Test commit

This reverts commit 30387af61c.

Causes breakage on Pixel 8pro for 1P apps.

.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

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:1.54.3'
 }
 ```
 
@@ -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', '~> 1.54.3'
 ```
 
-### 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
@@ -176,6 +166,7 @@ steps:
   - [x] KHR_materials_unlit
   - [x] KHR_materials_variants
   - [x] KHR_materials_volume
+  - [x] KHR_materials_specular
   - [x] KHR_mesh_quantization
   - [x] KHR_texture_basisu
   - [x] KHR_texture_transform

RELEASE_NOTES.md

@@ -7,6 +7,36 @@ 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
+
+
 ## v1.53.1
 
 

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

@@ -571,3 +571,9 @@ Java_com_google_android_filament_Engine_nBuilderBuild(JNIEnv*, jclass, jlong nat
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
     return (jlong) builder->build();
 }
+
+extern "C"
+JNIEXPORT jlong JNICALL
+Java_com_google_android_filament_Engine_getSteadyClockTimeNano(JNIEnv *env, jclass clazz) {
+    return (jlong)Engine::getSteadyClockTimeNano();
+}

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

@@ -25,12 +25,17 @@ using namespace filament;
 
 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Material_nBuilderBuild(JNIEnv *env, jclass,
-        jlong nativeEngine, jobject buffer_, jint size) {
+        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount) {
     Engine* engine = (Engine*) nativeEngine;
     AutoBuffer buffer(env, buffer_, size);
-    Material* material = Material::Builder()
+    auto builder = Material::Builder();
+    if (shBandCount) {
+        builder.sphericalHarmonicsBandCount(shBandCount);
+    }
+    Material* material = builder
             .package(buffer.getData(), buffer.getSize())
             .build(*engine);
+
     return (jlong) material;
 }
 

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) {
@@ -187,3 +195,10 @@ Java_com_google_android_filament_Renderer_nSetPresentationTime(JNIEnv *, jclass
     Renderer *renderer = (Renderer *) nativeRenderer;
     renderer->setPresentationTime(monotonicClockNanos);
 }
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Renderer_nSetVsyncTime(JNIEnv *, jclass,
+    jlong nativeRenderer, jlong steadyClockTimeNano) {
+    Renderer *renderer = (Renderer *) nativeRenderer;
+    renderer->setVsyncTime(steadyClockTimeNano);
+}

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.
@@ -1292,7 +1295,6 @@ public class Engine {
     /**
      * Switch the command queue to unprotected mode. Protected mode can be activated via
      * Renderer::beginFrame() using a protected SwapChain.
-
      * @see Renderer
      * @see SwapChain
      */
@@ -1300,6 +1302,13 @@ public class Engine {
         nUnprotected(getNativeObject());
     }
 
+    /**
+     * Get the current time. This is a convenience function that simply returns the
+     * time in nanosecond since epoch of std::chrono::steady_clock.
+     * @return current time in nanosecond since epoch of std::chrono::steady_clock.
+     * @see Renderer#beginFrame
+     */
+    public static native long getSteadyClockTimeNano();
 
     @UsedByReflection("TextureHelper.java")
     public long getNativeObject() {

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

@@ -346,6 +346,7 @@ public class Material {
     public static class Builder {
         private Buffer mBuffer;
         private int mSize;
+        private int mShBandCount = 0;
 
         /**
          * Specifies the material data. The material data is a binary blob produced by
@@ -361,6 +362,22 @@ public class Material {
             return this;
         }
 
+        /**
+         * Sets the quality of the indirect lights computations. This is only taken into account
+         * if this material is lit and in the surface domain. This setting will affect the
+         * IndirectLight computation if one is specified on the Scene and Spherical Harmonics
+         * are used for the irradiance.
+         *
+         * @param shBandCount Number of spherical harmonic bands. Must be 1, 2 or 3 (default).
+         * @return Reference to this Builder for chaining calls.
+         * @see IndirectLight
+         */
+        @NonNull
+        public Builder sphericalHarmonicsBandCount(@IntRange(from = 0) int shBandCount) {
+            mShBandCount = shBandCount;
+            return this;
+        }
+
         /**
          * Creates and returns the Material object.
          *
@@ -372,7 +389,8 @@ public class Material {
          */
         @NonNull
         public Material build(@NonNull Engine engine) {
-            long nativeMaterial = nBuilderBuild(engine.getNativeObject(), mBuffer, mSize);
+            long nativeMaterial = nBuilderBuild(engine.getNativeObject(),
+                mBuffer, mSize, mShBandCount);
             if (nativeMaterial == 0) throw new IllegalStateException("Couldn't create Material");
             return new Material(nativeMaterial);
         }
@@ -1023,7 +1041,7 @@ public class Material {
         mNativeObject = 0;
     }
 
-    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size);
+    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount);
     private static native long nCreateInstance(long nativeMaterial);
     private static native long nCreateInstanceWithName(long nativeMaterial, @NonNull String name);
     private static native long nGetDefaultInstance(long nativeMaterial);

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

@@ -284,6 +284,33 @@ public class Renderer {
         nSetPresentationTime(getNativeObject(), monotonicClockNanos);
     }
 
+    /**
+     * The use of this method is optional. It sets the VSYNC time expressed as the duration in
+     * nanosecond since epoch of std::chrono::steady_clock.
+     * If called, passing 0 to frameTimeNanos in Renderer.BeginFrame will use this
+     * time instead.
+     * @param steadyClockTimeNano duration in nanosecond since epoch of std::chrono::steady_clock
+     * @see Engine#getSteadyClockTimeNano
+     * @see Renderer#beginFrame
+     */
+    public void setVsyncTime(long steadyClockTimeNano) {
+        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
@@ -702,6 +729,8 @@ 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=1.54.3
 
 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

filament/backend/include/backend/DriverEnums.h

@@ -1058,7 +1058,7 @@ struct RasterState {
             bool inverseFrontFaces                      : 1;        // 31
 
             //! padding, must be 0
-            uint8_t padding                             : 1;        // 32
+            bool depthClamp                             : 1;        // 32
         };
         uint32_t u = 0;
     };
@@ -1244,7 +1244,7 @@ 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

filament/backend/include/backend/Platform.h

@@ -75,6 +75,8 @@ public:
          */
         size_t textureUseAfterFreePoolSize = 0;
 
+        size_t metalUploadBufferSizeBytes = 512 * 1024;
+
         /**
          * Set to `true` to forcibly disable parallel shader compilation in the backend.
          * Currently only honored by the GL and Metal backends.

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
  * -----------------------
@@ -305,6 +309,7 @@ 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)
@@ -439,12 +444,10 @@ DECL_DRIVER_API_N(setPushConstant,
         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/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>
 
@@ -39,7 +40,7 @@
 
 #define HandleAllocatorGL  HandleAllocator<32,  64, 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 {
 
@@ -173,14 +174,28 @@ 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();
             }
         }
 
         return static_cast<Dp>(p);
     }
 
+    template<typename B>
+    bool is_valid(Handle<B>& handle) {
+        if (handle && isPoolHandle(handle.getId())) {
+            auto [p, tag] = handleToPointer(handle.getId());
+            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;
+            return expectedAge == age;
+        }
+        return true;
+    }
+
     template<typename Dp, typename B>
     inline typename std::enable_if_t<
             std::is_pointer_v<Dp> &&
@@ -189,6 +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>
@@ -306,12 +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;
@@ -326,7 +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 };
         }
@@ -341,7 +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;
     }
@@ -351,6 +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/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/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;
 
@@ -204,6 +206,15 @@ public:
 
 private:
 
+    enum class UploadStrategy {
+        POOL,
+        BUMP_ALLOCATOR,
+    };
+
+    void uploadWithPoolBuffer(void* src, size_t size, size_t byteOffset) const;
+    void uploadWithBumpAllocator(void* src, size_t size, size_t byteOffset) const;
+
+    UploadStrategy mUploadStrategy;
     TrackedMetalBuffer mBuffer;
     size_t mBufferSize = 0;
     void* mCpuBuffer = nullptr;

filament/backend/src/metal/MetalBuffer.mm

@@ -27,7 +27,16 @@ MetalPlatform* TrackedMetalBuffer::platform = nullptr;
 MetalPlatform* ScopedAllocationTimer::platform = nullptr;
 
 MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
-        size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
+        size_t size, bool forceGpuBuffer)
+    : mBufferSize(size), mContext(context) {
+    const MetalBumpAllocator& allocator = *mContext.bumpAllocator;
+    // VERTEX is also used for index buffers
+    if (allocator.getCapacity() > 0 && bindingType == BufferObjectBinding::VERTEX) {
+        mUploadStrategy = UploadStrategy::BUMP_ALLOCATOR;
+    } else {
+        mUploadStrategy = UploadStrategy::POOL;
+    }
+
     // 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.
@@ -44,8 +53,8 @@ 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() {
@@ -61,34 +70,23 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     FILAMENT_CHECK_PRECONDITION(size + byteOffset <= mBufferSize)
             << "Attempting to copy " << size << " bytes into a buffer of size " << mBufferSize
             << " at offset " << byteOffset;
+    // The copy blit requires that byteOffset be a multiple of 4.
+    FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3)) << "byteOffset must be a multiple of 4";
 
-    // Either copy into the Metal buffer or into our cpu buffer.
+    // If we have a cpu buffer, we can directly copy into it.
     if (mCpuBuffer) {
         memcpy(static_cast<uint8_t*>(mCpuBuffer) + byteOffset, src, size);
         return;
     }
 
-    // Acquire a staging buffer to hold the contents of this update.
-    MetalBufferPool* bufferPool = mContext.bufferPool;
-    const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
-    memcpy(staging->buffer.get().contents, src, size);
-
-    // The blit below requires that byteOffset be a multiple of 4.
-    FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3u)) << "byteOffset must be a multiple of 4";
-
-    // Encode a blit from the staging buffer into the private GPU buffer.
-    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
-    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
-    blitEncoder.label = @"Buffer upload blit";
-    [blitEncoder copyFromBuffer:staging->buffer.get()
-                   sourceOffset:0
-                       toBuffer:mBuffer.get()
-              destinationOffset:byteOffset
-                           size:size];
-    [blitEncoder endEncoding];
-    [cmdBuffer addCompletedHandler:^(id<MTLCommandBuffer> cb) {
-        bufferPool->releaseBuffer(staging);
-    }];
+    switch (mUploadStrategy) {
+        case UploadStrategy::BUMP_ALLOCATOR:
+            uploadWithBumpAllocator(src, size, byteOffset);
+            break;
+        case UploadStrategy::POOL:
+            uploadWithPoolBuffer(src, size, byteOffset);
+            break;
+    }
 }
 
 void MetalBuffer::copyIntoBufferUnsynchronized(void* src, size_t size, size_t byteOffset) {
@@ -199,5 +197,42 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
     }
 }
 
+void MetalBuffer::uploadWithPoolBuffer(void* src, size_t size, size_t byteOffset) const {
+    MetalBufferPool* bufferPool = mContext.bufferPool;
+    const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
+    memcpy(staging->buffer.get().contents, src, size);
+
+    // Encode a blit from the staging buffer into the private GPU buffer.
+    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
+    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
+    blitEncoder.label = @"Buffer upload blit - pool buffer";
+    [blitEncoder copyFromBuffer:staging->buffer.get()
+                   sourceOffset:0
+                       toBuffer:mBuffer.get()
+              destinationOffset:byteOffset
+                           size:size];
+    [blitEncoder endEncoding];
+    [cmdBuffer addCompletedHandler:^(id<MTLCommandBuffer> cb) {
+        bufferPool->releaseBuffer(staging);
+    }];
+}
+
+void MetalBuffer::uploadWithBumpAllocator(void* src, size_t size, size_t byteOffset) const {
+    MetalBumpAllocator& allocator = *mContext.bumpAllocator;
+    auto [buffer, offset] = allocator.allocateStagingArea(size);
+    memcpy(static_cast<char*>(buffer.contents) + offset, src, size);
+
+    // Encode a blit from the staging buffer into the private GPU buffer.
+    id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
+    id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
+    blitEncoder.label = @"Buffer upload blit - bump allocator";
+    [blitEncoder copyFromBuffer:buffer
+                   sourceOffset:offset
+                       toBuffer:mBuffer.get()
+              destinationOffset:byteOffset
+                           size:size];
+    [blitEncoder endEncoding];
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalBufferPool.h

@@ -38,6 +38,28 @@ struct MetalBufferPoolEntry {
     mutable uint32_t referenceCount;
 };
 
+class MetalBumpAllocator {
+public:
+    MetalBumpAllocator(id<MTLDevice> device, size_t capacity);
+
+    /**
+     * Allocates a staging area of the given size. Returns a pair of the buffer and the offset
+     * within the buffer. The buffer is guaranteed to be at least the given size, but may be larger.
+     * Clients must not write to the buffer beyond the returned offset + size.
+     * Clients are responsible for holding a reference to the returned buffer.
+     * Allocations are guaranteed to be aligned to 4 bytes.
+     */
+    std::pair<id<MTLBuffer>, size_t> allocateStagingArea(size_t size);
+
+    size_t getCapacity() const noexcept { return mCapacity; }
+
+private:
+    id<MTLDevice> mDevice;
+    TrackedMetalBuffer mCurrentUploadBuffer = nil;
+    size_t mHead = 0;
+    size_t mCapacity;
+};
+
 // Manages a pool of Metal buffers, periodically releasing ones that have been unused for awhile.
 class MetalBufferPool {
 public:

filament/backend/src/metal/MetalBufferPool.mm

@@ -116,5 +116,39 @@ void MetalBufferPool::reset() noexcept {
     mFreeStages.clear();
 }
 
+MetalBumpAllocator::MetalBumpAllocator(id<MTLDevice> device, size_t capacity)
+    : mDevice(device), mCapacity(capacity) {
+    if (mCapacity > 0) {
+        mCurrentUploadBuffer = { [device newBufferWithLength:capacity options:MTLStorageModeShared],
+            TrackedMetalBuffer::Type::STAGING };
+    }
+}
+
+std::pair<id<MTLBuffer>, size_t> MetalBumpAllocator::allocateStagingArea(size_t size) {
+    if (size == 0) {
+        return { nil, 0 };
+    }
+    if (size > mCapacity) {
+        return { [mDevice newBufferWithLength:size options:MTLStorageModeShared], 0 };
+    }
+    assert_invariant(mCurrentUploadBuffer);
+
+    // Align the head to a 4-byte boundary.
+    mHead = (mHead + 3) & ~3;
+
+    if (UTILS_LIKELY(mHead + size <= mCapacity)) {
+        const size_t oldHead = mHead;
+        mHead += size;
+        return { mCurrentUploadBuffer.get(), oldHead };
+    }
+
+    // We're finished with the current allocation.
+    mCurrentUploadBuffer = { [mDevice newBufferWithLength:mCapacity options:MTLStorageModeShared],
+        TrackedMetalBuffer::Type::STAGING };
+    mHead = size;
+
+    return { mCurrentUploadBuffer.get(), 0 };
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalContext.h

@@ -44,6 +44,7 @@ namespace backend {
 class MetalDriver;
 class MetalBlitter;
 class MetalBufferPool;
+class MetalBumpAllocator;
 class MetalRenderTarget;
 class MetalSamplerGroup;
 class MetalSwapChain;
@@ -92,7 +93,7 @@ struct MetalContext {
     } highestSupportedGpuFamily;
 
     struct {
-        bool a8xStaticTextureTargetError;
+        bool staticTextureTargetError;
     } bugs;
 
