Add missing VulkanDriver.h header

Keep it's original behavior of no-op

.github/workflows/postsubmit.yml

@@ -21,8 +21,9 @@ jobs:
       - name: Run update script
         env:
           GH_TOKEN: ${{ secrets.FILAMENTBOT_TOKEN }}
+          COMMIT_MESSAGE: ${{ steps.get_commit_msg.outputs.msg }}
         run: |
-          GOLDEN_BRANCH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | python3 test/renderdiff/src/commit_msg.py)
+          GOLDEN_BRANCH=$(echo "${COMMIT_MESSAGE}" | python3 test/renderdiff/src/commit_msg.py)
           COMMIT_HASH="${{ steps.get_commit_msg.outputs.hash }}"
           if [[ "${GOLDEN_BRANCH}" != "main" ]]; then
               git config --global user.email "filament.bot@gmail.com"

.github/workflows/presubmit.yml

@@ -134,18 +134,17 @@ jobs:
         env:
           COMMIT_MESSAGE: ${{ steps.get_commit_msg.outputs.msg }}
         run: |
-          ls ./gltf/Models
           TEST_DIR=test/renderdiff
           source ${TEST_DIR}/src/preamble.sh
-          start_
+          set -eux
           GOLDEN_BRANCH=$(echo "${COMMIT_MESSAGE}" | python3 ${TEST_DIR}/src/commit_msg.py)
-          bash ${TEST_DIR}/generate.sh && \
+          bash ${TEST_DIR}/generate.sh
-              python3 ${TEST_DIR}/src/golden_manager.py \
+          python3 ${TEST_DIR}/src/golden_manager.py \
                   --branch=${GOLDEN_BRANCH} \
                   --output=${GOLDEN_OUTPUT_DIR}
 
-          # Note that we need to upload the output even if comparison fails, so we undo `set -ex`
+          # Note that we need to upload the output even if comparison fails, so we undo `set -eux`
-          end_
+          set +eux
 
           python3 ${TEST_DIR}/src/compare.py \
               --src=${GOLDEN_OUTPUT_DIR} \
@@ -158,6 +157,7 @@ jobs:
              cat compare_output.txt >> "$GITHUB_OUTPUT"
              echo "$DELIMITER" >> "$GITHUB_OUTPUT"
           fi
+        shell: bash
       - uses: actions/upload-artifact@v4
         with:
           name: presubmit-renderdiff-result

CMakeLists.txt

@@ -872,7 +872,9 @@ add_subdirectory(${LIBRARIES}/utils)
 add_subdirectory(${LIBRARIES}/viewer)
 add_subdirectory(${FILAMENT}/shaders)
 add_subdirectory(${EXTERNAL}/abseil/tnt)
-add_subdirectory(${EXTERNAL}/basisu/tnt)
+# Add zstd before basisu to force it to use the external zstd target,
+# preventing a duplicate symbol conflict with its bundled version.
+add_subdirectory(${EXTERNAL}/zstd/tnt)
 add_subdirectory(${EXTERNAL}/civetweb/tnt)
 add_subdirectory(${EXTERNAL}/imgui/tnt)
 add_subdirectory(${EXTERNAL}/robin-map/tnt)
@@ -886,7 +888,7 @@ add_subdirectory(${EXTERNAL}/jsmn/tnt)
 add_subdirectory(${EXTERNAL}/stb/tnt)
 add_subdirectory(${EXTERNAL}/getopt)
 add_subdirectory(${EXTERNAL}/perfetto/tnt)
-add_subdirectory(${EXTERNAL}/zstd/tnt)
+add_subdirectory(${EXTERNAL}/basisu/tnt)
 
 
 # Note that this has to be placed after mikktspace in order for combine_static_libs to work.

NEW_RELEASE_NOTES.md

@@ -6,7 +6,3 @@
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).
 
 ## Release notes for next branch cut
-
-- engine: add `View::getLastDynamicResolutionScale()` (b/457753622)
-
-- materials: Make Material Instances' UBO descriptor use dynamic offsets. [⚠️ **Recompile Materials**]

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.67.0'
+    implementation 'com.google.android.filament:filament-android:1.68.0'
 }
 ```
 
@@ -51,7 +51,7 @@ Here are all the libraries available in the group `com.google.android.filament`:
 iOS projects can use CocoaPods to install the latest release:
 
