Add missing VulkanDriver.h header

Keep it's original behavior of no-op

.github/workflows/postsubmit.yml

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

.github/workflows/presubmit.yml

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

CMakeLists.txt

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

android/build.gradle

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

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

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

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

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

build.sh

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

build/android/build.sh

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

build/common/versions

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

build/toolchain-aarch64-linux-android.cmake

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

build/toolchain-arm7-linux-android.cmake

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

docs/dup/intro.html

@@ -181,7 +181,7 @@ important for <code>matc</code> (material compiler).</p>
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.67.1'
+    implementation 'com.google.android.filament:filament-android:1.68.0'
 }
 </code></pre>
 <p>Here are all the libraries available in the group <code>com.google.android.filament</code>:</p>
@@ -196,7 +196,7 @@ dependencies {
 </div>
 <h3 id="ios"><a class="header" href="#ios">iOS</a></h3>
 <p>iOS projects can use CocoaPods to install the latest release:</p>
-<pre><code class="language-shell">pod 'Filament', '~&gt; 1.67.1'
+<pre><code class="language-shell">pod 'Filament', '~&gt; 1.68.0'
 </code></pre>
 <h2 id="documentation"><a class="header" href="#documentation">Documentation</a></h2>
 <ul>

docs_src/build/requirements.txt

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

filament/backend/CMakeLists.txt

@@ -318,6 +318,8 @@ if (FILAMENT_SUPPORTS_WEBGPU)
             src/webgpu/WebGPURenderPrimitive.h
             src/webgpu/WebGPURenderTarget.cpp
             src/webgpu/WebGPURenderTarget.h
+            src/webgpu/WebGPUStagePool.cpp
+            src/webgpu/WebGPUStagePool.h
             src/webgpu/WebGPUStrings.h
             src/webgpu/WebGPUSwapChain.cpp
             src/webgpu/WebGPUSwapChain.h

filament/backend/include/backend/BufferObjectStreamDescriptor.h

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

filament/backend/include/backend/DriverEnums.h

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

filament/backend/include/backend/Platform.h

@@ -255,6 +255,28 @@ public:
         REALTIME,
     };
 
+    /**
+     * Defines how asynchronous operations are handled by the engine.
+     */
+    enum class AsynchronousMode : uint8_t {
+        /**
+         * Asynchronous operations are disabled. This is the default.
+         */
+        NONE,
+
+        /**
+         * Attempts to use a dedicated worker thread for asynchronous tasks. If threading is not
+         * supported by the platform, it automatically falls back to using an amortization strategy.
+         */
+        THREAD_PREFERRED,
+
+        /**
+         * Uses an amortization strategy, processing a small number of asynchronous tasks during
+         * each engine update cycle.
+         */
+        AMORTIZATION,
+    };
+
     struct DriverConfig {
         /**
          * Size of handle arena in bytes. Setting to 0 indicates default value is to be used.
@@ -324,6 +346,11 @@ public:
          *      - VulkanPlatform
          */
         bool vulkanEnableStagingBufferBypass = false;
+
+        /**
+         * Asynchronous mode for the engine. Defines how asynchronous operations are handled.
+         */
+        AsynchronousMode asynchronousMode = AsynchronousMode::NONE;
     };
 
     Platform() noexcept;

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

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

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

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

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

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

filament/backend/src/metal/MetalDriver.mm

@@ -1144,8 +1144,12 @@ bool MetalDriver::isRenderTargetFormatSupported(TextureFormat format) {
 bool MetalDriver::isFrameBufferFetchSupported() {
     // FrameBuffer fetch is achievable via "programmable blending" in Metal, and only supported on
     // Apple GPUs with readWriteTextureSupport.
-    return mContext->highestSupportedGpuFamily.apple >= 1 &&
-            mContext->device.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;
 }
 
 bool MetalDriver::isFrameBufferFetchMultiSampleSupported() {
@@ -1286,10 +1290,6 @@ void MetalDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&
     scheduleDestroy(std::move(data));
 }
 
-void MetalDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh, BufferObjectStreamDescriptor&& streams) {
-    // Noop
-}
-
 void MetalDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& data,
         uint32_t byteOffset) {
     FILAMENT_CHECK_PRECONDITION(!isInRenderPass(mContext))

filament/backend/src/metal/MetalShaderCompiler.mm

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

filament/backend/src/noop/NoopDriver.cpp

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

filament/backend/src/opengl/OpenGLDriver.cpp

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

filament/backend/src/opengl/OpenGLDriver.h

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

filament/backend/src/opengl/gl_headers.h

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

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

@@ -178,20 +178,30 @@ void PlatformEGLAndroid::beginFrame(
         int64_t const monotonic_clock_ns,
         int64_t refreshIntervalNs,
         uint32_t const frameId) noexcept {
-    // associate the user frameid with the system frame id
-    setPresentFrameId(mCurrentDrawSwapChain, frameId);
+    // 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)
 
-    if (mPerformanceHintSession.isValid()) {
+    // associate the user frameid with the system frame id.
+    if (frameId && mCurrentDrawSwapChain) {
+        // mCurrentDrawSwapChain could be null if we're called from renderStandaloneView
+        setPresentFrameId(mCurrentDrawSwapChain, frameId);
+    }
+
+    if (frameId && mPerformanceHintSession.isValid()) {
         if (refreshIntervalNs <= 0) {
             // we're not provided with a target time, use the display period, if everything fails,
             // assume 16.67ms
             refreshIntervalNs = 16'666'667;
 
-            CompositorTiming compositorTiming{};
-            bool const hasCompositorTiming =
-                    queryCompositorTiming(mCurrentDrawSwapChain, &compositorTiming);
-            if (hasCompositorTiming && compositorTiming.compositeInterval > 0) {
-                refreshIntervalNs = compositorTiming.compositeInterval;
+            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;
+                }
             }
         }
         mStartTimeOfActualWork = clock::time_point(std::chrono::nanoseconds(monotonic_clock_ns));
@@ -440,6 +450,10 @@ Platform::SwapChain* PlatformEGLAndroid::createSwapChain(
 void PlatformEGLAndroid::destroySwapChain(SwapChain* swapChain) noexcept {
     if (swapChain) {
         SwapChainEGLAndroid* const sc = static_cast<SwapChainEGLAndroid*>(swapChain);
+        if (mCurrentDrawSwapChain == sc) {
+            // don't keep a dangling pointer around
+            mCurrentDrawSwapChain = nullptr;
+        }
         sc->terminate(*this);
         delete sc;
     }

