immediately GetMappedRange after MapAsync still would return null pointer

Use a stage pool for WebGPU, but somehow GetMappedRange would sometimes return null, which breaks the code.
2025-11-21 16:53:03 -05:00 · 2025-11-21 10:57:02 -05:00
296 changed files with 12720 additions and 35602 deletions
@@ -21,9 +21,8 @@ jobs:
      - name: Run update script
        env:
          GH_TOKEN: ${{ secrets.FILAMENTBOT_TOKEN }}
-          COMMIT_MESSAGE: ${{ steps.get_commit_msg.outputs.msg }}
        run: |
-          GOLDEN_BRANCH=$(echo "${COMMIT_MESSAGE}" | python3 test/renderdiff/src/commit_msg.py)
+          GOLDEN_BRANCH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | python3 test/renderdiff/src/commit_msg.py)
          COMMIT_HASH="${{ steps.get_commit_msg.outputs.hash }}"
          if [[ "${GOLDEN_BRANCH}" != "main" ]]; then
              git config --global user.email "filament.bot@gmail.com"
@@ -134,17 +134,18 @@ jobs:
        env:
          COMMIT_MESSAGE: ${{ steps.get_commit_msg.outputs.msg }}
        run: |
+          ls ./gltf/Models
          TEST_DIR=test/renderdiff
          source ${TEST_DIR}/src/preamble.sh
-          set -eux
+          start_
          GOLDEN_BRANCH=$(echo "${COMMIT_MESSAGE}" | python3 ${TEST_DIR}/src/commit_msg.py)
-          bash ${TEST_DIR}/generate.sh
-          python3 ${TEST_DIR}/src/golden_manager.py \
+          bash ${TEST_DIR}/generate.sh && \
+              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 -eux`
-          set +eux
+          # Note that we need to upload the output even if comparison fails, so we undo `set -ex`
+          end_

          python3 ${TEST_DIR}/src/compare.py \
              --src=${GOLDEN_OUTPUT_DIR} \
@@ -157,7 +158,6 @@ jobs:
             cat compare_output.txt >> "$GITHUB_OUTPUT"
             echo "$DELIMITER" >> "$GITHUB_OUTPUT"
          fi
-        shell: bash
      - uses: actions/upload-artifact@v4
        with:
          name: presubmit-renderdiff-result
@@ -872,9 +872,7 @@ add_subdirectory(${LIBRARIES}/utils)
 add_subdirectory(${LIBRARIES}/viewer)
 add_subdirectory(${FILAMENT}/shaders)
 add_subdirectory(${EXTERNAL}/abseil/tnt)
-# Add zstd before basisu to force it to use the external zstd target,
-# preventing a duplicate symbol conflict with its bundled version.
-add_subdirectory(${EXTERNAL}/zstd/tnt)
+add_subdirectory(${EXTERNAL}/basisu/tnt)
 add_subdirectory(${EXTERNAL}/civetweb/tnt)
 add_subdirectory(${EXTERNAL}/imgui/tnt)
 add_subdirectory(${EXTERNAL}/robin-map/tnt)
@@ -888,7 +886,7 @@ add_subdirectory(${EXTERNAL}/jsmn/tnt)
 add_subdirectory(${EXTERNAL}/stb/tnt)
 add_subdirectory(${EXTERNAL}/getopt)
 add_subdirectory(${EXTERNAL}/perfetto/tnt)
-add_subdirectory(${EXTERNAL}/basisu/tnt)
+add_subdirectory(${EXTERNAL}/zstd/tnt)


 # Note that this has to be placed after mikktspace in order for combine_static_libs to work.
@@ -31,7 +31,7 @@ repositories {
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.68.0'
+    implementation 'com.google.android.filament:filament-android:1.67.1'
 }
 ```

@@ -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', '~> 1.68.0'
+pod 'Filament', '~> 1.67.1'
 ```

 ## Documentation
@@ -7,9 +7,6 @@ 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)
@@ -100,7 +100,7 @@ buildscript {
        'kotlin': '2.0.21',
        'kotlin_coroutines': '1.9.0',
        'buildTools': '35.0.0',
-        'ndk': '29.0.14206865',
+        'ndk': '27.0.11718014',
        'androidx_core': '1.13.1',
        'androidx_annotations': '1.9.0'
    ]
@@ -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, jint uboBatchingMode) {
+        jlong nativeEngine, jobject buffer_, jint size, jint shBandCount, jint shadowQuality) {
    Engine* engine = (Engine*) nativeEngine;
    AutoBuffer buffer(env, buffer_, size);
    auto builder = Material::Builder();
@@ -33,7 +33,6 @@ 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);
@@ -261,20 +261,6 @@ 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;
@@ -386,7 +372,6 @@ public class Material {
        private int mSize;
        private int mShBandCount = 0;
        private ShadowSamplingQuality mShadowSamplingQuality = ShadowSamplingQuality.LOW;
-        private UboBatchingMode mUboBatchingMode = UboBatchingMode.DEFAULT;


        /**
@@ -431,17 +416,6 @@ 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.
         *
@@ -454,7 +428,7 @@ public class Material {
        @NonNull
        public Material build(@NonNull Engine engine) {
            long nativeMaterial = nBuilderBuild(engine.getNativeObject(),
-                mBuffer, mSize, mShBandCount, mShadowSamplingQuality.ordinal(), mUboBatchingMode.ordinal());
+                mBuffer, mSize, mShBandCount, mShadowSamplingQuality.ordinal());
            if (nativeMaterial == 0) throw new IllegalStateException("Couldn't create Material");
            return new Material(nativeMaterial);
        }
@@ -1120,7 +1094,7 @@ public class Material {
        mNativeObject = 0;
    }

-    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount, int shadowQuality, int uboBatchingMode);
+    private static native long nBuilderBuild(long nativeEngine, @NonNull Buffer buffer, int size, int shBandCount, int shadowQuality);
    private static native long nCreateInstance(long nativeMaterial);
    private static native long nCreateInstanceWithName(long nativeMaterial, @NonNull String name);
    private static native long nGetDefaultInstance(long nativeMaterial);
@@ -16,9 +16,6 @@

 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
@@ -29,7 +26,6 @@ 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
@@ -123,8 +119,6 @@ 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()
@@ -311,39 +305,10 @@ 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
@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.68.0
+VERSION_NAME=1.67.1

 POM_DESCRIPTION=Real-time physically based rendering engine for Android.

@@ -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=$(cat `dirname $0`/build/common/versions | grep GITHUB_NDK_VERSION | sed s/GITHUB_NDK_VERSION=//g | cut -f 1 -d ".")
+FILAMENT_NDK_VERSION=${FILAMENT_NDK_VERSION:-$(cat `dirname $0`/build/common/versions | grep GITHUB_NDK_VERSION | sed s/GITHUB_NDK_VERSION=//g | cut -f 1 -d ".")}

 # Internal variables
 ISSUE_CLEAN=false
@@ -18,8 +18,8 @@ if [[ "$GITHUB_WORKFLOW" ]]; then
    fi
 fi

-# Unless explicitly specified, NDK version will be selected as highest available version within same major release chain
-FILAMENT_NDK_VERSION=$(cat `dirname $0`/../common/versions | grep GITHUB_NDK_VERSION | sed s/GITHUB_NDK_VERSION=//g)
+# Unless explicitly specified, NDK version will be set to match exactly the required one
+FILAMENT_NDK_VERSION=${GITHUB_NDK_VERSION:-27.0.11718014}

 (! 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" &&
@@ -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=29.0.14206865
+GITHUB_NDK_VERSION=27.0.11718014
 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
@@ -75,7 +75,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 #   C_FLAGS += -Wl,-pie
 #   CXX_FLAGS += -lstdc++

-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIE -march=armv8-a -mtune=cortex-a78" CACHE STRING "Toolchain CFLAGS")
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIE -mcpu=cortex-a57" 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 "${CMAKE_SHARED_LINKER_FLAGS} -static-libstdc++" CACHE STRING "Toolchain LDFLAGS")
@@ -87,7 +87,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 # for hardfp: CFLAGS must have -mhard-float
 #             LDFLAGS must have -Wl,--no-warn-mismatch
 #
-set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS} -mthumb -march=armv7-a -mfloat-abi=softfp -mfpu=neon-vfpv4 -fPIE" CACHE STRING "Toolchain CFLAGS")
+set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS} -mthumb -march=armv7-a -mcpu=cortex-a15 -mfloat-abi=softfp -mfpu=neon-vfpv4 -fPIE" CACHE STRING "Toolchain CFLAGS")
 set(CMAKE_CXX_FLAGS "${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")
@@ -181,7 +181,7 @@ important for <code>matc</code> (material compiler).</p>
 }

