Use MapAsync and AllowProcessEvents for callback

This allows for static call from java, e.g.:

```
static {
    Utils.init();
}
```

NEW_RELEASE_NOTES.md

@@ -8,3 +8,5 @@ appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).
 ## Release notes for next branch cut
 
 - engine: add `View::getLastDynamicResolutionScale()` (b/457753622)
+
+- materials: Make Material Instances' UBO descriptor use dynamic offsets. [⚠️ **Recompile Materials**]

android/filament-utils-android/src/main/java/com/google/android/filament/utils/Utils.kt

@@ -22,6 +22,7 @@ object Utils {
     /**
      * Initializes the utils JNI layer. Must be called before using any utils functionality.
      */
+    @JvmStatic
     fun init() {
         // Load Filament first to ensure that the NioUtils Java class is available in the JNIEnv.
         Filament.init()

android/samples/sample-gltf-viewer/build.gradle

@@ -17,12 +17,20 @@ filamentTools {
 }
 
 // don't forget to update MainACtivity.kt when/if changing this.
-tasks.register('copyMesh', Copy) {
+tasks.register('copyDamagedHelmetGltf', Copy) {
+    from file("../../../third_party/models/DamagedHelmet/DamagedHelmet.glb")
+    into file("src/main/assets/models")
+    rename {String fileName -> "helmet.glb"}
+}
+
+// don't forget to update MainACtivity.kt when/if changing this.
+tasks.register('copyBusterGltf', Copy) {
     from "../../../third_party/models/BusterDrone"
     into "src/main/assets/models"
 }
 
-preBuild.dependsOn copyMesh
+preBuild.dependsOn copyDamagedHelmetGltf
+preBuild.dependsOn copyBusterGltf
 
 clean.doFirst {
     delete "src/main/assets"

filament/backend/include/backend/platforms/VulkanPlatform.h

@@ -444,6 +444,11 @@ protected:
 
     virtual VkExternalFenceHandleTypeFlagBits getFenceExportFlags() const noexcept;
 
+    /**
+     * Query if transient attachments are supported by the backend.
+     */
+    bool isTransientAttachmentSupported() const noexcept;
+
 private:
     friend struct VulkanPlatformPrivate;
 };

filament/backend/src/vulkan/VulkanConstants.h

@@ -214,4 +214,10 @@ constexpr static const int FVK_MAX_PIPELINE_AGE = FVK_MAX_COMMAND_BUFFERS;
 // destroying any unused pipeline object.
 static_assert(FVK_MAX_PIPELINE_AGE >= FVK_MAX_COMMAND_BUFFERS);
 
+// Indicates if the backend must be setup to allow doing a RenderDoc capture.
+//
+// If this is true, the features not supported by RenderDoc must be disabled, otherwise
+// when using RenderDoc the application will crash or will fail to do a capture.
+constexpr static const int FVK_RENDERDOC_CAPTURE_MODE = false;
+
 #endif

filament/backend/src/vulkan/VulkanContext.h

@@ -154,6 +154,10 @@ public:
         return mStagingBufferBypassEnabled;
     }
 
+    inline bool pipelineCreationFeedbackSupported() const noexcept {
+        return mPipelineCreationFeedbackSupported;
+    }
+
 private:
     VkPhysicalDeviceMemoryProperties mMemoryProperties = {};
     VkPhysicalDeviceProperties2 mPhysicalDeviceProperties = {
@@ -181,6 +185,7 @@ private:
     bool mProtectedMemorySupported = false;
     bool mIsUnifiedMemoryArchitecture = false;
     bool mStagingBufferBypassEnabled = false;
+    bool mPipelineCreationFeedbackSupported = false;
 
     fvkutils::VkFormatList mDepthStencilFormats;
     fvkutils::VkFormatList mBlittableDepthStencilFormats;

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -231,7 +231,7 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext& context,
               mPlatform->getGraphicsQueueFamilyIndex(), mPlatform->getProtectedGraphicsQueue(),
               mPlatform->getProtectedGraphicsQueueFamilyIndex(), mContext, &mSemaphoreManager),
       mPipelineLayoutCache(mPlatform->getDevice()),
-      mPipelineCache(mPlatform->getDevice()),
+      mPipelineCache(mPlatform->getDevice(), mContext),
       mStagePool(mAllocator, &mResourceManager, &mCommands, &mContext.getPhysicalDeviceLimits()),
       mBufferCache(mContext, mResourceManager, mAllocator),
       mFramebufferCache(mPlatform->getDevice()),