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

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

filament/backend/src/ostream.cpp

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

filament/backend/src/vulkan/VulkanBufferProxy.cpp

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

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -150,31 +150,6 @@ static CallbackHandler::Callback syncCallbackWrapper = [](void* userData) {
     cbData->cb(cbData->sync, cbData->userData);
 };
 
-// This needs to be discussed because Vulkan has different default Matrix formats
-// right now this is following GL.
-static void copyMat3f(void* addr, size_t const offset, const math::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
-}
-
 }// anonymous namespace
 
 #if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
@@ -447,12 +422,14 @@ void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
         int64_t refreshIntervalNs, uint32_t frameId) {
     FVK_PROFILE_MARKER(PROFILE_NAME_BEGINFRAME);
 
-    if (mCurrentSwapChain) { // This should be guaranteed
+    // if frameId is 0, it means we're not associated to a particular frame, which is the case
+    // for standalone views. And in this case we skip the frame info timing collection.
+    if (frameId && mCurrentSwapChain) { // This should be guaranteed
         mPlatform->setPresentFrameId(mCurrentSwapChain->swapChain, frameId);
     }
 
     // Check if any command have finished and reset all its used resources. The resources
-    // wont be destroyed but their reference count will decreased if the command is already
+    // won't be destroyed but their reference count will be decreased if the command is already
     // completed.
     //
     // This will let us check if any VulkanBuffer is currently in flight or not.
@@ -528,10 +505,7 @@ void VulkanDriver::copyToMemoryMappedBuffer(MemoryMappedBufferHandle mmbh, size_
     assert_invariant(any(mmb->access & MapBufferAccessFlags::WRITE_BIT));
     assert_invariant(offset + data.size <= mmb->size);
 
-    // TODO: this isa zero-effort implementation of copyToMemoryMappedBuffer(), where we just
-    //       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.
+    // On UMA systems, this is a memcpy into the memory-mapped buffer, otherwise use a staging buffer.
     updateBufferObject(mmb->boh, std::move(data), mmb->offset + offset);
 }
 
@@ -1182,7 +1156,7 @@ void VulkanDriver::destroyDescriptorSet(Handle<HwDescriptorSet> dsh) {
 }
 
 Handle<HwStream> VulkanDriver::createStreamNative(void* nativeStream, utils::ImmutableCString tag) {
-    FILAMENT_CHECK_PRECONDITION(false) << "createStreamNative not supported in Vulkan.";
+    FVK_LOGW << "createStreamNative not supported in Vulkan.";
     return {};
 }
 
@@ -1206,7 +1180,6 @@ void VulkanDriver::setAcquiredImage(Handle<HwStream> sh, void* image, const math
         scheduleRelease(stream->takePrevious());
     }
     stream->acquire({ image, cb, userData, handler});
-    stream->setFrontEndTransform(transform);
     mStreamsWithPendingAcquiredImage.push_back(stream);
 }
 
@@ -1234,10 +1207,8 @@ void VulkanDriver::updateStreams(CommandStream* driver) {
             // 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,
-                                        transform = stream->getFrontEndTransform()]() {
+                                         image = stream->getAcquired().image]() {
                 auto texture = s->getTexture(image);
-                s->setBackendTransform(transform);
                 if (!texture) {
                     auto externalImage = fvkutils::createExternalImageFromRaw(mPlatform, image, false);
                     auto metadata = mPlatform->extractExternalImageMetadata(externalImage);
@@ -1583,13 +1554,6 @@ void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor
     scheduleDestroy(std::move(p));
 }
 
-void VulkanDriver::registerBufferObjectStreams(Handle<HwBufferObject> boh,
-        BufferObjectStreamDescriptor&& streams) {
-
-    auto bo = resource_ptr<VulkanBufferObject>::cast(&mResourceManager, boh);
-    mStreamUniformDescriptors[bo.get()] = std::move(streams);
-}
-
 void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& bd,
         uint32_t byteOffset) {
     VulkanCommandBuffer& commands = mCommands.get();
@@ -1597,33 +1561,6 @@ void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescript
     auto bo = resource_ptr<VulkanBufferObject>::cast(&mResourceManager, boh);
     commands.acquire(bo);
 
-    if (UTILS_UNLIKELY(!mStreamUniformDescriptors.empty())) {
-        auto streamDescriptors = mStreamUniformDescriptors.find(bo.get());
-        if (streamDescriptors != mStreamUniformDescriptors.end()) {
-            for (auto const& [offset, streamHandle, associationType]:
-                    streamDescriptors->second.mStreams) {
-                if (associationType == BufferObjectStreamAssociationType::TRANSFORM_MATRIX) {
-                    math::mat3f transform;
-                    if (streamHandle) {
-                        auto stream =
-                                resource_ptr<VulkanStream>::cast(&mResourceManager, streamHandle);
-                        if (stream->streamType == StreamType::NATIVE) {
-                            FILAMENT_CHECK_PRECONDITION(false)
-                                    << "Native Stream not supported in Vulkan.";
-                        } else {
-                            // Backend call since getStreamTransformMatrix is called from async
-                            // updateBufferObject.
-                            transform = stream->getBackEndTransform();
-                        }
-                    }
-
-                    copyMat3f(bd.buffer, offset, transform);
-                }
-            }
-            mStreamUniformDescriptors.erase(streamDescriptors);
-        }
-    }
-
     bo->loadFromCpu(commands, bd.buffer, byteOffset, bd.size);
 
     scheduleDestroy(std::move(bd));

filament/backend/src/vulkan/VulkanDriver.h

@@ -41,6 +41,7 @@
 #include "vulkan/utils/Definitions.h"
 
 #include "backend/DriverEnums.h"
+#include "backend/BufferObjectStreamDescriptor.h"
 
 #include "DriverBase.h"
 #include "private/backend/Driver.h"
@@ -160,10 +161,6 @@ private:
     VulkanExternalImageManager mExternalImageManager;
     VulkanStreamedImageManager mStreamedImageManager;
 
-    // Stream transforms
-    std::unordered_map<VulkanBufferObject*, BufferObjectStreamDescriptor> mStreamUniformDescriptors;
-    math::mat3f getStreamTransformMatrix(Handle<HwStream> sh);
-
 