     // sampleCountLookup[requestedSamples] gives a <= sample count supported by the device.
@@ -111,6 +112,7 @@ struct MetalContext {
     std::array<BufferState, MAX_SSBO_COUNT> ssboState;
     CullModeStateTracker cullModeState;
     WindingStateTracker windingState;
+    DepthClampStateTracker depthClampState;
     Handle<HwRenderPrimitive> currentRenderPrimitive;
 
     // State caches.
@@ -141,6 +143,7 @@ struct MetalContext {
     utils::FixedCircularBuffer<Handle<HwTexture>> texturesToDestroy;
 
     MetalBufferPool* bufferPool;
+    MetalBumpAllocator* bumpAllocator;
 
     MetalSwapChain* currentDrawSwapChain = nil;
     MetalSwapChain* currentReadSwapChain = nil;

filament/backend/src/metal/MetalDriver.h

@@ -57,7 +57,6 @@ class MetalDriver final : public DriverBase {
 
 public:
     static Driver* create(MetalPlatform* platform, const Platform::DriverConfig& driverConfig);
-    void runAtNextTick(const std::function<void()>& fn) noexcept;
 
 private:
 
@@ -73,10 +72,12 @@ 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;
 
     /*
      * Driver interface

filament/backend/src/metal/MetalDriver.mm

@@ -53,14 +53,14 @@ Driver* MetalDriverFactory::create(MetalPlatform* const platform, const Platform
     //    MetalRenderPrimitive         :  24       many
     //    MetalVertexBuffer            :  32       moderate
     // -- less than or equal 32 bytes
-    //    MetalIndexBuffer             :  40       moderate
     //    MetalFence                   :  48       few
-    //    MetalBufferObject            :  48       many
-    // -- less than or equal 48 bytes
+    //    MetalIndexBuffer             :  56       moderate
+    //    MetalBufferObject            :  64       many
+    // -- less than or equal 64 bytes
     //    MetalSamplerGroup            : 112       few
     //    MetalProgram                 : 152       moderate
     //    MetalTexture                 : 152       moderate
-    //    MetalSwapChain               : 184       few
+    //    MetalSwapChain               : 208       few
     //    MetalRenderTarget            : 272       few
     //    MetalVertexBufferInfo        : 552       moderate
     // -- less than or equal to 552 bytes
@@ -162,8 +162,10 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
         sc[s] = [mContext->device supportsTextureSampleCount:s] ? s : sc[s - 1];
     }
 
-    mContext->bugs.a8xStaticTextureTargetError =
-            [mContext->device.name containsString:@"Apple A8X GPU"];
+    mContext->bugs.staticTextureTargetError =
+            [mContext->device.name containsString:@"Apple A8X GPU"] ||
+            [mContext->device.name containsString:@"Apple A8 GPU"] ||
+            [mContext->device.name containsString:@"Apple A7 GPU"];
 
     mContext->commandQueue = mPlatform.createCommandQueue(mContext->device);
     mContext->pipelineStateCache.setDevice(mContext->device);
@@ -171,6 +173,8 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
     mContext->samplerStateCache.setDevice(mContext->device);
     mContext->argumentEncoderCache.setDevice(mContext->device);
     mContext->bufferPool = new MetalBufferPool(*mContext);
+    mContext->bumpAllocator =
+            new MetalBumpAllocator(mContext->device, driverConfig.metalUploadBufferSizeBytes);
     mContext->blitter = new MetalBlitter(*mContext);
 
     if (@available(iOS 12, *)) {
@@ -209,6 +213,7 @@ MetalDriver::~MetalDriver() noexcept {
     mContext->emptyTexture = nil;
     CFRelease(mContext->textureCache);
     delete mContext->bufferPool;
+    delete mContext->bumpAllocator;
     delete mContext->blitter;
     delete mContext->timerQueryImpl;
     delete mContext->shaderCompiler;
@@ -315,17 +320,40 @@ void MetalDriver::createVertexBufferR(Handle<HwVertexBuffer> vbh,
         uint32_t vertexCount, Handle<HwVertexBufferInfo> vbih) {
     MetalVertexBufferInfo const* const vbi = handle_cast<const MetalVertexBufferInfo>(vbih);
     construct_handle<MetalVertexBuffer>(vbh, *mContext, vertexCount, vbi->bufferCount, vbih);
+    // No actual GPU memory is allocated here, so no need to check for allocation success.
 }
 
 void MetalDriver::createIndexBufferR(Handle<HwIndexBuffer> ibh, ElementType elementType,
         uint32_t indexCount, BufferUsage usage) {
-    auto elementSize = (uint8_t) getElementTypeSize(elementType);
-    construct_handle<MetalIndexBuffer>(ibh, *mContext, usage, elementSize, indexCount);
+    auto elementSize = (uint8_t)getElementTypeSize(elementType);
+    auto* indexBuffer =
+            construct_handle<MetalIndexBuffer>(ibh, *mContext, usage, elementSize, indexCount);
+    auto& buffer = indexBuffer->buffer;
+    // If the allocation was not successful, postpone the error message until the next tick, to give
+    // Filament a chance to call setDebugTag on the handle; this way we get a nicer error message.
+    if (UTILS_UNLIKELY(!buffer.wasAllocationSuccessful())) {
+        const size_t byteCount = buffer.getSize();
+        runAtNextTick([byteCount, this, ibh]() {
+            FILAMENT_CHECK_POSTCONDITION(false)
+                    << "Could not allocate Metal index buffer of size " << byteCount
+                    << ", tag=" << mHandleAllocator.getHandleTag(ibh.getId()).c_str_safe();
+        });
+    }
 }
 
 void MetalDriver::createBufferObjectR(Handle<HwBufferObject> boh, uint32_t byteCount,
         BufferObjectBinding bindingType, BufferUsage usage) {
-    construct_handle<MetalBufferObject>(boh, *mContext, bindingType, usage, byteCount);
+    auto* bufferObject =
+            construct_handle<MetalBufferObject>(boh, *mContext, bindingType, usage, byteCount);
+    // If the allocation was not successful, postpone the error message until the next tick, to give
+    // Filament a chance to call setDebugTag on the handle; this way we get a nicer error message.
+    if (UTILS_UNLIKELY(!bufferObject->getBuffer()->wasAllocationSuccessful())) {
+        runAtNextTick([byteCount, this, boh]() {
+            FILAMENT_CHECK_POSTCONDITION(false)
+                    << "Could not allocate Metal buffer of size " << byteCount
+                    << ", tag=" << mHandleAllocator.getHandleTag(boh.getId()).c_str_safe();
+        });
+    }
 }
 
 void MetalDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint8_t levels,
@@ -829,6 +857,10 @@ bool MetalDriver::isProtectedTexturesSupported() {
     return false;
 }
 
+bool MetalDriver::isDepthClampSupported() {
+    return true;
+}
+
 bool MetalDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU:
@@ -837,8 +869,8 @@ bool MetalDriver::isWorkaroundNeeded(Workaround workaround) {
             return true;
         case Workaround::ADRENO_UNIFORM_ARRAY_CRASH:
             return false;
-        case Workaround::A8X_STATIC_TEXTURE_TARGET_ERROR:
-            return mContext->bugs.a8xStaticTextureTargetError;
+        case Workaround::METAL_STATIC_TEXTURE_TARGET_ERROR:
+            return mContext->bugs.staticTextureTargetError;
         case Workaround::DISABLE_BLIT_INTO_TEXTURE_ARRAY:
             return false;
         default:
@@ -1270,11 +1302,11 @@ void MetalDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
     pushConstants.setPushConstant(value, index);
 }
 
-void MetalDriver::insertEventMarker(const char* string, uint32_t len) {
+void MetalDriver::insertEventMarker(const char* string) {
 
 }
 
-void MetalDriver::pushGroupMarker(const char* string, uint32_t len) {
+void MetalDriver::pushGroupMarker(const char* string) {
     mContext->groupMarkers.push(string);
 }
 
@@ -1746,6 +1778,13 @@ void MetalDriver::bindPipeline(PipelineState const& ps) {
         [mContext->currentRenderPassEncoder setFrontFacingWinding:winding];
     }
 
+    // depth clip mode
+    MTLDepthClipMode depthClipMode = rs.depthClamp ? MTLDepthClipModeClamp : MTLDepthClipModeClip;
+    mContext->depthClampState.updateState(depthClipMode);
+    if (mContext->depthClampState.stateChanged()) {
+        [mContext->currentRenderPassEncoder setDepthClipMode:depthClipMode];
+    }
+
     // Set the depth-stencil state, if a state change is needed.
     DepthStencilState depthState;
     if (depthAttachment) {
@@ -1755,6 +1794,8 @@ void MetalDriver::bindPipeline(PipelineState const& ps) {
     if (stencilAttachment) {
         const auto& ss = ps.stencilState;
 
+        // Ben test
+
         auto& front = depthState.front;
         front.stencilCompare = getMetalCompareFunction(ss.front.stencilFunc);
         front.stencilOperationStencilFail = getMetalStencilOperation(ss.front.stencilOpStencilFail);
@@ -2056,16 +2097,17 @@ void MetalDriver::enumerateBoundBuffers(BufferObjectBinding bindingType,
 void MetalDriver::resetState(int) {
 }
 
+void MetalDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
+    mHandleAllocator.associateTagToHandle(handleId, std::move(tag));
+}
+
 void MetalDriver::runAtNextTick(const std::function<void()>& fn) noexcept {
-    std::lock_guard<std::mutex> const lock(mTickOpsLock);
     mTickOps.push_back(fn);
 }
 
 void MetalDriver::executeTickOps() noexcept {
     std::vector<std::function<void()>> ops;
-    mTickOpsLock.lock();
     std::swap(ops, mTickOps);
-    mTickOpsLock.unlock();
     for (const auto& f : ops) {
         f();
     }

filament/backend/src/metal/MetalHandles.mm

@@ -257,10 +257,6 @@ 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 +275,10 @@ 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:
@@ -556,8 +548,6 @@ 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 +562,6 @@ 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 +574,6 @@ 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
@@ -596,6 +582,12 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             break;
     }
 
+    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?";
+
     // 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 &&

filament/backend/src/metal/MetalState.h

@@ -382,6 +382,7 @@ using SamplerStateCache = StateCache<SamplerState, id<MTLSamplerState>, SamplerS
 
 using CullModeStateTracker = StateTracker<MTLCullMode>;
 using WindingStateTracker = StateTracker<MTLWinding>;
+using DepthClampStateTracker = StateTracker<MTLDepthClipMode>;
 
 // Argument encoder
 

filament/backend/src/noop/NoopDriver.cpp

@@ -202,6 +202,10 @@ bool NoopDriver::isProtectedTexturesSupported() {
     return true;
 }
 
+bool NoopDriver::isDepthClampSupported() {
+    return false;
+}
+
 bool NoopDriver::isWorkaroundNeeded(Workaround) {
     return false;
 }
@@ -316,10 +320,10 @@ 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 +392,7 @@ void NoopDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
 void NoopDriver::resetState(int) {
 }
 
+void NoopDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
+}
+
 } // namespace filament

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

@@ -220,8 +220,9 @@ 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 +240,10 @@ 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;
@@ -627,6 +628,7 @@ 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.cpp

@@ -83,12 +83,20 @@
 #define HAS_MAPBUFFERS 1
 #endif
 
-#define DEBUG_MARKER_NONE       0x00    // no debug marker
-#define DEBUG_MARKER_OPENGL     0x01    // markers in the gl command queue (req. driver support)
-#define DEBUG_MARKER_BACKEND    0x02    // markers on the backend side (systrace)
-#define DEBUG_MARKER_ALL        0x03    // all markers
+#define DEBUG_GROUP_MARKER_NONE       0x00    // no debug marker
+#define DEBUG_GROUP_MARKER_OPENGL     0x01    // markers in the gl command queue (req. driver support)
+#define DEBUG_GROUP_MARKER_BACKEND    0x02    // markers on the backend side (systrace)
+#define DEBUG_GROUP_MARKER_ALL        0x03    // all markers
 
-// set to the desired debug marker level
+#define DEBUG_MARKER_NONE             0x00    // no debug marker
+#define DEBUG_MARKER_OPENGL           0x01    // markers in the gl command queue (req. driver support)
+#define DEBUG_MARKER_BACKEND          0x02    // markers on the backend side (systrace)
+#define DEBUG_MARKER_ALL              0x03    // all markers
+
+// set to the desired debug marker level (for user markers [default: All])
+#define DEBUG_GROUP_MARKER_LEVEL      DEBUG_GROUP_MARKER_ALL
+
+// set to the desired debug level (for internal debugging [Default: None])
 #define DEBUG_MARKER_LEVEL      DEBUG_MARKER_NONE
 
 #if DEBUG_MARKER_LEVEL > DEBUG_MARKER_NONE
@@ -196,11 +204,37 @@ Driver* OpenGLDriver::create(OpenGLPlatform* const platform,
 
 OpenGLDriver::DebugMarker::DebugMarker(OpenGLDriver& driver, const char* string) noexcept
         : driver(driver) {
-    driver.pushGroupMarker(string, strlen(string));
+#ifndef __EMSCRIPTEN__
+#ifdef GL_EXT_debug_marker
+#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_OPENGL
+    if (UTILS_LIKELY(driver.getContext().ext.EXT_debug_marker)) {
+        glPushGroupMarkerEXT(GLsizei(strlen(string)), string);
+    }
+#endif
+#endif
+
+#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_BACKEND
+    SYSTRACE_CONTEXT();
+    SYSTRACE_NAME_BEGIN(string);
+#endif
+#endif
 }
 
 OpenGLDriver::DebugMarker::~DebugMarker() noexcept {
-    driver.popGroupMarker();
+#ifndef __EMSCRIPTEN__
+#ifdef GL_EXT_debug_marker
+#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_OPENGL
+    if (UTILS_LIKELY(driver.getContext().ext.EXT_debug_marker)) {
+        glPopGroupMarkerEXT();
+    }
+#endif
+#endif
+
+#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_BACKEND
+    SYSTRACE_CONTEXT();
+    SYSTRACE_NAME_END();
+#endif
+#endif
 }
 
 // ------------------------------------------------------------------------------------------------
@@ -304,6 +338,13 @@ void OpenGLDriver::bindSampler(GLuint unit, GLuint sampler) noexcept {
 void OpenGLDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
         backend::PushConstantVariant value) {
     assert_invariant(stage == ShaderStage::VERTEX || stage == ShaderStage::FRAGMENT);
+
+#if FILAMENT_ENABLE_MATDBG
+    if (UTILS_UNLIKELY(!mValidProgram)) {
+        return;
+    }
+#endif
+
     utils::Slice<std::pair<GLint, ConstantType>> constants;
     if (stage == ShaderStage::VERTEX) {
         constants = mCurrentPushConstants->vertexConstants;
@@ -340,15 +381,11 @@ void OpenGLDriver::bindTexture(GLuint unit, GLTexture const* t) noexcept {
 }
 
 bool OpenGLDriver::useProgram(OpenGLProgram* p) noexcept {
-    if (UTILS_UNLIKELY(!p->isValid())) {
-        // If the program is not valid, we can't call use().
-        return false;
-    }
-
     // set-up textures and samplers in the proper TMUs (as specified in setSamplers)
-    p->use(this, mContext);
+    bool const success = p->use(this, mContext);
+    assert_invariant(success == p->isValid());
 
