT_R16G16B16A16_SFLOAT,
+        VK_FORMAT_R32_UINT,
+        VK_FORMAT_R32_SINT,
+        VK_FORMAT_R32_SFLOAT,
+        VK_FORMAT_R32G32_UINT,
+        VK_FORMAT_R32G32_SINT,
+        VK_FORMAT_R32G32_SFLOAT,
+        VK_FORMAT_R32G32B32_UINT,
+        VK_FORMAT_R32G32B32_SINT,
+        VK_FORMAT_R32G32B32_SFLOAT,
+        VK_FORMAT_R32G32B32A32_UINT,
+        VK_FORMAT_R32G32B32A32_SINT,
+        VK_FORMAT_R32G32B32A32_SFLOAT,
+        VK_FORMAT_R64_UINT,
+        VK_FORMAT_R64_SINT,
+        VK_FORMAT_R64_SFLOAT,
+        VK_FORMAT_R64G64_UINT,
+        VK_FORMAT_R64G64_SINT,
+        VK_FORMAT_R64G64_SFLOAT,
+        VK_FORMAT_R64G64B64_UINT,
+        VK_FORMAT_R64G64B64_SINT,
+        VK_FORMAT_R64G64B64_SFLOAT,
+        VK_FORMAT_R64G64B64A64_UINT,
+        VK_FORMAT_R64G64B64A64_SINT,
+        VK_FORMAT_R64G64B64A64_SFLOAT,
+        VK_FORMAT_B10G11R11_UFLOAT_PACK32,
+        VK_FORMAT_E5B9G9R9_UFLOAT_PACK32,
+        VK_FORMAT_D16_UNORM,
+        VK_FORMAT_X8_D24_UNORM_PACK32,
+        VK_FORMAT_D32_SFLOAT,
+        VK_FORMAT_S8_UINT,
+        VK_FORMAT_D16_UNORM_S8_UINT,
+        VK_FORMAT_D24_UNORM_S8_UINT,
+        VK_FORMAT_D32_SFLOAT_S8_UINT,
+        VK_FORMAT_BC1_RGB_UNORM_BLOCK,
+        VK_FORMAT_BC1_RGB_SRGB_BLOCK,
+        VK_FORMAT_BC1_RGBA_UNORM_BLOCK,
+        VK_FORMAT_BC1_RGBA_SRGB_BLOCK,
+        VK_FORMAT_BC2_UNORM_BLOCK,
+        VK_FORMAT_BC2_SRGB_BLOCK,
+        VK_FORMAT_BC3_UNORM_BLOCK,
+        VK_FORMAT_BC3_SRGB_BLOCK,
+        VK_FORMAT_BC4_UNORM_BLOCK,
+        VK_FORMAT_BC4_SNORM_BLOCK,
+        VK_FORMAT_BC5_UNORM_BLOCK,
+        VK_FORMAT_BC5_SNORM_BLOCK,
+        VK_FORMAT_BC6H_UFLOAT_BLOCK,
+        VK_FORMAT_BC6H_SFLOAT_BLOCK,
+        VK_FORMAT_BC7_UNORM_BLOCK,
+        VK_FORMAT_BC7_SRGB_BLOCK,
+        VK_FORMAT_ETC2_R8G8B8_UNORM_BLOCK,
+        VK_FORMAT_ETC2_R8G8B8_SRGB_BLOCK,
+        VK_FORMAT_ETC2_R8G8B8A1_UNORM_BLOCK,
+        VK_FORMAT_ETC2_R8G8B8A1_SRGB_BLOCK,
+        VK_FORMAT_ETC2_R8G8B8A8_UNORM_BLOCK,
+        VK_FORMAT_ETC2_R8G8B8A8_SRGB_BLOCK,
+        VK_FORMAT_EAC_R11_UNORM_BLOCK,
+        VK_FORMAT_EAC_R11_SNORM_BLOCK,
+        VK_FORMAT_EAC_R11G11_UNORM_BLOCK,
+        VK_FORMAT_EAC_R11G11_SNORM_BLOCK,
+        VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
+        VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
+        VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
+        VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
+        VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
+        VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
+        VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
+        VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
+        VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
+        VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
+        VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
+        VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
+        VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
+        VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
+        VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
+        VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
+        VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
+        VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
+        VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
+        VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
+        VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
+        VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
+        VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
+        VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
+        VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
+        VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
+        VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
+        VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
+        VK_FORMAT_G8B8G8R8_422_UNORM,
+        VK_FORMAT_B8G8R8G8_422_UNORM,
+        VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM,
+        VK_FORMAT_G8_B8R8_2PLANE_420_UNORM,
+        VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM,
+        VK_FORMAT_G8_B8R8_2PLANE_422_UNORM,
+        VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM,
+        VK_FORMAT_R10X6_UNORM_PACK16,
+        VK_FORMAT_R10X6G10X6_UNORM_2PACK16,
+        VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16,
+        VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16,
+        VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16,
+        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16,
+        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16,
+        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16,
+        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16,
+        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16,
+        VK_FORMAT_R12X4_UNORM_PACK16,
+        VK_FORMAT_R12X4G12X4_UNORM_2PACK16,
+        VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16,
+        VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16,
+        VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16,
+        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16,
+        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16,
+        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16,
+        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16,
+        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16,
+        VK_FORMAT_G16B16G16R16_422_UNORM,
+        VK_FORMAT_B16G16R16G16_422_UNORM,
+        VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM,
+        VK_FORMAT_G16_B16R16_2PLANE_420_UNORM,
+        VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM,
+        VK_FORMAT_G16_B16R16_2PLANE_422_UNORM,
+        VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM,
+        VK_FORMAT_G8_B8R8_2PLANE_444_UNORM,
+        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16,
+        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16,
+        VK_FORMAT_G16_B16R16_2PLANE_444_UNORM,
+        VK_FORMAT_A4R4G4B4_UNORM_PACK16,
+        VK_FORMAT_A4B4G4R4_UNORM_PACK16,
+        VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_5x4_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_5x5_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_6x5_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_6x6_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_8x5_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_8x6_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_8x8_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_10x5_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_10x6_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_10x8_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_10x10_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_12x10_SFLOAT_BLOCK,
+        VK_FORMAT_ASTC_12x12_SFLOAT_BLOCK,
+        VK_FORMAT_PVRTC1_2BPP_UNORM_BLOCK_IMG,
+        VK_FORMAT_PVRTC1_4BPP_UNORM_BLOCK_IMG,
+        VK_FORMAT_PVRTC2_2BPP_UNORM_BLOCK_IMG,
+        VK_FORMAT_PVRTC2_4BPP_UNORM_BLOCK_IMG,
+        VK_FORMAT_PVRTC1_2BPP_SRGB_BLOCK_IMG,
+        VK_FORMAT_PVRTC1_4BPP_SRGB_BLOCK_IMG,
+        VK_FORMAT_PVRTC2_2BPP_SRGB_BLOCK_IMG,
+        VK_FORMAT_PVRTC2_4BPP_SRGB_BLOCK_IMG,
+// Not supported (yet) by bluevk
+//        VK_FORMAT_R16G16_SFIXED5_NV,
+//        VK_FORMAT_A1B5G5R5_UNORM_PACK16_KHR,
+//        VK_FORMAT_A8_UNORM_KHR,
+//        VK_FORMAT_A8_UNORM,
+        VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_5x4_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_5x5_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_6x5_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_6x6_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_8x5_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_8x6_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_8x8_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_10x5_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_10x6_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_10x8_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_10x10_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_12x10_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_ASTC_12x12_SFLOAT_BLOCK_EXT,
+        VK_FORMAT_G8B8G8R8_422_UNORM_KHR,
+        VK_FORMAT_B8G8R8G8_422_UNORM_KHR,
+        VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM_KHR,
+        VK_FORMAT_G8_B8R8_2PLANE_420_UNORM_KHR,
+        VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM_KHR,
+        VK_FORMAT_G8_B8R8_2PLANE_422_UNORM_KHR,
+        VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM_KHR,
+        VK_FORMAT_R10X6_UNORM_PACK16_KHR,
+        VK_FORMAT_R10X6G10X6_UNORM_2PACK16_KHR,
+        VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16_KHR,
+        VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16_KHR,
+        VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16_KHR,
+        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16_KHR,
+        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16_KHR,
+        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16_KHR,
+        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16_KHR,
+        VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16_KHR,
+        VK_FORMAT_R12X4_UNORM_PACK16_KHR,
+        VK_FORMAT_R12X4G12X4_UNORM_2PACK16_KHR,
+        VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16_KHR,
+        VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16_KHR,
+        VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16_KHR,
+        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16_KHR,
+        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16_KHR,
+        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16_KHR,
+        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16_KHR,
+        VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16_KHR,
+        VK_FORMAT_G16B16G16R16_422_UNORM_KHR,
+        VK_FORMAT_B16G16R16G16_422_UNORM_KHR,
+        VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM_KHR,
+        VK_FORMAT_G16_B16R16_2PLANE_420_UNORM_KHR,
+        VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM_KHR,
+        VK_FORMAT_G16_B16R16_2PLANE_422_UNORM_KHR,
+        VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM_KHR,
+        VK_FORMAT_G8_B8R8_2PLANE_444_UNORM_EXT,
+        VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16_EXT,
+        VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16_EXT,
+        VK_FORMAT_G16_B16R16_2PLANE_444_UNORM_EXT,
+        VK_FORMAT_A4R4G4B4_UNORM_PACK16_EXT,
+        VK_FORMAT_A4B4G4R4_UNORM_PACK16_EXT,
+        VK_FORMAT_R16G16_S10_5_NV,
+};
+
 // An Array that will be fixed capacity, but the "size" (as in user added elements) is variable.
 // Note that this class is movable.
 template<typename T, uint16_t CAPACITY>
@@ -122,14 +434,14 @@ public:
         clear();
     }
 