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -34,8 +34,41 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-VulkanPipelineCache::VulkanPipelineCache(VkDevice device)
+namespace {
-    : mDevice(device) {
+
+#if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+void printPipelineFeedbackInfo(VkPipelineCreationFeedbackCreateInfo const& feedbackInfo) {
+    VkPipelineCreationFeedback const& pipelineInfo = *feedbackInfo.pPipelineCreationFeedback;
+    if (!(pipelineInfo.flags & VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT)) {
+        return;
+    }
+
+    bool const isCacheHit =
+            (pipelineInfo.flags & VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT);
+    FVK_LOGD << "Pipeline build stats - Cache hit: " << isCacheHit
+             << ", Time: " << pipelineInfo.duration / 1000000.0 << "ms";
+
+    for (uint32_t i = 0; i < feedbackInfo.pipelineStageCreationFeedbackCount; ++i) {
+        VkPipelineCreationFeedback const& stageInfo = feedbackInfo.pPipelineStageCreationFeedbacks[i];
+        if (!(stageInfo.flags & VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT)) {
+            continue;
+        }
+
+        bool const isVertexShader = (i == 0);
+        bool const isCacheHit = (stageInfo.flags &
+                                 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT);
+        FVK_LOGD << (isVertexShader ? "Vertex" : "Fragment")
+                 << " shader build stats - Cache hit: " << isCacheHit
+                 << ", Time: " << stageInfo.duration / 1000000.0 << "ms";
+    }
+}
+#endif
+
+} // namespace
+
+VulkanPipelineCache::VulkanPipelineCache(VkDevice device, VulkanContext const& context)
+        : mDevice(device),
+          mContext(context) {
     VkPipelineCacheCreateInfo createInfo = {
         .sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO,
     };
@@ -217,14 +250,39 @@ VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() n
     }
 
 #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
-        FVK_LOGD << "vkCreateGraphicsPipelines with shaders = ("
+    FVK_LOGD << "vkCreateGraphicsPipelines with shaders = (" << shaderStages[0].module << ", "
-                 << shaderStages[0].module << ", " << shaderStages[1].module << ")";
+             << shaderStages[1].module << ")";
+
+    VkPipelineCreationFeedback stageFeedbacks[SHADER_MODULE_COUNT] = {};
+    VkPipelineCreationFeedback pipelineFeedback = {};
+    VkPipelineCreationFeedbackCreateInfo feedbackInfo = {
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT,
+        .pNext = nullptr,
+        .pPipelineCreationFeedback = &pipelineFeedback,
+        .pipelineStageCreationFeedbackCount = hasFragmentShader ? SHADER_MODULE_COUNT : 1,
+        .pPipelineStageCreationFeedbacks = stageFeedbacks,
+    };
+
+    if (mContext.pipelineCreationFeedbackSupported()) {
+        feedbackInfo.pNext = pipelineCreateInfo.pNext;
+        pipelineCreateInfo.pNext = &feedbackInfo;
+    }
 #endif
     PipelineCacheEntry cacheEntry = {
         .lastUsed = mCurrentTime,
     };
     VkResult error = vkCreateGraphicsPipelines(mDevice, mPipelineCache, 1, &pipelineCreateInfo,
             VKALLOC, &cacheEntry.handle);
+
+#if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+    FVK_LOGD << "vkCreateGraphicsPipelines with shaders = (" << shaderStages[0].module << ", "
+             << shaderStages[1].module << ")";
+
+    if (mContext.pipelineCreationFeedbackSupported()) {
+        printPipelineFeedbackInfo(feedbackInfo);
+    }
+#endif
+
     assert_invariant(error == VK_SUCCESS);
     if (error != VK_SUCCESS) {
         FVK_LOGE << "vkCreateGraphicsPipelines error " << error;

filament/backend/src/vulkan/VulkanPipelineCache.h

@@ -86,7 +86,7 @@ public:
 
     static_assert(sizeof(RasterState) == 16, "RasterState must not have implicit padding.");
 
-    VulkanPipelineCache(VkDevice device);
+    VulkanPipelineCache(VkDevice device, VulkanContext const& context);
 
     void bindLayout(VkPipelineLayout layout) noexcept;
 
@@ -210,6 +210,8 @@ private:
 