-    if (UTILS_UNLIKELY(mContext.isES2())) {
+    if (UTILS_UNLIKELY(mContext.isES2() && success)) {
         for (uint32_t i = 0; i < Program::UNIFORM_BINDING_COUNT; i++) {
             auto [id, buffer, age] = mContext.getEs2UniformBinding(i);
             if (buffer) {
@@ -359,7 +396,8 @@ bool OpenGLDriver::useProgram(OpenGLProgram* p) noexcept {
         // when mPlatform.isSRGBSwapChainSupported() is false (no need to check though).
         p->setRec709ColorSpace(mRec709OutputColorspace);
     }
-    return true;
+
+    return success;
 }
 
 
@@ -413,6 +451,14 @@ void OpenGLDriver::setRasterState(RasterState rs) noexcept {
     } else {
         gl.disable(GL_SAMPLE_ALPHA_TO_COVERAGE);
     }
+
+    if (gl.ext.EXT_depth_clamp) {
+        if (rs.depthClamp) {
+            gl.enable(GL_DEPTH_CLAMP);
+        } else {
+            gl.disable(GL_DEPTH_CLAMP);
+        }
+    }
 }
 
 void OpenGLDriver::setStencilState(StencilState ss) noexcept {
@@ -2081,6 +2127,10 @@ bool OpenGLDriver::isProtectedTexturesSupported() {
     return getContext().ext.EXT_protected_textures;
 }
 
+bool OpenGLDriver::isDepthClampSupported() {
+    return getContext().ext.EXT_depth_clamp;
+}
+
 bool OpenGLDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU:
@@ -2172,6 +2222,10 @@ void OpenGLDriver::makeCurrent(Handle<HwSwapChain> schDraw, Handle<HwSwapChain>
 // Updating driver objects
 // ------------------------------------------------------------------------------------------------
 
+void OpenGLDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {
+    mHandleAllocator.associateTagToHandle(handleId, std::move(tag));
+}
+
 void OpenGLDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh,
         uint32_t index, Handle<HwBufferObject> boh) {
    DEBUG_MARKER()
@@ -2332,11 +2386,20 @@ void OpenGLDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
     auto const* const pSamplers = (SamplerDescriptor const*)data.buffer;
     for (size_t i = 0, c = sb->textureUnitEntries.size(); i < c; i++) {
         GLuint samplerId = 0u;
-        const GLTexture* t = nullptr;
-        if (UTILS_LIKELY(pSamplers[i].t)) {
-            t = handle_cast<const GLTexture*>(pSamplers[i].t);
+        Handle<HwTexture> th = pSamplers[i].t;
+        if (UTILS_LIKELY(th)) {
+            GLTexture const* const t = handle_cast<const GLTexture*>(th);
             assert_invariant(t);
 
+            if (UTILS_UNLIKELY(es2)
+#if defined(GL_EXT_texture_filter_anisotropic)
+                    || UTILS_UNLIKELY(anisotropyWorkaround)
+ #endif
+                    ) {
+                // We must set texture parameters on the texture itself.
+                bindTexture(OpenGLContext::DUMMY_TEXTURE_BINDING, t);
+            }
+
             SamplerParams params = pSamplers[i].s;
             if (UTILS_UNLIKELY(t->target == SamplerType::SAMPLER_EXTERNAL)) {
                 // From OES_EGL_image_external spec:
@@ -2390,7 +2453,7 @@ void OpenGLDriver::updateSamplerGroup(Handle<HwSamplerGroup> sbh,
             // which is not an error.
         }
 
-        sb->textureUnitEntries[i] = { t, samplerId };
+        sb->textureUnitEntries[i] = { th, samplerId };
     }
     scheduleDestroy(std::move(data));
 }
@@ -3133,28 +3196,28 @@ void OpenGLDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
     CHECK_GL_ERROR(utils::slog.e)
 }
 
-void OpenGLDriver::insertEventMarker(char const* string, uint32_t len) {
+void OpenGLDriver::insertEventMarker(char const* string) {
 #ifndef __EMSCRIPTEN__
 #ifdef GL_EXT_debug_marker
     auto& gl = mContext;
     if (gl.ext.EXT_debug_marker) {
-        glInsertEventMarkerEXT(GLsizei(len ? len : strlen(string)), string);
+        glInsertEventMarkerEXT(GLsizei(strlen(string)), string);
     }
 #endif
 #endif
 }
 
-void OpenGLDriver::pushGroupMarker(char const* string, uint32_t len) {
+void OpenGLDriver::pushGroupMarker(char const* string) {
 #ifndef __EMSCRIPTEN__
 #ifdef GL_EXT_debug_marker
-#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_OPENGL
+#if DEBUG_GROUP_MARKER_LEVEL & DEBUG_GROUP_MARKER_OPENGL
     if (UTILS_LIKELY(mContext.ext.EXT_debug_marker)) {
-        glPushGroupMarkerEXT(GLsizei(len ? len : strlen(string)), string);
+        glPushGroupMarkerEXT(GLsizei(strlen(string)), string);
     }
 #endif
 #endif
 
-#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_BACKEND
+#if DEBUG_GROUP_MARKER_LEVEL & DEBUG_GROUP_MARKER_BACKEND
     SYSTRACE_CONTEXT();
     SYSTRACE_NAME_BEGIN(string);
 #endif
@@ -3164,14 +3227,14 @@ void OpenGLDriver::pushGroupMarker(char const* string, uint32_t len) {
 void OpenGLDriver::popGroupMarker(int) {
 #ifndef __EMSCRIPTEN__
 #ifdef GL_EXT_debug_marker
-#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_OPENGL
+#if DEBUG_GROUP_MARKER_LEVEL & DEBUG_GROUP_MARKER_OPENGL
     if (UTILS_LIKELY(mContext.ext.EXT_debug_marker)) {
         glPopGroupMarkerEXT();
     }
 #endif
 #endif
 
-#if DEBUG_MARKER_LEVEL & DEBUG_MARKER_BACKEND
+#if DEBUG_GROUP_MARKER_LEVEL & DEBUG_GROUP_MARKER_BACKEND
     SYSTRACE_CONTEXT();
     SYSTRACE_NAME_END();
 #endif
@@ -3877,6 +3940,7 @@ void OpenGLDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
 }
 
 void OpenGLDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+    DEBUG_MARKER()
     GLRenderPrimitive const* const rp = mBoundRenderPrimitive;
     if (UTILS_UNLIKELY(!rp || !mValidProgram)) {
         return;
@@ -3898,6 +3962,7 @@ void OpenGLDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t ins
 }
 
 void OpenGLDriver::draw2GLES2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+    DEBUG_MARKER()
     GLRenderPrimitive const* const rp = mBoundRenderPrimitive;
     if (UTILS_UNLIKELY(!rp || !mValidProgram)) {
         return;
@@ -3918,6 +3983,7 @@ void OpenGLDriver::draw2GLES2(uint32_t indexOffset, uint32_t indexCount, uint32_
 }
 
 void OpenGLDriver::scissor(Viewport scissor) {
+    DEBUG_MARKER()
     setScissor(scissor);
 }
 
@@ -3937,6 +4003,7 @@ void OpenGLDriver::draw(PipelineState state, Handle<HwRenderPrimitive> rph,
 }
 
 void OpenGLDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGroupCount) {
+    DEBUG_MARKER()
     getShaderCompilerService().tick();
 
     OpenGLProgram* const p = handle_cast<OpenGLProgram*>(program);

filament/backend/src/opengl/OpenGLDriver.h

@@ -81,7 +81,7 @@ public:
             const Platform::DriverConfig& driverConfig) noexcept;
 
     class DebugMarker {
-        OpenGLDriver& driver;
+        UTILS_UNUSED OpenGLDriver& driver;
     public:
         DebugMarker(OpenGLDriver& driver, const char* string) noexcept;
         ~DebugMarker() noexcept;
@@ -126,7 +126,7 @@ public:
     struct GLSamplerGroup : public HwSamplerGroup {
         using HwSamplerGroup::HwSamplerGroup;
         struct Entry {
-            GLTexture const* texture = nullptr;
+            Handle<HwTexture> th;
             GLuint sampler = 0u;
         };
         utils::FixedCapacityVector<Entry> textureUnitEntries;
@@ -256,6 +256,11 @@ private:
         return mHandleAllocator.handle_cast<Dp, B>(handle);
     }
 
+    template<typename B>
+    bool is_valid(Handle<B>& handle) {
+        return mHandleAllocator.is_valid(handle);
+    }
+
     template<typename Dp, typename B>
     inline typename std::enable_if_t<
             std::is_pointer_v<Dp> &&

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -20,21 +20,24 @@
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"
 
+#include <backend/DriverEnums.h>
 #include <backend/Program.h>
+#include <backend/Handle.h>
 
-#include <private/backend/BackendUtils.h>
-
-#include <utils/debug.h>
 #include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Log.h>
 #include <utils/Systrace.h>
 
+#include <algorithm>
 #include <array>
 #include <string_view>
 #include <utility>
 #include <new>
 
 #include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
@@ -241,8 +244,9 @@ void OpenGLProgram::updateSamplers(OpenGLDriver* const gld) const noexcept {
         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
-            const GLTexture* const t = sb->textureUnitEntries[j].texture;
-            if (t) { // program may not use all samplers of sampler group
+            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)) {

filament/backend/src/opengl/OpenGLProgram.h

@@ -53,11 +53,21 @@ public:
 
     bool isValid() const noexcept { return mToken || gl.program != 0; }
 
-    void use(OpenGLDriver* const gld, OpenGLContext& context) noexcept {
-        if (UTILS_UNLIKELY(!gl.program)) {
+    bool use(OpenGLDriver* const gld, OpenGLContext& context) noexcept {
+        // both non-null is impossible by construction
+        assert_invariant(!mToken || !gl.program);
+
+        if (UTILS_UNLIKELY(mToken && !gl.program)) {
+            // first time a program is used
             initialize(*gld);
         }
 
+        if (UTILS_UNLIKELY(!gl.program)) {
+            // compilation failed (token should be null)
+            assert_invariant(!mToken);
+            return false;
+        }
+
         context.useProgram(gl.program);
         if (UTILS_UNLIKELY(mUsedBindingsCount)) {
             // We rely on GL state tracking to avoid unnecessary glBindTexture / glBindSampler
@@ -74,6 +84,7 @@ public:
 
             updateSamplers(gld);
         }
+        return true;
     }
 
     // For ES2 only

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;
@@ -359,7 +371,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
 GLuint ShaderCompilerService::getProgram(ShaderCompilerService::program_token_t& token) {
     GLuint const program = initialize(token);
     assert_invariant(token == nullptr);
-#ifndef FILAMENT_ENABLE_MATDBG
+#if !FILAMENT_ENABLE_MATDBG
     assert_invariant(program);
 #endif
     return program;
@@ -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
@@ -572,24 +590,38 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 
             // split shader source, so we can insert the specialization constants and the packing
             // functions
-            auto const [prolog, body] = splitShaderSource({ shader_src, shader_len });
+            auto [version, prolog, body] = splitShaderSource({ shader_src, shader_len });
 
-            const std::array<const char*, 4> sources = {
-                    prolog.data(),
-                    specializationConstantString.c_str(),
-                    packingFunctions.data(),
-                    body.data()
+            // enable ESSL 3.10 if available
+            if (context.isAtLeastGLES<3, 1>()) {
+                version = "#version 310 es\n";
+            }
+
+            std::array<std::string_view, 5> sources = {
+                version,
+                prolog,
+                specializationConstantString,
+                packingFunctions,
+                { body.data(), body.size() - 1 }  // null-terminated
             };
 
-            const std::array<GLint, 4> lengths = {
-                    (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
@@ -661,6 +693,7 @@ void ShaderCompilerService::process_OVR_multiview2(OpenGLContext& context,
 
 // Tragically, OpenGL 4.1 doesn't support unpackHalf2x16 (appeared in 4.2) and
 // macOS doesn't support GL_ARB_shading_language_packing
+// Also GLES3.0 didn't have the full set of packing/unpacking functions
 std::string_view ShaderCompilerService::process_ARB_shading_language_packing(OpenGLContext& context) noexcept {
     using namespace std::literals;
 #ifdef BACKEND_OPENGL_VERSION_GL
@@ -700,31 +733,102 @@ highp uint packHalf2x16(vec2 v) {
     highp uint y = fp32tou16(v.y);
     return (y << 16u) | x;
 }
+highp uint packUnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, 0.0, 1.0) * 255.0);
+    highp uint a = uint(v.x);
+    highp uint b = uint(v.y) <<  8;
+    highp uint c = uint(v.z) << 16;
+    highp uint d = uint(v.w) << 24;
+    return (a|b|c|d);
+}
+highp uint packSnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, -1.0, 1.0) * 127.0);
+    highp uint a = uint((int(v.x) & 0xff));
+    highp uint b = uint((int(v.y) & 0xff)) <<  8;
+    highp uint c = uint((int(v.z) & 0xff)) << 16;
+    highp uint d = uint((int(v.w) & 0xff)) << 24;
+    return (a|b|c|d);
+}
+mediump vec4 unpackUnorm4x8(highp uint v) {
+    return vec4(float((v & 0x000000ffu)      ),
+                float((v & 0x0000ff00u) >>  8),
+                float((v & 0x00ff0000u) >> 16),
+                float((v & 0xff000000u) >> 24)) / 255.0;
+}
+mediump vec4 unpackSnorm4x8(highp uint v) {
+    int a = int(((v       ) & 0xffu) << 24u) >> 24 ;
+    int b = int(((v >>  8u) & 0xffu) << 24u) >> 24 ;
+    int c = int(((v >> 16u) & 0xffu) << 24u) >> 24 ;
+    int d = int(((v >> 24u) & 0xffu) << 24u) >> 24 ;
+    return clamp(vec4(float(a), float(b), float(c), float(d)) / 127.0, -1.0, 1.0);
+}
 )"sv;
     }
 #endif // BACKEND_OPENGL_VERSION_GL
+
+#ifdef BACKEND_OPENGL_VERSION_GLES
+    if (!context.isES2() && !context.isAtLeastGLES<3, 1>()) {
+        return R"(
+
+highp uint packUnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, 0.0, 1.0) * 255.0);
+    highp uint a = uint(v.x);
+    highp uint b = uint(v.y) <<  8;
+    highp uint c = uint(v.z) << 16;
+    highp uint d = uint(v.w) << 24;
+    return (a|b|c|d);
+}
+highp uint packSnorm4x8(mediump vec4 v) {
+    v = round(clamp(v, -1.0, 1.0) * 127.0);
+    highp uint a = uint((int(v.x) & 0xff));
+    highp uint b = uint((int(v.y) & 0xff)) <<  8;
+    highp uint c = uint((int(v.z) & 0xff)) << 16;
+    highp uint d = uint((int(v.w) & 0xff)) << 24;
+    return (a|b|c|d);
+}
+mediump vec4 unpackUnorm4x8(highp uint v) {
+    return vec4(float((v & 0x000000ffu)      ),
+                float((v & 0x0000ff00u) >>  8),
+                float((v & 0x00ff0000u) >> 16),
+                float((v & 0xff000000u) >> 24)) / 255.0;
+}
+mediump vec4 unpackSnorm4x8(highp uint v) {
+    int a = int(((v       ) & 0xffu) << 24u) >> 24 ;
+    int b = int(((v >>  8u) & 0xffu) << 24u) >> 24 ;
+    int c = int(((v >> 16u) & 0xffu) << 24u) >> 24 ;
+    int d = int(((v >> 24u) & 0xffu) << 24u) >> 24 ;
+    return clamp(vec4(float(a), float(b), float(c), float(d)) / 127.0, -1.0, 1.0);
+}
+)"sv;
+    }
+#endif // BACKEND_OPENGL_VERSION_GLES
     return ""sv;
 }
 