-    inline const_iterator begin() {
+    inline const_iterator begin() const {
         if (mInd == 0) {
             return mArray.cend();
         }
         return mArray.cbegin();
     }
 
-    inline const_iterator end() {
+    inline const_iterator end() const {
         if (mInd > 0 && mInd < CAPACITY) {
             return mArray.begin() + mInd;
         }
@@ -170,10 +482,26 @@ public:
         return mArray[ind];
     }
 
-    inline size_t size() {
+    inline T const& operator[](uint16_t ind) const {
+        return mArray[ind];
+    }
+
+    inline uint32_t size() const {
         return mInd;
     }
 
+    T* data() {
+        return mArray.data();
+    }
+
+    T const* data() const {
+        return mArray.data();
+    }
+
+    bool operator==(CappedArray const& b) const {
+        return this->mArray == b.mArray;
+    }
+
 private:
     void swap(CappedArray& rhs) {
         std::swap(mArray, rhs.mArray);
@@ -181,9 +509,129 @@ private:
     }
 
     FixedSizeArray mArray;
-    size_t mInd = 0;
+    uint32_t mInd = 0;
 };
 
+// TODO: ok to remove once Filament-side API is complete
+namespace descset {
+
+// Used to describe the descriptor binding in shader stages. We assume that the binding index does
+// not exceed 31. We also assume that we have two shader stages - vertex and fragment.  The below
+// types and struct are used across VulkanDescriptorSet and VulkanProgram.
+using UniformBufferBitmask = uint32_t;
+using SamplerBitmask = uint64_t;
+
+// We only have at most one input attachment, so this bitmask exists only to make the code more
+// general.
+using InputAttachmentBitmask = uint8_t;
+
+constexpr UniformBufferBitmask UBO_VERTEX_STAGE = 0x1;
+constexpr UniformBufferBitmask UBO_FRAGMENT_STAGE = (0x1ULL << (sizeof(UniformBufferBitmask) * 4));
+constexpr SamplerBitmask SAMPLER_VERTEX_STAGE = 0x1;
+constexpr SamplerBitmask SAMPLER_FRAGMENT_STAGE = (0x1ULL << (sizeof(SamplerBitmask) * 4));
+constexpr InputAttachmentBitmask INPUT_ATTACHMENT_VERTEX_STAGE = 0x1;
+constexpr InputAttachmentBitmask INPUT_ATTACHMENT_FRAGMENT_STAGE =
+        (0x1ULL << (sizeof(InputAttachmentBitmask) * 4));
+
+template<typename Bitmask>
+static constexpr Bitmask getVertexStage() noexcept {
+    if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
+        return UBO_VERTEX_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
+        return SAMPLER_VERTEX_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
+        return INPUT_ATTACHMENT_VERTEX_STAGE;
+    }
+}
+
+template<typename Bitmask>
+static constexpr Bitmask getFragmentStage() noexcept {
+    if constexpr (std::is_same_v<Bitmask, UniformBufferBitmask>) {
+        return UBO_FRAGMENT_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, SamplerBitmask>) {
+        return SAMPLER_FRAGMENT_STAGE;
+    }
+    if constexpr (std::is_same_v<Bitmask, InputAttachmentBitmask>) {
+        return INPUT_ATTACHMENT_FRAGMENT_STAGE;
+    }
+}
+
+typedef enum ShaderStageFlags2 : uint8_t {
+    NONE        =    0,
+    VERTEX      =    0x1,
+    FRAGMENT    =    0x2,
+} ShaderStageFlags2;
+
+enum class DescriptorType : uint8_t {
+    UNIFORM_BUFFER,
+    SAMPLER,
+    INPUT_ATTACHMENT,
+};
+
+enum class DescriptorFlags : uint8_t {
+    NONE = 0x00,
+    DYNAMIC_OFFSET = 0x01
+};
+
+struct DescriptorSetLayoutBinding {
+    DescriptorType type;
+    ShaderStageFlags2 stageFlags;
+    uint8_t binding;
+    DescriptorFlags flags;
+    uint16_t count;
+};
+
+struct DescriptorSetLayout {
+    utils::FixedCapacityVector<DescriptorSetLayoutBinding> bindings;
+};
+
+} // namespace descset
+
+namespace {
+// Use constexpr to statically generate a bit count table for 8-bit numbers.
+struct _BitCountHelper {
+    constexpr _BitCountHelper() : data{} {
+        for (uint16_t i = 0; i < 256; ++i) {
+            data[i] = 0;
+            for (auto j = i; j > 0; j /= 2) {
+                if (j & 1) {
+                    data[i]++;
+                }
+            }
+        }
+    }
+
+    template<typename MaskType>
+    constexpr uint8_t count(MaskType num) {
+        uint8_t count = 0;
+        for (uint8_t i = 0; i < sizeof(MaskType) * 8; i+=8) {
+            count += data[(num >> i) & 0xFF];
+        }
+        return count;
+    }
+
+private:
+    uint8_t data[256];
+};
+} // namespace anonymous
+
+template<typename MaskType>
+inline uint8_t countBits(MaskType num) {
+    static _BitCountHelper BitCounter = {};
+    return BitCounter.count(num);
+}
+
+// This is useful for counting the total number of descriptors for both vertex and fragment stages.
+template<typename MaskType>
+inline MaskType collapseStages(MaskType mask) {
+    constexpr uint8_t NBITS_DIV_2 = sizeof(MaskType) * 4;
+    // First zero out the top-half and then or the bottom-half against the original top-half.
+    return ((mask << NBITS_DIV_2) >> NBITS_DIV_2) | (mask >> NBITS_DIV_2);
+}
+
 } // namespace filament::backend
 