     // Current bindings for the pipeline and descriptor sets.
     PipelineKey mBoundPipeline = {};
+
+    [[maybe_unused]] VulkanContext const& mContext;
 };
 
 } // namespace filament::backend

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

@@ -228,6 +228,10 @@ ExtensionSet getDeviceExtensions(VkPhysicalDevice device) {
         VK_KHR_PORTABILITY_SUBSET_EXTENSION_NAME,
 #endif
         VK_KHR_MULTIVIEW_EXTENSION_NAME,
+
+#if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+        VK_EXT_PIPELINE_CREATION_FEEDBACK_EXTENSION_NAME,
+#endif
     };
     ExtensionSet exts;
     // Identify supported physical device extensions
@@ -865,6 +869,8 @@ Driver* VulkanPlatform::createDriver(void* sharedContext,
     if (!mImpl->mSharedContext) {
         context.mDebugUtilsSupported = setContains(instExts, VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
         context.mDebugMarkersSupported = setContains(deviceExts, VK_EXT_DEBUG_MARKER_EXTENSION_NAME);
+        context.mPipelineCreationFeedbackSupported =
+                setContains(deviceExts, VK_EXT_PIPELINE_CREATION_FEEDBACK_EXTENSION_NAME);
     } else {
         VulkanSharedContext const* scontext = (VulkanSharedContext const*) sharedContext;
         context.mDebugUtilsSupported = scontext->debugUtilsSupported;
@@ -873,6 +879,8 @@ Driver* VulkanPlatform::createDriver(void* sharedContext,
 
     // Check the availability of lazily allocated memory
     context.mLazilyAllocatedMemorySupported = false;
+    // RenderDoc doesn't support lazy allocated memory
+    if constexpr (!FVK_RENDERDOC_CAPTURE_MODE) {
         for (uint32_t i = 0, typeCount = context.mMemoryProperties.memoryTypeCount; i < typeCount;
                 ++i) {
             VkMemoryType const type = context.mMemoryProperties.memoryTypes[i];
@@ -882,6 +890,7 @@ Driver* VulkanPlatform::createDriver(void* sharedContext,
                 break;
             }
         }
+    }
 
     context.mIsUnifiedMemoryArchitecture = hasUnifiedMemoryArchitecture(context.mMemoryProperties);
 
@@ -1022,4 +1031,8 @@ VkExternalFenceHandleTypeFlagBits VulkanPlatform::getFenceExportFlags() const no
     return static_cast<VkExternalFenceHandleTypeFlagBits>(0);
 }
 
+bool VulkanPlatform::isTransientAttachmentSupported() const noexcept {
+    return mImpl->mContext.isLazilyAllocatedMemorySupported();
+}
+
 } // namespace filament::backend

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

@@ -158,6 +158,46 @@ std::pair<TextureFormat, TextureUsage> getFilamentFormatAndUsage(const AHardware
     };
 }
 
+uint32_t selectMemoryTypeForExternalImage(VkPhysicalDevice physicalDevice, VkDevice device,
+        VkImage image, uint32_t types, VkFlags requiredMemoryFlags) {
+    VkPhysicalDeviceMemoryProperties memoryProperties;
+    vkGetPhysicalDeviceMemoryProperties(physicalDevice, &memoryProperties);
+    uint32_t const memoryTypeIndex =
+            VulkanContext::selectMemoryType(memoryProperties, types, requiredMemoryFlags);
+
+    if constexpr (FVK_RENDERDOC_CAPTURE_MODE) {
+        // RenderDoc will replay external resources as non-external.
+        // Adjust properties that will trip it up when replaying, even though these are not valid.
+        // Update memory type index if necessary so that we can replay the capture.
+
+        VkMemoryRequirements imageMemoryRequirements;
+        vkGetImageMemoryRequirements(device, image, &imageMemoryRequirements);
+
+        uint32_t const imageMemoryTypeBits = imageMemoryRequirements.memoryTypeBits;
+        bool const isMemoryTypeSupported = ((1 << memoryTypeIndex) & imageMemoryTypeBits) != 0;
+        if (isMemoryTypeSupported) {
+            return memoryTypeIndex;
+        }
+
+        // Current memory type will not be replayable by RenderDoc.
+        // Attempt to change the memory type index
+        VkMemoryPropertyFlags const kRenderDocFallBackReqs = 0;
+        uint32_t commonMemoryTypeBits = types & imageMemoryTypeBits;
+        uint32_t commonTypeIndex = VulkanContext::selectMemoryType(memoryProperties,
+                commonMemoryTypeBits, kRenderDocFallBackReqs);
+        if (commonMemoryTypeBits && commonTypeIndex != VK_MAX_MEMORY_TYPES) {
+            return commonTypeIndex;
+        }
+        uint32_t imageTypeIndex = VulkanContext::selectMemoryType(memoryProperties,
+                imageMemoryTypeBits, kRenderDocFallBackReqs);
+        if (imageTypeIndex != VK_MAX_MEMORY_TYPES) {
+            return imageTypeIndex;
+        }
+    }
+
+    return memoryTypeIndex;
+}
+
 } // namespace
 