-// split shader source code in two, the first section goes from the start to the line after the
-// last #extension, and the 2nd part goes from there to the end.
-std::array<std::string_view, 2> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
-    auto start = source.find("#version");
-    assert_invariant(start != std::string_view::npos);
+// split shader source code in three:
+// - the version line
+// - extensions
+// - everything else
+std::array<std::string_view, 3> ShaderCompilerService::splitShaderSource(std::string_view source) noexcept {
+    auto version_start = source.find("#version");
+    assert_invariant(version_start != std::string_view::npos);
 
-    auto pos = source.rfind("\n#extension");
-    if (pos == std::string_view::npos) {
-        pos = start;
+    auto version_eol = source.find('\n', version_start) + 1;
+    assert_invariant(version_eol != std::string_view::npos);
+
+    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, eol - start);
-    std::string_view const body = source.substr(version.length(), source.length() - version.length());
-    return { version, body };
+    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/ShaderCompilerService.h

@@ -146,7 +146,7 @@ private:
 
     static std::string_view process_ARB_shading_language_packing(OpenGLContext& context) noexcept;
 
-    static std::array<std::string_view, 2> splitShaderSource(std::string_view source) noexcept;
+    static std::array<std::string_view, 3> splitShaderSource(std::string_view source) noexcept;
 
     static GLuint linkProgram(OpenGLContext& context,
             std::array<GLuint, Program::SHADER_TYPE_COUNT> shaders,

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/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/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

@@ -595,6 +595,8 @@ 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 +611,8 @@ 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 +625,8 @@ 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();
@@ -637,7 +643,7 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
 
     auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth, mPlatform->getDevice(),
             mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, width, height,
-            samples, colorTargets, depthStencil, mStagePool);
+            samples, colorTargets, depthStencil, mStagePool, layerCount);
     mResourceManager.acquire(renderTarget);
 }
 
@@ -938,6 +944,10 @@ bool VulkanDriver::isProtectedTexturesSupported() {
     return false;
 }
 
+bool VulkanDriver::isDepthClampSupported() {
+    return mContext.isDepthClampSupported();
+}
+
 bool VulkanDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU: {
@@ -986,15 +996,15 @@ 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;
     }
 
@@ -1267,6 +1277,8 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
         }
     }
 
+    uint8_t const renderTargetLayerCount = rt->getLayerCount();
+
     // Create the VkRenderPass or fetch it from cache.
     VulkanFboCache::RenderPassKey rpkey = {
         .initialColorLayoutMask = 0,
@@ -1277,10 +1289,12 @@ 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);
         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) {
@@ -1592,13 +1606,13 @@ 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);
@@ -1806,6 +1820,7 @@ 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,
@@ -2022,6 +2037,10 @@ 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/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

@@ -16,7 +16,6 @@
 
 #include "VulkanHandles.h"
 
-#include "VulkanDriver.h"
 #include "VulkanConstants.h"
 #include "VulkanDriver.h"
 #include "VulkanMemory.h"
@@ -323,11 +322,12 @@ VulkanRenderTarget::VulkanRenderTarget(VkDevice device, VkPhysicalDevice physica
         VulkanContext const& context, VmaAllocator allocator, VulkanCommands* commands,
         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];
     }

filament/backend/src/vulkan/VulkanHandles.h

@@ -304,7 +304,7 @@ struct VulkanRenderTarget : private HwRenderTarget, VulkanResource {
             VulkanContext const& context, VmaAllocator allocator,
             VulkanCommands* commands, 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 +319,7 @@ 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 +331,7 @@ private:
     VulkanAttachment mMsaaDepthAttachment = {};
     const bool mOffscreen : 1;
     uint8_t mSamples : 7;
+    uint8_t mLayerCount = 1;
 };
 
 struct VulkanBufferObject;

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -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;

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/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/VulkanTexture.cpp

@@ -97,13 +97,11 @@ 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)) {
@@ -121,7 +119,7 @@ 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 +128,9 @@ 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.
@@ -387,12 +384,15 @@ void VulkanTexture::setPrimaryRange(uint32_t minMiplevel, uint32_t maxMiplevel)
 }
 
 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);
 }

filament/backend/src/vulkan/VulkanTexture.h

@@ -73,6 +73,10 @@ struct VulkanTexture : public HwTexture, VulkanResource {
     // 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

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

@@ -172,7 +172,7 @@ public:
         UTILS_UNUSED VkResult result = vkAllocateDescriptorSets(mDevice, &allocInfo, &vkSet);
         FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
                 << "Failed to allocate descriptor set code=" << result << " size=" << mSize
-                << " capacity=" << mCapacity << " count=" << "%s";
+                << " capacity=" << mCapacity << " count=" << mUnusedCount;
         mSize++;
         return createSet(layout->bitmask, vkSet);
     }

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

@@ -33,11 +33,11 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
 
     // build the push constant layout key
     uint32_t pushConstantRangeCount = program->getPushConstantRangeCount();
-    auto const& pushCostantRanges = program->getPushConstantRanges();    
+    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 +52,8 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
         }
     }
 
-    auto iter = mPipelineLayouts.find(key);
-    if (iter != mPipelineLayouts.end()) {
-        PipelineLayoutCacheEntry& entry = mPipelineLayouts[key];
+    if (PipelineLayoutMap::iterator iter = mPipelineLayouts.find(key); iter != mPipelineLayouts.end()) {
+        PipelineLayoutCacheEntry& entry = iter.value();
         entry.lastUsed = mTimestamp++;
         return entry.handle;
     }
@@ -65,7 +64,7 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
         .setLayoutCount = (uint32_t) descSetLayoutCount,
         .pSetLayouts = key.descSetLayouts.data(),
         .pushConstantRangeCount = pushConstantRangeCount,
-        .pPushConstantRanges = pushCostantRanges,
+        .pPushConstantRanges = pushConstantRanges,
     };
 
     VkPipelineLayout layout;

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/docs/SpirvDebugging.md

@@ -64,14 +64,15 @@ First determine the Filament material and variant that causes the problem.
 What we want is the "raw" GLSL version of the shader, before any optimizations / cross-compilation
 happens.
 
-The easiest way I've found to do this is by adding a `std::cout` statement inside `MaterialBuilder`,
-right after `sg.createFragmentProgram` or `sg.createVertexProgram` is called. For example,
+We can use the `--save-raw-variants` debug flag in matc to export each GLSL
+shader to a file. For example:
 
 ```
-if (v.variant.key == 8u) { std::cout << shader << std::endl; }
+matc --save-raw-variants --optimize-size --variant-filter fog,ssr,vsm,stereo \
+        -a all -p all -o mymaterial.filamat mymaterial.mat
 ```
 
-Then, I invoke `matc` and redirect stdout to a .frag or .vert file.
+Files will be named like `mymaterial_0x05.frag` or `mymaterial_0x05.vert`.
 
 Note that gltfio material "templates" first go through a build step. After building gltfio, the
 gltfio Filament materials are output to:

filament/include/filament/BufferObject.h

@@ -54,7 +54,7 @@ public:
     using BufferDescriptor = backend::BufferDescriptor;
     using BindingType = backend::BufferObjectBinding;
 
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -78,6 +78,21 @@ public:
          */
         Builder& bindingType(BindingType bindingType) noexcept;
 
+        /**
+         * Associate an optional name with this BufferObject for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this BufferObject
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the BufferObject and returns a pointer to it. After creation, the buffer
          * object is uninitialized. Use BufferObject::setBuffer() to initialize it.

filament/include/filament/Engine.h

@@ -298,6 +298,24 @@ public:
          */
         size_t textureUseAfterFreePoolSize = 0;
 
+        /**
+         * When uploading vertex or index data, the Filament Metal backend copies data
+         * into a shared staging area before transferring it to the GPU. This setting controls
+         * the total size of the buffer used to perform these allocations.
+         *
+         * Higher values can improve performance when performing many uploads across a small
+         * number of frames.
+         *
+         * This buffer remains alive throughout the lifetime of the Engine, so this size adds to the
+         * memory footprint of the app and should be set as conservative as possible.
+         *
+         * A value of 0 disables the shared staging buffer entirely; uploads will acquire an
+         * individual buffer from a pool of shared buffers.
+         *
+         * Only respected by the Metal backend.
+         */
+        size_t metalUploadBufferSizeBytes = 512 * 1024;
+
         /**
          * Set to `true` to forcibly disable parallel shader compilation in the backend.
          * Currently only honored by the GL and Metal backends.
@@ -332,9 +350,12 @@ public:
         uint32_t 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.
          */
-        uint32_t resourceAllocatorCacheMaxAge = 2;
+        uint32_t resourceAllocatorCacheMaxAge = 1;
 
         /*
          * Disable backend handles use-after-free checks.
@@ -875,6 +896,29 @@ public:
     /** Tells whether an InstanceBuffer object is valid */
     bool isValid(const InstanceBuffer* UTILS_NULLABLE p) const;
 
+    /**
+     * Retrieve the count of each resource tracked by Engine.
+     * This is intended for debugging.
+     * @{
+     */
+    size_t getBufferObjectCount() const noexcept;
+    size_t getViewCount() const noexcept;
+    size_t getSceneCount() const noexcept;
+    size_t getSwapChainCount() const noexcept;
+    size_t getStreamCount() const noexcept;
+    size_t getIndexBufferCount() const noexcept;
+    size_t getSkinningBufferCount() const noexcept;
+    size_t getMorphTargetBufferCount() const noexcept;
+    size_t getInstanceBufferCount() const noexcept;
+    size_t getVertexBufferCount() const noexcept;
+    size_t getIndirectLightCount() const noexcept;
+    size_t getMaterialCount() const noexcept;
+    size_t getTextureCount() const noexcept;
+    size_t getSkyboxeCount() const noexcept;
+    size_t getColorGradingCount() const noexcept;
+    size_t getRenderTargetCount() const noexcept;
+    /**  @} */
+
     /**
      * Kicks the hardware thread (e.g. the OpenGL, Vulkan or Metal thread) and blocks until
      * all commands to this point are executed. Note that does guarantee that the
@@ -1021,6 +1065,21 @@ public:
     void resetBackendState() noexcept;
 #endif
 
+    /**
+     * Get the current time. This is a convenience function that simply returns the
+     * time in nanosecond since epoch of std::chrono::steady_clock.
+     * A possible implementation is:
+     *
+     * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+     *     return std::chrono::steady_clock::now().time_since_epoch().count();
+     * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+     *
+     * @return current time in nanosecond since epoch of std::chrono::steady_clock.
+     * @see Renderer::beginFrame()
+     */
+    static uint64_t getSteadyClockTimeNano() noexcept;
+
+
     DebugRegistry& getDebugRegistry() noexcept;
 
 protected:

filament/include/filament/FilamentAPI.h

@@ -19,6 +19,7 @@
 
 #include <utils/compiler.h>
 #include <utils/PrivateImplementation.h>
+#include <utils/CString.h>
 
 #include <stddef.h>
 
@@ -54,6 +55,23 @@ public:
 template<typename T>
 using BuilderBase = utils::PrivateImplementation<T>;
 
+// This needs to be public because it is used in the following template.
+UTILS_PUBLIC void builderMakeName(utils::CString& outName, const char* name, size_t len) noexcept;
+
+template <typename Builder>
+class UTILS_PUBLIC BuilderNameMixin {
+public:
+    Builder& name(const char* name, size_t len) noexcept {
+        builderMakeName(mName, name, len);
+        return static_cast<Builder&>(*this);
+    }
+
+    utils::CString const& getName() const noexcept { return mName; }
+
+private:
+    utils::CString mName;
+};
+
 } // namespace filament
 
 #endif // TNT_FILAMENT_FILAMENTAPI_H

filament/include/filament/IndexBuffer.h

@@ -59,7 +59,7 @@ public:
         UINT = uint8_t(backend::ElementType::UINT),      //!< 32-bit indices
     };
 
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -83,6 +83,21 @@ public:
          */
         Builder& bufferType(IndexType indexType) noexcept;
 
+        /**
+         * Associate an optional name with this IndexBuffer for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this IndexBuffer
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the IndexBuffer object and returns a pointer to it. After creation, the index
          * buffer is uninitialized. Use IndexBuffer::setBuffer() to initialize the IndexBuffer.

filament/include/filament/IndirectLight.h

@@ -158,6 +158,8 @@ public:
          *
          * @return This Builder, for chaining calls.
          *
+         * @see Material::Builder::sphericalHarmonicsBandCount()
+         *
          * @note
          * Because the coefficients are pre-scaled, `sh[0]` is the environment's
          * average irradiance.

filament/include/filament/InstanceBuffer.h

@@ -38,7 +38,7 @@ class UTILS_PUBLIC InstanceBuffer : public FilamentAPI {
     struct BuilderDetails;
 
 public:
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
 
     public:
@@ -70,6 +70,21 @@ public:
          */
         Builder& localTransforms(math::mat4f const* UTILS_NULLABLE localTransforms) noexcept;
 
+        /**
+         * Associate an optional name with this InstanceBuffer for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this InstanceBuffer
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the InstanceBuffer object and returns a pointer to it.
          */

filament/include/filament/Material.h

@@ -140,6 +140,18 @@ public:
             return constant(name, strlen(name), value);
         }
 
+        /**
+         * Sets the quality of the indirect lights computations. This is only taken into account
+         * if this material is lit and in the surface domain. This setting will affect the
+         * IndirectLight computation if one is specified on the Scene and Spherical Harmonics
+         * are used for the irradiance.
+         *
+         * @param shBandCount Number of spherical harmonic bands. Must be 1, 2 or 3 (default).
+         * @return Reference to this Builder for chaining calls.
+         * @see IndirectLight
+         */
+        Builder& sphericalHarmonicsBandCount(size_t shBandCount) noexcept;
+
         /**
          * Creates the Material object and returns a pointer to it.
          *

filament/include/filament/MorphTargetBuffer.h

@@ -39,7 +39,7 @@ class UTILS_PUBLIC MorphTargetBuffer : public FilamentAPI {
     struct BuilderDetails;
 
 public:
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -63,6 +63,21 @@ public:
          */
         Builder& count(size_t count) noexcept;
 
+        /**
+         * Associate an optional name with this MorphTargetBuffer for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this MorphTargetBuffer
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the MorphTargetBuffer object and returns a pointer to it.
          *

filament/include/filament/RenderableManager.h

@@ -489,43 +489,20 @@ public:
          * For standard morphing, A MorphTargetBuffer must be provided.
          * Standard morphing supports up to \c CONFIG_MAX_MORPH_TARGET_COUNT morph targets.
          *
-         * For legacy morphing, the attached VertexBuffer must provide data in the
-         * appropriate VertexAttribute slots (\c MORPH_POSITION_0 etc). Legacy morphing only
-         * supports up to 4 morph targets and will be deprecated in the future. Legacy morphing must
-         * be enabled on the material definition: either via the legacyMorphing material attribute
-         * or by calling filamat::MaterialBuilder::useLegacyMorphing().
-         *
          * See also RenderableManager::setMorphWeights(), which can be called on a per-frame basis
          * to advance the animation.
          */
         Builder& morphing(MorphTargetBuffer* UTILS_NONNULL morphTargetBuffer) noexcept;
 