 #endif // TNT_FILAMENT_BACKEND_VULKANUTILITY_H

filament/backend/src/vulkan/caching/VulkanDescriptorSet.cpp

@@ -1,647 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "VulkanDescriptorSet.h"
-
-#include <vulkan/VulkanConstants.h>
-#include <vulkan/VulkanResources.h>
-
-#include <type_traits>
-
-namespace filament::backend {
-
-namespace {
-
-template<typename MaskType>
-using LayoutMap = tsl::robin_map<MaskType, VkDescriptorSetLayout>;
-using LayoutArray = VulkanDescriptorSetManager::LayoutArray;
-using UniformBufferBitmask = VulkanDescriptorSetManager::UniformBufferBitmask;
-using SamplerBitmask = VulkanDescriptorSetManager::SamplerBitmask;
-
-constexpr uint8_t EXPECTED_IN_FLIGHT_FRAMES = 3; // Asssume triple buffering
-
-#define PORT_CONSTANT(NameSpace, K) constexpr decltype(NameSpace::K) K = NameSpace::K;
-
-PORT_CONSTANT(VulkanDescriptorSetManager, MAX_SUPPORTED_SHADER_STAGE)
-PORT_CONSTANT(VulkanDescriptorSetManager, VERTEX_STAGE)
-PORT_CONSTANT(VulkanDescriptorSetManager, FRAGMENT_STAGE)
-
-PORT_CONSTANT(Program, UNIFORM_BINDING_COUNT)
-PORT_CONSTANT(Program, SAMPLER_BINDING_COUNT)
-
-#undef PORT_CONSTANT
-
-// Use constexpr to statically generate a bit count table for 8-bit numbers.
-struct BitCountHelper {
-    constexpr BitCountHelper() : data{} {
-        for (uint16_t i = 0; i < 256; ++i) {
-            data[i] = 0;
-            for (auto j = i; j > 0; j /= 2) {
-                if (j & 1) {
-                    data[i]++;
-                }
-            }
-        }
-    }
-    uint8_t data[256];
-} BitCountImpl;
-
-inline uint8_t countBits(uint32_t num) {
-    static constexpr uint8_t const* BIT_COUNT = BitCountImpl.data;
-    return BIT_COUNT[num & 0xFF] + BIT_COUNT[(num >> 8) & 0xFF] + BIT_COUNT[(num >> 16) & 0xFF] +
-           BIT_COUNT[(num >> 24) & 0xFF];
-}
-
-struct SamplerSet {
-    using Mask = UniformBufferBitmask;
-    struct Key {
-        Mask mask;
-        static_assert(sizeof(Mask) == 4);
-        uint32_t padding;// We need a padding to ensure 8-byte alignement.
-        VkDescriptorSetLayout layout;
-        VkSampler sampler[SAMPLER_BINDING_COUNT];
-        VkImageView imageView[SAMPLER_BINDING_COUNT];
-        VkImageLayout imageLayout[SAMPLER_BINDING_COUNT];
-    };
-
-    struct Equal {
-        bool operator()(Key const& k1, Key const& k2) const {
-            if (k1.mask != k2.mask) return false;
-            if (k1.layout != k2.layout) return false;
-
-            for (uint8_t i = 0, bitCount = countBits(k1.mask); i < bitCount; ++i) {
-                if (k1.sampler[i] != k2.sampler[i] || k1.imageView[i] != k2.imageView[i] ||
-                        k1.imageLayout[i] != k2.imageLayout[i]) {
-                    return false;
-                }
-            }
-            return true;
-        }
-    };
-    using HashFn = utils::hash::MurmurHashFn<Key>;
-    using Cache = std::unordered_map<Key, VulkanDescriptorSet*, HashFn, Equal>;
-};
-
-struct UBOSet {
-    using Mask = SamplerBitmask;
-
-    struct Key {
-        Mask mask;
-        static_assert(sizeof(Mask) == 4);
-        uint32_t padding;// We need a padding to ensure 8-byte alignement.
-        VkDescriptorSetLayout layout;
-        VkBuffer buffers[UNIFORM_BINDING_COUNT];    //   80     0
-        VkDeviceSize offsets[UNIFORM_BINDING_COUNT];//   40  1592
-        VkDeviceSize sizes[UNIFORM_BINDING_COUNT];  //   40  1632
-    };
-
-    struct Equal {
-        bool operator()(Key const& k1, Key const& k2) const {
-            if (k1.mask != k2.mask) return false;
-            if (k1.layout != k2.layout) return false;
-
-            for (uint8_t i = 0, bitCount = countBits(k1.mask); i < bitCount; ++i) {
-                if (k1.buffers[i] != k2.buffers[i] || k1.offsets[i] != k2.offsets[i] ||
-                        k1.sizes[i] != k2.sizes[i]) {
-                    return false;
-                }
-            }
-            return true;
-        }
-    };
-    using HashFn = utils::hash::MurmurHashFn<Key>;
-    using Cache = std::unordered_map<Key, VulkanDescriptorSet*, HashFn, Equal>;
-};
-
-constexpr uint8_t MAX_BINDINGS = UNIFORM_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE;
-static_assert(MAX_BINDINGS >= SAMPLER_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE);
-
-template<typename TYPE>
-class DescriptorPool {
-private:
-    static constexpr uint32_t COUNT_MULTIPLIER =
-            MAX_SUPPORTED_SHADER_STAGE * EXPECTED_IN_FLIGHT_FRAMES;
-    static constexpr uint32_t CAPACITY =
-            std::is_same_v<TYPE, UBOSet> ? Program::UNIFORM_BINDING_COUNT * COUNT_MULTIPLIER
-                                         : Program::SAMPLER_BINDING_COUNT * COUNT_MULTIPLIER;
-
-public:
-    DescriptorPool()
-        : mDevice(VK_NULL_HANDLE),
-          mResourceAllocator(nullptr) {}
-
-    DescriptorPool(VkDevice device, VulkanResourceAllocator* allocator)
-        : mDevice(device),
-          mResourceAllocator(allocator) {
-        VkDescriptorPoolSize sizes[1] = {
-            {
-                .type = std::is_same_v<TYPE, UBOSet> ? VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
-                                                     : VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                .descriptorCount = CAPACITY,
-            },
-        };
-        VkDescriptorPoolCreateInfo info{
-            .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
-            .maxSets = EXPECTED_IN_FLIGHT_FRAMES,
-            .poolSizeCount = 1,
-            .pPoolSizes = sizes,
-        };
-        vkCreateDescriptorPool(mDevice, &info, VKALLOC, &mPool);
-    }
-
-    DescriptorPool(DescriptorPool const&) = delete;
-    DescriptorPool& operator=(DescriptorPool const&) = delete;
-
-    ~DescriptorPool() {
-        vkDestroyDescriptorPool(mDevice, mPool, VKALLOC);
-    }
-
-    VulkanDescriptorSet* obtainSet(VkDescriptorSetLayout layout) {
-        if (mCount == CAPACITY) {
-            return nullptr;
-        }
-
-        if (auto unused = mUnused.find(layout);
-                unused != mUnused.end() && !unused->second.empty()) {
-            auto& unusedList = unused->second;
-            auto set = unusedList.back();
-            unusedList.pop_back();
-            mCount++;
-            return set;
-        }
-
-        // Creating a new set
-        VkDescriptorSetLayout layouts[1] = {layout};
-        VkDescriptorSetAllocateInfo allocInfo = {
-            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
-            .descriptorPool = mPool,
-            .descriptorSetCount = 1,
-            .pSetLayouts = layouts,
-        };
-        VkDescriptorSet vkSet;
-        UTILS_UNUSED VkResult result = vkAllocateDescriptorSets(mDevice, &allocInfo, &vkSet);
-        ASSERT_POSTCONDITION(result == VK_SUCCESS, "Cannot allocate descriptor set. error=%d",
-                (int) result);
-
-        mCount++;
-        return createSet(layout, vkSet);
-    }
-
-private:
-    inline VulkanDescriptorSet* createSet(VkDescriptorSetLayout layout, VkDescriptorSet vkSet) {
-        return VulkanDescriptorSet::create(mResourceAllocator, vkSet, layout,
-                [this](VulkanDescriptorSet* set) {
-                    auto const layout = set->layout;
-                    auto const vkSet = set->vkSet;
-                    auto listIter = mUnused.find(layout);
-                    assert_invariant(listIter != mUnused.end());
-                    auto& unusedList = listIter->second;
-
-                    // At this point, we know that the pointer will be stale, and we need to remove
-                    // any non-ref-counted references to it.
-                    // TODO: this will be inefficient
-                    auto iter = std::find(unusedList.begin(), unusedList.end(), set);
-                    unusedList.erase(iter);
-