     // This maps a VulkanSwapchain to a native swapchain. VulkanSwapchain should have a copy of the
     // Platform::Swapchain pointer, but queryFrameTimestamps() and queryCompositorTiming() are

filament/backend/src/vulkan/VulkanSwapChain.cpp

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

filament/backend/src/vulkan/VulkanTexture.h

@@ -52,12 +52,8 @@ struct VulkanStream : public HwStream, fvkmemory::Resource {
         return previous;
     }
     const AcquiredImage& getAcquired() const { return user_thread.mAcquired; }
-    void setFrontEndTransform(const math::mat3f& transform) { user_thread.mTransform = transform; }
-    const math::mat3f& getFrontEndTransform() const { return user_thread.mTransform; }
 
     //-- These methods are only called from the backend thread
-    void setBackendTransform(const math::mat3f& transform) { mTransform = transform; }
-    const math::mat3f& getBackEndTransform() const { return mTransform; }
     fvkmemory::resource_ptr<VulkanTexture> getTexture(void* ahb) {
         if (auto itr = mTextures.find(ahb); itr != mTextures.end()) {
             return itr->second;
@@ -71,11 +67,9 @@ private:
     struct {
         AcquiredImage mAcquired;
         AcquiredImage mPrevious;
-        math::mat3f mTransform;
     } user_thread;
 
     // #TODO b/442937292
-    math::mat3f mTransform;
     std::unordered_map<void*, fvkmemory::resource_ptr<VulkanTexture>> mTextures;
 };
 

filament/backend/src/webgpu/WebGPUBufferBase.cpp

@@ -18,6 +18,7 @@
 
 #include "WebGPUConstants.h"
 #include "WebGPUQueueManager.h"
+#include "WebGPUStagePool.h"
 
 #include "DriverBase.h"
 #include <backend/BufferDescriptor.h>
@@ -65,7 +66,7 @@ WebGPUBufferBase::WebGPUBufferBase(wgpu::Device const& device, const wgpu::Buffe
 // of 4 by padding with zeros.
 void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
         const uint32_t byteOffset, wgpu::Device const& device,
-        WebGPUQueueManager* const webGPUQueueManager) {
+        WebGPUQueueManager* const webGPUQueueManager, WebGPUStagePool* const webGPUStagePool) {
     FILAMENT_CHECK_PRECONDITION(bufferDescriptor.buffer)
             << "updateGPUBuffer called with a null buffer";
     FILAMENT_CHECK_PRECONDITION(bufferDescriptor.size + byteOffset <= mBuffer.GetSize())
@@ -85,15 +86,12 @@ void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
     const size_t stagingBufferSize =
             remainder == 0 ? bufferDescriptor.size : mainBulk + FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS;
 
-    // create a staging buffer
-    wgpu::BufferDescriptor descriptor{
-        .label = "Filament WebGPU Staging Buffer",
-        .usage = wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc,
-        .size = stagingBufferSize,
-        .mappedAtCreation = true };
-    wgpu::Buffer stagingBuffer = device.CreateBuffer(&descriptor);
+    wgpu::Buffer stagingBuffer = webGPUStagePool->acquireBuffer(stagingBufferSize,
+            webGPUQueueManager->getLatestSubmissionState());
 
     void* mappedRange = stagingBuffer.GetMappedRange();
+    assert_invariant(mappedRange);
+
     memcpy(mappedRange, bufferDescriptor.buffer, bufferDescriptor.size);
 
     // Make sure the padded memory is set to 0 to have deterministic behaviors
@@ -106,7 +104,9 @@ void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
 
     // Copy the staging buffer contents to the destination buffer.
     webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(stagingBuffer, 0, mBuffer,
-            byteOffset, stagingBufferSize);
+            byteOffset,
+            remainder == 0 ? bufferDescriptor.size
+                           : mainBulk + FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS);
 }
 
 } // namespace filament::backend

filament/backend/src/webgpu/WebGPUBufferBase.h

@@ -25,6 +25,7 @@ namespace filament::backend {
 
 class BufferDescriptor;
 class WebGPUQueueManager;
+class WebGPUStagePool;
 
 /**
   * A base class for WebGPU buffer objects, providing common functionality for creating and
@@ -40,7 +41,7 @@ public:
      * ensures the calls happen in the expected sequence.
      */
     void updateGPUBuffer(BufferDescriptor const&, uint32_t byteOffset, wgpu::Device const& device,
-            WebGPUQueueManager* const webGPUQueueManager);
+            WebGPUQueueManager* const webGPUQueueManager, WebGPUStagePool* const webGPUStagePool);
 
     [[nodiscard]] wgpu::Buffer const& getBuffer() const { return mBuffer; }
 

filament/backend/src/webgpu/WebGPUDriver.cpp

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

filament/backend/src/webgpu/WebGPUDriver.h

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

filament/backend/src/webgpu/WebGPUQueueManager.cpp

@@ -64,6 +64,7 @@ wgpu::CommandEncoder WebGPUQueueManager::getCommandEncoder() {
         };
         mCommandEncoder = mDevice.CreateCommandEncoder(&commandEncoderDescriptor);
         ASSERT_POSTCONDITION(mCommandEncoder, "Failed to create command encoder.");
+        mLatestSubmissionState = std::make_shared<WebGPUSubmissionState>();
     }
     return mCommandEncoder;
 }
@@ -94,8 +95,6 @@ void WebGPUQueueManager::submit() {
         return;
     }
 
