Check for invalid texture inside MaterialInstance

DescriptorSetManager's commit() is the inner most of the
rendering loop. We want to avoid any hash map look up here.
We just store a pointer to the history object where appropriate.

.github/actions/android-continuous/action.yml

@@ -0,0 +1,17 @@
+name: 'Android Continuous'
+inputs:
+  build-abi:
+    description: 'The target platform ABI'
+    required: true
+    default: 'armeabi-v7a'
+runs:
+  using: "composite"
+  steps:
+    - uses: actions/setup-java@v3
+      with:
+        distribution: 'temurin'
+        java-version: '17'
+    - name: Run build script
+      run: |
+        cd build/android && printf "y" | ./build.sh continuous ${{ inputs.build-abi }}
+      shell: bash

.github/workflows/android-continuous.yml

@@ -8,32 +8,35 @@ on:
       - rc/**
 
 jobs:
-  build-android:
-    name: build-android
+  build-android-armv7:
+    name: build-android-armv7
     runs-on: macos-14
 
     steps:
       - uses: actions/checkout@v4.1.6
-      - uses: actions/setup-java@v3
+      - name: Run Android Continuous
+        uses: ./.github/actions/android-continuous
         with:
-          distribution: 'temurin'
-          java-version: '17'
-      - name: Run build script
-        run: |
-          cd build/android && printf "y" | ./build.sh continuous
-      - uses: actions/upload-artifact@v1.0.0
+          build-abi: armeabi-v7a
+
+  build-android-armv8a:
+    name: build-android-armv8a
+    runs-on: macos-14
+
+    steps:
+      - uses: actions/checkout@v4.1.6
+      - name: Run Android Continuous
+        uses: ./.github/actions/android-continuous
         with:
-          name: filament-android
-          path: out/filament-android-release.aar
-      - uses: actions/upload-artifact@v1.0.0
+          build-abi: arm64-v8a
+
+  build-android-x86_64:
+    name: build-android-x86_64
+    runs-on: macos-14
+
+    steps:
+      - uses: actions/checkout@v4.1.6
+      - name: Run Android Continuous
+        uses: ./.github/actions/android-continuous
         with:
-          name: filamat-android-full
-          path: out/filamat-android-release.aar
-      - uses: actions/upload-artifact@v1.0.0
-        with:
-          name: gltfio-android-release
-          path: out/gltfio-android-release.aar
-      - uses: actions/upload-artifact@v1.0.0
-        with:
-          name: filament-utils-android-release
-          path: out/filament-utils-android-release.aar
+          build-abi: x86_64

.github/workflows/presubmit.yml

@@ -49,8 +49,10 @@ jobs:
           distribution: 'temurin'
           java-version: '17'
       - name: Run build script
+        # Only build 1 64 bit target during presubmit to cut down build times during presubmit
+        # Continuous builds will build everything
         run: |
-          cd build/android && printf "y" | ./build.sh presubmit
+          cd build/android && printf "y" | ./build.sh presubmit arm64-v8a
 
   build-ios:
     name: build-iOS

.github/workflows/release.yml

@@ -120,7 +120,7 @@ jobs:
         env:
           TAG: ${{ steps.git_ref.outputs.tag }}
         run: |
-          cd build/android && printf "y" | ./build.sh release
+          cd build/android && printf "y" | ./build.sh release armeabi-v7a,arm64-v8a,x86,x86_64
           cd ../..
           mv out/filament-android-release.aar out/filament-${TAG}-android.aar
           mv out/filamat-android-release.aar out/filamat-${TAG}-android.aar

NEW_RELEASE_NOTES.md

@@ -7,3 +7,5 @@ for next branch cut* header.
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).
 
 ## Release notes for next branch cut
+
+- Add support for multi-layered render target with array textures.

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.53.1'
+    implementation 'com.google.android.filament:filament-android:1.54.2'
 }
 ```
 
@@ -51,19 +51,9 @@ 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', '~> 1.53.1'
+pod 'Filament', '~> 1.54.2'
 ```
 
-### Snapshots
-
-If you prefer to live on the edge, you can download a continuous build by following the following
-steps:
-
-1. Find the [commit](https://github.com/google/filament/commits/main) you're interested in.
-2. Click the green check mark under the commit message.
-3. Click on the _Details_ link for the platform you're interested in.
-4. On the top left click _Summary_, then in the _Artifacts_ section choose the desired artifact.
-
 ## Documentation
 
 - [Filament](https://google.github.io/filament/Filament.html), an in-depth explanation of

RELEASE_NOTES.md

@@ -7,6 +7,33 @@ 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).
 
+## v1.54.3
+
+
+## v1.54.2
+
+- Add a `name` API to Filament objects for debugging handle use-after-free assertions
+
+## v1.54.1
+
+
+## v1.54.0
+
+- materials: add a new `stereoscopicType` material parameter. [⚠️ **New Material Version**]
+- Fix a crash when compiling shaders on IMG devices
+
+## v1.53.5
+
+- engine: Fix bug causing certain sampler parameters to not be applied correctly in GLES 2.0 and on
+  certain GLES 3.0 drivers.
+
+## v1.53.4
+
+
+## v1.53.3
+
+- Add drag and drop support for IBL files for desktop gltf_viewer.
+
 ## v1.53.2
 
 

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

@@ -253,16 +253,10 @@ Java_com_google_android_filament_RenderableManager_nBuilderMorphingStandard(JNIE
 }
 
 extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nBuilderSetMorphTargetBufferAt(JNIEnv*, jclass,
-        jlong nativeBuilder, int level, int primitiveIndex, jlong nativeMorphTargetBuffer,
-        int offset, int count) {
+Java_com_google_android_filament_RenderableManager_nBuilderSetMorphTargetBufferOffsetAt(JNIEnv*, jclass,
+        jlong nativeBuilder, int level, int primitiveIndex, int offset) {
     RenderableManager::Builder *builder = (RenderableManager::Builder *) nativeBuilder;
-    if (nativeMorphTargetBuffer) {
-        MorphTargetBuffer *morphTargetBuffer = (MorphTargetBuffer *) nativeMorphTargetBuffer;
-        builder->morphing(level, primitiveIndex, morphTargetBuffer, offset, count);
-    } else {
-        builder->morphing(level, primitiveIndex, offset, count);
-    }
+    builder->morphing(level, primitiveIndex, offset);
 }
 
 extern "C" JNIEXPORT void JNICALL
@@ -334,18 +328,12 @@ Java_com_google_android_filament_RenderableManager_nSetMorphWeights(JNIEnv* env,
 }
 
 extern "C" JNIEXPORT void JNICALL
-Java_com_google_android_filament_RenderableManager_nSetMorphTargetBufferAt(JNIEnv*,
+Java_com_google_android_filament_RenderableManager_nSetMorphTargetBufferOffsetAt(JNIEnv*,
         jclass, jlong nativeRenderableManager, jint i, int level, jint primitiveIndex,
-        jlong nativeMorphTargetBuffer, jint offset, jint count) {
+        jlong, jint offset) {
     RenderableManager *rm = (RenderableManager *) nativeRenderableManager;
-    if (nativeMorphTargetBuffer) {
-        MorphTargetBuffer *morphTargetBuffer = (MorphTargetBuffer *) nativeMorphTargetBuffer;
-        rm->setMorphTargetBufferAt((RenderableManager::Instance) i, (uint8_t) level,
-                (size_t) primitiveIndex, morphTargetBuffer, (size_t) offset, (size_t) count);
-    } else {
-        rm->setMorphTargetBufferAt((RenderableManager::Instance) i, (uint8_t) level,
-                (size_t) primitiveIndex, (size_t) offset, (size_t) count);
-    }
+    rm->setMorphTargetBufferOffsetAt((RenderableManager::Instance) i, (uint8_t) level,
+            (size_t) primitiveIndex, (size_t) offset);
 }
 
 extern "C" JNIEXPORT jint JNICALL

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

@@ -28,6 +28,14 @@
 using namespace filament;
 using namespace backend;
 
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Renderer_nSkipFrame(JNIEnv *, jclass, jlong nativeRenderer,
+        jlong vsyncSteadyClockTimeNano) {
+    Renderer *renderer = (Renderer *) nativeRenderer;
+    renderer->skipFrame(uint64_t(vsyncSteadyClockTimeNano));
+}
+
 extern "C" JNIEXPORT jboolean JNICALL
 Java_com_google_android_filament_Renderer_nBeginFrame(JNIEnv *, jclass, jlong nativeRenderer,
         jlong nativeSwapChain, jlong frameTimeNanos) {

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

@@ -531,3 +531,12 @@ Java_com_google_android_filament_View_nGetFogEntity(JNIEnv *env, jclass clazz,
     View *view = (View *) nativeView;
     return (jint)view->getFogEntity().getId();
 }
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_View_nClearFrameHistory(JNIEnv *env, jclass clazz,
+        jlong nativeView, jlong nativeEngine) {
+    View *view = (View *) nativeView;
+    Engine *engine = (Engine *) nativeEngine;
+    view->clearFrameHistory(*engine);
+}

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

@@ -425,9 +425,12 @@ public class Engine {
         public long resourceAllocatorCacheSizeMB = 64;
 
         /*
-         * This value determines for how many frames are texture entries kept in the cache.
+         * This value determines how many frames texture entries are kept for in the cache. This
+         * is a soft limit, meaning some texture older than this are allowed to stay in the cache.
+         * Typically only one texture is evicted per frame.
+         * The default is 1.
          */
-        public long resourceAllocatorCacheMaxAge = 2;
+        public long resourceAllocatorCacheMaxAge = 1;
 
         /*
          * Disable backend handles use-after-free checks.

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

@@ -74,7 +74,7 @@ public class MorphTargetBuffer {
          *
          * @exception IllegalStateException if the MorphTargetBuffer could not be created
          *
-         * @see #setMorphTargetBufferAt
+         * @see #setMorphTargetBufferOffsetAt
          */
         @NonNull
         public MorphTargetBuffer build(@NonNull Engine engine) {

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

@@ -569,40 +569,13 @@ public class RenderableManager {
          *
          * @param level the level of detail (lod), only 0 can be specified
          * @param primitiveIndex zero-based index of the primitive, must be less than the count passed to Builder constructor
-         * @param morphTargetBuffer specifies the morph target buffer
          * @param offset specifies where in the morph target buffer to start reading (expressed as a number of vertices)
-         * @param count number of vertices in the morph target buffer to read, must equal the geometry's count (for triangles, this should be a multiple of 3)
          */
         @NonNull
         public Builder morphing(@IntRange(from = 0) int level,
                                 @IntRange(from = 0) int primitiveIndex,
-                                @IntRange(from = 0) int offset,
-                                @IntRange(from = 0) int count) {
-            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex, 0, offset, count);
-            return this;
-        }
-
-        /** @deprecated */
-        @Deprecated
-        @NonNull
-        public Builder morphing(@IntRange(from = 0) int level,
-                                @IntRange(from = 0) int primitiveIndex,
-                                @NonNull MorphTargetBuffer morphTargetBuffer,
-                                @IntRange(from = 0) int offset,
-                                @IntRange(from = 0) int count) {
-            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex,
-                    morphTargetBuffer.getNativeObject(), offset, count);
-            return this;
-        }
-
-        /** @deprecated */
-        @Deprecated
-        @NonNull
-        public Builder morphing(@IntRange(from = 0) int level,
-                                @IntRange(from = 0) int primitiveIndex,
-                                @NonNull MorphTargetBuffer morphTargetBuffer) {
-            nBuilderSetMorphTargetBufferAt(mNativeBuilder, level, primitiveIndex,
-                    morphTargetBuffer.getNativeObject(), 0, morphTargetBuffer.getVertexCount());
+                                @IntRange(from = 0) int offset) {
+            nBuilderSetMorphTargetBufferOffsetAt(mNativeBuilder, level, primitiveIndex, offset);
             return this;
         }
 
@@ -705,34 +678,11 @@ public class RenderableManager {
      *
      * @see Builder#morphing
      */
-    public void setMorphTargetBufferAt(@EntityInstance int i,
+    public void setMorphTargetBufferOffsetAt(@EntityInstance int i,
                                        @IntRange(from = 0) int level,
                                        @IntRange(from = 0) int primitiveIndex,
-                                       @IntRange(from = 0) int offset,
-                                       @IntRange(from = 0) int count) {
-        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex, 0, offset, count);
-    }
-
-    /** @deprecated */
-    @Deprecated
-    public void setMorphTargetBufferAt(@EntityInstance int i,
-                                       @IntRange(from = 0) int level,
-                                       @IntRange(from = 0) int primitiveIndex,
-                                       @NonNull MorphTargetBuffer morphTargetBuffer,
-                                       @IntRange(from = 0) int offset,
-                                       @IntRange(from = 0) int count) {
-        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex,
-                morphTargetBuffer.getNativeObject(), offset, count);
-    }
-
-    /** @deprecated */
-    @Deprecated
-    public void setMorphTargetBufferAt(@EntityInstance int i,
-                                       @IntRange(from = 0) int level,
-                                       @IntRange(from = 0) int primitiveIndex,
-                                       @NonNull MorphTargetBuffer morphTargetBuffer) {
-        nSetMorphTargetBufferAt(mNativeObject, i, level, primitiveIndex,
-                morphTargetBuffer.getNativeObject(), 0, morphTargetBuffer.getVertexCount());
+                                       @IntRange(from = 0) int offset) {
+        nSetMorphTargetBufferOffsetAt(mNativeObject, i, level, primitiveIndex, 0, offset);
     }
 
     /**
@@ -1033,7 +983,7 @@ public class RenderableManager {
     private static native void nBuilderSkinningBuffer(long nativeBuilder, long nativeSkinningBuffer, int boneCount, int offset);
     private static native void nBuilderMorphing(long nativeBuilder, int targetCount);
     private static native void nBuilderMorphingStandard(long nativeBuilder, long nativeMorphTargetBuffer);
-    private static native void nBuilderSetMorphTargetBufferAt(long nativeBuilder, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset, int count);
+    private static native void nBuilderSetMorphTargetBufferOffsetAt(long nativeBuilder, int level, int primitiveIndex, int offset);
     private static native void nBuilderEnableSkinningBuffers(long nativeBuilder, boolean enabled);
     private static native void nBuilderFog(long nativeBuilder, boolean enabled);
     private static native void nBuilderLightChannel(long nativeRenderableManager, int channel, boolean enable);
@@ -1043,7 +993,7 @@ public class RenderableManager {
     private static native int nSetBonesAsMatrices(long nativeObject, int i, Buffer matrices, int remaining, int boneCount, int offset);
     private static native int nSetBonesAsQuaternions(long nativeObject, int i, Buffer quaternions, int remaining, int boneCount, int offset);
     private static native void nSetMorphWeights(long nativeObject, int instance, float[] weights, int offset);
-    private static native void nSetMorphTargetBufferAt(long nativeObject, int i, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset, int count);
+    private static native void nSetMorphTargetBufferOffsetAt(long nativeObject, int i, int level, int primitiveIndex, long nativeMorphTargetBuffer, int offset);
     private static native int nGetMorphTargetCount(long nativeObject, int i);
     private static native void nSetAxisAlignedBoundingBox(long nativeRenderableManager, int i, float cx, float cy, float cz, float ex, float ey, float ez);
     private static native void nSetLayerMask(long nativeRenderableManager, int i, int select, int value);

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

@@ -297,6 +297,20 @@ public class Renderer {
         nSetVsyncTime(getNativeObject(), steadyClockTimeNano);
     }
 
+    /**
+     * Call skipFrame when momentarily skipping frames, for instance if the content of the
+     * scene doesn't change.
+     *
+     * @param vsyncSteadyClockTimeNano The time in nanoseconds when the frame started being rendered,
+     *                       in the {@link System#nanoTime()} timebase. Divide this value by 1000000 to
+     *                       convert it to the {@link android.os.SystemClock#uptimeMillis()}
+     *                       time base. This typically comes from
+     *                       {@link android.view.Choreographer.FrameCallback}.
+     */
+    public void skipFrame(long vsyncSteadyClockTimeNano) {
+        nSkipFrame(getNativeObject(), vsyncSteadyClockTimeNano);
+    }
+
     /**
      * Sets up a frame for this <code>Renderer</code>.
      * <p><code>beginFrame</code> manages frame pacing, and returns whether or not a frame should be
@@ -716,6 +730,7 @@ public class Renderer {
 
     private static native void nSetPresentationTime(long nativeObject, long monotonicClockNanos);
     private static native void nSetVsyncTime(long nativeObject, long steadyClockTimeNano);
+    private static native void nSkipFrame(long nativeObject, long vsyncSteadyClockTimeNano);
     private static native boolean nBeginFrame(long nativeRenderer, long nativeSwapChain, long frameTimeNanos);
     private static native void nEndFrame(long nativeRenderer);
     private static native void nRender(long nativeRenderer, long nativeView);

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

@@ -1233,6 +1233,18 @@ public class View {
         return nGetFogEntity(getNativeObject());
     }
 
+    /**
+     * When certain temporal features are used (e.g.: TAA or Screen-space reflections), the view
+     * keeps a history of previous frame renders associated with the Renderer the view was last
+     * used with. When switching Renderer, it may be necessary to clear that history by calling
+     * this method. Similarly, if the whole content of the screen change, like when a cut-scene
+     * starts, clearing the history might be needed to avoid artifacts due to the previous frame
+     * being very different.
+     */
+    public void clearFrameHistory(Engine engine) {
+        nClearFrameHistory(getNativeObject(), engine.getNativeObject());
+    }
+
     public long getNativeObject() {
         if (mNativeObject == 0) {
             throw new IllegalStateException("Calling method on destroyed View");
@@ -1294,7 +1306,7 @@ public class View {
     private static native void nSetMaterialGlobal(long nativeView, int index, float x, float y, float z, float w);
     private static native void nGetMaterialGlobal(long nativeView, int index, float[] out);
     private static native int nGetFogEntity(long nativeView);
-
+    private static native void nClearFrameHistory(long nativeView, long nativeEngine);
 
     /**
      * List of available ambient occlusion techniques.

android/filament-utils-android/src/main/cpp/Manipulator.cpp

@@ -125,6 +125,11 @@ extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBu
     builder->groundPlane(a, b, c, d);
 }
 
+extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBuilderPanning(JNIEnv*, jclass, jlong nativeBuilder, jboolean enabled) {
+    Builder* builder = (Builder*) nativeBuilder;
+    builder->panning(enabled);
+}
+
 extern "C" JNIEXPORT long Java_com_google_android_filament_utils_Manipulator_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder, jint mode) {
     Builder* builder = (Builder*) nativeBuilder;
     return (jlong) builder->build((Mode) mode);

android/filament-utils-android/src/main/java/com/google/android/filament/utils/Manipulator.java

@@ -274,6 +274,17 @@ public class Manipulator {
             return this;
         }
 
+        /**
+         * Sets whether panning is enabled in the manipulator.
+         *
+         * @return this <code>Builder</code> object for chaining calls
+         */
+        @NonNull
+        public Builder panning(Boolean enabled) {
+            nBuilderPanning(mNativeBuilder, enabled);
+            return this;
+        }
+
         /**
          * Creates and returns the <code>Manipulator</code> object.
          *
@@ -483,6 +494,7 @@ public class Manipulator {
     private static native void nBuilderFlightPanSpeed(long nativeBuilder, float x, float y);
     private static native void nBuilderFlightMoveDamping(long nativeBuilder, float damping);
     private static native void nBuilderGroundPlane(long nativeBuilder, float a, float b, float c, float d);
+    private static native void nBuilderPanning(long nativeBuilder, Boolean enabled);
     private static native long nBuilderBuild(long nativeBuilder, int mode);
 
     private static native void nDestroyManipulator(long nativeManip);

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.53.1
+VERSION_NAME=1.54.2
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

build/android/build.sh

@@ -60,13 +60,7 @@ if [[ ! -d "${ANDROID_HOME}/ndk/$FILAMENT_NDK_VERSION" ]]; then
     yes | ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager --licenses
     ${ANDROID_HOME}/cmdline-tools/latest/bin/sdkmanager "ndk;$FILAMENT_NDK_VERSION"
 fi
-
-# Only build 1 64 bit target during presubmit to cut down build times during presubmit
-# Continuous builds will build everything
 ANDROID_ABIS=
-if [[ "$TARGET" == "presubmit" ]]; then
-  ANDROID_ABIS="-q arm64-v8a"
-fi
 
 # Build the Android sample-gltf-viewer APK during release.
 BUILD_SAMPLES=
@@ -74,5 +68,19 @@ if [[ "$TARGET" == "release" ]]; then
     BUILD_SAMPLES="-k sample-gltf-viewer"
 fi
 
+function build_android() {
+    local ABI=$1
+
+    # Do the following in two steps so that we do not run out of space
+    if [[ -n "${BUILD_DEBUG}" ]]; then
+        FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android -q ${ABI} -c ${BUILD_SAMPLES} ${GENERATE_ARCHIVES} ${BUILD_DEBUG}
+        rm -rf out/cmake-android-debug-*
+    fi
+    if [[ -n "${BUILD_RELEASE}" ]]; then
+        FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android -q ${ABI} -c ${BUILD_SAMPLES} ${GENERATE_ARCHIVES} ${BUILD_RELEASE}
+        rm -rf out/cmake-android-release-*
+    fi
+}
+
 pushd `dirname $0`/../.. > /dev/null
-FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION} ./build.sh -p android $ANDROID_ABIS -c $BUILD_SAMPLES $GENERATE_ARCHIVES $BUILD_DEBUG $BUILD_RELEASE
+build_android $2

docs/Materials.md.html

@@ -1308,7 +1308,12 @@ Description
       declare a variable called `eyeDirection` you can access it in the fragment shader using
       `variable_eyeDirection`. In the vertex shader, the interpolant name is simply a member of
       the `MaterialVertexInputs` structure (`material.eyeDirection` in your example). Each
-      interpolant is of type `float4` (`vec4`) in the shaders.
+      interpolant is of type `float4` (`vec4`) in the shaders. By default the precision of the
+      interpolant is `highp` in *both* the vertex and fragment shaders.
+      An alternate syntax can be used to specify both the name and precision of the interpolant.
+      In this case the specified precision is used as-is in both fragment and vertex stages, in
+      particular if `default` is specified the default precision is used is the fragment shader
+      (`mediump`) and in the vertex shader (`highp`).
 
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ JSON
 material {
@@ -1320,7 +1325,11 @@ material {
         }
     ],
     variables : [
-         eyeDirection
+         eyeDirection,
+         {
+            name : eyeColor,
+            precision : medium
+         }
     ],
     vertexDomain : device,
     depthWrite : false,

filament/CMakeLists.txt

@@ -61,6 +61,7 @@ set(SRCS
         src/Engine.cpp
         src/Exposure.cpp
         src/Fence.cpp
+        src/FilamentBuilder.cpp
         src/FrameInfo.cpp
         src/FrameSkipper.cpp
         src/Froxelizer.cpp
@@ -75,8 +76,6 @@ set(SRCS
         src/MaterialInstance.cpp
         src/MaterialParser.cpp
         src/MorphTargetBuffer.cpp
-        src/PerViewUniforms.cpp
-        src/PerShadowMapUniforms.cpp
         src/PostProcessManager.cpp
         src/RenderPass.cpp
         src/RenderPrimitive.cpp
@@ -125,6 +124,12 @@ set(SRCS
         src/details/Texture.cpp
         src/details/VertexBuffer.cpp
         src/details/View.cpp
+        src/ds/ColorPassDescriptorSet.cpp
+        src/ds/DescriptorSet.cpp
+        src/ds/DescriptorSetLayout.cpp
+        src/ds/PostProcessDescriptorSet.cpp
+        src/ds/ShadowMapDescriptorSet.cpp
+        src/ds/SsrPassDescriptorSet.cpp
         src/fg/Blackboard.cpp
         src/fg/DependencyGraph.cpp
         src/fg/FrameGraph.cpp
@@ -152,19 +157,16 @@ set(PRIVATE_HDRS
         src/HwVertexBufferInfoFactory.h
         src/Intersections.h
         src/MaterialParser.h
-        src/PerViewUniforms.h
-        src/PerShadowMapUniforms.h
         src/PIDController.h
         src/PostProcessManager.h
-        src/RendererUtils.h
         src/RenderPass.h
         src/RenderPrimitive.h
+        src/RendererUtils.h
         src/ResourceAllocator.h
         src/ResourceList.h
         src/ShadowMap.h
         src/ShadowMapManager.h
         src/SharedHandle.h
-        src/TypedUniformBuffer.h
         src/UniformBuffer.h
         src/components/CameraManager.h
         src/components/LightManager.h
@@ -192,6 +194,14 @@ set(PRIVATE_HDRS
         src/details/Texture.h
         src/details/VertexBuffer.h
         src/details/View.h
+        src/downcast.h
+        src/ds/ColorPassDescriptorSet.h
+        src/ds/DescriptorSetLayout.h
+        src/ds/PostProcessDescriptorSet.h
+        src/ds/ShadowMapDescriptorSet.h
+        src/ds/SsrPassDescriptorSet.h
+        src/ds/TypedBuffer.h
+        src/ds/TypedUniformBuffer.h
         src/fg/Blackboard.h
         src/fg/FrameGraph.h
         src/fg/FrameGraphId.h
@@ -209,7 +219,6 @@ set(PRIVATE_HDRS
         src/materials/fsr/ffx_a.h
         src/materials/fsr/ffx_fsr1.h
         src/materials/fsr/ffx_fsr1_mobile.fs
-        src/downcast.h
 )
 
 set(MATERIAL_SRCS

filament/backend/CMakeLists.txt

@@ -12,6 +12,7 @@ set(PUBLIC_HDRS
         include/backend/AcquiredImage.h
         include/backend/BufferDescriptor.h
         include/backend/CallbackHandler.h
+        include/backend/DescriptorSetOffsetArray.h
         include/backend/DriverApiForward.h
         include/backend/DriverEnums.h
         include/backend/Handle.h
@@ -69,9 +70,13 @@ set(PRIVATE_HDRS
 if (FILAMENT_SUPPORTS_OPENGL AND NOT FILAMENT_USE_EXTERNAL_GLES3)
     list(APPEND SRCS
             include/backend/platforms/OpenGLPlatform.h
+            src/opengl/BindingMap.h
             src/opengl/gl_headers.cpp
             src/opengl/gl_headers.h
             src/opengl/GLBufferObject.h
+            src/opengl/GLDescriptorSet.cpp
+            src/opengl/GLDescriptorSet.h
+            src/opengl/GLDescriptorSetLayout.h
             src/opengl/GLTexture.h
             src/opengl/GLUtils.cpp
             src/opengl/GLUtils.h
@@ -495,21 +500,38 @@ endif()
 
 # ==================================================================================================
 # Compute tests
+#
+#if (NOT IOS AND NOT WEBGL)
+#
+#add_executable(compute_test
+#        test/ComputeTest.cpp
+#        test/Arguments.cpp
+#        test/test_ComputeBasic.cpp
+#        )
+#
+#target_link_libraries(compute_test PRIVATE
+#        backend
+#        getopt
+#        gtest
+#        )
+#
+#set_target_properties(compute_test PROPERTIES FOLDER Tests)
+#
+#endif()
 
-if (NOT IOS AND NOT WEBGL)
+# ==================================================================================================
+# Metal utils tests
 
-add_executable(compute_test
-        test/ComputeTest.cpp
-        test/Arguments.cpp
-        test/test_ComputeBasic.cpp
-        )
+if (APPLE AND NOT IOS)
 
-target_link_libraries(compute_test PRIVATE
+add_executable(metal_utils_test test/MetalTest.mm)
+
+target_link_libraries(metal_utils_test PRIVATE
         backend
         getopt
         gtest
         )
 
-set_target_properties(compute_test PROPERTIES FOLDER Tests)
+set_target_properties(metal_utils_test PROPERTIES FOLDER Tests)
 
 endif()

filament/backend/include/backend/DescriptorSetOffsetArray.h

@@ -0,0 +1,101 @@
+/*
+ * 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_COMMANDSTREAMVECTOR_H
+#define TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H
+
+#include <backend/DriverApiForward.h>
+
+#include <initializer_list>
+#include <memory>
+
+#include <stddef.h>
+#include <stdint.h>
+
+
+namespace filament::backend {
+
+void* allocateFromCommandStream(DriverApi& driver, size_t size, size_t alignment) noexcept;
+
+class DescriptorSetOffsetArray {
+public:
+    using value_type = uint32_t;
+    using reference = value_type&;
+    using const_reference = value_type const&;
+    using size_type = uint32_t;
+    using difference_type = int32_t;
+    using pointer = value_type*;
+    using const_pointer = value_type const*;
+    using iterator = pointer;
+    using const_iterator = const_pointer;
+
+    DescriptorSetOffsetArray() noexcept = default;
+
+    ~DescriptorSetOffsetArray() noexcept = default;
+
+    DescriptorSetOffsetArray(size_type size, DriverApi& driver) noexcept {
+        mOffsets = (value_type *)allocateFromCommandStream(driver,
+                size * sizeof(value_type), alignof(value_type));
+        std::uninitialized_fill_n(mOffsets, size, 0);
+    }
+
+    DescriptorSetOffsetArray(std::initializer_list<uint32_t> list, DriverApi& driver) noexcept {
+        mOffsets = (value_type *)allocateFromCommandStream(driver,
+                list.size() * sizeof(value_type), alignof(value_type));
+        std::uninitialized_copy(list.begin(), list.end(), mOffsets);
+    }
+
+    DescriptorSetOffsetArray(DescriptorSetOffsetArray const&) = delete;
+    DescriptorSetOffsetArray& operator=(DescriptorSetOffsetArray const&) = delete;
+
+    DescriptorSetOffsetArray(DescriptorSetOffsetArray&& rhs) noexcept
+            : mOffsets(rhs.mOffsets) {
+        rhs.mOffsets = nullptr;
+    }
+
+    DescriptorSetOffsetArray& operator=(DescriptorSetOffsetArray&& rhs) noexcept {
+        if (this != &rhs) {
+            mOffsets = rhs.mOffsets;
+            rhs.mOffsets = nullptr;
+        }
+        return *this;
+    }
+
+    bool empty() const noexcept { return mOffsets == nullptr; }
+
+    value_type* data() noexcept { return mOffsets; }
+    const value_type* data() const noexcept { return mOffsets; }
+
+
+    reference operator[](size_type n) noexcept {
+        return *(data() + n);
+    }
+
+    const_reference operator[](size_type n) const noexcept {
+        return *(data() + n);
+    }
+
+    void clear() noexcept {
+        mOffsets = nullptr;
+    }
+
+private:
+    value_type *mOffsets = nullptr;
+};
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_COMMANDSTREAMVECTOR_H

filament/backend/include/backend/DriverEnums.h

@@ -19,13 +19,16 @@
 #ifndef TNT_FILAMENT_BACKEND_DRIVERENUMS_H
 #define TNT_FILAMENT_BACKEND_DRIVERENUMS_H
 
-#include <utils/BitmaskEnum.h>
 #include <utils/unwindows.h> // Because we define ERROR in the FenceStatus enum.
 
 #include <backend/Platform.h>
 #include <backend/PresentCallable.h>
 
+#include <utils/BitmaskEnum.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Invocable.h>
+#include <utils/compiler.h>
+#include <utils/debug.h>
 #include <utils/ostream.h>
 
 #include <math/vec4.h>
@@ -97,6 +100,8 @@ static constexpr size_t MAX_VERTEX_ATTRIBUTE_COUNT  = 16;   // This is guarantee
 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_DESCRIPTOR_SET_COUNT    = 4;    // This is guaranteed by Vulkan.
+static constexpr size_t MAX_DESCRIPTOR_COUNT        = 64;   // per set
 
 static constexpr size_t MAX_PUSH_CONSTANT_COUNT     = 32;   // Vulkan 1.1 spec allows for 128-byte
                                                             // of push constant (we assume 4-byte
@@ -191,6 +196,61 @@ static constexpr const char* shaderLanguageToString(ShaderLanguage shaderLanguag
     }
 }
 
+enum class ShaderStage : uint8_t {
+    VERTEX = 0,
+    FRAGMENT = 1,
+    COMPUTE = 2
+};
+
+static constexpr size_t PIPELINE_STAGE_COUNT = 2;
+enum class ShaderStageFlags : uint8_t {
+    NONE        =    0,
+    VERTEX      =    0x1,
+    FRAGMENT    =    0x2,
+    COMPUTE     =    0x4,
+    ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
+};
+
+static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
+    switch (type) {
+        case ShaderStage::VERTEX:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
+        case ShaderStage::FRAGMENT:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::FRAGMENT));
+        case ShaderStage::COMPUTE:
+            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::COMPUTE));
+    }
+}
+
+enum class DescriptorType : uint8_t {
+    UNIFORM_BUFFER,
+    SHADER_STORAGE_BUFFER,
+    SAMPLER,
+    INPUT_ATTACHMENT,
+};
+
+enum class DescriptorFlags : uint8_t {
+    NONE = 0x00,
+    DYNAMIC_OFFSET = 0x01
+};
+
+using descriptor_set_t = uint8_t;
+
+using descriptor_binding_t = uint8_t;
+
+struct DescriptorSetLayoutBinding {
+    DescriptorType type;
+    ShaderStageFlags stageFlags;
+    descriptor_binding_t binding;
+    DescriptorFlags flags;
+    uint16_t count;
+};
+
+struct DescriptorSetLayout {
+    utils::FixedCapacityVector<DescriptorSetLayoutBinding> bindings;
+};
+
+
 /**
  * Bitmask for selecting render buffers
  */
@@ -270,15 +330,6 @@ enum class FenceStatus : int8_t {
     TIMEOUT_EXPIRED = 1,        //!< wait()'s timeout expired. The Fence condition is not satisfied.
 };
 
-/**
- * Status codes for sync objects
- */
-enum class SyncStatus : int8_t {
-    ERROR = -1,          //!< An error occurred. The Sync is not signaled.
-    SIGNALED = 0,        //!< The Sync is signaled.
-    NOT_SIGNALED = 1,    //!< The Sync is not signaled yet
-};
-
 static constexpr uint64_t FENCE_WAIT_FOR_EVER = uint64_t(-1);
 
 /**
@@ -368,6 +419,18 @@ enum class SamplerType : uint8_t {
     SAMPLER_CUBEMAP_ARRAY,  //!< Cube map array texture (feature level 2)
 };
 
+inline const char* stringify(SamplerType samplerType) {
+    switch (samplerType) {
+        case SamplerType::SAMPLER_2D: return "SAMPLER_2D";
+        case SamplerType::SAMPLER_2D_ARRAY: return "SAMPLER_2D_ARRAY";
+        case SamplerType::SAMPLER_CUBEMAP: return "SAMPLER_CUBEMAP";
+        case SamplerType::SAMPLER_EXTERNAL: return "SAMPLER_EXTERNAL";
+        case SamplerType::SAMPLER_3D: return "SAMPLER_3D";
+        case SamplerType::SAMPLER_CUBEMAP_ARRAY: return "SAMPLER_CUBEMAP_ARRAY";
+    }
+    return "UNKNOWN";
+}
+
 //! Subpass type
 enum class SubpassType : uint8_t {
     SUBPASS_INPUT
@@ -696,6 +759,23 @@ enum class TextureUsage : uint16_t {
     DEFAULT             = UPLOADABLE | SAMPLEABLE   //!< Default texture usage
 };
 
+inline const char* stringify(TextureUsage usage) {
+    switch (usage) {
+        case TextureUsage::NONE: return "NONE";
+        case TextureUsage::COLOR_ATTACHMENT: return "COLOR_ATTACHMENT";
+        case TextureUsage::DEPTH_ATTACHMENT: return "DEPTH_ATTACHMENT";
+        case TextureUsage::STENCIL_ATTACHMENT: return "STENCIL_ATTACHMENT";
+        case TextureUsage::UPLOADABLE: return "UPLOADABLE";
+        case TextureUsage::SAMPLEABLE: return "SAMPLEABLE";
+        case TextureUsage::SUBPASS_INPUT: return "SUBPASS_INPUT";
+        case TextureUsage::BLIT_SRC: return "BLIT_SRC";
+        case TextureUsage::BLIT_DST: return "BLIT_DST";
+        case TextureUsage::PROTECTED: return "PROTECTED";
+        case TextureUsage::DEFAULT: return "DEFAULT";
+        default: return "UNKNOWN";
+    }
+}
+
 //! Texture swizzle
 enum class TextureSwizzle : uint8_t {
     SUBSTITUTE_ZERO,
@@ -887,6 +967,9 @@ struct SamplerParams { // NOLINT
 
     struct EqualTo {
         bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
+            assert_invariant(lhs.padding0 == 0);
+            assert_invariant(lhs.padding1 == 0);
+            assert_invariant(lhs.padding2 == 0);
             auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
             auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
             return *pLhs == *pRhs;
@@ -895,6 +978,9 @@ struct SamplerParams { // NOLINT
 
     struct LessThan {
         bool operator()(SamplerParams lhs, SamplerParams rhs) const noexcept {
+            assert_invariant(lhs.padding0 == 0);
+            assert_invariant(lhs.padding1 == 0);
+            assert_invariant(lhs.padding2 == 0);
             auto* pLhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&lhs));
             auto* pRhs = reinterpret_cast<uint32_t const*>(reinterpret_cast<char const*>(&rhs));
             return *pLhs == *pRhs;
@@ -902,6 +988,12 @@ struct SamplerParams { // NOLINT
     };
 
 private:
+    friend inline bool operator == (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return SamplerParams::EqualTo{}(lhs, rhs);
+    }
+    friend inline bool operator != (SamplerParams lhs, SamplerParams rhs) noexcept {
+        return  !SamplerParams::EqualTo{}(lhs, rhs);
+    }
     friend inline bool operator < (SamplerParams lhs, SamplerParams rhs) noexcept {
         return SamplerParams::LessThan{}(lhs, rhs);
     }
@@ -1058,7 +1150,7 @@ struct RasterState {
             bool inverseFrontFaces                      : 1;        // 31
 
             //! padding, must be 0
-            uint8_t padding                             : 1;        // 32
+            bool depthClamp                             : 1;        // 32
         };
         uint32_t u = 0;
     };
@@ -1069,32 +1161,6 @@ struct RasterState {
  * \privatesection
  */
 
-enum class ShaderStage : uint8_t {
-    VERTEX = 0,
-    FRAGMENT = 1,
-    COMPUTE = 2
-};
-
-static constexpr size_t PIPELINE_STAGE_COUNT = 2;
-enum class ShaderStageFlags : uint8_t {
-    NONE        =    0,
-    VERTEX      =    0x1,
-    FRAGMENT    =    0x2,
-    COMPUTE     =    0x4,
-    ALL_SHADER_STAGE_FLAGS = VERTEX | FRAGMENT | COMPUTE
-};
-
-static inline constexpr bool hasShaderType(ShaderStageFlags flags, ShaderStage type) noexcept {
-    switch (type) {
-        case ShaderStage::VERTEX:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::VERTEX));
-        case ShaderStage::FRAGMENT:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::FRAGMENT));
-        case ShaderStage::COMPUTE:
-            return bool(uint8_t(flags) & uint8_t(ShaderStageFlags::COMPUTE));
-    }
-}
-
 /**
  * Selects which buffers to clear at the beginning of the render pass, as well as which buffers
  * can be discarded at the beginning and end of the render pass.
@@ -1244,7 +1310,7 @@ enum class Workaround : uint16_t {
     ADRENO_UNIFORM_ARRAY_CRASH,
     // Workaround a Metal pipeline compilation error with the message:
     // "Could not statically determine the target of a texture". See light_indirect.fs
-    A8X_STATIC_TEXTURE_TARGET_ERROR,
+    METAL_STATIC_TEXTURE_TARGET_ERROR,
     // Adreno drivers sometimes aren't able to blit into a layer of a texture array.
     DISABLE_BLIT_INTO_TEXTURE_ARRAY,
     // Multiple workarounds needed for PowerVR GPUs
@@ -1259,6 +1325,8 @@ template<> struct utils::EnableBitMaskOperators<filament::backend::ShaderStageFl
         : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::TargetBufferFlags>
         : public std::true_type {};
+template<> struct utils::EnableBitMaskOperators<filament::backend::DescriptorFlags>
+        : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::TextureUsage>
         : public std::true_type {};
 template<> struct utils::EnableBitMaskOperators<filament::backend::StencilFace>

filament/backend/include/backend/Handle.h

@@ -41,6 +41,8 @@ struct HwTexture;
 struct HwTimerQuery;
 struct HwVertexBufferInfo;
 struct HwVertexBuffer;
+struct HwDescriptorSetLayout;
+struct HwDescriptorSet;
 
 /*
  * A handle to a backend resource. HandleBase is for internal use only.
@@ -130,19 +132,21 @@ private:
 
 // Types used by the command stream
 // (we use this renaming because the macro-system doesn't deal well with "<" and ">")
-using BufferObjectHandle     = Handle<HwBufferObject>;
-using FenceHandle            = Handle<HwFence>;
-using IndexBufferHandle      = Handle<HwIndexBuffer>;
-using ProgramHandle          = Handle<HwProgram>;
-using RenderPrimitiveHandle  = Handle<HwRenderPrimitive>;
-using RenderTargetHandle     = Handle<HwRenderTarget>;
-using SamplerGroupHandle     = Handle<HwSamplerGroup>;
-using StreamHandle           = Handle<HwStream>;
-using SwapChainHandle        = Handle<HwSwapChain>;
-using TextureHandle          = Handle<HwTexture>;
-using TimerQueryHandle       = Handle<HwTimerQuery>;
-using VertexBufferHandle     = Handle<HwVertexBuffer>;
-using VertexBufferInfoHandle = Handle<HwVertexBufferInfo>;
+using BufferObjectHandle        = Handle<HwBufferObject>;
+using FenceHandle               = Handle<HwFence>;
+using IndexBufferHandle         = Handle<HwIndexBuffer>;
+using ProgramHandle             = Handle<HwProgram>;
+using RenderPrimitiveHandle     = Handle<HwRenderPrimitive>;
+using RenderTargetHandle        = Handle<HwRenderTarget>;
+using SamplerGroupHandle        = Handle<HwSamplerGroup>;
+using StreamHandle              = Handle<HwStream>;
+using SwapChainHandle           = Handle<HwSwapChain>;
+using TextureHandle             = Handle<HwTexture>;
+using TimerQueryHandle          = Handle<HwTimerQuery>;
+using VertexBufferHandle        = Handle<HwVertexBuffer>;
+using VertexBufferInfoHandle    = Handle<HwVertexBufferInfo>;
+using DescriptorSetLayoutHandle = Handle<HwDescriptorSetLayout>;
+using DescriptorSetHandle       = Handle<HwDescriptorSet>;
 
 } // namespace filament::backend
 

filament/backend/include/backend/PipelineState.h

@@ -22,15 +22,23 @@
 
 #include <utils/ostream.h>
 
+#include <array>
+
 #include <stdint.h>
 
 namespace filament::backend {
 
 //! \privatesection
 
+struct PipelineLayout {
+    using SetLayout = std::array<Handle<HwDescriptorSetLayout>, MAX_DESCRIPTOR_SET_COUNT>;
+    SetLayout setLayout;      // 16
+};
+
 struct PipelineState {
     Handle<HwProgram> program;                                              //  4
     Handle<HwVertexBufferInfo> vertexBufferInfo;                            //  4
+    PipelineLayout pipelineLayout;                                          // 16
     RasterState rasterState;                                                //  4
     StencilState stencilState;                                              // 12
     PolygonOffset polygonOffset;                                            //  8

filament/backend/include/backend/Program.h

@@ -24,9 +24,11 @@
 
 #include <backend/DriverEnums.h>
 
-#include <array>    // FIXME: STL headers are not allowed in public headers
-#include <utility>  // FIXME: STL headers are not allowed in public headers
-#include <variant>  // FIXME: STL headers are not allowed in public headers
+#include <array>
+#include <unordered_map>
+#include <tuple>
+#include <utility>
+#include <variant>
 
 #include <stddef.h>
 #include <stdint.h>
@@ -40,29 +42,36 @@ public:
     static constexpr size_t UNIFORM_BINDING_COUNT = CONFIG_UNIFORM_BINDING_COUNT;
     static constexpr size_t SAMPLER_BINDING_COUNT = CONFIG_SAMPLER_BINDING_COUNT;
 
-    struct Sampler {
-        utils::CString name = {};   // name of the sampler in the shader
-        uint32_t binding = 0;       // binding point of the sampler in the shader
+    struct Descriptor {
+        utils::CString name;
+        backend::DescriptorType type;
+        backend::descriptor_binding_t binding;
     };
 