-                    // We are recycling - release the vk resource back into the pool. Note that the
-                    // vk handle has not changed, but we need to change the backend handle to allow
-                    // for proper refcounting of resources referenced in this set.
-                    unusedList.push_back(createSet(layout, vkSet));
-                    mCount--;
-                });
-    }
-
-    VkDevice mDevice;
-    VulkanResourceAllocator* mResourceAllocator;
-    VkDescriptorPool mPool;
-
-    uint32_t mCount = 0;
-    std::unordered_map<VkDescriptorSetLayout, std::vector<VulkanDescriptorSet*>> mUnused;
-};
-
-template<typename TYPE>
-class DescriptorInfinitePool {
-public:
-    DescriptorInfinitePool(VkDevice device, VulkanResourceAllocator* allocator)
-        : mDevice(device),
-          mResourceAllocator(allocator) {}
-
-    ~DescriptorInfinitePool() {
-        for (auto pool: mPools) {
-            delete pool;
-        }
-    }
-
-    VulkanDescriptorSet* obtainSet(VkDescriptorSetLayout layout) {
-        for (auto pool: mPools) {
-            auto set = pool->obtainSet(layout);
-            if (set) {
-                return set;
-            }
-        }
-        // We need to increase the number of pools
-        mPools.push_back(new DescriptorPool<TYPE>{mDevice, mResourceAllocator});
-        auto pool = mPools.back();
-        return pool->obtainSet(layout);
-    }
-
-private:
-    VkDevice mDevice;
-    VulkanResourceAllocator* mResourceAllocator;
-    std::vector<DescriptorPool<TYPE>*> mPools;
-};
-
-// This holds a cache of descriptor sets of a TYPE (UBO or Sampler). The Key is defined as the
-// layout and the content (for example specific samplers).
-template<typename TYPE, typename Cache, typename Key>
-class DescriptorSetCache {
-public:
-    DescriptorSetCache(VkDevice device, VulkanResourceAllocator* allocator)
-        : mPool(device, allocator),
-          mResourceManager(allocator) {}
-
-    std::pair<VulkanDescriptorSet*, bool> obtainSet(Key const& key) {
-        mAge++;
-        if (auto result = mCache.find(key); result != mCache.end()) {
-            auto set = result->second;
-            auto const oldAge = mReverseHistory[set];
-            mHistory.erase(oldAge);
-            mHistory[mAge] = set;
-            mReverseHistory[set] = mAge;
-            return {set, true};
-        }
-
-        VulkanDescriptorSet* set = mPool.obtainSet(key.layout);
-        mCache[key] = set;
-        mReverseCache[set] = key;
-        mHistory[mAge] = set;
-        mReverseHistory[set] = mAge;
-        mResourceManager.acquire(set);
-        return {set, false};
-    }
-
-    // We need to periodically purge the descriptor sets so that we're not holding on to resources
-    // unnecessarily. The strategy for purging needs to be examined more.
-    void gc() noexcept {
-        constexpr uint32_t ALLOWED_ENTRIES = EXPECTED_IN_FLIGHT_FRAMES * 3;
-        int32_t const toCutCount = mHistory.size() - ALLOWED_ENTRIES;
-
-        if (toCutCount <= 0) {
-            return;
-        }
-
-        std::vector<uint64_t> remove;
-        for (auto entry = mHistory.begin(); entry != mHistory.end(); entry++) {
-            auto const& set = entry->second;
-            Key const& key = mReverseCache[set];
-            mCache.erase(key);
-            mReverseCache.erase(set);
-            mReverseHistory.erase(set);
-
-            remove.push_back(entry->first);
-            mResourceManager.release(set);
-        }
-        for (auto removeAge: remove) {
-            mHistory.erase(removeAge);
-        }
-    }
-
-private:
-    DescriptorInfinitePool<TYPE> mPool;
-
-    // TODO: combine some of these data structures
-    Cache mCache;
-    std::unordered_map<VulkanDescriptorSet*, Key> mReverseCache;
-
-    // Use the ordering for purging if needed;
-    std::map<uint64_t, VulkanDescriptorSet*> mHistory;
-    std::unordered_map<VulkanDescriptorSet*, uint64_t> mReverseHistory;
-    uint64_t mAge;
-
-    // Note that instead of owning the resources (i.e. descriptor set) in the pools, we keep them
-    // here since all the allocated sets have to pass through this class, and this class has better
-    // knowledge to make decisions about gc.
-    VulkanResourceManager mResourceManager;
-};
-
-using UBOSetCache = DescriptorSetCache<UBOSet, UBOSet::Cache, UBOSet::Key>;
-using SamplerSetCache = DescriptorSetCache<SamplerSet, SamplerSet::Cache, SamplerSet::Key>;
-
-template<typename MaskType>
-class LayoutCache {
-public:
-    explicit LayoutCache(VkDevice device)
-        : mDevice(device) {}
-
-    ~LayoutCache() {
-        for (auto [mask, layout]: mLayouts) {
-            vkDestroyDescriptorSetLayout(mDevice, layout, VKALLOC);
-        }
-        mLayouts.clear();
-    }
-
-    VkDescriptorSetLayout getLayout(MaskType stageMask) noexcept {
-        if (stageMask == 0) {
-            return VK_NULL_HANDLE;
-        }
-        if (auto iter = mLayouts.find(stageMask); iter != mLayouts.end()) {
-            return iter->second;
-        }
-        VkDescriptorType descriptorType;
-        if constexpr (std::is_same_v<MaskType, UBOSet::Mask>) {
-            descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
-        } else if constexpr (std::is_same_v<MaskType, SamplerSet::Mask>) {
-            descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
-        }
-
-        VkDescriptorSetLayoutBinding toBind[MAX_BINDINGS];
-        uint32_t count = 0;
-        for (uint8_t i = 0, maxBindings = sizeof(stageMask) * 4; i < maxBindings; i++) {
-            VkShaderStageFlags stages = 0;
-            if ((stageMask & (VERTEX_STAGE << i)) != 0) {
-                stages |= VK_SHADER_STAGE_VERTEX_BIT;
-            }
-            if ((stageMask & (FRAGMENT_STAGE << i)) == 0) {
-                stages |= VK_SHADER_STAGE_FRAGMENT_BIT;
-            }
-            if (stages == 0) {
-                continue;
-            }
-
-            auto& bindInfo = toBind[count++];
-            bindInfo = {
-                .binding = i,
-                .descriptorType = descriptorType,
-                .descriptorCount = 1,
-                .stageFlags = stages,
-            };
-        }
-        VkDescriptorSetLayoutCreateInfo dlinfo = {
-            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
-            .bindingCount = count,
-            .pBindings = toBind,
-        };
-
-        VkDescriptorSetLayout handle;
-        vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &handle);
-        mLayouts[stageMask] = handle;
-
-        return handle;
-    }
-
-private:
-    VkDevice mDevice;
-    LayoutMap<MaskType> mLayouts;
-};
-
-using UBOSetLayoutCache = LayoutCache<UBOSet::Mask>;
-using SamplerSetLayoutCache = LayoutCache<SamplerSet::Mask>;
-
-} // anonymous namespace
-
-VulkanDescriptorSet::VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
-        VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn)
-    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
-      resources(allocator),
-      vkSet(rawSet),
-      layout(layout),
-      mOnRecycleFn(std::move(onRecycleFn)) {}
-
-VulkanDescriptorSet::~VulkanDescriptorSet() {
-    if (mOnRecycleFn) {
-        mOnRecycleFn(this);
-    }
-}
-
-VulkanDescriptorSet* VulkanDescriptorSet::create(VulkanResourceAllocator* allocator,
-        VkDescriptorSet rawSet, VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn) {
-    auto handle = allocator->allocHandle<VulkanDescriptorSet>();
-    auto set = allocator->construct<VulkanDescriptorSet>(handle, allocator, rawSet, layout,
-            std::move(onRecycleFn));
-    return set;
-}
-
-class VulkanDescriptorSetManager::Impl {
-public:
-    Impl(VkDevice device, VulkanResourceAllocator* allocator)
-        : mDevice(device),
-          mUBOLayoutCache(device),
-          mUBOCache(device, allocator),