-        /**
-         * @deprecated Use morphing(uint8_t level, size_t primitiveIndex, size_t offset, size_t count) instead
-         */
-        Builder& morphing(uint8_t level, size_t primitiveIndex,
-                MorphTargetBuffer* UTILS_NONNULL morphTargetBuffer,
-                size_t offset, size_t count) noexcept;
-
-        /**
-         * @deprecated Use morphing(uint8_t level, size_t primitiveIndex, size_t offset, size_t count) instead
-         */
-        inline Builder& morphing(uint8_t level, size_t primitiveIndex,
-                MorphTargetBuffer* UTILS_NONNULL morphTargetBuffer) noexcept {
-            return morphing(level, primitiveIndex, morphTargetBuffer, 0,
-                    morphTargetBuffer->getVertexCount());
-        }
-
         /**
          * Specifies the the range of the MorphTargetBuffer to use with this primitive.
          *
          * @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 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)
          */
-        Builder& morphing(uint8_t level, size_t primitiveIndex,
-                size_t offset, size_t count) noexcept;
+        RenderableManager::Builder& morphing(uint8_t level,
+                size_t primitiveIndex, size_t offset) noexcept;
 
 
         /**
@@ -786,25 +763,13 @@ public:
     /**
      * Associates a MorphTargetBuffer to the given primitive.
      */
-    void setMorphTargetBufferAt(Instance instance, uint8_t level, size_t primitiveIndex,
-            size_t offset, size_t count);
-
-    /** @deprecated */
-    void setMorphTargetBufferAt(Instance instance, uint8_t level, size_t primitiveIndex,
-            MorphTargetBuffer* UTILS_NONNULL morphTargetBuffer, size_t offset, size_t count);
-
-    /** @deprecated */
-    inline void setMorphTargetBufferAt(Instance instance, uint8_t level, size_t primitiveIndex,
-            MorphTargetBuffer* UTILS_NONNULL morphTargetBuffer) {
-        setMorphTargetBufferAt(instance, level, primitiveIndex, morphTargetBuffer, 0,
-                morphTargetBuffer->getVertexCount());
-    }
+    void setMorphTargetBufferOffsetAt(Instance instance, uint8_t level, size_t primitiveIndex,
+            size_t offset);
 
     /**
-     * Get a MorphTargetBuffer to the given primitive or null if it doesn't exist.
+     * Get a MorphTargetBuffer to the given renderable or null if it doesn't exist.
      */
-    MorphTargetBuffer* UTILS_NULLABLE getMorphTargetBufferAt(Instance instance,
-            uint8_t level, size_t primitiveIndex) const noexcept;
+    MorphTargetBuffer* UTILS_NULLABLE getMorphTargetBuffer(Instance instance) const noexcept;
 
     /**
      * Gets the number of morphing in the given entity.

filament/include/filament/Renderer.h

@@ -22,9 +22,11 @@
 #include <filament/FilamentAPI.h>
 
 #include <utils/compiler.h>
+#include <utils/FixedCapacityVector.h>
 
 #include <math/vec4.h>
 
+#include <stddef.h>
 #include <stdint.h>
 
 namespace filament {
@@ -81,6 +83,37 @@ public:
         UTILS_DEPRECATED uint64_t vsyncOffsetNanos = 0;
     };
 
+    /**
+     * Timing information about a frame
+     * @see getFrameInfoHistory()
+     */
+    struct FrameInfo {
+        using time_point_ns = int64_t;
+        using duration_ns = int64_t;
+        uint32_t frameId;                   //!< monotonically increasing frame identifier
+        duration_ns frameTime;              //!< frame duration on the GPU in nanosecond [ns]
+        duration_ns denoisedFrameTime;      //!< denoised frame duration on the GPU in [ns]
+        time_point_ns beginFrame;           //!< Renderer::beginFrame() time since epoch [ns]
+        time_point_ns endFrame;             //!< Renderer::endFrame() time since epoch [ns]
+        time_point_ns backendBeginFrame;    //!< Backend thread time of frame start since epoch [ns]
+        time_point_ns backendEndFrame;      //!< Backend thread time of frame end since epoch [ns]
+    };
+
+    /**
+     * Retrieve an historic of frame timing information. The maximum frame history size is
+     * given by getMaxFrameHistorySize().
+     * @param historySize requested history size. The returned vector could be smaller.
+     * @return A vector of FrameInfo.
+     */
+    utils::FixedCapacityVector<Renderer::FrameInfo> getFrameInfoHistory(
+            size_t historySize = 1) const noexcept;
+
+    /**
+     * @return the maximum supported frame history size.
+     * @see getFrameInfoHistory()
+     */
+    size_t getMaxFrameHistorySize() const noexcept;
+
     /**
      * Use FrameRateOptions to set the desired frame rate and control how quickly the system
      * reacts to GPU load changes.
@@ -227,6 +260,25 @@ public:
     static constexpr CopyFrameFlag CLEAR = 0x4;
 
 
+    /**
+     * The use of this method is optional. It sets the VSYNC time expressed as the duration in
+     * nanosecond since epoch of std::chrono::steady_clock.
+     * If called, passing 0 to vsyncSteadyClockTimeNano in Renderer::BeginFrame will use this
+     * time instead.
+     * @param steadyClockTimeNano duration in nanosecond since epoch of std::chrono::steady_clock
+     * @see Engine::getSteadyClockTimeNano()
+     * @see Renderer::BeginFrame()
+     */
+    void setVsyncTime(uint64_t steadyClockTimeNano) noexcept;
+
+    /**
+     * Call skipFrame when momentarily skipping frames, for instance if the content of the
+     * scene doesn't change.
+     *
+     * @param vsyncSteadyClockTimeNano
+     */
+    void skipFrame(uint64_t vsyncSteadyClockTimeNano = 0u);
+
     /**
      * Set-up a frame for this Renderer.
      *

filament/include/filament/SkinningBuffer.h

@@ -39,7 +39,7 @@ class UTILS_PUBLIC SkinningBuffer : public FilamentAPI {
     struct BuilderDetails;
 
 public:
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -69,6 +69,21 @@ public:
          */
         Builder& initialize(bool initialize = true) noexcept;
 
+        /**
+         * Associate an optional name with this SkinningBuffer for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this SkinningBuffer
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the SkinningBuffer object and returns a pointer to it.
          *

filament/include/filament/Stream.h

@@ -94,7 +94,7 @@ public:
      *
      * To create a NATIVE stream, call the <pre>stream</pre> method on the builder.
      */
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -136,6 +136,21 @@ public:
          */
         Builder& height(uint32_t height) noexcept;
 
+        /**
+         * Associate an optional name with this Stream for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this Stream
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the Stream object and returns a pointer to it.
          *

filament/include/filament/Texture.h

@@ -112,7 +112,7 @@ public:
 
 
     //! Use Builder to construct a Texture object instance
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -202,6 +202,21 @@ public:
          */
         Builder& swizzle(Swizzle r, Swizzle g, Swizzle b, Swizzle a) noexcept;
 
+        /**
+         * Associate an optional name with this Texture for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this Texture
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+         // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the Texture object and returns a pointer to it.
          *

filament/include/filament/VertexBuffer.h

@@ -61,7 +61,7 @@ public:
     using AttributeType = backend::ElementType;
     using BufferDescriptor = backend::BufferDescriptor;
 
-    class Builder : public BuilderBase<BuilderDetails> {
+    class Builder : public BuilderBase<BuilderDetails>, public BuilderNameMixin<Builder> {
         friend struct BuilderDetails;
     public:
         Builder() noexcept;
@@ -158,6 +158,21 @@ public:
          */
         Builder& advancedSkinning(bool enabled) noexcept;
 
+        /**
+         * Associate an optional name with this VertexBuffer for debugging purposes.
+         *
+         * name will show in error messages and should be kept as short as possible. The name is
+         * truncated to a maximum of 128 characters.
+         *
+         * The name string is copied during this method so clients may free its memory after
+         * the function returns.
+         *
+         * @param name A string to identify this VertexBuffer
+         * @param len Length of name, should be less than or equal to 128
+         * @return This Builder, for chaining calls.
+         */
+        // Builder& name(const char* UTILS_NONNULL name, size_t len) noexcept; // inherited
+
         /**
          * Creates the VertexBuffer object and returns a pointer to it.
          *

filament/include/filament/View.h

@@ -24,6 +24,7 @@
 
 #include <utils/compiler.h>
 #include <utils/Entity.h>
+#include <utils/FixedCapacityVector.h>
 
 #include <math/mathfwd.h>
 
@@ -40,6 +41,7 @@ class CallbackHandler;
 
 class Camera;
 class ColorGrading;
+class Engine;
 class MaterialInstance;
 class RenderTarget;
 class Scene;
@@ -726,7 +728,7 @@ public:
     void setDebugCamera(Camera* UTILS_NULLABLE camera) noexcept;
 
     //! debugging: returns a Camera from the point of view of *the* dominant directional light used for shadowing.
-    Camera const* UTILS_NULLABLE getDirectionalShadowCamera() const noexcept;
+    utils::FixedCapacityVector<Camera const*> getDirectionalShadowCameras() const noexcept;
 
 
     /** Result of a picking query */
@@ -878,6 +880,17 @@ public:
      */
     utils::Entity getFogEntity() const noexcept;
 
+
+    /**
+     * 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.
+     */
+    void clearFrameHistory(Engine& engine) noexcept;
+
     /**
      * List of available ambient occlusion techniques
      * @deprecated use AmbientOcclusionOptions::enabled instead

filament/src/Engine.cpp

@@ -16,6 +16,8 @@
 
 #include "details/Engine.h"
 
+#include "ResourceAllocator.h"
+
 #include "details/BufferObject.h"
 #include "details/Camera.h"
 #include "details/Fence.h"
@@ -39,6 +41,8 @@
 #include <utils/compiler.h>
 #include <utils/Panic.h>
 
+#include <chrono>
+
 #include <stddef.h>
 #include <stdint.h>
 
@@ -264,6 +268,71 @@ bool Engine::isValid(const InstanceBuffer* p) const {
     return downcast(this)->isValid(downcast(p));
 }
 
+size_t Engine::getBufferObjectCount() const noexcept {
+    return downcast(this)->getBufferObjectCount();
+}
+
+size_t Engine::getViewCount() const noexcept {
+    return downcast(this)->getViewCount();
+}
+
+size_t Engine::getSceneCount() const noexcept {
+    return downcast(this)->getSceneCount();
+}
+
+size_t Engine::getSwapChainCount() const noexcept {
+    return downcast(this)->getSwapChainCount();
+}
+
+size_t Engine::getStreamCount() const noexcept {
+    return downcast(this)->getStreamCount();
+}
+
+size_t Engine::getIndexBufferCount() const noexcept {
+    return downcast(this)->getIndexBufferCount();
+}
+
+size_t Engine::getSkinningBufferCount() const noexcept {
+    return downcast(this)->getSkinningBufferCount();
+}
+
+size_t Engine::getMorphTargetBufferCount() const noexcept {
+    return downcast(this)->getMorphTargetBufferCount();
+}
+
+size_t Engine::getInstanceBufferCount() const noexcept {
+    return downcast(this)->getInstanceBufferCount();
+}
+
+size_t Engine::getVertexBufferCount() const noexcept {
+    return downcast(this)->getVertexBufferCount();
+}
+
+size_t Engine::getIndirectLightCount() const noexcept {
+    return downcast(this)->getIndirectLightCount();
+}
+
+size_t Engine::getMaterialCount() const noexcept {
+    return downcast(this)->getMaterialCount();
+}
+
+size_t Engine::getTextureCount() const noexcept {
+    return downcast(this)->getTextureCount();
+}
+
+size_t Engine::getSkyboxeCount() const noexcept {
+    return downcast(this)->getSkyboxeCount();
+}
+
+size_t Engine::getColorGradingCount() const noexcept {
+    return downcast(this)->getColorGradingCount();
+}
+
+size_t Engine::getRenderTargetCount() const noexcept {
+    return downcast(this)->getRenderTargetCount();
+}
+
+
 void Engine::flushAndWait() {
     downcast(this)->flushAndWait();
 }
@@ -361,7 +430,7 @@ const Engine::Config& Engine::getConfig() const noexcept {
     return downcast(this)->getConfig();
 }
 
-bool Engine::isStereoSupported(StereoscopicType stereoscopicType) const noexcept {
+bool Engine::isStereoSupported(StereoscopicType) const noexcept {
     return downcast(this)->isStereoSupported();
 }
 
@@ -369,6 +438,10 @@ size_t Engine::getMaxStereoscopicEyes() noexcept {
     return FEngine::getMaxStereoscopicEyes();
 }
 
+uint64_t Engine::getSteadyClockTimeNano() noexcept {
+    return std::chrono::steady_clock::now().time_since_epoch().count();
+}
+
 #if defined(__EMSCRIPTEN__)
 void Engine::resetBackendState() noexcept {
     downcast(this)->resetBackendState();

filament/src/FilamentBuilder.cpp

@@ -0,0 +1,31 @@
+/*
+ * 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 <filament/FilamentAPI.h>
+
+#include <algorithm>
+
+namespace filament {
+
+void builderMakeName(utils::CString& outName, const char* name, size_t len) noexcept {
+    if (!name) {
+        return;
+    }
+    size_t const length = std::min(len, size_t { 128u });
+    outName = utils::CString(name, length);
+}
+
+} // namespace filament

filament/src/FrameHistory.h

@@ -21,6 +21,9 @@
 #include <fg/FrameGraphTexture.h>
 
 #include <math/mat4.h>
+#include <math/vec2.h>
+
+#include <stdint.h>
 
 namespace filament {
 

filament/src/FrameInfo.cpp

@@ -16,30 +16,34 @@
 
 #include "FrameInfo.h"
 
+#include <filament/Renderer.h>
+
+#include <backend/DriverEnums.h>
+
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Log.h>
+#include <utils/ostream.h>
 #include <utils/Systrace.h>
 
-#include <math/scalar.h>
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <memory>
+#include <ratio>
 
-#include <cmath>
+#include <stdint.h>
+#include <stddef.h>
 
 namespace filament {
 
 using namespace utils;
 using namespace backend;
 
-// this is to avoid a call to memmove
-template<class InputIterator, class OutputIterator>
-static inline
-void move_backward(InputIterator first, InputIterator last, OutputIterator result) {
-    while (first != last) {
-        *--result = *--last;
-    }
-}
-
 FrameInfoManager::FrameInfoManager(DriverApi& driver) noexcept {
     for (auto& query : mQueries) {
-        query = driver.createTimerQuery();
+        query.handle = driver.createTimerQuery();
     }
 }
 
@@ -47,65 +51,155 @@ FrameInfoManager::~FrameInfoManager() noexcept = default;
 
 void FrameInfoManager::terminate(DriverApi& driver) noexcept {
     for (auto& query : mQueries) {
-        driver.destroyTimerQuery(query);
+        driver.destroyTimerQuery(query.handle);
     }
 }
 