-    struct SamplerGroupData {
-        utils::FixedCapacityVector<Sampler> samplers;
-        ShaderStageFlags stageFlags = ShaderStageFlags::ALL_SHADER_STAGE_FLAGS;
+    struct SpecializationConstant {
+        using Type = std::variant<int32_t, float, bool>;
+        uint32_t id;    // id set in glsl
+        Type value;     // value and type
     };
 
-    struct Uniform {
+    struct Uniform { // For ES2 support
         utils::CString name;    // full qualified name of the uniform field
         uint16_t offset;        // offset in 'uint32_t' into the uniform buffer
         uint8_t size;           // >1 for arrays
         UniformType type;       // uniform type
     };
 
-    using UniformBlockInfo = std::array<utils::CString, UNIFORM_BINDING_COUNT>;
-    using UniformInfo = utils::FixedCapacityVector<Uniform>;
-    using SamplerGroupInfo = std::array<SamplerGroupData, SAMPLER_BINDING_COUNT>;
+    using DescriptorBindingsInfo = utils::FixedCapacityVector<Descriptor>;
+    using DescriptorSetInfo = std::array<DescriptorBindingsInfo, MAX_DESCRIPTOR_SET_COUNT>;
+    using SpecializationConstantsInfo = utils::FixedCapacityVector<SpecializationConstant>;
     using ShaderBlob = utils::FixedCapacityVector<uint8_t>;
     using ShaderSource = std::array<ShaderBlob, SHADER_TYPE_COUNT>;
 
+    using AttributesInfo = utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>>;
+    using UniformInfo = utils::FixedCapacityVector<Uniform>;
+    using BindingUniformsInfo = utils::FixedCapacityVector<
+            std::tuple<uint8_t, utils::CString, Program::UniformInfo>>;
+
     Program() noexcept;
 
     Program(const Program& rhs) = delete;
@@ -79,43 +88,19 @@ public:
     Program& diagnostics(utils::CString const& name,
             utils::Invocable<utils::io::ostream&(utils::io::ostream& out)>&& logger);
 
-    // sets one of the program's shader (e.g. vertex, fragment)
+    // Sets one of the program's shader (e.g. vertex, fragment)
     // string-based shaders are null terminated, consequently the size parameter must include the
     // 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)
+    // 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.
-    Program& uniformBlockBindings(
-            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept;
+    // Descriptor binding (set, binding, type -> shader name) info
+    Program& descriptorBindings(backend::descriptor_set_t set,
+            DescriptorBindingsInfo descriptorBindings) noexcept;
 
-    // Note: This is only needed for GLES2.0, this is used to emulate UBO. This function tells
-    // the program everything it needs to know about the uniforms at a given binding
-    Program& uniforms(uint32_t index, UniformInfo const& uniforms) noexcept;
-
-    // Note: This is only needed for GLES2.0.
-    Program& attributes(
-            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes) noexcept;
-
-    // sets the 'bindingPoint' sampler group descriptor for this program.
-    // 'samplers' can be destroyed after this call.
-    // This effectively associates a set of (BindingPoints, index) to a texture unit in the shader.
-    // Or more precisely, what layout(binding=) is set to in GLSL.
-    Program& setSamplerGroup(size_t bindingPoint, ShaderStageFlags stageFlags,
-            Sampler const* samplers, size_t count) noexcept;
-
-    struct SpecializationConstant {
-        using Type = std::variant<int32_t, float, bool>;
-        uint32_t id;                                // id set in glsl
-        Type value;                                 // value and type
-    };
-
-    Program& specializationConstants(
-            utils::FixedCapacityVector<SpecializationConstant> specConstants) noexcept;
+    Program& specializationConstants(SpecializationConstantsInfo specConstants) noexcept;
 
     struct PushConstant {
         utils::CString name;
@@ -129,33 +114,40 @@ public:
 
     Program& multiview(bool multiview) noexcept;
 
+    // For ES2 support only...
+    Program& uniforms(uint32_t index, utils::CString name, UniformInfo uniforms) noexcept;
+    Program& attributes(AttributesInfo attributes) noexcept;
+
+    //
+    // Getters for program construction...
+    //
+
     ShaderSource const& getShadersSource() const noexcept { return mShadersSource; }
     ShaderSource& getShadersSource() noexcept { return mShadersSource; }
 
-    UniformBlockInfo const& getUniformBlockBindings() const noexcept { return mUniformBlocks; }
-    UniformBlockInfo& getUniformBlockBindings() noexcept { return mUniformBlocks; }
-
-    SamplerGroupInfo const& getSamplerGroupInfo() const { return mSamplerGroups; }
-    SamplerGroupInfo& getSamplerGroupInfo() { return mSamplerGroups; }
-
-    auto const& getBindingUniformInfo() const { return mBindingUniformInfo; }
-    auto& getBindingUniformInfo() { return mBindingUniformInfo; }
-
-    auto const& getAttributes() const { return mAttributes; }
-    auto& getAttributes() { return mAttributes; }
-
     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 {
+    uint64_t getCacheId() const noexcept { return mCacheId; }
+
+    bool isMultiview() const noexcept { return mMultiview; }
+
+    CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
+
+    SpecializationConstantsInfo const& getSpecializationConstants() const noexcept {
         return mSpecializationConstants;
     }
-    utils::FixedCapacityVector<SpecializationConstant>& getSpecializationConstants() noexcept {
+
+    SpecializationConstantsInfo& getSpecializationConstants() noexcept {
         return mSpecializationConstants;
     }
 
+    DescriptorSetInfo& getDescriptorBindings() noexcept {
+        return mDescriptorBindings;
+    }
+
     utils::FixedCapacityVector<PushConstant> const& getPushConstants(
             ShaderStage stage) const noexcept {
         return mPushConstants[static_cast<uint8_t>(stage)];
@@ -165,27 +157,29 @@ public:
         return mPushConstants[static_cast<uint8_t>(stage)];
     }
 
-    uint64_t getCacheId() const noexcept { return mCacheId; }
+    auto const& getBindingUniformInfo() const { return mBindingUniformsInfo; }
+    auto& getBindingUniformInfo() { return mBindingUniformsInfo; }
 
-    bool isMultiview() const noexcept { return mMultiview; }
-
-    CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
+    auto const& getAttributes() const { return mAttributes; }
+    auto& getAttributes() { return mAttributes; }
 
 private:
     friend utils::io::ostream& operator<<(utils::io::ostream& out, const Program& builder);
 
-    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;
+    utils::Invocable<utils::io::ostream&(utils::io::ostream& out)> mLogger;
+    SpecializationConstantsInfo mSpecializationConstants;
+    std::array<utils::FixedCapacityVector<PushConstant>, SHADER_TYPE_COUNT> mPushConstants;
+    DescriptorSetInfo mDescriptorBindings;
+
+    // For ES2 support only
+    AttributesInfo mAttributes;
+    BindingUniformsInfo mBindingUniformsInfo;
+
     // Indicates the current engine was initialized with multiview stereo, and the variant for this
     // program contains STE flag. This will be referred later for the OpenGL shader compiler to
     // determine whether shader code replacement for the num_views should be performed.

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

@@ -18,6 +18,7 @@
 #define TNT_FILAMENT_BACKEND_PRIVATE_DRIVER_H
 
 #include <backend/CallbackHandler.h>
+#include <backend/DescriptorSetOffsetArray.h>
 #include <backend/DriverApiForward.h>
 #include <backend/DriverEnums.h>
 #include <backend/Handle.h>

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

@@ -162,6 +162,10 @@ DECL_DRIVER_API_0(finish)
 // reset state tracking, if the driver does any state tracking (e.g. GL)
 DECL_DRIVER_API_0(resetState)
 
+DECL_DRIVER_API_N(setDebugTag,
+        backend::HandleBase::HandleId, handleId,
+        utils::CString, tag)
+
 /*
  * Creating driver objects
  * -----------------------
@@ -196,20 +200,33 @@ DECL_DRIVER_API_R_N(backend::TextureHandle, createTexture,
         uint32_t, depth,
         backend::TextureUsage, usage)
 
-DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureSwizzled,
-        backend::SamplerType, target,
-        uint8_t, levels,
-        backend::TextureFormat, format,
-        uint8_t, samples,
-        uint32_t, width,
-        uint32_t, height,
-        uint32_t, depth,
-        backend::TextureUsage, usage,
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureView,
+        backend::TextureHandle, texture,
+        uint8_t, baseLevel,
+        uint8_t, levelCount)
+
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureViewSwizzle,
+        backend::TextureHandle, texture,
         backend::TextureSwizzle, r,
         backend::TextureSwizzle, g,
         backend::TextureSwizzle, b,
         backend::TextureSwizzle, a)
 
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureExternalImage,
+        backend::TextureFormat, format,
+        uint32_t, width,
+        uint32_t, height,
+        backend::TextureUsage, usage,
+        void*, image)
+
+DECL_DRIVER_API_R_N(backend::TextureHandle, createTextureExternalImagePlane,
+        backend::TextureFormat, format,
+        uint32_t, width,
+        uint32_t, height,
+        backend::TextureUsage, usage,
+        void*, image,
+        uint32_t, plane)
+
 DECL_DRIVER_API_R_N(backend::TextureHandle, importTexture,
         intptr_t, id,
         backend::SamplerType, target,
@@ -221,9 +238,6 @@ DECL_DRIVER_API_R_N(backend::TextureHandle, importTexture,
         uint32_t, depth,
         backend::TextureUsage, usage)
 
-DECL_DRIVER_API_R_N(backend::SamplerGroupHandle, createSamplerGroup,
-        uint32_t, size, utils::FixedSizeString<32>, debugName)
-
 DECL_DRIVER_API_R_N(backend::RenderPrimitiveHandle, createRenderPrimitive,
         backend::VertexBufferHandle, vbh,
         backend::IndexBufferHandle, ibh,
@@ -257,25 +271,53 @@ DECL_DRIVER_API_R_N(backend::SwapChainHandle, createSwapChainHeadless,
 
 DECL_DRIVER_API_R_0(backend::TimerQueryHandle, createTimerQuery)
 
+DECL_DRIVER_API_R_N(backend::DescriptorSetLayoutHandle, createDescriptorSetLayout,
+        backend::DescriptorSetLayout&&, info)
+
+DECL_DRIVER_API_R_N(backend::DescriptorSetHandle, createDescriptorSet,
+        backend::DescriptorSetLayoutHandle, dslh)
+
+DECL_DRIVER_API_N(updateDescriptorSetBuffer,
+        backend::DescriptorSetHandle, dsh,
+        backend::descriptor_binding_t, binding,
+        backend::BufferObjectHandle, boh,
+        uint32_t, offset,
+        uint32_t, size
+)
+
+DECL_DRIVER_API_N(updateDescriptorSetTexture,
+        backend::DescriptorSetHandle, dsh,
+        backend::descriptor_binding_t, binding,
+        backend::TextureHandle, th,
+        SamplerParams, params
+)
+
+DECL_DRIVER_API_N(bindDescriptorSet,
+        backend::DescriptorSetHandle, dsh,
+        backend::descriptor_set_t, set,
+        backend::DescriptorSetOffsetArray&&, offsets
+)
+
 
 /*
  * Destroying driver objects
  * -------------------------
  */
 
-DECL_DRIVER_API_N(destroyVertexBuffer,    backend::VertexBufferHandle, vbh)
-DECL_DRIVER_API_N(destroyVertexBufferInfo,backend::VertexBufferInfoHandle, vbih)
-DECL_DRIVER_API_N(destroyIndexBuffer,     backend::IndexBufferHandle, ibh)
-DECL_DRIVER_API_N(destroyBufferObject,    backend::BufferObjectHandle, ibh)
-DECL_DRIVER_API_N(destroyRenderPrimitive, backend::RenderPrimitiveHandle, rph)
-DECL_DRIVER_API_N(destroyProgram,         backend::ProgramHandle, ph)
-DECL_DRIVER_API_N(destroySamplerGroup,    backend::SamplerGroupHandle, sbh)
-DECL_DRIVER_API_N(destroyTexture,         backend::TextureHandle, th)
-DECL_DRIVER_API_N(destroyRenderTarget,    backend::RenderTargetHandle, rth)
-DECL_DRIVER_API_N(destroySwapChain,       backend::SwapChainHandle, sch)
-DECL_DRIVER_API_N(destroyStream,          backend::StreamHandle, sh)
-DECL_DRIVER_API_N(destroyTimerQuery,      backend::TimerQueryHandle, sh)
-DECL_DRIVER_API_N(destroyFence,           backend::FenceHandle, fh)
+DECL_DRIVER_API_N(destroyVertexBuffer,          backend::VertexBufferHandle, vbh)
+DECL_DRIVER_API_N(destroyVertexBufferInfo,      backend::VertexBufferInfoHandle, vbih)
+DECL_DRIVER_API_N(destroyIndexBuffer,           backend::IndexBufferHandle, ibh)
+DECL_DRIVER_API_N(destroyBufferObject,          backend::BufferObjectHandle, ibh)
+DECL_DRIVER_API_N(destroyRenderPrimitive,       backend::RenderPrimitiveHandle, rph)
+DECL_DRIVER_API_N(destroyProgram,               backend::ProgramHandle, ph)
+DECL_DRIVER_API_N(destroyTexture,               backend::TextureHandle, th)
+DECL_DRIVER_API_N(destroyRenderTarget,          backend::RenderTargetHandle, rth)
+DECL_DRIVER_API_N(destroySwapChain,             backend::SwapChainHandle, sch)
+DECL_DRIVER_API_N(destroyStream,                backend::StreamHandle, sh)
+DECL_DRIVER_API_N(destroyTimerQuery,            backend::TimerQueryHandle, sh)
+DECL_DRIVER_API_N(destroyFence,                 backend::FenceHandle, fh)
+DECL_DRIVER_API_N(destroyDescriptorSetLayout,   backend::DescriptorSetLayoutHandle, dslh)
+DECL_DRIVER_API_N(destroyDescriptorSet,         backend::DescriptorSetHandle, dsh)
 
 /*
  * Synchronous APIs
@@ -305,6 +347,7 @@ 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(bool, isDepthClampSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(uint8_t, getMaxDrawBuffers)
 DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getMaxUniformBufferSize)
 DECL_DRIVER_API_SYNCHRONOUS_0(math::float2, getClipSpaceParams)
@@ -341,15 +384,6 @@ DECL_DRIVER_API_N(updateBufferObjectUnsynchronized,
 DECL_DRIVER_API_N(resetBufferObject,
         backend::BufferObjectHandle, ibh)
 
-DECL_DRIVER_API_N(updateSamplerGroup,
-        backend::SamplerGroupHandle, ubh,
-        backend::BufferDescriptor&&, data)
-
-DECL_DRIVER_API_N(setMinMaxLevels,
-        backend::TextureHandle, th,
-        uint32_t, minLevel,
-        uint32_t, maxLevel)
-
 DECL_DRIVER_API_N(update3DImage,
         backend::TextureHandle, th,
         uint32_t, level,
@@ -364,10 +398,12 @@ DECL_DRIVER_API_N(update3DImage,
 DECL_DRIVER_API_N(generateMipmaps,
         backend::TextureHandle, th)
 
+// Deprecated
 DECL_DRIVER_API_N(setExternalImage,
         backend::TextureHandle, th,
         void*, image)
 
+// Deprecated
 DECL_DRIVER_API_N(setExternalImagePlane,
         backend::TextureHandle, th,
         void*, image,
@@ -414,37 +450,16 @@ DECL_DRIVER_API_N(commit,
  * -----------------------
  */
 
-DECL_DRIVER_API_N(bindUniformBuffer,
-        uint32_t, index,
-        backend::BufferObjectHandle, ubh)
-
-DECL_DRIVER_API_N(bindBufferRange,
-        BufferObjectBinding, bindingType,
-        uint32_t, index,
-        backend::BufferObjectHandle, ubh,
-        uint32_t, offset,
-        uint32_t, size)
-
-DECL_DRIVER_API_N(unbindBuffer,
-        BufferObjectBinding, bindingType,
-        uint32_t, index)
-
-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)
+        const char*, string)
 
 DECL_DRIVER_API_N(pushGroupMarker,
-        const char*, string,
-        uint32_t, len = 0)
+        const char*, string)
 
 DECL_DRIVER_API_0(popGroupMarker)
 

filament/backend/include/private/backend/DriverApi.h

@@ -18,6 +18,17 @@
 #define TNT_FILAMENT_BACKEND_PRIVATE_DRIVERAPI_H
 
 #include "backend/DriverApiForward.h"
+
 #include "private/backend/CommandStream.h"
 
+#include <stddef.h>
+
+namespace filament::backend {
+
+inline void* allocateFromCommandStream(DriverApi& driver, size_t size, size_t alignment) noexcept {
+    return driver.allocate(size, alignment);
+}
+
+} // namespace filament::backend
+
 #endif // TNT_FILAMENT_BACKEND_PRIVATE_DRIVERAPI_H

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

@@ -20,11 +20,12 @@
 #include <backend/Handle.h>
 
 #include <utils/Allocator.h>
+#include <utils/CString.h>
 #include <utils/Log.h>
+#include <utils/Panic.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/ostream.h>
-#include <utils/Panic.h>
 
 #include <tsl/robin_map.h>
 
@@ -37,9 +38,9 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#define HandleAllocatorGL  HandleAllocator<32,  64, 136>    // ~4520 / pool / MiB
+#define HandleAllocatorGL  HandleAllocator<32,  96, 136>    // ~4520 / pool / MiB
 #define HandleAllocatorVK  HandleAllocator<64, 160, 312>    // ~1820 / pool / MiB
-#define HandleAllocatorMTL HandleAllocator<32,  48, 552>    // ~1660 / pool / MiB
+#define HandleAllocatorMTL HandleAllocator<32,  64, 552>    // ~1660 / pool / MiB
 
 namespace filament::backend {
 
@@ -173,8 +174,10 @@ public:
                 uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
                 auto const pNode = static_cast<typename Allocator::Node*>(p);
                 uint8_t const expectedAge = pNode[-1].age;
-                FILAMENT_CHECK_POSTCONDITION(expectedAge == age) <<
-                        "use-after-free of Handle with id=" << handle.getId();
+                // getHandleTag() is only called if the check fails.
+                FILAMENT_CHECK_POSTCONDITION(expectedAge == age)
+                        << "use-after-free of Handle with id=" << handle.getId()
+                        << ", tag=" << getHandleTag(handle.getId()).c_str_safe();
             }
         }
 
@@ -201,6 +204,29 @@ public:
         return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
     }
 
+    void associateTagToHandle(HandleBase::HandleId id, utils::CString&& tag) noexcept {
+        // TODO: for now, only pool handles check for use-after-free, so we only keep tags for
+        // those
+        if (isPoolHandle(id)) {
+            // Truncate the age to get the debug tag
+            uint32_t const key = id & ~(HANDLE_DEBUG_TAG_MASK ^ HANDLE_AGE_MASK);
+            // This line is the costly part. In the future, we could potentially use a custom
+            // allocator.
+            mDebugTags[key] = std::move(tag);
+        }
+    }
+
+    utils::CString getHandleTag(HandleBase::HandleId id) const noexcept {
+        if (!isPoolHandle(id)) {
+            return "(no tag)";
+        }
+        uint32_t const key = id & ~(HANDLE_DEBUG_TAG_MASK ^ HANDLE_AGE_MASK);
+        if (auto pos = mDebugTags.find(key); pos != mDebugTags.end()) {
+            return pos->second;
+        }
+        return "(no tag)";
+    }
+
 private:
 
     template<typename D>
@@ -318,12 +344,24 @@ private:
         }
     }
 
-    // we handle a 4 bits age per address
-    static constexpr uint32_t HANDLE_HEAP_FLAG      = 0x80000000u;      // pool vs heap handle
-    static constexpr uint32_t HANDLE_AGE_MASK       = 0x78000000u;      // handle's age
-    static constexpr uint32_t HANDLE_INDEX_MASK     = 0x07FFFFFFu;      // handle index
-    static constexpr uint32_t HANDLE_TAG_MASK       = HANDLE_AGE_MASK;
-    static constexpr uint32_t HANDLE_AGE_SHIFT      = 27;
+    // number if bits allotted to the handle's age (currently 4 max)
+    static constexpr uint32_t HANDLE_AGE_BIT_COUNT = 4;
+    // number if bits allotted to the handle's debug tag (HANDLE_AGE_BIT_COUNT max)
+    static constexpr uint32_t HANDLE_DEBUG_TAG_BIT_COUNT = 2;
+    // bit shift for both the age and debug tag
+    static constexpr uint32_t HANDLE_AGE_SHIFT = 27;
+    // mask for the heap (vs pool) flag
+    static constexpr uint32_t HANDLE_HEAP_FLAG = 0x80000000u;
+    // mask for the age
+    static constexpr uint32_t HANDLE_AGE_MASK =
+            ((1 << HANDLE_AGE_BIT_COUNT) - 1) << HANDLE_AGE_SHIFT;
+    // mask for the debug tag
+    static constexpr uint32_t HANDLE_DEBUG_TAG_MASK =
+            ((1 << HANDLE_DEBUG_TAG_BIT_COUNT) - 1) << HANDLE_AGE_SHIFT;
+    // mask for the index
+    static constexpr uint32_t HANDLE_INDEX_MASK = 0x07FFFFFFu;
+
+    static_assert(HANDLE_DEBUG_TAG_BIT_COUNT <= HANDLE_AGE_BIT_COUNT);
 
     static bool isPoolHandle(HandleBase::HandleId id) noexcept {
         return (id & HANDLE_HEAP_FLAG) == 0u;
@@ -338,7 +376,7 @@ private:
         // a non-pool handle.
         if (UTILS_LIKELY(isPoolHandle(id))) {
             char* const base = (char*)mHandleArena.getArea().begin();
-            uint32_t const tag = id & HANDLE_TAG_MASK;
+            uint32_t const tag = id & HANDLE_AGE_MASK;
             size_t const offset = (id & HANDLE_INDEX_MASK) * Allocator::getAlignment();
             return { static_cast<void*>(base + offset), tag };
         }
@@ -353,7 +391,7 @@ private:
         size_t const offset = (char*)p - base;
         assert_invariant((offset % Allocator::getAlignment()) == 0);
         auto id = HandleBase::HandleId(offset / Allocator::getAlignment());
-        id |= tag & HANDLE_TAG_MASK;
+        id |= tag & HANDLE_AGE_MASK;
         assert_invariant((id & HANDLE_HEAP_FLAG) == 0);
         return id;
     }
@@ -363,6 +401,7 @@ private:
     // Below is only used when running out of space in the HandleArena
     mutable utils::Mutex mLock;
     tsl::robin_map<HandleBase::HandleId, void*> mOverflowMap;
+    tsl::robin_map<HandleBase::HandleId, utils::CString> mDebugTags;
     HandleBase::HandleId mId = 0;
     bool mUseAfterFreeCheckDisabled = false;
 };

filament/backend/src/CircularBuffer.cpp

@@ -24,7 +24,7 @@
 #include <utils/debug.h>
 #include <utils/ostream.h>
 
-#if !defined(WIN32) && !defined(__EMSCRIPTEN__) && !defined(IOS)
+#if !defined(WIN32) && !defined(__EMSCRIPTEN__)
 #    include <sys/mman.h>
 #    include <unistd.h>
 #    define HAS_MMAP 1
@@ -32,10 +32,11 @@
 #    define HAS_MMAP 0
 #endif
 
-#include <stdint.h>
 #include <stddef.h>
-#include <stdlib.h>
+#include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 
 using namespace utils;
 
@@ -81,6 +82,9 @@ void* CircularBuffer::alloc(size_t size) noexcept {
             // map the circular buffer once...
             vaddr = mmap(reserve_vaddr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
             if (vaddr != MAP_FAILED) {
+                // populate the address space with pages (because this is a circular buffer,
+                // all the pages will be allocated eventually, might as well do it now)
+                memset(vaddr, 0, size);
                 // and map the circular buffer again, behind the previous copy...
                 vaddr_shadow = mmap((char*)vaddr + size, size,
                         PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
@@ -101,7 +105,7 @@ void* CircularBuffer::alloc(size_t size) noexcept {
     if (UTILS_UNLIKELY(mAshmemFd < 0)) {
         // ashmem failed
         if (vaddr_guard != MAP_FAILED) {
-            munmap(vaddr_guard, size);
+            munmap(vaddr_guard, BLOCK_SIZE);
         }
 
         if (vaddr_shadow != MAP_FAILED) {
@@ -119,12 +123,11 @@ void* CircularBuffer::alloc(size_t size) noexcept {
         data = mmap(nullptr, size * 2 + BLOCK_SIZE,
                 PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 
-        FILAMENT_CHECK_POSTCONDITION(data) <<
+        FILAMENT_CHECK_POSTCONDITION(data != MAP_FAILED) <<
                 "couldn't allocate " << (size * 2 / 1024) <<
                 " KiB of virtual address space for the command buffer";
 
-        slog.d << "WARNING: Using soft CircularBuffer (" << (size * 2 / 1024) << " KiB)"
-               << io::endl;
+        slog.w << "Using 'soft' CircularBuffer (" << (size * 2 / 1024) << " KiB)" << io::endl;
 
         // guard page at the end
         void* guard = (void*)(uintptr_t(data) + size * 2);

filament/backend/src/CommandBufferQueue.cpp

@@ -101,9 +101,8 @@ void CommandBufferQueue::flush() noexcept {
     size_t const used = std::distance(
             static_cast<char const*>(begin), static_cast<char const*>(end));
 
+
     std::unique_lock<utils::Mutex> lock(mLock);
-    mCommandBuffersToExecute.push_back({ begin, end });
-    mCondition.notify_one();
 
     // circular buffer is too small, we corrupted the stream
     FILAMENT_CHECK_POSTCONDITION(used <= mFreeSpace) <<
@@ -112,10 +111,12 @@ void CommandBufferQueue::flush() noexcept {
             "Space used at this time: " << used <<
             " bytes, overflow: " << used - mFreeSpace << " bytes";
 
-    // wait until there is enough space in the buffer
     mFreeSpace -= used;
-    if (UTILS_UNLIKELY(mFreeSpace < requiredSize)) {
+    mCommandBuffersToExecute.push_back({ begin, end });
+    mCondition.notify_one();
 
+    // wait until there is enough space in the buffer
+    if (UTILS_UNLIKELY(mFreeSpace < requiredSize)) {
 
 #ifndef NDEBUG
         size_t const totalUsed = circularBuffer.size() - mFreeSpace;
@@ -153,8 +154,10 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
 }
 
 void CommandBufferQueue::releaseBuffer(CommandBufferQueue::Range const& buffer) {
+    size_t const used = std::distance(
+            static_cast<char const*>(buffer.begin), static_cast<char const*>(buffer.end));
     std::lock_guard<utils::Mutex> const lock(mLock);
-    mFreeSpace += uintptr_t(buffer.end) - uintptr_t(buffer.begin);
+    mFreeSpace += used;
     mCondition.notify_one();
 }
 

filament/backend/src/CommandStream.cpp

@@ -20,11 +20,16 @@
 #include <utils/CallStack.h>
 #endif
 
+#include <utils/compiler.h>
 #include <utils/Log.h>
+#include <utils/ostream.h>
 #include <utils/Profiler.h>
 #include <utils/Systrace.h>
 
+#include <cstddef>
 #include <functional>
+#include <string>
+#include <utility>
 
 #ifdef __ANDROID__
 #include <sys/system_properties.h>
@@ -74,8 +79,8 @@ CommandStream::CommandStream(Driver& driver, CircularBuffer& buffer) noexcept
 }
 
 void CommandStream::execute(void* buffer) {
-    SYSTRACE_CALL();
-    SYSTRACE_CONTEXT();
+    // NOTE: we can't use SYSTRACE_CALL() or similar here because, execute() below, also
+    // uses systrace BEGIN/END and the END is not guaranteed to be happening in this scope.
 
     Profiler profiler;
 
@@ -100,6 +105,7 @@ void CommandStream::execute(void* buffer) {
             // we want to remove all this when tracing is completely disabled
             profiler.stop();
             UTILS_UNUSED Profiler::Counters const counters = profiler.readCounters();
+            SYSTRACE_CONTEXT();
             SYSTRACE_VALUE32("GLThread (I)", counters.getInstructions());
             SYSTRACE_VALUE32("GLThread (C)", counters.getCpuCycles());
             SYSTRACE_VALUE32("GLThread (CPI x10)", counters.getCPI() * 10);

filament/backend/src/DriverBase.h

@@ -101,6 +101,14 @@ struct HwProgram : public HwBase {
     HwProgram() noexcept = default;
 };
 
+struct HwDescriptorSetLayout : public HwBase {
+    HwDescriptorSetLayout() noexcept = default;
+};
+
+struct HwDescriptorSet : public HwBase {
+    HwDescriptorSet() noexcept = default;
+};
+
 struct HwSamplerGroup : public HwBase {
     HwSamplerGroup() noexcept = default;
 };

filament/backend/src/HandleAllocator.cpp

@@ -80,6 +80,9 @@ HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size,
         bool disableUseAfterFreeCheck) noexcept
     : mHandleArena(name, size, disableUseAfterFreeCheck),
       mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
+    // Reserve initial space for debug tags. This prevents excessive calls to malloc when the first
+    // few tags are set.
+    mDebugTags.reserve(512);
 }
 
 template <size_t P0, size_t P1, size_t P2>

filament/backend/src/Program.cpp

@@ -14,7 +14,18 @@
  * limitations under the License.
  */
 
-#include "backend/Program.h"
+#include <backend/Program.h>
+#include <backend/DriverEnums.h>
+
+#include <utils/debug.h>
+#include <utils/CString.h>
+#include <utils/ostream.h>
+#include <utils/Invocable.h>
+
+#include <utility>
+
+#include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
@@ -52,41 +63,24 @@ Program& Program::shaderLanguage(ShaderLanguage shaderLanguage) {
     return *this;
 }
 
-Program& Program::uniformBlockBindings(
-        FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& uniformBlockBindings) noexcept {
-    for (auto const& item : uniformBlockBindings) {
-        assert_invariant(item.second < UNIFORM_BINDING_COUNT);
-        mUniformBlocks[item.second] = item.first;
-    }
+Program& Program::descriptorBindings(backend::descriptor_set_t set,
+        DescriptorBindingsInfo descriptorBindings) noexcept {
+    mDescriptorBindings[set] = std::move(descriptorBindings);
     return *this;
 }
 
-Program& Program::uniforms(uint32_t index, UniformInfo const& uniforms) noexcept {
-    assert_invariant(index < UNIFORM_BINDING_COUNT);
-    mBindingUniformInfo[index] = uniforms;
+Program& Program::uniforms(uint32_t index, utils::CString name, UniformInfo uniforms) noexcept {
+    mBindingUniformsInfo.reserve(mBindingUniformsInfo.capacity() + 1);
+    mBindingUniformsInfo.emplace_back(index, std::move(name), std::move(uniforms));
     return *this;
 }
 
-
-Program& Program::attributes(
-        utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes) noexcept {
+Program& Program::attributes(AttributesInfo attributes) noexcept {
     mAttributes = std::move(attributes);
     return *this;
 }
 
-Program& Program::setSamplerGroup(size_t bindingPoint, ShaderStageFlags stageFlags,
-        const Program::Sampler* samplers, size_t count) noexcept {
-    auto& groupData = mSamplerGroups[bindingPoint];
-    groupData.stageFlags = stageFlags;
-    auto& samplerList = groupData.samplers;
-    samplerList.reserve(count);
-    samplerList.resize(count);
-    std::copy_n(samplers, count, samplerList.data());
-    return *this;
-}
-
-Program& Program::specializationConstants(
-        FixedCapacityVector<SpecializationConstant> specConstants) noexcept {
+Program& Program::specializationConstants(SpecializationConstantsInfo specConstants) noexcept {
     mSpecializationConstants = std::move(specConstants);
     return *this;
 }

filament/backend/src/metal/MetalBuffer.h

@@ -160,6 +160,8 @@ public:
          size_t size, bool forceGpuBuffer = false);
     ~MetalBuffer();
 
+    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer || mCpuBuffer; }
+
     MetalBuffer(const MetalBuffer& rhs) = delete;
     MetalBuffer& operator=(const MetalBuffer& rhs) = delete;
 
@@ -180,7 +182,7 @@ public:
      * is no device allocation.
      *
      */
-    id<MTLBuffer> getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept;
+    id<MTLBuffer> getGpuBufferForDraw() noexcept;
 
     void* getCpuBuffer() const noexcept { return mCpuBuffer; }
 

filament/backend/src/metal/MetalBuffer.mm

@@ -40,12 +40,15 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
     // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
     // buffer. Instead, we use immediate command encoder methods like setVertexBytes:length:atIndex:.
     // This won't work for SSBOs, since they are read/write.
+
+    /*
     if (size <= 4 * 1024 && bindingType != BufferObjectBinding::SHADER_STORAGE &&
             usage == BufferUsage::DYNAMIC && !forceGpuBuffer) {
         mBuffer = nil;
         mCpuBuffer = malloc(size);
         return;
     }
+    */
 
     // Otherwise, we allocate a private GPU buffer.
     {
@@ -53,8 +56,8 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
         mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
             TrackedMetalBuffer::Type::GENERIC };
     }
-    FILAMENT_CHECK_POSTCONDITION(mBuffer)
-            << "Could not allocate Metal buffer of size " << size << ".";
+    // mBuffer might fail to be allocated. Clients can check for this by calling
+    // wasAllocationSuccessful().
 }
 
 MetalBuffer::~MetalBuffer() {
@@ -94,7 +97,7 @@ void MetalBuffer::copyIntoBufferUnsynchronized(void* src, size_t size, size_t by
     copyIntoBuffer(src, size, byteOffset);
 }
 
-id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) noexcept {
+id<MTLBuffer> MetalBuffer::getGpuBufferForDraw() noexcept {
     // If there's a CPU buffer, then we return nil here, as the CPU-side buffer will be bound
     // separately.
     if (mCpuBuffer) {
@@ -137,7 +140,7 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
         }
         // getGpuBufferForDraw() might return nil, which means there isn't a device allocation for
         // this buffer. In this case, we'll bind the buffer below with the CPU-side memory.
-        id<MTLBuffer> gpuBuffer = buffer->getGpuBufferForDraw(cmdBuffer);
+        id<MTLBuffer> gpuBuffer = buffer->getGpuBufferForDraw();
         if (!gpuBuffer) {
             continue;
         }

filament/backend/src/metal/MetalBufferPool.mm

@@ -145,6 +145,7 @@ std::pair<id<MTLBuffer>, size_t> MetalBumpAllocator::allocateStagingArea(size_t
     // We're finished with the current allocation.
     mCurrentUploadBuffer = { [mDevice newBufferWithLength:mCapacity options:MTLStorageModeShared],
         TrackedMetalBuffer::Type::STAGING };
+    mHead = size;
 
     return { mCurrentUploadBuffer.get(), 0 };
 }

filament/backend/src/metal/MetalContext.h

@@ -21,6 +21,8 @@
 #include "MetalShaderCompiler.h"
 #include "MetalState.h"
 
+#include <backend/DriverEnums.h>
+
 #include <CoreVideo/CVMetalTextureCache.h>
 #include <Metal/Metal.h>
 #include <QuartzCore/QuartzCore.h>
@@ -46,13 +48,13 @@ class MetalBlitter;
 class MetalBufferPool;
 class MetalBumpAllocator;
 class MetalRenderTarget;
-class MetalSamplerGroup;
 class MetalSwapChain;
 class MetalTexture;
 class MetalTimerQueryInterface;
 struct MetalUniformBuffer;
 struct MetalIndexBuffer;
 struct MetalVertexBuffer;
+struct MetalDescriptorSet;
 
 constexpr static uint8_t MAX_SAMPLE_COUNT = 8;  // Metal devices support at most 8 MSAA samples
 
@@ -68,6 +70,53 @@ private:
     bool mDirty = false;
 };
 
+class MetalDynamicOffsets {
+public:
+    void setOffsets(uint32_t set, const uint32_t* offsets, uint32_t count) {
+        assert(set < MAX_DESCRIPTOR_SET_COUNT);
+
+        auto getStartIndexForSet = [&](uint32_t s) {
+            uint32_t startIndex = 0;
+            for (uint32_t i = 0; i < s; i++) {
+                startIndex += mCounts[i];
+            }
+            return startIndex;
+        };
+
+        const bool resizeNecessary = mCounts[set] != count;
+        if (UTILS_UNLIKELY(resizeNecessary)) {
+            int delta = count - mCounts[set];
+
+            auto thisSetStart = mOffsets.begin() + getStartIndexForSet(set);
+            if (delta > 0) {
+                mOffsets.insert(thisSetStart, delta, 0);
+            } else {
+                mOffsets.erase(thisSetStart, thisSetStart - delta);
+            }
+
+            mCounts[set] = count;
+        }
+
+        if (resizeNecessary ||
+                !std::equal(
+                        offsets, offsets + count, mOffsets.begin() + getStartIndexForSet(set))) {
+            std::copy(offsets, offsets + count, mOffsets.begin() + getStartIndexForSet(set));
+            mDirty = true;
+        }
+    }
+    bool isDirty() const { return mDirty; }
+    void setDirty(bool dirty) { mDirty = dirty; }
+
+    std::pair<uint32_t, const uint32_t*> getOffsets() const {
+        return { mOffsets.size(), mOffsets.data() };
+    }
+
+private:
+    std::array<uint32_t, MAX_DESCRIPTOR_SET_COUNT> mCounts = { 0 };
+    std::vector<uint32_t> mOffsets;
+    bool mDirty = false;
+};
+
 struct MetalContext {
     explicit MetalContext(size_t metalFreedTextureListSize)
         : texturesToDestroy(metalFreedTextureListSize) {}
@@ -76,8 +125,12 @@ struct MetalContext {
     id<MTLDevice> device = nullptr;
     id<MTLCommandQueue> commandQueue = nullptr;
 
-    id<MTLCommandBuffer> pendingCommandBuffer = nullptr;
-    id<MTLRenderCommandEncoder> currentRenderPassEncoder = nullptr;
+    // The ID of pendingCommandBuffer (or the next command buffer, if pendingCommandBuffer is nil).
+    uint64_t pendingCommandBufferId = 1;
+    // read from driver thread, set from completion handlers
+    std::atomic<uint64_t> latestCompletedCommandBufferId = 0;
+    id<MTLCommandBuffer> pendingCommandBuffer = nil;
+    id<MTLRenderCommandEncoder> currentRenderPassEncoder = nil;
 
     std::atomic<bool> memorylessLimitsReached = false;
 
@@ -93,7 +146,7 @@ struct MetalContext {
     } highestSupportedGpuFamily;
 
     struct {
-        bool a8xStaticTextureTargetError;
+        bool staticTextureTargetError;
     } bugs;
 
     // sampleCountLookup[requestedSamples] gives a <= sample count supported by the device.
@@ -108,10 +161,9 @@ struct MetalContext {
     // State trackers.
     PipelineStateTracker pipelineState;
     DepthStencilStateTracker depthStencilState;
-    std::array<BufferState, Program::UNIFORM_BINDING_COUNT> uniformState;
-    std::array<BufferState, MAX_SSBO_COUNT> ssboState;
     CullModeStateTracker cullModeState;
     WindingStateTracker windingState;
+    DepthClampStateTracker depthClampState;
     Handle<HwRenderPrimitive> currentRenderPrimitive;
 
     // State caches.
@@ -124,13 +176,15 @@ struct MetalContext {
 
     std::array<MetalPushConstantBuffer, Program::SHADER_TYPE_COUNT> currentPushConstants;
 
-    MetalSamplerGroup* samplerBindings[Program::SAMPLER_BINDING_COUNT] = {};
+    // Keeps track of descriptor sets we've finalized for the current render pass.
+    tsl::robin_set<MetalDescriptorSet*> finalizedDescriptorSets;
+    std::array<MetalDescriptorSet*, MAX_DESCRIPTOR_SET_COUNT> currentDescriptorSets = {};
+    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::VERTEX> vertexDescriptorBindings;
+    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::FRAGMENT> fragmentDescriptorBindings;
+    MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::COMPUTE> computeDescriptorBindings;
+    MetalDynamicOffsets dynamicOffsets;
 
-    // Keeps track of sampler groups we've finalized for the current render pass.
-    tsl::robin_set<MetalSamplerGroup*> finalizedSamplerGroups;
-
-    // Keeps track of all alive sampler groups, textures.
-    tsl::robin_set<MetalSamplerGroup*> samplerGroups;
+    // Keeps track of all alive textures.
     tsl::robin_set<MetalTexture*> textures;
 
     // This circular buffer implements delayed destruction for Metal texture handles. It keeps a
@@ -153,6 +207,7 @@ struct MetalContext {
 
     // Empty texture used to prevent GPU errors when a sampler has been bound without a texture.
     id<MTLTexture> emptyTexture = nil;
+    id<MTLBuffer> emptyBuffer = nil;
 
     MetalBlitter* blitter = nullptr;
 

filament/backend/src/metal/MetalContext.mm

@@ -101,9 +101,14 @@ id<MTLCommandBuffer> getPendingCommandBuffer(MetalContext* context) {
     context->pendingCommandBuffer = [context->commandQueue commandBuffer];
     // It's safe for this block to capture the context variable. MetalDriver::terminate will ensure
     // all frames and their completion handlers finish before context is deallocated.
+    uint64_t thisCommandBufferId = context->pendingCommandBufferId;
     [context->pendingCommandBuffer addCompletedHandler:^(id <MTLCommandBuffer> buffer) {
         context->resourceTracker.clearResources((__bridge void*) buffer);
-        
+
+        // Command buffers should complete in order, so latestCompletedCommandBufferId will only
+        // ever increase.
+        context->latestCompletedCommandBufferId = thisCommandBufferId;
+
         auto errorCode = (MTLCommandBufferError)buffer.error.code;
         if (@available(macOS 11.0, *)) {
             if (errorCode == MTLCommandBufferErrorMemoryless) {
@@ -125,6 +130,7 @@ void submitPendingCommands(MetalContext* context) {
     assert_invariant(context->pendingCommandBuffer.status != MTLCommandBufferStatusCommitted);
     [context->pendingCommandBuffer commit];
     context->pendingCommandBuffer = nil;
+    context->pendingCommandBufferId++;
 }
 
 id<MTLTexture> getOrCreateEmptyTexture(MetalContext* context) {
@@ -167,7 +173,6 @@ void MetalPushConstantBuffer::setPushConstant(PushConstantVariant value, uint8_t
 
 void MetalPushConstantBuffer::setBytes(id<MTLCommandEncoder> encoder, ShaderStage stage) {
     constexpr size_t PUSH_CONSTANT_SIZE_BYTES = 4;
-    constexpr size_t PUSH_CONSTANT_BUFFER_INDEX = 26;
 
     static char buffer[MAX_PUSH_CONSTANT_COUNT * PUSH_CONSTANT_SIZE_BYTES];
     assert_invariant(mPushConstants.size() <= MAX_PUSH_CONSTANT_COUNT);

filament/backend/src/metal/MetalDriver.h

@@ -32,6 +32,7 @@
 #include <functional>
 #include <mutex>
 #include <vector>
+#include <deque>
 
 namespace filament {
 namespace backend {
@@ -57,11 +58,11 @@ class MetalDriver final : public DriverBase {
 
 public:
     static Driver* create(MetalPlatform* platform, const Platform::DriverConfig& driverConfig);
-    void runAtNextTick(const std::function<void()>& fn) noexcept;
 
 private:
 
     friend class MetalSwapChain;
+    friend struct MetalDescriptorSet;
 
     MetalPlatform& mPlatform;
     MetalContext* mContext;
@@ -73,10 +74,23 @@ private:
 
     /*
      * Tasks run regularly on the driver thread.
+     * Not thread-safe; tasks are run from the driver thead and must be enqueued from the driver
+     * thread.
      */
+    void runAtNextTick(const std::function<void()>& fn) noexcept;
     void executeTickOps() noexcept;
     std::vector<std::function<void()>> mTickOps;
-    std::mutex mTickOpsLock;
+
+    // Tasks regularly executed on the driver thread after a command buffer has completed
+    struct DeferredTask {
+        DeferredTask(uint64_t commandBufferId, utils::Invocable<void()>&& fn) noexcept
+            : commandBufferId(commandBufferId), fn(std::move(fn)) {}
+        uint64_t commandBufferId;     // after this command buffer completes
+        utils::Invocable<void()> fn;  // execute this task
+    };
+    void executeAfterCurrentCommandBufferCompletes(utils::Invocable<void()>&& fn) noexcept;
+    void executeDeferredOps() noexcept;
+    std::deque<DeferredTask> mDeferredTasks;
 
     /*
      * Driver interface
@@ -137,7 +151,6 @@ private:
     inline void setRenderPrimitiveBuffer(Handle<HwRenderPrimitive> rph, PrimitiveType pt,
             Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh);
 