-          mSamplerLayoutCache(device),
-          mSamplerCache(device, allocator),
-          mResources(allocator) {}
-
-    void setUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
-            VkDeviceSize offset, VkDeviceSize size) noexcept {
-        mUbos.push_back({bindingIndex, bufferObject, offset, size});
-
-        // Between "set" and "commit", we need to ref the buffer object to avoid it being collected.
-        mResources.acquire(bufferObject);
-    }
-
-    void setSamplers(SamplerArray&& samplers) {
-        mSamplers = std::move(samplers);
-        for (auto const& sampler: mSamplers) {
-            mResources.acquire(sampler.texture);
-        }
-    }
-
-    void gc() noexcept {
-        mUBOCache.gc();
-        mSamplerCache.gc();
-    }
-
-    // This will write/update all of the descriptor set (and create a set if a one of the same
-    // layout is not available).
-    void bind(VulkanCommandBuffer* commands, GetPipelineLayoutFunction& getPipelineLayoutFn) {
-        LayoutArray layouts = {VK_NULL_HANDLE, VK_NULL_HANDLE};
-        std::array<VkDescriptorSet, 2> descSets = {VK_NULL_HANDLE, VK_NULL_HANDLE};
-        uint8_t descSetCount = 0;
-
-        VkWriteDescriptorSet descriptorWrites[UNIFORM_BINDING_COUNT + SAMPLER_BINDING_COUNT];
-        VkDescriptorBufferInfo uboWrite[UNIFORM_BINDING_COUNT];
-        VkDescriptorImageInfo samplerWrite[SAMPLER_BINDING_COUNT];
-        uint32_t nwrites = 0;
-
-        UBOSet::Key uboKey = {};
-        auto& uboMask = uboKey.mask;
-        uint8_t uboCount = 0;
-        for (auto const& ubo: mUbos) {
-            auto const& binding = std::get<0>(ubo);
-
-            uboKey.buffers[uboCount] = std::get<1>(ubo)->buffer.getGpuBuffer();
-            uboKey.offsets[uboCount] = std::get<2>(ubo);
-            uboKey.sizes[uboCount] = std::get<3>(ubo);
-            uboCount++;
-
-            // Currently we let ubo be visible in all stages.
-            uboMask |= (VERTEX_STAGE << binding);
-            uboMask |= (FRAGMENT_STAGE << binding);
-        }
-
-        if (uboMask) {
-            layouts[UBO_SET_INDEX] = uboKey.layout = mUBOLayoutCache.getLayout(uboMask);
-
-            auto [descriptorSet, cached] = mUBOCache.obtainSet(uboKey);
-            descSets[descSetCount++] = descriptorSet->vkSet;
-
-            // We need to write to the descriptor set since it wasn't cached.
-            if (!cached) {
-                // If it wasn't cached before, we need to ref the resources that it touches.
-                for (auto const& ubo: mUbos) {
-                    auto const buffer = std::get<1>(ubo);
-                    descriptorSet->resources.acquire(buffer);
-                }
-
-                for (uint8_t i = 0; i < uboCount; ++i) {
-                    auto const& binding = std::get<0>(mUbos[i]);
-                    VkWriteDescriptorSet& descriptorWrite = descriptorWrites[nwrites++];
-                    auto& writeInfo = uboWrite[i];
-                    writeInfo = {
-                        .buffer = uboKey.buffers[i],
-                        .offset = uboKey.offsets[i],
-                        .range = uboKey.sizes[i],
-                    };
-                    descriptorWrite = {
-                        .dstSet = descriptorSet->vkSet,
-                        .dstBinding = binding,
-                        .descriptorCount = 1,
-                        .descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
-                        .pBufferInfo = &writeInfo,
-                    };
-                }
-            }
-            commands->acquire(descriptorSet);
-        }
-
-        SamplerSet::Key samplerKey = {};
-        auto& samplerMask = samplerKey.mask;
-        uint8_t samplerCount = 0;
-        for (auto const sampler: mSamplers) {
-            samplerMask |= (sampler.stage << sampler.binding);
-
-            samplerKey.sampler[samplerCount] = sampler.info.sampler;
-            samplerKey.imageView[samplerCount] = sampler.info.imageView;
-            samplerKey.imageLayout[samplerCount] = sampler.info.imageLayout;
-            samplerCount++;
-        }
-
-        if (samplerMask) {
-            layouts[SAMPLER_SET_INDEX] = samplerKey.layout =
-                    mSamplerLayoutCache.getLayout(samplerMask);
-
-            auto [descriptorSet, cached] = mSamplerCache.obtainSet(samplerKey);
-            descSets[descSetCount++] = descriptorSet->vkSet;
-
-            if (!cached) {
-                for (auto const& sampler: mSamplers) {
-                    descriptorSet->resources.acquire(sampler.texture);
-                }
-
-                for (uint8_t i = 0; i < samplerCount; ++i) {
-                    auto const& binding = mSamplers[i].binding;
-                    VkWriteDescriptorSet& descriptorWrite = descriptorWrites[nwrites++];
-                    auto& writeInfo = samplerWrite[i];
-                    writeInfo = mSamplers[i].info;
-                    descriptorWrite = {
-                        .dstSet = descriptorSet->vkSet,
-                        .dstBinding = binding,
-                        .descriptorCount = 1,
-                        .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                        .pImageInfo = &writeInfo,
-                    };
-                }
-            }
-            commands->acquire(descriptorSet);
-        }
-        if (nwrites) {
-            vkUpdateDescriptorSets(mDevice, nwrites, descriptorWrites, 0, nullptr);
-        }
-
-        VkPipelineLayout const pipelineLayout = getPipelineLayoutFn(layouts);
-        VkCommandBuffer const cmdbuffer = commands->buffer();
-
-        BoundState state {
-            .cmdbuf = cmdbuffer,
-            .pipelineLayout = pipelineLayout,
-        };
-
-        if (state == mPreviousBoundState) {
-            return;
-        }
-
-        vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0,
-                descSetCount, descSets.data(), 0, nullptr);
-
-        mPreviousBoundState = state;
-
-        // Once bound, the resources are now ref'd in the descriptor set and the references in this
-        // class can be released and the descriptor set is ref'd by the command buffer.
-        mResources.clear();
-    }
-
-private:
-    struct BoundState {
-        VkCommandBuffer cmdbuf;
-        VkPipelineLayout pipelineLayout;
-
-        bool operator==(BoundState const& b) const {
-            return cmdbuf == b.cmdbuf && pipelineLayout == b.pipelineLayout;
-        }
-    };
-
-    VkDevice mDevice;
-
-    UBOSetLayoutCache mUBOLayoutCache;
-    UBOSetCache mUBOCache;
-
-    SamplerSetLayoutCache mSamplerLayoutCache;
-    SamplerSetCache mSamplerCache;
-
-    std::vector<std::tuple<uint32_t, VulkanBufferObject*, VkDeviceSize, VkDeviceSize>> mUbos;
-    SamplerArray mSamplers;
-    VulkanAcquireOnlyResourceManager mResources;
-
-    BoundState mPreviousBoundState;
-};
-
-VulkanDescriptorSetManager::VulkanDescriptorSetManager(VkDevice device,
-        VulkanResourceAllocator* resourceAllocator)
-    : mImpl(new Impl(device, resourceAllocator)) {}
-
-void VulkanDescriptorSetManager::terminate() noexcept {
-    assert_invariant(mImpl);
-    delete mImpl;
-    mImpl = nullptr;
-}
-
-void VulkanDescriptorSetManager::gc() noexcept { mImpl->gc(); }
-
-void VulkanDescriptorSetManager::bind(VulkanCommandBuffer* commands,
-        GetPipelineLayoutFunction& getPipelineLayoutFn) {
-    mImpl->bind(commands, getPipelineLayoutFn);
-}
-
-void VulkanDescriptorSetManager::setUniformBufferObject(uint32_t bindingIndex,
-        VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept {
-    mImpl->setUniformBufferObject(bindingIndex, bufferObject, offset, size);
-}
-
-void VulkanDescriptorSetManager::setSamplers(SamplerArray&& samplers) {
-    mImpl->setSamplers(std::move(samplers));
-}
-
-} // namespace filament::backend