 VulkanPlatformAndroid::ExternalImageVulkanAndroid::~ExternalImageVulkanAndroid() {
@@ -354,8 +394,6 @@ VulkanPlatform::ImageData VulkanPlatformAndroid::createVkImageFromExternal(
             .image = image,
             .buffer = VK_NULL_HANDLE,
         };
-        VkPhysicalDeviceMemoryProperties memoryProperties;
-        vkGetPhysicalDeviceMemoryProperties(physicalDevice, &memoryProperties);
         VkMemoryPropertyFlags requiredMemoryFlags =
                 !isExternal && any(metadata.filamentUsage & TextureUsage::UPLOADABLE)
                         ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
@@ -365,8 +403,8 @@ VulkanPlatform::ImageData VulkanPlatformAndroid::createVkImageFromExternal(
             requiredMemoryFlags |= VK_MEMORY_PROPERTY_PROTECTED_BIT;
         }
 
-        uint32_t const memoryTypeIndex = VulkanContext::selectMemoryType(memoryProperties,
+        uint32_t const memoryTypeIndex = selectMemoryTypeForExternalImage(physicalDevice, device,
-                metadata.memoryTypeBits, requiredMemoryFlags);
+                image, metadata.memoryTypeBits, requiredMemoryFlags);
 
         VkMemoryAllocateInfo const allocInfo = {
             .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,

filament/backend/src/webgpu/WebGPUBufferBase.cpp

@@ -63,7 +63,7 @@ WebGPUBufferBase::WebGPUBufferBase(wgpu::Device const& device, const wgpu::Buffe
 // WebGPU requires that the size of the data copied from the staging buffer to the GPU buffer is a
 // multiple of 4. This function handles cases where the buffer descriptor's size is not a multiple
 // of 4 by padding with zeros.
-void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
+void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor&& bufferDescriptor,
         const uint32_t byteOffset, wgpu::Device const& device,
         WebGPUQueueManager* const webGPUQueueManager) {
     FILAMENT_CHECK_PRECONDITION(bufferDescriptor.buffer)
@@ -89,24 +89,49 @@ void WebGPUBufferBase::updateGPUBuffer(BufferDescriptor const& bufferDescriptor,
     wgpu::BufferDescriptor descriptor{
         .label = "Filament WebGPU Staging Buffer",
         .usage = wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc,
-        .size = stagingBufferSize,
+        .size = stagingBufferSize};
-        .mappedAtCreation = true };
     wgpu::Buffer stagingBuffer = device.CreateBuffer(&descriptor);
 