-void FrameInfoManager::beginFrame(DriverApi& driver,Config const& config, uint32_t) noexcept {
-    driver.beginTimerQuery(mQueries[mIndex]);
-    uint64_t elapsed = 0;
-    TimerQueryResult const result = driver.getTimerQueryValue(mQueries[mLast], &elapsed);
-    switch (result) {
-        case TimerQueryResult::NOT_READY:
-            // nothing to do
+void FrameInfoManager::beginFrame(DriverApi& driver, Config const& config, uint32_t frameId) noexcept {
+    auto& history = mFrameTimeHistory;
+    // don't exceed the capacity, drop the oldest entry
+    if (UTILS_LIKELY(history.size() == history.capacity())) {
+        history.pop_back();
+    }
+
+    // create a new entry
+    auto& front = history.emplace_front(frameId);
+
+    // store the current time
+    front.beginFrame = std::chrono::steady_clock::now();
+
+    // references are not invalidated by CircularQueue<>, so we can associate a reference to
+    // the slot we created to the timer query used to find the frame time.
+    mQueries[mIndex].pInfo = std::addressof(front);
+    // issue the timer query
+    driver.beginTimerQuery(mQueries[mIndex].handle);
+    // issue the custom backend command to get the backend time
+    driver.queueCommand([&front](){
+        front.backendBeginFrame = std::chrono::steady_clock::now();
+    });
+
+    // now is a good time to check the oldest active query
+    while (mLast != mIndex) {
+        uint64_t elapsed = 0;
+        TimerQueryResult const result = driver.getTimerQueryValue(mQueries[mLast].handle, &elapsed);
+        switch (result) {
+            case TimerQueryResult::NOT_READY:
+                // nothing to do
+                break;
+            case TimerQueryResult::ERROR:
+                mLast = (mLast + 1) % POOL_COUNT;
+                break;
+            case TimerQueryResult::AVAILABLE: {
+                SYSTRACE_CONTEXT();
+                SYSTRACE_VALUE32("FrameInfo::elapsed", uint32_t(elapsed));
+                // conversion to our duration happens here
+                pFront = mQueries[mLast].pInfo;
+                pFront->frameTime = std::chrono::duration<uint64_t, std::nano>(elapsed);
+                mLast = (mLast + 1) % POOL_COUNT;
+                denoiseFrameTime(history, config);
+                break;
+            }
+        }
+        if (result != TimerQueryResult::AVAILABLE) {
             break;
-        case TimerQueryResult::ERROR:
-            mLast = (mLast + 1) % POOL_COUNT;
-            break;
-        case TimerQueryResult::AVAILABLE:
-            mLast = (mLast + 1) % POOL_COUNT;
-            // conversion to our duration happens here
-            mFrameTime = std::chrono::duration<uint64_t, std::nano>(elapsed);
-            break;
+        }
+        // read the pending timer queries until we find one that's not ready
+    }
+
+    // keep this just for debugging
+    if constexpr (false) {
+        using namespace utils;
+        auto h = getFrameInfoHistory(1);
+        if (!h.empty()) {
+            slog.d << frameId << ": "
+                   << h[0].frameId << " (" << frameId - h[0].frameId << ")"
+                   << ", Dm=" << h[0].endFrame - h[0].beginFrame
+                   << ", L =" << h[0].backendBeginFrame - h[0].beginFrame
+                   << ", Db=" << h[0].backendEndFrame - h[0].backendBeginFrame
+                   << ", T =" << h[0].frameTime
+                   << io::endl;
+        }
     }
-    update(config, mFrameTime);
 }
 
 void FrameInfoManager::endFrame(DriverApi& driver) noexcept {
-    driver.endTimerQuery(mQueries[mIndex]);
+    auto& front = mFrameTimeHistory.front();
+    // close the timer query
+    driver.endTimerQuery(mQueries[mIndex].handle);
+    // queue custom backend command to query the current time
+    driver.queueCommand([&front](){
+        // backend frame end-time
+        front.backendEndFrame = std::chrono::steady_clock::now();
+        // signal that the data is available
+        front.ready.store(true, std::memory_order_release);
+    });
+    // and finally acquire the time on the main thread
+    front.endFrame = std::chrono::steady_clock::now();
     mIndex = (mIndex + 1) % POOL_COUNT;
 }
 
-void FrameInfoManager::update(Config const& config,
-        FrameInfoManager::duration lastFrameTime) noexcept {
-    // keep an history of frame times
-    auto& history = mFrameTimeHistory;
+void FrameInfoManager::denoiseFrameTime(FrameHistoryQueue& history, Config const& config) noexcept {
+    assert_invariant(!history.empty());
 
-    // this is like doing { pop_back(); push_front(); }
-    filament::move_backward(history.begin(), history.end() - 1, history.end());
-    history[0].frameTime = lastFrameTime;
-
-    mFrameTimeHistorySize = std::min(++mFrameTimeHistorySize, uint32_t(MAX_FRAMETIME_HISTORY));
-    if (UTILS_UNLIKELY(mFrameTimeHistorySize < 3)) {
-        // not enough history to do anything useful
-        history[0].valid = false;
-        return;
+    // find the first slot that has a valid frame duration
+    size_t first = history.size();
+    for (size_t i = 0, c = history.size(); i < c; ++i) {
+        if (history[i].frameTime != duration(0)) {
+            first = i;
+            break;
+        }
     }
+    assert_invariant(first != history.size());
 
-    // apply a median filter to get a good representation of the frame time of the last
-    // N frames.
-    std::array<duration, MAX_FRAMETIME_HISTORY> median; // NOLINT -- it's initialized below
-    size_t const size = std::min(mFrameTimeHistorySize,
-            std::min(config.historySize, (uint32_t)median.size()));
-    for (size_t i = 0; i < size; ++i) {
-        median[i] = history[i].frameTime;
-    }
-    std::sort(median.begin(), median.begin() + size);
-    duration const denoisedFrameTime = median[size / 2];
+    // we need at least 3 valid frame time to calculate the median
+    if (history.size() >= first + 3) {
+        // apply a median filter to get a good representation of the frame time of the last
+        // N frames.
+        std::array<duration, MAX_FRAMETIME_HISTORY> median; // NOLINT -- it's initialized below
+        size_t const size = std::min({
+            history.size() - first,
+            median.size(),
+            size_t(config.historySize) });
 
-    history[0].denoisedFrameTime = denoisedFrameTime;
-    history[0].valid = true;
+        for (size_t i = 0; i < size; ++i) {
+            median[i] = history[first + i].frameTime;
+        }
+        std::sort(median.begin(), median.begin() + size);
+        duration const denoisedFrameTime = median[size / 2];
+
+        history[first].denoisedFrameTime = denoisedFrameTime;
+        history[first].valid = true;
+     }
 }
 
+utils::FixedCapacityVector<Renderer::FrameInfo> FrameInfoManager::getFrameInfoHistory(
+        size_t historySize) const noexcept {
+    auto result = utils::FixedCapacityVector<Renderer::FrameInfo>::with_capacity(MAX_FRAMETIME_HISTORY);
+    auto const& history = mFrameTimeHistory;
+    size_t i = 0;
+    size_t const c = history.size();
+    for (; i < c; ++i) {
+        auto const& entry = history[i];
+        if (entry.ready.load(std::memory_order_acquire) && entry.valid) {
+            // once we found an entry ready,
+            // we know by construction that all following ones are too
+            break;
+        }
+    }
+    for (; i < c && historySize; ++i, --historySize) {
+        auto const& entry = history[i];
+        using namespace std::chrono;
+        result.push_back({
+                entry.frameId,
+                duration_cast<nanoseconds>(entry.frameTime).count(),
+                duration_cast<nanoseconds>(entry.denoisedFrameTime).count(),
+                duration_cast<nanoseconds>(entry.beginFrame.time_since_epoch()).count(),
+                duration_cast<nanoseconds>(entry.endFrame.time_since_epoch()).count(),
+                duration_cast<nanoseconds>(entry.backendBeginFrame.time_since_epoch()).count(),
+                duration_cast<nanoseconds>(entry.backendEndFrame.time_since_epoch()).count()
+        });
+    }
+    return result;
+}
 
 } // namespace filament

filament/src/FrameInfo.h

@@ -17,30 +17,136 @@
 #ifndef TNT_FILAMENT_FRAMEINFO_H
 #define TNT_FILAMENT_FRAMEINFO_H
 
-#include "backend/Handle.h"
+#include <filament/Renderer.h>
+
+#include <backend/Handle.h>
+
 #include <private/backend/DriverApi.h>
 
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
+
 #include <array>
+#include <atomic>
 #include <chrono>
+#include <ratio>
+#include <type_traits>
 
 #include <stdint.h>
+#include <stddef.h>
 
 namespace filament {
 class FEngine;
 
+namespace details {
 struct FrameInfo {
     using duration = std::chrono::duration<float, std::milli>;
     duration frameTime{};            // frame period
     duration denoisedFrameTime{};    // frame period (median filter)
-    bool valid = false;
+    bool valid = false;              // true if the data of the structure is valid
+};
+} // namespace details
+
+struct FrameInfoImpl : public details::FrameInfo {
+    using clock = std::chrono::steady_clock;
+    using time_point = clock::time_point;
+    uint32_t const frameId;
+    time_point beginFrame;           // main thread beginFrame time
+    time_point endFrame;             // main thread endFrame time
+    time_point backendBeginFrame;    // backend thread beginFrame time (makeCurrent time)
+    time_point backendEndFrame;      // backend thread endFrame time (present time)
+    std::atomic_bool ready{};        // true once backend thread has populated its data
+    explicit FrameInfoImpl(uint32_t frameId) noexcept
+        : frameId(frameId) {
+    }
+};
+
+template<typename T, size_t CAPACITY>
+class CircularQueue {
+public:
+    using value_type = T;
+    using reference = value_type&;
+    using const_reference = value_type const&;
+
+    size_t capacity() const {
+        return CAPACITY;
+    }
+
+    size_t size() const {
+        return mSize;
+    }
+
+    bool empty() const noexcept {
+        return !size();
+    }
+
+    void pop_back() noexcept {
+        assert_invariant(!empty());
+        --mSize;
+        std::destroy_at(&mStorage[(mFront - mSize) % CAPACITY]);
+    }
+
+    void push_front(T const& v) noexcept {
+        assert_invariant(size() < CAPACITY);
+        mFront = advance(mFront);
+        new(&mStorage[mFront]) T(v);
+        ++mSize;
+    }
+
+    void push_front(T&& v) noexcept {
+        assert_invariant(size() < CAPACITY);
+        mFront = advance(mFront);
+        new(&mStorage[mFront]) T(std::move(v));
+        ++mSize;
+    }
+
+    template<typename ...Args>
+    T& emplace_front(Args&&... args) noexcept {
+        assert_invariant(size() < CAPACITY);
+        mFront = advance(mFront);
+        new(&mStorage[mFront]) T(std::forward<Args>(args)...);
+        ++mSize;
+        return front();
+    }
+
+    T& operator[](size_t pos) noexcept {
+        assert_invariant(pos < size());
+        size_t const index = (mFront + CAPACITY - pos) % CAPACITY;
+        return *std::launder(reinterpret_cast<T*>(&mStorage[index]));
+    }
+
+    T const& operator[](size_t pos) const noexcept {
+        return const_cast<CircularQueue&>(*this)[pos];
+    }
+
+    T const& front() const noexcept {
+        assert_invariant(!empty());
+        return operator[](0);
+    }
+
+    T& front() noexcept {
+        assert_invariant(!empty());
+        return operator[](0);
+    }
+
+private:
+    using Storage = std::aligned_storage_t<sizeof(T), alignof(T)>;
+    Storage mStorage[CAPACITY];
+    uint32_t mFront = 0;    // always index 0
+    uint32_t mSize = 0;
+    [[nodiscard]] inline uint32_t advance(uint32_t v) noexcept {
+        return (v + 1) % CAPACITY;
+    }
 };
 
 class FrameInfoManager {
     static constexpr size_t POOL_COUNT = 4;
-    static constexpr size_t MAX_FRAMETIME_HISTORY = 31u;
+    static constexpr size_t MAX_FRAMETIME_HISTORY = 16u;
 
 public:
-    using duration = FrameInfo::duration;
+    using duration = FrameInfoImpl::duration;
+    using clock = FrameInfoImpl::clock;
 
     struct Config {
         uint32_t historySize;
@@ -57,23 +163,25 @@ public:
     // call this immediately before "swap buffers"
     void endFrame(backend::DriverApi& driver) noexcept;
 
-    FrameInfo const& getLastFrameInfo() const noexcept {
-        return mFrameTimeHistory[0];
+    details::FrameInfo getLastFrameInfo() const noexcept {
+        // if pFront is not set yet, return FrameInfo(). But the `valid` field will be false in this case.
+        return pFront ? *pFront : details::FrameInfo{};
     }
 
-    duration getLastFrameTime() const noexcept {
-        return getLastFrameInfo().frameTime;
-    }
+    utils::FixedCapacityVector<Renderer::FrameInfo> getFrameInfoHistory(size_t historySize) const noexcept;
 
 private:
-    void update(Config const& config, duration lastFrameTime) noexcept;
-    backend::Handle<backend::HwTimerQuery> mQueries[POOL_COUNT];
-    duration mFrameTime{};
-    uint32_t mIndex = 0;
-    uint32_t mLast = 0;
-
-    std::array<FrameInfo, MAX_FRAMETIME_HISTORY> mFrameTimeHistory;
-    uint32_t mFrameTimeHistorySize = 0;
+    using FrameHistoryQueue = CircularQueue<FrameInfoImpl, MAX_FRAMETIME_HISTORY>;
+    static void denoiseFrameTime(FrameHistoryQueue& history, Config const& config) noexcept;
+    struct Query {
+        backend::Handle<backend::HwTimerQuery> handle{};
+        FrameInfoImpl* pInfo = nullptr;
+    };
+    std::array<Query, POOL_COUNT> mQueries;
+    uint32_t mIndex = 0;                // index of current query
+    uint32_t mLast = 0;                 // index of oldest query still active
+    FrameInfoImpl* pFront = nullptr;    // the most recent slot with a valid frame time
+    FrameHistoryQueue mFrameTimeHistory;
 };
 
 

filament/src/FrameSkipper.cpp

@@ -16,16 +16,22 @@
 
 #include "FrameSkipper.h"
 
-#include <utils/Log.h>
+#include <backend/DriverEnums.h>
+
+#include <utils/compiler.h>
 #include <utils/debug.h>
 
+#include <algorithm>
+
+#include <stddef.h>
+
 namespace filament {
 
 using namespace utils;
 using namespace backend;
 
 FrameSkipper::FrameSkipper(size_t latency) noexcept
-        : mLast(latency - 1) {
+        : mLast(std::max(latency, MAX_FRAME_LATENCY) - 1) {
     assert_invariant(latency <= MAX_FRAME_LATENCY);
 }
 