-    void finalizeSamplerGroup(MetalSamplerGroup* sg);
     void enumerateBoundBuffers(BufferObjectBinding bindingType,
             const std::function<void(const BufferState&, MetalBuffer*, uint32_t)>& f);
 

filament/backend/src/metal/MetalExternalImage.h

@@ -32,100 +32,75 @@ struct MetalContext;
  * texture.
  */
 class MetalExternalImage {
-
 public:
+    MetalExternalImage() = default;
 
-    MetalExternalImage(MetalContext& context,
-            TextureSwizzle r = TextureSwizzle::CHANNEL_0,
-            TextureSwizzle g = TextureSwizzle::CHANNEL_1,
-            TextureSwizzle b = TextureSwizzle::CHANNEL_2,
-            TextureSwizzle a = TextureSwizzle::CHANNEL_3) noexcept;
+    MetalExternalImage(MetalExternalImage&&);
+    MetalExternalImage& operator=(MetalExternalImage&&);
+    ~MetalExternalImage() noexcept;
+
+    MetalExternalImage(const MetalExternalImage&) = delete;
+    MetalExternalImage& operator=(const MetalExternalImage&) = delete;
 
     /**
-     * @return true, if this MetalExternalImage is holding a live external image. Returns false
-     * until set has been called with a valid CVPixelBuffer. The image can be cleared via
-     * set(nullptr), and isValid will return false again.
+     * While the texture is used for rendering, this MetalExternalImage must be kept alive.
      */
-    bool isValid() const noexcept;
+    id<MTLTexture> getMtlTexture() const noexcept;
+
+    bool isValid() const noexcept {
+        return mImage != nil || mRgbTexture != nullptr;
+    }
+
+    NSUInteger getWidth() const noexcept;
+    NSUInteger getHeight() const noexcept;
 
     /**
-     * Set this external image to the passed-in CVPixelBuffer. Future calls to
-     * getMetalTextureForDraw will return a texture backed by this CVPixelBuffer. Previous
-     * CVPixelBuffers and related resources will be released when all GPU work using them has
-     * finished.
+     * Create an external image with the passed-in CVPixelBuffer.
      *
-     * Calling set with a YCbCr image will encode a compute pass to convert the image from YCbCr to
-     * RGB.
+     * Ownership is taken of the CVPixelBuffer, which will be released when the returned
+     * MetalExternalImage is destroyed (or, in the case of a YCbCr image, after the conversion has
+     * completed).
+     *
+     * Calling set with a YCbCr image will encode a compute pass to convert the image from
+     * YCbCr to RGB.
      */
-    void set(CVPixelBufferRef image) noexcept;
+    static MetalExternalImage createFromImage(MetalContext& context, CVPixelBufferRef image);
 
     /**
-     * Set this external image to a specific plane of the passed-in CVPixelBuffer. Future calls to
-     * getMetalTextureForDraw will return a texture backed by a single plane of this CVPixelBuffer.
-     * Previous CVPixelBuffers and related resources will be released when all GPU work using them
-     * has finished.
+     * Create an external image with a specific plane of the passed-in CVPixelBuffer.
+     *
+     * Ownership is taken of the CVPixelBuffer, which will be released when the returned
+     * MetalExternalImage is destroyed.
      */
-    void set(CVPixelBufferRef image, size_t plane) noexcept;
+    static MetalExternalImage createFromImagePlane(
+            MetalContext& context, CVPixelBufferRef image, uint32_t plane);
 
-    /**
-     * Returns the width of the external image, or 0 if one is not set. For YCbCr images, returns
-     * the width of the luminance plane.
-     */
-    size_t getWidth() const noexcept { return mWidth; }
-
-    /**
-     * Returns the height of the external image, or 0 if one is not set. For YCbCr images, returns
-     * the height of the luminance plane.
-     */
-    size_t getHeight() const noexcept { return mHeight; }
-
-    /**
-     * Get a Metal texture used to draw this image and denote that it is used for the current frame.
-     * For future frames that use this external image, getMetalTextureForDraw must be called again.
-     */
-    id<MTLTexture> getMetalTextureForDraw() const noexcept;
+    static void assertWritableImage(CVPixelBufferRef image);
 
     /**
      * Free resources. Should be called at least once when no further calls to set will occur.
      */
     static void shutdown(MetalContext& context) noexcept;
 
-    static void assertWritableImage(CVPixelBufferRef image);
-
 private:
+    MetalExternalImage(CVPixelBufferRef image, CVMetalTextureRef texture) noexcept
+        : mImage(image), mTexture(texture) {}
+    explicit MetalExternalImage(id<MTLTexture> texture) noexcept : mRgbTexture(texture) {}
 
-    void unset();
-
-    CVMetalTextureRef createTextureFromImage(CVPixelBufferRef image, MTLPixelFormat format,
-            size_t plane);
-    id<MTLTexture> createRgbTexture(size_t width, size_t height);
-    id<MTLTexture> createSwizzledTextureView(id<MTLTexture> texture) const;
-    id<MTLTexture> createSwizzledTextureView(CVMetalTextureRef texture) const;
-    void ensureComputePipelineState();
-    id<MTLCommandBuffer> encodeColorConversionPass(id<MTLTexture> inYPlane, id<MTLTexture>
-            inCbCrTexture, id<MTLTexture> outTexture);
+    static id<MTLTexture> createRgbTexture(id<MTLDevice> device, size_t width, size_t height);
+    static CVMetalTextureRef createTextureFromImage(CVMetalTextureCacheRef textureCache,
+            CVPixelBufferRef image, MTLPixelFormat format, size_t plane);
+    static void ensureComputePipelineState(MetalContext& context);
+    static id<MTLCommandBuffer> encodeColorConversionPass(MetalContext& context,
+            id<MTLTexture> inYPlane, id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture);
 
     static constexpr size_t Y_PLANE = 0;
     static constexpr size_t CBCR_PLANE = 1;
 
-    MetalContext& mContext;
-
-    // If the external image has a single plane, mImage and mTexture hold references to the image
-    // and created Metal texture, respectively.
-    // mTextureView is a view of mTexture with any swizzling applied.
+    // TODO: this could probably be a union.
     CVPixelBufferRef mImage = nullptr;
     CVMetalTextureRef mTexture = nullptr;
-    id<MTLTexture> mTextureView = nullptr;
-    size_t mWidth = 0;
-    size_t mHeight = 0;
-
-    // If the external image is in the YCbCr format, this holds the result of the converted RGB
-    // texture.
     id<MTLTexture> mRgbTexture = nil;
-
-    struct {
-        TextureSwizzle r, g, b, a;
-    } mSwizzle;
 };
 
 } // namespace backend

filament/backend/src/metal/MetalExternalImage.mm

@@ -34,10 +34,6 @@
 namespace filament {
 namespace backend {
 
-static const auto cvBufferDeleter = [](const void* buffer) {
-    CVBufferRelease((CVMetalTextureRef) buffer);
-};
-
 static const char* kernel = R"(
 #include <metal_stdlib>
 #include <simd/simd.h>
@@ -71,18 +67,30 @@ ycbcrToRgb(texture2d<half, access::read>  inYTexture    [[texture(0)]],
 }
 )";
 
-MetalExternalImage::MetalExternalImage(MetalContext& context, TextureSwizzle r, TextureSwizzle g,
-        TextureSwizzle b, TextureSwizzle a) noexcept : mContext(context), mSwizzle{r, g, b, a} { }
-
-bool MetalExternalImage::isValid() const noexcept {
-    return mRgbTexture != nil || mImage != nullptr;
+NSUInteger MetalExternalImage::getWidth() const noexcept {
+    if (mImage) {
+        return CVPixelBufferGetWidth(mImage);
+    }
+    if (mRgbTexture) {
+        return mRgbTexture.width;
+    }
+    return 0;
 }
 
-void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
-    unset();
+NSUInteger MetalExternalImage::getHeight() const noexcept {
+    if (mImage) {
+        return CVPixelBufferGetHeight(mImage);
+    }
+    if (mRgbTexture) {
+        return mRgbTexture.height;
+    }
+    return 0;
+}
 
+MetalExternalImage MetalExternalImage::createFromImage(
+        MetalContext& context, CVPixelBufferRef image) {
     if (!image) {
-        return;
+        return {};
     }
 
     OSType formatType = CVPixelBufferGetPixelFormatType(image);
@@ -96,30 +104,29 @@ void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
             << ".";
 
     if (planeCount == 0) {
-        mImage = image;
-        mTexture = createTextureFromImage(image, MTLPixelFormatBGRA8Unorm, 0);
-        mTextureView = createSwizzledTextureView(mTexture);
-        mWidth = CVPixelBufferGetWidth(image);
-        mHeight = CVPixelBufferGetHeight(image);
+        CVMetalTextureRef texture =
+                createTextureFromImage(context.textureCache, image, MTLPixelFormatBGRA8Unorm, 0);
+        return { CVPixelBufferRetain(image), texture };
     }
 
     if (planeCount == 2) {
-        CVMetalTextureRef yPlane = createTextureFromImage(image, MTLPixelFormatR8Unorm, Y_PLANE);
-        CVMetalTextureRef cbcrPlane = createTextureFromImage(image, MTLPixelFormatRG8Unorm,
-                CBCR_PLANE);
+        CVPixelBufferRetain(image);
+
+        CVMetalTextureRef yPlane =
+                createTextureFromImage(context.textureCache, image, MTLPixelFormatR8Unorm, Y_PLANE);
+        CVMetalTextureRef cbcrPlane =
+                createTextureFromImage(context.textureCache, image, MTLPixelFormatRG8Unorm, CBCR_PLANE);
 
         // Get the size of luminance plane.
-        mWidth = CVPixelBufferGetWidthOfPlane(image, Y_PLANE);
-        mHeight = CVPixelBufferGetHeightOfPlane(image, Y_PLANE);
+        NSUInteger width = CVPixelBufferGetWidthOfPlane(image, Y_PLANE);
+        NSUInteger height = CVPixelBufferGetHeightOfPlane(image, Y_PLANE);
 
-        id<MTLTexture> rgbTexture = createRgbTexture(mWidth, mHeight);
-        id<MTLCommandBuffer> commandBuffer = encodeColorConversionPass(
+        id<MTLTexture> rgbTexture = createRgbTexture(context.device, width, height);
+        id<MTLCommandBuffer> commandBuffer = encodeColorConversionPass(context,
                 CVMetalTextureGetTexture(yPlane),
                 CVMetalTextureGetTexture(cbcrPlane),
                 rgbTexture);
 
-        mRgbTexture = createSwizzledTextureView(rgbTexture);
-
         [commandBuffer addCompletedHandler:^(id <MTLCommandBuffer> o) {
             CVBufferRelease(yPlane);
             CVBufferRelease(cbcrPlane);
@@ -127,70 +134,83 @@ void MetalExternalImage::set(CVPixelBufferRef image) noexcept {
         }];
 
         [commandBuffer commit];
+        return MetalExternalImage { rgbTexture };
     }
+
+    return {};
 }
 
-void MetalExternalImage::set(CVPixelBufferRef image, size_t plane) noexcept {
-    unset();
-
+MetalExternalImage MetalExternalImage::createFromImagePlane(
+        MetalContext& context, CVPixelBufferRef image, uint32_t plane) {
     if (!image) {
-        return;
+        return {};
     }
 
     const OSType formatType = CVPixelBufferGetPixelFormatType(image);
     FILAMENT_CHECK_POSTCONDITION(formatType == kCVPixelFormatType_420YpCbCr8BiPlanarFullRange)
             << "Metal planar external images must be in the 420f format.";
+    FILAMENT_CHECK_POSTCONDITION(plane == 0 || plane == 1)
+            << "Metal planar external images must be created from planes 0 or 1.";
 
-    mImage = image;
-
-    auto getPlaneFormat = [] (size_t plane) {
-      // Right now Metal only supports kCVPixelFormatType_420YpCbCr8BiPlanarFullRange planar
-      // external images, so we can make the following assumptions about the format of each plane.
-      if (plane == 0) {
-        return MTLPixelFormatR8Unorm; // luminance
-      }
-      if (plane == 1) {
-        // CbCr
-        return MTLPixelFormatRG8Unorm; // CbCr
-      }
-      return MTLPixelFormatInvalid;
+    auto getPlaneFormat = [](size_t plane) {
+        // Right now Metal only supports kCVPixelFormatType_420YpCbCr8BiPlanarFullRange planar
+        // external images, so we can make the following assumptions about the format of each plane.
+        if (plane == 0) {
+            return MTLPixelFormatR8Unorm;  // luminance
+        }
+        if (plane == 1) {
+            return MTLPixelFormatRG8Unorm;  // CbCr
+        }
+        return MTLPixelFormatInvalid;
     };
 
     const MTLPixelFormat format = getPlaneFormat(plane);
     assert_invariant(format != MTLPixelFormatInvalid);
-    mTexture = createTextureFromImage(image, format, plane);
-    mTextureView = createSwizzledTextureView(mTexture);
+    CVMetalTextureRef mTexture = createTextureFromImage(context.textureCache, image, format, plane);
+    return { CVPixelBufferRetain(image), mTexture };
 }
 
-id<MTLTexture> MetalExternalImage::getMetalTextureForDraw() const noexcept {
+MetalExternalImage::MetalExternalImage(MetalExternalImage&& rhs) {
+    std::swap(mImage, rhs.mImage);
+    std::swap(mTexture, rhs.mTexture);
+    std::swap(mRgbTexture, rhs.mRgbTexture);
+}
+
+MetalExternalImage& MetalExternalImage::operator=(MetalExternalImage&& rhs) {
+    CVPixelBufferRelease(mImage);
+    CVBufferRelease(mTexture);
+    mImage = nullptr;
+    mTexture = nullptr;
+    mRgbTexture = nullptr;
+    std::swap(mImage, rhs.mImage);
+    std::swap(mTexture, rhs.mTexture);
+    std::swap(mRgbTexture, rhs.mRgbTexture);
+    return *this;
+}
+
+MetalExternalImage::~MetalExternalImage() noexcept {
+    CVPixelBufferRelease(mImage);
+    CVBufferRelease(mTexture);
+}
+
+id<MTLTexture> MetalExternalImage::getMtlTexture() const noexcept {
     if (mRgbTexture) {
         return mRgbTexture;
     }
-
-    // Retain the image and Metal texture until the GPU has finished with this frame. This does
-    // not need to be done for the RGB texture, because it is an Objective-C object whose
-    // lifetime is automatically managed by Metal.
-    auto& tracker = mContext.resourceTracker;
-    auto commandBuffer = getPendingCommandBuffer(&mContext);
-    if (tracker.trackResource((__bridge void*) commandBuffer, mImage, cvBufferDeleter)) {
-        CVPixelBufferRetain(mImage);
+    if (mTexture) {
+        return CVMetalTextureGetTexture(mTexture);
     }
-    if (tracker.trackResource((__bridge void*) commandBuffer, mTexture, cvBufferDeleter)) {
-        CVBufferRetain(mTexture);
-    }
-
-    assert_invariant(mTextureView);
-    return mTextureView;
+    return nil;
 }
 
-CVMetalTextureRef MetalExternalImage::createTextureFromImage(CVPixelBufferRef image,
-        MTLPixelFormat format, size_t plane) {
+CVMetalTextureRef MetalExternalImage::createTextureFromImage(CVMetalTextureCacheRef textureCache,
+        CVPixelBufferRef image, MTLPixelFormat format, size_t plane) {
     const size_t width = CVPixelBufferGetWidthOfPlane(image, plane);
     const size_t height = CVPixelBufferGetHeightOfPlane(image, plane);
 
     CVMetalTextureRef texture;
-    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault,
-            mContext.textureCache, image, nullptr, format, width, height, plane, &texture);
+    CVReturn result = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault, textureCache,
+            image, nullptr, format, width, height, plane, &texture);
     FILAMENT_CHECK_POSTCONDITION(result == kCVReturnSuccess)
             << "Could not create a CVMetalTexture from CVPixelBuffer.";
 
@@ -201,58 +221,19 @@ void MetalExternalImage::shutdown(MetalContext& context) noexcept {
     context.externalImageComputePipelineState = nil;
 }
 
-void MetalExternalImage::assertWritableImage(CVPixelBufferRef image) {
-    OSType formatType = CVPixelBufferGetPixelFormatType(image);
-    FILAMENT_CHECK_PRECONDITION(formatType == kCVPixelFormatType_32BGRA)
-            << "Metal SwapChain images must be in the 32BGRA format.";
-}
-
-void MetalExternalImage::unset() {
-    CVPixelBufferRelease(mImage);
-    CVBufferRelease(mTexture);
-
-    mImage = nullptr;
-    mTexture = nullptr;
-    mTextureView = nil;
-    mRgbTexture = nil;
-    mWidth = 0;
-    mHeight = 0;
-}
-
-id<MTLTexture> MetalExternalImage::createRgbTexture(size_t width, size_t height) {
+id<MTLTexture> MetalExternalImage::createRgbTexture(
+        id<MTLDevice> device, size_t width, size_t height) {
     MTLTextureDescriptor *descriptor =
             [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
                                                                width:width
                                                               height:height
                                                            mipmapped:NO];
     descriptor.usage = MTLTextureUsageShaderWrite | MTLTextureUsageShaderRead;
-    return [mContext.device newTextureWithDescriptor:descriptor];
+    return [device newTextureWithDescriptor:descriptor];
 }
 
-id<MTLTexture> MetalExternalImage::createSwizzledTextureView(id<MTLTexture> texture) const {
-    const bool isDefaultSwizzle =
-            mSwizzle.r == TextureSwizzle::CHANNEL_0 &&
-            mSwizzle.g == TextureSwizzle::CHANNEL_1 &&
-            mSwizzle.b == TextureSwizzle::CHANNEL_2 &&
-            mSwizzle.a == TextureSwizzle::CHANNEL_3;
-    if (!isDefaultSwizzle && mContext.supportsTextureSwizzling) {
-        // Even though we've already checked supportsTextureSwizzling, we still need to guard these
-        // calls with @availability, otherwise the API usage will generate compiler warnings.
-        if (@available(iOS 13, *)) {
-            texture = createTextureViewWithSwizzle(texture,
-                    getSwizzleChannels(mSwizzle.r, mSwizzle.g, mSwizzle.b, mSwizzle.a));
-        }
-    }
-    return texture;
-}
-
-id<MTLTexture> MetalExternalImage::createSwizzledTextureView(CVMetalTextureRef ref) const {
-    id<MTLTexture> texture = CVMetalTextureGetTexture(ref);
-    return createSwizzledTextureView(texture);
-}
-
-void MetalExternalImage::ensureComputePipelineState() {
-    if (mContext.externalImageComputePipelineState != nil) {
+void MetalExternalImage::ensureComputePipelineState(MetalContext& context) {
+    if (context.externalImageComputePipelineState != nil) {
         return;
     }
 
@@ -260,29 +241,28 @@ void MetalExternalImage::ensureComputePipelineState() {
 
     NSString* objcSource = [NSString stringWithCString:kernel
                                               encoding:NSUTF8StringEncoding];
-    id<MTLLibrary> library = [mContext.device newLibraryWithSource:objcSource
-                                                            options:nil
-                                                              error:&error];
+    id<MTLLibrary> library = [context.device newLibraryWithSource:objcSource
+                                                          options:nil
+                                                            error:&error];
     NSERROR_CHECK("Unable to compile Metal shading library.");
 
     id<MTLFunction> kernelFunction = [library newFunctionWithName:@"ycbcrToRgb"];
 
-    mContext.externalImageComputePipelineState =
-            [mContext.device newComputePipelineStateWithFunction:kernelFunction
-                                                           error:&error];
+    context.externalImageComputePipelineState =
+            [context.device newComputePipelineStateWithFunction:kernelFunction error:&error];
     NSERROR_CHECK("Unable to create Metal compute pipeline state.");
 }
 
-id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(id<MTLTexture> inYPlane,
-        id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture) {
-    ensureComputePipelineState();
+id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(MetalContext& context,
+        id<MTLTexture> inYPlane, id<MTLTexture> inCbCrTexture, id<MTLTexture> outTexture) {
+    ensureComputePipelineState(context);
 
-    id<MTLCommandBuffer> commandBuffer = [mContext.commandQueue commandBuffer];
+    id<MTLCommandBuffer> commandBuffer = [context.commandQueue commandBuffer];
     commandBuffer.label = @"YCbCr to RGB conversion";
 
     id<MTLComputeCommandEncoder> computeEncoder = [commandBuffer computeCommandEncoder];
 
-    [computeEncoder setComputePipelineState:mContext.externalImageComputePipelineState];
+    [computeEncoder setComputePipelineState:context.externalImageComputePipelineState];
     [computeEncoder setTexture:inYPlane atIndex:0];
     [computeEncoder setTexture:inCbCrTexture atIndex:1];
     [computeEncoder setTexture:outTexture atIndex:2];
@@ -300,5 +280,11 @@ id<MTLCommandBuffer> MetalExternalImage::encodeColorConversionPass(id<MTLTexture
     return commandBuffer;
 }
 
+void MetalExternalImage::assertWritableImage(CVPixelBufferRef image) {
+    OSType formatType = CVPixelBufferGetPixelFormatType(image);
+    FILAMENT_CHECK_PRECONDITION(formatType == kCVPixelFormatType_32BGRA)
+            << "Metal SwapChain images must be in the 32BGRA format.";
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalHandles.h

@@ -44,6 +44,7 @@
 #include <condition_variable>
 #include <memory>
 #include <type_traits>
+#include <vector>
 
 namespace filament {
 namespace backend {
@@ -84,6 +85,8 @@ public:
     NSUInteger getSurfaceWidth() const;
     NSUInteger getSurfaceHeight() const;
 
+    bool isPixelBuffer() const { return type == SwapChainType::CVPIXELBUFFERREF; }
+
 private:
 
     enum class SwapChainType {
@@ -93,7 +96,6 @@ private:
     };
     bool isCaMetalLayer() const { return type == SwapChainType::CAMETALLAYER; }
     bool isHeadless() const { return type == SwapChainType::HEADLESS; }
-    bool isPixelBuffer() const { return type == SwapChainType::CVPIXELBUFFERREF; }
 
     void scheduleFrameScheduledCallback();
     void scheduleFrameCompletedCallback();
@@ -138,12 +140,6 @@ public:
     void updateBufferUnsynchronized(void* data, size_t size, uint32_t byteOffset);
     MetalBuffer* getBuffer() { return &buffer; }
 
-    // Tracks which uniform/ssbo buffers this buffer object is bound into.
-    static_assert(Program::UNIFORM_BINDING_COUNT <= 32);
-    static_assert(MAX_SSBO_COUNT <= 32);
-    utils::bitset32 boundUniformBuffers;
-    utils::bitset32 boundSsbos;
-
 private:
     MetalBuffer buffer;
 };
@@ -200,12 +196,10 @@ public:
     MetalProgram(MetalContext& context, Program&& program) noexcept;
 
     const MetalShaderCompiler::MetalFunctionBundle& getFunctions();
-    const Program::SamplerGroupInfo& getSamplerGroupInfo() { return samplerGroupInfo; }
 
 private:
     void initialize();
 
-    Program::SamplerGroupInfo samplerGroupInfo;
     MetalContext& mContext;
     MetalShaderCompiler::MetalFunctionBundle mFunctionBundle;
     MetalShaderCompiler::program_token_t mToken;
@@ -227,43 +221,42 @@ struct PixelBufferShape {
 class MetalTexture : public HwTexture {
 public:
     MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
-            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
-            TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a)
-            noexcept;
+            uint8_t samples, uint32_t width, uint32_t height, uint32_t depth,
+            TextureUsage usage) noexcept;
+
+    // constructors for creating texture views
+    MetalTexture(MetalContext& context, MetalTexture const* src, uint8_t baseLevel,
+            uint8_t levelCount) noexcept;
+    MetalTexture(MetalContext& context, MetalTexture const* src, TextureSwizzle r, TextureSwizzle g,
+            TextureSwizzle b, TextureSwizzle a) noexcept;
 
     // Constructor for importing an id<MTLTexture> outside of Filament.
     MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
             uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
             id<MTLTexture> texture) noexcept;
 
-    ~MetalTexture();
+    // Constructors for importing external images.
+    MetalTexture(MetalContext& context, TextureFormat format, uint32_t width, uint32_t height,
+            TextureUsage usage, CVPixelBufferRef image) noexcept;
+    MetalTexture(MetalContext& context, TextureFormat format, uint32_t width, uint32_t height,
+            TextureUsage usage, CVPixelBufferRef image, uint32_t plane) noexcept;
 
-    // Returns an id<MTLTexture> suitable for reading in a shader, taking into account swizzle and
-    // LOD clamping.
-    id<MTLTexture> getMtlTextureForRead() noexcept;
+    // Returns an id<MTLTexture> suitable for reading in a shader, taking into account swizzle.
+    id<MTLTexture> getMtlTextureForRead() const noexcept;
 
     // Returns the id<MTLTexture> for attaching to a render pass.
-    id<MTLTexture> getMtlTextureForWrite() noexcept {
+    id<MTLTexture> getMtlTextureForWrite() const noexcept {
         return texture;
     }
 
+    std::shared_ptr<MetalExternalImage> getExternalImage() const noexcept { return externalImage; }
+
     void loadImage(uint32_t level, MTLRegion region, PixelBufferDescriptor& p) noexcept;
     void generateMipmaps() noexcept;
 
-    // A texture starts out with none of its mip levels (also referred to as LODs) available for
-    // reading. 4 actions update the range of LODs available:
-    // - calling loadImage
-    // - calling generateMipmaps
-    // - using the texture as a render target attachment
-    // - calling setMinMaxLevels
-    // A texture's available mips are consistent throughout a render pass.
-    void setLodRange(uint16_t minLevel, uint16_t maxLevel);
-    void extendLodRangeTo(uint16_t level);
-
     static MTLPixelFormat decidePixelFormat(MetalContext* context, TextureFormat format);
 
     MetalContext& context;
-    MetalExternalImage externalImage;
 
     // A "sidecar" texture used to implement automatic MSAA resolve.
     // This is created by MetalRenderTarget and stored here so it can be used with multiple
@@ -302,97 +295,16 @@ private:
 
     id<MTLTexture> texture = nil;
 
+    std::shared_ptr<MetalExternalImage> externalImage;
+
     // If non-nil, a swizzled texture view to use instead of "texture".
     // Filament swizzling only affects texture reads, so this should not be used when the texture is
     // bound as a render target attachment.
     id<MTLTexture> swizzledTextureView = nil;
-    id<MTLTexture> lodTextureView = nil;
-
-    uint16_t minLod = std::numeric_limits<uint16_t>::max();
-    uint16_t maxLod = 0;
 
     bool terminated = false;
 };
 
-class MetalSamplerGroup : public HwSamplerGroup {
-public:
-    explicit MetalSamplerGroup(size_t size, utils::FixedSizeString<32> name) noexcept
-        : size(size),
-          debugName(name),
-          textureHandles(size, Handle<HwTexture>()),
-          textures(size, nil),
-          samplers(size, nil) {}
-
-    inline void setTextureHandle(size_t index, Handle<HwTexture> th) {
-        assert_invariant(!finalized);
-        textureHandles[index] = th;
-    }
-
-    // This method is only used for debugging, to ensure all texture handles are alive.
-    const auto& getTextureHandles() const {
-        return textureHandles;
-    }
-
-    // Encode a MTLTexture into this SamplerGroup at the given index.
-    inline void setFinalizedTexture(size_t index, id<MTLTexture> t) {
-        assert_invariant(!finalized);
-        textures[index] = t;
-    }
-
-    // Encode a MTLSamplerState into this SamplerGroup at the given index.
-    inline void setFinalizedSampler(size_t index, id<MTLSamplerState> s) {
-        assert_invariant(!finalized);
-        samplers[index] = s;
-    }
-
-    // A SamplerGroup is "finalized" when all of its textures have been set and is ready for use in
-    // a draw call.
-    // Once a SamplerGroup is finalized, it must be reset or mutated to be written into again.
-    void finalize();
-    bool isFinalized() const noexcept { return finalized; }
-
-    // Both of these methods "unfinalize" a SamplerGroup, allowing it to be updated via calls to
-    // setFinalizedTexture or setFinalizedSampler. The difference is that when reset is called, all
-    // the samplers/textures must be rebound. The MTLArgumentEncoder must be specified, in case
-    // the texture types have changed.
-    // Mutate re-encodes the current set of samplers/textures into the new argument
-    // buffer.
-    void reset(id<MTLCommandBuffer> cmdBuffer, id<MTLArgumentEncoder> e, id<MTLDevice> device);
-    void mutate(id<MTLCommandBuffer> cmdBuffer);
-
-    id<MTLBuffer> getArgumentBuffer() const {
-        assert_invariant(finalized);
-        return argBuffer->getCurrentAllocation().first;
-    }
-
-    NSUInteger getArgumentBufferOffset() const {
-        return argBuffer->getCurrentAllocation().second;
-    }
-
-    inline std::pair<Handle<HwTexture>, id<MTLTexture>> getFinalizedTexture(size_t index) {
-        return {textureHandles[index], textures[index]};
-    }
-
-    // Calls the Metal useResource:usage:stages: method for all the textures in this SamplerGroup.
-    void useResources(id<MTLRenderCommandEncoder> renderPassEncoder);
-
-    size_t size;
-    utils::FixedSizeString<32> debugName;
-
-public:
-
-    // These vectors are kept in sync with one another.
-    utils::FixedCapacityVector<Handle<HwTexture>> textureHandles;
-    utils::FixedCapacityVector<id<MTLTexture>> textures;
-    utils::FixedCapacityVector<id<MTLSamplerState>> samplers;
-
-    id<MTLArgumentEncoder> encoder;
-
-    std::unique_ptr<MetalRingBuffer> argBuffer = nullptr;
-
-    bool finalized = false;
-};
-
 class MetalRenderTarget : public HwRenderTarget {
 public:
 
@@ -547,6 +459,61 @@ struct MetalTimerQuery : public HwTimerQuery {
     std::shared_ptr<Status> status;
 };
 
+class MetalDescriptorSetLayout : public HwDescriptorSetLayout {
+public:
+    MetalDescriptorSetLayout(DescriptorSetLayout&& layout) noexcept;
+
+    const auto& getBindings() const noexcept { return mLayout.bindings; }
+
+    size_t getDynamicOffsetCount() const noexcept { return mDynamicOffsetCount; }
+
+    /**
+     * Get an argument encoder for this descriptor set and shader stage.
+     * textureTypes should only include the textures present in the corresponding shader stage.
+     */
+    id<MTLArgumentEncoder> getArgumentEncoder(id<MTLDevice> device, ShaderStage stage,
+            utils::FixedCapacityVector<MTLTextureType> const& textureTypes);
+
+private:
+    id<MTLArgumentEncoder> getArgumentEncoderSlow(id<MTLDevice> device, ShaderStage stage,
+            utils::FixedCapacityVector<MTLTextureType> const& textureTypes);
+
+    DescriptorSetLayout mLayout;
+    size_t mDynamicOffsetCount = 0;
+    std::array<id<MTLArgumentEncoder>, Program::SHADER_TYPE_COUNT> mCachedArgumentEncoder = { nil };
+    std::array<utils::FixedCapacityVector<MTLTextureType>, Program::SHADER_TYPE_COUNT>
+            mCachedTextureTypes;
+};
+
+struct MetalDescriptorSet : public HwDescriptorSet {
+    MetalDescriptorSet(MetalDescriptorSetLayout* layout) noexcept;
+
+    void finalize(MetalDriver* driver);
+
+    id<MTLBuffer> finalizeAndGetBuffer(MetalDriver* driver, ShaderStage stage);
+
+    MetalDescriptorSetLayout* layout;
+
+    struct BufferBinding {
+        id<MTLBuffer> buffer;
+        uint32_t offset;
+        uint32_t size;
+    };
+    struct TextureBinding {
+        id<MTLTexture> texture;
+        SamplerParams sampler;
+    };
+    tsl::robin_map<descriptor_binding_t, BufferBinding> buffers;
+    tsl::robin_map<descriptor_binding_t, TextureBinding> textures;
+
+    std::vector<id<MTLResource>> vertexResources;
+    std::vector<id<MTLResource>> fragmentResources;
+
+    std::vector<std::shared_ptr<MetalExternalImage>> externalImages;
+
+    std::array<id<MTLBuffer>, Program::SHADER_TYPE_COUNT> cachedBuffer = { nil };
+};
+
 } // namespace backend
 } // namespace filament
 

filament/backend/src/metal/MetalHandles.mm

@@ -74,7 +74,6 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, CAMetalLayer* nativeWindow
       depthStencilFormat(decideDepthStencilFormat(flags)),
       layer(nativeWindow),
       layerDrawableMutex(std::make_shared<std::mutex>()),
-      externalImage(context),
       type(SwapChainType::CAMETALLAYER) {
 
     if (!(flags & SwapChain::CONFIG_TRANSPARENT) && !nativeWindow.opaque) {
@@ -100,17 +99,15 @@ MetalSwapChain::MetalSwapChain(MetalContext& context, int32_t width, int32_t hei
       depthStencilFormat(decideDepthStencilFormat(flags)),
       headlessWidth(width),
       headlessHeight(height),
-      externalImage(context),
       type(SwapChainType::HEADLESS) {}
 
 MetalSwapChain::MetalSwapChain(MetalContext& context, CVPixelBufferRef pixelBuffer, uint64_t flags)
     : context(context),
       depthStencilFormat(decideDepthStencilFormat(flags)),
-      externalImage(context),
+      externalImage(MetalExternalImage::createFromImage(context, pixelBuffer)),
       type(SwapChainType::CVPIXELBUFFERREF) {
     assert_invariant(flags & SWAP_CHAIN_CONFIG_APPLE_CVPIXELBUFFER);
     MetalExternalImage::assertWritableImage(pixelBuffer);
-    externalImage.set(pixelBuffer);
     assert_invariant(externalImage.isValid());
 }
 
@@ -121,7 +118,6 @@ MTLPixelFormat MetalSwapChain::decideDepthStencilFormat(uint64_t flags) {
 }
 
 MetalSwapChain::~MetalSwapChain() {
-    externalImage.set(nullptr);
 }
 
 NSUInteger MetalSwapChain::getSurfaceWidth() const {
@@ -171,7 +167,7 @@ id<MTLTexture> MetalSwapChain::acquireDrawable() {
     }
 
     if (isPixelBuffer()) {
-        return externalImage.getMetalTextureForDraw();
+        return externalImage.getMtlTexture();
     }
 
     assert_invariant(isCaMetalLayer());
@@ -257,10 +253,6 @@ void MetalSwapChain::present() {
     }
 }
 
-#ifndef FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD
-#define FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD 1
-#endif
-
 class PresentDrawableData {
 public:
     PresentDrawableData() = delete;
@@ -279,14 +271,10 @@ public:
            [that->mDrawable present];
         }
 
-#if FILAMENT_RELEASE_PRESENT_DRAWABLE_MAIN_THREAD == 1
         // mDrawable is acquired on the driver thread. Typically, we would release this object on
         // the same thread, but after receiving consistent crash reports from within
         // [CAMetalDrawable dealloc], we suspect this object requires releasing on the main thread.
         dispatch_async(dispatch_get_main_queue(), ^{ cleanupAndDestroy(that); });
-#else
-        that->mDriver->runAtNextTick([that]() { cleanupAndDestroy(that); });
-#endif
     }
 
 private:
@@ -482,11 +470,6 @@ void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
 
 MetalProgram::MetalProgram(MetalContext& context, Program&& program) noexcept
     : HwProgram(program.getName()), mContext(context) {
-
-    // Save this program's SamplerGroupInfo, it's used during draw calls to bind sampler groups to
-    // the appropriate stage(s).
-    samplerGroupInfo = program.getSamplerGroupInfo();
-
     mToken = context.shaderCompiler->createProgram(program.getName(), std::move(program));
     assert_invariant(mToken);
 }
@@ -506,10 +489,9 @@ void MetalProgram::initialize() {
 
 MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t levels,
         TextureFormat format, uint8_t samples, uint32_t width, uint32_t height, uint32_t depth,
-        TextureUsage usage, TextureSwizzle r, TextureSwizzle g, TextureSwizzle b,
-        TextureSwizzle a) noexcept
-    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
-        externalImage(context, r, g, b, a) {
+        TextureUsage usage) noexcept
+    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context) {
+    assert_invariant(target != SamplerType::SAMPLER_EXTERNAL);
 
     devicePixelFormat = decidePixelFormat(&context, format);
     FILAMENT_CHECK_POSTCONDITION(devicePixelFormat != MTLPixelFormatInvalid)
@@ -556,8 +538,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_CUBEMAP:
         case SamplerType::SAMPLER_CUBEMAP_ARRAY:
@@ -572,8 +552,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_3D:
             descriptor = [MTLTextureDescriptor new];
@@ -586,8 +564,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_EXTERNAL:
             // If we're using external textures (CVPixelBufferRefs), we don't need to make any
@@ -596,15 +572,33 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             break;
     }
 
-    // If swizzling is set, set up a swizzled texture view that we'll use when sampling this texture.
-    const bool isDefaultSwizzle =
-            r == TextureSwizzle::CHANNEL_0 &&
-            g == TextureSwizzle::CHANNEL_1 &&
-            b == TextureSwizzle::CHANNEL_2 &&
-            a == TextureSwizzle::CHANNEL_3;
-    // If texture is nil, then it must be a SAMPLER_EXTERNAL texture.
-    // Swizzling for external textures is handled inside MetalExternalImage.
-    if (!isDefaultSwizzle && texture && context.supportsTextureSwizzling) {
+    FILAMENT_CHECK_POSTCONDITION(target == SamplerType::SAMPLER_EXTERNAL || texture != nil)
+            << "Could not create Metal texture (SamplerType = " << int(target)
+            << ", levels = " << int(levels) << ", MTLPixelFormat = " << int(devicePixelFormat)
+            << ", width = " << width << ", height = " << height << ", depth = " << depth
+            << "). Out of memory?";
+}
+
+MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, uint8_t baseLevel,
+        uint8_t levelCount) noexcept
+    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
+              src->format, src->usage),
+      context(context),
+      devicePixelFormat(src->devicePixelFormat),
+      externalImage(src->externalImage) {
+    texture = createTextureViewWithLodRange(
+            src->getMtlTextureForRead(), baseLevel, baseLevel + levelCount - 1);
+}
+
+MetalTexture::MetalTexture(MetalContext& context, MetalTexture const* src, TextureSwizzle r,
+        TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) noexcept
+    : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
+              src->format, src->usage),
+      context(context),
+      devicePixelFormat(src->devicePixelFormat),
+      externalImage(src->externalImage) {
+    texture = src->getMtlTextureForRead();
+    if (context.supportsTextureSwizzling) {
         // Even though we've already checked context.supportsTextureSwizzling, we still need to
         // guard these calls with @availability, otherwise the API usage will generate compiler
         // warnings.
@@ -618,44 +612,38 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
 MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t levels, TextureFormat format,
         uint8_t samples, uint32_t width, uint32_t height, uint32_t depth, TextureUsage usage,
         id<MTLTexture> metalTexture) noexcept
-    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context),
-        externalImage(context) {
+    : HwTexture(target, levels, samples, width, height, depth, format, usage), context(context) {
     texture = metalTexture;
-    setLodRange(0, levels - 1);
+}
+
+MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t width,
+        uint32_t height, TextureUsage usage, CVPixelBufferRef image) noexcept
+    : HwTexture(SamplerType::SAMPLER_EXTERNAL, 1, 1, width, height, 1, format, usage),
+      context(context),
+      externalImage(std::make_shared<MetalExternalImage>(
+              MetalExternalImage::createFromImage(context, image))) {
+    texture = externalImage->getMtlTexture();
+}
+
+MetalTexture::MetalTexture(MetalContext& context, TextureFormat format, uint32_t width,
+        uint32_t height, TextureUsage usage, CVPixelBufferRef image, uint32_t plane) noexcept
+    : HwTexture(SamplerType::SAMPLER_EXTERNAL, 1, 1, width, height, 1, format, usage),
+      context(context),
+      externalImage(std::make_shared<MetalExternalImage>(
+              MetalExternalImage::createFromImagePlane(context, image, plane))) {
+    texture = externalImage->getMtlTexture();
 }
 
 void MetalTexture::terminate() noexcept {
     texture = nil;
     swizzledTextureView = nil;
-    lodTextureView = nil;
     msaaSidecar = nil;
-    externalImage.set(nullptr);
+    externalImage = nullptr;
     terminated = true;
 }
 
-MetalTexture::~MetalTexture() {
-    externalImage.set(nullptr);
-}
-
-id<MTLTexture> MetalTexture::getMtlTextureForRead() noexcept {
-    if (lodTextureView) {
-        return lodTextureView;
-    }
-    // The texture's swizzle remains constant throughout its lifetime, however its LOD range can
-    // change. We'll cache the LOD view, and set lodTextureView to nil if minLod or maxLod is
-    // updated.
-    id<MTLTexture> t = swizzledTextureView ? swizzledTextureView : texture;
-    if (!t) {
-        return nil;
-    }
-    if (UTILS_UNLIKELY(minLod > maxLod)) {
-        // If the texture does not have any available LODs, provide a view of only level 0.
-        // Filament should prevent this from ever occurring.
-        lodTextureView = createTextureViewWithLodRange(t, 0, 0);
-        return lodTextureView;
-    }
-    lodTextureView = createTextureViewWithLodRange(t, minLod, maxLod);
-    return lodTextureView;
+id<MTLTexture> MetalTexture::getMtlTextureForRead() const noexcept {
+    return swizzledTextureView ? swizzledTextureView : texture;
 }
 
 MTLPixelFormat MetalTexture::decidePixelFormat(MetalContext* context, TextureFormat format) {
@@ -774,15 +762,12 @@ void MetalTexture::loadImage(uint32_t level, MTLRegion region, PixelBufferDescri
             assert_invariant(false);
         }
     }
-
-    extendLodRangeTo(level);
 }
 
 void MetalTexture::generateMipmaps() noexcept {
     id <MTLBlitCommandEncoder> blitEncoder = [getPendingCommandBuffer(&context) blitCommandEncoder];
     [blitEncoder generateMipmapsForTexture:texture];
     [blitEncoder endEncoding];
-    setLodRange(0, texture.mipmapLevelCount - 1);
 }
 
 void MetalTexture::loadSlice(uint32_t level, MTLRegion region, uint32_t byteOffset, uint32_t slice,
@@ -906,98 +891,6 @@ void MetalTexture::loadWithBlit(uint32_t level, uint32_t slice, MTLRegion region
     context.blitter->blit(getPendingCommandBuffer(&context), args, "Texture upload blit");
 }
 
-void MetalTexture::extendLodRangeTo(uint16_t level) {
-    assert_invariant(!isInRenderPass(&context));
-    minLod = std::min(minLod, level);
-    maxLod = std::max(maxLod, level);
-    lodTextureView = nil;
-}
-
-void MetalTexture::setLodRange(uint16_t min, uint16_t max) {
-    assert_invariant(!isInRenderPass(&context));
-    assert_invariant(min <= max);
-    minLod = min;
-    maxLod = max;
-    lodTextureView = nil;
-}
-
-void MetalSamplerGroup::finalize() {
-    assert_invariant(encoder);
-    // TODO: we should be able to encode textures and samplers inside setFinalizedTexture and
-    // setFinalizedSampler as they become available, but Metal doesn't seem to like this; the arg
-    // buffer gets encoded incorrectly. This warrants more investigation.
-
-    auto [buffer, offset] = argBuffer->getCurrentAllocation();
-    [encoder setArgumentBuffer:buffer offset:offset];
-
-    // Encode all textures and samplers.
-    for (size_t s = 0; s < size; s++) {
-        [encoder setTexture:textures[s] atIndex:(s * 2 + 0)];
-        [encoder setSamplerState:samplers[s] atIndex:(s * 2 + 1)];
-    }
-
-    finalized = true;
-}
-
-void MetalSamplerGroup::reset(id<MTLCommandBuffer> cmdBuffer, id<MTLArgumentEncoder> e,
-        id<MTLDevice> device) {
-    encoder = e;
-
-    // The number of slots in the ring buffer we use to manage argument buffer allocations.
-    // This number was chosen to avoid running out of slots and having to allocate a "fallback"
-    // buffer when SamplerGroups are updated multiple times a frame. This value can reduced after
-    // auditing Filament's calls to updateSamplerGroup, which should be as few times as possible.
-    // For example, the bloom downsample pass should be refactored to maintain two separate
-    // MaterialInstances instead of "ping ponging" between two texture bindings, which causes a
-    // single SamplerGroup to be updated many times a frame.
-    static constexpr auto METAL_ARGUMENT_BUFFER_SLOTS = 32;
-
-    MTLSizeAndAlign argBufferLayout;
-    argBufferLayout.size = encoder.encodedLength;
-    argBufferLayout.align = encoder.alignment;
-    // Chances are, even though the MTLArgumentEncoder might change, the required size and alignment
-    // probably won't. So we can re-use the previous ring buffer.
-    if (UTILS_UNLIKELY(!argBuffer || !argBuffer->canAccomodateLayout(argBufferLayout))) {
-        argBuffer = std::make_unique<MetalRingBuffer>(device, MTLResourceStorageModeShared,
-                argBufferLayout, METAL_ARGUMENT_BUFFER_SLOTS);
-    } else {
-        argBuffer->createNewAllocation(cmdBuffer);
-    }
-
-    // Clear all textures and samplers.
-    assert_invariant(textureHandles.size() == textures.size());
-    assert_invariant(textures.size() == samplers.size());
-    for (size_t s = 0; s < textureHandles.size(); s++) {
-        textureHandles[s] = {};
-        textures[s] = nil;
-        samplers[s] = nil;
-    }
-
-    finalized = false;
-}
-
-void MetalSamplerGroup::mutate(id<MTLCommandBuffer> cmdBuffer) {
-    assert_invariant(finalized);    // only makes sense to mutate if this sampler group is finalized
-    assert_invariant(argBuffer);
-    argBuffer->createNewAllocation(cmdBuffer);
-    finalized = false;
-}
-
-void MetalSamplerGroup::useResources(id<MTLRenderCommandEncoder> renderPassEncoder) {
-    assert_invariant(finalized);
-    if (@available(iOS 13, *)) {
-        // TODO: pass only the appropriate stages to useResources.
-        [renderPassEncoder useResources:textures.data()
-                                  count:textures.size()
-                                  usage:MTLResourceUsageRead | MTLResourceUsageSample
-                                 stages:MTLRenderStageFragment | MTLRenderStageVertex];
-    } else {
-        [renderPassEncoder useResources:textures.data()
-                                  count:textures.size()
-                                  usage:MTLResourceUsageRead | MTLResourceUsageSample];
-    }
-}
-
 MetalRenderTarget::MetalRenderTarget(MetalContext* context, uint32_t width, uint32_t height,
         uint8_t samples, Attachment colorAttachments[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
         Attachment depthAttachment, Attachment stencilAttachment) :
@@ -1343,5 +1236,189 @@ FenceStatus MetalFence::wait(uint64_t timeoutNs) {
     return FenceStatus::ERROR;
 }
 