filament/backend/src/vulkan/caching/VulkanDescriptorSet.h

@@ -1,111 +0,0 @@
-/*
- * Copyright (C) 2024 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
-#define TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
-
-#include <vulkan/VulkanResourceAllocator.h>
-#include <vulkan/VulkanTexture.h>
-#include <vulkan/VulkanUtility.h>
-
-#include <backend/DriverEnums.h>
-#include <backend/Program.h>
-#include <backend/TargetBufferInfo.h>
-
-#include <utils/bitset.h>
-
-#include <bluevk/BlueVK.h>
-#include <tsl/robin_map.h>
-
-namespace filament::backend {
-
-// We need to make this class public to enable allocation on the HandleAllocator.
-struct VulkanDescriptorSet : public VulkanResourceBase {
-public:
-    // Because we need to recycle descriptor set not used, we allow for a callback that the "Pool"
-    // can use to repackage the vk handle.
-    using OnRecycle = std::function<void(VulkanDescriptorSet*)>;
-
-    VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
-            VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn);
-
-    ~VulkanDescriptorSet();
-
-    static VulkanDescriptorSet* create(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
-            VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn);
-
-    // TODO: maybe change to fixed size for performance.
-    VulkanAcquireOnlyResourceManager resources;
-
-    VkDescriptorSet const vkSet;
-    VkDescriptorSetLayout const layout;
-
-private:
-    OnRecycle mOnRecycleFn;
-};
-
-// Abstraction over the pool and the layout cache.
-class VulkanDescriptorSetManager {
-public:
-    using StageBitMask = uint32_t;
-    using UniformBufferBitmask = StageBitMask;
-    using SamplerBitmask = StageBitMask;
-
-    static constexpr uint8_t CONCURRENT_DESCRIPTOR_SET_COUNT = 2;// UBO and samplers
-    static constexpr uint8_t MAX_SUPPORTED_SHADER_STAGE = 2;     // Vertex and fragment.
-
-    static_assert(sizeof(UniformBufferBitmask) * 8 >=
-                  (Program::UNIFORM_BINDING_COUNT) *MAX_SUPPORTED_SHADER_STAGE);
-    static_assert(sizeof(SamplerBitmask) * 8 >=
-                  Program::SAMPLER_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE);
-
-    static constexpr StageBitMask VERTEX_STAGE = 0x1;
-    static constexpr StageBitMask FRAGMENT_STAGE = (0x1 << (sizeof(StageBitMask) / 4));
-    static constexpr uint8_t UBO_SET_INDEX = 0;
-    static constexpr uint8_t SAMPLER_SET_INDEX = 1;
-
-    using LayoutArray = std::array<VkDescriptorSetLayout, CONCURRENT_DESCRIPTOR_SET_COUNT>;
-    struct SamplerBundle {
-        VkDescriptorImageInfo info = {};
-        VulkanTexture* texture = nullptr;
-        uint8_t binding = 0;
-        SamplerBitmask stage = 0;
-    };
-    using SamplerArray = CappedArray<SamplerBundle, Program::SAMPLER_BINDING_COUNT>;
-    using GetPipelineLayoutFunction = std::function<VkPipelineLayout(LayoutArray)>;
-
-    VulkanDescriptorSetManager(VkDevice device, VulkanResourceAllocator* resourceAllocator);
-
-    void terminate() noexcept;
-
-    void gc() noexcept;
-
-    // This will write/update all of the descriptor set.
-    void bind(VulkanCommandBuffer* commands, GetPipelineLayoutFunction& getPipelineFn);
-
-    void setUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
-            VkDeviceSize offset, VkDeviceSize size) noexcept;
-
-    void setSamplers(SamplerArray&& samplers);
-
-private:
-    class Impl;
-    Impl* mImpl;
-};
-
-} // namespace filament::backend
-
-#endif// TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H

filament/backend/src/vulkan/caching/VulkanDescriptorSetManager.h

@@ -0,0 +1,111 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
+#define TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
+
+#include <vulkan/VulkanResourceAllocator.h>
+#include <vulkan/VulkanTexture.h>
+#include <vulkan/VulkanUtility.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Program.h>
+#include <backend/TargetBufferInfo.h>
+
+#include <utils/bitset.h>
+
+#include <bluevk/BlueVK.h>
+#include <tsl/robin_map.h>
+
+#include <functional>
+
+namespace filament::backend {
+
+using namespace descset;
+
+// [GDSR]: Great-Descriptor-Set-Refactor: As of 03/20/24, the Filament frontend is planning to
+// introduce descriptor set. This PR will arrive before that change is complete. As such, some of
+// the methods introduced here will be obsolete, and certain logic will be generalized.
+
+// Abstraction over the pool and the layout cache.
+class VulkanDescriptorSetManager {
+public:
+    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT =
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT;
+    using GetPipelineLayoutFunction = std::function<VkPipelineLayout(
+            VulkanDescriptorSetLayoutList const&, VulkanProgram* program)>;
+
+    VulkanDescriptorSetManager(VkDevice device, VulkanResourceAllocator* resourceAllocator);
+
+    void terminate() noexcept;
+
+    void gc();
+
+    // TODO: Obsolete after [GDSR].
+    // This will write/update/bind all of the descriptor set. After [GDSR], the binding for
+    // descriptor sets will not depend on the layout described in the program.
+    VkPipelineLayout bind(VulkanCommandBuffer* commands, VulkanProgram* program,
+            GetPipelineLayoutFunction& getPipelineLayoutFn);
+
+    // TODO: Obsolete after [GDSR].
+    // This is to "dynamically" bind UBOs that might have offsets changed between pipeline binding
+    // and the draw call. We do this because UBOs for primitives that are part of the same
+    // renderable can be stored within one buffer. This can be a no-op if there were no range
+    // changes between the pipeline bind and the draw call. We will re-use applicable states
+    // provided within the bind() call, including the UBO descriptor set layout. TODO: make it a
+    // proper dynamic binding when Filament-side descriptor changes are completed.
+    void dynamicBind(VulkanCommandBuffer* commands, Handle<VulkanDescriptorSetLayout> uboLayout);
+
+    // TODO: Obsolete after [GDSR].
+    // Since we use program pointer as cache key, we need to clear the cache when it's freed.
+    void clearProgram(VulkanProgram* program) noexcept;
+
+    Handle<VulkanDescriptorSetLayout> createLayout(descset::DescriptorSetLayout const& layout);
+
+    void destroyLayout(Handle<VulkanDescriptorSetLayout> layout);
+
+    void updateBuffer(Handle<VulkanDescriptorSet> set, uint8_t binding,
+            VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept;
+
+    void updateSampler(Handle<VulkanDescriptorSet> set, uint8_t binding,
+            VulkanTexture* texture, VkSampler sampler) noexcept;
+
+    void updateInputAttachment(Handle<VulkanDescriptorSet> set, VulkanAttachment attachment) noexcept;
+
+    void clearBuffer(uint32_t bindingIndex);
+
+    void setPlaceHolders(VkSampler sampler, VulkanTexture* texture,
+            VulkanBufferObject* bufferObject) noexcept;
+
+    void clearState() noexcept;
+
+    // TODO: To be completed after [GDSR]
+    Handle<VulkanDescriptorSet> createSet(Handle<VulkanDescriptorSetLayout> layout) {
+        return Handle<VulkanDescriptorSet>();
+    }
+
+    // TODO: To be completed after [GDSR]
+    void destroySet(Handle<VulkanDescriptorSet> set) {}
+
+private:
+    class Impl;
+    Impl* mImpl;
+};
+
+}// namespace filament::backend
+
+#endif// TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
+