 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.68.0'
+    implementation 'com.google.android.filament:filament-android:1.67.1'
 }
 </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; 1.68.0'
+<pre><code class="language-shell">pod 'Filament', '~&gt; 1.67.1'
 </code></pre>
 <h2 id="documentation"><a class="header" href="#documentation">Documentation</a></h2>
 <ul>
@@ -10,6 +10,6 @@ sortedcontainers==2.4.0
 trio==0.31.0
 trio-websocket==0.12.2
 typing_extensions==4.15.0
-urllib3==2.6.0
+urllib3==2.5.0
 websocket-client==1.9.0
 wsproto==1.2.0
@@ -206,8 +206,6 @@ 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
@@ -0,0 +1,63 @@
+/*
+ * 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
@@ -1713,8 +1713,6 @@ using FrameTimestamps = Platform::FrameTimestamps;

 using CompositorTiming = Platform::CompositorTiming;

-using AsynchronousMode = Platform::AsynchronousMode;
-
 } // namespace filament::backend

 template<> struct utils::EnableBitMaskOperators<filament::backend::ShaderStageFlags>
@@ -74,9 +74,6 @@ 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; }
@@ -255,28 +252,6 @@ 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.
@@ -346,11 +321,6 @@ public:
         *      - VulkanPlatform
         */
        bool vulkanEnableStagingBufferBypass = false;
-
-        /**
-         * Asynchronous mode for the engine. Defines how asynchronous operations are handled.
-         */
-        AsynchronousMode asynchronousMode = AsynchronousMode::NONE;
    };

    Platform() noexcept;
@@ -30,6 +30,8 @@

 #include <math/mat3.h>

+#include "AndroidNativeWindow.h"
+
 #include <chrono>

 #include <stddef.h>
@@ -22,6 +22,7 @@
 #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>
@@ -51,6 +51,7 @@
 namespace filament::backend {

 class BufferDescriptor;
+class BufferObjectStreamDescriptor;
 class CallbackHandler;
 class PixelBufferDescriptor;
 class Program;
@@ -392,6 +392,10 @@ 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,
@@ -61,16 +61,6 @@ 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) {
@@ -32,7 +32,6 @@ 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,
@@ -52,7 +51,6 @@ 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);
@@ -19,9 +19,6 @@

 #include <android/native_window.h>

-#include <utils/compiler.h>
-#include <utils/Logger.h>
-
 #include <cstddef>
 #include <cstdint>
 #include <limits>
@@ -39,23 +36,11 @@ bool AndroidSwapChainHelper::setPresentFrameId(
    int const status = NativeWindow::getNextFrameId(anw, &sysFrameId);
    if (status == 0) {
        std::lock_guard const lock(mLock);
-        // frameIds must be strictly monotonic, if that's not the case (i.e. the new frameId is
-        // less or equal to the last one in the map), we have to clear the map, because the
-        // map's find() assume sorted keys.
-        // 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();
+        auto const pos = mFrameIdToSystemFrameId.find(frameId);
+        if (pos && *pos != sysFrameId) {
+            // we're trying to associate the same frame id to a different frame!
+            return false;
        }
-
        // oldest entry is removed
        mFrameIdToSystemFrameId.insert(frameId, sysFrameId);
        return true;
@@ -146,12 +146,12 @@ void Platform::setBlobFunc(InsertBlobFunc&& insertBlob, RetrieveBlobFunc&& retri

 bool Platform::hasInsertBlobFunc() const noexcept {
    std::lock_guard<decltype(mMutex)> lock(mMutex);
-    return mInsertBlob && bool(*mInsertBlob);
+    return bool(mInsertBlob);
 }

 bool Platform::hasRetrieveBlobFunc() const noexcept {
    std::lock_guard<decltype(mMutex)> lock(mMutex);
-    return mRetrieveBlob && bool(*mRetrieveBlob);
+    return bool(mRetrieveBlob);
 }

 void Platform::insertBlob(void const* key, size_t keySize, void const* value, size_t valueSize) {
@@ -184,7 +184,7 @@ void Platform::setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noe

 bool Platform::hasDebugUpdateStatFunc() const noexcept {
    std::lock_guard<decltype(mMutex)> lock(mMutex);
-    return mDebugUpdateStat && bool(*mDebugUpdateStat);
+    return mDebugUpdateStat != nullptr;
 }

 void Platform::debugUpdateStat(const char* key, uint64_t intValue) {
@@ -165,7 +165,7 @@ public:
         size_t size, bool forceGpuBuffer = false);
    ~MetalBuffer();

-    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer; }
+    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer || mCpuBuffer; }

    MetalBuffer(const MetalBuffer& rhs) = delete;
    MetalBuffer& operator=(const MetalBuffer& rhs) = delete;
@@ -185,12 +185,14 @@ 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, it never returns
-     * nil.
+     * @return The MTLBuffer representing the current state of the buffer to bind, or nil if there
+     * is no device allocation.
     *
     */
    id<MTLBuffer> getGpuBufferForDraw() noexcept;

+    void* getCpuBuffer() const noexcept { return mCpuBuffer; }
+
    void setLabel(const utils::ImmutableCString& label) {
 #if FILAMENT_METAL_DEBUG_LABELS
        if (label.empty()) {
@@ -233,6 +235,7 @@ private:
    UploadStrategy mUploadStrategy;
    TrackedMetalBuffer mBuffer;
    size_t mBufferSize = 0;
+    void* mCpuBuffer = nullptr;
    MetalContext& mContext;
 };

@@ -39,30 +39,34 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
        mUploadStrategy = UploadStrategy::POOL;
    }

-    MTLResourceOptions options = MTLResourceStorageModePrivate;
+    // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
+    // buffer. Instead, we use immediate command encoder methods like setVertexBytes:length:atIndex:.
+    // This won't work for SSBOs, since they are read/write.

-    // The buffer will be memory mapped for write operations.
-    if (any(usage & BufferUsage::SHARED_WRITE_BIT)) {
-#if defined(FILAMENT_IOS) || defined(__arm64__) || defined(__aarch64__)
-        // iOS and Apple Silicon devices use UMA (Unified Memory Architecture), so we use Shared memory.
-        options = MTLResourceStorageModeShared;
-#else
-        // Intel Macs require Managed memory for CPU/GPU synchronization.
-        options = MTLResourceStorageModeManaged;
-#endif
+    /*
+    if (size <= 4 * 1024 && bindingType != BufferObjectBinding::SHADER_STORAGE &&
+            usage == BufferUsage::DYNAMIC && !forceGpuBuffer) {
+        mBuffer = nil;
+        mCpuBuffer = malloc(size);
+        return;
    }
+    */

+    // Otherwise, we allocate a private GPU buffer.
    {
        ScopedAllocationTimer timer("generic");
-        mBuffer = { [context.device newBufferWithLength:size options:options],
+        mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
            TrackedMetalBuffer::Type::GENERIC };
    }
-
    // mBuffer might fail to be allocated. Clients can check for this by calling
    // wasAllocationSuccessful().
 }

-MetalBuffer::~MetalBuffer() = default;
+MetalBuffer::~MetalBuffer() {
+    if (mCpuBuffer) {
+        free(mCpuBuffer);
+    }
+}

 void MetalBuffer::copyIntoBuffer(
        void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
@@ -79,6 +83,12 @@ 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));
@@ -96,6 +106,11 @@ 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();
 }
@@ -156,6 +171,41 @@ 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(
@@ -61,8 +61,6 @@ public:

    MetalContext* getContext() { return mContext; }

-    using DriverBase::scheduleDestroy;
-
 private:

    friend class MetalSwapChain;
@@ -1144,12 +1144,8 @@ 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.
-    // On macOS, framebuffer fetch requires MSL 2.3, which is only available with macOS 11.0.
-    if (@available(macOS 11.0, *)) {
-        return mContext->highestSupportedGpuFamily.apple >= 1 &&
-               mContext->device.readWriteTextureSupport;
-    }
-    return false;
+    return mContext->highestSupportedGpuFamily.apple >= 1 &&
+            mContext->device.readWriteTextureSupport;
 }

 bool MetalDriver::isFrameBufferFetchMultiSampleSupported() {
@@ -1290,6 +1286,10 @@ 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))
@@ -2275,10 +2275,7 @@ MemoryMappedBufferHandle MetalDriver::mapBufferS() noexcept {
 void MetalDriver::mapBufferR(MemoryMappedBufferHandle mmbh,
        BufferObjectHandle boh, size_t offset,
        size_t size, MapBufferAccessFlags access, utils::ImmutableCString&& tag) {
-    assert_invariant(boh);
-    MetalBufferObject* bo = mHandleAllocator.handle_cast<MetalBufferObject*>(boh);
-    assert_invariant(bo);
-    construct_handle<MetalMemoryMappedBuffer>(mmbh, bo, offset, size, access);
+    construct_handle<MetalMemoryMappedBuffer>(mmbh, boh, offset, size, access);
    mHandleAllocator.associateTagToHandle(mmbh.getId(), std::move(tag));
 }