-    mLatestSubmissionState = std::make_shared<WebGPUSubmissionState>();
-
     wgpu::CommandBufferDescriptor commandBufferDescriptor{
         .label = "Filament Command Buffer",
     };

filament/backend/src/webgpu/WebGPUStagePool.cpp

@@ -0,0 +1,92 @@
+/*
+* Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "WebGPUStagePool.h"
+
+#include "WebGPUConstants.h"
+#include "WebGPUQueueManager.h"
+
+namespace filament::backend {
+
+WebGPUStagePool::WebGPUStagePool(wgpu::Device const& device) : mDevice(device) {}
+
+WebGPUStagePool::~WebGPUStagePool() = default;
+
+wgpu::Buffer WebGPUStagePool::acquireBuffer(size_t requiredSize,
+        std::shared_ptr<WebGPUSubmissionState> submissionState) {
+    wgpu::Buffer buffer;
+    {
+        std::lock_guard<std::mutex> lock(mMutex);
+        auto iter = mBuffers.lower_bound(requiredSize);
+        if (iter != mBuffers.end()) {
+            buffer = iter->second;
+            mBuffers.erase(iter);
+        }
+    }
+    if (!buffer.Get()) {
+        buffer = createNewBuffer(requiredSize);
+    }
+    mInProgress.push_back({submissionState, buffer});
+    return buffer;
+}
+
+void WebGPUStagePool::recycleBuffer(wgpu::Buffer buffer) {
+    struct UserData final {
+        wgpu::Buffer buffer;
+        WebGPUStagePool* webGPUStagePool;
+    };
+    auto userData =
+            std::make_unique<UserData>(UserData{ .buffer = buffer, .webGPUStagePool = this });
+    buffer.MapAsync(wgpu::MapMode::Write, 0, buffer.GetSize(), wgpu::CallbackMode::AllowSpontaneous,
+            [data = std::move(userData)](wgpu::MapAsyncStatus status, const char* message) {
+                if (UTILS_LIKELY(status == wgpu::MapAsyncStatus::Success)) {
+                    if (!data->webGPUStagePool) {
+                        return;
+                    }
+                    std::lock_guard<std::mutex> lock(data->webGPUStagePool->mMutex);
+                    data->webGPUStagePool->mBuffers.insert(
+                            { data->buffer.GetSize(), data->buffer });
+                } else {
+                    FWGPU_LOGE << "Failed to MapAsync when recycling staging buffer: " << message;
+                }
+            });
+}
+
+void WebGPUStagePool::gc() {
+    // We found that MapAsync would sometimes lead to GetMappedRange returning nullptr if the
+    // command using that staging buffer has not finished executing, so here we only recycle those
+    // buffers that are not still being used by any command
+    std::vector<std::pair<std::shared_ptr<WebGPUSubmissionState>, wgpu::Buffer>> stillInProgress;
+    for (auto& [st, buffer]: mInProgress) {
+        if (st->getStatus() == FenceStatus::CONDITION_SATISFIED) {
+            recycleBuffer(buffer);
+        } else {
+            stillInProgress.push_back({st, buffer});
+        }
+    }
+    std::swap(mInProgress, stillInProgress);
+}
+
+wgpu::Buffer WebGPUStagePool::createNewBuffer(size_t bufferSize) {
+    wgpu::BufferDescriptor descriptor{
+        .label = "Filament WebGPU Staging Buffer",
+        .usage = wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc,
+        .size = bufferSize,
+        .mappedAtCreation = true };
+    return mDevice.CreateBuffer(&descriptor);
+}
+
+} // namespace filament::backend

filament/backend/src/webgpu/WebGPUStagePool.h

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

filament/include/filament/Engine.h

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

filament/include/filament/Material.h

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

filament/src/Camera.cpp

@@ -20,6 +20,8 @@
 
 #include <math/mat4.h>
 
+#include <utils/Panic.h>
+
 #include <math/vec2.h>
 #include <math/vec3.h>
 #include <math/vec4.h>
@@ -34,6 +36,10 @@ using namespace math;
 void Camera::setProjection(double const fovInDegrees, double const aspect,
         double const near, double const far, Fov const direction) {
 
+    FILAMENT_CHECK_PRECONDITION(near > 0 && far > near)
+            << "Camera preconditions not met in setProjection(): near <= 0 or far <= near, near="
+            << near << ", far=" << far;
+
     setCustomProjection(
             projection(direction, fovInDegrees, aspect, near),
             projection(direction, fovInDegrees, aspect, near, far),
@@ -43,6 +49,10 @@ void Camera::setProjection(double const fovInDegrees, double const aspect,
 void Camera::setLensProjection(double const focalLengthInMillimeters,
         double const aspect, double const near, double const far) {
 
+    FILAMENT_CHECK_PRECONDITION(near > 0 && far > near)
+        << "Camera preconditions not met in setLensProjection(): near <= 0 or far <= near, near="
+        << near << ", far=" << far;
+
     setCustomProjection(
             projection(focalLengthInMillimeters, aspect, near),
             projection(focalLengthInMillimeters, aspect, near, far),

filament/src/Engine.cpp

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

filament/src/FrameInfo.cpp

@@ -101,11 +101,24 @@ void FrameInfoManager::beginFrame(FSwapChain* swapChain, DriverApi& driver,
     auto& history = mFrameTimeHistory;
     // don't exceed the capacity, drop the oldest entry
     if (UTILS_LIKELY(history.size() == history.capacity())) {
-        if (!mDisableGpuFrameComplete) {
-            assert_invariant(history.back().fence);
-            driver.destroyFence(std::move(history.back().fence));
+        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;
         }
-        history.pop_back();
     }
 
     // create a new entry
@@ -193,6 +206,14 @@ void FrameInfoManager::beginFrame(FSwapChain* swapChain, DriverApi& driver,
 }
 
 void FrameInfoManager::endFrame(DriverApi& driver) noexcept {
+    if (mLastBeginFrameSkipped) {
+        // if we had to skip the last beginFrame(), endFrame() needs to be skipped too
+        // because history.front() now references the wrong frame.
+        // It is guaranteed that if beginFrame() is called, endFrame() will be called too.
+        mLastBeginFrameSkipped = false;
+        return;
+    }
+
     auto& front = mFrameTimeHistory.front();
     front.endFrame = std::chrono::steady_clock::now();
 

filament/src/FrameInfo.h

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

filament/src/details/Camera.cpp

@@ -95,13 +95,8 @@ mat4 FCamera::projection(double const focalLengthInMillimeters,
 void UTILS_NOINLINE FCamera::setCustomProjection(mat4 const& projection,
         mat4 const& projectionForCulling, double const near, double const far) noexcept {
 