+MetalDescriptorSetLayout::MetalDescriptorSetLayout(DescriptorSetLayout&& l) noexcept
+    : mLayout(std::move(l)) {
+    size_t dynamicBindings = 0;
+    for (const auto& binding : mLayout.bindings) {
+        if (any(binding.flags & DescriptorFlags::DYNAMIC_OFFSET)) {
+            dynamicBindings++;
+        }
+    }
+    mDynamicOffsetCount = dynamicBindings;
+}
+
+id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoder(id<MTLDevice> device, ShaderStage stage,
+        utils::FixedCapacityVector<MTLTextureType> const& textureTypes) {
+    auto const index = static_cast<size_t>(stage);
+    assert_invariant(index < mCachedArgumentEncoder.size());
+    if (mCachedArgumentEncoder[index] &&
+            std::equal(
+                    textureTypes.begin(), textureTypes.end(), mCachedTextureTypes[index].begin())) {
+        return mCachedArgumentEncoder[index];
+    }
+    mCachedArgumentEncoder[index] = getArgumentEncoderSlow(device, stage, textureTypes);
+    mCachedTextureTypes[index] = textureTypes;
+    return mCachedArgumentEncoder[index];
+}
+
+id<MTLArgumentEncoder> MetalDescriptorSetLayout::getArgumentEncoderSlow(id<MTLDevice> device,
+        ShaderStage stage, utils::FixedCapacityVector<MTLTextureType> const& textureTypes) {
+    auto const& bindings = getBindings();
+    NSMutableArray<MTLArgumentDescriptor*>* arguments = [NSMutableArray new];
+    // Important! The bindings must be sorted by binding number. This has already been done inside
+    // createDescriptorSetLayout.
+    size_t textureIndex = 0;
+    for (auto const& binding : bindings) {
+        if (!hasShaderType(binding.stageFlags, stage)) {
+            continue;
+        }
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER:
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                MTLArgumentDescriptor* bufferArgument = [MTLArgumentDescriptor argumentDescriptor];
+                bufferArgument.index = binding.binding * 2;
+                bufferArgument.dataType = MTLDataTypePointer;
+                bufferArgument.access = MTLArgumentAccessReadOnly;
+                [arguments addObject:bufferArgument];
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
+                textureArgument.index = binding.binding * 2;
+                textureArgument.dataType = MTLDataTypeTexture;
+                MTLTextureType textureType = MTLTextureType2D;
+                if (textureIndex < textureTypes.size()) {
+                    textureType = textureTypes[textureIndex++];
+                }
+                textureArgument.textureType = textureType;
+                textureArgument.access = MTLArgumentAccessReadOnly;
+                [arguments addObject:textureArgument];
+
+                MTLArgumentDescriptor* samplerArgument = [MTLArgumentDescriptor argumentDescriptor];
+                samplerArgument.index = binding.binding * 2 + 1;
+                samplerArgument.dataType = MTLDataTypeSampler;
+                textureArgument.access = MTLArgumentAccessReadOnly;
+                [arguments addObject:samplerArgument];
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT:
+                // TODO: support INPUT_ATTACHMENT
+                assert_invariant(false);
+                break;
+        }
+    }
+    return [device newArgumentEncoderWithArguments:arguments];
+}
+
+MetalDescriptorSet::MetalDescriptorSet(MetalDescriptorSetLayout* layout) noexcept
+    : layout(layout) {}
+
+void MetalDescriptorSet::finalize(MetalDriver* driver) {
+    [driver->mContext->currentRenderPassEncoder useResource:driver->mContext->emptyBuffer
+                                                      usage:MTLResourceUsageRead];
+    [driver->mContext->currentRenderPassEncoder
+            useResource:getOrCreateEmptyTexture(driver->mContext)
+                  usage:MTLResourceUsageRead];
+
+    if (@available(iOS 13.0, *)) {
+        [driver->mContext->currentRenderPassEncoder useResources:vertexResources.data()
+                                                           count:vertexResources.size()
+                                                           usage:MTLResourceUsageRead
+                                                          stages:MTLRenderStageVertex];
+        [driver->mContext->currentRenderPassEncoder useResources:fragmentResources.data()
+                                                           count:fragmentResources.size()
+                                                           usage:MTLResourceUsageRead
+                                                          stages:MTLRenderStageFragment];
+    } else {
+        [driver->mContext->currentRenderPassEncoder useResources:vertexResources.data()
+                                                           count:vertexResources.size()
+                                                           usage:MTLResourceUsageRead];
+        [driver->mContext->currentRenderPassEncoder useResources:fragmentResources.data()
+                                                           count:fragmentResources.size()
+                                                           usage:MTLResourceUsageRead];
+    }
+}
+
+id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, ShaderStage stage) {
+    auto const index = static_cast<size_t>(stage);
+    assert_invariant(index < cachedBuffer.size());
+    auto& buffer = cachedBuffer[index];
+
+    if (buffer) {
+        return buffer;
+    }
+
+    // Map all the texture bindings to their respective texture types.
+    auto const& bindings = layout->getBindings();
+    auto textureTypes = utils::FixedCapacityVector<MTLTextureType>::with_capacity(bindings.size());
+    for (auto const& binding : bindings) {
+        if (!hasShaderType(binding.stageFlags, stage)) {
+            continue;
+        }
+        MTLTextureType textureType = MTLTextureType2D;
+        if (auto found = textures.find(binding.binding); found != textures.end()) {
+            auto const& textureBinding = textures[binding.binding];
+            textureType = textureBinding.texture.textureType;
+        }
+        textureTypes.push_back(textureType);
+    }
+
+    MetalContext const& context = *driver->mContext;
+
+    id<MTLArgumentEncoder> encoder =
+            layout->getArgumentEncoder(context.device, stage, textureTypes);
+
+    buffer = [context.device newBufferWithLength:encoder.encodedLength
+                                         options:MTLResourceStorageModeShared];
+    [encoder setArgumentBuffer:buffer offset:0];
+
+    for (auto const& binding : bindings) {
+        if (!hasShaderType(binding.stageFlags, stage)) {
+            continue;
+        }
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER:
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                auto found = buffers.find(binding.binding);
+                if (found == buffers.end()) {
+                    [encoder setBuffer:driver->mContext->emptyBuffer
+                                offset:0
+                               atIndex:binding.binding * 2];
+                    continue;
+                }
+
+                auto const& bufferBinding = buffers[binding.binding];
+                [encoder setBuffer:bufferBinding.buffer
+                            offset:bufferBinding.offset
+                           atIndex:binding.binding * 2];
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                auto found = textures.find(binding.binding);
+                if (found == textures.end()) {
+                    [encoder setTexture:driver->mContext->emptyTexture atIndex:binding.binding * 2];
+                    id<MTLSamplerState> sampler =
+                            driver->mContext->samplerStateCache.getOrCreateState({});
+                    [encoder setSamplerState:sampler atIndex:binding.binding * 2 + 1];
+                    continue;
+                }
+
+                auto const& textureBinding = textures[binding.binding];
+                [encoder setTexture:textureBinding.texture atIndex:binding.binding * 2];
+                SamplerState samplerState { .samplerParams = textureBinding.sampler };
+                id<MTLSamplerState> sampler =
+                        driver->mContext->samplerStateCache.getOrCreateState(samplerState);
+                [encoder setSamplerState:sampler atIndex:binding.binding * 2 + 1];
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT:
+                assert_invariant(false);
+                break;
+        }
+    }
+
+    return buffer;
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalState.h

@@ -33,32 +33,28 @@
 namespace filament {
 namespace backend {
 
-inline bool operator==(const SamplerParams& lhs, const SamplerParams& rhs) {
-    return SamplerParams::EqualTo{}(lhs, rhs);
-}
-
 //   Rasterization Bindings
 //   ----------------------
 //   Bindings    Buffer name                          Count
 //   ------------------------------------------------------
 //   0           Zero buffer (placeholder vertex buffer)  1
 //   1-16        Filament vertex buffers                 16   limited by MAX_VERTEX_BUFFER_COUNT
-//   17-25       Uniform buffers                          9   Program::UNIFORM_BINDING_COUNT
-//   26          Push constants                           1
-//   27-30       Sampler groups (argument buffers)        4   Program::SAMPLER_BINDING_COUNT
+//   20          Push constants                           1
+//   21-24       Descriptor sets (argument buffers)       4   limited by MAX_DESCRIPTOR_SET_COUNT
+//   25          Dynamic offset buffer                    1
 //
-//   Total                                               31
+//   Total                                               23
 
 //   Compute Bindings
 //   ----------------------
 //   Bindings    Buffer name                          Count
 //   ------------------------------------------------------
 //   0-3         SSBO buffers                             4   MAX_SSBO_COUNT
-//   17-25       Uniform buffers                          9   Program::UNIFORM_BINDING_COUNT
-//   26          Push constants                           1
-//   27-30       Sampler groups (argument buffers)        4   Program::SAMPLER_BINDING_COUNT
+//   20          Push constants                           1
+//   21-24       Descriptor sets (argument buffers)       4   limited by MAX_DESCRIPTOR_SET_COUNT
+//   25          Dynamic offset buffer                    1
 //
-//   Total                                               18
+//   Total                                               10
 
 // The total number of vertex buffer "slots" that the Metal backend can bind.
 // + 1 to account for the zero buffer, a placeholder buffer used internally by the Metal backend.
@@ -71,10 +67,11 @@ static constexpr uint32_t ZERO_VERTEX_BUFFER_BINDING = 0u;
 
 static constexpr uint32_t USER_VERTEX_BUFFER_BINDING_START = 1u;
 
+
 // These constants must match the equivalent in CodeGenerator.h.
-static constexpr uint32_t UNIFORM_BUFFER_BINDING_START = 17u;
-static constexpr uint32_t SSBO_BINDING_START = 0u;
-static constexpr uint32_t SAMPLER_GROUP_BINDING_START = 27u;
+static constexpr uint32_t PUSH_CONSTANT_BUFFER_INDEX = 20u;
+static constexpr uint32_t DESCRIPTOR_SET_BINDING_START = 21u;
+static constexpr uint32_t DYNAMIC_OFFSET_BINDING = 25u;
 
 // Forward declarations necessary here, definitions at end of file.
 inline bool operator==(const MTLViewport& lhs, const MTLViewport& rhs);
@@ -382,18 +379,22 @@ using SamplerStateCache = StateCache<SamplerState, id<MTLSamplerState>, SamplerS
 
 using CullModeStateTracker = StateTracker<MTLCullMode>;
 using WindingStateTracker = StateTracker<MTLWinding>;
+using DepthClampStateTracker = StateTracker<MTLDepthClipMode>;
 
 // Argument encoder
 
 struct ArgumentEncoderState {
+    NSUInteger bufferCount;
     utils::FixedCapacityVector<MTLTextureType> textureTypes;
 
-    explicit ArgumentEncoderState(utils::FixedCapacityVector<MTLTextureType>&& types)
-        : textureTypes(std::move(types)) {}
+    explicit ArgumentEncoderState(
+            NSUInteger bufferCount, utils::FixedCapacityVector<MTLTextureType>&& types)
+        : bufferCount(bufferCount), textureTypes(std::move(types)) {}
 
     bool operator==(const ArgumentEncoderState& rhs) const noexcept {
         return std::equal(textureTypes.begin(), textureTypes.end(), rhs.textureTypes.begin(),
-                rhs.textureTypes.end());
+                       rhs.textureTypes.end()) &&
+                bufferCount == rhs.bufferCount;
     }
 
     bool operator!=(const ArgumentEncoderState& rhs) const noexcept {
@@ -415,6 +416,30 @@ struct ArgumentEncoderCreator {
 using ArgumentEncoderCache = StateCache<ArgumentEncoderState, id<MTLArgumentEncoder>,
         ArgumentEncoderCreator, ArgumentEncoderHasher>;
 
+template <NSUInteger N, ShaderStage stage>
+class MetalBufferBindings {
+public:
+    MetalBufferBindings() { invalidate(); }
+
+    void invalidate() {
+        mDirtyBuffers.reset();
+        mDirtyOffsets.reset();
+        for (int i = 0; i < int(N); i++) {
+            mDirtyBuffers.set(i, true);
+            mDirtyOffsets.set(i, true);
+        }
+    }
+    void setBuffer(const id<MTLBuffer> buffer, NSUInteger offset, NSUInteger index);
+    void bindBuffers(id<MTLCommandEncoder> encoder, NSUInteger startIndex);
+
+private:
+    static_assert(N <= 8);
+    std::array<__unsafe_unretained id<MTLBuffer>, N> mBuffers = { nil };
+    std::array<NSUInteger, N> mOffsets = { 0 };
+    utils::bitset8 mDirtyBuffers;
+    utils::bitset8 mDirtyOffsets;
+};
+
 } // namespace backend
 } // namespace filament
 

filament/backend/src/metal/MetalState.mm

@@ -166,28 +166,40 @@ id<MTLSamplerState> SamplerStateCreator::operator()(id<MTLDevice> device,
 id<MTLArgumentEncoder> ArgumentEncoderCreator::operator()(id<MTLDevice> device,
         const ArgumentEncoderState &state) noexcept {
     const auto& textureTypes = state.textureTypes;
-    const auto& count = textureTypes.size();
-    assert_invariant(count > 0);
+    const auto& textureCount = textureTypes.size();
+    const auto& bufferCount = state.bufferCount;
+    assert_invariant(textureCount > 0);
 
     // Metal has separate data types for textures versus samplers, so the argument buffer layout
     // alternates between texture and sampler, i.e.:
+    // buffer0
+    // buffer1
     // textureA
     // samplerA
     // textureB
     // samplerB
     // etc
     NSMutableArray<MTLArgumentDescriptor*>* arguments =
-            [NSMutableArray arrayWithCapacity:(count * 2)];
-    for (size_t i = 0; i < count; i++) {
+            [NSMutableArray arrayWithCapacity:(bufferCount + textureCount * 2)];
+    size_t i = 0;
+    for (size_t j = 0; j < bufferCount; j++) {
+        MTLArgumentDescriptor* bufferArgument = [MTLArgumentDescriptor argumentDescriptor];
+        bufferArgument.index = i++;
+        bufferArgument.dataType = MTLDataTypePointer;
+        bufferArgument.access = MTLArgumentAccessReadOnly;
+        [arguments addObject:bufferArgument];
+    }
+
+    for (size_t j = 0; j < textureCount; j++) {
         MTLArgumentDescriptor* textureArgument = [MTLArgumentDescriptor argumentDescriptor];
-        textureArgument.index = i * 2 + 0;
+        textureArgument.index = i++;
         textureArgument.dataType = MTLDataTypeTexture;
         textureArgument.textureType = textureTypes[i];
         textureArgument.access = MTLArgumentAccessReadOnly;
         [arguments addObject:textureArgument];
 
         MTLArgumentDescriptor* samplerArgument = [MTLArgumentDescriptor argumentDescriptor];
-        samplerArgument.index = i * 2 + 1;
+        samplerArgument.index = i++;
         samplerArgument.dataType = MTLDataTypeSampler;
         textureArgument.access = MTLArgumentAccessReadOnly;
         [arguments addObject:samplerArgument];
@@ -196,5 +208,64 @@ id<MTLArgumentEncoder> ArgumentEncoderCreator::operator()(id<MTLDevice> device,
     return [device newArgumentEncoderWithArguments:arguments];
 }
 
+template <NSUInteger N, ShaderStage stage>
+void MetalBufferBindings<N, stage>::setBuffer(const id<MTLBuffer> buffer, NSUInteger offset, NSUInteger index) {
+    assert_invariant(offset + 1 <= N);
+
+    if (mBuffers[index] != buffer) {
+        mBuffers[index] = buffer;
+        mDirtyBuffers.set(index);
+    }
+
+    if (mOffsets[index] != offset) {
+        mOffsets[index] = offset;
+        mDirtyOffsets.set(index);
+    }
+}
+
+template <NSUInteger N, ShaderStage stage>
+void MetalBufferBindings<N, stage>::bindBuffers(
+        id<MTLCommandEncoder> encoder, NSUInteger startIndex) {
+    if (mDirtyBuffers.none() && mDirtyOffsets.none()) {
+        return;
+    }
+
+    utils::bitset8 onlyOffsetDirty = mDirtyOffsets & ~mDirtyBuffers;
+    onlyOffsetDirty.forEachSetBit([&](size_t i) {
+        if constexpr (stage == ShaderStage::VERTEX) {
+            [(id<MTLRenderCommandEncoder>)encoder setVertexBufferOffset:mOffsets[i]
+                                                                atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::FRAGMENT) {
+            [(id<MTLRenderCommandEncoder>)encoder setFragmentBufferOffset:mOffsets[i]
+                                                                  atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::COMPUTE) {
+            [(id<MTLComputeCommandEncoder>)encoder setBufferOffset:mOffsets[i]
+                                                           atIndex:i + startIndex];
+        }
+    });
+    mDirtyOffsets.reset();
+
+    mDirtyBuffers.forEachSetBit([&](size_t i) {
+        if constexpr (stage == ShaderStage::VERTEX) {
+            [(id<MTLRenderCommandEncoder>)encoder setVertexBuffer:mBuffers[i]
+                                                           offset:mOffsets[i]
+                                                          atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::FRAGMENT) {
+            [(id<MTLRenderCommandEncoder>)encoder setFragmentBuffer:mBuffers[i]
+                                                             offset:mOffsets[i]
+                                                            atIndex:i + startIndex];
+        } else if constexpr (stage == ShaderStage::COMPUTE) {
+            [(id<MTLComputeCommandEncoder>)encoder setBuffer:mBuffers[i]
+                                                      offset:mOffsets[i]
+                                                     atIndex:i + startIndex];
+        }
+    });
+    mDirtyBuffers.reset();
+}
+
+template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::VERTEX>;
+template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::FRAGMENT>;
+template class MetalBufferBindings<MAX_DESCRIPTOR_SET_COUNT, ShaderStage::COMPUTE>;
+
 } // namespace backend
 } // namespace filament

filament/backend/src/noop/NoopDriver.cpp

@@ -99,9 +99,6 @@ void NoopDriver::destroyProgram(Handle<HwProgram> ph) {
 void NoopDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
 }
 
-void NoopDriver::destroySamplerGroup(Handle<HwSamplerGroup> sbh) {
-}
-
 void NoopDriver::destroySwapChain(Handle<HwSwapChain> sch) {
 }
 
@@ -111,6 +108,12 @@ void NoopDriver::destroyStream(Handle<HwStream> sh) {
 void NoopDriver::destroyTimerQuery(Handle<HwTimerQuery> tqh) {
 }
 
+void NoopDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> tqh) {
+}
+
+void NoopDriver::destroyDescriptorSet(Handle<HwDescriptorSet> tqh) {
+}
+
 Handle<HwStream> NoopDriver::createStreamNative(void* nativeStream) {
     return {};
 }
@@ -202,6 +205,10 @@ bool NoopDriver::isProtectedTexturesSupported() {
     return true;
 }
 
+bool NoopDriver::isDepthClampSupported() {
+    return false;
+}
+
 bool NoopDriver::isWorkaroundNeeded(Workaround) {
     return false;
 }
@@ -244,9 +251,6 @@ void NoopDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t inde
         Handle<HwBufferObject> boh) {
 }
 
-void NoopDriver::setMinMaxLevels(Handle<HwTexture> th, uint32_t minLevel, uint32_t maxLevel) {
-}
-
 void NoopDriver::update3DImage(Handle<HwTexture> th,
         uint32_t level, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset,
         uint32_t width, uint32_t height, uint32_t depth,
@@ -272,11 +276,6 @@ void NoopDriver::setExternalStream(Handle<HwTexture> th, Handle<HwStream> sh) {
 
 void NoopDriver::generateMipmaps(Handle<HwTexture> th) { }
 
-void NoopDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
-        BufferDescriptor&& data) {
-    scheduleDestroy(std::move(data));
-}
-
 void NoopDriver::compilePrograms(CompilerPriorityQueue priority,
         CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
     if (callback) {
@@ -299,27 +298,14 @@ void NoopDriver::makeCurrent(Handle<HwSwapChain> drawSch, Handle<HwSwapChain> re
 void NoopDriver::commit(Handle<HwSwapChain> sch) {
 }
 
-void NoopDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> ubh) {
-}
-
-void NoopDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t index,
-        Handle<HwBufferObject> ubh, uint32_t offset, uint32_t size) {
-}
-
-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) {
+void NoopDriver::insertEventMarker(char const* string) {
 }
 
-void NoopDriver::pushGroupMarker(char const* string,  uint32_t len) {
+void NoopDriver::pushGroupMarker(char const* string) {
 }
 
 void NoopDriver::popGroupMarker(int) {
@@ -388,4 +374,28 @@ void NoopDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
 void NoopDriver::resetState(int) {
 }
 
+void NoopDriver::updateDescriptorSetBuffer(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_binding_t binding,
+        backend::BufferObjectHandle boh,
+        uint32_t offset,
+        uint32_t size) {
+}
+
+void NoopDriver::updateDescriptorSetTexture(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_binding_t binding,
+        backend::TextureHandle th,
+        SamplerParams params) {
+}
+
+void NoopDriver::bindDescriptorSet(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_set_t set,
+        backend::DescriptorSetOffsetArray&& offsets) {
+}
+
+void NoopDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
+}
+
 } // namespace filament

filament/backend/src/opengl/BindingMap.h

@@ -0,0 +1,89 @@
+/*
+ * 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_BINDINGMAP_H
+#define TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H
+
+#include <backend/DriverEnums.h>
+
+#include "gl_headers.h"
+
+#include <utils/bitset.h>
+#include <utils/debug.h>
+
+#include <new>
+
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+namespace filament::backend {
+
+class BindingMap {
+    struct CompressedBinding {
+        // this is in fact a GLuint, but we only want 8-bits
+        uint8_t binding : 7;
+        uint8_t sampler : 1;
+    };
+
+    CompressedBinding (*mStorage)[MAX_DESCRIPTOR_COUNT];
+
+    utils::bitset64 mActiveDescriptors[MAX_DESCRIPTOR_SET_COUNT];
+
+public:
+    BindingMap() noexcept
+            : mStorage(new (std::nothrow) CompressedBinding[MAX_DESCRIPTOR_SET_COUNT][MAX_DESCRIPTOR_COUNT]) {
+#ifndef NDEBUG
+        memset(mStorage, 0xFF, sizeof(CompressedBinding[MAX_DESCRIPTOR_SET_COUNT][MAX_DESCRIPTOR_COUNT]));
+#endif
+    }
+
+    ~BindingMap() noexcept {
+        delete [] mStorage;
+    }
+
+    BindingMap(BindingMap const&) noexcept = delete;
+    BindingMap(BindingMap&&) noexcept = delete;
+    BindingMap& operator=(BindingMap const&) noexcept = delete;
+    BindingMap& operator=(BindingMap&&) noexcept = delete;
+
+    struct Binding {
+        GLuint binding;
+        DescriptorType type;
+    };
+
+    void insert(descriptor_set_t set, descriptor_binding_t binding, Binding entry) noexcept {
+        assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
+        assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
+        assert_invariant(entry.binding < 128); // we reserve 1 bit for the type right now
+        mStorage[set][binding] = { (uint8_t)entry.binding, entry.type == DescriptorType::SAMPLER };
+        mActiveDescriptors[set].set(binding);
+    }
+
+    GLuint get(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
+        assert_invariant(set < MAX_DESCRIPTOR_SET_COUNT);
+        assert_invariant(binding < MAX_DESCRIPTOR_COUNT);
+        return mStorage[set][binding].binding;
+    }
+
+    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
+        return mActiveDescriptors[set];
+    }
+};
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_OPENGL_BINDINGMAP_H

filament/backend/src/opengl/GLDescriptorSet.cpp

@@ -0,0 +1,361 @@
+/*
+ * 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 "GLDescriptorSet.h"
+
+#include "GLBufferObject.h"
+#include "GLDescriptorSetLayout.h"
+#include "GLTexture.h"
+#include "GLUtils.h"
+#include "OpenGLDriver.h"
+#include "OpenGLContext.h"
+#include "OpenGLProgram.h"
+
+#include "gl_headers.h"
+
+#include <private/backend/HandleAllocator.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+
+#include <utils/BitmaskEnum.h>
+#include <utils/Log.h>
+#include <utils/Panic.h>
+#include <utils/bitset.h>
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <algorithm>
+
+#include <type_traits>
+#include <utility>
+#include <variant>
+
+#include <stddef.h>
+#include <stdint.h>
+
+namespace filament::backend {
+
+GLDescriptorSet::GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle dslh,
+        GLDescriptorSetLayout const* layout) noexcept
+        : descriptors(layout->maxDescriptorBinding + 1),
+          dslh(std::move(dslh)) {
+
+    // We have allocated enough storage for all descriptors. Now allocate the empty descriptor
+    // themselves.
+    for (auto const& entry : layout->bindings) {
+        size_t const index = entry.binding;
+
+        // now we'll initialize the alternative for each way we can handle this descriptor.
+        auto& desc = descriptors[index].desc;
+        switch (entry.type) {
+            case DescriptorType::UNIFORM_BUFFER: {
+                // A uniform buffer can have dynamic offsets or not and have special handling for
+                // ES2 (where we need to emulate it). That's four alternatives.
+                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
+                dynamicBuffers.set(index, dynamicOffset);
+                if (UTILS_UNLIKELY(gl.isES2())) {
+                    dynamicBufferCount++;
+                    desc.emplace<BufferGLES2>(dynamicOffset);
+                } else {
+                    auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::UNIFORM);
+                    if (dynamicOffset) {
+                        dynamicBufferCount++;
+                        desc.emplace<DynamicBuffer>(type);
+                    } else {
+                        desc.emplace<Buffer>(type);
+                    }
+                }
+                break;
+            }
+            case DescriptorType::SHADER_STORAGE_BUFFER: {
+                // shader storage buffers are not supported on ES2, So that's two alternatives.
+                bool const dynamicOffset = any(entry.flags & DescriptorFlags::DYNAMIC_OFFSET);
+                dynamicBuffers.set(index, dynamicOffset);
+                auto const type = GLUtils::getBufferBindingType(BufferObjectBinding::SHADER_STORAGE);
+                if (dynamicOffset) {
+                    dynamicBufferCount++;
+                    desc.emplace<DynamicBuffer>(type);
+                } else {
+                    desc.emplace<Buffer>(type);
+                }
+                break;
+            }
+            case DescriptorType::SAMPLER:
+                if (UTILS_UNLIKELY(gl.isES2())) {
+                    desc.emplace<SamplerGLES2>();
+                } else {
+                    const bool anisotropyWorkaround =
+                            gl.ext.EXT_texture_filter_anisotropic &&
+                            gl.bugs.texture_filter_anisotropic_broken_on_sampler;
+                    if (anisotropyWorkaround) {
+                        desc.emplace<SamplerWithAnisotropyWorkaround>();
+                    } else {
+                        desc.emplace<Sampler>();
+                    }
+                }
+                break;
+            case DescriptorType::INPUT_ATTACHMENT:
+                break;
+        }
+    }
+}
+
+void GLDescriptorSet::update(OpenGLContext&,
+        descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept {
+    assert_invariant(binding < descriptors.size());
+    std::visit([=](auto&& arg) {
+        using T = std::decay_t<decltype(arg)>;
+        if constexpr (std::is_same_v<T, Buffer> || std::is_same_v<T, DynamicBuffer>) {
+            assert_invariant(arg.target != 0);
+            arg.id = bo ? bo->gl.id : 0;
+            arg.offset = uint32_t(offset);
+            arg.size = uint32_t(size);
+            assert_invariant(arg.id || (!arg.size && !offset));
+        } else if constexpr (std::is_same_v<T, BufferGLES2>) {
+            arg.bo = bo;
+            arg.offset = uint32_t(offset);
+        } else {
+            // API usage error. User asked to update the wrong type of descriptor.
+            PANIC_PRECONDITION("descriptor %d is not a buffer", +binding);
+        }
+    }, descriptors[binding].desc);
+}
+
+void GLDescriptorSet::update(OpenGLContext& gl,
+        descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept {
+    assert_invariant(binding < descriptors.size());
+    std::visit([=, &gl](auto&& arg) mutable {
+        using T = std::decay_t<decltype(arg)>;
+        if constexpr (std::is_same_v<T, Sampler> ||
+                      std::is_same_v<T, SamplerWithAnisotropyWorkaround> ||
+                      std::is_same_v<T, SamplerGLES2>) {
+            if (UTILS_UNLIKELY(t && t->target == SamplerType::SAMPLER_EXTERNAL)) {
+                // From OES_EGL_image_external spec:
+                // "The default s and t wrap modes are CLAMP_TO_EDGE, and it is an INVALID_ENUM
+                //  error to set the wrap mode to any other value."
+                params.wrapS = SamplerWrapMode::CLAMP_TO_EDGE;
+                params.wrapT = SamplerWrapMode::CLAMP_TO_EDGE;
+                params.wrapR = SamplerWrapMode::CLAMP_TO_EDGE;
+            }
+            // GLES3.x specification forbids depth textures to be filtered.
+            if (t && isDepthFormat(t->format)
+                    && params.compareMode == SamplerCompareMode::NONE) {
+                params.filterMag = SamplerMagFilter::NEAREST;
+                switch (params.filterMin) {
+                    case SamplerMinFilter::LINEAR:
+                        params.filterMin = SamplerMinFilter::NEAREST;
+                        break;
+                    case SamplerMinFilter::LINEAR_MIPMAP_NEAREST:
+                    case SamplerMinFilter::NEAREST_MIPMAP_LINEAR:
+                    case SamplerMinFilter::LINEAR_MIPMAP_LINEAR:
+                        params.filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST;
+                        break;
+                    default:
+                        break;
+                }
+            }
+
+            arg.target = t ? t->gl.target : 0;
+            arg.id = t ? t->gl.id : 0;
+            if constexpr (std::is_same_v<T, Sampler> ||
+                          std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
+                if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
+                    arg.anisotropy = float(1u << params.anisotropyLog2);
+                }
+                if (t) {
+                    arg.ref = t->ref;
+                    arg.baseLevel = t->gl.baseLevel;
+                    arg.maxLevel = t->gl.maxLevel;
+                    arg.swizzle = t->gl.swizzle;
+                }
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+                arg.sampler = gl.getSampler(params);
+#else
+                (void)gl;
+#endif
+            } else {
+                arg.params = params;
+            }
+        } else {
+            // API usage error. User asked to update the wrong type of descriptor.
+            PANIC_PRECONDITION("descriptor %d is not a texture", +binding);
+        }
+    }, descriptors[binding].desc);
+}
+
+template<typename T>
+void GLDescriptorSet::updateTextureView(OpenGLContext& gl,
+        HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept {
+    // The common case is that we don't have a ref handle (we only have one if
+    // the texture ever had a View on it).
+    assert_invariant(desc.ref);
+    GLTextureRef* const ref = handleAllocator.handle_cast<GLTextureRef*>(desc.ref);
+    if (UTILS_UNLIKELY((desc.baseLevel != ref->baseLevel || desc.maxLevel != ref->maxLevel))) {
+        // If we have views, then it's still uncommon that we'll switch often
+        // handle the case where we reset to the original texture
+        GLint baseLevel = GLint(desc.baseLevel); // NOLINT(*-signed-char-misuse)
+        GLint maxLevel = GLint(desc.maxLevel); // NOLINT(*-signed-char-misuse)
+        if (baseLevel > maxLevel) {
+            baseLevel = 0;
+            maxLevel = 1000; // per OpenGL spec
+        }
+        // that is very unfortunate that we have to call activeTexture here
+        gl.activeTexture(unit);
+        glTexParameteri(desc.target, GL_TEXTURE_BASE_LEVEL, baseLevel);
+        glTexParameteri(desc.target, GL_TEXTURE_MAX_LEVEL,  maxLevel);
+        ref->baseLevel = desc.baseLevel;
+        ref->maxLevel = desc.maxLevel;
+    }
+    if (UTILS_UNLIKELY(desc.swizzle != ref->swizzle)) {
+        using namespace GLUtils;
+        gl.activeTexture(unit);
+#if !defined(__EMSCRIPTEN__)  && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_R, (GLint)getSwizzleChannel(desc.swizzle[0]));
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_G, (GLint)getSwizzleChannel(desc.swizzle[1]));
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_B, (GLint)getSwizzleChannel(desc.swizzle[2]));
+        glTexParameteri(desc.target, GL_TEXTURE_SWIZZLE_A, (GLint)getSwizzleChannel(desc.swizzle[3]));
+#endif
+        ref->swizzle = desc.swizzle;
+    }
+}
+
+void GLDescriptorSet::bind(
+        OpenGLContext& gl,
+        HandleAllocatorGL& handleAllocator,
+        OpenGLProgram const& p,
+        descriptor_set_t set, uint32_t const* offsets, bool offsetsOnly) const noexcept {
+    // TODO: check that offsets is sized correctly
+    size_t dynamicOffsetIndex = 0;
+
+    utils::bitset64 activeDescriptorBindings = p.getActiveDescriptors(set);
+    if (offsetsOnly) {
+        activeDescriptorBindings &= dynamicBuffers;
+    }
+
+    // loop only over the active indices for this program
+    activeDescriptorBindings.forEachSetBit(
+            [this,&gl, &handleAllocator, &p, set, offsets, &dynamicOffsetIndex]
+            (size_t binding) {
+
+        // This would fail here if we're trying to set a descriptor that doesn't exist in the
+        // program. In other words, a mismatch between the program's layout and this descriptor-set.
+        assert_invariant(binding < descriptors.size());
+
+        auto const& entry = descriptors[binding];
+        std::visit(
+                [&gl, &handleAllocator, &p, &dynamicOffsetIndex, set, binding, offsets]
+                (auto&& arg) {
+            using T = std::decay_t<decltype(arg)>;
+            if constexpr (std::is_same_v<T, Buffer>) {
+                GLuint const bindingPoint = p.getBufferBinding(set, binding);
+                GLintptr const offset = arg.offset;
+                assert_invariant(arg.id || (!arg.size && !offset));
+                gl.bindBufferRange(arg.target, bindingPoint, arg.id, offset, arg.size);
+            } else if constexpr (std::is_same_v<T, DynamicBuffer>) {
+                GLuint const bindingPoint = p.getBufferBinding(set, binding);
+                GLintptr const offset = arg.offset + offsets[dynamicOffsetIndex++];
+                assert_invariant(arg.id || (!arg.size && !offset));
+                gl.bindBufferRange(arg.target, bindingPoint, arg.id, offset, arg.size);
+            } else if constexpr (std::is_same_v<T, BufferGLES2>) {
+                GLuint const bindingPoint = p.getBufferBinding(set, binding);
+                GLintptr offset = arg.offset;
+                if (arg.dynamicOffset) {
+                    offset += offsets[dynamicOffsetIndex++];
+                }
+                if (arg.bo) {
+                    auto buffer = static_cast<char const*>(arg.bo->gl.buffer) + offset;
+                    p.updateUniforms(bindingPoint, arg.bo->gl.id, buffer, arg.bo->age);
+                }
+            } else if constexpr (std::is_same_v<T, Sampler>) {
+                GLuint const unit = p.getTextureUnit(set, binding);
+                if (arg.target) {
+                    gl.bindTexture(unit, arg.target, arg.id);
+                    gl.bindSampler(unit, arg.sampler);
+                    if (UTILS_UNLIKELY(arg.ref)) {
+                        updateTextureView(gl, handleAllocator, unit, arg);
+                    }
+                } else {
+                    gl.unbindTextureUnit(unit);
+                }
+            } else if constexpr (std::is_same_v<T, SamplerWithAnisotropyWorkaround>) {
+                GLuint const unit = p.getTextureUnit(set, binding);
+                if (arg.target) {
+                    gl.bindTexture(unit, arg.target, arg.id);
+                    gl.bindSampler(unit, arg.sampler);
+                    if (UTILS_UNLIKELY(arg.ref)) {
+                        updateTextureView(gl, handleAllocator, unit, arg);
+                    }
+#if defined(GL_EXT_texture_filter_anisotropic)
+                    // Driver claims to support anisotropic filtering, but it fails when set on
+                    // the sampler, we have to set it on the texture instead.
+                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                            std::min(gl.gets.max_anisotropy, float(arg.anisotropy)));
+#endif
+                } else {
+                    gl.unbindTextureUnit(unit);
+                }
+            } else if constexpr (std::is_same_v<T, SamplerGLES2>) {
+                // in ES2 the sampler parameters need to be set on the texture itself
+                GLuint const unit = p.getTextureUnit(set, binding);
+                if (arg.target) {
+                    gl.bindTexture(unit, arg.target, arg.id);
+                    SamplerParams const params = arg.params;
+                    glTexParameteri(arg.target, GL_TEXTURE_MIN_FILTER,
+                            (GLint)GLUtils::getTextureFilter(params.filterMin));
+                    glTexParameteri(arg.target, GL_TEXTURE_MAG_FILTER,
+                            (GLint)GLUtils::getTextureFilter(params.filterMag));
+                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_S,
+                            (GLint)GLUtils::getWrapMode(params.wrapS));
+                    glTexParameteri(arg.target, GL_TEXTURE_WRAP_T,
+                            (GLint)GLUtils::getWrapMode(params.wrapT));
+#if defined(GL_EXT_texture_filter_anisotropic)
+                    glTexParameterf(arg.target, GL_TEXTURE_MAX_ANISOTROPY_EXT,
+                            std::min(gl.gets.max_anisotropy, arg.anisotropy));
+#endif
+                } else {
+                    gl.unbindTextureUnit(unit);
+                }
+            }
+        }, entry.desc);
+    });
+    CHECK_GL_ERROR(utils::slog.e)
+}
+
+void GLDescriptorSet::validate(HandleAllocatorGL& allocator,
+        DescriptorSetLayoutHandle pipelineLayout) const {
+
+    if (UTILS_UNLIKELY(dslh != pipelineLayout)) {
+        auto* const dsl = allocator.handle_cast < GLDescriptorSetLayout const * > (dslh);
+        auto* const cur = allocator.handle_cast < GLDescriptorSetLayout const * > (pipelineLayout);
+
+        UTILS_UNUSED_IN_RELEASE
+        bool const pipelineLayoutMatchesDescriptorSetLayout = std::equal(
+                dsl->bindings.begin(), dsl->bindings.end(),
+                cur->bindings.begin(),
+                [](DescriptorSetLayoutBinding const& lhs,
+                        DescriptorSetLayoutBinding const& rhs) {
+                    return lhs.type == rhs.type &&
+                           lhs.stageFlags == rhs.stageFlags &&
+                           lhs.binding == rhs.binding &&
+                           lhs.flags == rhs.flags &&
+                           lhs.count == rhs.count;
+                });
+
+        assert_invariant(pipelineLayoutMatchesDescriptorSetLayout);
+    }
+}
+
+} // namespace filament::backend

filament/backend/src/opengl/GLDescriptorSet.h

@@ -0,0 +1,171 @@
+/*
+ * 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_GLDESCRIPTORSET_H
+#define TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H
+
+#include "DriverBase.h"
+
+#include "gl_headers.h"
+
+#include <private/backend/HandleAllocator.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+
+#include <utils/bitset.h>
+#include <utils/FixedCapacityVector.h>
+
+#include <math/half.h>
+
+#include <array>
+#include <variant>
+
+#include <stddef.h>
+#include <stdint.h>
+
+namespace filament::backend {
+
+struct GLBufferObject;
+struct GLTexture;
+struct GLTextureRef;
+struct GLDescriptorSetLayout;
+class OpenGLProgram;
+class OpenGLContext;
+class OpenGLDriver;
+
+struct GLDescriptorSet : public HwDescriptorSet {
+
+    using HwDescriptorSet::HwDescriptorSet;
+
+    GLDescriptorSet(OpenGLContext& gl, DescriptorSetLayoutHandle dslh,
+            GLDescriptorSetLayout const* layout) noexcept;
+
+    // update a buffer descriptor in the set
+    void update(OpenGLContext& gl,
+            descriptor_binding_t binding, GLBufferObject* bo, size_t offset, size_t size) noexcept;
+
+    // update a sampler descriptor in the set
+    void update(OpenGLContext& gl,
+            descriptor_binding_t binding, GLTexture* t, SamplerParams params) noexcept;
+
+    // conceptually bind the set to the command buffer
+    void bind(
+            OpenGLContext& gl,
+            HandleAllocatorGL& handleAllocator,
+            OpenGLProgram const& p,
+            descriptor_set_t set, uint32_t const* offsets, bool offsetsOnly) const noexcept;
+
+    uint32_t getDynamicBufferCount() const noexcept {
+        return dynamicBufferCount;
+    }
+
+    void validate(HandleAllocatorGL& allocator, DescriptorSetLayoutHandle pipelineLayout) const;
+
+private:
+    // a Buffer Descriptor such as SSBO or UBO with static offset
+    struct Buffer {
+        Buffer() = default;
+        explicit Buffer(GLenum target) noexcept : target(target) { }
+        GLenum target;                          // 4
+        GLuint id = 0;                          // 4
+        uint32_t offset = 0;                    // 4
+        uint32_t size = 0;                      // 4
+    };
+
+    // a Buffer Descriptor such as SSBO or UBO with dynamic offset
+    struct DynamicBuffer {
+        DynamicBuffer() = default;
+        explicit DynamicBuffer(GLenum target) noexcept : target(target) { }
+        GLenum target;                          // 4
+        GLuint id = 0;                          // 4
+        uint32_t offset = 0;                    // 4
+        uint32_t size = 0;                      // 4
+    };
+
+    // a UBO descriptor for ES2
+    struct BufferGLES2 {
+        BufferGLES2() = default;
+        explicit BufferGLES2(bool dynamicOffset) noexcept : dynamicOffset(dynamicOffset) { }
+        GLBufferObject const* bo = nullptr;     // 8
+        uint32_t offset = 0;                    // 4
+        bool dynamicOffset = false;             // 4
+    };
+
+    // A sampler descriptor
+    struct Sampler {
+        GLenum target = 0;                      // 4
+        GLuint id = 0;                          // 4
+        GLuint sampler = 0;                     // 4
+        Handle<GLTextureRef> ref;               // 4
+        int8_t baseLevel = 0x7f;                // 1
+        int8_t maxLevel = -1;                   // 1
+        std::array<TextureSwizzle, 4> swizzle{  // 4
+                TextureSwizzle::CHANNEL_0,
+                TextureSwizzle::CHANNEL_1,
+                TextureSwizzle::CHANNEL_2,
+                TextureSwizzle::CHANNEL_3
+        };
+    };
+
+    struct SamplerWithAnisotropyWorkaround {
+        GLenum target = 0;                      // 4
+        GLuint id = 0;                          // 4
+        GLuint sampler = 0;                     // 4
+        Handle<GLTextureRef> ref;               // 4
+        math::half anisotropy = 1.0f;           // 2
+        int8_t baseLevel = 0x7f;                // 1
+        int8_t maxLevel = -1;                   // 1
+        std::array<TextureSwizzle, 4> swizzle{  // 4
+                TextureSwizzle::CHANNEL_0,
+                TextureSwizzle::CHANNEL_1,
+                TextureSwizzle::CHANNEL_2,
+                TextureSwizzle::CHANNEL_3
+        };
+    };
+
+    // A sampler descriptor for ES2
+    struct SamplerGLES2 {
+        GLenum target = 0;                      // 4
+        GLuint id = 0;                          // 4
+        SamplerParams params{};                 // 4
+        float anisotropy = 1.0f;                // 4
+    };
+    struct Descriptor {
+        std::variant<
+                Buffer,
+                DynamicBuffer,
+                BufferGLES2,
+                Sampler,
+                SamplerWithAnisotropyWorkaround,
+                SamplerGLES2> desc;
+    };
+    static_assert(sizeof(Descriptor) <= 32);
+
+    template<typename T>
+    static void updateTextureView(OpenGLContext& gl,
+            HandleAllocatorGL& handleAllocator, GLuint unit, T const& desc) noexcept;
+
+    utils::FixedCapacityVector<Descriptor> descriptors;     // 16
+    utils::bitset64 dynamicBuffers;                         // 8
+    DescriptorSetLayoutHandle dslh;                         // 4
+    uint8_t dynamicBufferCount = 0;                         // 1
+};
+static_assert(sizeof(GLDescriptorSet) <= 32);
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSET_H

filament/backend/src/opengl/GLDescriptorSetLayout.h

@@ -0,0 +1,52 @@
+/*
+ * 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_GLDESCRIPTORSETLAYOUT_H
+#define TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H
+
+#include "DriverBase.h"
+
+#include <backend/DriverEnums.h>
+
+#include <algorithm>
+#include <utility>
+
+#include <stdint.h>
+
+namespace filament::backend {
+
+struct GLDescriptorSetLayout : public HwDescriptorSetLayout, public DescriptorSetLayout {
+    using HwDescriptorSetLayout::HwDescriptorSetLayout;
+    explicit GLDescriptorSetLayout(DescriptorSetLayout&& layout) noexcept
+            : DescriptorSetLayout(std::move(layout)) {
+
+        std::sort(bindings.begin(), bindings.end(),
+                [](auto&& lhs, auto&& rhs){
+            return lhs.binding < rhs.binding;
+        });
+
+        auto p = std::max_element(bindings.cbegin(), bindings.cend(),
+                [](auto const& lhs, auto const& rhs) {
+            return lhs.binding < rhs.binding;
+        });
+        maxDescriptorBinding = p->binding;
+    }
+    uint8_t maxDescriptorBinding = 0;
+};
+
+} // namespace filament::backend
+
+#endif //TNT_FILAMENT_BACKEND_OPENGL_GLDESCRIPTORSETLAYOUT_H

filament/backend/src/opengl/GLTexture.h

@@ -21,12 +21,32 @@
 
 #include "gl_headers.h"
 