-    void* mappedRange = stagingBuffer.GetMappedRange();
+    struct UserData final {
-    memcpy(mappedRange, bufferDescriptor.buffer, bufferDescriptor.size);
+        uint32_t byteOffset;
-
+        size_t stagingBufferSize;
-    // Make sure the padded memory is set to 0 to have deterministic behaviors
+        size_t remainder;
-    if (remainder != 0) {
+        BufferDescriptor srcBufferDescriptor;
-        uint8_t* paddingStart = static_cast<uint8_t*>(mappedRange) + bufferDescriptor.size;
+        wgpu::Buffer stagingBuffer;
-        memset(paddingStart, 0, FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS - remainder);
+        WebGPUQueueManager* const webGPUQueueManager;
+        wgpu::Buffer dstBuffer;
+    };
+    auto userData = std::make_unique<UserData>(UserData{
+        .byteOffset = byteOffset,
+        .stagingBufferSize = stagingBufferSize,
+        .remainder = remainder,
+        .srcBufferDescriptor = std::move(bufferDescriptor),
+        .stagingBuffer = stagingBuffer,
+        .webGPUQueueManager = webGPUQueueManager,
+        .dstBuffer = mBuffer});
+    stagingBuffer.MapAsync(
+            wgpu::MapMode::Write, 0, stagingBufferSize, wgpu::CallbackMode::AllowProcessEvents,
+            [](wgpu::MapAsyncStatus status, const char* message, UserData* userdata) {
+                std::unique_ptr<UserData> data(static_cast<UserData*>(userdata));
+                if (UTILS_LIKELY(status == wgpu::MapAsyncStatus::Success)) {
+                    void* mappedRange = data->stagingBuffer.GetMappedRange();
+                    memcpy(mappedRange, data->srcBufferDescriptor.buffer,
+                            data->srcBufferDescriptor.size);
+                    if (data->remainder != 0) {
+                        uint8_t* paddingStart =
+                                static_cast<uint8_t*>(mappedRange) + data->srcBufferDescriptor.size;
+                        memset(paddingStart, 0,
+                                FILAMENT_WEBGPU_BUFFER_SIZE_MODULUS - data->remainder);
                     }
-
+                    data->stagingBuffer.Unmap();
-    stagingBuffer.Unmap();
+                    data->webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(
-
+                            data->stagingBuffer, 0, data->dstBuffer, data->byteOffset,
-    // Copy the staging buffer contents to the destination buffer.
+                            data->stagingBufferSize);
-    webGPUQueueManager->getCommandEncoder().CopyBufferToBuffer(stagingBuffer, 0, mBuffer,
+                } else {
-            byteOffset, stagingBufferSize);
+                    FWGPU_LOGE << "Failed to map staging buffer for readPixels: " << message;
+                }
+            },
+            userData.release());
 }
 
 } // namespace filament::backend

filament/backend/src/webgpu/WebGPUBufferBase.h

@@ -39,7 +39,7 @@ public:
      * happen after draw commands encoded in the encoder. Submitting any commands up to this point
      * ensures the calls happen in the expected sequence.
      */
-    void updateGPUBuffer(BufferDescriptor const&, uint32_t byteOffset, wgpu::Device const& device,
+    void updateGPUBuffer(BufferDescriptor&&, uint32_t byteOffset, wgpu::Device const& device,
             WebGPUQueueManager* const webGPUQueueManager);
 
     [[nodiscard]] wgpu::Buffer const& getBuffer() const { return mBuffer; }

filament/backend/src/webgpu/WebGPUDriver.cpp

@@ -856,7 +856,7 @@ void WebGPUDriver::updateIndexBuffer(Handle<HwIndexBuffer> indexBufferHandle,
     // draw calls are made.
     flush();
     handleCast<WebGPUIndexBuffer>(indexBufferHandle)
-            ->updateGPUBuffer(bufferDescriptor, byteOffset, mDevice, &mQueueManager);
+            ->updateGPUBuffer(std::move(bufferDescriptor), byteOffset, mDevice, &mQueueManager);
     scheduleDestroy(std::move(bufferDescriptor));
 }
 
@@ -867,14 +867,14 @@ void WebGPUDriver::updateBufferObject(Handle<HwBufferObject> bufferObjectHandle,
     // draw calls are made.
     flush();
     handleCast<WebGPUBufferObject>(bufferObjectHandle)
-            ->updateGPUBuffer(bufferDescriptor, byteOffset, mDevice, &mQueueManager);
+            ->updateGPUBuffer(std::move(bufferDescriptor), byteOffset, mDevice, &mQueueManager);
     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(std::move(bufferDescriptor), byteOffset, mDevice, &mQueueManager);
     scheduleDestroy(std::move(bufferDescriptor));
 }
 

filament/backend/src/webgpu/WebGPUQueueManager.cpp

@@ -21,6 +21,7 @@
 #include <chrono>
 #include <cstdint>
 #include <thread>
+#include <iostream>
 
 namespace filament::backend {
 
@@ -58,6 +59,7 @@ WebGPUQueueManager::WebGPUQueueManager(wgpu::Device const& device)
 WebGPUQueueManager::~WebGPUQueueManager() = default;
 
 wgpu::CommandEncoder WebGPUQueueManager::getCommandEncoder() {
+    // std::unique_lock<std::mutex> lock(mLock);
     if (!mCommandEncoder) {
         wgpu::CommandEncoderDescriptor commandEncoderDescriptor = {
             .label = "Filament Command Encoder",
@@ -90,7 +92,9 @@ std::shared_ptr<WebGPUSubmissionState> WebGPUQueueManager::getLatestSubmissionSt
 }
 