-    FILAMENT_CHECK_PRECONDITION(near > 0)
-            << "Camera preconditions not met in setCustomProjection(): near <= 0, near=" << near;
-    auto const& featureFlags = mEngine.features.engine.debug;
-    FILAMENT_FLAG_GUARDED_CHECK_PRECONDITION(far > near,
-            featureFlags.assert_camera_projection_near_far)
-            << "Camera preconditions not met in setCustomProjection(): far <= near, near="
-            << near << ", far=" << far;
+    FILAMENT_CHECK_PRECONDITION(near != far)
+            << "Camera preconditions not met in setCustomProjection(): near = far = " << near;
 
     for (auto& eyeProjection: mEyeProjection) {
         eyeProjection = projection;
@@ -115,13 +110,8 @@ void UTILS_NOINLINE FCamera::setCustomEyeProjection(mat4 const* projection, size
         mat4 const& projectionForCulling, double const near, double const far) {
     const Engine::Config& config = mEngine.getConfig();
 
-    FILAMENT_CHECK_PRECONDITION(near > 0)
-            << "Camera preconditions not met in setCustomEyeProjection(): near <= 0, near=" << near;
-    auto const& featureFlags = mEngine.features.engine.debug;
-    FILAMENT_FLAG_GUARDED_CHECK_PRECONDITION(far > near,
-            featureFlags.assert_camera_projection_near_far)
-            << "Camera preconditions not met in setCustomEyeProjection(): far <= near, near="
-            << near << ", far=" << far;
+    FILAMENT_CHECK_PRECONDITION(near != far)
+            << "Camera preconditions not met in setCustomEyeProjection(): near = far = " << near;
 
     FILAMENT_CHECK_PRECONDITION(count >= config.stereoscopicEyeCount)
             << "All eye projections must be supplied together, count must be >= "

filament/src/details/Engine.cpp

@@ -98,7 +98,7 @@ using namespace filaflat;
 namespace {
 
 backend::Platform::DriverConfig getDriverConfig(FEngine* instance) {
-    return {
+    Platform::DriverConfig driverConfig {
         .handleArenaSize = instance->getRequestedDriverHandleArenaSize(),
         .metalUploadBufferSizeBytes = instance->getConfig().metalUploadBufferSizeBytes,
         .disableParallelShaderCompile = instance->features.backend.disable_parallel_shader_compile,
@@ -116,7 +116,11 @@ backend::Platform::DriverConfig getDriverConfig(FEngine* instance) {
         .gpuContextPriority = instance->getConfig().gpuContextPriority,
         .vulkanEnableStagingBufferBypass =
                 instance->features.backend.vulkan.enable_staging_buffer_bypass,
+        .asynchronousMode = instance->features.backend.enable_asynchronous_operation ?
+                instance->getConfig().asynchronousMode : AsynchronousMode::NONE,
     };
+
+    return driverConfig;
 }
 
 } // anonymous
@@ -727,7 +731,6 @@ void FEngine::prepare() {
             // post-process materials instances must be commited explicitly because their
             // parameters are typically not set at this point in time.
             if (item->getMaterial()->getMaterialDomain() == MaterialDomain::SURFACE) {
-                item->commitStreamUniformAssociations(driver);
                 item->commit(driver, uboManager);
             }
         });
@@ -991,10 +994,9 @@ FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
     return p;
 }
 
-FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material, const char* name,
-        UboBatchingMode batchingMode) noexcept {
-    FMaterialInstance* p =
-            mHeapAllocator.make<FMaterialInstance>(*this, material, name, batchingMode);
+FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
+        const char* name) noexcept {
+    FMaterialInstance* p = mHeapAllocator.make<FMaterialInstance>(*this, material, name);
     if (UTILS_LIKELY(p)) {
         auto pos = mMaterialInstances.emplace(material, "MaterialInstance");
         pos.first->second.insert(p);
@@ -1503,6 +1505,11 @@ Engine::FeatureLevel FEngine::setActiveFeatureLevel(FeatureLevel featureLevel) {
     return (mActiveFeatureLevel = std::max(mActiveFeatureLevel, featureLevel));
 }
 
+bool FEngine::isAsynchronousOperationSupported() const noexcept {
+    return features.backend.enable_asynchronous_operation &&
+        mConfig.asynchronousMode != AsynchronousMode::NONE;
+}
+
 #if defined(__EMSCRIPTEN__)
 void FEngine::resetBackendState() noexcept {
     getDriverApi().resetState();

filament/src/details/Engine.h

@@ -216,6 +216,8 @@ public:
         return getDriver().isStereoSupported();
     }
 
+    bool isAsynchronousOperationSupported() const noexcept;
+
     static size_t getMaxStereoscopicEyes() noexcept {
         return CONFIG_MAX_STEREOSCOPIC_EYES;
     }
@@ -345,21 +347,10 @@ public:
 
     FRenderer* createRenderer() noexcept;
 
-    // Defines whether a material instance should use UBO batching or not.
-    enum class UboBatchingMode {
-        // For default, it follows the engine settings.
-        // If UBO batching is enabled on the engine and the material domain is not SURFACE, it
-        // turns on the UBO batching. Otherwise, it turns off the UBO batching.
-        DEFAULT,
-        NO_UBO_BATCHING,
-        UBO_BATCHING
-    };
-
     FMaterialInstance* createMaterialInstance(const FMaterial* material,
             const FMaterialInstance* other, const char* name) noexcept;
 
-    FMaterialInstance* createMaterialInstance(const FMaterial* material, const char* name,
-            UboBatchingMode batchingMode) noexcept;
+    FMaterialInstance* createMaterialInstance(const FMaterial* material, const char* name) noexcept;
 
     FScene* createScene() noexcept;
     FView* createView() noexcept;
@@ -778,7 +769,6 @@ public:
                 bool assert_material_instance_texture_descriptor_set_compatible =
                         CORRECTNESS_ASSERTION_DEFAULT;
                 bool assert_texture_can_generate_mipmap = CORRECTNESS_ASSERTION_DEFAULT;
-                bool assert_camera_projection_near_far = CORRECTNESS_ASSERTION_DEFAULT;
             } debug;
             struct {
                 bool disable_gpu_frame_complete_metric = true;
@@ -798,6 +788,7 @@ public:
             bool disable_amortized_shader_compile = true;
             bool disable_handle_use_after_free_check = false;
             bool disable_heap_handle_tags = true; // FIXME: this should be false
+            bool enable_asynchronous_operation = false;
         } backend;
         struct {
             bool check_crc32_after_loading = false;
@@ -818,6 +809,9 @@ public:
             { "backend.disable_heap_handle_tags",
               "Disable Handle<> tags for heap-allocated handles.",
               &features.backend.disable_heap_handle_tags, true },
+            { "backend.enable_asynchronous_operation",
+              "Enable asynchronous operation for resource management.",
+              &features.backend.enable_asynchronous_operation, true },
             { "backend.opengl.assert_native_window_is_valid",
               "Asserts that the ANativeWindow is valid when rendering starts.",
               &features.backend.opengl.assert_native_window_is_valid, true },
@@ -848,9 +842,6 @@ public:
             { "engine.debug.assert_texture_can_generate_mipmap",
               "Assert if a texture has the correct usage set for generating mipmaps.",
               &features.engine.debug.assert_texture_can_generate_mipmap, false },
-            { "engine.debug.assert_camera_projection_near_far",
-              "Assert that the near plane is smaller than the far plane.",
-              &features.engine.debug.assert_camera_projection_near_far, false },
             { "material.check_crc32_after_loading",
               "Verify the checksum of package data when a material is loaded.",
               &features.material.check_crc32_after_loading, false },