@@ -41,31 +47,32 @@ void FrameSkipper::terminate(DriverApi& driver) noexcept {
 
 bool FrameSkipper::beginFrame(DriverApi& driver) noexcept {
     auto& fences = mDelayedFences;
-    auto fence = fences.front();
-    if (fence) {
-        auto status = driver.getFenceStatus(fence);
-        if (status == FenceStatus::TIMEOUT_EXPIRED) {
-            // Sync not ready, skip frame
+    if (fences.front()) {
+        // Do we have a latency old fence?
+        auto status = driver.getFenceStatus(fences.front());
+        if (UTILS_UNLIKELY(status == FenceStatus::TIMEOUT_EXPIRED)) {
+            // The fence hasn't signaled yet, skip this frame
             return false;
         }
         assert_invariant(status == FenceStatus::CONDITION_SATISFIED);
-        driver.destroyFence(fence);
     }
-    // shift all fences down by 1
-    std::move(fences.begin() + 1, fences.end(), fences.begin());
-    fences.back() = {};
     return true;
 }
 
 void FrameSkipper::endFrame(DriverApi& driver) noexcept {
-    // If the user produced a new frame despite the fact that the previous one wasn't finished
-    // (i.e. FrameSkipper::beginFrame() returned false), we need to make sure to replace
-    // a fence that might be here already)
-    auto& fence = mDelayedFences[mLast];
-    if (fence) {
-        driver.destroyFence(fence);
+    auto& fences = mDelayedFences;
+    size_t const last = mLast;
+
+    // pop the oldest fence and advance the other ones
+    if (fences.front()) {
+        driver.destroyFence(fences.front());
     }
-    fence = driver.createFence();
+    std::move(fences.begin() + 1, fences.end(), fences.begin());
+
+    // add a new fence to the end
+    assert_invariant(!fences[last]);
+
+    fences[last] = driver.createFence();
 }
 
 } // namespace filament

filament/src/FrameSkipper.h

@@ -22,6 +22,9 @@
 
 #include <array>
 
+#include <stddef.h>
+#include <stdint.h>
+
 namespace filament {
 
 /*
@@ -60,7 +63,7 @@ public:
 private:
     using Container = std::array<backend::Handle<backend::HwFence>, MAX_FRAME_LATENCY>;
     mutable Container mDelayedFences{};
-    size_t mLast;
+    uint8_t const mLast;
 };
 
 } // namespace filament

filament/src/Froxelizer.cpp

@@ -26,6 +26,7 @@
 #include <filament/Viewport.h>
 
 #include <utils/BinaryTreeArray.h>
+#include <utils/Log.h>
 #include <utils/Systrace.h>
 #include <utils/debug.h>
 

filament/src/MaterialParser.cpp

@@ -354,6 +354,12 @@ bool MaterialParser::getSpecularAntiAliasingThreshold(float* value) const noexce
     return mImpl.getFromSimpleChunk(ChunkType::MaterialSpecularAntiAliasingThreshold, value);
 }
 
+bool MaterialParser::getStereoscopicType(backend::StereoscopicType* value) const noexcept {
+    static_assert(sizeof(StereoscopicType) == sizeof(uint8_t),
+            "StereoscopicType expected size is wrong");
+    return mImpl.getFromSimpleChunk(ChunkType::MaterialStereoscopicType, reinterpret_cast<uint8_t*>(value));
+}
+
 bool MaterialParser::getRequiredAttributes(AttributeBitset* value) const noexcept {
     uint32_t rawAttributes = 0;
     if (!mImpl.getFromSimpleChunk(ChunkType::MaterialRequiredAttributes, &rawAttributes)) {

filament/src/MaterialParser.h

@@ -122,6 +122,7 @@ public:
     bool hasSpecularAntiAliasing(bool* value) const noexcept;
     bool getSpecularAntiAliasingVariance(float* value) const noexcept;
     bool getSpecularAntiAliasingThreshold(float* value) const noexcept;
+    bool getStereoscopicType(backend::StereoscopicType*) const noexcept;
 
     bool getShader(filaflat::ShaderContent& shader, backend::ShaderModel shaderModel,
             Variant variant, backend::ShaderStage stage) noexcept;

filament/src/PerViewUniforms.cpp

@@ -459,10 +459,12 @@ void PerViewUniforms::bind(backend::DriverApi& driver) noexcept {
 
 void PerViewUniforms::unbindSamplers() noexcept {
     auto& samplerGroup = mSamplers;
+    samplerGroup.clearSampler(PerViewSib::SHADOW_MAP);
+    samplerGroup.clearSampler(PerViewSib::IBL_SPECULAR);
     samplerGroup.clearSampler(PerViewSib::SSAO);
     samplerGroup.clearSampler(PerViewSib::SSR);
     samplerGroup.clearSampler(PerViewSib::STRUCTURE);
-    samplerGroup.clearSampler(PerViewSib::SHADOW_MAP);
+    samplerGroup.clearSampler(PerViewSib::FOG);
 }
 
 } // namespace filament

filament/src/PostProcessManager.cpp

@@ -494,7 +494,6 @@ PostProcessManager::StructurePassOutput PostProcessManager::structure(FrameGraph
 
     struct StructureMipmapData {
         FrameGraphId<FrameGraphTexture> depth;
-        uint32_t rt[8];
     };
 
     fg.addPass<StructureMipmapData>("StructureMipmap",
@@ -505,7 +504,7 @@ PostProcessManager::StructurePassOutput PostProcessManager::structure(FrameGraph
                             .level = uint8_t(i)
                     });
                     out = builder.write(out, FrameGraphTexture::Usage::DEPTH_ATTACHMENT);
-                    data.rt[i - 1] = builder.declareRenderPass("Structure mip target", {
+                    builder.declareRenderPass("Structure mip target", {
                             .attachments = { .depth = out }
                     });
                 }
@@ -1965,8 +1964,8 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
         // - visible bloom size changes with dynamic resolution in non-homogenous mode
         // This allows us to use the 9 sample downsampling filter (instead of 13)
         // for at least 4 levels.
-        uint32_t width  = std::max(1u, uint32_t(std::floor(bloomWidth)));
-        uint32_t height = std::max(1u, uint32_t(std::floor(bloomHeight)));
+        uint32_t width  = std::max(16u, uint32_t(std::floor(bloomWidth)));
+        uint32_t height = std::max(16u, uint32_t(std::floor(bloomHeight)));
         width  &= ~((1 << 4) - 1);  // at least 4 levels
         height &= ~((1 << 4) - 1);
         bloomWidth  = float(width);
@@ -1978,6 +1977,8 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
     // we don't need to do the fireflies reduction if we have TAA (it already does it)
     bool fireflies = threshold && !taaOptions.enabled;
 
+    assert_invariant(bloomWidth && bloomHeight);
+
     while (2 * bloomWidth < float(desc.width) || 2 * bloomHeight < float(desc.height)) {
         if (inoutBloomOptions.quality == QualityLevel::LOW ||
             inoutBloomOptions.quality == QualityLevel::MEDIUM) {
@@ -2944,6 +2945,10 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscale(FrameGraph& fg, bool
                                     1.0f / outputDesc.height});
                     }
 
+                    if (blitterNames[index] == "blitLow") {
+                        mi->setParameter("levelOfDetail", 0.0f);
+                    }
+
                     mi->setParameter("viewport", float4{
                             (float)vp.left   / inputDesc.width,
                             (float)vp.bottom / inputDesc.height,
@@ -3002,9 +3007,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::blit(FrameGraph& fg, bool tr
         SamplerMagFilter filterMag,
         SamplerMinFilter filterMin) noexcept {
 
-    // TODO: add support for sub-resources
-    assert_invariant(fg.getSubResourceDescriptor(input).layer == 0);
-    assert_invariant(fg.getSubResourceDescriptor(input).level == 0);
+    uint32_t const layer = fg.getSubResourceDescriptor(input).layer;
+    float const levelOfDetail = fg.getSubResourceDescriptor(input).level;
 
     struct QuadBlitData {
         FrameGraphId<FrameGraphTexture> input;
@@ -3030,7 +3034,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::blit(FrameGraph& fg, bool tr
                 // --------------------------------------------------------------------------------
                 // set uniforms
 
-                PostProcessMaterial const& material = getPostProcessMaterial("blitLow");
+                PostProcessMaterial const& material =
+                        getPostProcessMaterial(layer ? "blitArray" : "blitLow");
                 auto* mi = material.getMaterialInstance(mEngine);
                 mi->setParameter("color", color, {
                         .filterMag = filterMag,
@@ -3042,6 +3047,10 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::blit(FrameGraph& fg, bool tr
                         float(vp.width)  / inputDesc.width,
                         float(vp.height) / inputDesc.height
                 });
+                mi->setParameter("levelOfDetail", levelOfDetail);
+                if (layer) {
+                    mi->setParameter("layerIndex", layer);
+                }
                 mi->commit(driver);
                 mi->use(driver);
 
@@ -3184,6 +3193,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::resolve(FrameGraph& fg,
                 assert_invariant(dst);
                 assert_invariant(srcDesc.format == dstDesc.format);
                 assert_invariant(srcDesc.width == dstDesc.width && srcDesc.height == dstDesc.height);
+                assert_invariant(srcDesc.samples > 1 && dstDesc.samples <= 1);
                 driver.resolve(
                         dst, dstSubDesc.level, dstSubDesc.layer,
                         src, srcSubDesc.level, srcSubDesc.layer);

filament/src/RenderPass.cpp

@@ -38,6 +38,7 @@
 #include <backend/PipelineState.h>
 
 #include "private/backend/CircularBuffer.h"
+#include "private/backend/CommandStream.h"
 
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -117,10 +118,11 @@ RenderPass::RenderPass(FEngine& engine, RenderPassBuilder const& builder) noexce
     uint32_t const customCommandCount =
             builder.mCustomCommands.has_value() ? builder.mCustomCommands->size() : 0;
 
-    Command* const curr = builder.mArena.alloc<Command>(commandCount + customCommandCount);
-    assert_invariant(curr);
+    Command* const commandBegin = builder.mArena.alloc<Command>(commandCount + customCommandCount);
+    Command* commandEnd = commandBegin + (commandCount + customCommandCount);
+    assert_invariant(commandBegin);
 
-    if (UTILS_UNLIKELY(builder.mArena.getAllocator().isHeapAllocation(curr))) {
+    if (UTILS_UNLIKELY(builder.mArena.getAllocator().isHeapAllocation(commandBegin))) {
         static bool sLogOnce = true;
         if (UTILS_UNLIKELY(sLogOnce)) {
             sLogOnce = false;
@@ -129,10 +131,7 @@ RenderPass::RenderPass(FEngine& engine, RenderPassBuilder const& builder) noexce
         }
     }
 
-    mCommandBegin = curr;
-    mCommandEnd = curr + commandCount + customCommandCount;
-
-    appendCommands(engine, { curr, commandCount },
+    appendCommands(engine, { commandBegin, commandCount },
             builder.mUboHandle,
             builder.mVisibleRenderables,
             builder.mCommandTypeFlags,
@@ -143,33 +142,36 @@ RenderPass::RenderPass(FEngine& engine, RenderPassBuilder const& builder) noexce
             builder.mCameraForwardVector);
 
     if (builder.mCustomCommands.has_value()) {
-        Command* p = curr + commandCount;
+        Command* p = commandBegin + commandCount;
         for (auto [channel, passId, command, order, fn]: builder.mCustomCommands.value()) {
             appendCustomCommand(p++, channel, passId, command, order, fn);
         }
     }
 
     // sort commands once we're done adding commands
-    sortCommands(builder.mArena);
+    commandEnd = resize(builder.mArena,
+            RenderPass::sortCommands(commandBegin, commandEnd));
 
     if (engine.isAutomaticInstancingEnabled()) {
-        uint32_t stereoscopicEyeCount = 1;
+        int32_t stereoscopicEyeCount = 1;
         if (builder.mFlags & IS_INSTANCED_STEREOSCOPIC) {
             stereoscopicEyeCount *= engine.getConfig().stereoscopicEyeCount;
         }
-        instanceify(engine, builder.mArena, stereoscopicEyeCount);
+        commandEnd = resize(builder.mArena,
+                instanceify(engine, commandBegin, commandEnd, stereoscopicEyeCount));
     }
+
+    // these are `const` from this point on...
+    mCommandBegin = commandBegin;
+    mCommandEnd = commandEnd;
 }
 
 // this destructor is actually heavy because it inlines ~vector<>
-RenderPass::~RenderPass() noexcept {
-}
+RenderPass::~RenderPass() noexcept = default;
 
-void RenderPass::resize(Arena& arena, size_t count) noexcept {
-    if (mCommandBegin) {
-        mCommandEnd = mCommandBegin + count;
-        arena.rewind(mCommandEnd);
-    }
+RenderPass::Command* RenderPass::resize(Arena& arena, Command* const last) noexcept {
+    arena.rewind(last);
+    return last;
 }
 
 void RenderPass::appendCommands(FEngine& engine,
@@ -221,7 +223,7 @@ void RenderPass::appendCommands(FEngine& engine,
         work(vr.first, vr.size());
     } else {
         auto* jobCommandsParallel = jobs::parallel_for(js, nullptr, vr.first, (uint32_t)vr.size(),
-                std::cref(work), jobs::CountSplitter<JOBS_PARALLEL_FOR_COMMANDS_COUNT, 5>());
+                std::cref(work), jobs::CountSplitter<JOBS_PARALLEL_FOR_COMMANDS_COUNT>());
         js.runAndWait(jobCommandsParallel);
     }
 
@@ -260,18 +262,19 @@ void RenderPass::appendCustomCommand(Command* commands,
     commands->key = cmd;
 }
 
-void RenderPass::sortCommands(Arena& arena) noexcept {
-    SYSTRACE_NAME("sort and trim commands");
+RenderPass::Command* RenderPass::sortCommands(
+        Command* const begin, Command* const end) noexcept {
+    SYSTRACE_NAME("sort commands");
 
-    std::sort(mCommandBegin, mCommandEnd);
+    std::sort(begin, end);
 
     // find the last command
-    Command const* const last = std::partition_point(mCommandBegin, mCommandEnd,
+    Command* const last = std::partition_point(begin, end,
             [](Command const& c) {
                 return c.key != uint64_t(Pass::SENTINEL);
             });
 
-    resize(arena, uint32_t(last - mCommandBegin));
+    return last;
 }
 
 void RenderPass::execute(RenderPass const& pass,
@@ -284,7 +287,9 @@ void RenderPass::execute(RenderPass const& pass,
     driver.endRenderPass();
 }
 
-void RenderPass::instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) noexcept {
+RenderPass::Command* RenderPass::instanceify(FEngine& engine,
+        Command* curr, Command* const last,
+        int32_t eyeCount) const noexcept {
     SYSTRACE_NAME("instanceify");
 
     // instanceify works by scanning the **sorted** command stream, looking for repeat draw
@@ -297,14 +302,12 @@ void RenderPass::instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) no
 
     UTILS_UNUSED uint32_t drawCallsSavedCount = 0;
 
-    Command* curr = mCommandBegin;
-    Command* const last = mCommandEnd;
-
     Command* firstSentinel = nullptr;
     PerRenderableData const* uboData = nullptr;
     PerRenderableData* stagingBuffer = nullptr;
     uint32_t stagingBufferSize = 0;
     uint32_t instancedPrimitiveOffset = 0;
+    size_t const count = last - curr;
 
     // TODO: for the case of instancing we could actually use 128 instead of 64 instances
     constexpr size_t maxInstanceCount = CONFIG_MAX_INSTANCES;
@@ -343,7 +346,6 @@ void RenderPass::instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) no
             if (UTILS_UNLIKELY(!stagingBuffer)) {
 
                 // create a temporary UBO for instancing
-                size_t const count = mCommandEnd - mCommandBegin;
                 mInstancedUboHandle = BufferObjectSharedHandle{
                         engine.getDriverApi().createBufferObject(
                                 count * sizeof(PerRenderableData) + sizeof(PerRenderableUib),
@@ -385,7 +387,7 @@ void RenderPass::instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) no
 
     if (UTILS_UNLIKELY(firstSentinel)) {
         //slog.d << "auto-instancing, saving " << drawCallsSavedCount << " draw calls, out of "
-        //       << mCommandEnd - mCommandBegin << io::endl;
+        //       << count << io::endl;
 