@@ -2286,16 +2283,21 @@ 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);
-    mmb->copy(*this, offset, std::move(data));
+    auto mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
+
+    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
@@ -569,22 +569,14 @@ 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{};
-    struct {
-        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;
+    uint32_t size = 0;
+    uint32_t offset = 0;
 };

 } // namespace backend
@@ -1664,44 +1664,5 @@ 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
@@ -111,22 +111,9 @@ 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
@@ -291,6 +291,8 @@ 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));
@@ -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 = vao != &mDefaultVAO &&
+        bool const recreateVaoName = p != &mDefaultVAO &&
                ((vao->vao[contextIndex] == 0) ||
                        (vao->nameVersion != state.age && contextIndex == 1));
        if (UTILS_UNLIKELY(recreateVaoName)) {
@@ -39,6 +39,7 @@
 #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"
@@ -2836,6 +2837,38 @@ 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()
@@ -2849,6 +2882,19 @@ 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);
@@ -36,6 +36,7 @@
 #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"
@@ -393,6 +394,8 @@ 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;
@@ -37,7 +37,7 @@
 */


-#if defined(__ANDROID__) || defined(FILAMENT_USE_EXTERNAL_GLES3) || defined(__EMSCRIPTEN__) || defined(FILAMENT_SUPPORTS_EGL_ON_LINUX)
+#if defined(__ANDROID__) || defined(FILAMENT_USE_EXTERNAL_GLES3) || defined(__EMSCRIPTEN__)

    #if defined(__EMSCRIPTEN__)
    #   include <GLES3/gl3.h>
@@ -106,8 +106,6 @@ struct PlatformEGLAndroid::SwapChainEGLAndroid : public SwapChainEGL {
    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{};
 };
@@ -178,30 +176,20 @@ void PlatformEGLAndroid::beginFrame(
        int64_t const monotonic_clock_ns,
        int64_t refreshIntervalNs,
        uint32_t const frameId) noexcept {
-    // 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 performance hint (since we wouldn't get
-    // the right timing anyway as well as the frame info timing collections)
+    // associate the user frameid with the system frame id
+    setPresentFrameId(mCurrentDrawSwapChain, frameId);

-    // 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 (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;

-            if (mCurrentDrawSwapChain) {
-                // mCurrentDrawSwapChain could be null if we're called from renderStandaloneView
-                CompositorTiming compositorTiming{};
-                bool const hasCompositorTiming =
-                        queryCompositorTiming(mCurrentDrawSwapChain, &compositorTiming);
-                if (hasCompositorTiming && compositorTiming.compositeInterval > 0) {
-                    refreshIntervalNs = compositorTiming.compositeInterval;
-                }
+            CompositorTiming compositorTiming{};
+            bool const hasCompositorTiming =
+                    queryCompositorTiming(mCurrentDrawSwapChain, &compositorTiming);
+            if (hasCompositorTiming && compositorTiming.compositeInterval > 0) {
+                refreshIntervalNs = compositorTiming.compositeInterval;
            }
        }
        mStartTimeOfActualWork = clock::time_point(std::chrono::nanoseconds(monotonic_clock_ns));
@@ -240,9 +228,9 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
                    "eglGetNativeClientBufferANDROID"));

    if (ext.egl.ANDROID_presentation_time) {
-        eglPresentationTimeANDROID =
-                PFNEGLPRESENTATIONTIMEANDROIDPROC(eglGetProcAddress(
-                        "eglPresentationTimeANDROID"));
+        eglGetNativeClientBufferANDROID =
+                PFNEGLGETNATIVECLIENTBUFFERANDROIDPROC(eglGetProcAddress(
+                        "eglGetNativeClientBufferANDROID"));
    }

    if (ext.egl.ANDROID_get_frame_timestamps) {
@@ -301,21 +289,11 @@ bool PlatformEGLAndroid::queryCompositorTiming(SwapChain const* swapchain,
    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 true;
+            return false;
        }

        std::array<EGLnsecsANDROID, 3> values;
@@ -326,16 +304,26 @@ bool PlatformEGLAndroid::queryCompositorTiming(SwapChain const* swapchain,
        };
        EGLBoolean const success = eglGetCompositorTimingANDROID(getEglDisplay(), sur,
                names.size(), names.data(), values.data());
-        if (UTILS_UNLIKELY(!success)) {
-            // reset current error to EGL_SUCCESS
-            eglGetError();
-        } else {
-            outCompositorTiming->compositeDeadline = values[0];
-            outCompositorTiming->compositeInterval = values[1];
-            outCompositorTiming->compositeToPresentLatency = values[2];
+        if (!success) {
+            return false;
        }
+        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,
@@ -360,10 +348,6 @@ 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) {
@@ -384,9 +368,7 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
        };
        EGLBoolean const success = eglGetFrameTimestampsANDROID(getEglDisplay(), sur, hwFrameId,
                names.size(), names.data(), values.data());
-        if (UTILS_UNLIKELY(!success)) {
-            // reset current error to EGL_SUCCESS
-            eglGetError();
+        if (!success) {
            return false;
        }
        outFrameTimestamps->requestedPresentTime = values[0];
@@ -400,44 +382,28 @@ 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;
 }

@@ -450,10 +416,6 @@ 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;
    }
@@ -763,6 +725,8 @@ 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);
    }
 }

@@ -16,7 +16,7 @@

 #include <backend/platforms/PlatformEGLHeadless.h>

-#include <bluegl/BlueGL.h>
+#include "opengl/GLUtils.h"

 #include <EGL/egl.h>
 #include <EGL/eglext.h>
@@ -35,31 +35,16 @@ PlatformEGLHeadless::PlatformEGLHeadless() noexcept
 }

 bool PlatformEGLHeadless::isOpenGL() const noexcept {
-#if defined(BACKEND_OPENGL_VERSION_GL)
-    return true;
-#else
-    return false;
-#endif  // defined(BACKEND_OPENGL_VERSION_GL)
+    return  true;
 }

 backend::Driver* PlatformEGLHeadless::createDriver(void* sharedContext,
        const Platform::DriverConfig& driverConfig) {
-    auto bindApiHelper = [](EGLenum api, const char* errorString) -> bool {
-        EGLBoolean bindAPI = eglBindAPI(api);
-        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)) {
+    EGLBoolean bindAPI = eglBindAPI(EGL_OPENGL_API);
+    if (UTILS_UNLIKELY(!bindAPI)) {
+        LOG(ERROR) << "eglBindAPI EGL_OPENGL_API failed";
        return nullptr;
    }
-
    int bindBlueGL = bluegl::bind();
    if (UTILS_UNLIKELY(bindBlueGL != 0)) {
        LOG(ERROR) << "bluegl bind failed";
@@ -15,6 +15,7 @@
 */

 #include <backend/BufferDescriptor.h>
+#include <backend/BufferObjectStreamDescriptor.h>
 #include <backend/DescriptorSetOffsetArray.h>
 #include <backend/DriverEnums.h>
 #include <backend/PipelineState.h>
@@ -433,6 +434,10 @@ 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
@@ -62,9 +62,8 @@ 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 && mStagingBufferBypassEnabled) || isStaticOrShared) &&
-            isMemcopyable;
+    bool const useMemcpy = (isUniformAndAvailable || isStaticOrShared) && isMemcopyable &&
+                           mStagingBufferBypassEnabled;
    if (useMemcpy) {
        char* dest = static_cast<char*>(mBuffer->getGpuBuffer()->allocationInfo.pMappedData) +
                     byteOffset;
@@ -19,7 +19,6 @@
 #include "VulkanCommands.h"
 #include "VulkanHandles.h"
 #include "VulkanConstants.h"
-#include "utils/compiler.h"

 #include <utils/FixedCapacityVector.h>
 #include <utils/Panic.h>
@@ -32,18 +31,6 @@ 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;

@@ -317,7 +304,8 @@ void VulkanDescriptorSetCache::commit(VulkanCommandBuffer* commands,
        // This code actually binds the descriptor sets.
        auto set = updateSets[index];
        VkCommandBuffer const cmdbuffer = commands->buffer();
-        VkDescriptorSet const vkset = set->getVkSet();
+        VkDescriptorSet vkset = useExternalSamplers[index] ? set->getExternalSamplerVkSet() :
+            set->getVkSet();
        commands->acquire(set);
        vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, index,
                1, &vkset, set->uniqueDynamicUboCount, set->getOffsets()->data());
@@ -353,39 +341,17 @@ 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::updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
-        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();
+void VulkanDescriptorSetCache::updateSamplerImpl(VkDescriptorSet vkset, uint8_t binding,
+        fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept {
    VkImageSubresourceRange range = texture->getPrimaryViewRange();
    VkImageViewType const expectedType = texture->getViewType();
    if (any(texture->usage & TextureUsage::DEPTH_ATTACHMENT) &&
@@ -410,6 +376,19 @@ void VulkanDescriptorSetCache::updateSampler(fvkmemory::resource_ptr<VulkanDescr
        .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);
 }