+#include <backend/Handle.h>
+#include <backend/DriverEnums.h>
 #include <backend/platforms/OpenGLPlatform.h>
 
+#include <array>
+
 #include <stdint.h>
 
 namespace filament::backend {
 
+struct GLTextureRef {
+    GLTextureRef() = default;
+    // view reference counter
+    uint16_t count = 1;
+    // current per-view values of the texture (in GL we can only have a single View active at
+    // a time, and this tracks that state). It's used to avoid unnecessarily change state.
+    int8_t baseLevel = 127;
+    int8_t maxLevel = -1;
+    std::array<TextureSwizzle, 4> swizzle{
+            TextureSwizzle::CHANNEL_0,
+            TextureSwizzle::CHANNEL_1,
+            TextureSwizzle::CHANNEL_2,
+            TextureSwizzle::CHANNEL_3
+    };
+};
+
 struct GLTexture : public HwTexture {
     using HwTexture::HwTexture;
     struct GL {
@@ -44,8 +64,14 @@ struct GLTexture : public HwTexture {
         bool imported           : 1;
         uint8_t sidecarSamples  : 4;
         uint8_t reserved1       : 3;
+        std::array<TextureSwizzle, 4> swizzle{
+                TextureSwizzle::CHANNEL_0,
+                TextureSwizzle::CHANNEL_1,
+                TextureSwizzle::CHANNEL_2,
+                TextureSwizzle::CHANNEL_3
+        };
     } gl;
-
+    mutable Handle<GLTextureRef> ref;
     OpenGLPlatform::ExternalTexture* externalTexture = nullptr;
 };
 

filament/backend/src/opengl/OpenGLContext.cpp

@@ -676,6 +676,7 @@ void OpenGLContext::initExtensionsGLES(Extensions* ext, GLint major, GLint minor
 #ifndef __EMSCRIPTEN__
     ext->EXT_debug_marker = exts.has("GL_EXT_debug_marker"sv);
 #endif
+    ext->EXT_depth_clamp = exts.has("GL_EXT_depth_clamp"sv);
     ext->EXT_discard_framebuffer = exts.has("GL_EXT_discard_framebuffer"sv);
 #ifndef __EMSCRIPTEN__
     ext->EXT_disjoint_timer_query = exts.has("GL_EXT_disjoint_timer_query"sv);
@@ -746,6 +747,7 @@ void OpenGLContext::initExtensionsGL(Extensions* ext, GLint major, GLint minor)
     ext->EXT_color_buffer_half_float = true;  // Assumes core profile.
     ext->EXT_clip_cull_distance = true;
     ext->EXT_debug_marker = exts.has("GL_EXT_debug_marker"sv);
+    ext->EXT_depth_clamp = true;
     ext->EXT_discard_framebuffer = false;
     ext->EXT_disjoint_timer_query = true;
     ext->EXT_multisampled_render_to_texture = false;

filament/backend/src/opengl/OpenGLContext.h

@@ -60,10 +60,19 @@ public:
     struct RenderPrimitive {
         static_assert(MAX_VERTEX_ATTRIBUTE_COUNT <= 16);
 
-        GLuint vao[2] = {};                                     // 4
+        GLuint vao[2] = {};                                     // 8
         GLuint elementArray = 0;                                // 4
+        GLenum indicesType = 0;                                 // 4
+
+        // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
+        // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
+        // immutable.
+        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4
+
         mutable utils::bitset<uint16_t> vertexAttribArray;      // 2
 
+        uint8_t reserved[2] = {};                               // 2
+
         // if this differs from vertexBufferWithObjects->bufferObjectsVersion, this VAO needs to
         // be updated (see OpenGLDriver::updateVertexArrayObject())
         uint8_t vertexBufferVersion = 0;                        // 1
@@ -76,16 +85,11 @@ public:
         // See OpenGLContext::bindVertexArray()
         uint8_t nameVersion = 0;                                // 1
 
-        // Size in bytes of indices in the index buffer
-        uint8_t indicesSize = 0;                                // 1
-
-        // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
-        // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
-        // immutable.
-        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4
+        // Size in bytes of indices in the index buffer (1 or 2)
+        uint8_t indicesShift = 0;                                // 1
 
         GLenum getIndicesType() const noexcept {
-            return indicesSize == 4 ? GL_UNSIGNED_INT : GL_UNSIGNED_SHORT;
+            return indicesType;
         }
     } gl;
 
@@ -220,8 +224,9 @@ public:
         bool EXT_color_buffer_float;
         bool EXT_color_buffer_half_float;
         bool EXT_debug_marker;
-        bool EXT_disjoint_timer_query;
+        bool EXT_depth_clamp;
         bool EXT_discard_framebuffer;
+        bool EXT_disjoint_timer_query;
         bool EXT_multisampled_render_to_texture2;
         bool EXT_multisampled_render_to_texture;
         bool EXT_protected_textures;
@@ -239,10 +244,10 @@ public:
         bool KHR_parallel_shader_compile;
         bool KHR_texture_compression_astc_hdr;
         bool KHR_texture_compression_astc_ldr;
-        bool OES_depth_texture;
-        bool OES_depth24;
-        bool OES_packed_depth_stencil;
         bool OES_EGL_image_external_essl3;
+        bool OES_depth24;
+        bool OES_depth_texture;
+        bool OES_packed_depth_stencil;
         bool OES_rgb8_rgba8;
         bool OES_standard_derivatives;
         bool OES_texture_npot;
@@ -473,12 +478,6 @@ public:
 
     void unbindEverything() noexcept;
     void synchronizeStateAndCache(size_t index) noexcept;
-    void setEs2UniformBinding(size_t index, GLuint id, void const* data, uint16_t age) noexcept {
-        mUniformBindings[index] = { id, data, age };
-    }
-    auto getEs2UniformBinding(size_t index) const noexcept {
-        return mUniformBindings[index];
-    }
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     GLuint getSamplerSlow(SamplerParams sp) const noexcept;
@@ -505,9 +504,6 @@ private:
     std::vector<std::function<void(OpenGLContext&)>> mDestroyWithNormalContext;
     RenderPrimitive mDefaultVAO;
     std::optional<GLuint> mDefaultFbo[2];
-    std::array<
-            std::tuple<GLuint, void const*, uint16_t>,
-            CONFIG_UNIFORM_BINDING_COUNT> mUniformBindings = {};
     mutable tsl::robin_map<SamplerParams, GLuint,
             SamplerParams::Hasher, SamplerParams::EqualTo> mSamplerMap;
 
@@ -627,6 +623,7 @@ constexpr size_t OpenGLContext::getIndexForCap(GLenum cap) noexcept { //NOLINT
 #ifdef BACKEND_OPENGL_VERSION_GL
         case GL_PROGRAM_POINT_SIZE:             index = 10; break;
 #endif
+        case GL_DEPTH_CLAMP:                    index = 11; break;
         default: break;
     }
     assert_invariant(index < state.enables.caps.size());

filament/backend/src/opengl/OpenGLDriver.h

@@ -21,6 +21,8 @@
 #include "OpenGLContext.h"
 #include "OpenGLTimerQuery.h"
 #include "GLBufferObject.h"
+#include "GLDescriptorSet.h"
+#include "GLDescriptorSetLayout.h"
 #include "GLTexture.h"
 #include "ShaderCompilerService.h"
 
@@ -36,6 +38,7 @@
 #include "private/backend/Driver.h"
 #include "private/backend/HandleAllocator.h"
 
+#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -52,6 +55,7 @@
 #include <tuple>
 #include <type_traits>
 #include <utility>
+#include <variant>
 #include <vector>
 
 #include <stddef.h>
@@ -81,7 +85,7 @@ public:
             const Platform::DriverConfig& driverConfig) noexcept;
 
     class DebugMarker {
-        OpenGLDriver& driver;
+        UTILS_UNUSED OpenGLDriver& driver;
     public:
         DebugMarker(OpenGLDriver& driver, const char* string) noexcept;
         ~DebugMarker() noexcept;
@@ -123,16 +127,6 @@ public:
         } gl;
     };
 
-    struct GLSamplerGroup : public HwSamplerGroup {
-        using HwSamplerGroup::HwSamplerGroup;
-        struct Entry {
-            Handle<HwTexture> th;
-            GLuint sampler = 0u;
-        };
-        utils::FixedCapacityVector<Entry> textureUnitEntries;
-        explicit GLSamplerGroup(size_t size) noexcept : textureUnitEntries(size) { }
-    };
-
     struct GLRenderPrimitive : public HwRenderPrimitive {
         using HwRenderPrimitive::HwRenderPrimitive;
         OpenGLContext::RenderPrimitive gl;
@@ -145,6 +139,10 @@ public:
 
     using GLTimerQuery = filament::backend::GLTimerQuery;
 
+    using GLDescriptorSetLayout = filament::backend::GLDescriptorSetLayout;
+
+    using GLDescriptorSet = filament::backend::GLDescriptorSet;
+
     struct GLStream : public HwStream {
         using HwStream::HwStream;
         struct Info {
@@ -317,10 +315,6 @@ private:
     void resolvePass(ResolveAction action, GLRenderTarget const* rt,
             TargetBufferFlags discardFlags) noexcept;
 
-    const std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT>& getSamplerBindings() const {
-        return mSamplerBindings;
-    }
-
     using AttachmentArray = std::array<GLenum, MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT + 2>;
     static GLsizei getAttachments(AttachmentArray& attachments, TargetBufferFlags buffers,
             bool isDefaultFramebuffer) noexcept;
@@ -333,8 +327,16 @@ private:
     GLboolean mRenderPassStencilWrite{};
 
     GLRenderPrimitive const* mBoundRenderPrimitive = nullptr;
+    OpenGLProgram* mBoundProgram = nullptr;
     bool mValidProgram = false;
+    utils::bitset8 mInvalidDescriptorSetBindings;
+    utils::bitset8 mInvalidDescriptorSetBindingOffsets;
+    void updateDescriptors(utils::bitset8 invalidDescriptorSets) noexcept;
 
+    struct {
+        backend::DescriptorSetHandle dsh;
+        std::array<uint32_t, CONFIG_UNIFORM_BINDING_COUNT> offsets;
+    } mBoundDescriptorSets[MAX_DESCRIPTOR_SET_COUNT];
 
     void clearWithRasterPipe(TargetBufferFlags clearFlags,
             math::float4 const& linearColor, GLfloat depth, GLint stencil) noexcept;
@@ -346,9 +348,6 @@ private:
     // ES2 only. Uniform buffer emulation binding points
     GLuint mLastAssignedEmulatedUboId = 0;
 
-    // sampler buffer binding points (nullptr if not used)
-    std::array<GLSamplerGroup*, Program::SAMPLER_BINDING_COUNT> mSamplerBindings = {};   // 4 pointers
-
     // this must be accessed from the driver thread only
     std::vector<GLTexture*> mTexturesWithStreamsAttached;
 
@@ -359,8 +358,6 @@ private:
     void detachStream(GLTexture* t) noexcept;
     void replaceStream(GLTexture* t, GLStream* stream) noexcept;
 
-    void updateTextureLodRange(GLTexture* texture, int8_t targetLevel) noexcept;
-
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     // tasks executed on the main thread after the fence signaled
     void whenGpuCommandsComplete(const std::function<void()>& fn) noexcept;
@@ -384,6 +381,7 @@ private:
     bool mRec709OutputColorspace = false;
 
     PushConstantBundle* mCurrentPushConstants = nullptr;
+    PipelineLayout::SetLayout mCurrentSetLayout;
 };
 
 // ------------------------------------------------------------------------------------------------

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -17,6 +17,7 @@
 #include "OpenGLProgram.h"
 
 #include "GLUtils.h"
+#include "GLTexture.h"
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"
 
@@ -24,6 +25,7 @@
 #include <backend/Program.h>
 #include <backend/Handle.h>
 
+#include <utils/BitmaskEnum.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
 #include <utils/FixedCapacityVector.h>
@@ -32,9 +34,10 @@
 
 #include <algorithm>
 #include <array>
+#include <algorithm>
+#include <new>
 #include <string_view>
 #include <utility>
-#include <new>
 
 #include <stddef.h>
 #include <stdint.h>
@@ -46,9 +49,8 @@ using namespace utils;
 using namespace backend;
 
 struct OpenGLProgram::LazyInitializationData {
-    Program::UniformBlockInfo uniformBlockInfo;
-    Program::SamplerGroupInfo samplerGroupInfo;
-    std::array<Program::UniformInfo, Program::UNIFORM_BINDING_COUNT> bindingUniformInfo;
+    Program::DescriptorSetInfo descriptorBindings;
+    Program::BindingUniformsInfo bindingUniformInfo;
     utils::FixedCapacityVector<Program::PushConstant> vertexPushConstants;
     utils::FixedCapacityVector<Program::PushConstant> fragmentPushConstants;
 };
@@ -57,16 +59,14 @@ struct OpenGLProgram::LazyInitializationData {
 OpenGLProgram::OpenGLProgram() noexcept = default;
 
 OpenGLProgram::OpenGLProgram(OpenGLDriver& gld, Program&& program) noexcept
-        : HwProgram(std::move(program.getName())) {
+        : HwProgram(std::move(program.getName())), mRec709Location(-1) {
     auto* const lazyInitializationData = new(std::nothrow) LazyInitializationData();
-    lazyInitializationData->samplerGroupInfo = std::move(program.getSamplerGroupInfo());
     if (UTILS_UNLIKELY(gld.getContext().isES2())) {
         lazyInitializationData->bindingUniformInfo = std::move(program.getBindingUniformInfo());
-    } else {
-        lazyInitializationData->uniformBlockInfo = std::move(program.getUniformBlockBindings());
     }
     lazyInitializationData->vertexPushConstants = std::move(program.getPushConstants(ShaderStage::VERTEX));
     lazyInitializationData->fragmentPushConstants = std::move(program.getPushConstants(ShaderStage::FRAGMENT));
+    lazyInitializationData->descriptorBindings = std::move(program.getDescriptorBindings());
 
     ShaderCompilerService& compiler = gld.getShaderCompilerService();
     mToken = compiler.createProgram(name, std::move(program));
@@ -124,36 +124,86 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
 
     SYSTRACE_CALL();
 
+    // from the pipeline layout we compute a mapping from {set, binding} to {binding}
+    // for both buffers and textures
+
+    for (auto&& entry: lazyInitializationData.descriptorBindings) {
+        std::sort(entry.begin(), entry.end(),
+                [](Program::Descriptor const& lhs, Program::Descriptor const& rhs) {
+                    return lhs.binding < rhs.binding;
+                });
+    }
+
+    GLuint tmu = 0;
+    GLuint binding = 0;
+
+    // needed for samplers
+    context.useProgram(program);
+
+    UTILS_NOUNROLL
+    for (backend::descriptor_set_t set = 0; set < MAX_DESCRIPTOR_SET_COUNT; set++) {
+        for (Program::Descriptor const& entry: lazyInitializationData.descriptorBindings[set]) {
+            switch (entry.type) {
+                case DescriptorType::UNIFORM_BUFFER:
+                case DescriptorType::SHADER_STORAGE_BUFFER: {
+                    if (!entry.name.empty()) {
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-    if (!context.isES2()) {
-        // Note: This is only needed, because the layout(binding=) syntax is not permitted in glsl
-        // (ES3.0 and GL4.1). The backend needs a way to associate a uniform block to a binding point.
-        UTILS_NOUNROLL
-        for (GLuint binding = 0, n = lazyInitializationData.uniformBlockInfo.size();
-                binding < n; binding++) {
-            auto const& name = lazyInitializationData.uniformBlockInfo[binding];
-            if (!name.empty()) {
-                GLuint const index = glGetUniformBlockIndex(program, name.c_str());
-                if (index != GL_INVALID_INDEX) {
-                    glUniformBlockBinding(program, index, binding);
-                }
-                CHECK_GL_ERROR(utils::slog.e)
-            }
-        }
-    } else
+                        if (UTILS_LIKELY(!context.isES2())) {
+                            GLuint const index = glGetUniformBlockIndex(program,
+                                    entry.name.c_str());
+                            if (index != GL_INVALID_INDEX) {
+                                // this can fail if the program doesn't use this descriptor
+                                glUniformBlockBinding(program, index, binding);
+                                mBindingMap.insert(set, entry.binding,
+                                        { binding, entry.type });
+                                ++binding;
+                            }
+                        } else
 #endif
-    {
+                        {
+                            auto pos = std::find_if(lazyInitializationData.bindingUniformInfo.begin(),
+                                    lazyInitializationData.bindingUniformInfo.end(),
+                                    [&name = entry.name](const auto& item) {
+                                return std::get<1>(item) == name;
+                            });
+                            if (pos != lazyInitializationData.bindingUniformInfo.end()) {
+                                binding = std::get<0>(*pos);
+                                mBindingMap.insert(set, entry.binding, { binding, entry.type });
+                            }
+                        }
+                    }
+                    break;
+                }
+                case DescriptorType::SAMPLER: {
+                    if (!entry.name.empty()) {
+                        GLint const loc = glGetUniformLocation(program, entry.name.c_str());
+                        if (loc >= 0) {
+                            // this can fail if the program doesn't use this descriptor
+                            mBindingMap.insert(set, entry.binding, { tmu, entry.type });
+                            glUniform1i(loc, GLint(tmu));
+                            ++tmu;
+                        }
+                    }
+                    break;
+                }
+                case DescriptorType::INPUT_ATTACHMENT:
+                    break;
+            }
+            CHECK_GL_ERROR(utils::slog.e)
+        }
+    }
+
+    if (context.isES2()) {
         // ES2 initialization of (fake) UBOs
         UniformsRecord* const uniformsRecords = new(std::nothrow) UniformsRecord[Program::UNIFORM_BINDING_COUNT];
         UTILS_NOUNROLL
-        for (GLuint binding = 0, n = Program::UNIFORM_BINDING_COUNT; binding < n; binding++) {
-            Program::UniformInfo& uniforms = lazyInitializationData.bindingUniformInfo[binding];
-            uniformsRecords[binding].locations.reserve(uniforms.size());
-            uniformsRecords[binding].locations.resize(uniforms.size());
+        for (auto&& [index, name, uniforms] : lazyInitializationData.bindingUniformInfo) {
+            uniformsRecords[index].locations.reserve(uniforms.size());
+            uniformsRecords[index].locations.resize(uniforms.size());
             for (size_t j = 0, c = uniforms.size(); j < c; j++) {
                 GLint const loc = glGetUniformLocation(program, uniforms[j].name.c_str());
-                uniformsRecords[binding].locations[j] = loc;
-                if (UTILS_UNLIKELY(binding == 0)) {
+                uniformsRecords[index].locations[j] = loc;
+                if (UTILS_UNLIKELY(index == 0)) {
                     // This is a bit of a gross hack here, we stash the location of
                     // "frameUniforms.rec709", which obviously the backend shouldn't know about,
                     // which is used for emulating the "rec709" colorspace in the shader.
@@ -165,51 +215,11 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
                     }
                 }
             }
-            uniformsRecords[binding].uniforms = std::move(uniforms);
+            uniformsRecords[index].uniforms = std::move(uniforms);
         }
         mUniformsRecords = uniformsRecords;
     }
 
-    uint8_t usedBindingCount = 0;
-    uint8_t tmu = 0;
-
-    UTILS_NOUNROLL
-    for (size_t i = 0, c = lazyInitializationData.samplerGroupInfo.size(); i < c; i++) {
-        auto const& samplers = lazyInitializationData.samplerGroupInfo[i].samplers;
-        if (samplers.empty()) {
-            // this binding point doesn't have any samplers, skip it.
-            continue;
-        }
-
-        // keep this in the loop, so we skip it in the rare case a program doesn't have
-        // sampler. The context cache will prevent repeated calls to GL.
-        context.useProgram(program);
-
-        bool atLeastOneSamplerUsed = false;
-        UTILS_NOUNROLL
-        for (const Program::Sampler& sampler: samplers) {
-            // find its location and associate a TMU to it
-            GLint const loc = glGetUniformLocation(program, sampler.name.c_str());
-            if (loc >= 0) {
-                // this can fail if the program doesn't use this sampler
-                glUniform1i(loc, tmu);
-                atLeastOneSamplerUsed = true;
-            }
-            tmu++;
-        }
-
-        // if this program doesn't use any sampler from this HwSamplerGroup, just cancel the
-        // whole group.
-        if (atLeastOneSamplerUsed) {
-            // Cache the sampler uniform locations for each interface block
-            mUsedSamplerBindingPoints[usedBindingCount] = i;
-            usedBindingCount++;
-        } else {
-            tmu -= samplers.size();
-        }
-    }
-    mUsedBindingsCount = usedBindingCount;
-
     auto& vertexConstants = lazyInitializationData.vertexPushConstants;
     auto& fragmentConstants = lazyInitializationData.fragmentPushConstants;
 
@@ -226,41 +236,8 @@ void OpenGLProgram::initializeProgramState(OpenGLContext& context, GLuint progra
     }
 }
 
-void OpenGLProgram::updateSamplers(OpenGLDriver* const gld) const noexcept {
-    using GLTexture = OpenGLDriver::GLTexture;
-
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-    bool const es2 = gld->getContext().isES2();
-#endif
-
-    // cache a few member variable locally, outside the loop
-    auto const& UTILS_RESTRICT samplerBindings = gld->getSamplerBindings();
-    auto const& UTILS_RESTRICT usedBindingPoints = mUsedSamplerBindingPoints;
-
-    for (uint8_t i = 0, tmu = 0, n = mUsedBindingsCount; i < n; i++) {
-        size_t const binding = usedBindingPoints[i];
-        assert_invariant(binding < Program::SAMPLER_BINDING_COUNT);
-        auto const * const sb = samplerBindings[binding];
-        assert_invariant(sb);
-        if (!sb) continue; // should never happen, this would be a user error.
-        for (uint8_t j = 0, m = sb->textureUnitEntries.size(); j < m; ++j, ++tmu) { // "<=" on purpose here
-            Handle<HwTexture> th = sb->textureUnitEntries[j].th;
-            if (th) { // program may not use all samplers of sampler group
-                GLTexture const* const t = gld->handle_cast<GLTexture const*>(th);
-                gld->bindTexture(tmu, t);
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-                if (UTILS_LIKELY(!es2)) {
-                    GLuint const s = sb->textureUnitEntries[j].sampler;
-                    gld->bindSampler(tmu, s);
-                }
-#endif
-            }
-        }
-    }
-    CHECK_GL_ERROR(utils::slog.e)
-}
-
-void OpenGLProgram::updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept {
+void OpenGLProgram::updateUniforms(
+        uint32_t index, GLuint id, void const* buffer, uint16_t age) const noexcept {
     assert_invariant(mUniformsRecords);
     assert_invariant(buffer);
 

filament/backend/src/opengl/OpenGLProgram.h

@@ -19,17 +19,20 @@
 
 #include "DriverBase.h"
 
+#include "BindingMap.h"
 #include "OpenGLContext.h"
 #include "ShaderCompilerService.h"
 
 #include <private/backend/Driver.h>
+
+#include <backend/DriverEnums.h>
 #include <backend/Program.h>
 
+#include <utils/bitset.h>
 #include <utils/compiler.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Slice.h>
 
-#include <array>
 #include <limits>
 
 #include <stddef.h>
@@ -69,31 +72,24 @@ public:
         }
 
         context.useProgram(gl.program);
-        if (UTILS_UNLIKELY(mUsedBindingsCount)) {
-            // We rely on GL state tracking to avoid unnecessary glBindTexture / glBindSampler
-            // calls.
-
-            // we need to do this if:
-            // - the content of mSamplerBindings has changed
-            // - the content of any bound sampler buffer has changed
-            // ... since last time we used this program
-
-            // Turns out the former might be relatively cheap to check, the latter requires
-            // a bit less. Compared to what updateSamplers() actually does, which is
-            // pretty little, I'm not sure if we'll get ahead.
-
-            updateSamplers(gld);
-        }
         return true;
     }
 
-    // For ES2 only
-    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) noexcept;
-    void setRec709ColorSpace(bool rec709) const noexcept;
+    GLuint getBufferBinding(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
+        return mBindingMap.get(set, binding);
+    }
 
-    struct {
-        GLuint program = 0;
-    } gl;                                               // 4 bytes
+    GLuint getTextureUnit(descriptor_set_t set, descriptor_binding_t binding) const noexcept {
+        return mBindingMap.get(set, binding);
+    }
+
+    utils::bitset64 getActiveDescriptors(descriptor_set_t set) const noexcept {
+        return mBindingMap.getActiveDescriptors(set);
+    }
+
+    // For ES2 only
+    void updateUniforms(uint32_t index, GLuint id, void const* buffer, uint16_t age) const noexcept;
+    void setRec709ColorSpace(bool rec709) const noexcept;
 
     PushConstantBundle getPushConstants() {
         auto fragBegin = mPushConstants.begin() + mPushConstantFragmentStageOffset;
@@ -112,22 +108,15 @@ private:
     void initializeProgramState(OpenGLContext& context, GLuint program,
             LazyInitializationData& lazyInitializationData) noexcept;
 
-    void updateSamplers(OpenGLDriver* gld) const noexcept;
-
-    // number of bindings actually used by this program
-    std::array<uint8_t, Program::SAMPLER_BINDING_COUNT> mUsedSamplerBindingPoints;   // 4 bytes
+    BindingMap mBindingMap;     // 8 bytes + out-of-line 256 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
+    // 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
 
     // only needed for ES2
-    GLint mRec709Location = -1;                         // 4 bytes
-
     using LocationInfo = utils::FixedCapacityVector<GLint>;
     struct UniformsRecord {
         Program::UniformInfo uniforms;
@@ -135,15 +124,20 @@ private:
         mutable GLuint id = 0;
         mutable uint16_t age = std::numeric_limits<uint16_t>::max();
     };
-    UniformsRecord const* mUniformsRecords = nullptr;               // 8 bytes
+    UniformsRecord const* mUniformsRecords = nullptr;
+    GLint mRec709Location : 24;     // 4 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
+    // Push constant array offset for fragment stage constants.
+    GLint mPushConstantFragmentStageOffset : 8;      // 1 byte
+
+public:
+    struct {
+        GLuint program = 0;
+    } gl;                                               // 4 bytes
 };
 
-// if OpenGLProgram is larger tha 64 bytes, it'll fall in a larger Handle bucket.
-static_assert(sizeof(OpenGLProgram) <= 64); // currently 64 bytes
+// if OpenGLProgram is larger than 96 bytes, it'll fall in a larger Handle bucket.
+static_assert(sizeof(OpenGLProgram) <= 96); // currently 96 bytes
 
 } // namespace filament::backend
 

filament/backend/src/opengl/OpenGLTimerQuery.cpp

@@ -143,21 +143,25 @@ void TimerQueryNativeFactory::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQ
 
     driver.runEveryNowAndThen([&context = mContext, weak]() -> bool {
         auto state = weak.lock();
-        if (state) {
-            GLuint available = 0;
-            context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
-            CHECK_GL_ERROR(utils::slog.e)
-            if (!available) {
-                // we need to try this one again later
-                return false;
-            }
-            GLuint64 elapsedTime = 0;
-            // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
-            context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
-            state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
-        } else {
-            state->elapsed.store(int64_t(TimerQueryResult::ERROR), std::memory_order_relaxed);
+        if (!state) {
+            // The timer query state has been destroyed on the way, very likely due to the IBL
+            // prefilter context destruction. We still return true to get this element removed from
+            // the query list.
+            return true;
         }
+
+        GLuint available = 0;
+        context.procs.getQueryObjectuiv(state->gl.query, GL_QUERY_RESULT_AVAILABLE, &available);
+        CHECK_GL_ERROR(utils::slog.e)
+        if (!available) {
+            // we need to try this one again later
+            return false;
+        }
+        GLuint64 elapsedTime = 0;
+        // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
+        context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
+        state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
+
         return true;
     });
 }

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -26,16 +26,28 @@
 
 #include <utils/compiler.h>
 #include <utils/CString.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/JobSystem.h>
 #include <utils/Log.h>
+#include <utils/ostream.h>
+#include <utils/Panic.h>
 #include <utils/Systrace.h>
 
+#include <array>
 #include <cctype>
 #include <chrono>
+#include <mutex>
+#include <memory>
 #include <string>
 #include <string_view>
+#include <thread>
+#include <utility>
 #include <variant>
 
+#include <stddef.h>
+#include <stdint.h>
+
 namespace filament::backend {
 
 using namespace utils;
@@ -476,6 +488,12 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
     // check status of program linking and shader compilation, logs error and free all resources
     // in case of error.
     bool const success = checkProgramStatus(token);
+
+    // Unless we have matdbg, we panic if a program is invalid. Otherwise, we'd get a UB.
+    // The compilation error has been logged to log.e by this point.
+    FILAMENT_CHECK_POSTCONDITION(FILAMENT_ENABLE_MATDBG || success)
+            << "OpenGL program " << token->name.c_str_safe() << " failed to link or compile";
+
     if (UTILS_LIKELY(success)) {
         program = token->gl.program;
         // no need to keep the shaders around
@@ -579,24 +597,31 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
                 version = "#version 310 es\n";
             }
 
-            const std::array<const char*, 5> sources = {
-                    version.data(),
-                    prolog.data(),
-                    specializationConstantString.c_str(),
-                    packingFunctions.data(),
-                    body.data()
+            std::array<std::string_view, 5> sources = {
+                version,
+                prolog,
+                specializationConstantString,
+                packingFunctions,
+                { body.data(), body.size() - 1 }  // null-terminated
             };
 
-            const std::array<GLint, 5> lengths = {
-                    (GLint)version.length(),
-                    (GLint)prolog.length(),
-                    (GLint)specializationConstantString.length(),
-                    (GLint)packingFunctions.length(),
-                    (GLint)body.length() - 1 // null terminated
-            };
+            // Some of the sources may be zero-length. Remove them as to avoid passing lengths of
+            // zero to glShaderSource(). glShaderSource should work with lengths of zero, but some
+            // drivers instead interpret zero as a sentinel for a null-terminated string.
+            auto partitionPoint = std::stable_partition(
+                    sources.begin(), sources.end(), [](std::string_view s) { return !s.empty(); });
+            size_t count = std::distance(sources.begin(), partitionPoint);
+
+            std::array<const char*, 5> shaderStrings;
+            std::array<GLint, 5> lengths;
+            for (size_t i = 0; i < count; i++) {
+                shaderStrings[i] = sources[i].data();
+                lengths[i] = sources[i].size();
+            }
 
             GLuint const shaderId = glCreateShader(glShaderType);
-            glShaderSource(shaderId, sources.size(), sources.data(), lengths.data());
+            glShaderSource(shaderId, count, shaderStrings.data(), lengths.data());
+
             glCompileShader(shaderId);
 
 #ifndef NDEBUG
@@ -785,25 +810,24 @@ mediump vec4 unpackSnorm4x8(highp uint v) {
 // - extensions
 // - everything else
 std::array<std::string_view, 3> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
-    auto start = source.find("#version");
-    assert_invariant(start != std::string_view::npos);
+    auto version_start = source.find("#version");
+    assert_invariant(version_start != std::string_view::npos);
 
-    auto version_eol = source.find('\n', start) + 1;
+    auto version_eol = source.find('\n', version_start) + 1;
     assert_invariant(version_eol != std::string_view::npos);
 
-    auto pos = source.rfind("\n#extension");
-    if (pos == std::string_view::npos) {
-        pos = version_eol;
+    auto prolog_start = version_eol;
+    auto prolog_eol = source.rfind("\n#extension"); // last #extension line
+    if (prolog_eol == std::string_view::npos) {
+        prolog_eol = prolog_start;
     } else {
-        ++pos;
+        prolog_eol = source.find('\n', prolog_eol + 1) + 1;
     }
+    auto body_start = prolog_eol;
 
-    auto eol = source.find('\n', pos) + 1;
-    assert_invariant(eol != std::string_view::npos);
-
-    std::string_view const version = source.substr(start, version_eol - start);
-    std::string_view const prolog = source.substr(version_eol, eol - version_eol);
-    std::string_view const body = source.substr(eol, source.length() - eol);
+    std::string_view const version = source.substr(version_start, version_eol - version_start);
+    std::string_view const prolog = source.substr(prolog_start, prolog_eol - prolog_start);
+    std::string_view const body = source.substr(body_start, source.length() - body_start);
     return { version, prolog, body };
 }
 

filament/backend/src/opengl/gl_headers.h

@@ -201,6 +201,12 @@ using namespace glext;
 #   define GL_CLIP_DISTANCE1                        0x3001
 #endif
 
+#if defined(GL_EXT_depth_clamp)
+#   define GL_DEPTH_CLAMP                           GL_DEPTH_CLAMP_EXT
+#else
+#   define GL_DEPTH_CLAMP                           0x864F
+#endif
+
 #if defined(GL_KHR_debug)
 #   define GL_DEBUG_OUTPUT                          GL_DEBUG_OUTPUT_KHR
 #   define GL_DEBUG_OUTPUT_SYNCHRONOUS              GL_DEBUG_OUTPUT_SYNCHRONOUS_KHR

filament/backend/src/vulkan/VulkanBlitter.cpp

@@ -15,6 +15,7 @@
  */
 
 #include "VulkanBlitter.h"
+#include "VulkanCommands.h"
 #include "VulkanContext.h"
 #include "VulkanFboCache.h"
 #include "VulkanHandles.h"
@@ -33,9 +34,10 @@ namespace filament::backend {
 
 namespace {
 
-inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect, VkFilter filter,
+inline void blitFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect, VkFilter filter,
         VulkanAttachment src, VulkanAttachment dst,
         const VkOffset3D srcRect[2], const VkOffset3D dstRect[2]) {
+    VkCommandBuffer const cmdbuf = commands->buffer();
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
         FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
@@ -49,8 +51,8 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
 
-    src.texture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, VulkanLayout::TRANSFER_DST);
+    src.texture->transitionLayout(commands, srcRange, VulkanLayout::TRANSFER_SRC);
+    dst.texture->transitionLayout(commands, dstRange, VulkanLayout::TRANSFER_DST);
 
     const VkImageBlit blitRegions[1] = {{
             .srcSubresource = { aspect, src.level, src.layer, 1 },
@@ -58,7 +60,7 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
             .dstSubresource = { aspect, dst.level, dst.layer, 1 },
             .dstOffsets = { dstRect[0], dstRect[1] },
     }};
-    vkCmdBlitImage(cmdbuffer,
+    vkCmdBlitImage(cmdbuf,
             src.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
             dst.getImage(), imgutil::getVkLayout(VulkanLayout::TRANSFER_DST),
             1, blitRegions, filter);
@@ -69,12 +71,13 @@ inline void blitFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
         oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
     }
-    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
+    src.texture->transitionLayout(commands, srcRange, oldSrcLayout);
+    dst.texture->transitionLayout(commands, dstRange, oldDstLayout);
 }
 
-inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspect,
+inline void resolveFast(VulkanCommandBuffer* commands, VkImageAspectFlags aspect,
         VulkanAttachment src, VulkanAttachment dst) {
+    VkCommandBuffer const cmdbuffer = commands->buffer();
     if constexpr (FVK_ENABLED(FVK_DEBUG_BLITTER)) {
         FVK_LOGD << "Fast blit from=" << src.texture->getVkImage() << ",level=" << (int) src.level
                       << " layout=" << src.getLayout()
@@ -88,8 +91,8 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
     VulkanLayout oldSrcLayout = src.getLayout();
     VulkanLayout oldDstLayout = dst.getLayout();
 
-    src.texture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, VulkanLayout::TRANSFER_DST);
+    src.texture->transitionLayout(commands, srcRange, VulkanLayout::TRANSFER_SRC);
+    dst.texture->transitionLayout(commands, dstRange, VulkanLayout::TRANSFER_DST);
 
     assert_invariant(
             aspect != VK_IMAGE_ASPECT_DEPTH_BIT && "Resolve with depth is not yet supported.");
@@ -111,8 +114,8 @@ inline void resolveFast(const VkCommandBuffer cmdbuffer, VkImageAspectFlags aspe
     if (oldDstLayout == VulkanLayout::UNDEFINED) {
         oldDstLayout = imgutil::getDefaultLayout(dst.texture->usage);
     }
-    src.texture->transitionLayout(cmdbuffer, srcRange, oldSrcLayout);
-    dst.texture->transitionLayout(cmdbuffer, dstRange, oldDstLayout);
+    src.texture->transitionLayout(commands, srcRange, oldSrcLayout);
+    dst.texture->transitionLayout(commands, dstRange, oldDstLayout);
 }
 
 struct BlitterUniforms {
@@ -149,10 +152,9 @@ void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {
 #endif
 
     VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
     commands.acquire(src.texture);
     commands.acquire(dst.texture);
-    resolveFast(cmdbuffer, aspect, src, dst);
+    resolveFast(&commands, aspect, src, dst);
 }
 
 void VulkanBlitter::blit(VkFilter filter,
@@ -175,10 +177,9 @@ void VulkanBlitter::blit(VkFilter filter,
     // src and dst should have the same aspect here
     VkImageAspectFlags const aspect = src.texture->getImageAspect();
     VulkanCommandBuffer& commands = mCommands->get();
-    VkCommandBuffer const cmdbuffer = commands.buffer();
     commands.acquire(src.texture);
     commands.acquire(dst.texture);
-    blitFast(cmdbuffer, aspect, filter, src, dst, srcRectPair, dstRectPair);
+    blitFast(&commands, aspect, filter, src, dst, srcRectPair, dstRectPair);
 }
 
 void VulkanBlitter::terminate() noexcept {

filament/backend/src/vulkan/VulkanBuffer.cpp

@@ -28,6 +28,7 @@ VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
     : mAllocator(allocator),
       mStagePool(stagePool),
       mUsage(usage),
+	  mUpdatedOffset(0),
       mUpdatedBytes(0) {
     // for now make sure that only 1 bit is set in usage
     // (because loadFromCpu() assumes that somewhat)
@@ -80,6 +81,7 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
             .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .buffer = mGpuBuffer,
+			.offset = mUpdatedOffset,
             .size = mUpdatedBytes,
         };
         vkCmdPipelineBarrier(cmdbuf, srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 1,
@@ -93,6 +95,7 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
     };
     vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);
 
+	mUpdatedOffset = byteOffset;
     mUpdatedBytes = numBytes;
 
     // Firstly, ensure that the copy finishes before the next draw call.

filament/backend/src/vulkan/VulkanBuffer.h

@@ -42,6 +42,7 @@ private:
     VmaAllocation mGpuMemory = VK_NULL_HANDLE;
     VkBuffer mGpuBuffer = VK_NULL_HANDLE;
     VkBufferUsageFlags mUsage = {};
+	uint32_t mUpdatedOffset = 0;
     uint32_t mUpdatedBytes = 0;
 };
 

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -297,11 +297,11 @@ bool VulkanCommands::flush() {
 #endif
 
     auto& cmdfence = currentbuf->fence;
-    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
-    cmdfence->status.store(VK_NOT_READY);
-    UTILS_UNUSED_IN_RELEASE VkResult result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->fence);
-    cmdfence->condition.notify_all();
-    lock.unlock();
+    UTILS_UNUSED_IN_RELEASE VkResult result = VK_SUCCESS;
+    {
+        auto scope = cmdfence->setValue(VK_NOT_READY);
+        result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->getFence());
+    }
 
 #if FVK_ENABLED(FVK_DEBUG_COMMAND_BUFFER)
     if (result != VK_SUCCESS) {
@@ -340,7 +340,7 @@ void VulkanCommands::wait() {
         auto wrapper = mStorage[i].get();
         if (wrapper->buffer() != VK_NULL_HANDLE
                 && mCurrentCommandBufferIndex != static_cast<int8_t>(i)) {
-            fences[count++] = wrapper->fence->fence;
+            fences[count++] = wrapper->fence->getFence();
         }
     }
     if (count > 0) {
@@ -361,12 +361,13 @@ void VulkanCommands::gc() {
         if (wrapper->buffer() == VK_NULL_HANDLE) {
             continue;
         }
-        VkResult const result = vkGetFenceStatus(mDevice, wrapper->fence->fence);
+        auto const vkfence = wrapper->fence->getFence();
+        VkResult const result = vkGetFenceStatus(mDevice, vkfence);
         if (result != VK_SUCCESS) {
             continue;
         }
-        fences[count++] = wrapper->fence->fence;
-        wrapper->fence->status.store(VK_SUCCESS);
+        fences[count++] = vkfence;
+        wrapper->fence->setValue(VK_SUCCESS);
         wrapper->reset();
         mAvailableBufferCount++;
     }
@@ -383,9 +384,9 @@ void VulkanCommands::updateFences() {
         if (wrapper->buffer() != VK_NULL_HANDLE) {
             VulkanCmdFence* fence = wrapper->fence.get();
             if (fence) {
-                VkResult status = vkGetFenceStatus(mDevice, fence->fence);
+                VkResult status = vkGetFenceStatus(mDevice, fence->getFence());
                 // This is either VK_SUCCESS, VK_NOT_READY, or VK_ERROR_DEVICE_LOST.
-                fence->status.store(status, std::memory_order_relaxed);
+                fence->setValue(status);
             }
         }
     }

filament/backend/src/vulkan/VulkanCommands.h

@@ -61,8 +61,40 @@ private:
 