 ```shell
-pod 'Filament', '~> 1.67.0'
+pod 'Filament', '~> 1.68.0'
 ```
 
 ## Documentation

RELEASE_NOTES.md

@@ -7,6 +7,15 @@ 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).
 
+## v1.68.1
+
+
+## v1.68.0
+
+- engine: add `View::getLastDynamicResolutionScale()` (b/457753622)
+- Metal: report GPU errors to the platform via `debugUpdateStat` (b/431665753).
+- materials: Make Material Instances' UBO descriptor use dynamic offsets. [⚠️ **Recompile Materials**]
+
 ## v1.67.1
 
 - Metal: Add support for the `SwapChain::CONFIG_MSAA_4_SAMPLES` flag.

android/build.gradle

@@ -100,7 +100,7 @@ buildscript {
         'kotlin': '2.0.21',
         'kotlin_coroutines': '1.9.0',
         'buildTools': '35.0.0',
-        'ndk': '27.0.11718014',
+        'ndk': '29.0.14206865',
         'androidx_core': '1.13.1',
         'androidx_annotations': '1.9.0'
     ]

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

@@ -25,7 +25,7 @@ using namespace filament;
 
 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Material_nBuilderBuild(JNIEnv *env, jclass,
-        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount, jint shadowQuality) {
+        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount, jint shadowQuality, jint uboBatchingMode) {
     Engine* engine = (Engine*) nativeEngine;
     AutoBuffer buffer(env, buffer_, size);
     auto builder = Material::Builder();
@@ -33,6 +33,7 @@ Java_com_google_android_filament_Material_nBuilderBuild(JNIEnv *env, jclass,
         builder.sphericalHarmonicsBandCount(shBandCount);
     }
     builder.shadowSamplingQuality((Material::Builder::ShadowSamplingQuality)shadowQuality);
+    builder.uboBatching((Material::UboBatchingMode)uboBatchingMode);
     Material* material = builder
             .package(buffer.getData(), buffer.getSize())
             .build(*engine);

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

@@ -261,6 +261,20 @@ public class Material {
         LOW
     }
 
+    /**
+     * Defines whether a material instance should use UBO batching or not.
+     */
+    public enum UboBatchingMode {
+        /**
+         * For default, it follows the engine settings.
+         * If UBO batching is enabled on the engine and the material domain is SURFACE, it
+         * turns on the UBO batching. Otherwise, it turns off the UBO batching.
+        */
+        DEFAULT,
+        /** Disable the Ubo Batching for this material */
+        DISABLED
+    }
+
     public static class UserVariantFilterBit {
         /** Directional lighting */
         public static int DIRECTIONAL_LIGHTING = 0x01;
@@ -372,6 +386,7 @@ public class Material {
         private int mSize;
         private int mShBandCount = 0;
         private ShadowSamplingQuality mShadowSamplingQuality = ShadowSamplingQuality.LOW;
+        private UboBatchingMode mUboBatchingMode = UboBatchingMode.DEFAULT;
 
 
         /**
@@ -416,6 +431,17 @@ public class Material {
             return this;
         }
 
+        /**
+         * Set the batching mode of the instances created from this material.
+         * @param uboBatchingMode
+         * @return Reference to this Builder for chaining calls.
+         */
+        @NonNull
+        public Builder uboBatching(UboBatchingMode mode) {
+            mUboBatchingMode = mode;
+            return this;
+        }
+
         /**
          * Creates and returns the Material object.
          *
@@ -428,7 +454,7 @@ public class Material {
         @NonNull
         public Material build(@NonNull Engine engine) {
             long nativeMaterial = nBuilderBuild(engine.getNativeObject(),
-                mBuffer, mSize, mShBandCount, mShadowSamplingQuality.ordinal());
+                mBuffer, mSize, mShBandCount, mShadowSamplingQuality.ordinal(), mUboBatchingMode.ordinal());
             if (nativeMaterial == 0) throw new IllegalStateException("Couldn't create Material");
             return new Material(nativeMaterial);
         }
@@ -1094,7 +1120,7 @@ public class Material {
         mNativeObject = 0;
     }
 
-    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount, int shadowQuality);
+    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount, int shadowQuality, int uboBatchingMode);
     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-utils-android/src/main/java/com/google/android/filament/utils/ModelViewer.kt

@@ -16,6 +16,9 @@
 
 package com.google.android.filament.utils
 
+import android.graphics.Bitmap
+import android.os.Handler
+import android.os.Looper
 import android.view.MotionEvent
 import android.view.Surface
 import android.view.SurfaceView
@@ -26,6 +29,7 @@ import com.google.android.filament.android.UiHelper
 import com.google.android.filament.gltfio.*
 import kotlinx.coroutines.*
 import java.nio.Buffer
+import java.nio.ByteBuffer
 
 private const val kNearPlane = 0.05f     // 5 cm
 private const val kFarPlane = 1000.0f    // 1 km
@@ -119,6 +123,8 @@ class ModelViewer(
     private val target = DoubleArray(3)
     private val upward = DoubleArray(3)
 
+    private var debugFrameCallback: ((Bitmap) -> Unit)? = null
+
     init {
         renderer = engine.createRenderer()
         scene = engine.createScene()
@@ -305,10 +311,39 @@ class ModelViewer(
         // Render the scene, unless the renderer wants to skip the frame.
         if (renderer.beginFrame(swapChain!!, frameTimeNanos)) {
             renderer.render(view)
+
+            debugFrameCallback?.let {
+                val viewport = view.viewport
+                val bitmap = Bitmap.createBitmap(viewport.width, viewport.height,
+                        Bitmap.Config.ARGB_8888)
+                val buffer = ByteBuffer.allocateDirect(viewport.width * viewport.height * 4)
+
+                val handler = Handler(Looper.getMainLooper())
+                val pixelBufferDescriptor = Texture.PixelBufferDescriptor(buffer,
+                        Texture.Format.RGBA, Texture.Type.UBYTE, 1, 0, 0, 0, handler) {
+                    buffer.rewind()
+                    bitmap.copyPixelsFromBuffer(buffer)
+                    it(bitmap)
+                }
+                renderer.readPixels(viewport.left, viewport.bottom, viewport.width,
+                                    viewport.height, pixelBufferDescriptor)
+                debugFrameCallback = null
+            }
+
             renderer.endFrame()
         }
     }
 
+    /*
+     * Sets a callback that will be invoked with the next rendered frame as a Bitmap. Note that this
+     * is a one-time callback.
+     *
+     *  @param callback   callback to be invoked with a rendered frame as [Bitmap]
+     */
+    fun debugGetNextFrameCallback(callback: (Bitmap) -> Unit) {
+        debugFrameCallback = callback
+    }
+
     private fun populateScene(asset: FilamentAsset) {
         val rcm = engine.renderableManager
         var count = 0

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.67.0
+VERSION_NAME=1.68.0
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

build.sh

@@ -158,7 +158,7 @@ function print_fgviewer_help {
 }
 
 # Unless explicitly specified, NDK version will be selected as highest available version within same major release chain
-FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION:-$(cat `dirname $0`/build/common/versions | grep GITHUB_NDK_VERSION | sed s/GITHUB_NDK_VERSION=//g | cut -f 1 -d ".")}
+FILAMENT_NDK_VERSION=$(cat `dirname $0`/build/common/versions | grep GITHUB_NDK_VERSION | sed s/GITHUB_NDK_VERSION=//g | cut -f 1 -d ".")
 
 # Internal variables
 ISSUE_CLEAN=false

build/android/build.sh

@@ -18,8 +18,8 @@ if [[ "$GITHUB_WORKFLOW" ]]; then
     fi
 fi
 
-# Unless explicitly specified, NDK version will be set to match exactly the required one
+# Unless explicitly specified, NDK version will be selected as highest available version within same major release chain
-FILAMENT_NDK_VERSION=${GITHUB_NDK_VERSION:-27.0.11718014}
+FILAMENT_NDK_VERSION=$(cat `dirname $0`/../common/versions | grep GITHUB_NDK_VERSION | sed s/GITHUB_NDK_VERSION=//g)
 
 (! grep "${FILAMENT_NDK_VERSION}" `dirname $0`/../../android/build.gradle > /dev/null) &&
     echo "Mismatch of NDK versions: want ${FILAMENT_NDK_VERSION} and not found in android/build.gradle" &&

build/common/versions

@@ -3,7 +3,7 @@ GITHUB_CMAKE_VERSION=3.22.1
 GITHUB_NINJA_VERSION=1.10.2
 GITHUB_MESA_VERSION=24.2.1
 GITHUB_LLVM_VERSION=16
-GITHUB_NDK_VERSION=27.0.11718014
+GITHUB_NDK_VERSION=29.0.14206865
 GITHUB_EMSDK_VERSION=3.1.60
 GITHUB_VULKANSDK_VERSION=1.4.321.0
-GITHUB_GLTF_SAMPLE_ASSETS_COMMIT=d441dfdb87413ff412c620849a649d61789a470f
+GITHUB_GLTF_SAMPLE_ASSETS_COMMIT=d441dfdb87413ff412c620849a649d61789a470f

build/toolchain-aarch64-linux-android.cmake

@@ -75,7 +75,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 #   C_FLAGS += -Wl,-pie
 #   CXX_FLAGS += -lstdc++
 
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIE -mcpu=cortex-a57" CACHE STRING "Toolchain CFLAGS")
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIE -march=armv8-a -mtune=cortex-a78" CACHE STRING "Toolchain CFLAGS")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "Toolchain CXXFLAGS")
 set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fPIE -pie -static-libstdc++" CACHE STRING "Toolchain LDFLAGS")
 set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -static-libstdc++" CACHE STRING "Toolchain LDFLAGS")

build/toolchain-arm7-linux-android.cmake

@@ -87,7 +87,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 # for hardfp: CFLAGS must have -mhard-float
 #             LDFLAGS must have -Wl,--no-warn-mismatch
 #
-set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS} -mthumb -march=armv7-a -mcpu=cortex-a15 -mfloat-abi=softfp -mfpu=neon-vfpv4 -fPIE" CACHE STRING "Toolchain CFLAGS")
+set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS} -mthumb -march=armv7-a -mfloat-abi=softfp -mfpu=neon-vfpv4 -fPIE" CACHE STRING "Toolchain CFLAGS")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "Toolchain CXXFLAGS")
 set(CMAKE_EXE_LINKER_FLAGS    "${CMAKE_EXE_LINKER_FLAGS}    -march=armv7-a -Wl,--no-warn-mismatch -L${TOOLCHAIN}/arm-linux-androideabi/lib/armv7-a -static-libstdc++ -fPIE -pie" CACHE STRING "Toolchain LDFLAGS")
 set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -march=armv7-a -Wl,--no-warn-mismatch -L${TOOLCHAIN}/arm-linux-androideabi/lib/armv7-a -static-libstdc++" CACHE STRING "Toolchain LDFLAGS")

docs/dup/intro.html

@@ -181,7 +181,7 @@ important for <code>matc</code> (material compiler).</p>
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.66.2'
+    implementation 'com.google.android.filament:filament-android:1.68.0'
 }
 </code></pre>
 <p>Here are all the libraries available in the group <code>com.google.android.filament</code>:</p>
@@ -196,7 +196,7 @@ dependencies {
 </div>
 <h3 id="ios"><a class="header" href="#ios">iOS</a></h3>
 <p>iOS projects can use CocoaPods to install the latest release:</p>
-<pre><code class="language-shell">pod 'Filament', '~&gt; 1.66.2'
+<pre><code class="language-shell">pod 'Filament', '~&gt; 1.68.0'
 </code></pre>
 <h2 id="documentation"><a class="header" href="#documentation">Documentation</a></h2>
 <ul>

docs_src/build/requirements.txt

@@ -10,6 +10,6 @@ sortedcontainers==2.4.0
 trio==0.31.0
 trio-websocket==0.12.2
 typing_extensions==4.15.0
-urllib3==2.5.0
+urllib3==2.6.0
 websocket-client==1.9.0
 wsproto==1.2.0

filament/backend/CMakeLists.txt

@@ -206,6 +206,8 @@ if (FILAMENT_SUPPORTS_VULKAN)
             src/vulkan/VulkanDriverFactory.h
             src/vulkan/VulkanExternalImageManager.cpp
             src/vulkan/VulkanExternalImageManager.h
+            src/vulkan/VulkanStreamedImageManager.cpp
+            src/vulkan/VulkanStreamedImageManager.h
             src/vulkan/VulkanFboCache.cpp
             src/vulkan/VulkanFboCache.h
             src/vulkan/VulkanHandles.cpp
@@ -316,6 +318,8 @@ if (FILAMENT_SUPPORTS_WEBGPU)
             src/webgpu/WebGPURenderPrimitive.h
             src/webgpu/WebGPURenderTarget.cpp
             src/webgpu/WebGPURenderTarget.h
+            src/webgpu/WebGPUStagePool.cpp
+            src/webgpu/WebGPUStagePool.h
             src/webgpu/WebGPUStrings.h
             src/webgpu/WebGPUSwapChain.cpp
             src/webgpu/WebGPUSwapChain.h

filament/backend/include/backend/BufferObjectStreamDescriptor.h

@@ -1,63 +0,0 @@
-/*
- * Copyright (C) 2025 The Android Open Source Project
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef TNT_FILAMENT_BACKEND_BUFFEROBJECTSTREAMDESCRIPTOR_H
-#define TNT_FILAMENT_BACKEND_BUFFEROBJECTSTREAMDESCRIPTOR_H
-
-#include <backend/Handle.h>
-
-#include <vector>
-
-namespace filament::backend {
-
-/**
- * The type of association between a buffer object and a stream.
- */
- enum class BufferObjectStreamAssociationType { TRANSFORM_MATRIX };
-
-/**
- * A descriptor for a buffer object to stream association.
- */
-class UTILS_PUBLIC BufferObjectStreamDescriptor {
-public:
-    //! creates an empty descriptor
-    BufferObjectStreamDescriptor() noexcept = default;
-
-    BufferObjectStreamDescriptor(const BufferObjectStreamDescriptor& rhs) = delete;
-    BufferObjectStreamDescriptor& operator=(const BufferObjectStreamDescriptor& rhs) = delete;
-
-    BufferObjectStreamDescriptor(BufferObjectStreamDescriptor&& rhs) = default;
-    BufferObjectStreamDescriptor& operator=(BufferObjectStreamDescriptor&& rhs) = default;
-
-    // --------------------------------------------------------------------------------------------
-
-    struct StreamDataDescriptor {
-        uint32_t offset;
-        Handle<HwStream> stream;
-        BufferObjectStreamAssociationType associationType;
-    };
-
-    std::vector<StreamDataDescriptor> mStreams;
-};
-
-} // namespace filament::backend
-
-#if !defined(NDEBUG)
-utils::io::ostream& operator<<(utils::io::ostream& out,
-        const filament::backend::BufferObjectStreamDescriptor& b);
-#endif
-
-#endif // TNT_FILAMENT_BACKEND_BUFFEROBJECTSTREAMDESCRIPTOR_H

filament/backend/include/backend/DriverEnums.h

@@ -1713,6 +1713,8 @@ using FrameTimestamps = Platform::FrameTimestamps;
 
 using CompositorTiming = Platform::CompositorTiming;
 
+using AsynchronousMode = Platform::AsynchronousMode;
+
 } // namespace filament::backend
 
 template<> struct utils::EnableBitMaskOperators<filament::backend::ShaderStageFlags>

filament/backend/include/backend/Platform.h

@@ -19,13 +19,17 @@
 #ifndef TNT_FILAMENT_BACKEND_PLATFORM_H
 #define TNT_FILAMENT_BACKEND_PLATFORM_H
 
+#include <utils/CString.h>
 #include <utils/compiler.h>
 #include <utils/Invocable.h>
+#include <utils/Mutex.h>
 
 #include <stddef.h>
 #include <stdint.h>
 
 #include <atomic>
+#include <memory>
+#include <mutex>
 
 namespace filament::backend {
 
@@ -70,6 +74,9 @@ public:
         ExternalImageHandle& operator=(ExternalImageHandle const& rhs) noexcept;
         ExternalImageHandle& operator=(ExternalImageHandle&& rhs) noexcept;
 
+        bool operator==(const ExternalImageHandle& rhs) const noexcept {
+            return mTarget == rhs.mTarget;
+        }
         explicit operator bool() const noexcept { return mTarget != nullptr; }
 
         ExternalImage* UTILS_NULLABLE get() noexcept { return mTarget; }
@@ -248,6 +255,28 @@ public:
         REALTIME,
     };
 
+    /**
+     * Defines how asynchronous operations are handled by the engine.
+     */
+    enum class AsynchronousMode : uint8_t {
+        /**
+         * Asynchronous operations are disabled. This is the default.
+         */
+        NONE,
+
+        /**
+         * Attempts to use a dedicated worker thread for asynchronous tasks. If threading is not
+         * supported by the platform, it automatically falls back to using an amortization strategy.
+         */
+        THREAD_PREFERRED,
+
+        /**
+         * Uses an amortization strategy, processing a small number of asynchronous tasks during
+         * each engine update cycle.
+         */
+        AMORTIZATION,
+    };
+
     struct DriverConfig {
         /**
          * Size of handle arena in bytes. Setting to 0 indicates default value is to be used.
@@ -317,6 +346,11 @@ public:
          *      - VulkanPlatform
          */
         bool vulkanEnableStagingBufferBypass = false;
+
+        /**
+         * Asynchronous mode for the engine. Defines how asynchronous operations are handled.
+         */
+        AsynchronousMode asynchronousMode = AsynchronousMode::NONE;
     };
 
     Platform() noexcept;
@@ -508,13 +542,22 @@ public:
     // --------------------------------------------------------------------------------------------
     // Debugging APIs
 
-    using DebugUpdateStatFunc = utils::Invocable<void(const char* UTILS_NONNULL key, uint64_t value)>;
+    using DebugUpdateStatFunc = utils::Invocable<void(const char* UTILS_NONNULL key,
+            uint64_t intValue, utils::CString stringValue)>;
 
     /**
      * Sets the callback function that the backend can use to update backend-specific statistics
      * to aid with debugging. This callback is guaranteed to be called on the Filament driver
      * thread.
      *
+     * The callback signature is (key, intValue, stringValue). Note that for any given call,
+     * only one of the value parameters (intValue or stringValue) will be meaningful, depending on
+     * the specific key.
+     *
+     * IMPORTANT_NOTE: because the callback is called on the driver thread, only quick, non-blocking
+     * work should be done inside it. Furthermore, no graphics API calls (such as GL calls) should
+     * be made, which could interfere with Filament's driver state.
+     *
      * @param debugUpdateStat   an Invocable that updates debug statistics
      */
     void setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept;
@@ -533,15 +576,32 @@ public:
      * This function is guaranteed to be called only on a single thread, the Filament driver
      * thread.
      *
-     * @param key          a null-terminated C-string with the key of the debug statistic
+     * @param key           a null-terminated C-string with the key of the debug statistic
-     * @param value        the updated value of key
+     * @param intValue      the updated integer value of key (the string value passed to the
+     *                      callback will be empty)
      */
-    void debugUpdateStat(const char* UTILS_NONNULL key, uint64_t value);
+    void debugUpdateStat(const char* UTILS_NONNULL key, uint64_t intValue);
+
+    /**
+     * To track backend-specific statistics, the backend implementation can call the
+     * application-provided callback function debugUpdateStatFunc to associate or update a value
+     * with a given key. It is possible for this function to be called multiple times with the
+     * same key, in which case newer values should overwrite older values.
+     *
+     * This function is guaranteed to be called only on a single thread, the Filament driver
+     * thread.
+     *
+     * @param key           a null-terminated C-string with the key of the debug statistic
+     * @param stringValue   the updated string value of key (the integer value passed to the
+     *                      callback will be 0)
+     */
+    void debugUpdateStat(const char* UTILS_NONNULL key, utils::CString stringValue);
 
 private:
-    InsertBlobFunc mInsertBlob;
+    std::shared_ptr<InsertBlobFunc> mInsertBlob;
-    RetrieveBlobFunc mRetrieveBlob;
+    std::shared_ptr<RetrieveBlobFunc> mRetrieveBlob;
-    DebugUpdateStatFunc mDebugUpdateStat;
+    std::shared_ptr<DebugUpdateStatFunc> mDebugUpdateStat;
+    mutable utils::Mutex mMutex;
 };
 
 } // namespace filament

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

@@ -17,8 +17,6 @@
 #ifndef TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_EGL_ANDROID_H
 #define TNT_FILAMENT_BACKEND_OPENGL_OPENGL_PLATFORM_EGL_ANDROID_H
 
-#include "AndroidSwapChainHelper.h"
-#include "AndroidFrameCallback.h"
 #include "AndroidNdk.h"
 
 #include <backend/AcquiredImage.h>
@@ -32,8 +30,6 @@
 
 #include <math/mat3.h>
 
-#include "AndroidNativeWindow.h"
-
 #include <chrono>
 
 #include <stddef.h>
@@ -171,19 +167,10 @@ protected:
             SwapChain* drawSwapChain,
             SwapChain* readSwapChain) override;
 
-    struct SwapChainEGLAndroid : public SwapChainEGL {
-        SwapChainEGLAndroid(PlatformEGLAndroid const& platform,
-                void* nativeWindow, uint64_t flags);
-        SwapChainEGLAndroid(PlatformEGLAndroid const& platform,
-                uint32_t width, uint32_t height, uint64_t flags);
-        void terminate(PlatformEGLAndroid& platform);
-        bool setPresentFrameId(uint64_t frameId) const noexcept;
-        uint64_t getFrameId(uint64_t frameId) const noexcept;
-    private:
-        AndroidSwapChainHelper mImpl{};
-    };
-
 private:
+    struct SwapChainEGLAndroid;
+    struct AndroidDetails;
+
     // prevent derived classes' implementations to call through
     [[nodiscard]] SwapChain* createSwapChain(void* nativeWindow, uint64_t flags) override;
     [[nodiscard]] SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) override;
@@ -203,17 +190,14 @@ private:
 
     int mOSVersion;
     ExternalStreamManagerAndroid& mExternalStreamManager;
+    AndroidDetails& mAndroidDetails;
     InitializeJvmForPerformanceManagerIfNeeded const mInitializeJvmForPerformanceManagerIfNeeded;
     utils::PerformanceHintManager mPerformanceHintManager;
     utils::PerformanceHintManager::Session mPerformanceHintSession;
-    SwapChainEGLAndroid* mCurrentDrawSwapChain{};
-
     using clock = std::chrono::high_resolution_clock;
     clock::time_point mStartTimeOfActualWork;
-    AndroidProducerThrottling mProducerThrottling;
+    SwapChainEGLAndroid* mCurrentDrawSwapChain{};
     bool mAssertNativeWindowIsValid = false;
-
-    AndroidFrameCallback mAndroidFrameCallback;
 };
 
 } // namespace filament::backend

filament/backend/include/backend/platforms/VulkanPlatformAndroid.h

@@ -17,12 +17,13 @@
 #ifndef TNT_FILAMENT_BACKEND_PLATFORMS_VULKAN_PLATFORM_ANDROID_H
 #define TNT_FILAMENT_BACKEND_PLATFORMS_VULKAN_PLATFORM_ANDROID_H
 
-#include "AndroidFrameCallback.h"
 #include "AndroidNdk.h"
 
 #include <backend/DriverEnums.h>
 #include <backend/platforms/VulkanPlatform.h>
 
+#include <utils/compiler.h>
+
 #include <android/hardware_buffer.h>
 
 namespace filament::backend {
@@ -91,6 +92,8 @@ protected:
     VkExternalFenceHandleTypeFlagBits getFenceExportFlags() const noexcept override;
 
 private:
+    struct AndroidDetails;
+
     struct ExternalImageVulkanAndroid : public ExternalImage {
         AHardwareBuffer* aHardwareBuffer = nullptr;
         bool sRGB = false;
@@ -99,7 +102,7 @@ private:
         ~ExternalImageVulkanAndroid() override;
     };
 
-    AndroidFrameCallback mAndroidFrameCallback;
+    AndroidDetails& mAndroidDetails;
     int mOSVersion{};
 };
 

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

@@ -22,7 +22,6 @@
 #include "private/backend/Driver.h"
 
 #include <backend/BufferDescriptor.h>
-#include <backend/BufferObjectStreamDescriptor.h>
 #include <backend/CallbackHandler.h>
 #include <backend/DriverEnums.h>
 #include <backend/Handle.h>

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

@@ -51,7 +51,6 @@
 namespace filament::backend {
 
 class BufferDescriptor;