@@ -426,7 +405,7 @@ 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,
+            mResourceManager, handle, layout->bitmask.dynamicUbo, layout->count.dynamicUbo,
            [vkSet, count, vklayout, this](
                    VulkanDescriptorSet*) { this->manualRecycle(count, vklayout, vkSet); },
            vkSet);
@@ -449,22 +428,4 @@ 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
@@ -55,8 +55,7 @@ public:
            VkDeviceSize size) noexcept;

    void updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
-            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler,
-            VkDescriptorSetLayout externalSamplerLayout = VK_NULL_HANDLE) noexcept;
+            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept;

    void updateSamplerForExternalSamplerSet(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
            fvkmemory::resource_ptr<VulkanTexture> texture) noexcept;
@@ -91,8 +90,8 @@ public:
    void resetCachedState() noexcept { mLastBoundInfo = {}; }

 private:
-    void copySet(VkDescriptorSet srcSet, VkDescriptorSet destSet,
-            fvkutils::SamplerBitmask copyBindings) const;
+    void updateSamplerImpl(VkDescriptorSet set, uint8_t binding,
+            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept;

    class DescriptorInfinitePool;

@@ -31,7 +31,6 @@
 #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>
@@ -245,7 +244,6 @@ 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) {
@@ -422,14 +420,12 @@ void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
        int64_t refreshIntervalNs, uint32_t frameId) {
    FVK_PROFILE_MARKER(PROFILE_NAME_BEGINFRAME);

-    // 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
+    if (mCurrentSwapChain) { // This should be guaranteed
        mPlatform->setPresentFrameId(mCurrentSwapChain->swapChain, frameId);
    }

    // Check if any command have finished and reset all its used resources. The resources
-    // won't be destroyed but their reference count will be decreased if the command is already
+    // wont be destroyed but their reference count will decreased if the command is already
    // completed.
    //
    // This will let us check if any VulkanBuffer is currently in flight or not.
@@ -483,9 +479,6 @@ 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,
@@ -493,7 +486,6 @@ void VulkanDriver::updateDescriptorSetTexture(
        VkSampler const vksampler = mSamplerCache.getSampler(cacheParams);
        mDescriptorSetCache.updateSampler(set, binding, texture, vksampler);
        mExternalImageManager.clearTextureBinding(set, binding);
-        mStreamedImageManager.unbindStreamedTexture(set, binding);
    }
 }

@@ -505,7 +497,10 @@ void VulkanDriver::copyToMemoryMappedBuffer(MemoryMappedBufferHandle mmbh, size_
    assert_invariant(any(mmb->access & MapBufferAccessFlags::WRITE_BIT));
    assert_invariant(offset + data.size <= mmb->size);

-    // On UMA systems, this is a memcpy into the memory-mapped buffer, otherwise use a staging buffer.
+    // 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);
 }

@@ -1113,12 +1108,6 @@ 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) {
@@ -1156,38 +1145,18 @@ 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) {
-    // @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;
+    return {};
 }

 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) {
@@ -1195,70 +1164,6 @@ 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) {
@@ -1554,13 +1459,17 @@ 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));
@@ -1628,12 +1537,6 @@ 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) {
@@ -34,14 +34,12 @@
 #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"
@@ -159,8 +157,6 @@ 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
@@ -206,10 +202,6 @@ private:
    bool const mIsSRGBSwapChainSupported;
    bool const mIsMSAASwapChainSupported;
    backend::StereoscopicType const mStereoscopicType;
-
-    // setAcquiredImage is a DECL_DRIVER_API_SYNCHRONOUS_N which means we don't necessarily have the
-    // 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
@@ -31,6 +31,9 @@ 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);
@@ -47,6 +50,31 @@ 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,
@@ -113,9 +141,6 @@ 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;
@@ -147,14 +172,38 @@ 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) {
-        // 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);
+        mDescriptorSetCache->updateSamplerForExternalSamplerSet(set, binding, image);
    }
 }

@@ -212,25 +261,13 @@ void VulkanExternalImageManager::bindExternallySampledTexture(
        fvkmemory::resource_ptr<VulkanTexture> image, SamplerParams samplerParams) {
    // Should we do duplicate validation here?
    auto& imageData = findImage(mImages, image);
-    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;
-    }
+    mSetBindings.push_back({ bindingPoint, imageData.image, set, samplerParams });
 }

 void VulkanExternalImageManager::addExternallySampledTexture(
       fvkmemory::resource_ptr<VulkanTexture> image,
        Platform::ExternalImageHandleRef platformHandleRef) {
-    // 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 });
+    mImages.push_back({ image, platformHandleRef, false });
 }

 void VulkanExternalImageManager::removeExternallySampledTexture(
@@ -31,7 +31,7 @@ class VulkanSamplerCache;
 class VulkanDescriptorSetLayoutCache;
 class VulkanDescriptorSetCache;

-// Manages the logic of external images and their quirks wrt Vulkan.
+// Manages the logic of external images and their quirks wrt Vulikan.
 class VulkanExternalImageManager {
 public:

@@ -40,11 +40,6 @@ 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:
@@ -197,70 +192,15 @@ VulkanDescriptorSetLayout::Bitmask VulkanDescriptorSetLayout::Bitmask::fromLayou
    return fromBackendLayout(layout);
 }

-VulkanDescriptorSet::~VulkanDescriptorSet() {
-    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));
-}
-
+// This method 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.
 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) {
@@ -28,7 +28,6 @@
 #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"

@@ -58,8 +57,6 @@ inline uint8_t collapsedCount(Bitmask const& mask) {

 } // anonymous namespace

-class VulkanDescriptorSetCache;
-
 struct VulkanBufferObject;

 struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Resource {
@@ -162,14 +159,38 @@ public:
    using OnRecycle = std::function<void(VulkanDescriptorSet*)>;

    VulkanDescriptorSet(
-            fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout,
-            OnRecycle&& onRecycleFn, VkDescriptorSet vkSet);
+            fvkutils::UniformBufferBitmask const& dynamicUboMask,
+            uint8_t uniqueDynamicUboCount,
+            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 mSets[mCurrentSetIndex].vkSet;
+        return mVkSet;
+    }
+
+    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 {
@@ -190,33 +211,17 @@ public:

    void referencedBy(VulkanCommandBuffer& commands);

-    bool isBound() const {
-        return bool(mSets[mCurrentSetIndex].fenceStatus);
-    }
-
-    fvkmemory::resource_ptr<VulkanDescriptorSetLayout> getLayout() const { return mLayout; }
-
-    fvkutils::UniformBufferBitmask const& dynamicUboMask;
+    fvkutils::UniformBufferBitmask const dynamicUboMask;
    uint8_t const uniqueDynamicUboCount;

 private:
-    friend class VulkanDescriptorSetCache;
+    VkDescriptorSet const mVkSet;
+    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;
-    uint8_t mCurrentSetIndex;
-    std::vector<InternalVkSet> mSets;
+    OnRecycle mOnRecycleFn;
+    OnRecycle mOnRecycleExternalSamplerFn;
    fvkutils::UniformBufferBitmask mUboMask;
 };

@@ -191,12 +191,9 @@ 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 = enableDepthTest ? VK_TRUE : VK_FALSE,
+        .depthTestEnable = VK_TRUE,
        .depthWriteEnable = raster.depthWriteEnable,
        .depthCompareOp = fvkutils::getCompareOp(raster.depthCompareOp),
        .depthBoundsTestEnable = VK_FALSE,
@@ -1,60 +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.
- */
-
-#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
@@ -1,64 +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_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
@@ -19,7 +19,6 @@
 #include "VulkanCommands.h"
 #include "VulkanTexture.h"

-#include <utils/debug.h>
 #include <utils/FixedCapacityVector.h>
 #include <utils/Panic.h>

@@ -106,11 +105,6 @@ 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{
@@ -167,15 +161,7 @@ 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);
@@ -36,43 +36,6 @@ 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,
@@ -111,8 +74,6 @@ 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;
@@ -144,7 +105,6 @@ private:
    using ImageViewHash = utils::hash::MurmurHashFn<ImageViewKey>;
    std::unordered_map<ImageViewKey, VkImageView, ImageViewHash> mCachedImageViews;

-
    friend struct VulkanTexture;
 };

@@ -222,14 +182,6 @@ 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;
    }
@@ -41,7 +41,6 @@ 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 {
@@ -105,9 +104,6 @@ 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;
 }