filament/src/details/Material.cpp

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

filament/src/details/Material.h

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

filament/src/details/MaterialInstance.cpp

@@ -58,18 +58,8 @@ using namespace utils;
 namespace filament {
 
 using namespace backend;
-using UboBatchingMode = FEngine::UboBatchingMode;
 
-bool shouldEnableBatching(FEngine& engine, UboBatchingMode batchingMode, MaterialDomain domain) {
-    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
+FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name) noexcept
         : mMaterial(material),
           mDescriptorSet("MaterialInstance", material->getDescriptorSetLayout()),
           mCulling(CullingMode::BACK),
@@ -80,12 +70,9 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
           mHasScissor(false),
           mIsDoubleSided(false),
           mIsDefaultInstance(false),
-          mUseUboBatching(
-                  shouldEnableBatching(engine, batchingMode, material->getMaterialDomain())),
+          mUseUboBatching(material->useUboBatching()),
           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
@@ -160,7 +147,6 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
           mUseUboBatching(other->mUseUboBatching),
           mScissorRect(other->mScissorRect),
           mName(name ? CString(name) : other->mName) {
-    assert_invariant(!mUseUboBatching || engine.isUboBatchingEnabled());
     FEngine::DriverApi& driver = engine.getDriverApi();
     FMaterial const* const material = other->getMaterial();
 
@@ -223,31 +209,6 @@ 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());
@@ -255,10 +216,9 @@ void FMaterialInstance::commit(FEngine& engine) const {
 }
 
 void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManager) const {
-    if (mUniforms.isDirty() || mHasStreamUniformAssociations) {
+    if (mUniforms.isDirty()) {
         mUniforms.clean();
-        if (mUseUboBatching) {
-            assert_invariant(uboManager != nullptr);
+        if (isUsingUboBatching()) {
             if (!BufferAllocator::isValid(getAllocationId())) {
                 // The allocation hasn't happened yet, return.
                 return;
@@ -289,7 +249,7 @@ void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManage
     // TODO: eventually we should remove this in RELEASE builds
     fixMissingSamplers();
 
-    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
+    if (isUsingUboBatching() && !BufferAllocator::isValid(getAllocationId())) {
         return;
     }
 
@@ -439,7 +399,7 @@ void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
         return;
     }
 
-    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
+    if (isUsingUboBatching() && !BufferAllocator::isValid(getAllocationId())) {
         return;
     }
 
@@ -476,7 +436,7 @@ void FMaterialInstance::assignUboAllocation(
         const Handle<HwBufferObject>& ubHandle,
         BufferAllocator::AllocationId id,
         BufferAllocator::allocation_size_t offset) {
-    assert_invariant(mUseUboBatching);
+    assert_invariant(isUsingUboBatching());
 
     mUboData = id;
     mUboOffset = offset;

filament/src/details/MaterialInstance.h

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

filament/src/details/Renderer.cpp

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

filament/src/details/Renderer.h

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

filament/test/CMakeLists.txt

@@ -13,11 +13,25 @@ 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
-        ${CMAKE_CURRENT_SOURCE_DIR}/test_material_transformname.filamat
+        ${TEST_MATERIAL_TRANSFORM_NAME_FILAMAT}
 )
 
 add_custom_command(

filament/test/update_test_material.sh

@@ -11,7 +11,6 @@ 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
 

libs/filabridge/include/private/filament/BufferInterfaceBlock.h

@@ -170,10 +170,6 @@ 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 {
@@ -201,7 +197,6 @@ 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;

libs/filabridge/src/BufferInterfaceBlock.cpp

@@ -122,8 +122,6 @@ 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) {
@@ -150,11 +148,6 @@ 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;
@@ -170,14 +163,6 @@ 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);

libs/filagui/include/filagui/ImGuiHelper.h

@@ -17,9 +17,6 @@
 #ifndef FILAGUI_IMGUIHELPER_H_
 #define FILAGUI_IMGUIHELPER_H_
 
-#include <vector>
-#include <functional>
-
 #include <filament/Engine.h>
 #include <filament/IndexBuffer.h>
 #include <filament/Material.h>
@@ -32,6 +29,10 @@
 #include <utils/Entity.h>
 #include <utils/Path.h>
 
+#include <functional>
+#include <vector>
+#include <unordered_set>
+
 struct ImDrawData;
 struct ImGuiIO;
 struct ImGuiContext;
@@ -100,6 +101,7 @@ public:
       filament::TextureSampler mSampler;
       bool mFlipVertical = false;
       utils::Path mSettingsPath;
+      std::unordered_set<filament::Texture*> mImGuiTextures;
 };
 
 } // namespace filagui

libs/filagui/src/ImGuiHelper.cpp

@@ -148,6 +148,12 @@ 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);
@@ -189,7 +195,6 @@ 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) {
@@ -203,7 +208,6 @@ void ImGuiHelper::processImGuiCommands(ImDrawData* commands, const ImGuiIO& io)
                         .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,
@@ -213,6 +217,7 @@ void ImGuiHelper::processImGuiCommands(ImDrawData* commands, const ImGuiIO& io)
 