-class BufferObjectStreamDescriptor;
 class CallbackHandler;
 class PixelBufferDescriptor;
 class Program;

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

@@ -392,10 +392,6 @@ DECL_DRIVER_API_N(updateBufferObject,
         backend::BufferDescriptor&&, data,
         uint32_t, byteOffset)
 
-DECL_DRIVER_API_N(registerBufferObjectStreams,
-        backend::BufferObjectHandle, boh,
-        backend::BufferObjectStreamDescriptor&&, streams)
-
 DECL_DRIVER_API_N(updateBufferObjectUnsynchronized,
         backend::BufferObjectHandle, boh,
         backend::BufferDescriptor&&, data,

filament/backend/src/AndroidNativeWindow.cpp

@@ -61,6 +61,16 @@ int NativeWindow::enableFrameTimestamps(ANativeWindow* anw, bool enable) {
     return pWindow->perform(anw, ENABLE_FRAME_TIMESTAMPS, enable);
 }
 
+int NativeWindow::frameTimestampsSupportsPresent(ANativeWindow* anw, bool* outSupportsPresent) {
+    NativeWindow const* pWindow = reinterpret_cast<NativeWindow const*>(anw);
+    int value = 0;
+    bool const success = pWindow->perform(anw, FRAME_TIMESTAMPS_SUPPORTS_PRESENT, &value);
+    if (success) {
+        *outSupportsPresent = bool(value);
+    }
+    return success;
+}
+
 int NativeWindow::getCompositorTiming(ANativeWindow* anw,
         int64_t* compositeDeadline, int64_t* compositeInterval,
         int64_t* compositeToPresentLatency) {

filament/backend/src/AndroidNativeWindow.h

@@ -32,6 +32,7 @@ struct NativeWindow {
     // is valid query enum value
     enum {
         IS_VALID                = 17,
+        FRAME_TIMESTAMPS_SUPPORTS_PRESENT = 18,
         GET_NEXT_FRAME_ID       = 24,
         ENABLE_FRAME_TIMESTAMPS = 25,
         GET_COMPOSITOR_TIMING   = 26,
@@ -51,6 +52,7 @@ struct NativeWindow {
 
     static int getNextFrameId(ANativeWindow* anw, uint64_t* frameId);
     static int enableFrameTimestamps(ANativeWindow* anw, bool enable);
+    static int frameTimestampsSupportsPresent(ANativeWindow* anw, bool* outSupportsPresent);
     static int getCompositorTiming(ANativeWindow* anw,
             int64_t* compositeDeadline, int64_t* compositeInterval,
             int64_t* compositeToPresentLatency);

filament/backend/src/AndroidNdk.cpp

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "AndroidNdk.h"
+#include <backend/platforms/AndroidNdk.h>
 
 #include <android/hardware_buffer.h>
 

filament/backend/src/AndroidSwapChainHelper.cpp

@@ -19,6 +19,9 @@
 
 #include <android/native_window.h>
 
+#include <utils/compiler.h>
+#include <utils/Logger.h>
+
 #include <cstddef>
 #include <cstdint>
 #include <limits>
@@ -36,11 +39,23 @@ bool AndroidSwapChainHelper::setPresentFrameId(
     int const status = NativeWindow::getNextFrameId(anw, &sysFrameId);
     if (status == 0) {
         std::lock_guard const lock(mLock);
-        auto const pos = mFrameIdToSystemFrameId.find(frameId);
+        // frameIds must be strictly monotonic, if that's not the case (i.e. the new frameId is
-        if (pos && *pos != sysFrameId) {
+        // less or equal to the last one in the map), we have to clear the map, because the
-            // we're trying to associate the same frame id to a different frame!
+        // map's find() assume sorted keys.
-            return false;
+        // This case can happen if two different filament::Renderer are used with the same
+        // ANativeWindow (the Renderer would have different frameIds). This is expected to
+        // be a rare case.
+        if (UTILS_UNLIKELY(!mFrameIdToSystemFrameId.empty() &&
+                frameId <= mFrameIdToSystemFrameId.back().first)) {
+            // this log is expected to happen very rarely
+            DLOG(INFO) << "clearing frame history anw=" << anw
+                    << ", frameId=" << frameId
+                    << ", previous=" << mFrameIdToSystemFrameId.back().first
+                    << ", sysFrameId=" << sysFrameId;
+            // clear the frame history
+            mFrameIdToSystemFrameId.clear();
         }
+
         // oldest entry is removed
         mFrameIdToSystemFrameId.insert(frameId, sysFrameId);
         return true;

filament/backend/src/Platform.cpp

@@ -139,42 +139,73 @@ bool Platform::queryFrameTimestamps(SwapChain const*, uint64_t, FrameTimestamps*
 }
 
 void Platform::setBlobFunc(InsertBlobFunc&& insertBlob, RetrieveBlobFunc&& retrieveBlob) noexcept {
-    mInsertBlob = std::move(insertBlob);
+    std::lock_guard<decltype(mMutex)> lock(mMutex);
-    mRetrieveBlob = std::move(retrieveBlob);
+    mInsertBlob = std::make_shared<InsertBlobFunc>(std::move(insertBlob));
+    mRetrieveBlob = std::make_shared<RetrieveBlobFunc>(std::move(retrieveBlob));
 }
 
 bool Platform::hasInsertBlobFunc() const noexcept {
-    return bool(mInsertBlob);
+    std::lock_guard<decltype(mMutex)> lock(mMutex);
+    return mInsertBlob && bool(*mInsertBlob);
 }
 
 bool Platform::hasRetrieveBlobFunc() const noexcept {
-    return bool(mRetrieveBlob);
+    std::lock_guard<decltype(mMutex)> lock(mMutex);
+    return mRetrieveBlob && bool(*mRetrieveBlob);
 }
 
 void Platform::insertBlob(void const* key, size_t keySize, void const* value, size_t valueSize) {
-    if (mInsertBlob) {
+    std::shared_ptr<InsertBlobFunc> callback;
-        mInsertBlob(key, keySize, value, valueSize);
+    {
+        std::unique_lock<decltype(mMutex)> lock(mMutex);
+        callback = mInsertBlob;
+    }
+    if (callback) {
+        (*callback)(key, keySize, value, valueSize);
     }
 }
 
 size_t Platform::retrieveBlob(void const* key, size_t keySize, void* value, size_t valueSize) {
-    if (mRetrieveBlob) {
+    std::shared_ptr<RetrieveBlobFunc> callback;
-        return mRetrieveBlob(key, keySize, value, valueSize);
+    {
+        std::unique_lock<decltype(mMutex)> lock(mMutex);
+        callback = mRetrieveBlob;
+    }
+    if (callback) {
+        return (*callback)(key, keySize, value, valueSize);
     }
     return 0;
 }
 
 void Platform::setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept {
-    mDebugUpdateStat = std::move(debugUpdateStat);
+    std::lock_guard<decltype(mMutex)> lock(mMutex);
+    mDebugUpdateStat = std::make_shared<DebugUpdateStatFunc>(std::move(debugUpdateStat));
 }
 
 bool Platform::hasDebugUpdateStatFunc() const noexcept {
-    return bool(mDebugUpdateStat);
+    std::lock_guard<decltype(mMutex)> lock(mMutex);
+    return mDebugUpdateStat && bool(*mDebugUpdateStat);
 }
 
-void Platform::debugUpdateStat(const char* key, uint64_t value) {
+void Platform::debugUpdateStat(const char* key, uint64_t intValue) {
-    if (mDebugUpdateStat) {
+    std::shared_ptr<DebugUpdateStatFunc> callback;
-        mDebugUpdateStat(key, value);
+    {
+        std::unique_lock<decltype(mMutex)> lock(mMutex);
+        callback = mDebugUpdateStat;
+    }
+    if (callback) {
+        (*callback)(key, intValue, "");
+    }
+}
+
+void Platform::debugUpdateStat(const char* key, utils::CString stringValue) {
+    std::shared_ptr<DebugUpdateStatFunc> callback;
+    {
+        std::unique_lock<decltype(mMutex)> lock(mMutex);
+        callback = mDebugUpdateStat;
+    }
+    if (callback) {
+        (*callback)(key, 0, stringValue);
     }
 }
 

filament/backend/src/metal/MetalBuffer.h

@@ -165,7 +165,7 @@ public:
          size_t size, bool forceGpuBuffer = false);
     ~MetalBuffer();
 
-    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer || mCpuBuffer; }
+    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer; }
 
     MetalBuffer(const MetalBuffer& rhs) = delete;
     MetalBuffer& operator=(const MetalBuffer& rhs) = delete;
@@ -185,14 +185,12 @@ public:
      * Denotes that this buffer is used for a draw call ensuring that its allocation remains valid
      * until the end of the current frame.
      *
-     * @return The MTLBuffer representing the current state of the buffer to bind, or nil if there
+     * @return The MTLBuffer representing the current state of the buffer to bind, it never returns
-     * is no device allocation.
+     * nil.
      *
      */
     id<MTLBuffer> getGpuBufferForDraw() noexcept;
 
-    void* getCpuBuffer() const noexcept { return mCpuBuffer; }
-
     void setLabel(const utils::ImmutableCString& label) {
 #if FILAMENT_METAL_DEBUG_LABELS
         if (label.empty()) {
@@ -235,7 +233,6 @@ private:
     UploadStrategy mUploadStrategy;
     TrackedMetalBuffer mBuffer;
     size_t mBufferSize = 0;
-    void* mCpuBuffer = nullptr;
     MetalContext& mContext;
 };
 

filament/backend/src/metal/MetalBuffer.mm

@@ -39,34 +39,30 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
         mUploadStrategy = UploadStrategy::POOL;
     }
 
-    // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
+    MTLResourceOptions options = MTLResourceStorageModePrivate;
-    // buffer. Instead, we use immediate command encoder methods like setVertexBytes:length:atIndex:.
-    // This won't work for SSBOs, since they are read/write.
 
-    /*
+    // The buffer will be memory mapped for write operations.
-    if (size <= 4 * 1024 && bindingType != BufferObjectBinding::SHADER_STORAGE &&
+    if (any(usage & BufferUsage::SHARED_WRITE_BIT)) {
-            usage == BufferUsage::DYNAMIC && !forceGpuBuffer) {
+#if defined(FILAMENT_IOS) || defined(__arm64__) || defined(__aarch64__)
-        mBuffer = nil;
+        // iOS and Apple Silicon devices use UMA (Unified Memory Architecture), so we use Shared memory.
-        mCpuBuffer = malloc(size);
+        options = MTLResourceStorageModeShared;
-        return;
+#else
+        // Intel Macs require Managed memory for CPU/GPU synchronization.
+        options = MTLResourceStorageModeManaged;
+#endif
     }
-    */
 
-    // Otherwise, we allocate a private GPU buffer.
     {
         ScopedAllocationTimer timer("generic");
-        mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
+        mBuffer = { [context.device newBufferWithLength:size options:options],
             TrackedMetalBuffer::Type::GENERIC };
     }
+
     // mBuffer might fail to be allocated. Clients can check for this by calling
     // wasAllocationSuccessful().
 }
 
-MetalBuffer::~MetalBuffer() {
+MetalBuffer::~MetalBuffer() = default;
-    if (mCpuBuffer) {
-        free(mCpuBuffer);
-    }
-}
 
 void MetalBuffer::copyIntoBuffer(
         void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
@@ -83,12 +79,6 @@ void MetalBuffer::copyIntoBuffer(
     FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3))
             << "byteOffset must be a multiple of 4, tag=" << getHandleTag();
 
-    // If we have a cpu buffer, we can directly copy into it.
-    if (mCpuBuffer) {
-        memcpy(static_cast<uint8_t*>(mCpuBuffer) + byteOffset, src, size);
-        return;
-    }
-
     switch (mUploadStrategy) {
         case UploadStrategy::BUMP_ALLOCATOR:
             uploadWithBumpAllocator(src, size, byteOffset, std::move(getHandleTag));
@@ -106,11 +96,6 @@ void MetalBuffer::copyIntoBufferUnsynchronized(
 }
 
 id<MTLBuffer> MetalBuffer::getGpuBufferForDraw() noexcept {
-    // If there's a CPU buffer, then we return nil here, as the CPU-side buffer will be bound
-    // separately.
-    if (mCpuBuffer) {
-        return nil;
-    }
     assert_invariant(mBuffer);
     return mBuffer.get();
 }
@@ -171,41 +156,6 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
                                                    offsets:metalOffsets.data()
                                                  withRange:bufferRange];
     }
-
-    for (size_t b = 0; b < count; b++) {
-        MetalBuffer* const buffer = buffers[b];
-        if (!buffer) {
-            continue;
-        }
-
-        const void* cpuBuffer = buffer->getCpuBuffer();
-        if (!cpuBuffer) {
-            continue;
-        }
-
-        const size_t bufferIndex = bufferStart + b;
-        const size_t offset = offsets[b];
-        auto* bytes = static_cast<const uint8_t*>(cpuBuffer);
-
-        if (stages & Stage::VERTEX) {
-            [(id<MTLRenderCommandEncoder>) encoder setVertexBytes:(bytes + offset)
-                                                           length:(buffer->getSize() - offset)
-                                                          atIndex:bufferIndex];
-        }
-        if (stages & Stage::FRAGMENT) {
-            [(id<MTLRenderCommandEncoder>) encoder setFragmentBytes:(bytes + offset)
-                                                             length:(buffer->getSize() - offset)
-                                                            atIndex:bufferIndex];
-        }
-        if (stages & Stage::COMPUTE) {
-            // TODO: using setBytes means the data is read-only, which currently isn't enforced.
-            // In practice this won't be an issue since MetalBuffer ensures all SSBOs are realized
-            // through actual id<MTLBuffer> allocations.
-            [(id<MTLComputeCommandEncoder>) encoder setBytes:(bytes + offset)
-                                                      length:(buffer->getSize() - offset)
-                                                     atIndex:bufferIndex];
-        }
-    }
 }
 
 void MetalBuffer::uploadWithPoolBuffer(

filament/backend/src/metal/MetalContext.h

@@ -17,6 +17,7 @@
 #ifndef TNT_METALCONTEXT_H
 #define TNT_METALCONTEXT_H
 
+#include "MetalErrorQueue.h"
 #include "MetalResourceTracker.h"
 #include "MetalShaderCompiler.h"
 #include "MetalState.h"
@@ -129,6 +130,7 @@ struct MetalContext {
     id<MTLCommandBuffer> pendingCommandBuffer = nil;
     id<MTLRenderCommandEncoder> currentRenderPassEncoder = nil;
     uint32_t currentFrame = 0;
+    MetalErrorQueue commandBufferErrors;
 
     std::atomic<bool> memorylessLimitsReached = false;
 

filament/backend/src/metal/MetalContext.mm

@@ -153,6 +153,7 @@ id<MTLCommandBuffer> getPendingCommandBuffer(MetalContext* context) {
 
         if (UTILS_UNLIKELY(errorCode != MTLCommandBufferErrorNone)) {
             logMTLCommandBufferError(errorCode);
+            context->commandBufferErrors.push(buffer.error);
         }
     }];
     FILAMENT_CHECK_POSTCONDITION(context->pendingCommandBuffer)

filament/backend/src/metal/MetalDriver.h

@@ -61,6 +61,8 @@ public:
 
     MetalContext* getContext() { return mContext; }
 
+    using DriverBase::scheduleDestroy;
+
 private:
 
     friend class MetalSwapChain;

filament/backend/src/metal/MetalDriver.mm

@@ -254,6 +254,18 @@ MetalDriver::~MetalDriver() noexcept {
 void MetalDriver::tick(int) {
     executeTickOps();
     executeDeferredOps();
+
+    // Notify platform of GPU errors.
+    auto& platform = mPlatform;
+    if (UTILS_UNLIKELY(!mContext->commandBufferErrors.isEmpty())) {
+        mContext->commandBufferErrors.flush([&platform](NSError* error) {
+            if (UTILS_VERY_UNLIKELY(!error)) {
+                return;
+            }
+            const utils::CString errorString(error.localizedDescription.UTF8String);
+            platform.debugUpdateStat("filament.metal.command_buffer_error", errorString);
+        });
+    }
 }
 
 void MetalDriver::beginFrame(int64_t monotonic_clock_ns,
@@ -1132,8 +1144,12 @@ bool MetalDriver::isRenderTargetFormatSupported(TextureFormat format) {
 bool MetalDriver::isFrameBufferFetchSupported() {
     // FrameBuffer fetch is achievable via "programmable blending" in Metal, and only supported on
     // Apple GPUs with readWriteTextureSupport.
-    return mContext->highestSupportedGpuFamily.apple >= 1 &&
+    // On macOS, framebuffer fetch requires MSL 2.3, which is only available with macOS 11.0.
-            mContext->device.readWriteTextureSupport;
+    if (@available(macOS 11.0, *)) {
+        return mContext->highestSupportedGpuFamily.apple >= 1 &&
+               mContext->device.readWriteTextureSupport;
+    }
+    return false;
 }
 
 bool MetalDriver::isFrameBufferFetchMultiSampleSupported() {
@@ -1274,10 +1290,6 @@ void MetalDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&
     scheduleDestroy(std::move(data));
 }
 
-void MetalDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh, BufferObjectStreamDescriptor&& streams) {
-    // Noop
-}
-
 void MetalDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& data,
         uint32_t byteOffset) {
     FILAMENT_CHECK_PRECONDITION(!isInRenderPass(mContext))
@@ -2263,7 +2275,10 @@ MemoryMappedBufferHandle MetalDriver::mapBufferS() noexcept {
 void MetalDriver::mapBufferR(MemoryMappedBufferHandle mmbh,
         BufferObjectHandle boh, size_t offset,
         size_t size, MapBufferAccessFlags access, utils::ImmutableCString&& tag) {
-    construct_handle<MetalMemoryMappedBuffer>(mmbh, boh, offset, size, access);
+    assert_invariant(boh);
+    MetalBufferObject* bo = mHandleAllocator.handle_cast<MetalBufferObject*>(boh);
+    assert_invariant(bo);
+    construct_handle<MetalMemoryMappedBuffer>(mmbh, bo, offset, size, access);
     mHandleAllocator.associateTagToHandle(mmbh.getId(), std::move(tag));
 }
 
@@ -2271,21 +2286,16 @@ void MetalDriver::unmapBuffer(MemoryMappedBufferHandle mmbh) {
     if (UTILS_UNLIKELY(!mmbh)) {
         return;
     }
+
+    auto* mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
+    mmb->unmap();
     destruct_handle<MetalMemoryMappedBuffer>(mmbh);
 }
 
 void MetalDriver::copyToMemoryMappedBuffer(MemoryMappedBufferHandle mmbh, size_t offset,
         BufferDescriptor&& data) {
-    auto mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
+    auto* mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
-
+    mmb->copy(*this, offset, std::move(data));
-    assert_invariant(any(mmb->access & MapBufferAccessFlags::WRITE_BIT));
-    assert_invariant(offset + data.size <= mmb->size);
-
-    // TODO: this isa zero-effort implementation of copyToMemoryMappedBuffer(), where we just
-    //       call updateBufferObject(). This could be a fallback implementation for when
-    //       shared memory is not available.
-    //       On UMA systems, this should just be a memcpy into the memory-mapped buffer.
-    updateBufferObject(mmb->boh, std::move(data), mmb->offset + offset);
 }
 
 // explicit instantiation of the Dispatcher

filament/backend/src/metal/MetalErrorQueue.h

@@ -0,0 +1,66 @@
+/*
+ * Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_DRIVER_METALERRORQUEUE_H
+#define TNT_FILAMENT_DRIVER_METALERRORQUEUE_H
+
+#import <Foundation/Foundation.h>
+
+#include <utils/compiler.h>
+
+#include <atomic>
+#include <functional>
+#include <mutex>
+#include <vector>
+
+class MetalErrorQueue {
+public:
+    bool isEmpty() const {
+        return !mHasErrors.load(std::memory_order_relaxed);
+    }
+
+    void push(NSError* error) {
+        std::lock_guard<std::mutex> lock(mMutex);
+        mErrors.push_back(error);
+        mHasErrors.store(true, std::memory_order_relaxed);
+    }
+
+    void flush(const std::function<void(NSError*)>& callback) {
+        if (UTILS_LIKELY(isEmpty())) {
+            return;
+        }
+        std::vector<NSError*> errors;
+        {
+            std::lock_guard<std::mutex> lock(mMutex);
+            std::swap(mErrors, errors);
+            mHasErrors.store(false, std::memory_order_relaxed);
+        }
+
+        for (const auto& error: errors) {
+            callback(error);
+        }
+    }
+
+private:
+    std::vector<NSError*> mErrors;
+    std::mutex mMutex;
+
+    // Optimization to avoid locking the mutex at each call to flush.
+    std::atomic<bool> mHasErrors;
+};
+
+
+#endif // TNT_FILAMENT_DRIVER_METALERRORQUEUE_H

filament/backend/src/metal/MetalHandles.h

@@ -569,14 +569,22 @@ struct MetalDescriptorSet : public HwDescriptorSet {
 
 
 struct MetalMemoryMappedBuffer : public HwMemoryMappedBuffer {
-    MetalMemoryMappedBuffer(BufferObjectHandle boh, size_t const offset,
-            size_t const size, MapBufferAccessFlags const access)
-        : boh(boh), access(access), size(size), offset(offset) {
-    }
-    BufferObjectHandle boh{};
     MapBufferAccessFlags access{};
-    uint32_t size = 0;
+    struct {
-    uint32_t offset = 0;
+        MetalBufferObject* bo;
+        void* vaddr = nullptr;
+        uint32_t size = 0;
+        uint32_t offset = 0;
+    } mtl;
+
+    MetalMemoryMappedBuffer(MetalBufferObject* bo, size_t offset, size_t size,
+            MapBufferAccessFlags access) noexcept;
+
+    ~MetalMemoryMappedBuffer();
+
+    void unmap();
+
+    void copy(MetalDriver& mtld, size_t offset, BufferDescriptor&& data) const;
 };
 
 } // namespace backend

filament/backend/src/metal/MetalHandles.mm

@@ -1664,5 +1664,44 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
     return buffer.get();
 }
 
+MetalMemoryMappedBuffer::MetalMemoryMappedBuffer(MetalBufferObject* bo, size_t offset, size_t size,
+    MapBufferAccessFlags access) noexcept : access(access) {
+    MetalBuffer* buffer = bo->getBuffer();
+    assert_invariant(buffer);
+    id<MTLBuffer> mtlBuffer = buffer->getGpuBufferForDraw();
+
+    assert_invariant(offset + size <= bo->byteCount);
+    assert_invariant(mtlBuffer.storageMode != MTLStorageModePrivate);
+
+    mtl.bo = bo;
+    mtl.vaddr = static_cast<char*>(mtlBuffer.contents) + offset;
+    mtl.size = size;
+    mtl.offset = offset;
+}
+
+MetalMemoryMappedBuffer::~MetalMemoryMappedBuffer() = default;
+
+void MetalMemoryMappedBuffer::unmap() {
+#if !defined(FILAMENT_IOS) && defined(__x86_64__)
+    // Managed memory requires didModifyRange to synchronize changes to the GPU. This is specific to Intel Macs.
+    MetalBuffer* buffer = mtl.bo->getBuffer();
+    id<MTLBuffer> mtlBuffer = buffer->getGpuBufferForDraw();
+    if (mtlBuffer && mtlBuffer.storageMode == MTLStorageModeManaged) {
+        [mtlBuffer didModifyRange:NSMakeRange(mtl.offset, mtl.size)];
+    }
+#endif
+    // Shared memory on UMA systems is coherent; no explicit synchronization is required.
+}
+
+void MetalMemoryMappedBuffer::copy(MetalDriver& mtld, size_t offset, BufferDescriptor&& data) const {
+    assert_invariant(any(access & MapBufferAccessFlags::WRITE_BIT));
+    assert_invariant(offset + data.size <= mtl.size);
+    assert_invariant(mtl.vaddr);
+
+    memcpy(static_cast<char*>(mtl.vaddr) + offset, data.buffer, data.size);
+
+    mtld.scheduleDestroy(std::move(data));
+}
+
 } // namespace backend
 } // namespace filament

filament/backend/src/metal/MetalShaderCompiler.mm

@@ -111,9 +111,22 @@ bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
         id<MTLLibrary> library = nil;
         switch (program.getShaderLanguage()) {
             case ShaderLanguage::MSL: {
-                // By default, Metal uses the most recent language version.
                 MTLCompileOptions* options = [MTLCompileOptions new];
                 options.fastMathEnabled = YES;
+                // The docs for MTLCompileOptions say that Metal uses the most
+                // recent language version if it is not specified, but this is
+                // not always true: it is possible for the environment to be set
+                // up in a way where an older version is used by default. So we
+                // explicitly set the minimum version we require.
+                #if defined(FILAMENT_IOS)
+                    options.languageVersion = MTLLanguageVersion2_0;
+                #else
+                    if (@available(macOS 11.0, *)) {
+                        options.languageVersion = MTLLanguageVersion2_3;