@@ -153,8 +149,6 @@ std::string_view getTypeStr(ResourceType type) {
            return "VulkanMemoryMappedBuffer";
        case ResourceType::SEMAPHORE:
            return "Semaphore";
-        case ResourceType::STREAM:
-            return "VulkanStream";
        case ResourceType::UNDEFINED_TYPE:
            return "";
    }
@@ -55,8 +55,7 @@ enum class ResourceType : uint8_t {
    SYNC = 17,
    MEMORY_MAPPED_BUFFER = 18,
    SEMAPHORE = 19,
-    STREAM = 20,
-    UNDEFINED_TYPE = 21,    // Must be the last enum because we use it for iterating over the enums.
+    UNDEFINED_TYPE = 20,    // Must be the last enum because we use it for iterating over the enums.
 };

 template<typename D>
@@ -123,9 +123,6 @@ 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;
    }
@@ -541,15 +541,9 @@ bool VulkanPlatformAndroid::queryCompositorTiming(SwapChain const* swapchain,
    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);
-    vulkanSwapchain->queryCompositorTiming(outCompositorTiming);
-    return true;
+    return vulkanSwapchain->queryCompositorTiming(outCompositorTiming);
 }

 bool VulkanPlatformAndroid::setPresentFrameId(SwapChain const* swapchain, uint64_t frameId) noexcept {
@@ -361,15 +361,13 @@ bool VulkanPlatformSurfaceSwapChain::queryCompositorTiming(
        CompositorTiming* outCompositorTiming) const {
 #ifdef __ANDROID__
    // fallback to private APIs
-    if (UTILS_VERY_LIKELY(mNativeWindow)) {
-        int const status = NativeWindow::getCompositorTiming(
-                static_cast<ANativeWindow*>(mNativeWindow),
-                &outCompositorTiming->compositeDeadline,
-                &outCompositorTiming->compositeInterval,
-                &outCompositorTiming->compositeToPresentLatency);
-        if (status == 0) {
-            return true;
-        }
+    int const status = NativeWindow::getCompositorTiming(
+            static_cast<ANativeWindow*>(mNativeWindow),
+            &outCompositorTiming->compositeDeadline,
+            &outCompositorTiming->compositeInterval,
+            &outCompositorTiming->compositeToPresentLatency);
+    if (status == 0) {
+        return true;
    }
 #endif
    return VulkanPlatformSwapChainBase::queryCompositorTiming(outCompositorTiming);
@@ -18,12 +18,6 @@

 #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>
@@ -240,19 +234,6 @@ 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) {
@@ -18,7 +18,6 @@
 #define TNT_FILAMENT_BACKEND_VULKAN_UTILS_IMAGE_H

 #include <backend/DriverEnums.h>
-#include <backend/platforms/VulkanPlatform.h>

 #include <utils/Log.h>

@@ -107,9 +106,6 @@ 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
@@ -30,6 +30,7 @@

 #include <cstdint>
 #include <cstring>
+#include <iostream>

 namespace filament::backend {

@@ -80,33 +81,54 @@ void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
    // This may have some performance implications. That should be investigated later.
    assert_invariant(mBuffer.GetUsage() & wgpu::BufferUsage::CopyDst);

-    // Calculate some alignment related sizes
+    // // Calculate some alignment related sizes
    const size_t remainder = bufferDescriptor.size % FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS;
    const size_t mainBulk = bufferDescriptor.size - remainder;
    const size_t stagingBufferSize =
            remainder == 0 ? bufferDescriptor.size : mainBulk + FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS;

-    wgpu::Buffer stagingBuffer = webGPUStagePool->acquireBuffer(stagingBufferSize,
-            webGPUQueueManager->getLatestSubmissionState());
+    Stage stage = webGPUStagePool->acquireBuffer(stagingBufferSize);

-    void* mappedRange = stagingBuffer.GetMappedRange();
-    assert_invariant(mappedRange);
+    std::string mappedRangeIsNull = stage.mappedRange
+            ? "no"
+            : "yes";
+    std::cout << "Run Yu: got mapped range on the staging buffer with size "
+              << stage.buffer.GetSize() << " and it is null? " <<  mappedRangeIsNull << std::endl;
+    memcpy(stage.mappedRange, bufferDescriptor.buffer, bufferDescriptor.size);

-    memcpy(mappedRange, bufferDescriptor.buffer, bufferDescriptor.size);
-
-    // Make sure the padded memory is set to 0 to have deterministic behaviors
-    if (remainder != 0) {
-        uint8_t* paddingStart = static_cast<uint8_t*>(mappedRange) + bufferDescriptor.size;
-        memset(paddingStart, 0, FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS - remainder);
-    }
-
-    stagingBuffer.Unmap();
+    stage.buffer.Unmap();

+    std::cout << "Run Yu: about to issue copy command with actual staging buffer of size "
+              << stage.buffer.GetSize() << ", and computed size of " << stagingBufferSize
+              << ". The mBuffer size is " << mBuffer.GetSize() << std::endl;
    // Copy the staging buffer contents to the destination buffer.
-    webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(stagingBuffer, 0, mBuffer,
+    webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(stage.buffer, 0, mBuffer,
            byteOffset,
            remainder == 0 ? bufferDescriptor.size
                           : mainBulk + FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS);
+    webGPUQueueManager->flush();
+
+    struct UserData final {
+        wgpu::Buffer stagingBuffer;
+        WebGPUStagePool* webGPUStagePool;
+    };
+    auto userData = std::make_unique<UserData>(
+            UserData{ .stagingBuffer = stage.buffer, .webGPUStagePool = webGPUStagePool });
+    stage.buffer.MapAsync(wgpu::MapMode::Write, 0, stagingBufferSize,
+            wgpu::CallbackMode::AllowSpontaneous,
+            [data = std::move(userData)](wgpu::MapAsyncStatus status, const char* message) {
+                if (UTILS_LIKELY(status == wgpu::MapAsyncStatus::Success)) {
+                    std::cout << "Run Yu: successfully mapped a buffer with size "
+                              << data->stagingBuffer.GetSize() << std::endl;
+                    void* mappedRange = data->stagingBuffer.GetMappedRange();
+                    if (!mappedRange) {
+                        std::cout << "Run Yu: MAPPED RANGE IS NULL RIGHT AWAY!!\n";
+                    }
+                    data->webGPUStagePool->addBufferToPool(data->stagingBuffer, mappedRange);
+                } else {
+                    std::cout << "Run Yu: MAPPING UNSUCCESSFUL!!\n";
+                }
+            });
 }

 } // namespace filament::backend
@@ -178,9 +178,6 @@ 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.
@@ -655,6 +652,11 @@ 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
 }
@@ -64,7 +64,6 @@ wgpu::CommandEncoder WebGPUQueueManager::getCommandEncoder() {
        };
        mCommandEncoder = mDevice.CreateCommandEncoder(&commandEncoderDescriptor);
        ASSERT_POSTCONDITION(mCommandEncoder, "Failed to create command encoder.");
-        mLatestSubmissionState = std::make_shared<WebGPUSubmissionState>();
    }
    return mCommandEncoder;
 }
@@ -95,6 +94,8 @@ void WebGPUQueueManager::submit() {
        return;
    }

+    mLatestSubmissionState = std::make_shared<WebGPUSubmissionState>();
+
    wgpu::CommandBufferDescriptor commandBufferDescriptor{
        .label = "Filament Command Buffer",
    };
@@ -17,7 +17,8 @@
 #include "WebGPUStagePool.h"

 #include "WebGPUConstants.h"
-#include "WebGPUQueueManager.h"
+
+#include <iostream>

 namespace filament::backend {

@@ -25,62 +26,55 @@ 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;
+Stage WebGPUStagePool::acquireBuffer(size_t requiredSize) {
+    std::cout << "Run Yu: required size in acquireBuffer: " << requiredSize << std::endl;
+    std::cout << "Run Yu: the pool size is " << mBuffers.size() << std::endl;
    {
        std::lock_guard<std::mutex> lock(mMutex);
        auto iter = mBuffers.lower_bound(requiredSize);
        if (iter != mBuffers.end()) {
-            buffer = iter->second;
+            const Stage& fromPool = iter->second;
+            std::cout << "Run Yu: found buffer in the pool with size " << fromPool.buffer.GetSize()
+                      << std::endl;
+            if (fromPool.buffer.GetMapState() != wgpu::BufferMapState::Mapped) {
+                std::cout << "Run Yu: buffer from pool is not mapped!!" << std::endl;
+            }
+
+            Stage result{ .buffer = fromPool.buffer, .mappedRange = fromPool.mappedRange };
            mBuffers.erase(iter);
+            return result;
        }
    }
-    if (!buffer.Get()) {
-        buffer = createNewBuffer(requiredSize);
-    }
-    mInProgress.push_back({submissionState, buffer});
-    return buffer;
+    wgpu::Buffer newBuffer = createNewBuffer(requiredSize);
+    return { .buffer = newBuffer, .mappedRange = newBuffer.GetMappedRange() };
 }