         // we have instanced primitives
         DriverApi& driver = engine.getDriverApi();
@@ -401,14 +403,15 @@ void RenderPass::instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) no
         stagingBuffer = nullptr;
 
         // remove all the canceled commands
-        auto lastCommand = std::remove_if(firstSentinel, mCommandEnd, [](auto const& command) {
+        auto lastCommand = std::remove_if(firstSentinel, last, [](auto const& command) {
             return command.key == uint64_t(Pass::SENTINEL);
         });
 
-        resize(arena, uint32_t(lastCommand - mCommandBegin));
+        return lastCommand;
     }
 
     assert_invariant(stagingBuffer == nullptr);
+    return last;
 }
 
 
@@ -416,7 +419,8 @@ void RenderPass::instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) no
 UTILS_ALWAYS_INLINE // This function exists only to make the code more readable. we want it inlined.
 inline              // and we don't need it in the compilation unit
 void RenderPass::setupColorCommand(Command& cmdDraw, Variant variant,
-        FMaterialInstance const* const UTILS_RESTRICT mi, bool inverseFrontFaces) noexcept {
+        FMaterialInstance const* const UTILS_RESTRICT mi,
+        bool inverseFrontFaces, bool hasDepthClamp) noexcept {
 
     FMaterial const * const UTILS_RESTRICT ma = mi->getMaterial();
     variant = Variant::filterVariant(variant, ma->isVariantLit());
@@ -457,6 +461,7 @@ void RenderPass::setupColorCommand(Command& cmdDraw, Variant variant,
     cmdDraw.info.rasterState.colorWrite = mi->isColorWriteEnabled();
     cmdDraw.info.rasterState.depthWrite = mi->isDepthWriteEnabled();
     cmdDraw.info.rasterState.depthFunc = mi->getDepthFunc();
+    cmdDraw.info.rasterState.depthClamp = hasDepthClamp;
     cmdDraw.info.materialVariant = variant;
     // we keep "RasterState::colorWrite" to the value set by material (could be disabled)
 }
@@ -555,6 +560,9 @@ RenderPass::Command* RenderPass::generateCommandsImpl(RenderPass::CommandTypeFla
     bool const hasInstancedStereo =
             renderFlags & IS_INSTANCED_STEREOSCOPIC;
 
+    bool const hasDepthClamp =
+            renderFlags & HAS_DEPTH_CLAMP;
+
     float const cameraPositionDotCameraForward = dot(cameraPosition, cameraForward);
 
     auto const* const UTILS_RESTRICT soaWorldAABBCenter = soa.data<FScene::WORLD_AABB_CENTER>();
@@ -574,6 +582,7 @@ RenderPass::Command* RenderPass::generateCommandsImpl(RenderPass::CommandTypeFla
         cmd.info.rasterState.depthWrite = true;
         cmd.info.rasterState.depthFunc = RasterState::DepthFunc::GE;
         cmd.info.rasterState.alphaToCoverage = false;
+        cmd.info.rasterState.depthClamp = hasDepthClamp;
     }
 
     for (uint32_t i = range.first; i < range.last; ++i) {
@@ -607,7 +616,7 @@ RenderPass::Command* RenderPass::generateCommandsImpl(RenderPass::CommandTypeFla
         //   Here, objects close to the camera (but behind) will be drawn first.
         // An alternative that keeps the mathematical ordering is given here:
         //   distanceBits ^= ((int32_t(distanceBits) >> 31) | 0x80000000u);
-        float const distance = -dot(soaWorldAABBCenter[i], cameraForward) - cameraPositionDotCameraForward;
+        float const distance = -(dot(soaWorldAABBCenter[i], cameraForward) - cameraPositionDotCameraForward);
         uint32_t const distanceBits = reinterpret_cast<uint32_t const&>(distance);
 
         // calculate the per-primitive face winding order inversion
@@ -671,7 +680,6 @@ RenderPass::Command* RenderPass::generateCommandsImpl(RenderPass::CommandTypeFla
          */
         for (size_t pi = 0, c = primitives.size(); pi < c; ++pi) {
             auto const& primitive = primitives[pi];
-            auto const& morphTargets = morphing.targets[pi];
             FMaterialInstance const* const mi = primitive.getMaterialInstance();
             FMaterial const* const ma = mi->getMaterial();
 
@@ -684,11 +692,13 @@ RenderPass::Command* RenderPass::generateCommandsImpl(RenderPass::CommandTypeFla
             cmd.info.indexOffset = primitive.getIndexOffset();
             cmd.info.indexCount = primitive.getIndexCount();
             cmd.info.type = primitive.getPrimitiveType();
-            cmd.info.morphTargetBuffer = morphTargets.buffer->getHwHandle();
-            cmd.info.morphingOffset = morphTargets.offset;
+            cmd.info.morphTargetBuffer = morphing.morphTargetBuffer ?
+                    morphing.morphTargetBuffer->getHwHandle() : SamplerGroupHandle{};
+            cmd.info.morphingOffset = primitive.getMorphingBufferOffset();
 
             if constexpr (isColorPass) {
-                RenderPass::setupColorCommand(cmd, renderableVariant, mi, inverseFrontFaces);
+                RenderPass::setupColorCommand(cmd, renderableVariant, mi,
+                        inverseFrontFaces, hasDepthClamp);
                 const bool blendPass = Pass(cmd.key & PASS_MASK) == Pass::BLENDED;
                 if (blendPass) {
                     // TODO: at least for transparent objects, AABB should be per primitive
@@ -885,14 +895,31 @@ void RenderPass::Executor::execute(FEngine& engine,
 
         // Maximum space occupied in the CircularBuffer by a single `Command`. This must be
         // reevaluated when the inner loop below adds DriverApi commands or when we change the
-        // CommandStream protocol. Currently, the maximum is 240 bytes, and we use 256 to be on
-        // the safer side.
-        size_t const maxCommandSizeInBytes = 256;
+        // CommandStream protocol. Currently, the maximum is 320 bytes.
+        // The batch size is calculated by adding the size of all commands that can possibly be
+        // emitted per draw call:
+        constexpr size_t const maxCommandSizeInBytes =
+                sizeof(CustomCommand) +
+                sizeof(COMMAND_TYPE(scissor)) +
+                sizeof(COMMAND_TYPE(bindUniformBuffer)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindBufferRange)) +
+                sizeof(COMMAND_TYPE(bindBufferRange)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindUniformBuffer)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindPipeline)) +
+                sizeof(COMMAND_TYPE(setPushConstant)) +
+                sizeof(COMMAND_TYPE(bindRenderPrimitive)) +
+                sizeof(COMMAND_TYPE(draw2));
+
 
         // Number of Commands that can be issued and guaranteed to fit in the current
         // CircularBuffer allocation. In practice, we'll have tons of headroom especially if
         // skinning and morphing aren't used. With a 2 MiB buffer (the default) a batch is
-        // 8192 commands (i.e. draw calls).
+        // 6553 commands (i.e. draw calls).
         size_t const batchCommandCount = capacity / maxCommandSizeInBytes;
         while(first != last) {
             Command const* const batchLast = std::min(first + batchCommandCount, last);

filament/src/RenderPass.h

@@ -284,6 +284,7 @@ public:
     static constexpr RenderFlags HAS_SHADOWING             = 0x01;
     static constexpr RenderFlags HAS_INVERSE_FRONT_FACES   = 0x02;
     static constexpr RenderFlags IS_INSTANCED_STEREOSCOPIC = 0x04;
+    static constexpr RenderFlags HAS_DEPTH_CLAMP           = 0x08;
 
     // Arena used for commands
     using Arena = utils::Arena<
@@ -408,17 +409,19 @@ private:
             uint8_t channel, Pass pass, CustomCommand custom, uint32_t order,
             Executor::CustomCommandFn command);
 
-    void resize(Arena& arena, size_t count) noexcept;
+    static Command* resize(Arena& arena, Command* const last) noexcept;
 
     // sorts commands then trims sentinels
-    void sortCommands(Arena& arena) noexcept;
+    static Command* sortCommands(
+            Command* begin, Command* end) noexcept;
 
     // instanceify commands then trims sentinels
-    void instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) noexcept;
+    RenderPass::Command* instanceify(FEngine& engine,
+            Command* begin, Command* end,
+            int32_t eyeCount) const noexcept;
 
     // We choose the command count per job to minimize JobSystem overhead.
-    // On a Pixel 4, 2048 commands is about half a millisecond of processing.
-    static constexpr size_t JOBS_PARALLEL_FOR_COMMANDS_COUNT = 2048;
+    static constexpr size_t JOBS_PARALLEL_FOR_COMMANDS_COUNT = 128;
     static constexpr size_t JOBS_PARALLEL_FOR_COMMANDS_SIZE  =
             sizeof(Command) * JOBS_PARALLEL_FOR_COMMANDS_COUNT;
 
@@ -442,17 +445,17 @@ private:
             uint8_t instancedStereoEyeCount) noexcept;
 
     static void setupColorCommand(Command& cmdDraw, Variant variant,
-            FMaterialInstance const* mi, bool inverseFrontFaces) noexcept;
+            FMaterialInstance const* mi, bool inverseFrontFaces, bool hasDepthClamp) noexcept;
 
     static void updateSummedPrimitiveCounts(
             FScene::RenderableSoa& renderableData, utils::Range<uint32_t> vr) noexcept;
 
     FScene::RenderableSoa const& mRenderableSoa;
     backend::Viewport const mScissorViewport;
-    Command* mCommandBegin = nullptr;   // Pointer to the first command
-    Command* mCommandEnd = nullptr;     // Pointer to one past the last command
+    Command const* /* const */ mCommandBegin = nullptr;   // Pointer to the first command
+    Command const* /* const */ mCommandEnd = nullptr;     // Pointer to one past the last command
     // a UBO for instanced primitives
-    BufferObjectSharedHandle mInstancedUboHandle;
+    mutable BufferObjectSharedHandle mInstancedUboHandle;
     // a vector for our custom commands
     using CustomCommandVector = std::vector<Executor::CustomCommandFn,
             utils::STLAllocator<Executor::CustomCommandFn, LinearAllocatorArena>>;
@@ -526,7 +529,9 @@ public:
     // like above but allows to set specific flags
     RenderPassBuilder& renderFlags(
             RenderPass::RenderFlags mask, RenderPass::RenderFlags value) noexcept {
-        mFlags = (mFlags & mask) | (value & mask);
+        value &= mask;
+        mFlags &= ~mask;
+        mFlags |= value;
         return *this;
     }
 

filament/src/RenderPrimitive.h

@@ -56,6 +56,7 @@ public:
     backend::VertexBufferInfoHandle getVertexBufferInfoHandle() const { return mVertexBufferInfoHandle; }
     uint32_t getIndexOffset() const noexcept { return mIndexOffset; }
     uint32_t getIndexCount() const noexcept { return mIndexCount; }
+    uint32_t getMorphingBufferOffset() const noexcept { return mMorphingBufferOffset; }
 
     backend::PrimitiveType getPrimitiveType() const noexcept { return mPrimitiveType; }
     AttributeBitset getEnabledAttributes() const noexcept { return mEnabledAttributes; }
@@ -72,6 +73,10 @@ public:
         mGlobalBlendOrderEnabled = enabled;
     }
 
+    void setMorphingBufferOffset(uint32_t offset) noexcept {
+        mMorphingBufferOffset = offset;
+    }
+
 private:
     // These first fields are dereferences from PrimitiveInfo, keep them together
     struct {
@@ -80,6 +85,7 @@ private:
         backend::Handle<backend::HwVertexBufferInfo> mVertexBufferInfoHandle = {};
         uint32_t mIndexOffset = 0;
         uint32_t mIndexCount = 0;
+        uint32_t mMorphingBufferOffset = 0;
     };
 
     AttributeBitset mEnabledAttributes = {};

filament/src/RenderableManager.cpp

@@ -158,20 +158,14 @@ void RenderableManager::setMorphWeights(Instance instance, float const* weights,
     downcast(this)->setMorphWeights(instance, weights, count, offset);
 }
 
-void RenderableManager::setMorphTargetBufferAt(Instance instance, uint8_t level, size_t primitiveIndex,
-        MorphTargetBuffer* morphTargetBuffer, size_t offset, size_t count) {
-    downcast(this)->setMorphTargetBufferAt(instance, level, primitiveIndex,
-            downcast(morphTargetBuffer), offset, count);
+void RenderableManager::setMorphTargetBufferOffsetAt(Instance instance, uint8_t level,
+        size_t primitiveIndex,
+        size_t offset) {
+    downcast(this)->setMorphTargetBufferOffsetAt(instance, level, primitiveIndex, offset);
 }
 
-void RenderableManager::setMorphTargetBufferAt(
-        Instance instance, uint8_t level, size_t primitiveIndex, size_t offset, size_t count) {
-    downcast(this)->setMorphTargetBufferAt(instance, level, primitiveIndex, offset, count);
-}
-
-MorphTargetBuffer* RenderableManager::getMorphTargetBufferAt(Instance instance, uint8_t level,
-        size_t primitiveIndex) const noexcept {
-    return downcast(this)->getMorphTargetBufferAt(instance, level, primitiveIndex);
+MorphTargetBuffer* RenderableManager::getMorphTargetBuffer(Instance instance) const noexcept {
+    return downcast(this)->getMorphTargetBuffer(instance);
 }
 
 size_t RenderableManager::getMorphTargetCount(Instance instance) const noexcept {

filament/src/Renderer.cpp

@@ -14,11 +14,21 @@
  * limitations under the License.
  */
 
-#include "details/Renderer.h"
+#include <filament/Renderer.h>
+
+#include "ResourceAllocator.h"
 
 #include "details/Engine.h"
+#include "details/Renderer.h"
 #include "details/View.h"
 
+#include <utils/FixedCapacityVector.h>
+
+#include <utility>
+
+#include <stddef.h>
+#include <stdint.h>
+
 namespace filament {
 
 using namespace math;
@@ -36,6 +46,10 @@ void Renderer::setPresentationTime(int64_t monotonic_clock_ns) {
     downcast(this)->setPresentationTime(monotonic_clock_ns);
 }
 
+void Renderer::skipFrame(uint64_t vsyncSteadyClockTimeNano) {
+    downcast(this)->skipFrame(vsyncSteadyClockTimeNano);
+}
+
 bool Renderer::beginFrame(SwapChain* swapChain, uint64_t vsyncSteadyClockTimeNano) {
     return downcast(this)->beginFrame(downcast(swapChain), vsyncSteadyClockTimeNano);
 }
@@ -89,4 +103,16 @@ void Renderer::renderStandaloneView(View const* view) {
     downcast(this)->renderStandaloneView(downcast(view));
 }
 
+void Renderer::setVsyncTime(uint64_t steadyClockTimeNano) noexcept {
+    downcast(this)->setVsyncTime(steadyClockTimeNano);
+}
+
+utils::FixedCapacityVector<Renderer::FrameInfo> Renderer::getFrameInfoHistory(size_t historySize) const noexcept {
+    return downcast(this)->getFrameInfoHistory(historySize);
+}
+
+size_t Renderer::getMaxFrameHistorySize() const noexcept {
+    return downcast(this)->getMaxFrameHistorySize();
+}
+
 } // namespace filament