+                    } else {
+                        options.languageVersion = MTLLanguageVersion2_2;
+                    }
+                #endif
 
                 assert_invariant(source[source.size() - 1] == '\0');
                 // the shader string is null terminated and the length includes the null character

filament/backend/src/noop/NoopDriver.cpp

@@ -291,8 +291,6 @@ void NoopDriver::updateBufferObject(Handle<HwBufferObject> ibh, BufferDescriptor
     scheduleDestroy(std::move(p));
 }
 
-void NoopDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh, BufferObjectStreamDescriptor&& streams) { }
-
 void NoopDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> ibh, BufferDescriptor&& p,
         uint32_t byteOffset) {
     scheduleDestroy(std::move(p));

filament/backend/src/opengl/OpenGLContext.h

@@ -735,7 +735,7 @@ void OpenGLContext::bindVertexArray(RenderPrimitive const* p) noexcept {
         // - the nameVersion is out of date *and* we're on the protected context, in this case:
         //      - the name must be stale from a previous use of this context because we always
         //        destroy the protected context when we're done with it.
-        bool const recreateVaoName = p != &mDefaultVAO &&
+        bool const recreateVaoName = vao != &mDefaultVAO &&
                 ((vao->vao[contextIndex] == 0) ||
                         (vao->nameVersion != state.age && contextIndex == 1));
         if (UTILS_UNLIKELY(recreateVaoName)) {

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -39,7 +39,6 @@
 #include <backend/Platform.h>
 #include <backend/Program.h>
 #include <backend/TargetBufferInfo.h>
-#include <backend/BufferObjectStreamDescriptor.h>
 
 #include "private/backend/CommandStream.h"
 #include "private/backend/Dispatcher.h"
@@ -2837,38 +2836,6 @@ void OpenGLDriver::updateIndexBuffer(
     CHECK_GL_ERROR()
 }
 
-void OpenGLDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh, BufferObjectStreamDescriptor&& streams) {
-    DEBUG_MARKER()
-
-    GLBufferObject const* bo = handle_cast<GLBufferObject*>(boh);
-    mStreamUniformDescriptors[bo->gl.id] = std::move(streams);
-}
-
-
-// specialization for mat3f (which has a different alignment, see std140 layout rules)
-static void copyMat3f(void* addr, size_t const offset, const mat3f& v) noexcept {
-    struct mat43 {
-        float v[3][4];
-    };
-
-    addr = static_cast<char*>(addr) + offset;
-    mat43& temp = *static_cast<mat43*>(addr);
-
-    temp.v[0][0] = v[0][0];
-    temp.v[0][1] = v[0][1];
-    temp.v[0][2] = v[0][2];
-
-    temp.v[1][0] = v[1][0];
-    temp.v[1][1] = v[1][1];
-    temp.v[1][2] = v[1][2];
-
-    temp.v[2][0] = v[2][0];
-    temp.v[2][1] = v[2][1];
-    temp.v[2][2] = v[2][2];
-
-    // don't store anything in temp.v[][3] because there could be uniforms packed there
-}
-
 void OpenGLDriver::updateBufferObject(
         Handle<HwBufferObject> boh, BufferDescriptor&& bd, uint32_t const byteOffset) {
     DEBUG_MARKER()
@@ -2882,19 +2849,6 @@ void OpenGLDriver::updateBufferObject(
         gl.bindVertexArray(nullptr);
     }
 
-    if (UTILS_UNLIKELY(!mStreamUniformDescriptors.empty())) {
-        auto const streamDescriptors = mStreamUniformDescriptors.find(bo->gl.id);
-        if (streamDescriptors != mStreamUniformDescriptors.end()) {
-            for (auto const& [offset, stream, associationType] : streamDescriptors->second.mStreams) {
-                if (associationType == BufferObjectStreamAssociationType::TRANSFORM_MATRIX) {
-                    auto transform = getStreamTransformMatrix(stream);
-                    copyMat3f(bd.buffer, offset, transform);
-                }
-            }
-            mStreamUniformDescriptors.erase(streamDescriptors);
-        }
-    }
-
     if (UTILS_UNLIKELY(bo->bindingType == BufferObjectBinding::UNIFORM && gl.isES2())) {
         assert_invariant(bo->gl.buffer);
         memcpy(static_cast<uint8_t*>(bo->gl.buffer) + byteOffset, bd.buffer, bd.size);

filament/backend/src/opengl/OpenGLDriver.h

@@ -36,7 +36,6 @@
 #include <backend/Platform.h>
 #include <backend/Program.h>
 #include <backend/TargetBufferInfo.h>
-#include <backend/BufferObjectStreamDescriptor.h>
 
 #include "private/backend/Driver.h"
 #include "private/backend/HandleAllocator.h"
@@ -394,8 +393,6 @@ private:
     // the must be accessed from the user thread only
     std::vector<GLStream*> mStreamsWithPendingAcquiredImage;
 
-    std::unordered_map<GLuint, BufferObjectStreamDescriptor> mStreamUniformDescriptors;
-
     void attachStream(GLTexture* t, GLStream* stream);
     void detachStream(GLTexture* t) noexcept;
     void replaceStream(GLTexture* t, GLStream* stream) noexcept;

filament/backend/src/opengl/gl_headers.h

@@ -37,7 +37,7 @@
  */
 
 
-#if defined(__ANDROID__) || defined(FILAMENT_USE_EXTERNAL_GLES3) || defined(__EMSCRIPTEN__)
+#if defined(__ANDROID__) || defined(FILAMENT_USE_EXTERNAL_GLES3) || defined(__EMSCRIPTEN__) || defined(FILAMENT_SUPPORTS_EGL_ON_LINUX)
 
     #if defined(__EMSCRIPTEN__)
     #   include <GLES3/gl3.h>

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

@@ -28,14 +28,13 @@
 #include <private/backend/VirtualMachineEnv.h>
 
 #include "AndroidNativeWindow.h"
+#include "AndroidFrameCallback.h"
 #include "AndroidSwapChainHelper.h"
 #include "ExternalStreamManagerAndroid.h"
 
 #include <android/api-level.h>
 #include <android/native_window.h>
 #include <android/hardware_buffer.h>
-#include <android/choreographer.h>
-#include <android/looper.h>
 
 #include <utils/android/PerformanceHintManager.h>
 #include <utils/compiler.h>
@@ -99,6 +98,27 @@ using namespace glext;
 
 // ---------------------------------------------------------------------------------------------
 
+struct PlatformEGLAndroid::SwapChainEGLAndroid : public SwapChainEGL {
+    SwapChainEGLAndroid(PlatformEGLAndroid const& platform,
+            void* nativeWindow, uint64_t flags);
+    SwapChainEGLAndroid(PlatformEGLAndroid const& platform,
+            uint32_t width, uint32_t height, uint64_t flags);
+    void terminate(PlatformEGLAndroid& platform);
+    bool setPresentFrameId(uint64_t frameId) const noexcept;
+    uint64_t getFrameId(uint64_t frameId) const noexcept;
+    bool compositorTimingSupported = false;
+    bool frameTimestampsSupported = false;
+private:
+    AndroidSwapChainHelper mImpl{};
+};
+
+struct PlatformEGLAndroid::AndroidDetails {
+    AndroidProducerThrottling producerThrottling;
+    AndroidFrameCallback androidFrameCallback;
+};
+
+// ---------------------------------------------------------------------------------------------
+
 PlatformEGLAndroid::InitializeJvmForPerformanceManagerIfNeeded::InitializeJvmForPerformanceManagerIfNeeded() {
     // PerformanceHintManager() needs the calling thread to be a Java thread; so we need
     // to attach this thread to the JVM before we initialize PerformanceHintManager.
@@ -112,17 +132,20 @@ PlatformEGLAndroid::InitializeJvmForPerformanceManagerIfNeeded::InitializeJvmFor
 // ---------------------------------------------------------------------------------------------
 
 PlatformEGLAndroid::PlatformEGLAndroid() noexcept
-        : mExternalStreamManager(ExternalStreamManagerAndroid::create()) {
+        : mExternalStreamManager(ExternalStreamManagerAndroid::create()),
+          mAndroidDetails(*(new(std::nothrow) AndroidDetails{})) {
     mOSVersion = android_get_device_api_level();
     if (mOSVersion < 0) {
         mOSVersion = __ANDROID_API_FUTURE__;
     }
 }
 
-PlatformEGLAndroid::~PlatformEGLAndroid() noexcept = default;
+PlatformEGLAndroid::~PlatformEGLAndroid() noexcept {
+    delete &mAndroidDetails;
+}
 
 void PlatformEGLAndroid::terminate() noexcept {
-    mAndroidFrameCallback.terminate();
+    mAndroidDetails.androidFrameCallback.terminate();
     ExternalStreamManagerAndroid::destroy(&mExternalStreamManager);
     PlatformEGL::terminate();
 }
@@ -155,20 +178,30 @@ void PlatformEGLAndroid::beginFrame(
         int64_t const monotonic_clock_ns,
         int64_t refreshIntervalNs,
         uint32_t const frameId) noexcept {
-    // associate the user frameid with the system frame id
+    // if frameId is 0, it means we're not associated to a particular frame, which is the case
-    setPresentFrameId(mCurrentDrawSwapChain, frameId);
+    // for standalone views. And in this case we skip the performance hint (since we wouldn't get
+    // the right timing anyway as well as the frame info timing collections)
 
-    if (mPerformanceHintSession.isValid()) {
+    // associate the user frameid with the system frame id.
+    if (frameId && mCurrentDrawSwapChain) {
+        // mCurrentDrawSwapChain could be null if we're called from renderStandaloneView
+        setPresentFrameId(mCurrentDrawSwapChain, frameId);
+    }
+
+    if (frameId && mPerformanceHintSession.isValid()) {
         if (refreshIntervalNs <= 0) {
             // we're not provided with a target time, use the display period, if everything fails,
             // assume 16.67ms
             refreshIntervalNs = 16'666'667;
 
-            CompositorTiming compositorTiming{};
+            if (mCurrentDrawSwapChain) {
-            bool const hasCompositorTiming =
+                // mCurrentDrawSwapChain could be null if we're called from renderStandaloneView
-                    queryCompositorTiming(mCurrentDrawSwapChain, &compositorTiming);
+                CompositorTiming compositorTiming{};
-            if (hasCompositorTiming && compositorTiming.compositeInterval > 0) {
+                bool const hasCompositorTiming =
-                refreshIntervalNs = compositorTiming.compositeInterval;
+                        queryCompositorTiming(mCurrentDrawSwapChain, &compositorTiming);
+                if (hasCompositorTiming && compositorTiming.compositeInterval > 0) {
+                    refreshIntervalNs = compositorTiming.compositeInterval;
+                }
             }
         }
         mStartTimeOfActualWork = clock::time_point(std::chrono::nanoseconds(monotonic_clock_ns));
@@ -207,9 +240,9 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
                     "eglGetNativeClientBufferANDROID"));
 
     if (ext.egl.ANDROID_presentation_time) {
-        eglGetNativeClientBufferANDROID =
+        eglPresentationTimeANDROID =
-                PFNEGLGETNATIVECLIENTBUFFERANDROIDPROC(eglGetProcAddress(
+                PFNEGLPRESENTATIONTIMEANDROIDPROC(eglGetProcAddress(
-                        "eglGetNativeClientBufferANDROID"));
+                        "eglPresentationTimeANDROID"));
     }
 
     if (ext.egl.ANDROID_get_frame_timestamps) {
@@ -240,7 +273,7 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
 
     mAssertNativeWindowIsValid = driverConfig.assertNativeWindowIsValid;
 
-    mAndroidFrameCallback.init();
+    mAndroidDetails.androidFrameCallback.init();
 
     return driver;
 }
@@ -264,15 +297,25 @@ bool PlatformEGLAndroid::queryCompositorTiming(SwapChain const* swapchain,
     }
 
     AndroidFrameCallback::Timeline const preferredTimeline{
-            mAndroidFrameCallback.getPreferredTimeline() };
+            mAndroidDetails.androidFrameCallback.getPreferredTimeline() };
     outCompositorTiming->frameTime = preferredTimeline.frameTime;
     outCompositorTiming->expectedPresentTime = preferredTimeline.expectedPresentTime;
     outCompositorTiming->frameTimelineDeadline = preferredTimeline.frameTimelineDeadline;
+    outCompositorTiming->compositeDeadline = CompositorTiming::INVALID;
+    outCompositorTiming->compositeInterval = CompositorTiming::INVALID;
+    outCompositorTiming->compositeToPresentLatency = CompositorTiming::INVALID;
+
+    // From this point on, we always return "success" because some timings were returned.
+
+    if (!static_cast<SwapChainEGLAndroid const *>(swapchain)->compositorTimingSupported) {
+        // if this surface doesn't support it, don't attempt to query the values.
+        return true;
+    }
 
     if (UTILS_LIKELY(ext.egl.ANDROID_get_frame_timestamps)) {
         EGLSurface const sur = static_cast<SwapChainEGL const *>(swapchain)->sur;
         if (sur == EGL_NO_SURFACE) {
-            return false;
+            return true;
         }
 
         std::array<EGLnsecsANDROID, 3> values;
@@ -283,26 +326,16 @@ bool PlatformEGLAndroid::queryCompositorTiming(SwapChain const* swapchain,
         };
         EGLBoolean const success = eglGetCompositorTimingANDROID(getEglDisplay(), sur,
                 names.size(), names.data(), values.data());
-        if (!success) {
+        if (UTILS_UNLIKELY(!success)) {
-            return false;
+            // reset current error to EGL_SUCCESS
+            eglGetError();
+        } else {
+            outCompositorTiming->compositeDeadline = values[0];
+            outCompositorTiming->compositeInterval = values[1];
+            outCompositorTiming->compositeToPresentLatency = values[2];
         }
-        outCompositorTiming->compositeDeadline = values[0];
-        outCompositorTiming->compositeInterval = values[1];
-        outCompositorTiming->compositeToPresentLatency = values[2];
-        return true;
     }
-
+    return true;
-    // fallback to private APIs
-    auto const anw = static_cast<SwapChainEGL const *>(swapchain)->nativeWindow;
-    int const status = NativeWindow::getCompositorTiming(anw,
-            &outCompositorTiming->compositeDeadline,
-            &outCompositorTiming->compositeInterval,
-            &outCompositorTiming->compositeToPresentLatency);
-    if (status == 0) {
-        return true;
-    }
-
-    return PlatformEGL::queryCompositorTiming(swapchain, outCompositorTiming);
 }
 
 bool PlatformEGLAndroid::setPresentFrameId(SwapChain const* swapchain,
@@ -327,6 +360,10 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
         return false;
     }
 
+    if (!static_cast<SwapChainEGLAndroid const *>(swapchain)->frameTimestampsSupported) {
+        return false;
+    }
+
     if (UTILS_LIKELY(ext.egl.ANDROID_get_frame_timestamps)) {
         EGLSurface const sur = sc->sur;
         if (sur == EGL_NO_SURFACE) {
@@ -347,7 +384,9 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
         };
         EGLBoolean const success = eglGetFrameTimestampsANDROID(getEglDisplay(), sur, hwFrameId,
                 names.size(), names.data(), values.data());
-        if (!success) {
+        if (UTILS_UNLIKELY(!success)) {
+            // reset current error to EGL_SUCCESS
+            eglGetError();
             return false;
         }
         outFrameTimestamps->requestedPresentTime = values[0];
@@ -361,28 +400,44 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
         outFrameTimestamps->releaseTime = values[8];
         return true;
     }
-
-    // fallback to private APIs
-    auto const anw = sc->nativeWindow;
-    int const status = NativeWindow::getFrameTimestamps(anw, hwFrameId,
-            &outFrameTimestamps->requestedPresentTime,
-            &outFrameTimestamps->acquireTime,
-            &outFrameTimestamps->latchTime,
-            &outFrameTimestamps->firstCompositionStartTime,
-            &outFrameTimestamps->lastCompositionStartTime,
-            &outFrameTimestamps->gpuCompositionDoneTime,
-            &outFrameTimestamps->displayPresentTime,
-            &outFrameTimestamps->dequeueReadyTime,
-            &outFrameTimestamps->releaseTime);
-    if (status == 0) {
-        return true;
-    }
-
     return PlatformEGL::queryFrameTimestamps(swapchain, frameId, outFrameTimestamps);
 }
 
 Platform::SwapChain* PlatformEGLAndroid::createSwapChain(void* nativeWindow, uint64_t const flags) {
     auto* const sc = new(std::nothrow) SwapChainEGLAndroid(*this, nativeWindow, flags);
+    if (UTILS_LIKELY(ext.egl.ANDROID_get_frame_timestamps)) {
+        EGLDisplay const dpy = getEglDisplay();
+        sc->compositorTimingSupported =
+                eglGetCompositorTimingSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITE_DEADLINE_ANDROID) &&
+                eglGetCompositorTimingSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITE_INTERVAL_ANDROID) &&
+                eglGetCompositorTimingSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITE_TO_PRESENT_LATENCY_ANDROID);
+        sc->frameTimestampsSupported =
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_REQUESTED_PRESENT_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_RENDERING_COMPLETE_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITION_LATCH_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_FIRST_COMPOSITION_START_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_LAST_COMPOSITION_START_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_FIRST_COMPOSITION_GPU_FINISHED_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_DISPLAY_PRESENT_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_DEQUEUE_READY_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_READS_DONE_TIME_ANDROID);
+    }
+    // This is expected to be a low frequency log, only turned on in debug builds
+    DLOG(INFO) << "anw: " << nativeWindow
+            << ", compositorTimingSupported=" << sc->compositorTimingSupported
+            << ", frameTimestampsSupported=" << sc->frameTimestampsSupported;
     return sc;
 }
 
@@ -395,6 +450,10 @@ Platform::SwapChain* PlatformEGLAndroid::createSwapChain(
 void PlatformEGLAndroid::destroySwapChain(SwapChain* swapChain) noexcept {
     if (swapChain) {
         SwapChainEGLAndroid* const sc = static_cast<SwapChainEGLAndroid*>(swapChain);
+        if (mCurrentDrawSwapChain == sc) {
+            // don't keep a dangling pointer around
+            mCurrentDrawSwapChain = nullptr;
+        }
         sc->terminate(*this);
         delete sc;
     }
@@ -675,12 +734,12 @@ AcquiredImage PlatformEGLAndroid::transformAcquiredImage(AcquiredImage const sou
 
 
 bool PlatformEGLAndroid::isProducerThrottlingControlSupported() const {
-    return mProducerThrottling.isSupported();
+    return mAndroidDetails.producerThrottling.isSupported();
 }
 
 int32_t PlatformEGLAndroid::setProducerThrottlingEnabled(
     EGLNativeWindowType const nativeWindow, bool const enabled) const {
-    return mProducerThrottling.setProducerThrottlingEnabled(nativeWindow, enabled);
+    return mAndroidDetails.producerThrottling.setProducerThrottlingEnabled(nativeWindow, enabled);
 }
 
 // ---------------------------------------------------------------------------------------------
@@ -704,8 +763,6 @@ PlatformEGLAndroid::SwapChainEGLAndroid::SwapChainEGLAndroid(PlatformEGLAndroid
             // we ignore the result, it doesn't matter much if it fails
             eglSurfaceAttrib(platform.getEglDisplay(), sur, EGL_TIMESTAMPS_ANDROID, EGL_TRUE);
         }
-    } else {
-        NativeWindow::enableFrameTimestamps(EGLNativeWindowType(nativeWindow), true);
     }
 }
 

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

@@ -16,7 +16,7 @@
 
 #include <backend/platforms/PlatformEGLHeadless.h>
 
-#include "opengl/GLUtils.h"
+#include <bluegl/BlueGL.h>
 
 #include <EGL/egl.h>
 #include <EGL/eglext.h>
@@ -35,16 +35,31 @@ PlatformEGLHeadless::PlatformEGLHeadless() noexcept
 }
 
 bool PlatformEGLHeadless::isOpenGL() const noexcept {
-    return  true;
+#if defined(BACKEND_OPENGL_VERSION_GL)
+    return true;
+#else
+    return false;
+#endif  // defined(BACKEND_OPENGL_VERSION_GL)
 }
 
 backend::Driver* PlatformEGLHeadless::createDriver(void* sharedContext,
         const Platform::DriverConfig& driverConfig) {
-    EGLBoolean bindAPI = eglBindAPI(EGL_OPENGL_API);
+    auto bindApiHelper = [](EGLenum api, const char* errorString) -> bool {
-    if (UTILS_UNLIKELY(!bindAPI)) {
+        EGLBoolean bindAPI = eglBindAPI(api);
-        LOG(ERROR) << "eglBindAPI EGL_OPENGL_API failed";
+        if (UTILS_UNLIKELY(bindAPI == EGL_FALSE || bindAPI == EGL_BAD_PARAMETER)) {
+            logEglError(errorString);
+            eglReleaseThread();
+            return false;
+        };
+        return true;
+    };
+
+    EGLenum api = isOpenGL() ? EGL_OPENGL_API : EGL_OPENGL_ES_API;
+    const char* apiString = isOpenGL() ? "eglBindAPI EGL_OPENGL_API" : "eglBindAPI EGL_OPENGL_ES_API";
+    if (!bindApiHelper(api, apiString)) {
         return nullptr;
     }
+
     int bindBlueGL = bluegl::bind();
     if (UTILS_UNLIKELY(bindBlueGL != 0)) {
         LOG(ERROR) << "bluegl bind failed";

filament/backend/src/ostream.cpp

@@ -15,7 +15,6 @@
  */
 
 #include <backend/BufferDescriptor.h>
-#include <backend/BufferObjectStreamDescriptor.h>
 #include <backend/DescriptorSetOffsetArray.h>
 #include <backend/DriverEnums.h>
 #include <backend/PipelineState.h>
@@ -434,10 +433,6 @@ io::ostream& operator<<(io::ostream& out, BufferDescriptor const& b) {
     << ", user=" << b.getUser() << " }";
 }
 
-io::ostream& operator<<(io::ostream& out, const BufferObjectStreamDescriptor& b) {
-    return out << "BufferObjectStreamDescriptor{ streams(" << b.mStreams.size() << ")=... }";
-}
-
 io::ostream& operator<<(io::ostream& out, PixelBufferDescriptor const& b) {
     BufferDescriptor const& base = static_cast<BufferDescriptor const&>(b);
     return out << "PixelBufferDescriptor{ " << base

filament/backend/src/vulkan/VulkanBufferProxy.cpp

@@ -62,8 +62,9 @@ void VulkanBufferProxy::loadFromCpu(VulkanCommandBuffer& commands, const void* c
     // a memcpy if its available.
     bool const isStaticOrShared =
             any(mUsage & (BufferUsage::STATIC | BufferUsage::SHARED_WRITE_BIT));
-    bool const useMemcpy = (isUniformAndAvailable || isStaticOrShared) && isMemcopyable &&
+    bool const useMemcpy =
-                           mStagingBufferBypassEnabled;
+            ((isUniformAndAvailable && mStagingBufferBypassEnabled) || isStaticOrShared) &&
+            isMemcopyable;
     if (useMemcpy) {
         char* dest = static_cast<char*>(mBuffer->getGpuBuffer()->allocationInfo.pMappedData) +
                      byteOffset;

filament/backend/src/vulkan/VulkanDescriptorSetCache.cpp

@@ -19,6 +19,7 @@
 #include "VulkanCommands.h"
 #include "VulkanHandles.h"
 #include "VulkanConstants.h"
+#include "utils/compiler.h"
 
 #include <utils/FixedCapacityVector.h>
 #include <utils/Panic.h>
@@ -31,6 +32,18 @@ namespace filament::backend {
 
 namespace {
 
+using Bitmask = fvkutils::UniformBufferBitmask;
+static_assert(sizeof(Bitmask) * 8 == fvkutils::MAX_DESCRIPTOR_SET_BITMASK_BITS);
+
+Bitmask foldBitsInHalf(Bitmask bitset) {
+    Bitmask outBitset;
+    bitset.forEachSetBit([&](size_t index) {
+        constexpr size_t BITMASK_LOWER_BITS_LEN = sizeof(outBitset) * 4;
+        outBitset.set(index % BITMASK_LOWER_BITS_LEN);
+    });
+    return outBitset;
+}
+
 using DescriptorSetLayoutArray = VulkanDescriptorSetCache::DescriptorSetLayoutArray;
 using DescriptorCount = VulkanDescriptorSetCache::DescriptorCount;
 
@@ -304,8 +317,7 @@ void VulkanDescriptorSetCache::commit(VulkanCommandBuffer* commands,
         // This code actually binds the descriptor sets.
         auto set = updateSets[index];
         VkCommandBuffer const cmdbuffer = commands->buffer();
-        VkDescriptorSet vkset = useExternalSamplers[index] ? set->getExternalSamplerVkSet() :
+        VkDescriptorSet const vkset = set->getVkSet();
-            set->getVkSet();
         commands->acquire(set);
         vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, index,
                 1, &vkset, set->uniqueDynamicUboCount, set->getOffsets()->data());
@@ -341,17 +353,39 @@ void VulkanDescriptorSetCache::updateBuffer(fvkmemory::resource_ptr<VulkanDescri
         .pBufferInfo = &info,
     };
     vkUpdateDescriptorSets(mDevice, 1, &descriptorWrite, 0, nullptr);
-
-    if (auto externalSamplerSet = set->getExternalSamplerVkSet();
-            externalSamplerSet != VK_NULL_HANDLE) {
-        descriptorWrite.dstSet = externalSamplerSet;
-        vkUpdateDescriptorSets(mDevice, 1, &descriptorWrite, 0, nullptr);
-    }
     set->acquire(bufferObject);
 }
 
-void VulkanDescriptorSetCache::updateSamplerImpl(VkDescriptorSet vkset, uint8_t binding,
+void VulkanDescriptorSetCache::updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
-        fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept {
+        uint8_t binding, fvkmemory::resource_ptr<VulkanTexture> texture,
+        VkSampler sampler, VkDescriptorSetLayout externalSamplerLayout) noexcept {
+
+    // We have to update a bound set for two use cases
+    //   - streaming API (a changing feed of AHardwareBuffer)
+    //   - external samplers - potential changing of dataspace per-frame
+    if (UTILS_UNLIKELY(set->isBound())) {
+        auto layout = set->getLayout();
+        // Build a new descriptor set from the new layout
+        auto genLayout = externalSamplerLayout != VK_NULL_HANDLE ?
+                externalSamplerLayout : layout->getVkLayout();
+        VkDescriptorSet const newSet = getVkSet(layout->count, genLayout);
+        Bitmask const ubo = layout->bitmask.ubo | layout->bitmask.dynamicUbo;
+        Bitmask samplers = layout->bitmask.sampler;
+        samplers.unset(binding);
+
+        // Each bitmask denotes a binding index, and separated into two stages - vertex and buffer
+        // We fold the two stages into just the lower half of the bits to denote a combined set of
+        // bindings.
+        Bitmask const copyBindings = foldBitsInHalf(ubo | samplers);
+        VkDescriptorSet const srcSet = set->getVkSet();
+        copySet(srcSet, newSet, copyBindings);
+        set->addNewSet(newSet,
+                [this, layoutCount = layout->count, genLayout, newSet](VulkanDescriptorSet*) {
+                    this->manualRecycle(layoutCount, genLayout, newSet);
+                });
+    }
+
+    VkDescriptorSet const vkset = set->getVkSet();
     VkImageSubresourceRange range = texture->getPrimaryViewRange();
     VkImageViewType const expectedType = texture->getViewType();
     if (any(texture->usage & TextureUsage::DEPTH_ATTACHMENT) &&
@@ -376,19 +410,6 @@ void VulkanDescriptorSetCache::updateSamplerImpl(VkDescriptorSet vkset, uint8_t
         .pImageInfo = &info,
     };
     vkUpdateDescriptorSets(mDevice, 1, &descriptorWrite, 0, nullptr);
-}
-
-void VulkanDescriptorSetCache::updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
-        uint8_t binding, fvkmemory::resource_ptr<VulkanTexture> texture,
-        VkSampler sampler) noexcept {
-    updateSamplerImpl(set->getVkSet(), binding, texture, sampler);
-    set->acquire(texture);
-}
-
-void VulkanDescriptorSetCache::updateSamplerForExternalSamplerSet(
-        fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
-        fvkmemory::resource_ptr<VulkanTexture> texture) noexcept {
-    updateSamplerImpl(set->getExternalSamplerVkSet(), binding, texture, VK_NULL_HANDLE);
     set->acquire(texture);
 }
 
@@ -405,7 +426,7 @@ fvkmemory::resource_ptr<VulkanDescriptorSet> VulkanDescriptorSetCache::createSet
     auto const& count = layout->count;
     auto const vklayout = layout->getVkLayout();
     auto set = fvkmemory::resource_ptr<VulkanDescriptorSet>::make(
-            mResourceManager, handle, layout->bitmask.dynamicUbo, layout->count.dynamicUbo,
+            mResourceManager, handle, layout,
             [vkSet, count, vklayout, this](
                     VulkanDescriptorSet*) { this->manualRecycle(count, vklayout, vkSet); },
             vkSet);
@@ -428,4 +449,22 @@ void VulkanDescriptorSetCache::manualRecycle(VulkanDescriptorSetLayout::Count co
 
 void VulkanDescriptorSetCache::gc() { mStashedSets = {}; }
 
+void VulkanDescriptorSetCache::copySet(VkDescriptorSet srcSet, VkDescriptorSet dstSet,
+        fvkutils::SamplerBitmask bindings) const {
+    // TODO: fix the size for better memory management
+    std::vector<VkCopyDescriptorSet> copies;
+    bindings.forEachSetBit([&](size_t index) {
+        copies.push_back({
+            .sType = VK_STRUCTURE_TYPE_COPY_DESCRIPTOR_SET,
+            .srcSet = srcSet,
+            .srcBinding = (uint32_t) index,
+            .dstSet = dstSet,
+            .dstBinding = (uint32_t) index,
+            .descriptorCount = 1,
+        });
+    });
+    vkUpdateDescriptorSets(mDevice, 0, nullptr, copies.size(), copies.data());
+}
+
+
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanDescriptorSetCache.h

@@ -55,7 +55,8 @@ public:
             VkDeviceSize size) noexcept;
 
     void updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
-            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept;
+            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler,
+            VkDescriptorSetLayout externalSamplerLayout = VK_NULL_HANDLE) noexcept;
 
     void updateSamplerForExternalSamplerSet(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
             fvkmemory::resource_ptr<VulkanTexture> texture) noexcept;
@@ -90,8 +91,8 @@ public:
     void resetCachedState() noexcept { mLastBoundInfo = {}; }
 
 private:
-    void updateSamplerImpl(VkDescriptorSet set, uint8_t binding,
+    void copySet(VkDescriptorSet srcSet, VkDescriptorSet destSet,
-            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept;
+            fvkutils::SamplerBitmask copyBindings) const;
 
     class DescriptorInfinitePool;
 

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -31,6 +31,7 @@
 #include "vulkan/memory/ResourcePointer.h"
 #include "vulkan/utils/Conversion.h"
 #include "vulkan/utils/Definitions.h"
+#include "vulkan/utils/Image.h"
 
 #include <backend/DriverEnums.h>
 #include <backend/platforms/VulkanPlatform.h>
@@ -244,6 +245,7 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext& context,
       mQueryManager(mPlatform->getDevice()),
       mExternalImageManager(platform, &mSamplerCache, &mYcbcrConversionCache, &mDescriptorSetCache,
               &mDescriptorSetLayoutCache),
+      mStreamedImageManager(&mExternalImageManager),
       mIsSRGBSwapChainSupported(mPlatform->getCustomization().isSRGBSwapChainSupported),
       mIsMSAASwapChainSupported(false), // TODO: support MSAA swapchain
       mStereoscopicType(driverConfig.stereoscopicType) {
@@ -420,12 +422,14 @@ void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
         int64_t refreshIntervalNs, uint32_t frameId) {
     FVK_PROFILE_MARKER(PROFILE_NAME_BEGINFRAME);
 
-    if (mCurrentSwapChain) { // This should be guaranteed
+    // if frameId is 0, it means we're not associated to a particular frame, which is the case
+    // for standalone views. And in this case we skip the frame info timing collection.
+    if (frameId && mCurrentSwapChain) { // This should be guaranteed
         mPlatform->setPresentFrameId(mCurrentSwapChain->swapChain, frameId);
     }
 
     // Check if any command have finished and reset all its used resources. The resources
-    // wont be destroyed but their reference count will decreased if the command is already
+    // won't be destroyed but their reference count will be decreased if the command is already
     // completed.
     //
     // This will let us check if any VulkanBuffer is currently in flight or not.
@@ -479,6 +483,9 @@ void VulkanDriver::updateDescriptorSetTexture(
     if (UTILS_UNLIKELY(mExternalImageManager.isExternallySampledTexture(texture))) {
         mExternalImageManager.bindExternallySampledTexture(set, binding, texture, params);
         mAppState.hasBoundExternalImages = true;
+    } else if (bool(texture->getStream())) {
+        mStreamedImageManager.bindStreamedTexture(set, binding, texture, params);
+        mAppState.hasBoundExternalImages = true;
     } else {
         VulkanSamplerCache::Params cacheParams = {
             .sampler = params,
@@ -486,6 +493,7 @@ void VulkanDriver::updateDescriptorSetTexture(
         VkSampler const vksampler = mSamplerCache.getSampler(cacheParams);
         mDescriptorSetCache.updateSampler(set, binding, texture, vksampler);
         mExternalImageManager.clearTextureBinding(set, binding);
+        mStreamedImageManager.unbindStreamedTexture(set, binding);
     }
 }
 
@@ -497,10 +505,7 @@ void VulkanDriver::copyToMemoryMappedBuffer(MemoryMappedBufferHandle mmbh, size_
     assert_invariant(any(mmb->access & MapBufferAccessFlags::WRITE_BIT));
     assert_invariant(offset + data.size <= mmb->size);
 
-    // TODO: this isa zero-effort implementation of copyToMemoryMappedBuffer(), where we just
+    // On UMA systems, this is a memcpy into the memory-mapped buffer, otherwise use a staging buffer.
-    //       call updateBufferObject(). This could be a fallback implementation for when
-    //       shared memory is not available.
-    //       On UMA systems, this should just be a memcpy into the memory-mapped buffer.
     updateBufferObject(mmb->boh, std::move(data), mmb->offset + offset);
 }
 
@@ -921,14 +926,6 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
         }
     }
 
-#if defined(__ANDROID__)
-    // on Android, disable producer throttling
-    if (mProducerThrottling.isSupported()) {
-        mProducerThrottling.setProducerThrottlingEnabled(
-                static_cast<ANativeWindow*>(nativeWindow), false);
-    }
-#endif
-
     auto swapChain = resource_ptr<VulkanSwapChain>::make(&mResourceManager, sch, mPlatform,
             mContext, &mResourceManager, mAllocator, &mCommands, mStagePool, nativeWindow, flags);
     swapChain.inc();
@@ -1116,6 +1113,12 @@ void VulkanDriver::destroySwapChain(Handle<HwSwapChain> sch) {
 }
 
 void VulkanDriver::destroyStream(Handle<HwStream> sh) {
+    FVK_SYSTRACE_SCOPE();
+    if (!sh) {
+        return;
+    }
+    auto stream = resource_ptr<VulkanStream>::cast(&mResourceManager, sh);
+    stream.dec();
 }
 
 void VulkanDriver::destroySync(Handle<HwSync> sh) {
@@ -1153,18 +1156,38 @@ void VulkanDriver::destroyDescriptorSet(Handle<HwDescriptorSet> dsh) {
 }
 
 Handle<HwStream> VulkanDriver::createStreamNative(void* nativeStream, utils::ImmutableCString tag) {
+    FVK_LOGW << "createStreamNative not supported in Vulkan.";
     return {};
 }
 
 Handle<HwStream> VulkanDriver::createStreamAcquired(utils::ImmutableCString tag) {
-    return {};
+    // @TODO This is still not thread-safe. We might have to revisit this question.
+    FVK_SYSTRACE_SCOPE();
+    auto handle = mResourceManager.allocHandle<VulkanStream>();
+    auto stream = resource_ptr<VulkanStream>::make(&mResourceManager, handle);
+    stream.inc();
+    return handle;
 }
 
 void VulkanDriver::setAcquiredImage(Handle<HwStream> sh, void* image, const math::mat3f& transform,
         CallbackHandler* handler, StreamCallback cb, void* userData) {
+    FVK_SYSTRACE_SCOPE();
+    auto stream = resource_ptr<VulkanStream>::cast(&mResourceManager, sh);
+    assert_invariant(stream->streamType == StreamType::ACQUIRED);
+    assert_invariant(image != nullptr);
+    // This covers the case where setAcquiredImage is called back to back (no updateStreams in between)
+    if (stream->previousNeedsRelease()) {
+        scheduleRelease(stream->takePrevious());
+    }
+    stream->acquire({ image, cb, userData, handler});
+    mStreamsWithPendingAcquiredImage.push_back(stream);
 }
 
 void VulkanDriver::setStreamDimensions(Handle<HwStream> sh, uint32_t width, uint32_t height) {
+    FVK_SYSTRACE_SCOPE();
+    auto stream = resource_ptr<VulkanStream>::cast(&mResourceManager, sh);
+    stream->width = width;
+    stream->height = height;
 }
 
 int64_t VulkanDriver::getStreamTimestamp(Handle<HwStream> sh) {
@@ -1172,6 +1195,70 @@ int64_t VulkanDriver::getStreamTimestamp(Handle<HwStream> sh) {
 }
 
 void VulkanDriver::updateStreams(CommandStream* driver) {
+    FVK_SYSTRACE_SCOPE();
+    if (UTILS_UNLIKELY(!mStreamsWithPendingAcquiredImage.empty())) {
+        for (auto& stream: mStreamsWithPendingAcquiredImage) {
+            if (stream->previousNeedsRelease()) {
+                scheduleRelease(stream->takePrevious());
+            }
+
+            // This executes on the backend thread (updateStreams is synchonous which means it
+            // executes on the user thread) Note: stream is captured by copy which is fine, this is
+            // a copy of a resource_ptr<VulkanStream>. We only need it find the associated stream
+            // inside the mStreamedImageManager texture bindings
+            driver->queueCommand([this, stream, s = stream.get(),
+                                         image = stream->getAcquired().image]() {
+                auto texture = s->getTexture(image);
+                if (!texture) {
+                    auto externalImage = fvkutils::createExternalImageFromRaw(mPlatform, image, false);
+                    auto metadata = mPlatform->extractExternalImageMetadata(externalImage);
+                    auto imgData = mPlatform->createVkImageFromExternal(externalImage);
+
+                    assert_invariant(imgData.internal.valid() || imgData.external.valid());
+
+                    VkFormat vkformat = metadata.format;
+                    VkImage vkimage = VK_NULL_HANDLE;
+                    VkDeviceMemory memory = VK_NULL_HANDLE;
+                    if (imgData.internal.valid()) {
+                        metadata.externalFormat = 0;
+                        vkimage = imgData.internal.image;
+                        memory = imgData.internal.memory;
+                    } else {
+                        vkformat = VK_FORMAT_UNDEFINED;
+                        vkimage = imgData.external.image;
+                        memory = imgData.external.memory;
+                    }
+
+                    VkSamplerYcbcrConversion const conversion =
+                            mExternalImageManager.getVkSamplerYcbcrConversion(metadata);
+
+                    auto newTexture = resource_ptr<VulkanTexture>::construct(&mResourceManager,
+                            mContext, mPlatform->getDevice(), mAllocator, &mResourceManager,
+                            &mCommands, vkimage, memory, vkformat, conversion, metadata.samples,
+                            metadata.width, metadata.height, metadata.layers,
+                            metadata.filamentUsage, mStagePool);
+
+                    auto& commands = mCommands.get();
+                    // Unlike uploaded textures or swapchains, we need to explicit transition this
+                    // texture into the read layout.
+                    newTexture->transitionLayout(&commands, newTexture->getPrimaryViewRange(),
+                            VulkanLayout::FRAG_READ);
+
+                    if (imgData.external.valid()) {
+                        mExternalImageManager.addExternallySampledTexture(newTexture,
+                                externalImage);
+                        // Cache the AHB backed image. Acquires the image here.
+                        s->pushImage(image, newTexture);
+                    }
+
+                    texture = newTexture;
+                }
+
+                mStreamedImageManager.onStreamAcquireImage(texture, stream);
+            });
+        }
+        mStreamsWithPendingAcquiredImage.clear();
+    }
 }
 
 void VulkanDriver::destroyFence(Handle<HwFence> fh) {
@@ -1467,17 +1554,13 @@ void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor
     scheduleDestroy(std::move(p));
 }
 
-void VulkanDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh,
-        BufferObjectStreamDescriptor&& streams) {
-    // Noop
-}
-
 void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& bd,
         uint32_t byteOffset) {
     VulkanCommandBuffer& commands = mCommands.get();
 
     auto bo = resource_ptr<VulkanBufferObject>::cast(&mResourceManager, boh);
     commands.acquire(bo);
+
     bo->loadFromCpu(commands, bd.buffer, byteOffset, bd.size);
 
     scheduleDestroy(std::move(bd));
@@ -1545,6 +1628,12 @@ TimerQueryResult VulkanDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint
 }
 
 void VulkanDriver::setExternalStream(Handle<HwTexture> th, Handle<HwStream> sh) {
+    FVK_SYSTRACE_SCOPE();
+    auto t = resource_ptr<VulkanTexture>::cast(&mResourceManager, th);
+    assert_invariant(t);
+    auto s = resource_ptr<VulkanStream>::cast(&mResourceManager, sh);
+    assert_invariant(s);
+    t->setStream(s);
 }
 
 void VulkanDriver::generateMipmaps(Handle<HwTexture> th) {

filament/backend/src/vulkan/VulkanDriver.h

@@ -34,12 +34,14 @@
 #include "vulkan/VulkanDescriptorSetCache.h"
 #include "vulkan/VulkanDescriptorSetLayoutCache.h"
 #include "vulkan/VulkanExternalImageManager.h"
+#include "vulkan/VulkanStreamedImageManager.h"
 #include "vulkan/VulkanPipelineLayoutCache.h"
 #include "vulkan/memory/ResourceManager.h"
 #include "vulkan/memory/ResourcePointer.h"
 #include "vulkan/utils/Definitions.h"
 
 #include "backend/DriverEnums.h"
+#include "backend/BufferObjectStreamDescriptor.h"
 
 #include "DriverBase.h"
 #include "private/backend/Driver.h"
@@ -48,10 +50,6 @@
 #include <utils/Allocator.h>
 #include <utils/compiler.h>
 
-#if defined(__ANDROID__)
-#include "AndroidNativeWindow.h"
-#endif
-
 namespace filament::backend {
 
 class VulkanPlatform;
@@ -161,6 +159,8 @@ private:
     VulkanDescriptorSetCache mDescriptorSetCache;
     VulkanQueryManager mQueryManager;
     VulkanExternalImageManager mExternalImageManager;
+    VulkanStreamedImageManager mStreamedImageManager;
+
 
     // This maps a VulkanSwapchain to a native swapchain. VulkanSwapchain should have a copy of the
     // Platform::Swapchain pointer, but queryFrameTimestamps() and queryCompositorTiming() are
@@ -206,9 +206,10 @@ private:
     bool const mIsSRGBSwapChainSupported;
     bool const mIsMSAASwapChainSupported;
     backend::StereoscopicType const mStereoscopicType;
-#if defined(__ANDROID__)
+
-    AndroidProducerThrottling mProducerThrottling;
+    // setAcquiredImage is a DECL_DRIVER_API_SYNCHRONOUS_N which means we don't necessarily have the
-#endif
+    // data to process it at call time. So we store it and process it during updateStreams.
+    std::vector<resource_ptr<VulkanStream>> mStreamsWithPendingAcquiredImage;
 };
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanExternalImageManager.cpp

@@ -31,9 +31,6 @@ namespace filament::backend {
 
 namespace {
 
-using Bitmask = fvkutils::UniformBufferBitmask;
-static_assert(sizeof(Bitmask) * 8 == fvkutils::MAX_DESCRIPTOR_SET_BITMASK_BITS);
-
 template<typename T>
 void erasep(std::vector<T>& v, std::function<bool(T const&)> f) {
     auto newEnd = std::remove_if(v.begin(), v.end(), f);
@@ -50,31 +47,6 @@ ImageData& findImage(std::vector<ImageData>& images,
     return *itr;
 }
 
-void copySet(VkDevice device, VkDescriptorSet srcSet, VkDescriptorSet dstSet, Bitmask bindings) {
-    // TODO: fix the size for better memory management
-    std::vector<VkCopyDescriptorSet> copies;
-    bindings.forEachSetBit([&](size_t index) {
-        copies.push_back({
-            .sType = VK_STRUCTURE_TYPE_COPY_DESCRIPTOR_SET,
-            .srcSet = srcSet,
-            .srcBinding = (uint32_t) index,
-            .dstSet = dstSet,
-            .dstBinding = (uint32_t) index,
-            .descriptorCount = 1,
-        });
-    });
-    vkUpdateDescriptorSets(device, 0, nullptr, copies.size(), copies.data());
-}
-
-Bitmask foldBitsInHalf(Bitmask bitset) {
-    Bitmask outBitset;
-    bitset.forEachSetBit([&](size_t index) {
-        constexpr size_t BITMASK_LOWER_BITS_LEN = sizeof(outBitset) * 4;
-        outBitset.set(index % BITMASK_LOWER_BITS_LEN);
-    });
-    return outBitset;
-}
-
 }// namespace
 
 VulkanExternalImageManager::VulkanExternalImageManager(VulkanPlatform* platform,
@@ -141,6 +113,9 @@ void VulkanExternalImageManager::updateSetAndLayout(
             continue;
         }
         auto& imageData = findImage(mImages, bindingInfo.image);
+        // For non YUV images (some ext images are NOT ext FMT)
+        // getVkSamplerYcbcrConversion(metadata) will return NULL, and image->conversion will be
+        // null
         updateImage(&imageData);
 
         auto samplerParams = bindingInfo.samplerParams;
@@ -172,38 +147,14 @@ void VulkanExternalImageManager::updateSetAndLayout(
     std::for_each(samplerAndBindings.begin(), samplerAndBindings.end(),
             [&](auto const& b) { outSamplers.push_back(std::get<1>(b)); });
 
-    VkDescriptorSetLayout const oldLayout = layout->getExternalSamplerVkLayout();
     VkDescriptorSetLayout const newLayout = mDescriptorSetLayoutCache->getVkLayout(layout->bitmask,
             actualExternalSamplers, outSamplers);
-
+    layout->setExternalSamplerVkLayout(newLayout);
-    // Need to copy the set
-    VkDescriptorSet const oldSet = set->getExternalSamplerVkSet();
-    if (oldLayout != newLayout || oldSet == VK_NULL_HANDLE) {
-        // Build a new descriptor set from the new layout
-        VkDescriptorSet const newSet = mDescriptorSetCache->getVkSet(layout->count, newLayout);
-        auto const ubo = layout->bitmask.ubo | layout->bitmask.dynamicUbo;
-        auto const samplers = layout->bitmask.sampler & (~actualExternalSamplers);
-
-        // Each bitmask denotes a binding index, and separated into two stages - vertex and buffer
-        // We fold the two stages into just the lower half of the bits to denote a combined set of
-        // bindings.
-        Bitmask const copyBindings = foldBitsInHalf(ubo | samplers);
-        VkDescriptorSet const srcSet = oldSet != VK_NULL_HANDLE ? oldSet : set->getVkSet();
-        copySet(mPlatform->getDevice(), srcSet, newSet, copyBindings);
-
-        set->setExternalSamplerVkSet(newSet,
-                [&descriptorSetCache = mDescriptorSetCache, layoutCount = layout->count, newLayout,
-                        newSet](VulkanDescriptorSet*) {
-                    descriptorSetCache->manualRecycle(layoutCount, newLayout, newSet);
-                });
-        if (oldLayout != newLayout) {
-            layout->setExternalSamplerVkLayout(newLayout);
-        }
-    }
-
     // Update the external samplers in the set
     for (auto& [binding, sampler, image]: samplerAndBindings) {
-        mDescriptorSetCache->updateSamplerForExternalSamplerSet(set, binding, image);
+        // We cannot call updateSamplerForExternalSamplerSet because some samplers are non NULL
+        // (RGB) and we cannot do a combined update with a NULL sampler.
+        mDescriptorSetCache->updateSampler(set, binding, image, sampler, newLayout);
     }
 }
 
@@ -261,13 +212,25 @@ void VulkanExternalImageManager::bindExternallySampledTexture(
         fvkmemory::resource_ptr<VulkanTexture> image, SamplerParams samplerParams) {
     // Should we do duplicate validation here?
     auto& imageData = findImage(mImages, image);