-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::addBufferToPool(wgpu::Buffer buffer, void* mappedRange) {
+    std::lock_guard<std::mutex> lock(mMutex);
+    std::cout << "Run Yu: adding buffer to the pool with size " << buffer.GetSize() << std::endl;
+    Stage stage {.buffer = buffer, .mappedRange = mappedRange};
+    mBuffers.emplace(buffer.GetSize(), stage);
+    std::cout << "Run Yu: added buffer to the pool with size " << buffer.GetSize() << std::endl;

-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});
+    bool allMapped = true;
+    for (const auto& pair : mBuffers) {
+        auto state = pair.second.buffer.GetMapState();
+        if (state != wgpu::BufferMapState::Mapped) {
+            allMapped = false;
+            std::cout << "Run Yu: the buffer with size " << pair.second.buffer.GetSize()
+                      << " is not mapped but somehow was added to the pool, its state is "
+                      << static_cast<int>(state) << std::endl;
        }
    }
-    std::swap(mInProgress, stillInProgress);
+    if (!allMapped) {
+        std::cout << "Run Yu: found buffers that are not mapped\n";
+    } else {
+        std::cout << "Run Yu: all buffers are mapped\n";
+    }
 }

 wgpu::Buffer WebGPUStagePool::createNewBuffer(size_t bufferSize) {
+    std::cout << "Run Yu: creating new buffer with size " << bufferSize << std::endl;
    wgpu::BufferDescriptor descriptor{
        .label = "Filament WebGPU Staging Buffer",
        .usage = wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc,
@@ -24,27 +24,26 @@

 namespace filament::backend {

-struct WebGPUSubmissionState;
+struct Stage {
+    wgpu::Buffer buffer;
+    void* mappedRange;
+};

 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();
-
+    Stage acquireBuffer(size_t requiredSize);
+    void addBufferToPool(wgpu::Buffer buffer, void* mappedRange);
 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;
+    std::multimap<uint32_t, Stage> mBuffers;
+    mutable std::mutex mMutex;

    wgpu::Device mDevice;
 };

-} // namespace filament::backend
+}

 #endif // TNT_FILAMENT_BACKEND_WEBGPUSTAGEPOOL_H
@@ -192,7 +192,6 @@ 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
@@ -429,11 +428,6 @@ public:
         * initial memory usage.
         */
        uint32_t sharedUboInitialSizeInBytes = 256 * 64;
-
-        /**
-         * Asynchronous mode for the engine. Defines how asynchronous operations are handled.
-         */
-        AsynchronousMode asynchronousMode = AsynchronousMode::NONE;
    };


@@ -748,14 +742,6 @@ 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.
     *
@@ -69,20 +69,6 @@ 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.
     */
@@ -179,13 +165,6 @@ 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.
         *
@@ -113,10 +113,8 @@ 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;
@@ -328,8 +326,6 @@ 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
@@ -341,8 +337,6 @@ 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()
@@ -455,10 +455,6 @@ 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();
 }
@@ -101,24 +101,11 @@ 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())) {
-        FrameInfoImpl& frameInfo = history.back();
-        if (frameInfo.ready.load(std::memory_order_relaxed)) {
-            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;
+        if (!mDisableGpuFrameComplete) {
+            assert_invariant(history.back().fence);
+            driver.destroyFence(std::move(history.back().fence));
        }
+        history.pop_back();
    }

    // create a new entry
@@ -206,14 +193,6 @@ 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();

@@ -95,8 +95,33 @@ struct FrameInfoImpl : public details::FrameInfo {
        assert_invariant(!fence);
    }

-    FrameInfoImpl(FrameInfoImpl& rhs) noexcept = delete;
-    FrameInfoImpl& operator=(FrameInfoImpl& rhs) noexcept = delete;
+    FrameInfoImpl(FrameInfoImpl&& rhs) noexcept : 
+        details::FrameInfo(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(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>
@@ -315,7 +340,6 @@ private:
    FrameHistoryQueue mFrameTimeHistory{};
    utils::AsyncJobQueue mJobQueue;
    FSwapChain* mLastSeenSwapChain = nullptr;
-    bool mLastBeginFrameSkipped = false;
    bool const mHasTimerQueries = false;
    bool const mDisableGpuFrameComplete = false;
 };
@@ -98,7 +98,7 @@ using namespace filaflat;
 namespace {

 backend::Platform::DriverConfig getDriverConfig(FEngine* instance) {
-    Platform::DriverConfig driverConfig {
+    return {
        .handleArenaSize = instance->getRequestedDriverHandleArenaSize(),
        .metalUploadBufferSizeBytes = instance->getConfig().metalUploadBufferSizeBytes,
        .disableParallelShaderCompile = instance->features.backend.disable_parallel_shader_compile,
@@ -116,11 +116,7 @@ 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
@@ -731,6 +727,7 @@ 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);
            }
        });
@@ -994,9 +991,10 @@ FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
    return p;
 }

-FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
-        const char* name) noexcept {
-    FMaterialInstance* p = mHeapAllocator.make<FMaterialInstance>(*this, material, name);
+FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material, const char* name,
+        UboBatchingMode batchingMode) noexcept {
+    FMaterialInstance* p =
+            mHeapAllocator.make<FMaterialInstance>(*this, material, name, batchingMode);
    if (UTILS_LIKELY(p)) {
        auto pos = mMaterialInstances.emplace(material, "MaterialInstance");
        pos.first->second.insert(p);
@@ -1505,11 +1503,6 @@ 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();
@@ -216,8 +216,6 @@ public:
        return getDriver().isStereoSupported();
    }

-    bool isAsynchronousOperationSupported() const noexcept;
-
    static size_t getMaxStereoscopicEyes() noexcept {
        return CONFIG_MAX_STEREOSCOPIC_EYES;
    }
@@ -347,10 +345,21 @@ 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) noexcept;
+    FMaterialInstance* createMaterialInstance(const FMaterial* material, const char* name,
+            UboBatchingMode batchingMode) noexcept;

    FScene* createScene() noexcept;
    FView* createView() noexcept;
@@ -771,7 +780,7 @@ public:
                bool assert_texture_can_generate_mipmap = CORRECTNESS_ASSERTION_DEFAULT;
            } debug;
            struct {
-                bool disable_gpu_frame_complete_metric = true;
+                bool disable_gpu_frame_complete_metric = false;
            } frame_info;
        } engine;
        struct {
@@ -788,7 +797,6 @@ 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;
@@ -809,9 +817,6 @@ 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 },
@@ -70,19 +70,10 @@

 namespace filament {

-namespace {
-
 using namespace backend;
 using namespace filaflat;
 using namespace utils;
-using UboBatchingMode = Material::UboBatchingMode;
-
-bool shouldEnableBatching(FEngine& engine, UboBatchingMode batchingMode, MaterialDomain domain) {
-    return batchingMode != UboBatchingMode::DISABLED && engine.isUboBatchingEnabled() &&
-           domain == MaterialDomain::SURFACE;
-}
-
-} // anonymous namespace
+using UboBatchingMode = FEngine::UboBatchingMode;

 struct Material::BuilderDetails {
    const void* mPayload = nullptr;
@@ -90,7 +81,6 @@ 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>,
@@ -125,11 +115,6 @@ 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";
@@ -163,12 +148,8 @@ 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);

@@ -301,14 +282,14 @@ 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);
+        return mEngine.createMaterialInstance(this, name, UboBatchingMode::DEFAULT);
    }
 }

 FMaterialInstance* FMaterial::getDefaultInstance() noexcept {
    if (UTILS_UNLIKELY(!mDefaultMaterialInstance)) {
        mDefaultMaterialInstance =
-                mEngine.createMaterialInstance(this, mDefinition.name.c_str());
+                mEngine.createMaterialInstance(this, mDefinition.name.c_str(), UboBatchingMode::DEFAULT);
        mDefaultMaterialInstance->setDefaultInstance(true);
    }
    return mDefaultMaterialInstance;
@@ -248,10 +248,6 @@ public:
        return mDefinition.perViewLayoutIndex;
    }

-    bool useUboBatching() const noexcept {
-        return mUseUboBatching;
-    }
-
 #if FILAMENT_ENABLE_MATDBG
    void applyPendingEdits() noexcept;

@@ -314,8 +310,6 @@ private:

    bool mIsDefaultMaterial = false;

-    bool mUseUboBatching = false;
-
    // reserve some space to construct the default material instance
    mutable FMaterialInstance* mDefaultMaterialInstance = nullptr;

@@ -58,8 +58,18 @@ using namespace utils;
 namespace filament {

 using namespace backend;
+using UboBatchingMode = FEngine::UboBatchingMode;

-FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name) noexcept
+bool shouldEnableBatching(FEngine& engine, UboBatchingMode batchingMode, MaterialDomain domain) {
+    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),
@@ -70,9 +80,12 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
          mHasScissor(false),
          mIsDoubleSided(false),
          mIsDefaultInstance(false),
-          mUseUboBatching(material->useUboBatching()),
+          mUseUboBatching(
+                  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
@@ -147,6 +160,7 @@ 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();

@@ -209,6 +223,31 @@ void FMaterialInstance::terminate(FEngine& engine) {
    }
 }

+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());
@@ -216,9 +255,10 @@ void FMaterialInstance::commit(FEngine& engine) const {
 }

 void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManager) const {
-    if (mUniforms.isDirty()) {
+    if (mUniforms.isDirty() || mHasStreamUniformAssociations) {
        mUniforms.clean();
-        if (isUsingUboBatching()) {
+        if (mUseUboBatching) {
+            assert_invariant(uboManager != nullptr);
            if (!BufferAllocator::isValid(getAllocationId())) {
                // The allocation hasn't happened yet, return.
                return;
@@ -249,7 +289,7 @@ void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManage
    // TODO: eventually we should remove this in RELEASE builds
    fixMissingSamplers();

-    if (isUsingUboBatching() && !BufferAllocator::isValid(getAllocationId())) {
+    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
        return;
    }

@@ -399,7 +439,7 @@ void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
        return;
    }

-    if (isUsingUboBatching() && !BufferAllocator::isValid(getAllocationId())) {
+    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
        return;
    }

@@ -436,7 +476,7 @@ void FMaterialInstance::assignUboAllocation(
        const Handle<HwBufferObject>& ubHandle,
        BufferAllocator::AllocationId id,
        BufferAllocator::allocation_size_t offset) {
-    assert_invariant(isUsingUboBatching());
+    assert_invariant(mUseUboBatching);

    mUboData = id;
    mUboOffset = offset;
@@ -56,7 +56,8 @@ class FTexture;

 class FMaterialInstance : public MaterialInstance {
 public:
-    FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name) noexcept;
+    FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name,
+            FEngine::UboBatchingMode batchingMode) noexcept;
    FMaterialInstance(FEngine& engine, FMaterialInstance const* other, const char* name);
    FMaterialInstance(const FMaterialInstance& rhs) = delete;
    FMaterialInstance& operator=(const FMaterialInstance& rhs) = delete;
@@ -66,6 +67,8 @@ public:
    ~FMaterialInstance() noexcept;

    void terminate(FEngine& engine);
+
+    void commitStreamUniformAssociations(FEngine::DriverApi& driver);
    
    void commit(FEngine& engine) const;

@@ -285,6 +288,7 @@ private:
    tsl::robin_map<backend::descriptor_binding_t, TextureParameter> mTextureParameters;
    mutable DescriptorSet mDescriptorSet;
    UniformBuffer mUniforms;
+    bool mHasStreamUniformAssociations = false;

    backend::PolygonOffset mPolygonOffset{};
    backend::StencilState mStencilState{};
@@ -292,11 +292,16 @@ 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

-    // Run the component managers' GC
-    auto& js = engine.getJobSystem();
-    js.runAndWait(jobs::createJob(js, nullptr, &FEngine::gc, &engine)); // gc all managers
+    // make sure we're done with the gcs
+    js.waitAndRelease(job);
 }

 bool FRenderer::shouldRenderFrame() const noexcept {
@@ -463,11 +468,16 @@ 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

-    // Run the component managers' GC
-    auto& js = engine.getJobSystem();
-    js.runAndWait(jobs::createJob(js, nullptr, &FEngine::gc, &engine)); // gc all managers
+    // make sure we're done with the gcs
+    js.waitAndRelease(job);
 }

 void FRenderer::readPixels(
@@ -561,6 +571,7 @@ 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;
@@ -570,7 +581,8 @@ 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), 0);
+                        1'000'000'000.0 / mDisplayInfo.refreshRate),
+                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
@@ -579,7 +591,7 @@ void FRenderer::renderStandaloneView(FView const* view) {

        engine.submitFrame();

-        driver.endFrame(0);
+        driver.endFrame(mFrameId);

        // 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
@@ -209,7 +209,7 @@ private:
    backend::Handle<backend::HwRenderTarget> mRenderTargetHandle;
    FSwapChain* mSwapChain = nullptr;
    size_t mCommandsHighWatermark = 0;
-    uint32_t mFrameId = 1; // id 0 is reserved for standalone views
+    uint32_t mFrameId = 0;
    FrameInfoManager mFrameInfoManager;
    backend::TextureFormat mHdrTranslucent;
    backend::TextureFormat mHdrQualityMedium;
@@ -39,7 +39,7 @@ using allocation_size_t = BufferAllocator::allocation_size_t;
 // FenceManager
 // ------------------------------------------------------------------------------------------------

-void UboManager::FenceManager::track(DriverApi& driver, AllocationIdContainer&& allocationIds) {
+void UboManager::FenceManager::track(DriverApi& driver, std::unordered_set<AllocationId>&& allocationIds) {
    if (allocationIds.empty()) {
        return;
    }
@@ -159,16 +159,15 @@ void UboManager::finishBeginFrame(DriverApi& driver) {
 }

 void UboManager::endFrame(DriverApi& driver) {
-    auto allocationIds =
-            FenceManager::AllocationIdContainer::with_capacity(mManagedInstances.size());
-    for (const auto* mi: mManagedInstances) {
+    std::unordered_set<AllocationId> allocationIds;
+    for (const auto* mi : mManagedInstances) {
        const AllocationId id = mi->getAllocationId();
        if (UTILS_UNLIKELY(!BufferAllocator::isValid(id))) {
            continue;
        }

        mAllocator.acquireGpu(id);
-        allocationIds.push_back(id);
+        allocationIds.insert(id);
    }

    mFenceManager.track(driver, std::move(allocationIds));
@@ -55,7 +55,6 @@ public:
    class FenceManager {
    public:
        using AllocationId = BufferAllocator::AllocationId;
-        using AllocationIdContainer = utils::FixedCapacityVector<AllocationId>;

        FenceManager() = default;
        ~FenceManager() = default;
@@ -66,7 +65,7 @@ public:

        // Creates a new fence to track a set of allocation IDs for the current frame.
        // This marks the beginning of GPU's usage of these resources.
-        void track(backend::DriverApi& driver, AllocationIdContainer&& allocationIds);
+        void track(backend::DriverApi& driver, std::unordered_set<AllocationId>&& allocationIds);


        // Checks all tracked fences and invokes a callback for resources associated with
@@ -81,7 +80,7 @@ public:
    private:
        // Not ideal, but we need to know which slots to decrement gpuUseCount for each frame.
        using FenceAndAllocations =
-                std::pair<backend::Handle<backend::HwFence>, AllocationIdContainer>;
+                std::pair<backend::Handle<backend::HwFence>, std::unordered_set<AllocationId>>;
        std::vector<FenceAndAllocations> mFenceAllocationList;
    };

@@ -13,25 +13,11 @@ set(GENERATION_ROOT ${CMAKE_CURRENT_BINARY_DIR})
 set(RESOURCE_DIR  "${GENERATION_ROOT}/resources")
 file(MAKE_DIRECTORY ${RESOURCE_DIR})

-set(MATERIAL_DIR  "${GENERATION_ROOT}/material")
-file(MAKE_DIRECTORY ${MATERIAL_DIR})
-
-set(SANDBOX_TRANSFORM_NAME_MAT ${CMAKE_CURRENT_SOURCE_DIR}/sandboxTransformName.mat)
-set(TEST_MATERIAL_TRANSFORM_NAME_FILAMAT ${MATERIAL_DIR}/test_material_transformname.filamat)
-
-add_custom_command(
-        OUTPUT ${TEST_MATERIAL_TRANSFORM_NAME_FILAMAT}
-        COMMAND matc ${MATC_BASE_FLAGS} -o ${TEST_MATERIAL_TRANSFORM_NAME_FILAMAT} ${SANDBOX_TRANSFORM_NAME_MAT}
-        MAIN_DEPENDENCY ${SANDBOX_TRANSFORM_NAME_MAT}
-        DEPENDS matc
-        COMMENT "Compiling material ${SANDBOX_TRANSFORM_NAME_MAT}"
-)
-
 get_resgen_vars(${RESOURCE_DIR} filament_test_resources)

 set(RESOURCE_BINS
        ${CMAKE_CURRENT_SOURCE_DIR}/test_material.filamat
-        ${TEST_MATERIAL_TRANSFORM_NAME_FILAMAT}
+        ${CMAKE_CURRENT_SOURCE_DIR}/test_material_transformname.filamat
 )

 add_custom_command(
@@ -11,6 +11,7 @@ OUTPUT_DIR="filament/test/"
 ./build.sh -p desktop release matc
 MATC=out/cmake-release/tools/matc/matc
 ${MATC} --platform all --api all -o ${OUTPUT_DIR}/test_material.filamat samples/materials/sandboxLit.mat
+${MATC} --platform all --api all -o ${OUTPUT_DIR}/test_material_transformname.filamat samples/materials/sandboxTransformName.mat

 set +x

@@ -1,12 +1,12 @@
 Pod::Spec.new do |spec|
  spec.name = "Filament"
-  spec.version = "1.68.0"
+  spec.version = "1.67.1"
  spec.license = { :type => "Apache 2.0", :file => "LICENSE" }
  spec.homepage = "https://google.github.io/filament"
  spec.authors = "Google LLC."
  spec.summary = "Filament is a real-time physically based rendering engine for Android, iOS, Windows, Linux, macOS, and WASM/WebGL."
  spec.platform = :ios, "11.0"
-  spec.source = { :http => "https://github.com/google/filament/releases/download/v1.68.0/filament-v1.68.0-ios.tgz" }
+  spec.source = { :http => "https://github.com/google/filament/releases/download/v1.67.1/filament-v1.67.1-ios.tgz" }