 void WebGPUQueueManager::submit() {
+    // std::unique_lock<std::mutex> lock(mLock);
     if (!mCommandEncoder) {
+        std::cout << "Run Yu: no mCommandEncoder found!\n";
         return;
     }
 

filament/backend/src/webgpu/WebGPUQueueManager.h

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

filament/src/PostProcessManager.cpp

@@ -258,6 +258,10 @@ void PostProcessManager::bindPerRenderableDescriptorSet(DriverApi& driver) const
             { { 0, 0 }, driver });
 }
 
+UboManager* PostProcessManager::getUboManager() const noexcept {
+    return mEngine.getUboManager();
+}
+
 UTILS_NOINLINE
 void PostProcessManager::registerPostProcessMaterial(std::string_view const name,
         StaticMaterialInfo const& info) {
@@ -513,7 +517,7 @@ UTILS_NOINLINE
 void PostProcessManager::commitAndRenderFullScreenQuad(DriverApi& driver,
         FrameGraphResources::RenderPassInfo const& out, FMaterialInstance const* mi,
         PostProcessVariant const variant) const noexcept {
-    mi->commit(driver);
+    mi->commit(driver, getUboManager());
     mi->use(driver);
     FMaterial const* const ma = mi->getMaterial();
     PipelineState const pipeline = getPipelineState(ma, variant);
@@ -645,7 +649,7 @@ PostProcessManager::StructurePassOutput PostProcessManager::structure(FrameGraph
                     auto th = driver.createTextureView(in, level, 1);
                     mi->setParameter("depth", th, SamplerParams{
                         .filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST });
-                    mi->commit(driver);
+                    mi->commit(driver, getUboManager());
                     mi->use(driver);
                     renderFullScreenQuad(out, pipeline, driver);
                     DescriptorSet::unbind(driver, DescriptorSetBindingPoints::PER_MATERIAL);
@@ -1079,7 +1083,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::screenSpaceAmbientOcclusion(
                 mi->setParameter("ssctRayCount",
                         float2{ options.ssct.rayCount, 1.0f / float(options.ssct.rayCount) });
 
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto pipeline = getPipelineState(ma);
@@ -1200,7 +1204,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::bilateralBlurPass(FrameGraph
                 mi->setParameter("sampleCount", kGaussianCount);
                 mi->setParameter("farPlaneOverEdgeDistance", -zf / config.bilateralThreshold);
 
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto pipeline = getPipelineState(ma);
@@ -1895,7 +1899,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::dof(FrameGraph& fg,
                             .filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST });
                     mi->setParameter("weightScale", 0.5f / float(1u << level));   // FIXME: halfres?
                     mi->setParameter("texelSize", float2{ 1.0f / w, 1.0f / h });
-                    mi->commit(driver);
+                    mi->commit(driver, getUboManager());
                     mi->use(driver);
 
                     renderFullScreenQuad(out, pipeline, driver);
@@ -2425,7 +2429,7 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
                     mi->setParameter("source", hwOutView, SamplerParams{
                             .filterMag = SamplerMagFilter::LINEAR,
                             .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST});
-                    mi->commit(driver);
+                    mi->commit(driver, getUboManager());
                     mi->use(driver);
                     renderFullScreenQuad(hwDstRT, pipeline, driver);
                     DescriptorSet::unbind(driver, DescriptorSetBindingPoints::PER_MATERIAL);
@@ -2533,7 +2537,7 @@ void PostProcessManager::colorGradingPrepareSubpass(DriverApi& driver,
     FMaterialInstance* const mi =
             configureColorGradingMaterial(material, colorGrading, colorGradingConfig,
                     vignetteOptions, width, height);
-    mi->commit(driver);
+    mi->commit(driver, getUboManager());
 }
 
 void PostProcessManager::colorGradingSubpass(DriverApi& driver,
@@ -2566,7 +2570,7 @@ void PostProcessManager::customResolvePrepareSubpass(DriverApi& driver, CustomRe
     auto [mi, fixedIndex] = mMaterialInstanceManager.getFixedMaterialInstance(ma);
     mFixedMaterialInstanceIndex.customResolve = fixedIndex;
     mi->setParameter("direction", op == CustomResolveOp::COMPRESS ? 1.0f : -1.0f),
-    mi->commit(driver);
+    mi->commit(driver, getUboManager());
     material.getMaterial(mEngine);
 }
 