-    mSetBindings.push_back({ bindingPoint, imageData.image, set, samplerParams });
+    auto itr = std::find_if(mSetBindings.begin(), mSetBindings.end(),
+            [&](SetBindingInfo const& binding) {
+                return (binding.set == set && binding.binding == bindingPoint);
+            });
+    if (itr == mSetBindings.end()) {
+        mSetBindings.push_back({ bindingPoint, imageData.image, set, samplerParams });
+    } else {
+        // override the image data in the binding point
+        itr->image = image;
+        itr->samplerParams = samplerParams;
+    }
 }
 
 void VulkanExternalImageManager::addExternallySampledTexture(
        fvkmemory::resource_ptr<VulkanTexture> image,
         Platform::ExternalImageHandleRef platformHandleRef) {
-    mImages.push_back({ image, platformHandleRef, false });
+    // By passing VK_NULL_HANDLE which is already there by default.
+    // We make it clear that all default images do NOT have a chroma conversion.
+    mImages.push_back({ image, platformHandleRef, false, VK_NULL_HANDLE });
 }
 
 void VulkanExternalImageManager::removeExternallySampledTexture(

filament/backend/src/vulkan/VulkanExternalImageManager.h

@@ -31,7 +31,7 @@ class VulkanSamplerCache;
 class VulkanDescriptorSetLayoutCache;
 class VulkanDescriptorSetCache;
 
-// Manages the logic of external images and their quirks wrt Vulikan.
+// Manages the logic of external images and their quirks wrt Vulkan.
 class VulkanExternalImageManager {
 public:
 

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -40,6 +40,11 @@ namespace filament::backend {
 
 namespace {
 
+// We don't have a good estimate for this magic number. This is used to remove empty
+// slots from a descriptor set (a VulkanDescriptorset is backed by multiple
+// VkDescriptorSet.
+size_t DESCRIPTOR_SET_GC_LIMIT = 10;
+
 inline VulkanBufferBinding getBufferObjectBinding(BufferObjectBinding bindingType) noexcept {
     switch (bindingType) {
         case BufferObjectBinding::VERTEX:
@@ -192,15 +197,70 @@ VulkanDescriptorSetLayout::Bitmask VulkanDescriptorSetLayout::Bitmask::fromLayou
     return fromBackendLayout(layout);
 }
 
-// This method will store an age associated with this command buffer into the VulkanBuffer, which
+VulkanDescriptorSet::~VulkanDescriptorSet() {
-// will allow us to determine whether a barrier is necessary or not.
+    for (auto setBundle: mSets) {
+        if (setBundle.onRecycleFn) {
+            setBundle.onRecycleFn(this);
+        }
+    }
+}
+
+VulkanDescriptorSet::VulkanDescriptorSet(fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout,
+        OnRecycle&& onRecycleFn, VkDescriptorSet vkSet)
+    : dynamicUboMask(layout->bitmask.dynamicUbo),
+      uniqueDynamicUboCount(layout->count.dynamicUbo),
+      mLayout(layout),
+      mCurrentSetIndex(0) {
+    addNewSet(vkSet, std::move(onRecycleFn));
+}
+
 void VulkanDescriptorSet::referencedBy(VulkanCommandBuffer& commands) {
+    // This will store an age associated with this command buffer into the VulkanBuffer, which
+    // will allow us to determine whether a barrier is necessary or not.
     mUboMask.forEachSetBit([this, &commands](size_t index) {
         auto& res = mResources[index];
         fvkmemory::resource_ptr<VulkanBufferObject> bo =
                 fvkmemory::resource_ptr<VulkanBufferObject>::cast((VulkanBufferObject*) res.get());
         bo->referencedBy(commands);
     });
+    mSets[mCurrentSetIndex].fenceStatus = commands.getFenceStatus();
+}
+
+void VulkanDescriptorSet::gc() {
+    size_t empty = 0;
+    for (auto& setBundle: mSets) {
+        assert_invariant(setBundle.onRecycleFn);
+        if (setBundle.vkSet == VK_NULL_HANDLE) {
+            empty++;
+            continue;
+        }
+        if (setBundle.fenceStatus && setBundle.fenceStatus->getStatus() == VK_SUCCESS) {
+            setBundle.onRecycleFn(this);
+            setBundle.vkSet = VK_NULL_HANDLE;
+            setBundle.fenceStatus = {};
+            empty++;
+        }
+    }
+
+    if (empty > DESCRIPTOR_SET_GC_LIMIT) {
+        std::vector<InternalVkSet> retainedSets;
+        for (auto& setBundle: mSets) {
+            if (setBundle.vkSet != VK_NULL_HANDLE) {
+                retainedSets.push_back({
+                    setBundle.vkSet,
+                    std::move(setBundle.onRecycleFn),
+                    std::move(setBundle.fenceStatus)
+                });
+            }
+        }
+        std::swap(mSets, retainedSets);
+    }
+}
+
+void VulkanDescriptorSet::addNewSet(VkDescriptorSet vkSet, OnRecycle&& onRecycleFn) {
+    gc();
+    mCurrentSetIndex = mSets.size();
+    mSets.push_back({ vkSet, std::move(onRecycleFn) });
 }
 
 PushConstantDescription::PushConstantDescription(backend::Program const& program) {

filament/backend/src/vulkan/VulkanHandles.h

@@ -28,6 +28,7 @@
 #include "VulkanTexture.h"
 #include "vulkan/VulkanCommands.h"
 #include "vulkan/memory/Resource.h"
+#include "vulkan/memory/ResourcePointer.h"
 #include "vulkan/utils/Definitions.h"
 #include "vulkan/utils/StaticVector.h"
 
@@ -57,6 +58,8 @@ inline uint8_t collapsedCount(Bitmask const& mask) {
 
 } // anonymous namespace
 
+class VulkanDescriptorSetCache;
+
 struct VulkanBufferObject;
 
 struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Resource {
@@ -159,38 +162,14 @@ public:
     using OnRecycle = std::function<void(VulkanDescriptorSet*)>;
 
     VulkanDescriptorSet(
-            fvkutils::UniformBufferBitmask const& dynamicUboMask,
+            fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout,
-            uint8_t uniqueDynamicUboCount,
+            OnRecycle&& onRecycleFn, VkDescriptorSet vkSet);
-            OnRecycle&& onRecycleFn, VkDescriptorSet vkSet)
-        : dynamicUboMask(dynamicUboMask),
-          uniqueDynamicUboCount(uniqueDynamicUboCount),
-          mVkSet(vkSet),
-          mOnRecycleFn(std::move(onRecycleFn)) {}
 
     // NOLINTNEXTLINE(bugprone-exception-escape)
-    ~VulkanDescriptorSet() {
+    ~VulkanDescriptorSet();
-        if (mOnRecycleFn) {
-            mOnRecycleFn(this);
-        }
-        if (mOnRecycleExternalSamplerFn) {
-            mOnRecycleExternalSamplerFn(this);
-        }
-    }
 
     VkDescriptorSet getVkSet() const noexcept {
-        return mVkSet;
+        return mSets[mCurrentSetIndex].vkSet;
-    }
-
-    VkDescriptorSet getExternalSamplerVkSet() const noexcept {
-        return mExternalSamplerVkSet;
-    }
-
-    void setExternalSamplerVkSet(VkDescriptorSet vkset, OnRecycle onRecycle) {
-        mExternalSamplerVkSet = vkset;
-        if (mOnRecycleExternalSamplerFn) {
-            mOnRecycleExternalSamplerFn(this);
-        }
-        mOnRecycleExternalSamplerFn = onRecycle;
     }
 
     void setOffsets(backend::DescriptorSetOffsetArray&& offsets) noexcept {
@@ -211,17 +190,33 @@ public:
 
     void referencedBy(VulkanCommandBuffer& commands);
 
-    fvkutils::UniformBufferBitmask const dynamicUboMask;
+    bool isBound() const {
+        return bool(mSets[mCurrentSetIndex].fenceStatus);
+    }
+
+    fvkmemory::resource_ptr<VulkanDescriptorSetLayout> getLayout() const { return mLayout; }
+
+    fvkutils::UniformBufferBitmask const& dynamicUboMask;
     uint8_t const uniqueDynamicUboCount;
 
 private:
-    VkDescriptorSet const mVkSet;
+    friend class VulkanDescriptorSetCache;
-    VkDescriptorSet mExternalSamplerVkSet = VK_NULL_HANDLE;
 
+    void addNewSet(VkDescriptorSet vkSet, OnRecycle&& onRecycleFn);
+
+    void gc();
+
+    struct InternalVkSet {
+        VkDescriptorSet vkSet = {};
+        OnRecycle onRecycleFn;
+        std::shared_ptr<VulkanCmdFence> fenceStatus;
+    };
+
+    fvkmemory::resource_ptr<VulkanDescriptorSetLayout> mLayout;
     backend::DescriptorSetOffsetArray mOffsets;
     std::vector<fvkmemory::resource_ptr<fvkmemory::Resource>> mResources;
-    OnRecycle mOnRecycleFn;
+    uint8_t mCurrentSetIndex;
-    OnRecycle mOnRecycleExternalSamplerFn;
+    std::vector<InternalVkSet> mSets;
     fvkutils::UniformBufferBitmask mUboMask;
 };
 

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -191,9 +191,12 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
         .alphaToCoverageEnable = raster.alphaToCoverageEnable,
         .alphaToOneEnable = VK_FALSE,
     };
+    bool const enableDepthTest =
+        raster.depthCompareOp != SamplerCompareFunc::A ||
+        raster.depthWriteEnable;
     VkPipelineDepthStencilStateCreateInfo vkDs = {
         .sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
-        .depthTestEnable = VK_TRUE,
+        .depthTestEnable = enableDepthTest ? VK_TRUE : VK_FALSE,
         .depthWriteEnable = raster.depthWriteEnable,
         .depthCompareOp = fvkutils::getCompareOp(raster.depthCompareOp),
         .depthBoundsTestEnable = VK_FALSE,

filament/backend/src/vulkan/VulkanStreamedImageManager.cpp

@@ -0,0 +1,60 @@
+/*
+ * Copyright (C) 2025 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 "VulkanStreamedImageManager.h"
+
+#include "VulkanDescriptorSetCache.h"
+#include "VulkanExternalImageManager.h"
+#include "VulkanDescriptorSetLayoutCache.h"
+#include "VulkanSamplerCache.h"
+
+namespace filament::backend {
+VulkanStreamedImageManager::VulkanStreamedImageManager(VulkanExternalImageManager* manager)
+        : mExternalImageManager(manager) {}
+
+VulkanStreamedImageManager::~VulkanStreamedImageManager() = default;
+
+void VulkanStreamedImageManager::terminate() { mStreamedTexturesBindings.clear(); }
+
+void VulkanStreamedImageManager::bindStreamedTexture(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+        uint8_t bindingPoint, fvkmemory::resource_ptr<VulkanTexture> image,
+        SamplerParams samplerParams) {
+    mStreamedTexturesBindings.push_back({ bindingPoint, image, set, samplerParams });
+}
+
+void VulkanStreamedImageManager::unbindStreamedTexture(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+        uint8_t bindingPoint) {
+    auto iter = std::remove_if(mStreamedTexturesBindings.begin(), mStreamedTexturesBindings.end(),
+            [&](StreamedTextureBinding& binding) {
+                return ((binding.set == set) && (binding.binding == bindingPoint));
+            });
+    mStreamedTexturesBindings.erase(iter, mStreamedTexturesBindings.end());
+}
+
+void VulkanStreamedImageManager::onStreamAcquireImage(fvkmemory::resource_ptr<VulkanTexture> image,
+        fvkmemory::resource_ptr<VulkanStream> stream) {
+    for (StreamedTextureBinding const& data: mStreamedTexturesBindings) {
+        // Find the right stream
+        if (data.image->getStream() == stream) {
+            mExternalImageManager->bindExternallySampledTexture(data.set, data.binding, image,
+                    data.samplerParams);
+        }
+    }
+}
+
+} // namespace filament::backend

filament/backend/src/vulkan/VulkanStreamedImageManager.h

@@ -0,0 +1,64 @@
+/*
+ * Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANSTREAMEDIMAGEMANAGER_H
+#define TNT_FILAMENT_BACKEND_CACHING_VULKANSTREAMEDIMAGEMANAGER_H
+
+#include "VulkanHandles.h"
+
+#include <backend/DriverEnums.h>
+
+#include <array>
+#include <vector>
+
+namespace filament::backend {
+
+class VulkanExternalImageManager;
+class VulkanDescriptorSetCache;
+class VulkanSamplerCache;
+
+// Manages the logic of streamed images.
+class VulkanStreamedImageManager {
+public:
+    VulkanStreamedImageManager(
+            VulkanExternalImageManager* manager);
+    ~VulkanStreamedImageManager();
+    void terminate();
+
+public:
+    void bindStreamedTexture(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t bindingPoint,
+            fvkmemory::resource_ptr<VulkanTexture> image, SamplerParams samplerParams);
+    void unbindStreamedTexture(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+            uint8_t bindingPoint);
+    void onStreamAcquireImage(fvkmemory::resource_ptr<VulkanTexture> image,
+            fvkmemory::resource_ptr<VulkanStream> stream);
+
+private:
+    struct StreamedTextureBinding {
+        uint8_t binding = 0;
+        fvkmemory::resource_ptr<VulkanTexture> image;
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set;
+        SamplerParams samplerParams;
+    };
+    // keep track of all the stream bindings
+    std::vector<StreamedTextureBinding> mStreamedTexturesBindings;
+
+    VulkanExternalImageManager* mExternalImageManager;
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_CACHING_VULKANSTREAMEDIMAGEMANAGER_H

filament/backend/src/vulkan/VulkanSwapChain.cpp

@@ -19,6 +19,7 @@
 #include "VulkanCommands.h"
 #include "VulkanTexture.h"
 
+#include <utils/debug.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Panic.h>
 
@@ -105,6 +106,11 @@ void VulkanSwapChain::update() {
 }
 
 void VulkanSwapChain::present(DriverBase& driver) {
+    // The last acquire failed, so just skip presenting.
+    if (!mAcquired) {
+        return;
+    }
+
     if (!mHeadless && mTransitionSwapChainImageLayoutForPresent) {
         VulkanCommandBuffer& commands = mCommands->get();
         VkImageSubresourceRange const subresources{
@@ -161,7 +167,15 @@ void VulkanSwapChain::acquire(bool& resized) {
 
     VulkanPlatform::ImageSyncData imageSyncData;
     VkResult const result = mPlatform->acquire(swapChain, &imageSyncData);
+
+    if (result != VK_SUCCESS) {
+        // We just don't set mAcquired here so the next present will just skip.
+        FVK_LOGD << "Failed to acquire next image in the swapchain result=" << (int) result;
+        return;
+    }
+
     mCurrentSwapIndex = imageSyncData.imageIndex;
+    assert_invariant(mCurrentSwapIndex < mFinishedDrawing.size());
     mFinishedDrawing[mCurrentSwapIndex] = {};
     FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS || result == VK_SUBOPTIMAL_KHR)
             << "Cannot acquire in swapchain. error=" << static_cast<int32_t>(result);

filament/backend/src/vulkan/VulkanTexture.h

@@ -36,6 +36,43 @@ namespace filament::backend {
 
 struct VulkanTexture;
 
+struct VulkanStream : public HwStream, fvkmemory::Resource {
+
+    //-- These methods are only called from the frontend
+    void acquire(const AcquiredImage& image) {
+        user_thread.mPrevious = user_thread.mAcquired;
+        user_thread.mAcquired = image;
+    }
+    bool previousNeedsRelease() const { return (user_thread.mPrevious.image != nullptr); }
+    // this function will null the previous once the caller takes it.
+    // It ensures we don't schedule for release twice.
+    AcquiredImage takePrevious() {
+        AcquiredImage previous = user_thread.mPrevious;
+        user_thread.mPrevious = {};
+        return previous;
+    }
+    const AcquiredImage& getAcquired() const { return user_thread.mAcquired; }
+
+    //-- These methods are only called from the backend thread
+    fvkmemory::resource_ptr<VulkanTexture> getTexture(void* ahb) {
+        if (auto itr = mTextures.find(ahb); itr != mTextures.end()) {
+            return itr->second;
+        }
+        return {};
+    }
+    void pushImage(void* ahb, fvkmemory::resource_ptr<VulkanTexture> tex) { mTextures[ahb] = tex; }
+
+private:
+    // These are only called from the frontend
+    struct {
+        AcquiredImage mAcquired;
+        AcquiredImage mPrevious;
+    } user_thread;
+
+    // #TODO b/442937292
+    std::unordered_map<void*, fvkmemory::resource_ptr<VulkanTexture>> mTextures;
+};
+
 struct VulkanTextureState : public fvkmemory::Resource {
     VulkanTextureState(VulkanStagePool& stagePool, VulkanCommands* commands, VmaAllocator allocator,
             VkDevice device, VkImage image, VkDeviceMemory deviceMemory, VkFormat format,
@@ -74,6 +111,8 @@ private:
 
     // The texture with the sidecar owns the sidecar.
     fvkmemory::resource_ptr<VulkanTexture> mSidecarMSAA;
+    // The stream this texture is associated with.
+    fvkmemory::resource_ptr<VulkanStream> mStream;
 
     VkImage const mTextureImage;
     VkDeviceMemory const mTextureImageMemory;
@@ -105,6 +144,7 @@ private:
     using ImageViewHash = utils::hash::MurmurHashFn<ImageViewKey>;
     std::unordered_map<ImageViewKey, VkImageView, ImageViewHash> mCachedImageViews;
 
+
     friend struct VulkanTexture;
 };
 
@@ -182,6 +222,14 @@ struct VulkanTexture : public HwTexture, fvkmemory::Resource {
         return mState->mSidecarMSAA;
     }
 
+    void setStream(fvkmemory::resource_ptr<VulkanStream> stream) {
+        mState->mStream = stream;
+    }
+
+    fvkmemory::resource_ptr<VulkanStream> getStream() const {
+        return mState->mStream;
+    }
+
     bool isTransientAttachment() const {
         return mState->mUsage & VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT;
     }

filament/backend/src/vulkan/memory/Resource.cpp

@@ -41,6 +41,7 @@ template ResourceType getTypeEnum<VulkanBuffer>() noexcept;
 template ResourceType getTypeEnum<VulkanSync>() noexcept;
 template ResourceType getTypeEnum<VulkanMemoryMappedBuffer>() noexcept;
 template ResourceType getTypeEnum<VulkanSemaphore>() noexcept;
+template ResourceType getTypeEnum<VulkanStream>() noexcept;
 
 template<typename D>
 ResourceType getTypeEnum() noexcept {
@@ -104,6 +105,9 @@ ResourceType getTypeEnum() noexcept {
     if constexpr (std::is_same_v<D, VulkanSemaphore>) {
         return ResourceType::SEMAPHORE;
     }
+    if constexpr (std::is_same_v<D, VulkanStream>) {
+        return ResourceType::STREAM;
+    }
     return ResourceType::UNDEFINED_TYPE;
 }
 
@@ -149,6 +153,8 @@ std::string_view getTypeStr(ResourceType type) {
             return "VulkanMemoryMappedBuffer";
         case ResourceType::SEMAPHORE:
             return "Semaphore";
+        case ResourceType::STREAM:
+            return "VulkanStream";
         case ResourceType::UNDEFINED_TYPE:
             return "";
     }

filament/backend/src/vulkan/memory/Resource.h

@@ -55,7 +55,8 @@ enum class ResourceType : uint8_t {
     SYNC = 17,
     MEMORY_MAPPED_BUFFER = 18,
     SEMAPHORE = 19,
-    UNDEFINED_TYPE = 20,    // Must be the last enum because we use it for iterating over the enums.
+    STREAM = 20,
+    UNDEFINED_TYPE = 21,    // Must be the last enum because we use it for iterating over the enums.
 };
 
 template<typename D>

filament/backend/src/vulkan/memory/ResourceManager.cpp

@@ -123,6 +123,9 @@ void ResourceManager::destroyWithType(ResourceType type, HandleId id) {
         case ResourceType::SEMAPHORE:
             destruct<VulkanSemaphore>(Handle<VulkanSemaphore>(id));
             break;
+        case ResourceType::STREAM:
+            destruct<VulkanStream>(Handle<VulkanStream>(id));
+            break;
         case ResourceType::UNDEFINED_TYPE:
             break;
     }

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

@@ -13,6 +13,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+
 #include <backend/platforms/VulkanPlatformAndroid.h>
 
 #include "vulkan/VulkanConstants.h"
@@ -20,6 +21,8 @@
 #include "vulkan/platform/VulkanPlatformSwapChainImpl.h"
 #include "vulkan/utils/Image.h"
 
+#include "AndroidFrameCallback.h"
+
 #include <private/backend/BackendUtilsAndroid.h>
 
 #include <backend/DriverEnums.h>
@@ -32,6 +35,7 @@
 #include <android/hardware_buffer.h>
 #include <android/native_window.h>
 
+#include <new>
 #include <utility>
 
 using namespace bluevk;
@@ -200,6 +204,14 @@ uint32_t selectMemoryTypeForExternalImage(VkPhysicalDevice physicalDevice, VkDev
 
 } // namespace
 
+// ---------------------------------------------------------------------------------------------
+
+struct VulkanPlatformAndroid::AndroidDetails {
+    AndroidFrameCallback androidFrameCallback;
+};
+
+// ---------------------------------------------------------------------------------------------
+
 VulkanPlatformAndroid::ExternalImageVulkanAndroid::~ExternalImageVulkanAndroid() {
     if (__builtin_available(android 26, *)) {
         if (aHardwareBuffer) {
@@ -208,14 +220,19 @@ VulkanPlatformAndroid::ExternalImageVulkanAndroid::~ExternalImageVulkanAndroid()
     }
 }
 
-VulkanPlatformAndroid::VulkanPlatformAndroid() {
+// ---------------------------------------------------------------------------------------------
+
+VulkanPlatformAndroid::VulkanPlatformAndroid() :
+        mAndroidDetails(*(new(std::nothrow) AndroidDetails{})) {
     mOSVersion = android_get_device_api_level();
     if (mOSVersion < 0) {
         mOSVersion = __ANDROID_API_FUTURE__;
     }
 }
 
-VulkanPlatformAndroid::~VulkanPlatformAndroid() noexcept = default;
+VulkanPlatformAndroid::~VulkanPlatformAndroid() noexcept {
+    delete &mAndroidDetails;
+}
 
 Platform::ExternalImageHandle VulkanPlatformAndroid::createExternalImage(
         AHardwareBuffer const* buffer, bool sRGB) noexcept {
@@ -497,14 +514,14 @@ int VulkanPlatformAndroid::getOSVersion() const noexcept {
 }
 
 void VulkanPlatformAndroid::terminate() {
-    mAndroidFrameCallback.terminate();
+    mAndroidDetails.androidFrameCallback.terminate();
     VulkanPlatform::terminate();
 }
 
 Driver* VulkanPlatformAndroid::createDriver(void* sharedContext, DriverConfig const& driverConfig) {
     Driver* driver = VulkanPlatform::createDriver(sharedContext, driverConfig);
     if (driver) {
-        mAndroidFrameCallback.init();
+        mAndroidDetails.androidFrameCallback.init();
     }
     return driver;
 }
@@ -520,13 +537,19 @@ bool VulkanPlatformAndroid::queryCompositorTiming(SwapChain const* swapchain,
     }
 
     AndroidFrameCallback::Timeline const preferredTimeline{
-        mAndroidFrameCallback.getPreferredTimeline() };
+        mAndroidDetails.androidFrameCallback.getPreferredTimeline() };
     outCompositorTiming->frameTime = preferredTimeline.frameTime;
     outCompositorTiming->expectedPresentTime = preferredTimeline.expectedPresentTime;
     outCompositorTiming->frameTimelineDeadline = preferredTimeline.frameTimelineDeadline;
+    outCompositorTiming->compositeDeadline = CompositorTiming::INVALID;
+    outCompositorTiming->compositeInterval = CompositorTiming::INVALID;
+    outCompositorTiming->compositeToPresentLatency = CompositorTiming::INVALID;
+
+    // From this point on, we always return "success" because some timings were returned.
 
     auto vulkanSwapchain = static_cast<VulkanPlatformSwapChainBase const *>(swapchain);
-    return vulkanSwapchain->queryCompositorTiming(outCompositorTiming);
+    vulkanSwapchain->queryCompositorTiming(outCompositorTiming);
+    return true;
 }
 
 bool VulkanPlatformAndroid::setPresentFrameId(SwapChain const* swapchain, uint64_t frameId) noexcept {

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

@@ -165,6 +165,11 @@ VulkanPlatformSurfaceSwapChain::~VulkanPlatformSurfaceSwapChain() {
 VkResult VulkanPlatformSurfaceSwapChain::create() {
 #ifdef __ANDROID__
     NativeWindow::enableFrameTimestamps(static_cast<ANativeWindow*>(mNativeWindow), true);
+    // on Android, disable producer throttling
+    if (mProducerThrottling.isSupported()) {
+        mProducerThrottling.setProducerThrottlingEnabled(
+                static_cast<ANativeWindow*>(mNativeWindow), false);
+    }
 #endif
 
     VkSurfaceFormatKHR surfaceFormat = {};
@@ -356,13 +361,15 @@ bool VulkanPlatformSurfaceSwapChain::queryCompositorTiming(
         CompositorTiming* outCompositorTiming) const {
 #ifdef __ANDROID__
     // fallback to private APIs
-    int const status = NativeWindow::getCompositorTiming(