 // Wrapper to enable use of shared_ptr for implementing shared ownership of low-level Vulkan fences.
 struct VulkanCmdFence {
+    struct SetValueScope {
+    public:
+        ~SetValueScope() {
+            mHolder->mutex.unlock();
+            mHolder->condition.notify_all();
+        }
+
+    private:
+        SetValueScope(VulkanCmdFence* fenceHolder, VkResult result) :
+            mHolder(fenceHolder) {
+            mHolder->mutex.lock();
+            mHolder->status.store(result);
+        }
+        VulkanCmdFence* mHolder;
+        friend struct VulkanCmdFence;
+    };
+
     VulkanCmdFence(VkFence ifence);
     ~VulkanCmdFence() = default;
+
+    SetValueScope setValue(VkResult value) {
+        return {this, value};
+    }
+
+    VkFence& getFence() {
+        return fence;
+    }
+
+    VkResult getStatus() {
+        std::unique_lock<utils::Mutex> lock(mutex);
+        return status.load(std::memory_order_acquire);
+    }
+
+private:
     VkFence fence;
     utils::Condition condition;
     utils::Mutex mutex;

filament/backend/src/vulkan/VulkanContext.cpp

@@ -59,7 +59,11 @@ VkExtent2D VulkanAttachment::getExtent2D() const {
 
 VkImageView VulkanAttachment::getImageView() {
     assert_invariant(texture);
-    return texture->getAttachmentView(getSubresourceRange());
+    VkImageSubresourceRange range = getSubresourceRange();
+    if (range.layerCount > 1) {
+        return texture->getMultiviewAttachmentView(range);
+    }
+    return texture->getAttachmentView(range);
 }
 
 bool VulkanAttachment::isDepth() const {
@@ -73,7 +77,7 @@ VkImageSubresourceRange VulkanAttachment::getSubresourceRange() const {
             .baseMipLevel = uint32_t(level),
             .levelCount = 1,
             .baseArrayLayer = uint32_t(layer),
-            .layerCount = 1,
+            .layerCount = layerCount,
     };
 }
 

filament/backend/src/vulkan/VulkanContext.h

@@ -43,6 +43,8 @@ struct VulkanCommandBuffer;
 struct VulkanAttachment {
     VulkanTexture* texture = nullptr;
     uint8_t level = 0;
+    uint8_t baseViewIndex = 0;
+    uint8_t layerCount = 1;
     uint16_t layer = 0;
 
     bool isDepth() const;
@@ -123,6 +125,10 @@ public:
         return mPhysicalDeviceFeatures.imageCubeArray == VK_TRUE;
     }
 
+    inline bool isDepthClampSupported() const noexcept {
+        return mPhysicalDeviceFeatures.depthClamp == VK_TRUE;
+    }
+
     inline bool isDebugMarkersSupported() const noexcept {
         return mDebugMarkersSupported;
     }

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -90,9 +90,9 @@ VmaAllocator createAllocator(VkInstance instance, VkPhysicalDevice physicalDevic
 
 VulkanTexture* createEmptyTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
-        VulkanStagePool& stagePool) {
+        VulkanResourceAllocator* handleAllocator, VulkanStagePool& stagePool) {
     VulkanTexture* emptyTexture = new VulkanTexture(device, physicalDevice, context, allocator,
-            commands, SamplerType::SAMPLER_2D, 1, TextureFormat::RGBA8, 1, 1, 1, 1,
+            commands, handleAllocator, SamplerType::SAMPLER_2D, 1, TextureFormat::RGBA8, 1, 1, 1, 1,
             TextureUsage::DEFAULT | TextureUsage::COLOR_ATTACHMENT | TextureUsage::SUBPASS_INPUT,
             stagePool, true /* heap allocated */);
     uint32_t black = 0;
@@ -149,6 +149,7 @@ VKAPI_ATTR VkBool32 VKAPI_CALL debugUtilsCallback(VkDebugUtilsMessageSeverityFla
 }
 #endif // FVK_EANBLED(FVK_DEBUG_DEBUG_UTILS)
 
+
 }// anonymous namespace
 
 #if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
@@ -256,15 +257,11 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
     mTimestamps = std::make_unique<VulkanTimestamps>(mPlatform->getDevice());
 
     mEmptyTexture = createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
-            mContext, mAllocator, &mCommands, mStagePool);
+            mContext, mAllocator, &mCommands, &mResourceAllocator, mStagePool);
     mEmptyBufferObject = createEmptyBufferObject(mAllocator, mStagePool, &mCommands);
 
     mDescriptorSetManager.setPlaceHolders(mSamplerCache.getSampler({}), mEmptyTexture,
             mEmptyBufferObject);
-
-    mGetPipelineFunction = [this](VulkanDescriptorSetLayoutList const& layouts, VulkanProgram* program) {
-        return mPipelineLayoutCache.getLayout(layouts, program);
-    };
 }
 
 VulkanDriver::~VulkanDriver() noexcept = default;
@@ -387,7 +384,6 @@ void VulkanDriver::collectGarbage() {
     mStagePool.gc();
     mFramebufferCache.gc();
     mPipelineCache.gc();
-    mDescriptorSetManager.gc();
 
 #if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
     mResourceAllocator.print();
@@ -422,6 +418,36 @@ void VulkanDriver::endFrame(uint32_t frameId) {
     FVK_SYSTRACE_END();
 }
 
+void VulkanDriver::updateDescriptorSetBuffer(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_binding_t binding,
+        backend::BufferObjectHandle boh,
+        uint32_t offset,
+        uint32_t size) {
+    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
+    VulkanBufferObject* obj = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
+    mDescriptorSetManager.updateBuffer(set, binding, obj, offset, size);
+}
+
+void VulkanDriver::updateDescriptorSetTexture(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_binding_t binding,
+        backend::TextureHandle th,
+        SamplerParams params) {
+    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
+    VulkanTexture* texture = mResourceAllocator.handle_cast<VulkanTexture*>(th);
+
+    // We need to make sure the initial layout transition has been completed before we can write
+    // the sampler descriptor. We flush and wait until the transition has been completed.
+    if (!texture->transitionReady()) {
+        mCommands.flush();
+        mCommands.wait();
+    }
+
+    VkSampler const vksampler = mSamplerCache.getSampler(params);
+    mDescriptorSetManager.updateSampler(set, binding, texture, vksampler);
+}
+
 void VulkanDriver::flush(int) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("flush");
@@ -440,12 +466,6 @@ void VulkanDriver::finish(int dummy) {
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::createSamplerGroupR(Handle<HwSamplerGroup> sbh, uint32_t count,
-        utils::FixedSizeString<32> debugName) {
-    auto sg = mResourceAllocator.construct<VulkanSamplerGroup>(sbh, count);
-    mResourceManager.acquire(sg);
-}
-
 void VulkanDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
         Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh,
         PrimitiveType pt) {
@@ -528,23 +548,56 @@ void VulkanDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
         TextureUsage usage) {
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
+            target, levels,
             format, samples, w, h, depth, usage, mStagePool);
     mResourceManager.acquire(vktexture);
 }
 
-void VulkanDriver::createTextureSwizzledR(Handle<HwTexture> th, SamplerType target, uint8_t levels,
-        TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
-        TextureUsage usage,
-        TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) {
-    TextureSwizzle swizzleArray[] = {r, g, b, a};
-    const VkComponentMapping swizzleMap = getSwizzleMap(swizzleArray);
+//void VulkanDriver::createTextureSwizzledR(Handle<HwTexture> th, SamplerType target, uint8_t levels,
+//        TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
+//        TextureUsage usage,
+//        TextureSwizzle r, TextureSwizzle g, TextureSwizzle b, TextureSwizzle a) {
+//    TextureSwizzle swizzleArray[] = {r, g, b, a};
+//    const VkComponentMapping swizzleMap = getSwizzleMap(swizzleArray);
+//    auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
+//            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
+//            target, levels, format, samples, w, h, depth, usage, mStagePool,
+//            false /*heap allocated */, swizzleMap);
+//    mResourceManager.acquire(vktexture);
+//}
+
+void VulkanDriver::createTextureViewR(Handle<HwTexture> th, Handle<HwTexture> srch,
+        uint8_t baseLevel, uint8_t levelCount) {
+    VulkanTexture const* src = mResourceAllocator.handle_cast<VulkanTexture const*>(srch);
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
-            format, samples, w, h, depth, usage, mStagePool, false /*heap allocated */, swizzleMap);
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
+            src, baseLevel, levelCount);
     mResourceManager.acquire(vktexture);
 }
 
+void VulkanDriver::createTextureViewSwizzleR(Handle<HwTexture> th, Handle<HwTexture> srch,
+        backend::TextureSwizzle r, backend::TextureSwizzle g, backend::TextureSwizzle b,
+        backend::TextureSwizzle a) {
+    TextureSwizzle const swizzleArray[] = {r, g, b, a};
+    VkComponentMapping const swizzle = getSwizzleMap(swizzleArray);
+
+    VulkanTexture const* src = mResourceAllocator.handle_cast<VulkanTexture const*>(srch);
+    auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, &mResourceAllocator,
+            src, swizzle);
+    mResourceManager.acquire(vktexture);
+}
+
+void VulkanDriver::createTextureExternalImageR(Handle<HwTexture> th, backend::TextureFormat format,
+        uint32_t width, uint32_t height, backend::TextureUsage usage, void* image) {
+}
+
+void VulkanDriver::createTextureExternalImagePlaneR(Handle<HwTexture> th,
+        backend::TextureFormat format, uint32_t width, uint32_t height, backend::TextureUsage usage,
+        void* image, uint32_t plane) {
+}
+
 void VulkanDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
         SamplerType target, uint8_t levels,
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
@@ -571,7 +624,6 @@ void VulkanDriver::destroyProgram(Handle<HwProgram> ph) {
         return;
     }
     auto vkprogram = mResourceAllocator.handle_cast<VulkanProgram*>(ph);
-    mDescriptorSetManager.clearProgram(vkprogram);
     mResourceManager.release(vkprogram);
 }
 
@@ -595,6 +647,8 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
             colorTargets[i] = {
                 .texture = mResourceAllocator.handle_cast<VulkanTexture*>(color[i].handle),
                 .level = color[i].level,
+                .baseViewIndex = color[i].baseViewIndex,
+                .layerCount = layerCount,
                 .layer = color[i].layer,
             };
             UTILS_UNUSED_IN_RELEASE VkExtent2D extent = colorTargets[i].getExtent2D();
@@ -609,6 +663,8 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         depthStencil[0] = {
             .texture = mResourceAllocator.handle_cast<VulkanTexture*>(depth.handle),
             .level = depth.level,
+            .baseViewIndex = depth.baseViewIndex,
+            .layerCount = layerCount,
             .layer = depth.layer,
         };
         UTILS_UNUSED_IN_RELEASE VkExtent2D extent = depthStencil[0].getExtent2D();
@@ -621,6 +677,8 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         depthStencil[1] = {
             .texture = mResourceAllocator.handle_cast<VulkanTexture*>(stencil.handle),
             .level = stencil.level,
+            .baseViewIndex = stencil.baseViewIndex,
+            .layerCount = layerCount,
             .layer = stencil.layer,
         };
         UTILS_UNUSED_IN_RELEASE VkExtent2D extent = depthStencil[1].getExtent2D();
@@ -635,9 +693,10 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
     assert_invariant(tmin == tmax);
     assert_invariant(tmin.x >= width && tmin.y >= height);
 
-    auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, width, height,
-            samples, colorTargets, depthStencil, mStagePool);
+    auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth,
+            mPlatform->getDevice(), mPlatform->getPhysicalDevice(), mContext, mAllocator,
+            &mCommands, &mResourceAllocator, width, height, samples, colorTargets, depthStencil,
+            mStagePool, layerCount);
     mResourceManager.acquire(renderTarget);
 }
 
@@ -665,7 +724,7 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
         flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
     }
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, &mCommands, mStagePool, nativeWindow, flags);
+            mAllocator, &mCommands, &mResourceAllocator, mStagePool, nativeWindow, flags);
     mResourceManager.acquire(swapChain);
 }
 
@@ -678,7 +737,8 @@ void VulkanDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t wi
     }
     assert_invariant(width > 0 && height > 0 && "Vulkan requires non-zero swap chain dimensions.");
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, &mCommands, mStagePool, nullptr, flags, VkExtent2D{width, height});
+            mAllocator, &mCommands, &mResourceAllocator, mStagePool,
+            nullptr, flags, VkExtent2D{width, height});
     mResourceManager.acquire(swapChain);
 }
 
@@ -686,6 +746,22 @@ void VulkanDriver::createTimerQueryR(Handle<HwTimerQuery> tqh, int) {
     // nothing to do, timer query was constructed in createTimerQueryS
 }
 
+void VulkanDriver::createDescriptorSetLayoutR(Handle<HwDescriptorSetLayout> dslh,
+        backend::DescriptorSetLayout&& info) {
+    VulkanDescriptorSetLayout* layout = mResourceAllocator.construct<VulkanDescriptorSetLayout>(
+            dslh, mPlatform->getDevice(), info);
+    mResourceManager.acquire(layout);
+}
+
+void VulkanDriver::createDescriptorSetR(Handle<HwDescriptorSet> dsh,
+        Handle<HwDescriptorSetLayout> dslh) {
+    auto layout = mResourceAllocator.handle_cast<VulkanDescriptorSetLayout*>(dslh);
+    mDescriptorSetManager.createSet(dsh, layout);
+
+    auto set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
+    mResourceManager.acquire(set);
+}
+
 Handle<HwVertexBufferInfo> VulkanDriver::createVertexBufferInfoS() noexcept {
     return mResourceAllocator.allocHandle<VulkanVertexBufferInfo>();
 }
@@ -706,7 +782,19 @@ Handle<HwTexture> VulkanDriver::createTextureS() noexcept {
     return mResourceAllocator.allocHandle<VulkanTexture>();
 }
 
-Handle<HwTexture> VulkanDriver::createTextureSwizzledS() noexcept {
+Handle<HwTexture> VulkanDriver::createTextureViewS() noexcept {
+    return mResourceAllocator.allocHandle<VulkanTexture>();
+}
+
+Handle<HwTexture> VulkanDriver::createTextureViewSwizzleS() noexcept {
+    return mResourceAllocator.allocHandle<VulkanTexture>();
+}
+
+Handle<HwTexture> VulkanDriver::createTextureExternalImageS() noexcept {
+    return mResourceAllocator.allocHandle<VulkanTexture>();
+}
+
+Handle<HwTexture> VulkanDriver::createTextureExternalImagePlaneS() noexcept {
     return mResourceAllocator.allocHandle<VulkanTexture>();
 }
 
@@ -714,10 +802,6 @@ Handle<HwTexture> VulkanDriver::importTextureS() noexcept {
     return mResourceAllocator.allocHandle<VulkanTexture>();
 }
 
-Handle<HwSamplerGroup> VulkanDriver::createSamplerGroupS() noexcept {
-    return mResourceAllocator.allocHandle<VulkanSamplerGroup>();
-}
-
 Handle<HwRenderPrimitive> VulkanDriver::createRenderPrimitiveS() noexcept {
     return mResourceAllocator.allocHandle<VulkanRenderPrimitive>();
 }
@@ -756,21 +840,12 @@ Handle<HwTimerQuery> VulkanDriver::createTimerQueryS() noexcept {
     return tqh;
 }
 
-void VulkanDriver::destroySamplerGroup(Handle<HwSamplerGroup> sbh) {
-    if (!sbh) {
-        return;
-    }
-    // Unlike most of the other "Hw" handles, the sampler buffer is an abstract concept and does
-    // not map to any Vulkan objects. To handle destruction, the only thing we need to do is
-    // ensure that the next draw call doesn't try to access a zombie sampler buffer. Therefore,
-    // simply replace all weak references with null.
-    auto* hwsb = mResourceAllocator.handle_cast<VulkanSamplerGroup*>(sbh);
-    for (auto& binding : mSamplerBindings) {
-        if (binding == hwsb) {
-            binding = nullptr;
-        }
-    }
-    mResourceManager.release(hwsb);
+Handle<HwDescriptorSetLayout> VulkanDriver::createDescriptorSetLayoutS() noexcept {
+    return mResourceAllocator.allocHandle<VulkanDescriptorSetLayout>();
+}
+
+Handle<HwDescriptorSet> VulkanDriver::createDescriptorSetS() noexcept {
+    return mResourceAllocator.allocHandle<VulkanDescriptorSet>();
 }
 
 void VulkanDriver::destroySwapChain(Handle<HwSwapChain> sch) {
@@ -795,6 +870,17 @@ void VulkanDriver::destroyTimerQuery(Handle<HwTimerQuery> tqh) {
     mThreadSafeResourceManager.release(vtq);
 }
 
+void VulkanDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> dslh) {
+    VulkanDescriptorSetLayout* layout = mResourceAllocator.handle_cast<VulkanDescriptorSetLayout*>(dslh);
+    mResourceManager.release(layout);
+}
+
+void VulkanDriver::destroyDescriptorSet(Handle<HwDescriptorSet> dsh) {
+    mDescriptorSetManager.destroySet(dsh);
+    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
+    mResourceManager.release(set);
+}
+
 Handle<HwStream> VulkanDriver::createStreamNative(void* nativeStream) {
     return {};
 }
@@ -832,8 +918,7 @@ FenceStatus VulkanDriver::getFenceStatus(Handle<HwFence> fh) {
 
     // Internally we use the VK_INCOMPLETE status to mean "not yet submitted".
     // When this fence gets submitted, its status changes to VK_NOT_READY.
-    std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
-    if (cmdfence->status.load() == VK_SUCCESS) {
+    if (cmdfence->getStatus() == VK_SUCCESS) {
         return FenceStatus::CONDITION_SATISFIED;
     }
 
@@ -884,7 +969,9 @@ bool VulkanDriver::isRenderTargetFormatSupported(TextureFormat format) {
 }
 
 bool VulkanDriver::isFrameBufferFetchSupported() {
-    return true;
+    // TODO: we must fix this before landing descriptor set change.  Otherwise, the scuba tests will fail.
+    //return true;
+    return false;
 }
 
 bool VulkanDriver::isFrameBufferFetchMultiSampleSupported() {
@@ -938,6 +1025,10 @@ bool VulkanDriver::isProtectedTexturesSupported() {
     return false;
 }
 
+bool VulkanDriver::isDepthClampSupported() {
+    return mContext.isDepthClampSupported();
+}
+
 bool VulkanDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU: {
@@ -986,15 +1077,15 @@ FeatureLevel VulkanDriver::getFeatureLevel() {
 
     // If the max sampler counts do not meet FL2 standards, then this is an FL1 device.
     const auto& fl2 = FEATURE_LEVEL_CAPS[+FeatureLevel::FEATURE_LEVEL_2];
-    if (fl2.MAX_VERTEX_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers ||
-        fl2.MAX_FRAGMENT_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers) {
+    if (limits.maxPerStageDescriptorSamplers < fl2.MAX_VERTEX_SAMPLER_COUNT  ||
+        limits.maxPerStageDescriptorSamplers < fl2.MAX_FRAGMENT_SAMPLER_COUNT) {
         return FeatureLevel::FEATURE_LEVEL_1;
     }
 
     // If the max sampler counts do not meet FL3 standards, then this is an FL2 device.
     const auto& fl3 = FEATURE_LEVEL_CAPS[+FeatureLevel::FEATURE_LEVEL_3];
-    if (fl3.MAX_VERTEX_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers ||
-        fl3.MAX_FRAGMENT_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers) {
+    if (limits.maxPerStageDescriptorSamplers < fl3.MAX_VERTEX_SAMPLER_COUNT||
+        limits.maxPerStageDescriptorSamplers < fl3.MAX_FRAGMENT_SAMPLER_COUNT) {
         return FeatureLevel::FEATURE_LEVEL_2;
     }
 
@@ -1068,10 +1159,6 @@ void VulkanDriver::resetBufferObject(Handle<HwBufferObject> boh) {
     // This is only useful if updateBufferObjectUnsynchronized() is implemented unsynchronizedly.
 }
 
-void VulkanDriver::setMinMaxLevels(Handle<HwTexture> th, uint32_t minLevel, uint32_t maxLevel) {
-    mResourceAllocator.handle_cast<VulkanTexture*>(th)->setPrimaryRange(minLevel, maxLevel);
-}
-
 void VulkanDriver::update3DImage(Handle<HwTexture> th, uint32_t level, uint32_t xoffset,
         uint32_t yoffset, uint32_t zoffset, uint32_t width, uint32_t height, uint32_t depth,
         PixelBufferDescriptor&& data) {
@@ -1157,23 +1244,6 @@ void VulkanDriver::generateMipmaps(Handle<HwTexture> th) {
         srcw = dstw;
         srch = dsth;
     } while ((srcw > 1 || srch > 1) && level < t->levels);
-    t->setPrimaryRange(0, t->levels - 1);
-}
-
-void VulkanDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
-        BufferDescriptor&& data) {
-    auto* sb = mResourceAllocator.handle_cast<VulkanSamplerGroup*>(sbh);
-
-    // FIXME: we shouldn't be using SamplerGroup here, instead the backend should create
-    //        a descriptor or any internal data-structure that represents the textures/samplers.
-    //        It's preferable to do as much work as possible here.
-    //        Here, we emulate the older backend API by re-creating a SamplerGroup from the
-    //        passed data.
-    SamplerGroup samplerGroup(data.size / sizeof(SamplerDescriptor));
-    memcpy(samplerGroup.data(), data.buffer, data.size);
-    *sb->sb = std::move(samplerGroup);
-
-    scheduleDestroy(std::move(data));
 }
 
 void VulkanDriver::compilePrograms(CompilerPriorityQueue priority,
@@ -1188,7 +1258,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     FVK_SYSTRACE_START("beginRenderPass");
 
     VulkanRenderTarget* const rt = mResourceAllocator.handle_cast<VulkanRenderTarget*>(rth);
-    const VkExtent2D extent = rt->getExtent();
+    VkExtent2D const extent = rt->getExtent();
     assert_invariant(rt == mDefaultRenderTarget || extent.width > 0 && extent.height > 0);
 
     // Filament has the expectation that the contents of the swap chain are not preserved on the
@@ -1218,36 +1288,6 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     // the non-sampling case.
     VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer const cmdbuffer = commands.buffer();
-
-    UTILS_NOUNROLL
-    for (uint8_t samplerGroupIdx = 0; samplerGroupIdx < Program::SAMPLER_BINDING_COUNT;
-            samplerGroupIdx++) {
-        VulkanSamplerGroup* vksb = mSamplerBindings[samplerGroupIdx];
-        if (!vksb) {
-            continue;
-        }
-        SamplerGroup* sb = vksb->sb.get();
-        for (size_t i = 0; i < sb->getSize(); i++) {
-            SamplerDescriptor const* boundSampler = sb->data() + i;
-            if (UTILS_LIKELY(boundSampler->t)) {
-                VulkanTexture* texture
-                        = mResourceAllocator.handle_cast<VulkanTexture*>(boundSampler->t);
-                if (!any(texture->usage & TextureUsage::DEPTH_ATTACHMENT)) {
-                    continue;
-                }
-                if (texture->getPrimaryImageLayout() == VulkanLayout::DEPTH_SAMPLER) {
-                    continue;
-                }
-                commands.acquire(texture);
-
-                // Transition the primary view, which is the sampler's view into the right layout.
-                texture->transitionLayout(cmdbuffer, texture->getPrimaryViewRange(),
-                        VulkanLayout::DEPTH_SAMPLER);
-                break;
-            }
-        }
-    }
-
     VulkanLayout currentDepthLayout = depth.getLayout();
 
     TargetBufferFlags clearVal = params.flags.clear;
@@ -1260,13 +1300,13 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         // 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;
-        }
+        depth.texture->transitionLayout(&commands, depth.getSubresourceRange(),
+                VulkanLayout::DEPTH_ATTACHMENT);
+        currentDepthLayout = VulkanLayout::DEPTH_ATTACHMENT;
     }
 
+    uint8_t const renderTargetLayerCount = rt->getLayerCount();
+
     // Create the VkRenderPass or fetch it from cache.
     VulkanFboCache::RenderPassKey rpkey = {
         .initialColorLayoutMask = 0,
@@ -1277,20 +1317,17 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         .discardEnd = discardEndVal,
         .samples = rt->getSamples(),
         .subpassMask = uint8_t(params.subpassMask),
+        .viewCount = renderTargetLayerCount,
     };
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
-        const VulkanAttachment& info = rt->getColor(i);
+        VulkanAttachment const& info = rt->getColor(i);
         if (info.texture) {
+            assert_invariant(info.layerCount == renderTargetLayerCount);
             rpkey.initialColorLayoutMask |= 1 << i;
             rpkey.colorFormat[i] = info.getFormat();
             if (rpkey.samples > 1 && info.texture->samples == 1) {
                 rpkey.needsResolveMask |= (1 << i);
             }
-            if (info.texture->getPrimaryImageLayout() != VulkanLayout::COLOR_ATTACHMENT) {
-                ((VulkanTexture*) info.texture)
-                        ->transitionLayout(cmdbuffer, info.getSubresourceRange(),
-                                VulkanLayout::COLOR_ATTACHMENT);
-            }
         } else {
             rpkey.colorFormat[i] = VK_FORMAT_UNDEFINED;
         }
@@ -1309,28 +1346,39 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     };
     auto& renderPassAttachments = mRenderPassFboInfo.attachments;
     for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; i++) {
-        if (!rt->getColor(i).texture) {
+        VulkanAttachment& attachment = rt->getColor(i);
+        if (!attachment.texture) {
             fbkey.color[i] = VK_NULL_HANDLE;
             fbkey.resolve[i] = VK_NULL_HANDLE;
-        } else if (fbkey.samples == 1) {
-            auto& colorAttachment = rt->getColor(i);
-            renderPassAttachments.insert(colorAttachment);
-            fbkey.color[i] = colorAttachment.getImageView();
+            continue;
+        }
+
+        if (fbkey.samples == 1) {
+            auto const& range = attachment.getSubresourceRange();
+            attachment.texture->transitionLayout(&commands,
+                    range, VulkanLayout::COLOR_ATTACHMENT);
+            renderPassAttachments.insert(attachment);
+            fbkey.color[i] = attachment.getImageView();
             fbkey.resolve[i] = VK_NULL_HANDLE;
             assert_invariant(fbkey.color[i]);
         } else {
             auto& msaaColorAttachment = rt->getMsaaColor(i);
+            auto const& msaaRange = attachment.getSubresourceRange();
+            msaaColorAttachment.texture->transitionLayout(&commands,
+                    msaaRange, VulkanLayout::COLOR_ATTACHMENT);
             renderPassAttachments.insert(msaaColorAttachment);
 
-            auto& colorAttachment = rt->getColor(i);
             fbkey.color[i] = msaaColorAttachment.getImageView();
 
-            VulkanTexture* texture = colorAttachment.texture;
+            VulkanTexture* texture = attachment.texture;
             if (texture->samples == 1) {
                 mRenderPassFboInfo.hasColorResolve = true;
 
-                renderPassAttachments.insert(colorAttachment);
-                fbkey.resolve[i] = colorAttachment.getImageView();
+                auto const& range = attachment.getSubresourceRange();
+                attachment.texture->transitionLayout(&commands,
+                        range, VulkanLayout::COLOR_ATTACHMENT);
+                renderPassAttachments.insert(attachment);
+                fbkey.resolve[i] = attachment.getImageView();
                 assert_invariant(fbkey.resolve[i]);
             }
             assert_invariant(fbkey.color[i]);
@@ -1448,57 +1496,27 @@ void VulkanDriver::endRenderPass(int) {
     assert_invariant(rt);
 
     // Since we might soon be sampling from the render target that we just wrote to, we need a
-    // pipeline barrier between framebuffer writes and shader reads. This is a memory barrier rather
-    // than an image barrier. If we were to use image barriers here, we would potentially need to
-    // 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.
+    // pipeline barrier between framebuffer writes and shader reads.
     if (!rt->isSwapChain()) {
-        for (auto const& attachment: mRenderPassFboInfo.attachments) {
-            bool const isDepth = attachment.isDepth();
-            VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
-
-            // This is a workaround around a validation issue (might not be an actual driver issue).
-            if (mRenderPassFboInfo.hasColorResolve && !isDepth) {
-                srcStageMask = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT;
-            }
-
-            VkPipelineStageFlags dstStageMask =
-                    VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
-            VkAccessFlags srcAccess = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-            VkAccessFlags dstAccess = VK_ACCESS_SHADER_READ_BIT;
-            VulkanLayout layout = VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT;
-            if (isDepth) {
-                srcAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
-                dstAccess = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
-                srcStageMask = VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
-                dstStageMask = VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT;
-                layout =  VulkanFboCache::FINAL_DEPTH_ATTACHMENT_LAYOUT;
-            }
-
-            auto const vkLayout = imgutil::getVkLayout(layout);
+        for (auto& attachment: mRenderPassFboInfo.attachments) {
             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);
+            bool const isDepth = attachment.isDepth();
+            auto texture = attachment.texture;
+            if (isDepth) {
+                texture->setLayout(range, VulkanFboCache::FINAL_DEPTH_ATTACHMENT_LAYOUT);
+                if (!texture->transitionLayout(&commands, range, VulkanLayout::DEPTH_SAMPLER)) {
+                    texture->attachmentToSamplerBarrier(&commands, range);
+                }
+            } else {
+                texture->setLayout(range, VulkanFboCache::FINAL_COLOR_ATTACHMENT_LAYOUT);
+                if (!texture->transitionLayout(&commands, range, VulkanLayout::READ_WRITE)) {
+                    texture->attachmentToSamplerBarrier(&commands, range);
+                }
+            }
         }
     }
 
     mRenderPassFboInfo.clear();
-    mDescriptorSetManager.clearState();
     mCurrentRenderPass.renderTarget = nullptr;
     mCurrentRenderPass.renderPass = VK_NULL_HANDLE;
     FVK_SYSTRACE_END();
@@ -1557,33 +1575,6 @@ void VulkanDriver::commit(Handle<HwSwapChain> sch) {
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::bindUniformBuffer(uint32_t index, Handle<HwBufferObject> boh) {
-    auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    VkDeviceSize const offset = 0;
-    VkDeviceSize const size = VK_WHOLE_SIZE;
-    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
-}
-
-void VulkanDriver::bindBufferRange(BufferObjectBinding bindingType, uint32_t index,
-        Handle<HwBufferObject> boh, uint32_t offset, uint32_t size) {
-
-    assert_invariant(bindingType == BufferObjectBinding::UNIFORM);
-
-    // TODO: implement BufferObjectBinding::SHADER_STORAGE case
-
-    auto* bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
-    mDescriptorSetManager.updateBuffer({}, (uint32_t) index, bo, offset, size);
-}
-
-void VulkanDriver::unbindBuffer(BufferObjectBinding bindingType, uint32_t index) {
-    mDescriptorSetManager.clearBuffer((uint32_t) index);
-}
-
-void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
-    auto* hwsb = mResourceAllocator.handle_cast<VulkanSamplerGroup*>(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");
@@ -1592,13 +1583,13 @@ void VulkanDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
             value);
 }
 
-void VulkanDriver::insertEventMarker(char const* string, uint32_t len) {
+void VulkanDriver::insertEventMarker(char const* string) {
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands.insertEventMarker(string, len);
+    mCommands.insertEventMarker(string, strlen(string));
 #endif
 }
 
-void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
+void VulkanDriver::pushGroupMarker(char const* string) {
     // Turns out all the markers are 0-terminated, so we can just pass it without len.
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
     mCommands.pushGroupMarker(string);
@@ -1784,8 +1775,8 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
             *mResourceAllocator.handle_cast<VulkanVertexBufferInfo*>(pipelineState.vertexBufferInfo);
 
     Handle<HwProgram> programHandle = pipelineState.program;
-    RasterState rasterState = pipelineState.rasterState;
-    PolygonOffset depthOffset = pipelineState.polygonOffset;
+    RasterState const& rasterState = pipelineState.rasterState;
+    PolygonOffset const& depthOffset = pipelineState.polygonOffset;
 
     auto* program = mResourceAllocator.handle_cast<VulkanProgram*>(programHandle);
     commands->acquire(program);
@@ -1806,6 +1797,7 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
         .dstAlphaBlendFactor = getBlendFactor(rasterState.blendFunctionDstAlpha),
         .colorWriteMask = (VkColorComponentFlags) (rasterState.colorWrite ? 0xf : 0x0),
         .rasterizationSamples = rt->getSamples(),
+        .depthClamp = rasterState.depthClamp,
         .colorTargetCount = rt->getColorTargetCount(mCurrentRenderPass),
         .colorBlendOp = rasterState.blendEquationRGB,
         .alphaBlendOp =  rasterState.blendEquationAlpha,
@@ -1828,59 +1820,26 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
     mPipelineCache.bindPrimitiveTopology(topology);
     mPipelineCache.bindVertexArray(attribDesc, bufferDesc, vbi.getAttributeCount());
 
-    // Query the program for the mapping from (SamplerGroupBinding,Offset) to (SamplerBinding),
-    // where "SamplerBinding" is the integer in the GLSL, and SamplerGroupBinding is the abstract
-    // Filament concept used to form groups of samplers.
+    auto& setLayouts = pipelineState.pipelineLayout.setLayout;
+    VulkanDescriptorSetLayout::DescriptorSetLayoutArray layoutList;
+    uint8_t layoutCount = 0;
+    std::transform(setLayouts.begin(), setLayouts.end(), layoutList.begin(),
+            [&](Handle<HwDescriptorSetLayout> handle) -> VkDescriptorSetLayout {
+                if (!handle) {
+                    return VK_NULL_HANDLE;
+                }
+                auto layout = mResourceAllocator.handle_cast<VulkanDescriptorSetLayout*>(handle);
+                layoutCount++;
+                return layout->vklayout;
+            });
+    auto pipelineLayout = mPipelineLayoutCache.getLayout(layoutList, program);
 
-    auto const& bindingToSamplerIndex = program->getBindingToSamplerIndex();
-#if FVK_ENABLED_DEBUG_SAMPLER_NAME
-    auto const& bindingToName = program->getBindingToName();
-#endif
+    constexpr uint8_t descriptorSetMaskTable[4] = {0x1, 0x3, 0x7, 0xF};
 
-    for (auto binding: program->getBindings()) {
-        uint16_t const indexPair = bindingToSamplerIndex[binding];
-        if (indexPair == 0xffff) {
-            continue;
-        }
-
-        uint16_t const samplerGroupInd = (indexPair >> 8) & 0xff;
-        uint16_t const samplerInd = (indexPair & 0xff);
-
-        VulkanSamplerGroup* vksb = mSamplerBindings[samplerGroupInd];
-        if (!vksb) {
-            continue;
-        }
-        SamplerDescriptor const* boundSampler = ((SamplerDescriptor*) vksb->sb->data()) + samplerInd;
-
-        if (UTILS_UNLIKELY(!boundSampler->t)) {
-            continue;
-        }
-        VulkanTexture* texture = mResourceAllocator.handle_cast<VulkanTexture*>(boundSampler->t);
-
-        // TODO: can this uninitialized check be checked in a higher layer?
-        // This fallback path is very flaky because the dummy texture might not have
-        // matching characteristics. (e.g. if the missing texture is a 3D texture)
-        if (UTILS_UNLIKELY(texture->getPrimaryImageLayout() == VulkanLayout::UNDEFINED)) {
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE) && FVK_ENABLED_DEBUG_SAMPLER_NAME
-            FVK_LOGW << "Uninitialized texture bound to '" << bindingToName[binding] << "'";
-            FVK_LOGW << " in material '" << program->name.c_str() << "'";
-            FVK_LOGW << " at binding point " << +binding << utils::io::endl;
-#endif
-            texture = mEmptyTexture;
-        }
-
-        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
-        mDescriptorSetManager.updateSampler({}, binding, texture, vksampler);
-    }
-
-    auto const pipelineLayout = mDescriptorSetManager.bind(commands, program, mGetPipelineFunction);
     mBoundPipeline = {
         .program = program,
         .pipelineLayout = pipelineLayout,
+        .descriptorSetMask = DescriptorSetMask(descriptorSetMaskTable[layoutCount]),
     };
 
     mPipelineCache.bindLayout(pipelineLayout);
@@ -1923,6 +1882,14 @@ void VulkanDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
     FVK_SYSTRACE_END();
 }
 
+void VulkanDriver::bindDescriptorSet(
+        backend::DescriptorSetHandle dsh,
+        backend::descriptor_set_t setIndex,
+        backend::DescriptorSetOffsetArray&& offsets) {
+    VulkanDescriptorSet* set = mResourceAllocator.handle_cast<VulkanDescriptorSet*>(dsh);
+    mDescriptorSetManager.bind(setIndex, set, std::move(offsets));
+}
+
 void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("draw2");
@@ -1930,8 +1897,8 @@ void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
     VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer cmdbuffer = commands.buffer();
 
-    // Bind "dynamic" UBOs if they need to change.
-    mDescriptorSetManager.dynamicBind(&commands, {});
+    mDescriptorSetManager.commit(&commands, mBoundPipeline.pipelineLayout,
+            mBoundPipeline.descriptorSetMask);
 
     // Finally, make the actual draw call. TODO: support subranges
     const uint32_t firstIndex = indexOffset;
@@ -1979,7 +1946,7 @@ void VulkanDriver::scissor(Viewport scissorBox) {
             .extent = { uint32_t(r - l), uint32_t(t - b) }
     };
 
-    const VulkanRenderTarget* rt = mCurrentRenderPass.renderTarget;
+    VulkanRenderTarget const* rt = mCurrentRenderPass.renderTarget;
     rt->transformClientRectToPlatform(&scissor);
     vkCmdSetScissor(cmdbuffer, 0, 1, &scissor);
 }
@@ -2022,6 +1989,10 @@ void VulkanDriver::debugCommandBegin(CommandStream* cmds, bool synchronous, cons
 void VulkanDriver::resetState(int) {
 }
 
+void VulkanDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
+    mResourceAllocator.associateHandle(handleId, std::move(tag));
+}
+
 // explicit instantiation of the Dispatcher
 template class ConcreteDispatcher<VulkanDriver>;
 

filament/backend/src/vulkan/VulkanDriver.h

@@ -159,12 +159,11 @@ private:
     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;
+        DescriptorSetMask descriptorSetMask;
     };
     BoundPipeline mBoundPipeline = {};
     RenderPassFboBundle mRenderPassFboInfo;

filament/backend/src/vulkan/VulkanFboCache.cpp

@@ -43,6 +43,7 @@ bool VulkanFboCache::RenderPassEq::operator()(const RenderPassKey& k1,
     if (k1.samples != k2.samples) return false;
     if (k1.needsResolveMask != k2.needsResolveMask) return false;
     if (k1.subpassMask != k2.subpassMask) return false;
+    if (k1.viewCount != k2.viewCount) return false;
     return true;
 }
 
@@ -186,6 +187,25 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
         .pDependencies = dependencies
     };
 
+    VkRenderPassMultiviewCreateInfo multiviewCreateInfo = {};
+    uint32_t const subpassViewMask = (1 << config.viewCount) - 1;
+    // Prepare a view mask array for the maximum number of subpasses. All subpasses have all views
+    // activated.
+    uint32_t const viewMasks[2] = {subpassViewMask, subpassViewMask};
+    if (config.viewCount > 1) {
+        // Fill the multiview create info.
+        multiviewCreateInfo.sType = VK_STRUCTURE_TYPE_RENDER_PASS_MULTIVIEW_CREATE_INFO;
+        multiviewCreateInfo.pNext = nullptr;
+        multiviewCreateInfo.subpassCount = hasSubpasses ? 2u : 1u;
+        multiviewCreateInfo.pViewMasks = viewMasks;
+        multiviewCreateInfo.dependencyCount = 0;
+        multiviewCreateInfo.pViewOffsets = nullptr;
+        multiviewCreateInfo.correlationMaskCount = 1;
+        multiviewCreateInfo.pCorrelationMasks = &subpassViewMask;
+
+        renderPassInfo.pNext = &multiviewCreateInfo;
+    }
+
     int attachmentIndex = 0;
 
     // Populate the Color Attachments.

filament/backend/src/vulkan/VulkanFboCache.h

@@ -64,7 +64,7 @@ public:
         uint8_t samples; // 1 byte
         uint8_t needsResolveMask; // 1 byte
         uint8_t subpassMask; // 1 byte
-        uint8_t padding2; // 1 byte
+        uint8_t viewCount; // 1 byte
     };
     struct RenderPassVal {
         VkRenderPass handle;

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -16,13 +16,11 @@
 
 #include "VulkanHandles.h"
 
-#include "VulkanDriver.h"
 #include "VulkanConstants.h"
-#include "VulkanDriver.h"
 #include "VulkanMemory.h"
+#include "VulkanResourceAllocator.h"
 #include "VulkanUtility.h"
 #include "spirv/VulkanSpirvUtils.h"
-#include "utils/Log.h"
 
 #include <backend/platforms/VulkanPlatform.h>
 
@@ -54,54 +52,13 @@ void clampToFramebuffer(VkRect2D* rect, uint32_t fbWidth, uint32_t fbHeight) {
 }
 
 template<typename Bitmask>
-static constexpr Bitmask fromStageFlags(ShaderStageFlags2 flags, uint8_t binding) {
-    Bitmask ret = 0;
-    if (flags & ShaderStageFlags2::VERTEX) {
-        ret |= (getVertexStage<Bitmask>() << binding);
+inline void fromStageFlags(backend::ShaderStageFlags stage, descriptor_binding_t binding,
+        Bitmask& mask) {
+    if ((bool) (stage & ShaderStageFlags::VERTEX)) {
+        mask.set(binding + getVertexStageShift<Bitmask>());
     }
-    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);
+    if ((bool) (stage & ShaderStageFlags::FRAGMENT)) {
+        mask.set(binding + getFragmentStageShift<Bitmask>());
     }
 }
 
@@ -123,22 +80,113 @@ inline VkShaderStageFlags getVkStage(backend::ShaderStage stage) {
     }
 }
 
+inline VkDescriptorSetLayoutCreateInfo getLayoutCreateInfo(DescriptorSetLayout const& layout) {
+    // Note that the following *needs* to be static so that VkDescriptorSetLayoutCreateInfo will not
+    // refer to stack memory.
+    static VkDescriptorSetLayoutBinding toBind[VulkanDescriptorSetLayout::MAX_BINDINGS];
+    uint32_t count = 0;
+
+    for (auto const& binding: layout.bindings) {
+        VkShaderStageFlags stages = 0;
+        VkDescriptorType type;
+
+        if ((binding.stageFlags & ShaderStageFlags::VERTEX) != ShaderStageFlags::NONE) {
+            stages |= VK_SHADER_STAGE_VERTEX_BIT;
+        }
+        if ((binding.stageFlags & ShaderStageFlags::FRAGMENT) != ShaderStageFlags::NONE) {
+            stages |= VK_SHADER_STAGE_FRAGMENT_BIT;
+        }
+        assert_invariant(stages != 0);
+
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER: {
+                type = (binding.flags & DescriptorFlags::DYNAMIC_OFFSET) != DescriptorFlags::NONE
+                               ? VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
+                               : VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT: {
+                type = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
+                break;
+            }
+            case DescriptorType::SHADER_STORAGE_BUFFER:
+                PANIC_POSTCONDITION("Shader storage is not supported.");
+                break;
+        }
+        toBind[count++] = {
+                .binding = binding.binding,
+                .descriptorType = type,
+                .descriptorCount = 1,
+                .stageFlags = stages,
+        };
+    }
+
+    assert_invariant(count != 0 && "Need at least one binding for descriptor set layout.");
+    VkDescriptorSetLayoutCreateInfo dlinfo = {
+            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+            .pNext = nullptr,
+            .bindingCount = count,
+            .pBindings = toBind,
+    };
+    return dlinfo;
+}
+
+using BitmaskGroup = VulkanDescriptorSetLayout::Bitmask;
+BitmaskGroup fromBackendLayout(DescriptorSetLayout const& layout) {
+    BitmaskGroup mask;
+    for (auto const& binding: layout.bindings) {
+        switch (binding.type) {
+            case DescriptorType::UNIFORM_BUFFER: {
+                if ((binding.flags & DescriptorFlags::DYNAMIC_OFFSET) != DescriptorFlags::NONE) {
+                    fromStageFlags(binding.stageFlags, binding.binding, mask.dynamicUbo);
+                } else {
+                    fromStageFlags(binding.stageFlags, binding.binding, mask.ubo);
+                }
+                break;
+            }
+            case DescriptorType::SAMPLER: {
+                fromStageFlags(binding.stageFlags, binding.binding, mask.sampler);
+                break;
+            }
+            case DescriptorType::INPUT_ATTACHMENT: {
+                fromStageFlags(binding.stageFlags, binding.binding, mask.inputAttachment);
+                break;
+            }
+            case DescriptorType::SHADER_STORAGE_BUFFER:
+                PANIC_POSTCONDITION("Shader storage is not supported");
+                break;
+        }
+    }
+    return mask;
+}
+
 } // 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)) {}
+        DescriptorSetLayout const& layout)
+    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET_LAYOUT), mDevice(device),
+      vklayout(createDescriptorSetLayout(device, getLayoutCreateInfo(layout))),
+      bitmask(fromBackendLayout(layout)),
+      count(Count::fromLayoutBitmask(bitmask)) {
+}
 
 VulkanDescriptorSetLayout::~VulkanDescriptorSetLayout() {
     vkDestroyDescriptorSetLayout(mDevice, vklayout, VKALLOC);
 }
 
+void VulkanDescriptorSet::acquire(VulkanTexture* texture) {
+    mResources.acquire(texture);
+    mTextures[mTextureCount++] = texture;
+}
+
+void VulkanDescriptorSet::acquire(VulkanBufferObject* bufferObject) {
+    mResources.acquire(bufferObject);
+}
+
 PushConstantDescription::PushConstantDescription(backend::Program const& program) noexcept {
     mRangeCount = 0;
     for (auto stage : { ShaderStage::VERTEX, ShaderStage::FRAGMENT, ShaderStage::COMPUTE }) {
@@ -190,24 +238,11 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
       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);
@@ -224,12 +259,6 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
             dataSize = shader.size() * 4;
         }
 
-        auto const [ubo, sampler, inputAttachment] = getProgramBindings(blob);
-        uboMask |= (static_cast<UniformBufferBitmask>(ubo) << (UBO_MODULE_OFFSET * i));
-        samplerMask |= (static_cast<SamplerBitmask>(sampler) << (SAMPLER_MODULE_OFFSET * i));
-        inputAttachmentMask |= (static_cast<InputAttachmentBitmask>(inputAttachment)
-                                << (INPUT_ATTACHMENT_MODULE_OFFSET * i));
-
         VkShaderModule& module = modules[i];
         VkShaderModuleCreateInfo moduleInfo = {
             .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
@@ -257,40 +286,6 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
     }
 
-    LayoutDescriptionList& layouts = mInfo->layouts;
-    layouts[0].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
-            countBits(collapseStages(uboMask)));
-    layouts[1].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
-            countBits(collapseStages(samplerMask)));
-    layouts[2].bindings = utils::FixedCapacityVector<DescriptorSetLayoutBinding>::with_capacity(
-            countBits(collapseStages(inputAttachmentMask)));
-
-    addDescriptors(uboMask, layouts[0].bindings);
-    addDescriptors(samplerMask, layouts[1].bindings);
-    addDescriptors(inputAttachmentMask, layouts[2].bindings);
-
-#if FVK_ENABLED_DEBUG_SAMPLER_NAME
-    auto& bindingToName = mInfo->bindingToName;
-#endif
-
-    auto& groupInfo = builder.getSamplerGroupInfo();
-    auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
-    auto& bindings = mInfo->bindings;
-    for (uint8_t groupInd = 0; groupInd < Program::SAMPLER_BINDING_COUNT; groupInd++) {
-        auto const& group = groupInfo[groupInd];
-        auto const& samplers = group.samplers;
-        for (size_t i = 0; i < samplers.size(); ++i) {
-            uint32_t const binding = samplers[i].binding;
-            bindingToSamplerIndex[binding] = (groupInd << 8) | (0xff & i);
-            assert_invariant(bindings.find(binding) == bindings.end());
-            bindings.insert(binding);
-
-#if FVK_ENABLED_DEBUG_SAMPLER_NAME
-            bindingToName[binding] = samplers[i].name.c_str();
-#endif
-        }
-    }
-
 #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
     FVK_LOGD << "Created VulkanProgram " << builder << ", shaders = (" << modules[0]
              << ", " << modules[1] << ")" << utils::io::endl;
@@ -321,13 +316,15 @@ void VulkanRenderTarget::bindToSwapChain(VulkanSwapChain& swapChain) {
 
 VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physicalDevice,
         VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
+        VulkanResourceAllocator* handleAllocator,
         uint32_t width, uint32_t height, uint8_t samples,
         VulkanAttachment color[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
-        VulkanAttachment depthStencil[2], VulkanStagePool& stagePool)
+        VulkanAttachment depthStencil[2], VulkanStagePool& stagePool, uint8_t layerCount)
     : HwRenderTarget(width, height),
       VulkanResource(VulkanResourceType::RENDER_TARGET),
       mOffscreen(true),
-      mSamples(samples) {
+      mSamples(samples),
+      mLayerCount(layerCount) {
     for (int index = 0; index < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; index++) {
         mColor[index] = color[index];
     }
@@ -353,6 +350,7 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica
             if (UTILS_UNLIKELY(!msTexture)) {
                 // TODO: This should be allocated with the ResourceAllocator.
                 msTexture = new VulkanTexture(device, physicalDevice, context, allocator, commands,
+                        handleAllocator,
                         texture->target, ((VulkanTexture const*) texture)->levels, texture->format,
                         samples, texture->width, texture->height, texture->depth, texture->usage,
                         stagePool, true /* heap allocated */);
@@ -382,7 +380,8 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica
     VulkanTexture* msTexture = depthTexture->getSidecar();
     if (UTILS_UNLIKELY(!msTexture)) {
         msTexture = new VulkanTexture(device, physicalDevice, context, allocator,
-                commands, depthTexture->target, msLevel, depthTexture->format, samples,
+                commands, handleAllocator,
+                depthTexture->target, msLevel, depthTexture->format, samples,
                 depthTexture->width, depthTexture->height, depthTexture->depth, depthTexture->usage,
                 stagePool, true /* heap allocated */);
         depthTexture->setSidecar(msTexture);
@@ -534,15 +533,7 @@ bool VulkanTimerQuery::isCompleted() noexcept {
     // timestamp has at least been written into a processed command buffer.
 