                 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,
@@ -222,11 +227,13 @@ void ImGuiHelper::processImGuiCommands(ImDrawData* commands, const ImGuiIO& io)
                 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);
+                    tex->UnusedFrames > 0) {
+
+                filament::Texture* ftex = (filament::Texture*)tex->TexID;
+                mEngine->destroy(ftex);
+                tex->SetTexID(ImTextureID_Invalid);
+                tex->SetStatus(ImTextureStatus_Destroyed);
+                mImGuiTextures.erase(ftex);
             }
         }
     }

libs/filamentapp/src/NativeWindowHelperLinux.cpp

@@ -38,19 +38,18 @@ void* getNativeWindow(SDL_Window* sdlWindow) {
         SDL_GetWindowSize(sdlWindow, &width, &height);
 
         // Static is used here to allocate the struct pointer for the lifetime of the program.
-        // Without static the valid struct quickyly goes out of scope, and ends with seemingly
-        // random segfaults.
+        // Without static the valid struct quickly goes out of scope, and ends with seemingly
+        // random segfaults. We must update the values on each call.
         static struct {
             struct wl_display *display;
             struct wl_surface *surface;
             uint32_t width;
             uint32_t height;
-        } wayland {
-            wmi.info.wl.display,
-            wmi.info.wl.surface,
-            static_cast<uint32_t>(width),
-            static_cast<uint32_t>(height)
-        };
+        } wayland;
+        wayland.display = wmi.info.wl.display;
+        wayland.surface = wmi.info.wl.surface;
+        wayland.width = static_cast<uint32_t>(width);
+        wayland.height = static_cast<uint32_t>(height);
         return (void *) &wayland;
 #endif
     }

libs/utils/benchmark/benchmark_memcpy.cpp

@@ -18,25 +18,125 @@
 
 #include <benchmark/benchmark.h>
 
+#include <utils/memalign.h>
 
-static void BM_memcpy(benchmark::State& state) {
-    char* src = new char[state.range(0)];
-    char* dst = new char[state.range(0)];
-    memset(src, 'x', (size_t)state.range(0));
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+#include <vector>
 
-    {
-        PerformanceCounters pc(state);
-        for (auto _ : state) {
-            memcpy(dst, src, (size_t)state.range(0));
-            benchmark::DoNotOptimize(dst);
-            benchmark::DoNotOptimize(src);
-            benchmark::ClobberMemory();
-        }
-    }
-    state.SetBytesProcessed(int64_t(state.iterations()) * int64_t(state.range(0)));
+// ------------------------------------------------------------------------------------------------
 
-    delete[] src;
-    delete[] dst;
+#ifdef __linux__
+#include <sched.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+#endif
+
+// Returns true if successful, false if failed
+static bool pinThreadToCore(int core_id) {
+#ifdef __linux__
+    cpu_set_t cpuset;
+    CPU_ZERO(&cpuset);
+    CPU_SET(core_id, &cpuset);
+
+    // 0 = current thread
+    int const result = sched_setaffinity(0, sizeof(cpu_set_t), &cpuset);
+    return (result == 0);
+#else
+    (void)core_id;
+    return false;
+#endif
 }
 
-BENCHMARK(BM_memcpy)->Range(8, 8192<<10)->Threads(1)->Threads(8);
+// ------------------------------------------------------------------------------------------------
+
+static constexpr size_t ALIGNMENT = 16 << 10;
+
+template<class T, size_t Align>
+struct AlignedAllocator {
+    using value_type = T;
+
+    template<class U>
+    struct rebind {
+        using other = AlignedAllocator<U, Align>;
+    };
+
+    AlignedAllocator() noexcept = default;
+
+    template<class U>
+    explicit AlignedAllocator(const AlignedAllocator<U, Align>&) noexcept {}
+
+    T* allocate(size_t const n) {
+        size_t const bytes = n * sizeof(T);
+        // std::aligned_alloc requires size to be a multiple of alignment.
+        size_t const aligned_bytes = ((bytes + Align - 1) / Align) * Align;
+        void* ptr = utils::aligned_alloc(aligned_bytes, Align);
+        return static_cast<T*>(ptr);
+    }
+
+    void deallocate(T* p, size_t) noexcept {
+        utils::aligned_free(p);
+    }
+};
+
+template<typename T>
+using PageAlignedVector = std::vector<T, AlignedAllocator<T, ALIGNMENT>>;
+
+// ------------------------------------------------------------------------------------------------
+
+static void BM_memcpy(benchmark::State& state) {
+    // pinThreadToCore(7);
+    int64_t const size = state.range(0);
+    PageAlignedVector<char> src(size);
+    PageAlignedVector<char> dst(size);
+
+    // make all these pages resident
+    memset(src.data(), 0, size);
+    memset(dst.data(), 0, size);
+    benchmark::ClobberMemory();
+
+    {
+        // PerformanceCounters const pc(state);
+        for (auto _: state) {
+            memcpy(dst.data(), src.data(), size);
+            benchmark::DoNotOptimize(dst);
+            benchmark::DoNotOptimize(src);
+        }
+    }
+    state.SetBytesProcessed(int64_t(state.iterations()) * int64_t(size));
+}
+
+static void BM_memset(benchmark::State& state) {
+    //pinThreadToCore(7);
+    int64_t const size = state.range(0);
+    PageAlignedVector<char> src(size);
+
+    // make all these pages resident
+    memset(src.data(), 0, size);
+    benchmark::ClobberMemory();
+
+    {
+        // PerformanceCounters const pc(state);
+        for (auto _: state) {
+            memset(src.data(), 0, size);
+            benchmark::DoNotOptimize(src);
+        }
+    }
+    state.SetBytesProcessed(int64_t(state.iterations()) * int64_t(size));
+}
+
+BENCHMARK(BM_memcpy)
+    ->DenseRange( 4<<10, 128<<10,   8<<10)
+    ->DenseRange(128<<10,  4<<20, 128<<10)
+    ->DenseRange(  4<<20, 16<<20,   1<<20)
+    ->DenseRange( 16<<20, 32<<20,   2<<20)
+    ;
+
+BENCHMARK(BM_memset)
+    ->DenseRange( 4<<10, 128<<10,   8<<10)
+    ->DenseRange(128<<10,  4<<20, 128<<10)
+    ->DenseRange(  4<<20, 16<<20,   1<<20)
+    ->DenseRange( 16<<20, 32<<20,   2<<20)
+    ;