+    if (UTILS_VERY_LIKELY(mNativeWindow)) {
-            static_cast<ANativeWindow*>(mNativeWindow),
+        int const status = NativeWindow::getCompositorTiming(
-            &outCompositorTiming->compositeDeadline,
+                static_cast<ANativeWindow*>(mNativeWindow),
-            &outCompositorTiming->compositeInterval,
+                &outCompositorTiming->compositeDeadline,
-            &outCompositorTiming->compositeToPresentLatency);
+                &outCompositorTiming->compositeInterval,
-    if (status == 0) {
+                &outCompositorTiming->compositeToPresentLatency);
-        return true;
+        if (status == 0) {
+            return true;
+        }
     }
 #endif
     return VulkanPlatformSwapChainBase::queryCompositorTiming(outCompositorTiming);

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

@@ -25,7 +25,8 @@
 #include <bluevk/BlueVK.h>
 
 #ifdef __ANDROID__
-#include <AndroidSwapChainHelper.h>
+#include "AndroidSwapChainHelper.h"
+#include "AndroidNativeWindow.h"
 #endif
 
 #include <unordered_map>
@@ -122,6 +123,7 @@ private:
 
 #ifdef __ANDROID__
     AndroidSwapChainHelper mImpl{};
+    AndroidProducerThrottling mProducerThrottling;
 #endif
 };
 

filament/backend/src/vulkan/utils/Image.cpp

@@ -18,6 +18,12 @@
 
 #include "vulkan/VulkanTexture.h"
 
+#if defined(__ANDROID__)
+#include "backend/platforms/VulkanPlatformAndroid.h"
+#include <android/hardware_buffer.h>
+#include <android/native_window.h>
+#endif
+
 #include <utils/Panic.h>
 #include <utils/algorithm.h>
 #include <utils/debug.h>
@@ -234,6 +240,19 @@ uint8_t reduceSampleCount(uint8_t sampleCount, VkSampleCountFlags mask) {
     return mostSignificantBit((sampleCount - 1) & mask);
 }
 
+filament::backend::Platform::ExternalImageHandle createExternalImageFromRaw(
+        filament::backend::VulkanPlatform* platform, void* image, bool sRGB) {
+
+#if defined(__ANDROID__)
+    if (__builtin_available(android 26, *)) {
+        return static_cast<filament::backend::VulkanPlatformAndroid*>(platform)
+                ->createExternalImage(reinterpret_cast<AHardwareBuffer const*>(image), sRGB);
+    }
+#endif
+    return {};
+
+}
+
 } // namespace filament::backend::fvkutils
 
 bool operator<(const VkImageSubresourceRange& a, const VkImageSubresourceRange& b) {

filament/backend/src/vulkan/utils/Image.h

@@ -18,6 +18,7 @@
 #define TNT_FILAMENT_BACKEND_VULKAN_UTILS_IMAGE_H
 
 #include <backend/DriverEnums.h>
+#include <backend/platforms/VulkanPlatform.h>
 
 #include <utils/Log.h>
 
@@ -106,6 +107,9 @@ VkImageAspectFlags getImageAspect(VkFormat format);
 
 uint8_t reduceSampleCount(uint8_t sampleCount, VkSampleCountFlags mask);
 
+Platform::ExternalImageHandle createExternalImageFromRaw(filament::backend::VulkanPlatform* platform,
+        void* image,
+        bool sRGB);
 } // namespace fvkutils
 
 } // namespace filament::backend

filament/backend/src/webgpu/WebGPUBufferBase.cpp

@@ -18,6 +18,7 @@
 
 #include "WebGPUConstants.h"
 #include "WebGPUQueueManager.h"
+#include "WebGPUStagePool.h"
 
 #include "DriverBase.h"
 #include <backend/BufferDescriptor.h>
@@ -63,9 +64,9 @@ WebGPUBufferBase::WebGPUBufferBase(wgpu::Device const& device, const wgpu::Buffe
 // WebGPU requires that the size of the data copied from the staging buffer to the GPU buffer is a
 // multiple of 4. This function handles cases where the buffer descriptor's size is not a multiple
 // of 4 by padding with zeros.
-void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor&& bufferDescriptor,
+void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
         const uint32_t byteOffset, wgpu::Device const& device,
-        WebGPUQueueManager* const webGPUQueueManager) {
+        WebGPUQueueManager* const webGPUQueueManager, WebGPUStagePool* const webGPUStagePool) {
     FILAMENT_CHECK_PRECONDITION(bufferDescriptor.buffer)
             << "updateGPUBuffer called with a null buffer";
     FILAMENT_CHECK_PRECONDITION(bufferDescriptor.size + byteOffset <= mBuffer.GetSize())
@@ -85,53 +86,27 @@ void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor&& bufferDescriptor,
     const size_t stagingBufferSize =
             remainder == 0 ? bufferDescriptor.size : mainBulk + FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS;
 
-    // create a staging buffer
+    wgpu::Buffer stagingBuffer = webGPUStagePool->acquireBuffer(stagingBufferSize,
-    wgpu::BufferDescriptor descriptor{
+            webGPUQueueManager->getLatestSubmissionState());
-        .label = "Filament WebGPU Staging Buffer",
-        .usage = wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc,
-        .size = stagingBufferSize};
-    wgpu::Buffer stagingBuffer = device.CreateBuffer(&descriptor);
 
-    struct UserData final {
+    void* mappedRange = stagingBuffer.GetMappedRange();
-        uint32_t byteOffset;
+    assert_invariant(mappedRange);
-        size_t stagingBufferSize;
+
-        size_t remainder;
+    memcpy(mappedRange, bufferDescriptor.buffer, bufferDescriptor.size);
-        BufferDescriptor srcBufferDescriptor;
+
-        wgpu::Buffer stagingBuffer;
+    // Make sure the padded memory is set to 0 to have deterministic behaviors
-        WebGPUQueueManager* const webGPUQueueManager;
+    if (remainder != 0) {
-        wgpu::Buffer dstBuffer;
+        uint8_t* paddingStart = static_cast<uint8_t*>(mappedRange) + bufferDescriptor.size;
-    };
+        memset(paddingStart, 0, FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS - remainder);
-    auto userData = std::make_unique<UserData>(UserData{
+    }
-        .byteOffset = byteOffset,
+
-        .stagingBufferSize = stagingBufferSize,
+    stagingBuffer.Unmap();
-        .remainder = remainder,
+
-        .srcBufferDescriptor = std::move(bufferDescriptor),
+    // Copy the staging buffer contents to the destination buffer.
-        .stagingBuffer = stagingBuffer,
+    webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(stagingBuffer, 0, mBuffer,
-        .webGPUQueueManager = webGPUQueueManager,
+            byteOffset,
-        .dstBuffer = mBuffer});
+            remainder == 0 ? bufferDescriptor.size
-    stagingBuffer.MapAsync(
+                           : mainBulk + FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS);
-            wgpu::MapMode::Write, 0, stagingBufferSize, wgpu::CallbackMode::AllowProcessEvents,
-            [](wgpu::MapAsyncStatus status, const char* message, UserData* userdata) {
-                std::unique_ptr<UserData> data(static_cast<UserData*>(userdata));
-                if (UTILS_LIKELY(status == wgpu::MapAsyncStatus::Success)) {
-                    void* mappedRange = data->stagingBuffer.GetMappedRange();
-                    memcpy(mappedRange, data->srcBufferDescriptor.buffer,
-                            data->srcBufferDescriptor.size);
-                    if (data->remainder != 0) {
-                        uint8_t* paddingStart =
-                                static_cast<uint8_t*>(mappedRange) + data->srcBufferDescriptor.size;
-                        memset(paddingStart, 0,
-                                FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS - data->remainder);
-                    }
-                    data->stagingBuffer.Unmap();
-                    data->webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(
-                            data->stagingBuffer, 0, data->dstBuffer, data->byteOffset,
-                            data->stagingBufferSize);
-                } else {
-                    FWGPU_LOGE << "Failed to map staging buffer for readPixels: " << message;
-                }
-            },
-            userData.release());
 }
 
 } // namespace filament::backend

filament/backend/src/webgpu/WebGPUBufferBase.h

@@ -25,6 +25,7 @@ namespace filament::backend {
 
 class BufferDescriptor;
 class WebGPUQueueManager;
+class WebGPUStagePool;
 
 /**
   * A base class for WebGPU buffer objects, providing common functionality for creating and
@@ -39,8 +40,8 @@ public:
      * happen after draw commands encoded in the encoder. Submitting any commands up to this point
      * ensures the calls happen in the expected sequence.
      */
-    void updateGPUBuffer(BufferDescriptor&&, uint32_t byteOffset, wgpu::Device const& device,
+    void updateGPUBuffer(BufferDescriptor const&, uint32_t byteOffset, wgpu::Device const& device,
-            WebGPUQueueManager* const webGPUQueueManager);
+            WebGPUQueueManager* const webGPUQueueManager, WebGPUStagePool* const webGPUStagePool);
 
     [[nodiscard]] wgpu::Buffer const& getBuffer() const { return mBuffer; }
 

filament/backend/src/webgpu/WebGPUDriver.cpp

@@ -107,6 +107,7 @@ WebGPUDriver::WebGPUDriver(WebGPUPlatform& platform,
       mAdapter{ mPlatform.requestAdapter(nullptr) },
       mDevice{ mPlatform.requestDevice(mAdapter) },
       mQueueManager{ mDevice },
+      mStagePool{ mDevice },
       mPipelineLayoutCache{ mDevice },
       mPipelineCache{ mDevice },
       mRenderPassMipmapGenerator{ mDevice, &mQueueManager },
@@ -177,6 +178,9 @@ void WebGPUDriver::endFrame(const uint32_t /* frameId */) {
     for (size_t i = 0; i < MAX_DESCRIPTOR_SET_COUNT; i++) {
         mCurrentDescriptorSets[i] = {};
     }
+
+    // Garbage collection (if necessary)
+    mStagePool.gc();
 }
 
 // If a command encoder is in flight then the encoder is finished and submitted to the GPU queue.
@@ -651,11 +655,6 @@ void WebGPUDriver::setAcquiredImage(Handle<HwStream> sh, void* image, const math
     //todo
 }
 
-void WebGPUDriver::registerBufferObjectStreams(Handle<HwBufferObject> bufferObjectHandle,
-        BufferObjectStreamDescriptor&& streams) {
-    //todo
-}
-
 void WebGPUDriver::setStreamDimensions(Handle<HwStream> sh, uint32_t width, uint32_t height) {
     //todo
 }
@@ -856,7 +855,7 @@ void WebGPUDriver::updateIndexBuffer(Handle<HwIndexBuffer> indexBufferHandle,
     // draw calls are made.
     flush();
     handleCast<WebGPUIndexBuffer>(indexBufferHandle)
-            ->updateGPUBuffer(std::move(bufferDescriptor), byteOffset, mDevice, &mQueueManager);
+            ->updateGPUBuffer(bufferDescriptor, byteOffset, mDevice, &mQueueManager, &mStagePool);
     scheduleDestroy(std::move(bufferDescriptor));
 }
 
@@ -867,14 +866,14 @@ void WebGPUDriver::updateBufferObject(Handle<HwBufferObject> bufferObjectHandle,
     // draw calls are made.
     flush();
     handleCast<WebGPUBufferObject>(bufferObjectHandle)
-            ->updateGPUBuffer(std::move(bufferDescriptor), byteOffset, mDevice, &mQueueManager);
+            ->updateGPUBuffer(bufferDescriptor, byteOffset, mDevice, &mQueueManager, &mStagePool);
     scheduleDestroy(std::move(bufferDescriptor));
 }
 
 void WebGPUDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> bufferObjectHandle,
         BufferDescriptor&& bufferDescriptor, const uint32_t byteOffset) {
     handleCast<WebGPUBufferObject>(bufferObjectHandle)
-            ->updateGPUBuffer(std::move(bufferDescriptor), byteOffset, mDevice, &mQueueManager);
+            ->updateGPUBuffer(bufferDescriptor, byteOffset, mDevice, &mQueueManager, &mStagePool);
     scheduleDestroy(std::move(bufferDescriptor));
 }
 

filament/backend/src/webgpu/WebGPUDriver.h

@@ -25,6 +25,7 @@
 #include "webgpu/WebGPUPipelineLayoutCache.h"
 #include "webgpu/WebGPURenderPassMipmapGenerator.h"
 #include "webgpu/WebGPUQueueManager.h"
+#include "webgpu/WebGPUStagePool.h"
 #include "webgpu/utils/AsyncTaskCounter.h"
 #include <backend/platforms/WebGPUPlatform.h>
 
@@ -81,6 +82,7 @@ private:
     wgpu::Device mDevice = nullptr;
     wgpu::Limits mDeviceLimits = {};
     WebGPUQueueManager mQueueManager;
+    WebGPUStagePool mStagePool;
     void* mNativeWindow = nullptr;
     WebGPUSwapChain* mSwapChain = nullptr;
     uint64_t mNextFakeHandle = 1;

filament/backend/src/webgpu/WebGPUQueueManager.cpp

@@ -21,7 +21,6 @@
 #include <chrono>
 #include <cstdint>
 #include <thread>
-#include <iostream>
 
 namespace filament::backend {
 
@@ -59,13 +58,13 @@ WebGPUQueueManager::WebGPUQueueManager(wgpu::Device const& device)
 WebGPUQueueManager::~WebGPUQueueManager() = default;
 
 wgpu::CommandEncoder WebGPUQueueManager::getCommandEncoder() {
-    // std::unique_lock<std::mutex> lock(mLock);
     if (!mCommandEncoder) {
         wgpu::CommandEncoderDescriptor commandEncoderDescriptor = {
             .label = "Filament Command Encoder",
         };
         mCommandEncoder = mDevice.CreateCommandEncoder(&commandEncoderDescriptor);
         ASSERT_POSTCONDITION(mCommandEncoder, "Failed to create command encoder.");
+        mLatestSubmissionState = std::make_shared<WebGPUSubmissionState>();
     }
     return mCommandEncoder;
 }
@@ -92,14 +91,10 @@ std::shared_ptr<WebGPUSubmissionState> WebGPUQueueManager::getLatestSubmissionSt
 }
 
 void WebGPUQueueManager::submit() {
-    // std::unique_lock<std::mutex> lock(mLock);
     if (!mCommandEncoder) {
-        std::cout << "Run Yu: no mCommandEncoder found!\n";
         return;
     }
 
-    mLatestSubmissionState = std::make_shared<WebGPUSubmissionState>();
-
     wgpu::CommandBufferDescriptor commandBufferDescriptor{
         .label = "Filament Command Buffer",
     };

filament/backend/src/webgpu/WebGPUQueueManager.h

@@ -74,7 +74,6 @@ private:
     wgpu::Queue mQueue;
     wgpu::CommandEncoder mCommandEncoder;
     std::shared_ptr<WebGPUSubmissionState> mLatestSubmissionState;
-    std::mutex mLock;
 };
 
 } // namespace filament::backend

filament/backend/src/webgpu/WebGPUStagePool.cpp

@@ -0,0 +1,92 @@
+/*
+* Copyright (C) 2025 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 "WebGPUStagePool.h"
+
+#include "WebGPUConstants.h"
+#include "WebGPUQueueManager.h"
+
+namespace filament::backend {
+
+WebGPUStagePool::WebGPUStagePool(wgpu::Device const& device) : mDevice(device) {}
+
+WebGPUStagePool::~WebGPUStagePool() = default;
+
+wgpu::Buffer WebGPUStagePool::acquireBuffer(size_t requiredSize,
+        std::shared_ptr<WebGPUSubmissionState> submissionState) {
+    wgpu::Buffer buffer;
+    {
+        std::lock_guard<std::mutex> lock(mMutex);
+        auto iter = mBuffers.lower_bound(requiredSize);
+        if (iter != mBuffers.end()) {
+            buffer = iter->second;
+            mBuffers.erase(iter);
+        }
+    }
+    if (!buffer.Get()) {
+        buffer = createNewBuffer(requiredSize);
+    }
+    mInProgress.push_back({submissionState, buffer});
+    return buffer;
+}
+
+void WebGPUStagePool::recycleBuffer(wgpu::Buffer buffer) {
+    struct UserData final {
+        wgpu::Buffer buffer;
+        WebGPUStagePool* webGPUStagePool;
+    };
+    auto userData =
+            std::make_unique<UserData>(UserData{ .buffer = buffer, .webGPUStagePool = this });
+    buffer.MapAsync(wgpu::MapMode::Write, 0, buffer.GetSize(), wgpu::CallbackMode::AllowSpontaneous,
+            [data = std::move(userData)](wgpu::MapAsyncStatus status, const char* message) {
+                if (UTILS_LIKELY(status == wgpu::MapAsyncStatus::Success)) {
+                    if (!data->webGPUStagePool) {
+                        return;
+                    }
+                    std::lock_guard<std::mutex> lock(data->webGPUStagePool->mMutex);
+                    data->webGPUStagePool->mBuffers.insert(
+                            { data->buffer.GetSize(), data->buffer });
+                } else {
+                    FWGPU_LOGE << "Failed to MapAsync when recycling staging buffer: " << message;
+                }
+            });
+}
+
+void WebGPUStagePool::gc() {
+    // We found that MapAsync would sometimes lead to GetMappedRange returning nullptr if the
+    // command using that staging buffer has not finished executing, so here we only recycle those
+    // buffers that are not still being used by any command
+    std::vector<std::pair<std::shared_ptr<WebGPUSubmissionState>, wgpu::Buffer>> stillInProgress;
+    for (auto& [st, buffer]: mInProgress) {
+        if (st->getStatus() == FenceStatus::CONDITION_SATISFIED) {
+            recycleBuffer(buffer);
+        } else {
+            stillInProgress.push_back({st, buffer});
+        }
+    }
+    std::swap(mInProgress, stillInProgress);
+}
+
+wgpu::Buffer WebGPUStagePool::createNewBuffer(size_t bufferSize) {
+    wgpu::BufferDescriptor descriptor{
+        .label = "Filament WebGPU Staging Buffer",
+        .usage = wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc,
+        .size = bufferSize,
+        .mappedAtCreation = true };
+    return mDevice.CreateBuffer(&descriptor);
+}
+
+} // namespace filament::backend

filament/backend/src/webgpu/WebGPUStagePool.h

@@ -0,0 +1,50 @@
+/*
+* Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_WEBGPUSTAGEPOOL_H
+#define TNT_FILAMENT_BACKEND_WEBGPUSTAGEPOOL_H
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <map>
+#include <mutex>
+
+namespace filament::backend {
+
+struct WebGPUSubmissionState;
+
+class WebGPUStagePool {
+public:
+    WebGPUStagePool(wgpu::Device const& device);
+    ~WebGPUStagePool();
+
+    wgpu::Buffer acquireBuffer(size_t requiredSize,
+            std::shared_ptr<WebGPUSubmissionState> submissionState);
+    void recycleBuffer(wgpu::Buffer buffer);
+    void gc();
+
+private:
+    wgpu::Buffer createNewBuffer(size_t bufferSize);
+    std::multimap<uint32_t, wgpu::Buffer> mBuffers;
+    std::vector<std::pair<std::shared_ptr<WebGPUSubmissionState>, wgpu::Buffer>> mInProgress;
+    std::mutex mMutex;
+
+    wgpu::Device mDevice;
+};
+
+} // namespace filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_WEBGPUSTAGEPOOL_H

filament/include/filament/Engine.h

@@ -192,6 +192,7 @@ public:
     using StereoscopicType = backend::StereoscopicType;
     using Driver = backend::Driver;
     using GpuContextPriority = backend::Platform::GpuContextPriority;
+    using AsynchronousMode = backend::AsynchronousMode;
 
     /**
      * Config is used to define the memory footprint used by the engine, such as the
@@ -428,6 +429,11 @@ public:
          * initial memory usage.
          */
         uint32_t sharedUboInitialSizeInBytes = 256 * 64;
+
+        /**
+         * Asynchronous mode for the engine. Defines how asynchronous operations are handled.
+         */
+        AsynchronousMode asynchronousMode = AsynchronousMode::NONE;
     };
 
 
@@ -742,6 +748,14 @@ public:
      */
     bool isStereoSupported(StereoscopicType stereoscopicType) const noexcept;
 
+    /**
+     * Checks if the engine is set up for asynchronous operation. If it returns true, the
+     * asynchronous versions of the APIs are available for use.
+     *
+     * @return true if the engine supports asynchronous operation.
+     */
+    bool isAsynchronousOperationSupported() const noexcept;
+
     /**
      * Retrieves the configuration settings of this Engine.
      *

filament/include/filament/Material.h

@@ -69,6 +69,20 @@ public:
     using ShaderModel = backend::ShaderModel;
     using SubpassType = backend::SubpassType;
 
+    /**
+     * Defines whether a material instance should use UBO batching or not.
+     */
+    enum class UboBatchingMode {
+        /**
+         * For default, it follows the engine settings.
+         * If UBO batching is enabled on the engine and the material domain is SURFACE, it
+         * turns on the UBO batching. Otherwise, it turns off the UBO batching.
+        */
+        DEFAULT,
+        //! Disable the Ubo Batching for this material
+        DISABLED,
+    };
+
     /**
      * Holds information about a material parameter.
      */
@@ -165,6 +179,13 @@ public:
          */
         Builder& shadowSamplingQuality(ShadowSamplingQuality quality) noexcept;
 
+        /**
+         * Set the batching mode of the instances created from this material.
+         * @param uboBatchingMode
+         * @return
+         */
+        Builder& uboBatching(UboBatchingMode uboBatchingMode) noexcept;
+
         /**
          * Creates the Material object and returns a pointer to it.
          *

filament/include/filament/Renderer.h

@@ -113,8 +113,10 @@ public:
     /**
      * Retrieve a history of frame timing information. The maximum frame history size is
      * given by getMaxFrameHistorySize().
+     * All or part of the history can be lost when using a different SwapChain in beginFrame().
      * @param historySize requested history size. The returned vector could be smaller.
      * @return A vector of FrameInfo.
+     * @see beginFrame()
      */
     utils::FixedCapacityVector<FrameInfo> getFrameInfoHistory(
             size_t historySize = 1) const noexcept;
@@ -326,6 +328,8 @@ public:
      *                                 or 0 if unknown. This value should be the timestamp of
      *                                 the last h/w vsync. It is expressed in the
      *                                 std::chrono::steady_clock time base.
+     *                                 On Android this should be the frame time received from
+     *                                 a Choreographer.
      * @param swapChain A pointer to the SwapChain instance to use.
      *
      * @return
@@ -337,6 +341,8 @@ public:
      *
      * @note
      * All calls to render() must happen *after* beginFrame().
+     * It is recommended to use the same swapChain for every call to beginFrame, failing to do
+     * so can result is losing all or part of the FrameInfo history.
      *
      * @see
      * endFrame()

filament/src/Engine.cpp

@@ -455,6 +455,10 @@ bool Engine::isStereoSupported(StereoscopicType) const noexcept {
     return downcast(this)->isStereoSupported();
 }
 
+bool Engine::isAsynchronousOperationSupported() const noexcept {
+    return downcast(this)->isAsynchronousOperationSupported();
+}
+
 size_t Engine::getMaxStereoscopicEyes() noexcept {
     return FEngine::getMaxStereoscopicEyes();
 }

filament/src/FrameInfo.cpp

@@ -101,11 +101,24 @@ void FrameInfoManager::beginFrame(FSwapChain* swapChain, DriverApi& driver,
     auto& history = mFrameTimeHistory;
     // don't exceed the capacity, drop the oldest entry
     if (UTILS_LIKELY(history.size() == history.capacity())) {
-        if (!mDisableGpuFrameComplete) {
+        FrameInfoImpl& frameInfo = history.back();
-            assert_invariant(history.back().fence);
+        if (frameInfo.ready.load(std::memory_order_relaxed)) {
-            driver.destroyFence(std::move(history.back().fence));
+            if (!mDisableGpuFrameComplete) {
+                assert_invariant(frameInfo.fence);
+                driver.destroyFence(std::move(frameInfo.fence));
+            }
+            history.pop_back();
+        } else {
+            // This is a big problem, we ran out of space in the circular queue and that entry
+            // hasn't been processed yet. Because the code below keeps a reference to the
+            // front element of the queue, we can't pop/push. Our only option is to not record
+            // a new entry for this frame, which will create a false skipped frame in the
+            // data.
+            LOG(WARNING) << "FrameInfo's circular queue is full, but the oldest item hasn't "
+                            "been processed yet. Skipping this frame, id = " << frameId;
+            mLastBeginFrameSkipped = true;
+            return;
         }
-        history.pop_back();
     }
 