     // This fence indicates that the corresponding buffer has been completed.
-    if (!mFence) {
-        return false;
-    }
-    VkResult status = mFence->status.load(std::memory_order_relaxed);
-    if (status != VK_SUCCESS) {
-        return false;
-    }
-
-    return true;
+    return mFence && mFence->getStatus() == VK_SUCCESS;
 }
 
 VulkanTimerQuery::~VulkanTimerQuery() = default;
@@ -558,35 +549,4 @@ VulkanRenderPrimitive::VulkanRenderPrimitive(VulkanResourceAllocator* resourceAl
     mResources.acquire(indexBuffer);
 }
 
-using Bitmask = VulkanDescriptorSetLayout::Bitmask;
-
-Bitmask Bitmask::fromBackendLayout(descset::DescriptorSetLayout const& layout) {
-    Bitmask mask;
-    for (auto const& binding: layout.bindings) {
-        switch (binding.type) {
-            case descset::DescriptorType::UNIFORM_BUFFER: {
-                if (binding.flags == descset::DescriptorFlags::DYNAMIC_OFFSET) {
-                    mask.dynamicUbo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
-                            binding.binding);
-                } else {
-                    mask.ubo |= fromStageFlags<UniformBufferBitmask>(binding.stageFlags,
-                            binding.binding);
-                }
-                break;
-            }
-            case descset::DescriptorType::SAMPLER: {
-                mask.sampler |= fromStageFlags<SamplerBitmask>(binding.stageFlags, binding.binding);
-                break;
-            }
-            case descset::DescriptorType::INPUT_ATTACHMENT: {
-                mask.inputAttachment |=
-                        fromStageFlags<InputAttachmentBitmask>(binding.stageFlags, binding.binding);
-                break;
-            }
-        }
-    }
-    return mask;
-}
-
-
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanHandles.h

@@ -29,39 +29,52 @@
 #include <private/backend/SamplerGroup.h>
 #include <backend/Program.h>
 
+#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Mutex.h>
 #include <utils/StructureOfArrays.h>
 
 namespace filament::backend {
 
-using namespace descset;
+namespace {
+// Counts the total number of descriptors for both vertex and fragment stages.
+template<typename Bitmask>
+inline uint8_t collapsedCount(Bitmask const& mask) {
+    static_assert(sizeof(mask) <= 64);
+    constexpr uint64_t VERTEX_MASK = (1ULL << getVertexStageShift<Bitmask>()) - 1ULL;
+    constexpr uint64_t FRAGMENT_MASK = (1ULL << getFragmentStageShift<Bitmask>()) - 1ULL;
+    uint64_t val = mask.getValue();
+    val = ((val | VERTEX_MASK) >> getVertexStageShift<Bitmask>()) |
+        ((val | FRAGMENT_MASK) >> getFragmentStageShift<Bitmask>());
+    return (uint8_t) Bitmask(val).count();
+}
+
+} // anonymous namespace
 
 class VulkanTimestamps;
+struct VulkanBufferObject;
 
-struct VulkanDescriptorSetLayout : public VulkanResource {
-    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 3;
+struct VulkanDescriptorSetLayout : public VulkanResource, HwDescriptorSetLayout {
+    static constexpr uint8_t UNIQUE_DESCRIPTOR_SET_COUNT = 4;
+    static constexpr uint8_t MAX_BINDINGS = 25;
+
+    using DescriptorSetLayoutArray = std::array<VkDescriptorSetLayout,
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
 
     // 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
+        // TODO: better utiltize the space below and use bitset instead.
+        UniformBufferBitmask ubo;         // 8 bytes
+        UniformBufferBitmask dynamicUbo;  // 8 bytes
+        SamplerBitmask sampler;           // 8 bytes
+        InputAttachmentBitmask inputAttachment; // 8 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);
     };
+    static_assert(sizeof(Bitmask) == 32);
 
     // This is a convenience struct to quickly check layout compatibility in terms of descriptor set
     // pools.
@@ -71,6 +84,10 @@ struct VulkanDescriptorSetLayout : public VulkanResource {
         uint32_t sampler = 0;
         uint32_t inputAttachment = 0;
 
+        inline uint32_t total() const {
+            return ubo + dynamicUbo + sampler + inputAttachment;
+        }
+
         bool operator==(Count const& right) const noexcept {
             return ubo == right.ubo && dynamicUbo == right.dynamicUbo && sampler == right.sampler &&
                    inputAttachment == right.inputAttachment;
@@ -78,10 +95,10 @@ struct VulkanDescriptorSetLayout : public VulkanResource {
 
         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)),
+                .ubo = collapsedCount(mask.ubo),
+                .dynamicUbo = collapsedCount(mask.dynamicUbo),
+                .sampler = collapsedCount(mask.sampler),
+                .inputAttachment = collapsedCount(mask.inputAttachment),
             };
         }
 
@@ -97,70 +114,27 @@ struct VulkanDescriptorSetLayout : public VulkanResource {
         }
     };
 
-    static_assert(sizeof(Bitmask) % 8 == 0);
-
-    explicit VulkanDescriptorSetLayout(VkDevice device, VkDescriptorSetLayoutCreateInfo const& info,
-            Bitmask const& bitmask);
+    VulkanDescriptorSetLayout(VkDevice device, DescriptorSetLayout const& layout);
 
     ~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 {
+struct VulkanDescriptorSet : public VulkanResource, HwDescriptorSet {
 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)
+    VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+            OnRecycle&& onRecycleFn)
         : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
-          resources(allocator),
           vkSet(rawSet),
+          mResources(allocator),
           mOnRecycleFn(std::move(onRecycleFn)) {}
 
     ~VulkanDescriptorSet() {
@@ -169,17 +143,25 @@ public:
         }
     }
 
+    void acquire(VulkanTexture* texture);
+
+    void acquire(VulkanBufferObject* texture);
+
+    bool hasTexture(VulkanTexture* texture) {
+        return std::any_of(mTextures.begin(), mTextures.end(),
+                [texture](auto t) { return t == texture; });
+    }
+
     // TODO: maybe change to fixed size for performance.
-    VulkanAcquireOnlyResourceManager resources;
     VkDescriptorSet const vkSet;
 
 private:
+    std::array<VulkanTexture*, 16> mTextures = { nullptr };
+    uint8_t mTextureCount = 0;
+    VulkanAcquireOnlyResourceManager mResources;
     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>;
 
@@ -224,16 +206,6 @@ struct VulkanProgram : public HwProgram, VulkanResource {
     // 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();
     }
@@ -273,10 +245,6 @@ private:
         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
@@ -302,9 +270,11 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
     // Creates an offscreen render target.
     VulkanRenderTarget(VkDevice device, VkPhysicalDevice physicalDevice,
             VulkanContext const& context, VmaAllocator allocator,
-            VulkanCommands* commands, uint32_t width, uint32_t height,
+            VulkanCommands* commands,
+            VulkanResourceAllocator* handleAllocator,
+            uint32_t width, uint32_t height,
             uint8_t samples, VulkanAttachment color[MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT],
-            VulkanAttachment depthStencil[2], VulkanStagePool& stagePool);
+            VulkanAttachment depthStencil[2], VulkanStagePool& stagePool, uint8_t layerCount);
 
     // Creates a special "default" render target (i.e. associated with the swap chain)
     explicit VulkanRenderTarget();
@@ -319,6 +289,7 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
     VulkanAttachment& getMsaaDepth();
     uint8_t getColorTargetCount(const VulkanRenderPass& pass) const;
     uint8_t getSamples() const { return mSamples; }
+    uint8_t getLayerCount() const { return mLayerCount; }
     bool hasDepth() const { return mDepth.texture; }
     bool isSwapChain() const { return !mOffscreen; }
     void bindToSwapChain(VulkanSwapChain& surf);
@@ -330,6 +301,7 @@ private:
     VulkanAttachment mMsaaDepthAttachment = {};
     const bool mOffscreen : 1;
     uint8_t mSamples : 7;
+    uint8_t mLayerCount = 1;
 };
 
 struct VulkanBufferObject;
@@ -482,6 +454,7 @@ private:
     utils::Mutex mFenceMutex;
 };
 
+
 inline constexpr VkBufferUsageFlagBits getBufferObjectUsage(
         BufferObjectBinding bindingType) noexcept {
     switch(bindingType) {

filament/backend/src/vulkan/VulkanImageUtility.cpp

@@ -131,14 +131,18 @@ getVkTransition(const VulkanLayoutTransition& transition) {
 
 }// anonymous namespace
 
-void transitionLayout(VkCommandBuffer cmdbuffer,
+bool transitionLayout(VkCommandBuffer cmdbuffer,
         VulkanLayoutTransition transition) {
     if (transition.oldLayout == transition.newLayout) {
-        return;
+        return false;
     }
     auto [srcAccessMask, dstAccessMask, srcStage, dstStage, oldLayout, newLayout]
             = getVkTransition(transition);
 
+    if (oldLayout == newLayout) {
+        return false;
+    }
+
     assert_invariant(transition.image != VK_NULL_HANDLE && "No image for transition");
     VkImageMemoryBarrier barrier = {
             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
@@ -152,6 +156,7 @@ void transitionLayout(VkCommandBuffer cmdbuffer,
             .subresourceRange = transition.subresources,
     };
     vkCmdPipelineBarrier(cmdbuffer, srcStage, dstStage, 0, 0, nullptr, 0, nullptr, 1, &barrier);
+    return true;
 }
 
 }// namespace filament::backend

filament/backend/src/vulkan/VulkanImageUtility.h

@@ -135,7 +135,9 @@ constexpr inline VkImageLayout getVkLayout(VulkanLayout layout) {
     }
 }
 
-void transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
+// Returns true if a transition has been added to the command buffer, false otherwis (where there is
+// no transition necessary).
+bool transitionLayout(VkCommandBuffer cmdbuffer, VulkanLayoutTransition transition);
 
 } // namespace imgutil
 

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -98,7 +98,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     VkPipelineColorBlendStateCreateInfo colorBlendState;
     colorBlendState = VkPipelineColorBlendStateCreateInfo{};
     colorBlendState.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO;
-    colorBlendState.attachmentCount = 1;
+    colorBlendState.attachmentCount = mPipelineRequirements.rasterState.colorTargetCount;    
     colorBlendState.pAttachments = colorBlendAttachments;
 
     // If we reach this point, we need to create and stash a brand new pipeline object.
@@ -184,6 +184,7 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     vkRaster.polygonMode = VK_POLYGON_MODE_FILL;
     vkRaster.cullMode = raster.cullMode;
     vkRaster.frontFace = raster.frontFace;
+    vkRaster.depthClampEnable = raster.depthClamp;
     vkRaster.depthBiasEnable = raster.depthBiasEnable;
     vkRaster.depthBiasConstantFactor = raster.depthBiasConstantFactor;
     vkRaster.depthBiasClamp = 0.0f;
@@ -209,8 +210,8 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     pipelineCreateInfo.pDynamicState = &dynamicState;
 
     // Filament assumes consistent blend state across all color attachments.
-    colorBlendState.attachmentCount = mPipelineRequirements.rasterState.colorTargetCount;
-    for (auto& target : colorBlendAttachments) {
+    for (uint8_t i = 0; i < colorBlendState.attachmentCount; ++i) {
+        auto& target = colorBlendAttachments[i];
         target.blendEnable = mPipelineRequirements.rasterState.blendEnable;
         target.srcColorBlendFactor = mPipelineRequirements.rasterState.srcColorBlendFactor;
         target.dstColorBlendFactor = mPipelineRequirements.rasterState.dstColorBlendFactor;

filament/backend/src/vulkan/VulkanPipelineCache.h

@@ -90,7 +90,8 @@ public:
         VkBlendFactor         srcAlphaBlendFactor : 5;
         VkBlendFactor         dstAlphaBlendFactor : 5;
         VkColorComponentFlags colorWriteMask : 4;
-        uint8_t               rasterizationSamples;    // offset = 4 bytes
+        uint8_t               rasterizationSamples : 4;// offset = 4 bytes
+        uint8_t               depthClamp : 4;
         uint8_t               colorTargetCount;        // offset = 5 bytes
         BlendEquation         colorBlendOp : 4;        // offset = 6 bytes
         BlendEquation         alphaBlendOp : 4;

filament/backend/src/vulkan/VulkanReadPixels.cpp

@@ -232,7 +232,7 @@ void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint
 
     VulkanAttachment const srcAttachment = srcTarget->getColor(0);
     VkImageSubresourceRange const srcRange = srcAttachment.getSubresourceRange();
-    srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::TRANSFER_SRC);
+    srcTexture->transitionLayout(cmdbuffer, {}, srcRange, VulkanLayout::TRANSFER_SRC);
 
     VkImageCopy const imageCopyRegion = {
         .srcSubresource = {
@@ -268,7 +268,7 @@ void VulkanReadPixels::run(VulkanRenderTarget* srcTarget, uint32_t const x, uint
             imgutil::getVkLayout(VulkanLayout::TRANSFER_DST), 1, &imageCopyRegion);
 
     // Restore the source image layout.
-    srcTexture->transitionLayout(cmdbuffer, srcRange, VulkanLayout::COLOR_ATTACHMENT);
+    srcTexture->transitionLayout(cmdbuffer, {}, srcRange, VulkanLayout::COLOR_ATTACHMENT);
 
     vkEndCommandBuffer(cmdbuffer);
 

filament/backend/src/vulkan/VulkanResourceAllocator.h

@@ -105,6 +105,10 @@ public:
         mHandleAllocatorImpl.deallocate(handle, obj);
     }
 
+    inline void associateHandle(HandleBase::HandleId id, utils::CString&& tag) noexcept {
+        mHandleAllocatorImpl.associateTagToHandle(id, std::move(tag));
+    }
+
 private:
     AllocatorImpl mHandleAllocatorImpl;
 

filament/backend/src/vulkan/VulkanSwapChain.cpp

@@ -26,12 +26,14 @@ using namespace utils;
 namespace filament::backend {
 
 VulkanSwapChain::VulkanSwapChain(VulkanPlatform* platform, VulkanContext const& context,
-        VmaAllocator allocator, VulkanCommands* commands, VulkanStagePool& stagePool,
+        VmaAllocator allocator, VulkanCommands* commands, VulkanResourceAllocator* handleAllocator,
+        VulkanStagePool& stagePool,
         void* nativeWindow, uint64_t flags, VkExtent2D extent)
     : VulkanResource(VulkanResourceType::SWAP_CHAIN),
       mPlatform(platform),
       mCommands(commands),
       mAllocator(allocator),
+      mHandleAllocator(handleAllocator),
       mStagePool(stagePool),
       mHeadless(extent.width != 0 && extent.height != 0 && !nativeWindow),
       mFlushAndWaitOnResize(platform->getCustomization().flushAndWaitOnWindowResize),
@@ -62,12 +64,12 @@ void VulkanSwapChain::update() {
     VkDevice const device = mPlatform->getDevice();
 
     for (auto const color: bundle.colors) {
-        mColors.push_back(std::make_unique<VulkanTexture>(device, mAllocator, mCommands, color,
-                bundle.colorFormat, 1, bundle.extent.width, bundle.extent.height,
+        mColors.push_back(std::make_unique<VulkanTexture>(device, mAllocator, mCommands, mHandleAllocator,
+                color, bundle.colorFormat, 1, bundle.extent.width, bundle.extent.height,
                 TextureUsage::COLOR_ATTACHMENT, mStagePool, true /* heap allocated */));
     }
-    mDepth = std::make_unique<VulkanTexture>(device, mAllocator, mCommands, bundle.depth,
-            bundle.depthFormat, 1, bundle.extent.width, bundle.extent.height,
+    mDepth = std::make_unique<VulkanTexture>(device, mAllocator, mCommands, mHandleAllocator,
+            bundle.depth, bundle.depthFormat, 1, bundle.extent.width, bundle.extent.height,
             TextureUsage::DEPTH_ATTACHMENT, mStagePool, true /* heap allocated */);
 
     mExtent = bundle.extent;
@@ -75,7 +77,7 @@ void VulkanSwapChain::update() {
 
 void VulkanSwapChain::present() {
     if (!mHeadless && mTransitionSwapChainImageLayoutForPresent) {
-        VkCommandBuffer const cmdbuf = mCommands->get().buffer();
+        VulkanCommandBuffer& commands = mCommands->get();
         VkImageSubresourceRange const subresources{
                 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
                 .baseMipLevel = 0,
@@ -83,7 +85,7 @@ void VulkanSwapChain::present() {
                 .baseArrayLayer = 0,
                 .layerCount = 1,
         };
-        mColors[mCurrentSwapIndex]->transitionLayout(cmdbuf, subresources, VulkanLayout::PRESENT);
+        mColors[mCurrentSwapIndex]->transitionLayout(&commands, subresources, VulkanLayout::PRESENT);
     }
 
     mCommands->flush();

filament/backend/src/vulkan/VulkanSwapChain.h

@@ -36,11 +36,13 @@ namespace filament::backend {
 
 struct VulkanHeadlessSwapChain;
 struct VulkanSurfaceSwapChain;
+class VulkanResourceAllocator;
 
 // A wrapper around the platform implementation of swapchain.
 struct VulkanSwapChain : public HwSwapChain, VulkanResource {
     VulkanSwapChain(VulkanPlatform* platform, VulkanContext const& context, VmaAllocator allocator,
-            VulkanCommands* commands, VulkanStagePool& stagePool,
+            VulkanCommands* commands, VulkanResourceAllocator* handleAllocator,
+            VulkanStagePool& stagePool,
             void* nativeWindow, uint64_t flags, VkExtent2D extent = {0, 0});
 
     ~VulkanSwapChain();
@@ -80,6 +82,7 @@ private:
     VulkanPlatform* mPlatform;
     VulkanCommands* mCommands;
     VmaAllocator mAllocator;
+    VulkanResourceAllocator* const mHandleAllocator;
     VulkanStagePool& mStagePool;
     bool const mHeadless;
     bool const mFlushAndWaitOnResize;

filament/backend/src/vulkan/VulkanTexture.cpp

@@ -15,6 +15,7 @@
  */
 
 #include "VulkanMemory.h"
+#include "VulkanResourceAllocator.h"
 #include "VulkanTexture.h"
 #include "VulkanUtility.h"
 
@@ -28,50 +29,138 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-VulkanTexture::VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
+namespace {
+
+inline uint8_t getLayerCount(SamplerType const target, uint32_t const depth) {
+    switch (target) {
+        case SamplerType::SAMPLER_2D:
+        case SamplerType::SAMPLER_3D:
+        case SamplerType::SAMPLER_EXTERNAL:
+            return 1;
+        case SamplerType::SAMPLER_CUBEMAP:
+            return 6;
+        case SamplerType::SAMPLER_CUBEMAP_ARRAY:
+            return depth * 6;
+        case SamplerType::SAMPLER_2D_ARRAY:
+            return depth;
+    }
+}
+
+VkComponentMapping composeSwizzle(VkComponentMapping const& prev, VkComponentMapping const& next) {
+    static constexpr VkComponentSwizzle IDENTITY[] = {
+        VK_COMPONENT_SWIZZLE_R,
+        VK_COMPONENT_SWIZZLE_G,
+        VK_COMPONENT_SWIZZLE_B,
+        VK_COMPONENT_SWIZZLE_A,
+    };
+
+    auto const compose = [](VkComponentSwizzle out, VkComponentMapping const& prev,
+                                 uint8_t channelIndex) {
+        // We need to first change all identities to its equivalent channel.
+        if (out == VK_COMPONENT_SWIZZLE_IDENTITY) {
+            out = IDENTITY[channelIndex];
+        }
+        switch (out) {
+            case VK_COMPONENT_SWIZZLE_R:
+                out = prev.r;
+                break;
+            case VK_COMPONENT_SWIZZLE_G:
+                out = prev.g;
+                break;
+            case VK_COMPONENT_SWIZZLE_B:
+                out = prev.b;
+                break;
+            case VK_COMPONENT_SWIZZLE_A:
+                out = prev.a;
+                break;
+            case VK_COMPONENT_SWIZZLE_IDENTITY:
+            case VK_COMPONENT_SWIZZLE_ZERO:
+            case VK_COMPONENT_SWIZZLE_ONE:
+                return out;
+            // Below is not exposed in Vulkan's API, but needs to be there for compilation.
+            case VK_COMPONENT_SWIZZLE_MAX_ENUM:
+                break;
+        }
+        // If the result correctly corresponds to the identity, just return identity.
+        if (IDENTITY[channelIndex] == out) {
+            return VK_COMPONENT_SWIZZLE_IDENTITY;
+        }
+        return out;
+    };
+
+    // Note that the channel index corresponds to the VkComponentMapping struct layout.
+    return {
+        compose(next.r, prev, 0),
+        compose(next.g, prev, 1),
+        compose(next.b, prev, 2),
+        compose(next.a, prev, 3),
+    };
+}
+
+} // anonymous namespace
+
+VulkanTextureState::VulkanTextureState(
+        VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
+        VulkanStagePool& stagePool,
+        VkFormat format, VkImageViewType viewType, uint8_t levels, uint8_t layerCount)
+        : VulkanResource(VulkanResourceType::HEAP_ALLOCATED),
+          mVkFormat(format),
+          mViewType(viewType),
+          mFullViewRange {
+                  filament::backend::getImageAspect(format), 0, levels, 0, layerCount
+          },
+          mStagePool(stagePool),
+          mDevice(device),
+          mAllocator(allocator),
+          mCommands(commands) {
+}
+
+VulkanTextureState* VulkanTexture::getSharedState() {
+    VulkanTextureState* state = mAllocator->handle_cast<VulkanTextureState*>(mState);
+    return state;
+}
+
+VulkanTextureState const* VulkanTexture::getSharedState() const {
+    VulkanTextureState const* state = mAllocator->handle_cast<VulkanTextureState const*>(mState);
+    return state;
+}
+
+VulkanTexture::VulkanTexture(
+        VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
+        VulkanResourceAllocator* handleAllocator,
         VkImage image, VkFormat format, uint8_t samples, uint32_t width, uint32_t height,
         TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
-    : HwTexture(SamplerType::SAMPLER_2D, 1, samples, width, height, 1, TextureFormat::UNUSED,
-            tusage),
-      VulkanResource(
-              heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
-      mVkFormat(format),
-      mViewType(imgutil::getViewType(target)),
-      mSwizzle({}),
-      mTextureImage(image),
-      mFullViewRange{
-              .aspectMask = getImageAspect(),
-              .baseMipLevel = 0,
-              .levelCount = 1,
-              .baseArrayLayer = 0,
-              .layerCount = 1,
-      },
-      mPrimaryViewRange(mFullViewRange),
-      mStagePool(stagePool),
-      mDevice(device),
-      mAllocator(allocator),
-      mCommands(commands) {}
+        : HwTexture(SamplerType::SAMPLER_2D, 1, samples, width, height, 1, TextureFormat::UNUSED,
+                tusage),
+          VulkanResource(
+                  heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
+          mAllocator(handleAllocator),
+          mState(handleAllocator->initHandle<VulkanTextureState>(
+                  device, allocator, commands, stagePool,
+                  format, imgutil::getViewType(SamplerType::SAMPLER_2D), 1, 1)) {
+    auto* const state = getSharedState();
+    state->mTextureImage = image;
+    mPrimaryViewRange = state->mFullViewRange;
+}
 
 VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
-        SamplerType target, uint8_t levels, TextureFormat tformat, uint8_t samples, uint32_t w,
-        uint32_t h, uint32_t depth, TextureUsage tusage, VulkanStagePool& stagePool,
-        bool heapAllocated, VkComponentMapping swizzle)
+        VulkanResourceAllocator* handleAllocator, SamplerType target, uint8_t levels,
+        TextureFormat tformat, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
+        TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated)
     : HwTexture(target, levels, samples, w, h, depth, tformat, tusage),
       VulkanResource(
               heapAllocated ? VulkanResourceType::HEAP_ALLOCATED : VulkanResourceType::TEXTURE),
-      mVkFormat(backend::getVkFormat(tformat)),
-      mViewType(imgutil::getViewType(target)),
-      mSwizzle(swizzle),
-      mStagePool(stagePool),
-      mDevice(device),
-      mAllocator(allocator),
-      mCommands(commands) {
+      mAllocator(handleAllocator),
+      mState(handleAllocator->initHandle<VulkanTextureState>(device, allocator, commands, stagePool,
+              backend::getVkFormat(tformat), imgutil::getViewType(target), levels,
+              getLayerCount(target, depth))) {
+    auto* const state = getSharedState();
 
     // Create an appropriately-sized device-only VkImage, but do not fill it yet.
     VkImageCreateInfo imageInfo{.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
             .imageType = target == SamplerType::SAMPLER_3D ? VK_IMAGE_TYPE_3D : VK_IMAGE_TYPE_2D,
-            .format = mVkFormat,
+            .format = state->mVkFormat,
             .extent = {w, h, depth},
             .mipLevels = levels,
             .arrayLayers = 1,
@@ -97,13 +186,11 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         imageInfo.extent.depth = 1;
     }
 
-    // Filament expects blit() to work with any texture, so we almost always set these usage flags.
-    // TODO: investigate performance implications of setting these flags.
-    constexpr VkImageUsageFlags blittable = VK_IMAGE_USAGE_TRANSFER_DST_BIT |
-            VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
-
-    if (any(usage & (TextureUsage::BLIT_DST | TextureUsage::BLIT_SRC))) {
-        imageInfo.usage |= blittable;
+    if (any(usage & TextureUsage::BLIT_SRC)) {
+        imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+    }
+    if (any(usage & TextureUsage::BLIT_DST)) {
+        imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
     }
 
     if (any(usage & TextureUsage::SAMPLEABLE)) {
@@ -111,9 +198,9 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
         // Validate that the format is actually sampleable.
         VkFormatProperties props;
-        vkGetPhysicalDeviceFormatProperties(physicalDevice, mVkFormat, &props);
+        vkGetPhysicalDeviceFormatProperties(physicalDevice, state->mVkFormat, &props);
         if (!(props.optimalTilingFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT)) {
-            FVK_LOGW << "Texture usage is SAMPLEABLE but format " << mVkFormat << " is not "
+            FVK_LOGW << "Texture usage is SAMPLEABLE but format " << state->mVkFormat << " is not "
                     "sampleable with optimal tiling." << utils::io::endl;
         }
 #endif
@@ -121,7 +208,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         imageInfo.usage |= VK_IMAGE_USAGE_SAMPLED_BIT;
     }
     if (any(usage & TextureUsage::COLOR_ATTACHMENT)) {
-        imageInfo.usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | blittable;
+        imageInfo.usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
         if (any(usage & TextureUsage::SUBPASS_INPUT)) {
             imageInfo.usage |= VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT;
         }
@@ -130,10 +217,9 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         imageInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
     }
     if (any(usage & TextureUsage::UPLOADABLE)) {
-        imageInfo.usage |= blittable;
+        imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
     }
     if (any(usage & TextureUsage::DEPTH_ATTACHMENT)) {
-        imageInfo.usage |= blittable;
         imageInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
 
         // Depth resolves uses a custom shader and therefore needs to be sampleable.
@@ -147,7 +233,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
     // any kind of attachment (color or depth).
     const auto& limits = context.getPhysicalDeviceLimits();
     if (imageInfo.usage & VK_IMAGE_USAGE_SAMPLED_BIT) {
-        samples = reduceSampleCount(samples, isVkDepthFormat(mVkFormat)
+        samples = reduceSampleCount(samples, isVkDepthFormat(state->mVkFormat)
                                                      ? limits.sampledImageDepthSampleCounts
                                                      : limits.sampledImageColorSampleCounts);
     }
@@ -161,12 +247,12 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
     this->samples = samples;
     imageInfo.samples = (VkSampleCountFlagBits) samples;
 
-    VkResult error = vkCreateImage(mDevice, &imageInfo, VKALLOC, &mTextureImage);
+    VkResult error = vkCreateImage(state->mDevice, &imageInfo, VKALLOC, &state->mTextureImage);
     if (error || FVK_ENABLED(FVK_DEBUG_TEXTURE)) {
         FVK_LOGD << "vkCreateImage: "
-            << "image = " << mTextureImage << ", "
+            << "image = " << state->mTextureImage << ", "
             << "result = " << error << ", "
-            << "handle = " << utils::io::hex << mTextureImage << utils::io::dec << ", "
+            << "handle = " << utils::io::hex << state->mTextureImage << utils::io::dec << ", "
             << "extent = " << w << "x" << h << "x"<< depth << ", "
             << "mipLevels = " << int(levels) << ", "
             << "TextureUsage = " << static_cast<int>(usage) << ", "
@@ -175,13 +261,13 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
             << "type = " << imageInfo.imageType << ", "
             << "flags = " << imageInfo.flags << ", "
             << "target = " << static_cast<int>(target) <<", "
-            << "format = " << mVkFormat << utils::io::endl;
+            << "format = " << state->mVkFormat << utils::io::endl;
     }
     FILAMENT_CHECK_POSTCONDITION(!error) << "Unable to create image.";
 
     // Allocate memory for the VkImage and bind it.
     VkMemoryRequirements memReqs = {};
-    vkGetImageMemoryRequirements(mDevice, mTextureImage, &memReqs);
+    vkGetImageMemoryRequirements(state->mDevice, state->mTextureImage, &memReqs);
 
     uint32_t memoryTypeIndex
             = context.selectMemoryType(memReqs.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
@@ -194,58 +280,67 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
         .allocationSize = memReqs.size,
         .memoryTypeIndex = memoryTypeIndex,
     };
-    error = vkAllocateMemory(mDevice, &allocInfo, nullptr, &mTextureImageMemory);
+    error = vkAllocateMemory(state->mDevice, &allocInfo, nullptr, &state->mTextureImageMemory);
     FILAMENT_CHECK_POSTCONDITION(!error) << "Unable to allocate image memory.";
-    error = vkBindImageMemory(mDevice, mTextureImage, mTextureImageMemory, 0);
+    error = vkBindImageMemory(state->mDevice, state->mTextureImage, state->mTextureImageMemory, 0);
     FILAMENT_CHECK_POSTCONDITION(!error) << "Unable to bind image.";
 
-    uint32_t layerCount = 0;
-    if (target == SamplerType::SAMPLER_CUBEMAP) {
-        layerCount = 6;
-    } else if (target == SamplerType::SAMPLER_CUBEMAP_ARRAY) {
-        layerCount = depth * 6;
-    } else if (target == SamplerType::SAMPLER_2D_ARRAY) {
-        layerCount = depth;
-    } else if (target == SamplerType::SAMPLER_3D) {
-        layerCount = 1;
-    } else {
-        layerCount = 1;
-    }
-
-    mFullViewRange = {
-        .aspectMask = getImageAspect(),
-        .baseMipLevel = 0,
-        .levelCount = levels,
-        .baseArrayLayer = 0,
-        .layerCount = layerCount,
-    };
-
     // Spec out the "primary" VkImageView that shaders use to sample from the image.
-    mPrimaryViewRange = mFullViewRange;
+    mPrimaryViewRange = state->mFullViewRange;
 
     // Go ahead and create the primary image view.
-    getImageView(mPrimaryViewRange, mViewType, mSwizzle);
+    getImageView(mPrimaryViewRange, state->mViewType, mSwizzle);
 
-    // Transition the layout of each image slice that might be used as a render target.
-    // We do not transition images that are merely SAMPLEABLE, this is deferred until upload time
-    // because we do not know how many layers and levels will actually be used.
-    if (imageInfo.usage
-        & (VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
-           | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) {
-        VulkanCommandBuffer& commands = mCommands->get();
-        VkCommandBuffer const cmdbuf = commands.buffer();
-        commands.acquire(this);
-        transitionLayout(cmdbuf, mFullViewRange, imgutil::getDefaultLayout(imageInfo.usage));
-    }
+    VulkanCommandBuffer& commandsBuf = state->mCommands->get();
+    commandsBuf.acquire(this);
+    transitionLayout(&commandsBuf, mPrimaryViewRange, imgutil::getDefaultLayout(imageInfo.usage));
+}
+
+VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
+        VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
+        VulkanResourceAllocator* handleAllocator,
+        VulkanTexture const* src, uint8_t baseLevel, uint8_t levelCount)
+        : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
+                src->format, src->usage),
+          VulkanResource(VulkanResourceType::TEXTURE),
+          mAllocator(handleAllocator)
+{
+    mState = src->mState;
+    auto* state = getSharedState();
+
+    state->refs++;
+    mPrimaryViewRange = src->mPrimaryViewRange;
+    mPrimaryViewRange.baseMipLevel = src->mPrimaryViewRange.baseMipLevel + baseLevel;
+    mPrimaryViewRange.levelCount = levelCount;
+}
+
+VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
+        VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
+        VulkanResourceAllocator* handleAllocator,
+        VulkanTexture const* src, VkComponentMapping swizzle)
+        : HwTexture(src->target, src->levels, src->samples, src->width, src->height, src->depth,
+                src->format, src->usage),
+          VulkanResource(VulkanResourceType::TEXTURE),
+          mAllocator(handleAllocator) {
+    mState = src->mState;
+    auto* state = getSharedState();
+    state->refs++;
+    mPrimaryViewRange = src->mPrimaryViewRange;
+    mSwizzle = composeSwizzle(src->mSwizzle, swizzle);
 }
 
 VulkanTexture::~VulkanTexture() {
-    if (mTextureImageMemory != VK_NULL_HANDLE) {
-        vkDestroyImage(mDevice, mTextureImage, VKALLOC);
-        vkFreeMemory(mDevice, mTextureImageMemory, VKALLOC);
-    }
-    for (auto entry : mCachedImageViews) {
-        vkDestroyImageView(mDevice, entry.second, VKALLOC);
+    auto* const state = getSharedState();
+    state->refs--;
+    if (state->refs == 0) {
+        if (state->mTextureImageMemory != VK_NULL_HANDLE) {
+            vkDestroyImage(state->mDevice, state->mTextureImage, VKALLOC);
+            vkFreeMemory(state->mDevice, state->mTextureImageMemory, VKALLOC);
+        }
+        for (auto entry: state->mCachedImageViews) {
+            vkDestroyImageView(state->mDevice, entry.second, VKALLOC);
+        }
+        mAllocator->destruct<VulkanTextureState>(mState);
     }
 }
 
@@ -254,7 +349,7 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
     assert_invariant(width <= this->width && height <= this->height);
     assert_invariant(depth <= this->depth * ((target == SamplerType::SAMPLER_CUBEMAP ||
                         target == SamplerType::SAMPLER_CUBEMAP_ARRAY) ? 6 : 1));
-
+    auto* const state = getSharedState();
     const PixelBufferDescriptor* hostData = &data;
     PixelBufferDescriptor reshapedData;
 
@@ -267,7 +362,7 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
 
     // If format conversion is both required and supported, use vkCmdBlitImage.
     const VkFormat hostFormat = backend::getVkFormat(hostData->format, hostData->type);
-    const VkFormat deviceFormat = getVkFormatLinear(mVkFormat);
+    const VkFormat deviceFormat = getVkFormatLinear(state->mVkFormat);
     if (hostFormat != deviceFormat && hostFormat != VK_FORMAT_UNDEFINED) {
         assert_invariant(xoffset == 0 && yoffset == 0 && zoffset == 0 &&
                 "Offsets not yet supported when format conversion is required.");
@@ -279,14 +374,14 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
 
     // Otherwise, use vkCmdCopyBufferToImage.
     void* mapped = nullptr;
-    VulkanStage const* stage = mStagePool.acquireStage(hostData->size);
+    VulkanStage const* stage = state->mStagePool.acquireStage(hostData->size);
     assert_invariant(stage->memory);
-    vmaMapMemory(mAllocator, stage->memory, &mapped);
+    vmaMapMemory(state->mAllocator, stage->memory, &mapped);
     memcpy(mapped, hostData->buffer, hostData->size);
-    vmaUnmapMemory(mAllocator, stage->memory);
-    vmaFlushAllocation(mAllocator, stage->memory, 0, hostData->size);
+    vmaUnmapMemory(state->mAllocator, stage->memory);
+    vmaFlushAllocation(state->mAllocator, stage->memory, 0, hostData->size);
 
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = state->mCommands->get();
     VkCommandBuffer const cmdbuf = commands.buffer();
     commands.acquire(this);
 
@@ -332,24 +427,25 @@ void VulkanTexture::updateImage(const PixelBufferDescriptor& data, uint32_t widt
         nextLayout = imgutil::getDefaultLayout(this->usage);
     }
 
-    transitionLayout(cmdbuf, transitionRange, newLayout);
+    transitionLayout(&commands, transitionRange, newLayout);
 
-    vkCmdCopyBufferToImage(cmdbuf, stage->buffer, mTextureImage, newVkLayout, 1, &copyRegion);
+    vkCmdCopyBufferToImage(cmdbuf, stage->buffer, state->mTextureImage, newVkLayout, 1, &copyRegion);
 
-    transitionLayout(cmdbuf, transitionRange, nextLayout);
+    transitionLayout(&commands, transitionRange, nextLayout);
 }
 
 void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, uint32_t width,
         uint32_t height, uint32_t depth, uint32_t miplevel) {
+    auto* const state = getSharedState();
     void* mapped = nullptr;
     VulkanStageImage const* stage
-            = mStagePool.acquireImage(hostData.format, hostData.type, width, height);
-    vmaMapMemory(mAllocator, stage->memory, &mapped);
+            = state->mStagePool.acquireImage(hostData.format, hostData.type, width, height);
+    vmaMapMemory(state->mAllocator, stage->memory, &mapped);
     memcpy(mapped, hostData.buffer, hostData.size);
-    vmaUnmapMemory(mAllocator, stage->memory);
-    vmaFlushAllocation(mAllocator, stage->memory, 0, hostData.size);
+    vmaUnmapMemory(state->mAllocator, stage->memory);
+    vmaFlushAllocation(state->mAllocator, stage->memory, 0, hostData.size);
 
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = state->mCommands->get();
     VkCommandBuffer const cmdbuf = commands.buffer();
     commands.acquire(this);
 
@@ -371,63 +467,68 @@ void VulkanTexture::updateImageWithBlit(const PixelBufferDescriptor& hostData, u
 
     VulkanLayout const newLayout = VulkanLayout::TRANSFER_DST;
     VulkanLayout const oldLayout = getLayout(layer, miplevel);
-    transitionLayout(cmdbuf, range, newLayout);
+    transitionLayout(&commands, range, newLayout);
 
     vkCmdBlitImage(cmdbuf, stage->image, imgutil::getVkLayout(VulkanLayout::TRANSFER_SRC),
-            mTextureImage, imgutil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
+            state->mTextureImage, imgutil::getVkLayout(newLayout), 1, blitRegions, VK_FILTER_NEAREST);
 
-    transitionLayout(cmdbuf, range, oldLayout);
-}
-
-void VulkanTexture::setPrimaryRange(uint32_t minMiplevel, uint32_t maxMiplevel) {
-    maxMiplevel = filament::math::min(int(maxMiplevel), int(this->levels - 1));
-    mPrimaryViewRange.baseMipLevel = minMiplevel;
-    mPrimaryViewRange.levelCount = maxMiplevel - minMiplevel + 1;
-    getImageView(mPrimaryViewRange, mViewType, mSwizzle);
+    transitionLayout(&commands, range, oldLayout);
 }
 
 VkImageView VulkanTexture::getAttachmentView(VkImageSubresourceRange range) {
-    // Attachments should only have one mipmap level and one layer.
     range.levelCount = 1;
     range.layerCount = 1;
     return getImageView(range, VK_IMAGE_VIEW_TYPE_2D, {});
 }
 