  # Fix linking error with Xcode 12; we do not yet support the simulator on Apple silicon.
  spec.pod_target_xcconfig = {
@@ -28,7 +28,7 @@
 namespace filament {

 // update this when a new version of filament wouldn't work with older materials
-static constexpr size_t MATERIAL_VERSION = 68;
+static constexpr size_t MATERIAL_VERSION = 67;

 /**
 * Supported shading models
@@ -170,6 +170,10 @@ public:
    // negative value if name doesn't exist or Panic if exceptions are enabled
    ssize_t getFieldOffset(std::string_view name, size_t index) const;

+    // returns offset in bytes of the transform matrix for the given external texture binding
+    // returns -1 if the field doesn't exist
+    ssize_t getTransformFieldOffset(uint8_t binding) const;
+
    FieldInfo const* getFieldInfo(std::string_view name) const;

    bool hasField(std::string_view name) const noexcept {
@@ -197,6 +201,7 @@ private:
    utils::CString mName;
    utils::FixedCapacityVector<FieldInfo> mFieldInfoList;
    std::unordered_map<std::string_view , uint32_t> mInfoMap;
+    std::unordered_map<uint8_t, uint32_t> mTransformOffsetMap;
    uint32_t mSize = 0; // size in bytes rounded to multiple of 4
    Alignment mAlignment = Alignment::std140;
    Target mTarget = Target::UNIFORM;
@@ -122,6 +122,8 @@ BufferInterfaceBlock::BufferInterfaceBlock(Builder const& builder) noexcept

    auto& uniformsInfoList = mFieldInfoList;

+    auto& transformOffsetMap = mTransformOffsetMap;
+
    uint32_t i = 0;
    uint16_t offset = 0;
    for (auto const& e : builder.mEntries) {
@@ -148,6 +150,11 @@ BufferInterfaceBlock::BufferInterfaceBlock(Builder const& builder) noexcept
        // record this uniform info
        infoMap[{ info.name.data(), info.name.size() }] = i;

+        // record the transform offset if this is an external sampler
+        if (e.associatedSampler != 0) {
+            transformOffsetMap[e.associatedSampler] = offset;
+        }
+
        // advance offset to next slot
        offset += stride * std::max(1u, e.size);
        ++i;
@@ -163,6 +170,14 @@ ssize_t BufferInterfaceBlock::getFieldOffset(std::string_view name, size_t index
    return (ssize_t)info->getBufferOffset(index);
 }

+ssize_t BufferInterfaceBlock::getTransformFieldOffset(uint8_t binding) const {
+    auto pos = mTransformOffsetMap.find(binding);
+    if (pos == mTransformOffsetMap.end()) {
+        return -1;
+    }
+    return (ssize_t)pos->second * sizeof(uint32_t);
+}
+
 BufferInterfaceBlock::FieldInfo const* BufferInterfaceBlock::getFieldInfo(
        std::string_view name) const {
    auto pos = mInfoMap.find(name);
@@ -17,6 +17,9 @@
 #ifndef FILAGUI_IMGUIHELPER_H_
 #define FILAGUI_IMGUIHELPER_H_

+#include <vector>
+#include <functional>
+
 #include <filament/Engine.h>
 #include <filament/IndexBuffer.h>
 #include <filament/Material.h>
@@ -29,10 +32,6 @@
 #include <utils/Entity.h>
 #include <utils/Path.h>

-#include <functional>
-#include <vector>
-#include <unordered_set>
-
 struct ImDrawData;
 struct ImGuiIO;
 struct ImGuiContext;
@@ -101,7 +100,6 @@ public:
      filament::TextureSampler mSampler;
      bool mFlipVertical = false;
      utils::Path mSettingsPath;
-      std::unordered_set<filament::Texture*> mImGuiTextures;
 };

 } // namespace filagui
@@ -148,12 +148,6 @@ ImGuiHelper::~ImGuiHelper() {
        mEngine->destroy(ib);
    }

-    // These are textures that have been created by imgui but were not explicitly freed before
-    // shutdown.
-    for (auto itextures: mImGuiTextures) {
-        mEngine->destroy(itextures);
-    }
-
    EntityManager& em = utils::EntityManager::get();
    em.destroy(mRenderable);
    em.destroy(mCameraEntity);
@@ -195,48 +189,6 @@ void ImGuiHelper::render(float timeStepInSeconds, Callback imguiCommands) {

 void ImGuiHelper::processImGuiCommands(ImDrawData* commands, const ImGuiIO& io) {
    ImGui::SetCurrentContext(mImGuiContext);
-    if (commands->Textures != nullptr) {
-        for (ImTextureData* tex : *commands->Textures) {
-            if (tex->Status == ImTextureStatus_OK) {
-                continue;
-            } else if (tex->Status == ImTextureStatus_WantCreate) {
-                IM_ASSERT(tex->TexID == 0 && tex->BackendUserData == nullptr);
-                Texture* ftex = Texture::Builder()
-                        .width(tex->Width)
-                        .height(tex->Height)
-                        .levels(1)
-                        .format(Texture::InternalFormat::RGBA8)
-                        .sampler(Texture::Sampler::SAMPLER_2D)
-                        .build(*mEngine);
-                IM_ASSERT(tex->Format == ImTextureFormat_RGBA32);
-                const int size = tex->Width * tex->Height * 4;
-                Texture::PixelBufferDescriptor pb(tex->GetPixels(), size,
-                                                Texture::Format::RGBA,
-                                                Texture::Type::UBYTE);
-                ftex->setImage(*mEngine, 0, std::move(pb));
-
-                tex->SetTexID((ImTextureID)ftex);
-                tex->SetStatus(ImTextureStatus_OK);
-                mImGuiTextures.insert(ftex);
-            } else if (tex->Status == ImTextureStatus_WantUpdates) {
-                const int size = tex->Width * tex->Height * 4;
-                Texture::PixelBufferDescriptor pb(tex->GetPixels(), size,
-                                                Texture::Format::RGBA,
-                                                Texture::Type::UBYTE);
-                filament::Texture* ftex = (filament::Texture*)tex->TexID;
-                ftex->setImage(*mEngine, 0, std::move(pb));
-                tex->SetStatus(ImTextureStatus_OK);
-            } else if (tex->Status == ImTextureStatus_WantDestroy &&
-                    tex->UnusedFrames > 0) {
-
-                filament::Texture* ftex = (filament::Texture*)tex->TexID;
-                mEngine->destroy(ftex);
-                tex->SetTexID(ImTextureID_Invalid);
-                tex->SetStatus(ImTextureStatus_Destroyed);
-                mImGuiTextures.erase(ftex);
-            }
-        }
-    }

    mHasSynced = false;
    auto& rcm = mEngine->getRenderableManager();
@@ -276,7 +228,7 @@ void ImGuiHelper::processImGuiCommands(ImDrawData* commands, const ImGuiIO& io)
            if (pcmd.UserCallback) {
                pcmd.UserCallback(cmds, &pcmd);
            } else {
-                auto texture = (Texture const*)pcmd.GetTexID();
+                auto texture = (Texture const*)pcmd.TextureId;
                MaterialInstance* materialInstance;
 #ifdef __ANDROID__
                if (texture && texture->getTarget() == Texture::Sampler::SAMPLER_EXTERNAL) {
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Run Yu	44ef0af4a4	immediately GetMappedRange after MapAsync still would return null pointer	2025-11-21 16:53:03 -05:00
Run Yu	c9b5bfc4a4	Use a stage pool for WebGPU, but somehow GetMappedRange would sometimes return null, which breaks the code.	2025-11-21 10:57:02 -05:00