     // create a new entry
@@ -193,6 +206,14 @@ void FrameInfoManager::beginFrame(FSwapChain* swapChain, DriverApi& driver,
 }
 
 void FrameInfoManager::endFrame(DriverApi& driver) noexcept {
+    if (mLastBeginFrameSkipped) {
+        // if we had to skip the last beginFrame(), endFrame() needs to be skipped too
+        // because history.front() now references the wrong frame.
+        // It is guaranteed that if beginFrame() is called, endFrame() will be called too.
+        mLastBeginFrameSkipped = false;
+        return;
+    }
+
     auto& front = mFrameTimeHistory.front();
     front.endFrame = std::chrono::steady_clock::now();
 

filament/src/FrameInfo.h

@@ -95,33 +95,8 @@ struct FrameInfoImpl : public details::FrameInfo {
         assert_invariant(!fence);
     }
 
-    FrameInfoImpl(FrameInfoImpl&& rhs) noexcept : 
+    FrameInfoImpl(FrameInfoImpl& rhs) noexcept = delete;
-        details::FrameInfo(rhs),
+    FrameInfoImpl& operator=(FrameInfoImpl& rhs) noexcept = delete;
-        frameId(rhs.frameId),
-        beginFrame(rhs.beginFrame),
-        endFrame(rhs.endFrame),
-        backendBeginFrame(rhs.backendBeginFrame),
-        backendEndFrame(rhs.backendEndFrame),
-        gpuFrameComplete(rhs.gpuFrameComplete),
-        vsync(rhs.vsync),
-        fence(rhs.fence),
-        ready(rhs.ready.load())
-    {
-    }
-
-    FrameInfoImpl& operator=(FrameInfoImpl&& rhs) noexcept {
-        details::FrameInfo::operator=(rhs);
-        frameId = rhs.frameId;
-        beginFrame = rhs.beginFrame;
-        endFrame = rhs.endFrame;
-        backendBeginFrame = rhs.backendBeginFrame;
-        backendEndFrame = rhs.backendEndFrame;
-        gpuFrameComplete = rhs.gpuFrameComplete;
-        vsync = rhs.vsync;
-        fence = rhs.fence;
-        ready.store(rhs.ready.load());
-        return *this;
-    }
 };
 
 template<typename T, size_t CAPACITY>
@@ -340,6 +315,7 @@ private:
     FrameHistoryQueue mFrameTimeHistory{};
     utils::AsyncJobQueue mJobQueue;
     FSwapChain* mLastSeenSwapChain = nullptr;
+    bool mLastBeginFrameSkipped = false;
     bool const mHasTimerQueries = false;
     bool const mDisableGpuFrameComplete = false;
 };

filament/src/details/BufferAllocator.cpp

@@ -16,6 +16,7 @@
 
 #include "details/BufferAllocator.h"
 
+#include <private/utils/Tracing.h>
 #include <utils/Panic.h>
 #include <utils/debug.h>
 
@@ -169,6 +170,7 @@ void BufferAllocator::releaseGpu(AllocationId id) {
 }
 
 void BufferAllocator::releaseFreeSlots() {
+    FILAMENT_TRACING_CALL(FILAMENT_TRACING_CATEGORY_FILAMENT);
     if (!mHasPendingFrees) {
         return;
     }

filament/src/details/Engine.cpp

@@ -98,7 +98,7 @@ using namespace filaflat;
 namespace {
 
 backend::Platform::DriverConfig getDriverConfig(FEngine* instance) {
-    return {
+    Platform::DriverConfig driverConfig {
         .handleArenaSize = instance->getRequestedDriverHandleArenaSize(),
         .metalUploadBufferSizeBytes = instance->getConfig().metalUploadBufferSizeBytes,
         .disableParallelShaderCompile = instance->features.backend.disable_parallel_shader_compile,
@@ -116,7 +116,11 @@ backend::Platform::DriverConfig getDriverConfig(FEngine* instance) {
         .gpuContextPriority = instance->getConfig().gpuContextPriority,
         .vulkanEnableStagingBufferBypass =
                 instance->features.backend.vulkan.enable_staging_buffer_bypass,
+        .asynchronousMode = instance->features.backend.enable_asynchronous_operation ?
+                instance->getConfig().asynchronousMode : AsynchronousMode::NONE,
     };
+
+    return driverConfig;
 }
 
 } // anonymous
@@ -718,8 +722,7 @@ void FEngine::prepare() {
 
     if (useUboBatching) {
         assert_invariant(mUboManager != nullptr);
-
+        mUboManager->beginFrame(driver);
-        mUboManager->beginFrame(driver, mMaterialInstances);
     }
 
     UboManager* uboManager = mUboManager;
@@ -728,7 +731,6 @@ void FEngine::prepare() {
             // post-process materials instances must be commited explicitly because their
             // parameters are typically not set at this point in time.
             if (item->getMaterial()->getMaterialDomain() == MaterialDomain::SURFACE) {
-                item->commitStreamUniformAssociations(driver);
                 item->commit(driver, uboManager);
             }
         });
@@ -758,7 +760,7 @@ void FEngine::gc() {
 void FEngine::submitFrame() {
     if (isUboBatchingEnabled()) {
         DriverApi& driver = getDriverApi();
-        getUboManager()->endFrame(driver, getMaterialInstanceResourceList());
+        getUboManager()->endFrame(driver);
     }
 }
 
@@ -992,10 +994,9 @@ FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
     return p;
 }
 
-FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material, const char* name,
+FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
-        UboBatchingMode batchingMode) noexcept {
+        const char* name) noexcept {
-    FMaterialInstance* p =
+    FMaterialInstance* p = mHeapAllocator.make<FMaterialInstance>(*this, material, name);
-            mHeapAllocator.make<FMaterialInstance>(*this, material, name, batchingMode);
     if (UTILS_LIKELY(p)) {
         auto pos = mMaterialInstances.emplace(material, "MaterialInstance");
         pos.first->second.insert(p);
@@ -1285,11 +1286,6 @@ UTILS_NOINLINE
 bool FEngine::destroy(const FMaterialInstance* p) {
     if (p == nullptr) return true;
 
-    if (p->isUsingUboBatching()) {
-        assert_invariant(isUboBatchingEnabled());
-        mUboManager->retireSlot(p->getAllocationId());
-    }
-
     // Check that the material instance we're destroying is not in use in the RenderableManager
     // To do this, we currently need to inspect all render primitives in the RenderableManager
     EntityManager const& em = mEntityManager;
@@ -1509,6 +1505,11 @@ Engine::FeatureLevel FEngine::setActiveFeatureLevel(FeatureLevel featureLevel) {
     return (mActiveFeatureLevel = std::max(mActiveFeatureLevel, featureLevel));
 }
 
+bool FEngine::isAsynchronousOperationSupported() const noexcept {
+    return features.backend.enable_asynchronous_operation &&
+        mConfig.asynchronousMode != AsynchronousMode::NONE;
+}
+
 #if defined(__EMSCRIPTEN__)
 void FEngine::resetBackendState() noexcept {
     getDriverApi().resetState();

filament/src/details/Engine.h

@@ -216,6 +216,8 @@ public:
         return getDriver().isStereoSupported();
     }
 
+    bool isAsynchronousOperationSupported() const noexcept;
+
     static size_t getMaxStereoscopicEyes() noexcept {
         return CONFIG_MAX_STEREOSCOPIC_EYES;
     }
@@ -345,21 +347,10 @@ public:
 
     FRenderer* createRenderer() noexcept;
 
-    // Defines whether a material instance should use UBO batching or not.
-    enum class UboBatchingMode {
-        // For default, it follows the engine settings.
-        // If UBO batching is enabled on the engine and the material domain is not SURFACE, it
-        // turns on the UBO batching. Otherwise, it turns off the UBO batching.
-        DEFAULT,
-        NO_UBO_BATCHING,
-        UBO_BATCHING
-    };
-
     FMaterialInstance* createMaterialInstance(const FMaterial* material,
             const FMaterialInstance* other, const char* name) noexcept;
 
-    FMaterialInstance* createMaterialInstance(const FMaterial* material, const char* name,
+    FMaterialInstance* createMaterialInstance(const FMaterial* material, const char* name) noexcept;
-            UboBatchingMode batchingMode) noexcept;
 
     FScene* createScene() noexcept;
     FView* createView() noexcept;
@@ -780,7 +771,7 @@ public:
                 bool assert_texture_can_generate_mipmap = CORRECTNESS_ASSERTION_DEFAULT;
             } debug;
             struct {
-                bool disable_gpu_frame_complete_metric = false;
+                bool disable_gpu_frame_complete_metric = true;
             } frame_info;
         } engine;
         struct {
@@ -797,6 +788,7 @@ public:
             bool disable_amortized_shader_compile = true;
             bool disable_handle_use_after_free_check = false;
             bool disable_heap_handle_tags = true; // FIXME: this should be false
+            bool enable_asynchronous_operation = false;
         } backend;
         struct {
             bool check_crc32_after_loading = false;
@@ -817,6 +809,9 @@ public:
             { "backend.disable_heap_handle_tags",
               "Disable Handle<> tags for heap-allocated handles.",
               &features.backend.disable_heap_handle_tags, true },
+            { "backend.enable_asynchronous_operation",
+              "Enable asynchronous operation for resource management.",
+              &features.backend.enable_asynchronous_operation, true },
             { "backend.opengl.assert_native_window_is_valid",
               "Asserts that the ANativeWindow is valid when rendering starts.",
               &features.backend.opengl.assert_native_window_is_valid, true },

filament/src/details/Material.cpp

@@ -70,10 +70,19 @@
 
 namespace filament {
 
+namespace {
+
 using namespace backend;
 using namespace filaflat;
 using namespace utils;
-using UboBatchingMode = FEngine::UboBatchingMode;
+using UboBatchingMode = Material::UboBatchingMode;
+
+bool shouldEnableBatching(FEngine& engine, UboBatchingMode batchingMode, MaterialDomain domain) {
+    return batchingMode != UboBatchingMode::DISABLED && engine.isUboBatchingEnabled() &&
+           domain == MaterialDomain::SURFACE;
+}
+
+} // anonymous namespace
 
 struct Material::BuilderDetails {
     const void* mPayload = nullptr;
@@ -81,6 +90,7 @@ struct Material::BuilderDetails {
     bool mDefaultMaterial = false;
     int32_t mShBandsCount = 3;
     Builder::ShadowSamplingQuality mShadowSamplingQuality = Builder::ShadowSamplingQuality::LOW;
+    UboBatchingMode mUboBatchingMode = UboBatchingMode::DEFAULT;
     std::unordered_map<
         CString,
         std::variant<int32_t, float, bool>,
@@ -115,6 +125,11 @@ Material::Builder& Material::Builder::shadowSamplingQuality(ShadowSamplingQualit
     return *this;
 }
 
+Material::Builder& Material::Builder::uboBatching(UboBatchingMode const mode) noexcept {
+    mImpl->mUboBatchingMode = mode;
+    return *this;
+}
+
 template<typename T, typename>
 Material::Builder& Material::Builder::constant(const char* name, size_t nameLength, T value) {
     FILAMENT_CHECK_PRECONDITION(name != nullptr) << "name cannot be null";
@@ -148,8 +163,12 @@ Material* Material::Builder::build(Engine& engine) const {
 FMaterial::FMaterial(FEngine& engine, const Builder& builder, MaterialDefinition const& definition)
         : mDefinition(definition),
           mIsDefaultMaterial(builder->mDefaultMaterial),
+          mUseUboBatching(shouldEnableBatching(engine, builder->mUboBatchingMode,
+                  definition.materialDomain)),
           mEngine(engine),
           mMaterialId(engine.getMaterialId()) {
+    FILAMENT_CHECK_PRECONDITION(!mUseUboBatching || engine.isUboBatchingEnabled())
+            << "UBO batching is not enabled.";
     mSpecializationConstants = processSpecializationConstants(builder);
     precacheDepthVariants(engine);
 
@@ -282,14 +301,14 @@ FMaterialInstance* FMaterial::createInstance(const char* name) const noexcept {
         return FMaterialInstance::duplicate(mDefaultMaterialInstance, name);
     } else {
         // but if we don't, just create an instance with all the default parameters
-        return mEngine.createMaterialInstance(this, name, UboBatchingMode::DEFAULT);
+        return mEngine.createMaterialInstance(this, name);
     }
 }
 
 FMaterialInstance* FMaterial::getDefaultInstance() noexcept {
     if (UTILS_UNLIKELY(!mDefaultMaterialInstance)) {
         mDefaultMaterialInstance =
-                mEngine.createMaterialInstance(this, mDefinition.name.c_str(), UboBatchingMode::DEFAULT);
+                mEngine.createMaterialInstance(this, mDefinition.name.c_str());
         mDefaultMaterialInstance->setDefaultInstance(true);
     }
     return mDefaultMaterialInstance;

filament/src/details/Material.h

@@ -248,6 +248,10 @@ public:
         return mDefinition.perViewLayoutIndex;
     }
 
+    bool useUboBatching() const noexcept {
+        return mUseUboBatching;
+    }
+
 #if FILAMENT_ENABLE_MATDBG
     void applyPendingEdits() noexcept;
 
@@ -310,6 +314,8 @@ private:
 
     bool mIsDefaultMaterial = false;
 
+    bool mUseUboBatching = false;
+
     // reserve some space to construct the default material instance
     mutable FMaterialInstance* mDefaultMaterialInstance = nullptr;
 

filament/src/details/MaterialInstance.cpp

@@ -58,18 +58,8 @@ using namespace utils;
 namespace filament {
 
 using namespace backend;
-using UboBatchingMode = FEngine::UboBatchingMode;
 
-bool shouldEnableBatching(FEngine& engine, UboBatchingMode batchingMode, MaterialDomain domain) {
+FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name) noexcept
-    if (batchingMode != UboBatchingMode::DEFAULT) {
-        return batchingMode == UboBatchingMode::UBO_BATCHING;
-    }
-
-    return engine.isUboBatchingEnabled() && domain == MaterialDomain::SURFACE;
-}
-
-FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name,
-        UboBatchingMode batchingMode) noexcept
         : mMaterial(material),
           mDescriptorSet("MaterialInstance", material->getDescriptorSetLayout()),
           mCulling(CullingMode::BACK),
@@ -80,12 +70,9 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
           mHasScissor(false),
           mIsDoubleSided(false),
           mIsDefaultInstance(false),
-          mUseUboBatching(
+          mUseUboBatching(material->useUboBatching()),
-                  shouldEnableBatching(engine, batchingMode, material->getMaterialDomain())),
           mTransparencyMode(TransparencyMode::DEFAULT),
           mName(name ? CString(name) : material->getName()) {
-    FILAMENT_CHECK_PRECONDITION(!mUseUboBatching || engine.isUboBatchingEnabled())
-            << "UBO batching is not enabled.";
     FEngine::DriverApi& driver = engine.getDriverApi();
 
     // even if the material doesn't have any parameters, we allocate a small UBO because it's
@@ -95,6 +82,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
 
     if (mUseUboBatching) {
         mUboData = BufferAllocator::UNALLOCATED;
+        engine.getUboManager()->manageMaterialInstance(this);
     } else {
         mUboData = driver.createBufferObject(mUniforms.getSize(), BufferObjectBinding::UNIFORM,
                 BufferUsage::STATIC, ImmutableCString{ material->getName().c_str_safe() });
@@ -159,7 +147,6 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
           mUseUboBatching(other->mUseUboBatching),
           mScissorRect(other->mScissorRect),
           mName(name ? CString(name) : other->mName) {
-    assert_invariant(!mUseUboBatching || engine.isUboBatchingEnabled());
     FEngine::DriverApi& driver = engine.getDriverApi();
     FMaterial const* const material = other->getMaterial();
 
@@ -167,6 +154,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
 
     if (mUseUboBatching) {
         mUboData = BufferAllocator::UNALLOCATED;
+        engine.getUboManager()->manageMaterialInstance(this);
     } else {
         mUboData = driver.createBufferObject(mUniforms.getSize(), BufferObjectBinding::UNIFORM,
                 BufferUsage::DYNAMIC, ImmutableCString{ material->getName().c_str_safe() });
@@ -211,37 +199,16 @@ FMaterialInstance::~FMaterialInstance() noexcept = default;
 void FMaterialInstance::terminate(FEngine& engine) {
     FEngine::DriverApi& driver = engine.getDriverApi();
     mDescriptorSet.terminate(driver);
+    if (mUseUboBatching) {
+        engine.getUboManager()->unmanageMaterialInstance(this);
+    }
+
     auto* ubHandle = std::get_if<Handle<HwBufferObject>>(&mUboData);
     if (ubHandle){
         driver.destroyBufferObject(*ubHandle);
     }
 }
 
-void FMaterialInstance::commitStreamUniformAssociations(FEngine::DriverApi& driver) {
-    mHasStreamUniformAssociations = false;
-    if (!mTextureParameters.empty()) {
-        BufferObjectStreamDescriptor descriptor;
-        for (auto const& [binding, p]: mTextureParameters) {
-            ssize_t offset = mMaterial->getUniformInterfaceBlock().getTransformFieldOffset(binding);
-            if (offset >= 0) {
-                if (auto stream = p.texture->getStream()) {
-                    mHasStreamUniformAssociations = true;
-                    descriptor.mStreams.push_back({ uint32_t(offset), stream->getHandle(),
-                        BufferObjectStreamAssociationType::TRANSFORM_MATRIX });
-                }
-            }
-        }
-        if (descriptor.mStreams.size() > 0) {
-            // UBO batching is incompatible with stream uniform associations because streams require a
-            // dedicated UBO handle, not a sub-allocation. Assert that any instance here uses its own UBO.
-            assert_invariant(!mUseUboBatching);
-            auto const* ubHandle = std::get_if<Handle<HwBufferObject>>(&mUboData);
-            assert_invariant(ubHandle);
-            driver.registerBufferObjectStreams(*ubHandle, std::move(descriptor));
-        }
-    }
-}
-
 void FMaterialInstance::commit(FEngine& engine) const {
     if (UTILS_LIKELY(mMaterial->getMaterialDomain() != MaterialDomain::SURFACE)) {
         commit(engine.getDriverApi(), engine.getUboManager());
@@ -249,10 +216,9 @@ void FMaterialInstance::commit(FEngine& engine) const {
 }
 
 void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManager) const {
-    if (mUniforms.isDirty() || mHasStreamUniformAssociations) {
+    if (mUniforms.isDirty()) {
         mUniforms.clean();
-        if (mUseUboBatching) {
+        if (isUsingUboBatching()) {
-            assert_invariant(uboManager != nullptr);
             if (!BufferAllocator::isValid(getAllocationId())) {
                 // The allocation hasn't happened yet, return.
                 return;
@@ -283,7 +249,7 @@ void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManage
     // TODO: eventually we should remove this in RELEASE builds
     fixMissingSamplers();
 
-    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
+    if (isUsingUboBatching() && !BufferAllocator::isValid(getAllocationId())) {
         return;
     }
 
@@ -433,7 +399,7 @@ void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
         return;
     }
 
-    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
+    if (isUsingUboBatching() && !BufferAllocator::isValid(getAllocationId())) {
         return;
     }
 
@@ -470,7 +436,7 @@ void FMaterialInstance::assignUboAllocation(
         const Handle<HwBufferObject>& ubHandle,
         BufferAllocator::AllocationId id,
         BufferAllocator::allocation_size_t offset) {
-    assert_invariant(mUseUboBatching);
+    assert_invariant(isUsingUboBatching());
 
     mUboData = id;
     mUboOffset = offset;

filament/src/details/MaterialInstance.h

@@ -56,8 +56,7 @@ class FTexture;
 
 class FMaterialInstance : public MaterialInstance {
 public:
-    FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name,
+    FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name) noexcept;
-            FEngine::UboBatchingMode batchingMode) noexcept;
     FMaterialInstance(FEngine& engine, FMaterialInstance const* other, const char* name);
     FMaterialInstance(const FMaterialInstance& rhs) = delete;
     FMaterialInstance& operator=(const FMaterialInstance& rhs) = delete;
@@ -67,8 +66,6 @@ public:
     ~FMaterialInstance() noexcept;
 
     void terminate(FEngine& engine);
-
-    void commitStreamUniformAssociations(FEngine::DriverApi& driver);
     
     void commit(FEngine& engine) const;
 
@@ -288,7 +285,6 @@ private:
     tsl::robin_map<backend::descriptor_binding_t, TextureParameter> mTextureParameters;
     mutable DescriptorSet mDescriptorSet;
     UniformBuffer mUniforms;
-    bool mHasStreamUniformAssociations = false;
 
     backend::PolygonOffset mPolygonOffset{};
     backend::StencilState mStencilState{};

filament/src/details/Renderer.cpp

@@ -292,16 +292,11 @@ void FRenderer::skipFrame(uint64_t vsyncSteadyClockTimeNano) {
     // do this before engine.flush()
     mResourceAllocator->gc(true);
 
-    // Run the component managers' GC in parallel
-    // WARNING: while doing this we can't access any component manager
-    auto& js = engine.getJobSystem();
-
-    auto *job = js.runAndRetain(jobs::createJob(js, nullptr, &FEngine::gc, &engine)); // gc all managers
-
     engine.flush();     // flush command stream
 
-    // make sure we're done with the gcs
+    // Run the component managers' GC
-    js.waitAndRelease(job);
+    auto& js = engine.getJobSystem();
+    js.runAndWait(jobs::createJob(js, nullptr, &FEngine::gc, &engine)); // gc all managers
 }
 
 bool FRenderer::shouldRenderFrame() const noexcept {
@@ -468,16 +463,11 @@ void FRenderer::endFrame() {
     // do this before engine.flush()
     mResourceAllocator->gc();
 
-    // Run the component managers' GC in parallel
-    // WARNING: while doing this we can't access any component manager
-    auto& js = engine.getJobSystem();
-
-    auto *job = js.runAndRetain(jobs::createJob(js, nullptr, &FEngine::gc, &engine)); // gc all managers
-
     engine.flush();     // flush command stream
 
-    // make sure we're done with the gcs
+    // Run the component managers' GC
-    js.waitAndRelease(job);
+    auto& js = engine.getJobSystem();
+    js.runAndWait(jobs::createJob(js, nullptr, &FEngine::gc, &engine)); // gc all managers
 }
 
 void FRenderer::readPixels(
@@ -571,7 +561,6 @@ void FRenderer::renderStandaloneView(FView const* view) {
 
     if (UTILS_LIKELY(view->getScene())) {
         mPreviousRenderTargets.clear();
-        mFrameId++;
 
         // ask the engine to do what it needs to (e.g. updates light buffer, materials...)
         FEngine& engine = mEngine;
@@ -581,8 +570,7 @@ void FRenderer::renderStandaloneView(FView const* view) {
         driver.beginFrame(
                 steady_clock::now().time_since_epoch().count(),
                 mDisplayInfo.refreshRate == 0.0 ? 0 : int64_t(
-                        1'000'000'000.0 / mDisplayInfo.refreshRate),
+                        1'000'000'000.0 / mDisplayInfo.refreshRate), 0);
-                mFrameId);
 
         // because we don't have a "present" call, we use flush so the driver can submit
         // the command buffer; we do this before driver.endFrame() to mimic what would
@@ -591,7 +579,7 @@ void FRenderer::renderStandaloneView(FView const* view) {
 
         engine.submitFrame();
 
-        driver.endFrame(mFrameId);
+        driver.endFrame(0);
 
         // engine.flush() has already been called by renderInternal(), but we need an extra one
         // for endFrame() above. This operation in actually not too heavy, it just kicks the

filament/src/details/Renderer.h

@@ -209,7 +209,7 @@ private:
     backend::Handle<backend::HwRenderTarget> mRenderTargetHandle;
     FSwapChain* mSwapChain = nullptr;
     size_t mCommandsHighWatermark = 0;
-    uint32_t mFrameId = 0;
+    uint32_t mFrameId = 1; // id 0 is reserved for standalone views
     FrameInfoManager mFrameInfoManager;
     backend::TextureFormat mHdrTranslucent;
     backend::TextureFormat mHdrQualityMedium;