+VkImageView VulkanTexture::getMultiviewAttachmentView(VkImageSubresourceRange range) {
+    return getImageView(range, VK_IMAGE_VIEW_TYPE_2D_ARRAY, {});
+}
+
 VkImageView VulkanTexture::getViewForType(VkImageSubresourceRange const& range, VkImageViewType type) {
     return getImageView(range, type, mSwizzle);
 }
 
 VkImageView VulkanTexture::getImageView(VkImageSubresourceRange range, VkImageViewType viewType,
         VkComponentMapping swizzle) {
-    ImageViewKey const key {range, viewType, swizzle};
-    auto iter = mCachedImageViews.find(key);
-    if (iter != mCachedImageViews.end()) {
+    auto* const state = getSharedState();
+    VulkanTextureState::ImageViewKey const key{ range, viewType, swizzle };
+    auto iter = state->mCachedImageViews.find(key);
+    if (iter != state->mCachedImageViews.end()) {
         return iter->second;
     }
     VkImageViewCreateInfo viewInfo = {
         .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
-        .image = mTextureImage,
+        .image = state->mTextureImage,
         .viewType = viewType,
-        .format = mVkFormat,
+        .format = state->mVkFormat,
         .components = swizzle,
         .subresourceRange = range,
     };
     VkImageView imageView;
-    vkCreateImageView(mDevice, &viewInfo, VKALLOC, &imageView);
-    mCachedImageViews.emplace(key, imageView);
+    vkCreateImageView(state->mDevice, &viewInfo, VKALLOC, &imageView);
+    state->mCachedImageViews.emplace(key, imageView);
     return imageView;
 }
 
 VkImageAspectFlags VulkanTexture::getImageAspect() const {
     // Helper function in VulkanUtility
-    return filament::backend::getImageAspect(mVkFormat);
+    auto* const state = getSharedState();
+    return filament::backend::getImageAspect(state->mVkFormat);
 }
 
-void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubresourceRange& range,
-        VulkanLayout newLayout) {
+bool VulkanTexture::transitionLayout(VulkanCommandBuffer* commands,
+        const VkImageSubresourceRange& range, VulkanLayout newLayout) {
+    return transitionLayout(commands->buffer(), commands->fence, range, newLayout);
+}
 
+bool VulkanTexture::transitionLayout(
+        VkCommandBuffer cmdbuf, std::shared_ptr<VulkanCmdFence> fence,
+        const VkImageSubresourceRange& range,
+        VulkanLayout newLayout) {
+    auto* const state = getSharedState();
     VulkanLayout const oldLayout = getLayout(range.baseArrayLayer, range.baseMipLevel);
 
     uint32_t const firstLayer = range.baseArrayLayer;
@@ -438,7 +539,7 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
     // If we are transitioning more than one layer/level (slice), we need to know whether they are
     // all of the same layer.  If not, we need to transition slice-by-slice. Otherwise it would
     // trigger the validation layer saying that the `oldLayout` provided is incorrect.
-    // TODO: transition by multiple slices with more sophiscated range finding.
+    // TODO: transition by multiple slices with more sophisticated range finding.
     bool transitionSliceBySlice = false;
     for (uint32_t i = firstLayer; i < lastLayer; ++i) {
         for (uint32_t j = firstLevel; j < lastLevel; ++j) {
@@ -449,50 +550,82 @@ void VulkanTexture::transitionLayout(VkCommandBuffer cmdbuf, const VkImageSubres
         }
     }
 
-#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
-    FVK_LOGD << "transition texture=" << mTextureImage
-                  << " (" << range.baseArrayLayer
-                  << "," << range.baseMipLevel << ")"
-                  << " count=(" << range.layerCount
-                  << "," << range.levelCount << ")"
-                  << " from=" << oldLayout << " to=" << newLayout
-                  << " format=" << mVkFormat
-                  << " depth=" << isVkDepthFormat(mVkFormat)
-                  << " slice-by-slice=" << transitionSliceBySlice
-                  << utils::io::endl;
-#endif
-
+    bool hasTransitions = false;
     if (transitionSliceBySlice) {
         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, {
-                        .image = mTextureImage,
-                        .oldLayout = layout,
-                        .newLayout = newLayout,
-                        .subresources = {
-                            .aspectMask = range.aspectMask,
-                            .baseMipLevel = j,
-                            .levelCount = 1,
-                            .baseArrayLayer = i,
-                            .layerCount = 1,
-                        },
-                    });
+                if (layout == newLayout) {
+                    continue;
+                }
+                hasTransitions = hasTransitions || imgutil::transitionLayout(cmdbuf, {
+                    .image = state->mTextureImage,
+                    .oldLayout = layout,
+                    .newLayout = newLayout,
+                    .subresources = {
+                        .aspectMask = range.aspectMask,
+                        .baseMipLevel = j,
+                        .levelCount = 1,
+                        .baseArrayLayer = i,
+                        .layerCount = 1,
+                    },
+                });
             }
         }
-    } else {
-        imgutil::transitionLayout(cmdbuf, {
-            .image = mTextureImage,
+    } else if (newLayout != oldLayout) {
+        hasTransitions = imgutil::transitionLayout(cmdbuf, {
+            .image = state->mTextureImage,
             .oldLayout = oldLayout,
             .newLayout = newLayout,
             .subresources = range,
         });
     }
 
-    setLayout(range, newLayout);
+    if (hasTransitions) {
+        state->mTransitionFence = fence;
+        setLayout(range, newLayout);
+
+#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
+        FVK_LOGD << "transition texture=" << state->mTextureImage << " (" << range.baseArrayLayer
+                 << "," << range.baseMipLevel << ")" << " count=(" << range.layerCount << ","
+                 << range.levelCount << ")" << " from=" << oldLayout << " to=" << newLayout
+                 << " format=" << state->mVkFormat << " depth=" << isVkDepthFormat(state->mVkFormat)
+                 << " slice-by-slice=" << transitionSliceBySlice << utils::io::endl;
+#endif
+    } else {
+#if FVK_ENABLED(FVK_DEBUG_LAYOUT_TRANSITION)
+        FVK_LOGD << "transition texture=" << state->mTextureImage << " (" << range.baseArrayLayer
+                 << "," << range.baseMipLevel << ")" << " count=(" << range.layerCount << ","
+                 << range.levelCount << ")" << " to=" << newLayout
+                 << " is skipped because of no change in layout" << utils::io::endl;
+#endif
+    }
+    return hasTransitions;
 }
 
-void VulkanTexture::setLayout(const VkImageSubresourceRange& range, VulkanLayout newLayout) {
+void VulkanTexture::attachmentToSamplerBarrier(VulkanCommandBuffer* commands,
+        VkImageSubresourceRange const& range) {
+    VkCommandBuffer const cmdbuf = commands->buffer();
+    auto* const state = getSharedState();
+    VkImageLayout const layout
+            = imgutil::getVkLayout(getLayout(range.baseArrayLayer, range.baseMipLevel));
+    VkImageMemoryBarrier barrier = {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
+            .oldLayout = layout,
+            .newLayout = layout,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = state->mTextureImage,
+            .subresourceRange = range,
+    };
+    vkCmdPipelineBarrier(cmdbuf, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+            VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &barrier);
+}
+
+void VulkanTexture::setLayout(VkImageSubresourceRange const& range, VulkanLayout newLayout) {
+    auto* const state = getSharedState();
     uint32_t const firstLayer = range.baseArrayLayer;
     uint32_t const lastLayer = firstLayer + range.layerCount;
     uint32_t const firstLevel = range.baseMipLevel;
@@ -505,28 +638,30 @@ void VulkanTexture::setLayout(const VkImageSubresourceRange& range, VulkanLayout
         for (uint32_t layer = firstLayer; layer < lastLayer; ++layer) {
             uint32_t const first = (layer << 16) | firstLevel;
             uint32_t const last = (layer << 16) | lastLevel;
-            mSubresourceLayouts.clear(first, last);
+            state->mSubresourceLayouts.clear(first, last);
         }
     } else {
         for (uint32_t layer = firstLayer; layer < lastLayer; ++layer) {
             uint32_t const first = (layer << 16) | firstLevel;
             uint32_t const last = (layer << 16) | lastLevel;
-            mSubresourceLayouts.add(first, last, newLayout);
+            state->mSubresourceLayouts.add(first, last, newLayout);
         }
     }
 }
 
 VulkanLayout VulkanTexture::getLayout(uint32_t layer, uint32_t level) const {
     assert_invariant(level <= 0xffff && layer <= 0xffff);
+    auto* const state = getSharedState();
     const uint32_t key = (layer << 16) | level;
-    if (!mSubresourceLayouts.has(key)) {
+    if (!state->mSubresourceLayouts.has(key)) {
         return VulkanLayout::UNDEFINED;
     }
-    return mSubresourceLayouts.get(key);
+    return state->mSubresourceLayouts.get(key);
 }
 
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
 void VulkanTexture::print() const {
+    auto* const state = getSharedState();
     uint32_t const firstLayer = 0;
     uint32_t const lastLayer = firstLayer + mFullViewRange.layerCount;
     uint32_t const firstLevel = 0;
@@ -539,16 +674,16 @@ void VulkanTexture::print() const {
                 layer < (mPrimaryViewRange.baseArrayLayer + mPrimaryViewRange.layerCount) &&
                 level >= mPrimaryViewRange.baseMipLevel &&
                 level < (mPrimaryViewRange.baseMipLevel + mPrimaryViewRange.levelCount);
-            FVK_LOGD << "[" << mTextureImage << "]: (" << layer << "," << level
+            FVK_LOGD << "[" << state->mTextureImage << "]: (" << layer << "," << level
                           << ")=" << getLayout(layer, level)
                           << " primary=" << primary
                           << utils::io::endl;
         }
     }
 
-    for (auto view: mCachedImageViews) {
+    for (auto view: state->mCachedImageViews) {
         auto& range = view.first.range;
-        FVK_LOGD << "[" << mTextureImage << ", imageView=" << view.second << "]=>"
+        FVK_LOGD << "[" << state->mTextureImage << ", imageView=" << view.second << "]=>"
                       << " (" << range.baseArrayLayer << "," << range.baseMipLevel << ")"
                       << " count=(" << range.layerCount << "," << range.levelCount << ")"
                       << " aspect=" << range.aspectMask << " viewType=" << view.first.type

filament/backend/src/vulkan/VulkanTexture.h

@@ -30,84 +30,12 @@
 
 namespace filament::backend {
 
-struct VulkanTexture : public HwTexture, VulkanResource {
-    // Standard constructor for user-facing textures.
-    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
-            VmaAllocator allocator, VulkanCommands* commands, SamplerType target, uint8_t levels,
-            TextureFormat tformat, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
-            TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated = false,
-            VkComponentMapping swizzle = {});
+class VulkanResourceAllocator;
 
-    // Specialized constructor for internally created textures (e.g. from a swap chain)
-    // The texture will never destroy the given VkImage, but it does manages its subresources.
-    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 = false);
-
-    ~VulkanTexture();
-
-    // Uploads data into a subregion of a 2D or 3D texture.
-    void updateImage(const PixelBufferDescriptor& data, uint32_t width, uint32_t height,
-            uint32_t depth, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset, uint32_t miplevel);
-
-    // Returns the primary image view, which is used for shader sampling.
-    VkImageView getPrimaryImageView() {
-        return getImageView(mPrimaryViewRange, mViewType, mSwizzle);
-    }
-
-    VkImageViewType getViewType() const { return mViewType; }
-
-    // Sets the min/max range of miplevels in the primary image view.
-    void setPrimaryRange(uint32_t minMiplevel, uint32_t maxMiplevel);
-
-    VkImageSubresourceRange getPrimaryViewRange() const { return mPrimaryViewRange; }
-
-    VkImageSubresourceRange getFullViewRange() const { return mFullViewRange; }
-
-    VulkanLayout getPrimaryImageLayout() const {
-        return getLayout(mPrimaryViewRange.baseArrayLayer, mPrimaryViewRange.baseMipLevel);
-    }
-
-    // Gets or creates a cached VkImageView for a single subresource that can be used as a render
-    // target attachment.  Unlike the primary image view, this always has type VK_IMAGE_VIEW_TYPE_2D
-    // and the identity swizzle.
-    VkImageView getAttachmentView(VkImageSubresourceRange range);
-
-    // This is a workaround for the first few frames where we're waiting for the texture to actually
-    // be uploaded.  In that case, we bind the sampler to an empty texture, but the corresponding
-    // imageView needs to be of the right type. Hence, we provide an option to indicate the
-    // view type. Swizzle option does not matter in this case.
-    VkImageView getViewForType(VkImageSubresourceRange const& range, VkImageViewType type);
-
-    VkFormat getVkFormat() const { return mVkFormat; }
-    VkImage getVkImage() const { return mTextureImage; }
-
-    VulkanLayout getLayout(uint32_t layer, uint32_t level) const;
-
-    void setSidecar(VulkanTexture* sidecar) {
-        mSidecarMSAA.reset(sidecar);
-    }
-
-    VulkanTexture* getSidecar() const {
-        return mSidecarMSAA.get();
-    }
-
-    void transitionLayout(VkCommandBuffer commands, const VkImageSubresourceRange& range,
-            VulkanLayout newLayout);
-
-    // Returns the preferred data plane of interest for all image views.
-    // For now this always returns either DEPTH or COLOR.
-    VkImageAspectFlags getImageAspect() const;
-
-    // For implicit transition like the end of a render pass, we need to be able to set the layout
-    // manually (outside of calls to transitionLayout).
-    void setLayout(const VkImageSubresourceRange& range, VulkanLayout newLayout);
-
-#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
-    void print() const;
-#endif
-
-private:
+struct VulkanTextureState : public VulkanResource {
+    VulkanTextureState(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
+            VulkanStagePool& stagePool,
+            VkFormat format, VkImageViewType viewType, uint8_t levels, uint8_t layerCount);
 
     struct ImageViewKey {
         VkImageSubresourceRange range;  // 4 * 5 bytes
@@ -130,6 +58,150 @@ private:
 
     using ImageViewHash = utils::hash::MurmurHashFn<ImageViewKey>;
 
+    uint32_t refs = 1;
+
+    // The texture with the sidecar owns the sidecar.
+    std::unique_ptr<VulkanTexture> mSidecarMSAA;
+    VkDeviceMemory mTextureImageMemory = VK_NULL_HANDLE;
+
+    VkFormat const mVkFormat;
+    VkImageViewType const mViewType;
+    VkImageSubresourceRange const mFullViewRange;
+
+    VkImage mTextureImage = VK_NULL_HANDLE;
+
+    // Track the image layout of each subresource using a sparse range map.
+    utils::RangeMap<uint32_t, VulkanLayout> mSubresourceLayouts;
+
+    std::unordered_map<ImageViewKey, VkImageView, ImageViewHash> mCachedImageViews;
+    VulkanStagePool& mStagePool;
+    VkDevice mDevice;
+    VmaAllocator mAllocator;
+    VulkanCommands* mCommands;
+    std::shared_ptr<VulkanCmdFence> mTransitionFence;
+};
+
+
+struct VulkanTexture : public HwTexture, VulkanResource {
+    // Standard constructor for user-facing textures.
+    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
+            VmaAllocator allocator, VulkanCommands* commands,
+            VulkanResourceAllocator* handleAllocator,
+            SamplerType target, uint8_t levels,
+            TextureFormat tformat, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
+            TextureUsage tusage, VulkanStagePool& stagePool, bool heapAllocated = false);
+
+    // Specialized constructor for internally created textures (e.g. from a swap chain)
+    // The texture will never destroy the given VkImage, but it does manages its subresources.
+    VulkanTexture(VkDevice device, VmaAllocator allocator, VulkanCommands* commands,
+            VulkanResourceAllocator* handleAllocator,
+            VkImage image,
+            VkFormat format, uint8_t samples, uint32_t width, uint32_t height, TextureUsage tusage,
+            VulkanStagePool& stagePool, bool heapAllocated = false);
+
+    // Constructor for creating a texture view for wrt specific mip range
+    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
+            VmaAllocator allocator, VulkanCommands* commands,
+            VulkanResourceAllocator* handleAllocator,
+            VulkanTexture const* src, uint8_t baseLevel, uint8_t levelCount);
+
+    // Constructor for creating a texture view for swizzle.
+    VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice, VulkanContext const& context,
+            VmaAllocator allocator, VulkanCommands* commands,
+            VulkanResourceAllocator* handleAllocator,
+            VulkanTexture const* src, VkComponentMapping swizzle);
+
+    ~VulkanTexture();
+
+    // Uploads data into a subregion of a 2D or 3D texture.
+    void updateImage(const PixelBufferDescriptor& data, uint32_t width, uint32_t height,
+            uint32_t depth, uint32_t xoffset, uint32_t yoffset, uint32_t zoffset, uint32_t miplevel);
+
+    // Returns the primary image view, which is used for shader sampling.
+    VkImageView getPrimaryImageView() {
+        VulkanTextureState* state = getSharedState();
+        return getImageView(mPrimaryViewRange, state->mViewType, mSwizzle);
+    }
+
+    VkImageViewType getViewType() const {
+        VulkanTextureState const* state = getSharedState();
+        return state->mViewType;
+    }
+
+    VkImageSubresourceRange getPrimaryViewRange() const { return mPrimaryViewRange; }
+
+    VulkanLayout getPrimaryImageLayout() const {
+        return getLayout(mPrimaryViewRange.baseArrayLayer, mPrimaryViewRange.baseMipLevel);
+    }
+
+    // Gets or creates a cached VkImageView for a single subresource that can be used as a render
+    // target attachment.  Unlike the primary image view, this always has type VK_IMAGE_VIEW_TYPE_2D
+    // and the identity swizzle.
+    VkImageView getAttachmentView(VkImageSubresourceRange range);
+
+    // Gets or creates a cached VkImageView for a single subresource that can be used as a render
+    // target attachment when rendering with multiview.
+    VkImageView getMultiviewAttachmentView(VkImageSubresourceRange range);
+
+    // This is a workaround for the first few frames where we're waiting for the texture to actually
+    // be uploaded.  In that case, we bind the sampler to an empty texture, but the corresponding
+    // imageView needs to be of the right type. Hence, we provide an option to indicate the
+    // view type. Swizzle option does not matter in this case.
+    VkImageView getViewForType(VkImageSubresourceRange const& range, VkImageViewType type);
+
+    VkFormat getVkFormat() const {
+        VulkanTextureState const* state = getSharedState();
+        return state->mVkFormat;
+    }
+    VkImage getVkImage() const {
+        VulkanTextureState const* state = getSharedState();
+        return state->mTextureImage;
+    }
+
+    VulkanLayout getLayout(uint32_t layer, uint32_t level) const;
+
+    void setSidecar(VulkanTexture* sidecar) {
+        VulkanTextureState* state = getSharedState();
+        state->mSidecarMSAA.reset(sidecar);
+    }
+
+    VulkanTexture* getSidecar() const {
+        VulkanTextureState const* state = getSharedState();
+        return state->mSidecarMSAA.get();
+    }
+
+    bool transitionLayout(VulkanCommandBuffer* commands, const VkImageSubresourceRange& range,
+            VulkanLayout newLayout);
+
+    bool transitionLayout(VkCommandBuffer cmdbuf, std::shared_ptr<VulkanCmdFence> fence,
+            VkImageSubresourceRange const& range, VulkanLayout newLayout);
+
+    void attachmentToSamplerBarrier(VulkanCommandBuffer* commands,
+            VkImageSubresourceRange const& range);
+
+    // Returns the preferred data plane of interest for all image views.
+    // For now this always returns either DEPTH or COLOR.
+    VkImageAspectFlags getImageAspect() const;
+
+    // For implicit transition like the end of a render pass, we need to be able to set the layout
+    // manually (outside of calls to transitionLayout).
+    void setLayout(VkImageSubresourceRange const& range, VulkanLayout newLayout);
+
+    bool transitionReady() {
+        VulkanTextureState* state = getSharedState();
+        auto res = !state->mTransitionFence || state->mTransitionFence->getStatus() == VK_SUCCESS;
+        state->mTransitionFence.reset();
+        return res;
+    }
+
+#if FVK_ENABLED(FVK_DEBUG_TEXTURE)
+    void print() const;
+#endif
+
+private:
+    VulkanTextureState* getSharedState();
+    VulkanTextureState const* getSharedState() const;
+
     // Gets or creates a cached VkImageView for a range of miplevels, array layers, viewType, and
     // swizzle (or not).
     VkImageView getImageView(VkImageSubresourceRange range, VkImageViewType viewType,
@@ -138,28 +210,15 @@ private:
     void updateImageWithBlit(const PixelBufferDescriptor& hostData, uint32_t width, uint32_t height,
             uint32_t depth, uint32_t miplevel);
 
-    // The texture with the sidecar owns the sidecar.
-    std::unique_ptr<VulkanTexture> mSidecarMSAA;
-    const VkFormat mVkFormat;
-    const VkImageViewType mViewType;
-    const VkComponentMapping mSwizzle;
-    VkImage mTextureImage = VK_NULL_HANDLE;
-    VkDeviceMemory mTextureImageMemory = VK_NULL_HANDLE;
+    VulkanResourceAllocator* const mAllocator;
 
-    // Track the image layout of each subresource using a sparse range map.
-    utils::RangeMap<uint32_t, VulkanLayout> mSubresourceLayouts;
-
-    VkImageSubresourceRange mFullViewRange;
+    Handle<VulkanTextureState> mState;
 
     // Track the range of subresources that define the "primary" image view, which is the special
     // image view that gets bound to an actual texture sampler.
     VkImageSubresourceRange mPrimaryViewRange;
 
-    std::unordered_map<ImageViewKey, VkImageView, ImageViewHash> mCachedImageViews;
-    VulkanStagePool& mStagePool;
-    VkDevice mDevice;
-    VmaAllocator mAllocator;
-    VulkanCommands* mCommands;
+    VkComponentMapping mSwizzle {};
 };
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanUtility.cpp

@@ -577,7 +577,7 @@ uint32_t getComponentCount(VkFormat format) {
     return {};
 }
 
-VkComponentMapping getSwizzleMap(TextureSwizzle swizzle[4]) {
+VkComponentMapping getSwizzleMap(TextureSwizzle const swizzle[4]) {
     VkComponentMapping map;
     VkComponentSwizzle* dst = &map.r;
     for (int i = 0; i < 4; ++i, ++dst) {

filament/backend/src/vulkan/VulkanUtility.h

@@ -19,6 +19,7 @@
 
 #include <backend/DriverEnums.h>
 
+#include <utils/bitset.h>
 #include <utils/FixedCapacityVector.h>
 
 #include <bluevk/BlueVK.h>
@@ -38,7 +39,7 @@ VkCullModeFlags getCullMode(CullingMode mode);
 VkFrontFace getFrontFace(bool inverseFrontFaces);
 PixelDataType getComponentType(VkFormat format);
 uint32_t getComponentCount(VkFormat format);
-VkComponentMapping getSwizzleMap(TextureSwizzle swizzle[4]);
+VkComponentMapping getSwizzleMap(TextureSwizzle const swizzle[4]);
 VkShaderStageFlags getShaderStageFlags(ShaderStageFlags stageFlags);
 
 bool equivalent(const VkRect2D& a, const VkRect2D& b);
@@ -405,12 +406,13 @@ constexpr VkFormat ALL_VK_FORMATS[] = {
         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.
+// An Array that will be statically fixed in capacity, but the "size" (as in user added elements) is
+// variable. Note that this class is movable.
 template<typename T, uint16_t CAPACITY>
 class CappedArray {
 private:
     using FixedSizeArray = std::array<T, CAPACITY>;
+
 public:
     using const_iterator = typename FixedSizeArray::const_iterator;
     using iterator = typename FixedSizeArray::iterator;
@@ -448,6 +450,20 @@ public:
         return mArray.cend();
     }
 
+    inline iterator begin() {
+        if (mInd == 0) {
+            return mArray.end();
+        }
+        return mArray.begin();
+    }
+
+    inline iterator end() {
+        if (mInd > 0 && mInd < CAPACITY) {
+            return mArray.begin() + mInd;
+        }
+        return mArray.end();
+    }
+
     inline T back() {
         assert_invariant(mInd > 0);
         return *(mArray.begin() + mInd);
@@ -512,125 +528,28 @@ private:
     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;
+using UniformBufferBitmask = utils::bitset64;
+using SamplerBitmask = utils::bitset64;
 
 // 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));
+using InputAttachmentBitmask = utils::bitset64;
 
 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;
-    }
+static constexpr uint8_t getVertexStageShift() noexcept {
+    // We assume the bottom half of bits are for vertex stages.
+    return 0;
 }
 
 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;
-    }
+static constexpr uint8_t getFragmentStageShift() noexcept {
+    // We assume the top half of bits are for fragment stages.
+    return sizeof(Bitmask) * 4;
 }
 
-typedef enum ShaderStageFlags2 : uint8_t {
-    NONE        =    0,
-    VERTEX      =    0x1,
-    FRAGMENT    =    0x2,
-} ShaderStageFlags2;
+// We have at most 4 descriptor sets. This is to indicate which ones are active.
+using DescriptorSetMask = utils::bitset8;
 
-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
 

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

@@ -30,12 +30,8 @@
 #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.
@@ -45,60 +41,32 @@ 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)>;
+
+    using DescriptorSetLayoutArray = VulkanDescriptorSetLayout::DescriptorSetLayoutArray;
 
     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,
+    void updateBuffer(VulkanDescriptorSet* set, uint8_t binding,
             VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept;
 
-    void updateSampler(Handle<VulkanDescriptorSet> set, uint8_t binding,
+    void updateSampler(VulkanDescriptorSet* set, uint8_t binding,
             VulkanTexture* texture, VkSampler sampler) noexcept;
 
-    void updateInputAttachment(Handle<VulkanDescriptorSet> set, VulkanAttachment attachment) noexcept;
+    void updateInputAttachment(VulkanDescriptorSet* set, VulkanAttachment attachment) noexcept;
 
-    void clearBuffer(uint32_t bindingIndex);
+    void bind(uint8_t setIndex, VulkanDescriptorSet* set, backend::DescriptorSetOffsetArray&& offsets);
+
+    void commit(VulkanCommandBuffer* commands, VkPipelineLayout pipelineLayout,
+            DescriptorSetMask const& setMask);
 
     void setPlaceHolders(VkSampler sampler, VulkanTexture* texture,
             VulkanBufferObject* bufferObject) noexcept;
 
-    void clearState() noexcept;
+    void createSet(Handle<HwDescriptorSet> handle, VulkanDescriptorSetLayout* layout);
 
-    // 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) {}
+    void destroySet(Handle<HwDescriptorSet> handle);
 
 private:
     class Impl;
@@ -108,4 +76,3 @@ private:
 }// namespace filament::backend
 
 #endif// TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSETMANAGER_H
-

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

@@ -21,23 +21,24 @@
 namespace filament::backend {
 
 VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
-        VulkanDescriptorSetLayoutList const& descriptorSetLayouts, VulkanProgram* program) {
+        DescriptorSetLayoutArray const& descriptorSetLayouts, VulkanProgram* program) {
     PipelineLayoutKey key = {};
     uint8_t descSetLayoutCount = 0;
+    key.descSetLayouts = descriptorSetLayouts;
     for (auto layoutHandle: descriptorSetLayouts) {
-        if (layoutHandle) {
-            auto layout = mAllocator->handle_cast<VulkanDescriptorSetLayout*>(layoutHandle);
-            key.descSetLayouts[descSetLayoutCount++] = layout->vklayout;
+        if (layoutHandle == VK_NULL_HANDLE) {
+            break;
         }
+        descSetLayoutCount++;
     }
 
     // build the push constant layout key
-    uint32_t pushConstantRangeCount = program->getPushConstantRangeCount();
-    auto const& pushCostantRanges = program->getPushConstantRanges();    
+    uint32_t const pushConstantRangeCount = program->getPushConstantRangeCount();
+    auto const& pushConstantRanges = 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 const& range = pushConstantRanges[i];
             auto& pushConstant = key.pushConstant[i];
             if (range.stageFlags &  VK_SHADER_STAGE_VERTEX_BIT) {
                 pushConstant.stage = static_cast<uint8_t>(ShaderStage::VERTEX);
@@ -52,9 +53,8 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
         }
     }
 
-    auto iter = mPipelineLayouts.find(key);
-    if (iter != mPipelineLayouts.end()) {
-        PipelineLayoutCacheEntry& entry = mPipelineLayouts[key];
+    if (auto iter = mPipelineLayouts.find(key); iter != mPipelineLayouts.end()) {
+        PipelineLayoutCacheEntry& entry = iter->second;
         entry.lastUsed = mTimestamp++;
         return entry.handle;
     }
@@ -65,7 +65,7 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
         .setLayoutCount = (uint32_t) descSetLayoutCount,
         .pSetLayouts = key.descSetLayouts.data(),
         .pushConstantRangeCount = pushConstantRangeCount,
-        .pPushConstantRanges = pushCostantRanges,
+        .pPushConstantRanges = pushConstantRanges,
     };
 
     VkPipelineLayout layout;

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

@@ -22,41 +22,40 @@
 
 #include <utils/Hash.h>
 
-#include <tsl/robin_map.h>
+#include <unordered_map>
 
 namespace filament::backend {
 
 class VulkanPipelineLayoutCache {
 public:
+    using DescriptorSetLayoutArray = VulkanDescriptorSetLayout::DescriptorSetLayoutArray;
+    
     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;
+        uint8_t stage = 0;// We have one set of push constant per shader stage (fragment, vertex, etc).
+        uint8_t size = 0;
         // 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
+        DescriptorSetLayoutArray descSetLayouts = {};                                     // 8 * 4
+        PushConstantKey pushConstant[Program::SHADER_TYPE_COUNT] = {};                    // 2 * 3
         uint16_t padding = 0;
     };
-    static_assert(sizeof(PipelineLayoutKey) == 32);
+    static_assert(sizeof(PipelineLayoutKey) == 40);
 
     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,
+    VkPipelineLayout getLayout(DescriptorSetLayoutArray const& descriptorSetLayouts,
             VulkanProgram* program);
 
 private:
@@ -73,15 +72,14 @@ private:
         }
     };
 
-    using PipelineLayoutMap = tsl::robin_map<PipelineLayoutKey, PipelineLayoutCacheEntry,
+    using PipelineLayoutMap = std::unordered_map<PipelineLayoutKey, PipelineLayoutCacheEntry,
             PipelineLayoutKeyHashFn, PipelineLayoutKeyEqual>;
 
     VkDevice mDevice;
-    VulkanResourceAllocator* mAllocator;
     Timestamp mTimestamp;
     PipelineLayoutMap mPipelineLayouts;
 };
 
-}
+} // filament::backend
 
 #endif // TNT_FILAMENT_BACKEND_VULKANPIPELINECACHE_H

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

@@ -325,6 +325,7 @@ VkDevice createLogicalDevice(VkPhysicalDevice physicalDevice,
     // We could simply enable all supported features, but since that may have performance
     // consequences let's just enable the features we need.
     VkPhysicalDeviceFeatures enabledFeatures{
+            .depthClamp = features.depthClamp,
             .samplerAnisotropy = features.samplerAnisotropy,
             .textureCompressionETC2 = features.textureCompressionETC2,
             .textureCompressionBC = features.textureCompressionBC,

filament/backend/test/BackendTest.cpp

@@ -112,9 +112,10 @@ void BackendTest::fullViewport(Viewport& viewport) {
 }
 
 void BackendTest::renderTriangle(
-        filament::backend::Handle<filament::backend::HwRenderTarget> renderTarget,
-        filament::backend::Handle<filament::backend::HwSwapChain> swapChain,
-        filament::backend::Handle<filament::backend::HwProgram> program) {
+        PipelineLayout const& pipelineLayout,
+        Handle<filament::backend::HwRenderTarget> renderTarget,
+        Handle<filament::backend::HwSwapChain> swapChain,
+        Handle<filament::backend::HwProgram> program) {
     RenderPassParams params = {};
     fullViewport(params);
     params.flags.clear = TargetBufferFlags::COLOR;
@@ -123,11 +124,15 @@ void BackendTest::renderTriangle(
     params.flags.discardEnd = TargetBufferFlags::NONE;
     params.viewport.height = 512;
     params.viewport.width = 512;
-    renderTriangle(renderTarget, swapChain, program, params);
+    renderTriangle(pipelineLayout, renderTarget, swapChain, program, params);
 }
 
-void BackendTest::renderTriangle(Handle<HwRenderTarget> renderTarget,
-        Handle<HwSwapChain> swapChain, Handle<HwProgram> program, const RenderPassParams& params) {
+void BackendTest::renderTriangle(
+        PipelineLayout const& pipelineLayout,
+        Handle<HwRenderTarget> renderTarget,
+        Handle<HwSwapChain> swapChain,
+        Handle<HwProgram> program,
+        const RenderPassParams& params) {
     auto& api = getDriverApi();
 
     TrianglePrimitive triangle(api);
@@ -138,6 +143,7 @@ void BackendTest::renderTriangle(Handle<HwRenderTarget> renderTarget,
 
     PipelineState state;
     state.program = program;
+    state.pipelineLayout = pipelineLayout;
     state.rasterState.colorWrite = true;
     state.rasterState.depthWrite = false;
     state.rasterState.depthFunc = RasterState::DepthFunc::A;

filament/backend/test/BackendTest.h

@@ -51,10 +51,15 @@ protected:
     static void fullViewport(filament::backend::RenderPassParams& params);
     static void fullViewport(filament::backend::Viewport& viewport);
 
-    void renderTriangle(filament::backend::Handle<filament::backend::HwRenderTarget> renderTarget,
+    void renderTriangle(
+            filament::backend::PipelineLayout const& pipelineLayout,
+            filament::backend::Handle<filament::backend::HwRenderTarget> renderTarget,
             filament::backend::Handle<filament::backend::HwSwapChain> swapChain,
             filament::backend::Handle<filament::backend::HwProgram> program);
-    void renderTriangle(filament::backend::Handle<filament::backend::HwRenderTarget> renderTarget,
+
+    void renderTriangle(
+            filament::backend::PipelineLayout const& pipelineLayout,
+            filament::backend::Handle<filament::backend::HwRenderTarget> renderTarget,
             filament::backend::Handle<filament::backend::HwSwapChain> swapChain,
             filament::backend::Handle<filament::backend::HwProgram> program,
             const filament::backend::RenderPassParams& params);

filament/backend/test/MetalTest.mm

@@ -0,0 +1,159 @@
+/*
+ * 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 <gtest/gtest.h>
+
+#include "../src/metal/MetalContext.h"
+
+namespace test {
+
+TEST(MetalDynamicOffsets, none) {
+    filament::backend::MetalDynamicOffsets dynamicOffsets;
+    const auto [count, offsets] = dynamicOffsets.getOffsets();
+    EXPECT_EQ(count, 0u);
+}
+
+TEST(MetalDynamicOffsets, basic) {
+    filament::backend::MetalDynamicOffsets dynamicOffsets;
+    {
+        const auto [count, offsets] = dynamicOffsets.getOffsets();
+        EXPECT_EQ(count, 0u);
+    }
+
+    {
+        uint32_t o[2] = { 1, 2 };
+        dynamicOffsets.setOffsets(0, o, 2);
+        const auto [count, offsets] = dynamicOffsets.getOffsets();
+        EXPECT_EQ(count, 2);
+        EXPECT_EQ(offsets[0], 1);
+        EXPECT_EQ(offsets[1], 2);
+    }
+
+    {
+        uint32_t o[3] = { 3, 4, 5 };
+        dynamicOffsets.setOffsets(1, o, 3);
+        const auto [count, offsets] = dynamicOffsets.getOffsets();
+        EXPECT_EQ(count, 5);
+        EXPECT_EQ(offsets[0], 1);
+        EXPECT_EQ(offsets[1], 2);
+        EXPECT_EQ(offsets[2], 3);
+        EXPECT_EQ(offsets[3], 4);
+        EXPECT_EQ(offsets[4], 5);
+    }
+
+    // skip descriptor set index 2
+
+    {
+        uint32_t o[1] = { 6 };
+        dynamicOffsets.setOffsets(3, o, 1);
+        const auto [count, offsets] = dynamicOffsets.getOffsets();
+        EXPECT_EQ(count, 6);
+        EXPECT_EQ(offsets[0], 1);
+        EXPECT_EQ(offsets[1], 2);
+        EXPECT_EQ(offsets[2], 3);
+        EXPECT_EQ(offsets[3], 4);
+        EXPECT_EQ(offsets[4], 5);
+        EXPECT_EQ(offsets[5], 6);
+    }
+}
+
+TEST(MetalDynamicOffsets, outOfOrder) {
+    filament::backend::MetalDynamicOffsets dynamicOffsets;
+    uint32_t o1[2] = { 2, 3 };
+    dynamicOffsets.setOffsets(1, o1, 2);
+    uint32_t o2[2] = { 0, 1 };
+    dynamicOffsets.setOffsets(0, o2, 2);
+    uint32_t o3[2] = { 4, 5 };
+    dynamicOffsets.setOffsets(2, o3, 2);
+
+    const auto [count, offsets] = dynamicOffsets.getOffsets();
+    EXPECT_EQ(count, 6);
+    EXPECT_EQ(offsets[0], 0);
+    EXPECT_EQ(offsets[1], 1);
+    EXPECT_EQ(offsets[2], 2);
+    EXPECT_EQ(offsets[3], 3);
+    EXPECT_EQ(offsets[4], 4);
+    EXPECT_EQ(offsets[5], 5);
+};
+
+TEST(MetalDynamicOffsets, removal) {
+    filament::backend::MetalDynamicOffsets dynamicOffsets;
+    uint32_t o1[2] = { 2, 3 };
+    dynamicOffsets.setOffsets(1, o1, 2);
+    uint32_t o2[2] = { 0, 1 };
+    dynamicOffsets.setOffsets(0, o2, 2);
+    uint32_t o3[2] = { 4, 5 };
+    dynamicOffsets.setOffsets(2, o3, 2);
+    dynamicOffsets.setOffsets(1, nullptr, 0);
+
+    const auto [count, offsets] = dynamicOffsets.getOffsets();
+    EXPECT_EQ(count, 4);
+    EXPECT_EQ(offsets[0], 0);
+    EXPECT_EQ(offsets[1], 1);
+    EXPECT_EQ(offsets[2], 4);
+    EXPECT_EQ(offsets[3], 5);
+};
+
+TEST(MetalDynamicOffsets, resize) {
+    filament::backend::MetalDynamicOffsets dynamicOffsets;
+    uint32_t o1[2] = { 2, 3 };
+    dynamicOffsets.setOffsets(1, o1, 2);
+    uint32_t o2[2] = { 0, 1 };
+    dynamicOffsets.setOffsets(0, o2, 2);
+    uint32_t o3[2] = { 6, 7 };
+    dynamicOffsets.setOffsets(2, o3, 2);
+    uint32_t o4[4] = { 2, 3, 4, 5 };
+    dynamicOffsets.setOffsets(1, o4, 4);
+
+    const auto [count, offsets] = dynamicOffsets.getOffsets();
+    EXPECT_EQ(count, 8);
+    EXPECT_EQ(offsets[0], 0);
+    EXPECT_EQ(offsets[1], 1);
+    EXPECT_EQ(offsets[2], 2);
+    EXPECT_EQ(offsets[3], 3);
+    EXPECT_EQ(offsets[4], 4);
+    EXPECT_EQ(offsets[5], 5);
+    EXPECT_EQ(offsets[6], 6);
+    EXPECT_EQ(offsets[7], 7);
+};
+
+TEST(MetalDynamicOffsets, dirty) {
+    filament::backend::MetalDynamicOffsets dynamicOffsets;
+    EXPECT_FALSE(dynamicOffsets.isDirty());
+
+    uint32_t o1[2] = { 2, 3 };
+    dynamicOffsets.setOffsets(1, o1, 2);
+    EXPECT_TRUE(dynamicOffsets.isDirty());
+
+    dynamicOffsets.setDirty(false);
+    EXPECT_FALSE(dynamicOffsets.isDirty());
+
+    // Setting the same offsets should not mark the offsets as dirty
+    dynamicOffsets.setOffsets(1, o1, 2);
+    EXPECT_FALSE(dynamicOffsets.isDirty());
+
+    // Resizing the offsets should mark the offsets as dirty
+    uint32_t o2[3] = { 4, 5, 6 };
+    dynamicOffsets.setOffsets(1, o2, 3);
+    EXPECT_TRUE(dynamicOffsets.isDirty());
+};
+
+};
+
+int main(int argc, char **argv) {
+    testing::InitGoogleTest(&argc, argv);
+    return RUN_ALL_TESTS();
+}

filament/backend/test/ShaderGenerator.cpp

@@ -73,12 +73,13 @@ void ShaderGenerator::shutdown() {
     FinalizeProcess();
 }
 
-ShaderGenerator::ShaderGenerator(std::string vertex, std::string fragment,
-        Backend backend, bool isMobile, const filament::SamplerInterfaceBlock* sib) noexcept
-        : mBackend(backend),
-          mVertexBlob(transpileShader(ShaderStage::VERTEX, std::move(vertex), backend, isMobile, sib)),
-          mFragmentBlob(transpileShader(ShaderStage::FRAGMENT, std::move(fragment), backend,
-                  isMobile, sib)) {
+ShaderGenerator::ShaderGenerator(std::string vertex, std::string fragment, Backend backend,
+        bool isMobile, filamat::DescriptorSets&& descriptorSets) noexcept
+    : mBackend(backend),
+      mVertexBlob(transpileShader(
+              ShaderStage::VERTEX, std::move(vertex), backend, isMobile, descriptorSets)),
+      mFragmentBlob(transpileShader(
+              ShaderStage::FRAGMENT, std::move(fragment), backend, isMobile, descriptorSets)) {
     switch (backend) {
         case Backend::OPENGL:
             mShaderLanguage = filament::backend::ShaderLanguage::ESSL3;
@@ -95,9 +96,8 @@ ShaderGenerator::ShaderGenerator(std::string vertex, std::string fragment,
     }
 }
 
-ShaderGenerator::Blob ShaderGenerator::transpileShader(
-        ShaderStage stage, std::string shader, Backend backend, bool isMobile,
-        const filament::SamplerInterfaceBlock* sib) noexcept {
+ShaderGenerator::Blob ShaderGenerator::transpileShader(ShaderStage stage, std::string shader,
+        Backend backend, bool isMobile, const filamat::DescriptorSets& descriptorSets) noexcept {
     TProgram program;
     const EShLanguage language = stage == ShaderStage::VERTEX ? EShLangVertex : EShLangFragment;
     TShader tShader(language);
@@ -161,9 +161,8 @@ ShaderGenerator::Blob ShaderGenerator::transpileShader(
         return { result.c_str(), result.c_str() + result.length() + 1 };
     } else if (backend == Backend::METAL) {
         const auto sm = isMobile ? ShaderModel::MOBILE : ShaderModel::DESKTOP;
-        filamat::SibVector sibs = filamat::SibVector::with_capacity(1);
-        if (sib) { sibs.emplace_back(0, sib); }
-        filamat::GLSLPostProcessor::spirvToMsl(&spirv, &result, sm, false, sibs, nullptr);
+        filamat::GLSLPostProcessor::spirvToMsl(
+                &spirv, &result, stage, sm, false, descriptorSets, nullptr);
         return { result.c_str(), result.c_str() + result.length() + 1 };
     } else if (backend == Backend::VULKAN) {
         return { (uint8_t*)spirv.data(), (uint8_t*)(spirv.data() + spirv.size()) };

filament/backend/test/ShaderGenerator.h

@@ -22,6 +22,7 @@
 #include "private/filament/SamplerInterfaceBlock.h"
 #include "private/backend/DriverApi.h"
 #include "backend/Program.h"
+#include "../src/GLSLPostProcessor.h"
 
 #include <string>
 
@@ -40,7 +41,7 @@ public:
      * @param fragment The fragment shader, written in GLSL 450 core.
      */
     ShaderGenerator(std::string vertex, std::string fragment, Backend backend, bool isMobile,
-            const filament::SamplerInterfaceBlock* sib = nullptr) noexcept;
+            filamat::DescriptorSets&& descriptorSets = {}) noexcept;
 
     ShaderGenerator(const ShaderGenerator& rhs) = delete;
     ShaderGenerator& operator=(const ShaderGenerator& rhs) = delete;
@@ -52,7 +53,7 @@ private:
 
     using Blob = std::vector<char>;
     static Blob transpileShader(ShaderStage stage, std::string shader, Backend backend,
-            bool isMobile, const filament::SamplerInterfaceBlock* sib = nullptr) noexcept;
+            bool isMobile, const filamat::DescriptorSets& descriptorSets) noexcept;
 
     Backend mBackend;