filament/backend/src/vulkan/caching/VulkanPipelineLayoutCache.cpp

@@ -0,0 +1,87 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "VulkanPipelineLayoutCache.h"
+
+#include <vulkan/VulkanResourceAllocator.h>
+
+namespace filament::backend {
+
+VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
+        VulkanDescriptorSetLayoutList const& descriptorSetLayouts, VulkanProgram* program) {
+    PipelineLayoutKey key = {};
+    uint8_t descSetLayoutCount = 0;
+    for (auto layoutHandle: descriptorSetLayouts) {
+        if (layoutHandle) {
+            auto layout = mAllocator->handle_cast<VulkanDescriptorSetLayout*>(layoutHandle);
+            key.descSetLayouts[descSetLayoutCount++] = layout->vklayout;
+        }
+    }
+
+    // build the push constant layout key
+    uint32_t pushConstantRangeCount = program->getPushConstantRangeCount();
+    auto const& pushCostantRanges = program->getPushConstantRanges();    
+    if (pushConstantRangeCount > 0) {
+        assert_invariant(pushConstantRangeCount <= Program::SHADER_TYPE_COUNT);
+        for (uint8_t i = 0; i < pushConstantRangeCount; ++i) {
+            auto const& range = pushCostantRanges[i];
+            auto& pushConstant = key.pushConstant[i];
+            if (range.stageFlags &  VK_SHADER_STAGE_VERTEX_BIT) {
+                pushConstant.stage = static_cast<uint8_t>(ShaderStage::VERTEX);
+            }
+            if (range.stageFlags &  VK_SHADER_STAGE_FRAGMENT_BIT) {
+                pushConstant.stage = static_cast<uint8_t>(ShaderStage::FRAGMENT);
+            }
+            if (range.stageFlags &  VK_SHADER_STAGE_COMPUTE_BIT) {
+                pushConstant.stage = static_cast<uint8_t>(ShaderStage::COMPUTE);
+            }
+            pushConstant.size = range.size;
+        }
+    }
+
+    auto iter = mPipelineLayouts.find(key);
+    if (iter != mPipelineLayouts.end()) {
+        PipelineLayoutCacheEntry& entry = mPipelineLayouts[key];
+        entry.lastUsed = mTimestamp++;
+        return entry.handle;
+    }
+
+    VkPipelineLayoutCreateInfo info{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+        .pNext = nullptr,
+        .setLayoutCount = (uint32_t) descSetLayoutCount,
+        .pSetLayouts = key.descSetLayouts.data(),
+        .pushConstantRangeCount = pushConstantRangeCount,
+        .pPushConstantRanges = pushCostantRanges,
+    };
+
+    VkPipelineLayout layout;
+    vkCreatePipelineLayout(mDevice, &info, VKALLOC, &layout);
+
+    mPipelineLayouts[key] = {
+        .handle = layout,
+        .lastUsed = mTimestamp++,
+    };
+    return layout;
+}
+
+void VulkanPipelineLayoutCache::terminate() noexcept {
+    for (auto const& [key, entry]: mPipelineLayouts) {
+        vkDestroyPipelineLayout(mDevice, entry.handle, VKALLOC);
+    }
+}
+
+}// namespace filament::backend

filament/backend/src/vulkan/caching/VulkanPipelineLayoutCache.h

@@ -0,0 +1,87 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_VULKANPIPELINELAYOUTCACHE_H
+#define TNT_FILAMENT_BACKEND_VULKANPIPELINELAYOUTCACHE_H
+
+#include <vulkan/VulkanHandles.h>
+#include <bluevk/BlueVK.h>
+
+#include <utils/Hash.h>
+
+#include <tsl/robin_map.h>
+
+namespace filament::backend {
+
+class VulkanPipelineLayoutCache {
+public:
+    VulkanPipelineLayoutCache(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mAllocator(allocator),
+          mTimestamp(0) {}
+
+    void terminate() noexcept;
+
+    struct PushConstantKey {
+        uint8_t stage;// We have one set of push constant per shader stage (fragment, vertex, etc).
+        uint8_t size;
+        // Note that there is also an offset parameter for push constants, but
+        // we always assume our update range will have the offset 0.
+    };
+
+    struct PipelineLayoutKey {
+        using DescriptorSetLayoutArray = std::array<VkDescriptorSetLayout,
+                VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+        DescriptorSetLayoutArray descSetLayouts = {};                   // 8 * 3
+        PushConstantKey pushConstant[Program::SHADER_TYPE_COUNT] = {};  // 2 * 3
+        uint16_t padding = 0;
+    };
+    static_assert(sizeof(PipelineLayoutKey) == 32);
+
+    VulkanPipelineLayoutCache(VulkanPipelineLayoutCache const&) = delete;
+    VulkanPipelineLayoutCache& operator=(VulkanPipelineLayoutCache const&) = delete;
+
+    // A pipeline layout depends on the descriptor set layout and the push constant ranges, which
+    // are described in the program.
+    VkPipelineLayout getLayout(VulkanDescriptorSetLayoutList const& descriptorSetLayouts,
+            VulkanProgram* program);
+
+private:
+    using Timestamp = uint64_t;
+    struct PipelineLayoutCacheEntry {
+        VkPipelineLayout handle;
+        Timestamp lastUsed;
+    };
+
+    using PipelineLayoutKeyHashFn = utils::hash::MurmurHashFn<PipelineLayoutKey>;
+    struct PipelineLayoutKeyEqual {
+        bool operator()(PipelineLayoutKey const& k1, PipelineLayoutKey const& k2) const {
+            return 0 == memcmp((const void*) &k1, (const void*) &k2, sizeof(PipelineLayoutKey));
+        }
+    };
+
+    using PipelineLayoutMap = tsl::robin_map<PipelineLayoutKey, PipelineLayoutCacheEntry,
+            PipelineLayoutKeyHashFn, PipelineLayoutKeyEqual>;
+
+    VkDevice mDevice;
+    VulkanResourceAllocator* mAllocator;
+    Timestamp mTimestamp;
+    PipelineLayoutMap mPipelineLayouts;
+};
+
+}
+
+#endif // TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H

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

@@ -137,18 +137,14 @@ void printDepthFormats(VkPhysicalDevice device) {
     const VkFormatFeatureFlags required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT
                                               | VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
     utils::slog.i << "Sampleable depth formats: ";
-    for (VkFormat format = (VkFormat) 1;;) {
+    for (VkFormat format : ALL_VK_FORMATS) {
         VkFormatProperties props;
         vkGetPhysicalDeviceFormatProperties(device, format, &props);
         if ((props.optimalTilingFeatures & required) == required) {
             utils::slog.i << format << " ";
         }
-        if (format == VK_FORMAT_ASTC_12x12_SRGB_BLOCK) {
-            utils::slog.i << utils::io::endl;
-            break;
-        }
-        format = (VkFormat) (1 + (int) format);
     }
+    utils::slog.i << utils::io::endl;
 }
 #endif
 
@@ -257,6 +253,7 @@ VkInstance createInstance(ExtensionSet const& requiredExts) {
     // Create the Vulkan instance.
     VkApplicationInfo appInfo = {};
     appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
+    appInfo.pEngineName = "Filament";
     appInfo.apiVersion
             = VK_MAKE_API_VERSION(0, FVK_REQUIRED_VERSION_MAJOR, FVK_REQUIRED_VERSION_MINOR, 0);
     instanceCreateInfo.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO;
@@ -343,13 +340,13 @@ std::tuple<ExtensionSet, ExtensionSet> pruneExtensions(VkPhysicalDevice device,
     ExtensionSet newInstExts = instExts;
     ExtensionSet newDeviceExts = deviceExts;
 
-#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)    
+#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
     // debugUtils and debugMarkers extensions are used mutually exclusively.
     if (newInstExts.find(VK_EXT_DEBUG_UTILS_EXTENSION_NAME) != newInstExts.end()
             && newDeviceExts.find(VK_EXT_DEBUG_MARKER_EXTENSION_NAME) != newDeviceExts.end()) {
         newDeviceExts.erase(VK_EXT_DEBUG_MARKER_EXTENSION_NAME);
     }
-#endif    
+#endif
 