samples/CMakeLists.txt

@@ -37,7 +37,6 @@ set(MATERIAL_SRCS
         materials/sandboxLitSolidRefractionSsr.mat
         materials/sandboxSpecGloss.mat
         materials/sandboxSubsurface.mat
-        materials/sandboxTransformName.mat
         materials/sandboxUnlit.mat
         materials/texturedLit.mat
 )

test/renderdiff/src/render.py

@@ -42,7 +42,8 @@ def _render_single_model(gltf_viewer, test_json_path, named_output_dir,
     env |= {
       'VK_ICD_FILENAMES': vk_icd,
       'VK_DRIVER_FILES': vk_icd,
-      'VK_LOADER_DEBUG': 'all',
+      # Uncomment when there's a problem finding the vk driver
+      # 'VK_LOADER_DEBUG': 'all',
     }
 
   out_name = f'{test_name}.{backend}.{model}'

test/renderdiff/src/viewer_requirements.txt

@@ -8,6 +8,6 @@ itsdangerous==2.2.0
 Jinja2==3.1.6
 MarkupSafe==3.0.2
 requests==2.32.5
-urllib3==2.5.0
+urllib3==2.6.0
 waitress==3.0.2
 Werkzeug==3.1.4

third_party/basisu/tnt/CMakeLists.txt

@@ -2,8 +2,6 @@ project(basisu)
 
 cmake_minimum_required(VERSION 3.19)
 
-set(ZSTD zstd-basisu)
-
 set(TRANSCODER_SRC
     ../transcoder/basisu_transcoder.cpp
 )
@@ -52,21 +50,28 @@ else()
     set (BASIS_CONFIG ${BASIS_CONFIG} BASISD_SUPPORT_DXT5A=0 BASISD_SUPPORT_DXT1=0)
 endif()
 
-add_library(${ZSTD} ../zstd/zstd.c)
+# If an external zstd target exists, use it.
+if(NOT TARGET zstd)
+    add_library(zstd-basisu ../zstd/zstd.c)
+    target_include_directories(zstd-basisu PUBLIC ../zstd)
+    set_target_properties(zstd-basisu PROPERTIES FOLDER ThirdParty)
+    set(ZSTD_TARGET zstd-basisu)
+else()
+    set(ZSTD_TARGET zstd)
+endif()
+
 add_library(basis_encoder ${ENCODER_SRC})
 add_library(basis_transcoder ${TRANSCODER_SRC})
 
-target_link_libraries(basis_transcoder PRIVATE ${ZSTD})
+target_link_libraries(basis_transcoder PRIVATE ${ZSTD_TARGET})
 target_link_libraries(basis_encoder basis_transcoder)
 
-target_include_directories(${ZSTD} PUBLIC ../zstd)
 target_include_directories(basis_encoder PUBLIC ../encoder)
 target_include_directories(basis_transcoder PUBLIC ../transcoder)
 
 target_compile_definitions(basis_encoder PRIVATE ${BASIS_CONFIG})
 target_compile_definitions(basis_transcoder PRIVATE ${BASIS_CONFIG})
 
-set_target_properties(${ZSTD} PROPERTIES FOLDER ThirdParty)
 set_target_properties(basis_encoder PROPERTIES FOLDER ThirdParty)
 set_target_properties(basis_transcoder PROPERTIES FOLDER ThirdParty)
 

web/filament-js/extensions.js

@@ -125,10 +125,12 @@ Filament.loadClassExtensions = function() {
 
     /// createMaterial ::method::
     /// package ::argument:: asset string, or Uint8Array, or [Buffer] with filamat contents
+    /// options ::argument:: optional dictionary with `uboBatching` key.
     /// ::retval:: an instance of [createMaterial]
-    Filament.Engine.prototype.createMaterial = function(buffer) {
+    Filament.Engine.prototype.createMaterial = function(buffer, options) {
         buffer = getBufferDescriptor(buffer);
-        const result = this._createMaterial(buffer);
+        const uboBatching = (options && options.uboBatching) || Filament.Material$UboBatchingMode.DEFAULT; // The default is DEFAULT
+        const result = this._createMaterial(buffer, uboBatching);
         buffer.delete();
         return result;
     };

web/filament-js/filament.d.ts

@@ -375,6 +375,10 @@ export class Material {
     public getName(): string;
 }
 
+export enum Material$UboBatchingMode {
+    DISABLED,
+    DEFAULT,
+}
 export class Frustum {
     constructor(pv: mat4);
     public setProjection(pv: mat4): void;
@@ -542,7 +546,7 @@ export class Engine {
     public static destroy(engine: Engine): void;
     public execute(): void;
     public createCamera(entity: Entity): Camera;
-    public createMaterial(urlOrBuffer: BufferReference): Material;
+    public createMaterial(urlOrBuffer: BufferReference, options?: { uboBatching?: Material$UboBatchingMode }): Material;
     public createRenderer(): Renderer;
     public createScene(): Scene;
     public createSwapChain(): SwapChain;

web/filament-js/jsbindings.cpp

@@ -495,8 +495,11 @@ class_<Engine>("Engine")
             (Engine* engine, utils::Entity camera) { engine->destroyCameraComponent(camera); },
             allow_raw_pointers())
 
-    .function("_createMaterial", EMBIND_LAMBDA(Material*, (Engine* engine, BufferDescriptor mbd), {
-        return Material::Builder().package(mbd.bd->buffer, mbd.bd->size).build(*engine);
+    .function("_createMaterial", EMBIND_LAMBDA(Material*, (Engine* engine, BufferDescriptor mbd, Material::UboBatchingMode uboBatching), {
+        return Material::Builder()
+                .package(mbd.bd->buffer, mbd.bd->size)
+                .uboBatching(uboBatching)
+                .build(*engine);
     }), allow_raw_pointers())
     /// destroyMaterial ::method::
     /// material ::argument:: an instance of [Material]
@@ -1360,6 +1363,10 @@ class_<Material>("Material")
         return transformName ? std::string(transformName) : std::string();
     }), allow_raw_pointers());
 
+enum_<Material::UboBatchingMode>("Material$UboBatchingMode")
+    .value("DISABLED", Material::UboBatchingMode::DISABLED)
+    .value("DEFAULT", Material::UboBatchingMode::DEFAULT);
+
 class_<MaterialInstance>("MaterialInstance")
     .function("getName", EMBIND_LAMBDA(std::string, (MaterialInstance* self), {
         return std::string(self->getName());