 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
     // debugMarker must also request debugReport the instance extension. So check if that's present.
@@ -404,7 +401,7 @@ inline int deviceTypeOrder(VkPhysicalDeviceType deviceType) {
         case VK_PHYSICAL_DEVICE_TYPE_OTHER:
             return 1;
         default:
-            utils::slog.w << "devcieTypeOrder: Unexpected deviceType: " << deviceType
+            utils::slog.w << "deviceTypeOrder: Unexpected deviceType: " << deviceType
                           << utils::io::endl;
             return -1;
     }
@@ -499,7 +496,7 @@ VkPhysicalDevice selectPhysicalDevice(VkInstance instance,
     return device;
 }
 
-VkFormatList findAttachmentDepthFormats(VkPhysicalDevice device) {
+VkFormatList findAttachmentDepthStencilFormats(VkPhysicalDevice device) {
     VkFormatFeatureFlags const features = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
 
     // The ordering here indicates the preference of choosing depth+stencil format.
@@ -523,6 +520,24 @@ VkFormatList findAttachmentDepthFormats(VkPhysicalDevice device) {
     return ret;
 }
 
+VkFormatList findBlittableDepthStencilFormats(VkPhysicalDevice device) {
+    std::vector<VkFormat> selectedFormats;
+    VkFormatFeatureFlags const required = VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT |
+            VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
+    for (VkFormat format : ALL_VK_FORMATS) {
+        if (isVkDepthFormat(format)) {
+            VkFormatProperties props;
+            vkGetPhysicalDeviceFormatProperties(device, format, &props);
+            if ((props.optimalTilingFeatures & required) == required) {
+                selectedFormats.push_back(format);
+            }
+        }
+    }
+    VkFormatList ret(selectedFormats.size());
+    std::copy(selectedFormats.begin(), selectedFormats.end(), ret.begin());
+    return ret;
+}
+
 }// anonymous namespace
 
 using SwapChainPtr = VulkanPlatform::SwapChainPtr;
@@ -668,9 +683,11 @@ Driver* VulkanPlatform::createDriver(void* sharedContext,
             "Debug utils should not be enabled in release build.");
 #endif
 
-    context.mDepthFormats = findAttachmentDepthFormats(mImpl->mPhysicalDevice);
+    context.mDepthStencilFormats = findAttachmentDepthStencilFormats(mImpl->mPhysicalDevice);
+    context.mBlittableDepthStencilFormats =
+            findBlittableDepthStencilFormats(mImpl->mPhysicalDevice);
 
-    assert_invariant(context.mDepthFormats.size() > 0);
+    assert_invariant(context.mDepthStencilFormats.size() > 0);
 
 #if FVK_ENABLED(FVK_DEBUG_VALIDATION)
     printDepthFormats(mImpl->mPhysicalDevice);

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

@@ -243,7 +243,7 @@ VkResult VulkanPlatformSurfaceSwapChain::create() {
     mSwapChainBundle.colors = enumerate(vkGetSwapchainImagesKHR, mDevice, mSwapchain);
     mSwapChainBundle.colorFormat = surfaceFormat.format;
     mSwapChainBundle.depthFormat =
-            selectDepthFormat(mContext.getAttachmentDepthFormats(), mHasStencil);
+            selectDepthFormat(mContext.getAttachmentDepthStencilFormats(), mHasStencil);
     mSwapChainBundle.depth = createImage(mSwapChainBundle.extent, mSwapChainBundle.depthFormat);
 
     slog.i << "vkCreateSwapchain"
@@ -330,7 +330,7 @@ VulkanPlatformHeadlessSwapChain::VulkanPlatformHeadlessSwapChain(VulkanContext c
 
     bool const hasStencil = (flags & backend::SWAP_CHAIN_HAS_STENCIL_BUFFER) != 0;
     mSwapChainBundle.depthFormat =
-            selectDepthFormat(mContext.getAttachmentDepthFormats(), hasStencil);
+            selectDepthFormat(mContext.getAttachmentDepthStencilFormats(), hasStencil);
     mSwapChainBundle.depth = createImage(extent, mSwapChainBundle.depthFormat);
 }
 

filament/backend/src/vulkan/spirv/VulkanSpirvUtils.cpp

@@ -24,6 +24,7 @@
 
 #include <unordered_map>
 #include <variant>
+#include <vector>
 
 namespace filament::backend {
 
@@ -72,7 +73,7 @@ void workaroundSpecConstant(Program::ShaderBlob const& blob,
         std::vector<uint32_t>& output) {
     using WordMap = std::unordered_map<uint32_t, uint32_t>;
     using SpecValueMap = std::unordered_map<uint32_t, SpecConstantValue>;
-    constexpr size_t const HEADER_SIZE = 5;
+    constexpr size_t HEADER_SIZE = 5;
 
     WordMap varToIdMap;
     SpecValueMap idToValue;
@@ -145,4 +146,61 @@ void workaroundSpecConstant(Program::ShaderBlob const& blob,
     output.resize(outputCursor);
 }
 
+std::tuple<uint32_t,uint32_t, uint32_t> getProgramBindings(Program::ShaderBlob const& blob) {
+    std::unordered_map<uint32_t, uint32_t> targetToSet, targetToBinding;
+
+    constexpr size_t HEADER_SIZE = 5;
+
+    size_t const dataSize = blob.size() / 4;
+    uint32_t const* data = (uint32_t*) blob.data();
+
+    for (uint32_t cursor = HEADER_SIZE, cursorEnd = dataSize; cursor < cursorEnd;) {
+        uint32_t const firstWord = data[cursor];
+        uint32_t const wordCount = firstWord >> 16;
+        uint32_t const op = firstWord & 0x0000FFFF;
+
+        switch(op) {
+            case spv::Op::OpDecorate: {
+                if (data[cursor + 2] == spv::Decoration::DecorationDescriptorSet) {
+                    uint32_t const targetVar = data[cursor + 1];
+                    uint32_t const setId = data[cursor + 3];
+                    targetToSet[targetVar] = setId;
+                } else if (data[cursor + 2] == spv::Decoration::DecorationBinding) {
+                    uint32_t const targetVar = data[cursor + 1];
+                    uint32_t const binding = data[cursor + 3];
+                    targetToBinding[targetVar] = binding;
+                }
+                break;
+            }
+            default:
+                break;
+        }
+        cursor += wordCount;
+    }
+
+    constexpr uint32_t UBO = 0;
+    constexpr uint32_t SAMPLER = 1;
+    constexpr uint32_t IATTACHMENT = 2;
+    uint32_t ubo = 0, sampler = 0, inputAttachment = 0;
+
+    for (auto const& [target, setId]: targetToSet) {
+        uint32_t const binding = targetToBinding[target];
+        assert_invariant(binding < 32);
+        switch(setId) {
+            case UBO:
+                ubo |= (1 << binding);
+                break;
+            case SAMPLER:
+                sampler |= (1 << binding);
+                break;
+            case IATTACHMENT:
+                inputAttachment |= (1 << binding);
+                break;
+            default:
+                PANIC_POSTCONDITION("unexpected %d", (int) setId);
+        }
+    }
+    return {ubo, sampler, inputAttachment};
+}
+
 } // namespace filament::backend

filament/backend/src/vulkan/spirv/VulkanSpirvUtils.h

@@ -17,10 +17,11 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANSPIRVUTILS_H
 #define TNT_FILAMENT_BACKEND_VULKANSPIRVUTILS_H
 
-#include "backend/Program.h"
+#include <backend/Program.h>
 
 #include <utils/FixedCapacityVector.h>
 
+#include <tuple>
 #include <vector>
 
 namespace filament::backend {
@@ -42,6 +43,9 @@ void workaroundSpecConstant(Program::ShaderBlob const& blob,
         utils::FixedCapacityVector<Program::SpecializationConstant> const& specConstants,
         std::vector<uint32_t>& output);
 
+// bindings for UBO, samplers, input attachment
+std::tuple<uint32_t, uint32_t, uint32_t> getProgramBindings(Program::ShaderBlob const& blob);
+
 } // namespace filament::backend
 
 #endif // TNT_FILAMENT_BACKEND_VULKANSPIRVUTILS_H