@@ -2619,7 +2623,7 @@ void PostProcessManager::clearAncillaryBuffersPrepare(DriverApi& driver) noexcep
     auto ma = material.getMaterial(mEngine, PostProcessVariant::OPAQUE);
     auto [mi, fixedIndex] = mMaterialInstanceManager.getFixedMaterialInstance(ma);
     mFixedMaterialInstanceIndex.clearDepth = fixedIndex;
-    mi->commit(driver);
+    mi->commit(driver, getUboManager());
     material.getMaterial(mEngine);
 }
 
@@ -3111,7 +3115,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::taa(FrameGraph& fg,
                         mat4f{ historyProjection * inverse(current.projection) } *
                         normalizedToClip);
 
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 if (colorGradingConfig.asSubpass) {
@@ -3196,7 +3200,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::rcas(
                 mi->setParameter("resolution", float4{
                         outputDesc.width, outputDesc.height,
                         1.0f / outputDesc.width, 1.0f / outputDesc.height });
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto pipeline = getPipelineState(material.getMaterial(mEngine), variant);
@@ -3287,7 +3291,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscaleBilinear(FrameGraph&
                         float(vp.width)  / inputDesc.width,
                         float(vp.height) / inputDesc.height
                 });
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto out = resources.getRenderPassInfo();
@@ -3374,7 +3378,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscaleSGSR1(FrameGraph& fg,
                         float(inputDesc.height)
                 });
 
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto const out = resources.getRenderPassInfo();
@@ -3469,7 +3473,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscaleFSR1(FrameGraph& fg,
                     mi->setParameter("resolution",
                             float4{ outputDesc.width, outputDesc.height,
                                     1.0f / outputDesc.width, 1.0f / outputDesc.height });
-                    mi->commit(driver);
+                    mi->commit(driver, getUboManager());
                     mi->use(driver);
                 }
 
@@ -3496,7 +3500,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscaleFSR1(FrameGraph& fg,
                             float(vp.width)  / inputDesc.width,
                             float(vp.height) / inputDesc.height
                     });
-                    mi->commit(driver);
+                    mi->commit(driver, getUboManager());
                     mi->use(driver);
                 }
 
@@ -3582,7 +3586,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::blit(FrameGraph& fg, bool co
                 if (layer) {
                     mi->setParameter("layerIndex", layer);
                 }
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto pipeline = getPipelineState(ma);
@@ -3840,7 +3844,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::vsmMipmapPass(FrameGraph& fg
                 });
                 mi->setParameter("layer", uint32_t(layer));
                 mi->setParameter("uvscale", 1.0f / float(dim));
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 renderFullScreenQuadWithScissor(out, pipeline, scissor, driver);
@@ -3942,7 +3946,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::debugCombineArrayTexture(Fra
                         float(vp.width) / inputDesc.width,
                         float(vp.height) / inputDesc.height
                     });
-                mi->commit(driver);
+                mi->commit(driver, getUboManager());
                 mi->use(driver);
 
                 auto pipeline = getPipelineState(ma);
@@ -3959,7 +3963,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::debugCombineArrayTexture(Fra
                 // Render all layers of the texture to the screen side-by-side.
                 for (uint32_t i = 0; i < inputTextureDesc.depth; ++i) {
                     mi->setParameter("layerIndex", i);
-                    mi->commit(driver);
+                    mi->commit(driver, getUboManager());
                     renderFullScreenQuad(out, pipeline, driver);
                     DescriptorSet::unbind(driver, DescriptorSetBindingPoints::PER_MATERIAL);
                     // From the second draw, don't clear the targetbuffer.

filament/src/PostProcessManager.h

@@ -65,6 +65,7 @@ class FMaterialInstance;
 class FrameGraph;
 class RenderPass;
 class RenderPassBuilder;
+class UboManager;
 struct CameraInfo;
 
 class PostProcessManager {
@@ -422,6 +423,8 @@ private:
         return getMaterialInstance(ma);
     }
 
+    UboManager* getUboManager() const noexcept;
+
     backend::RenderPrimitiveHandle mFullScreenQuadRph;
     backend::VertexBufferInfoHandle mFullScreenQuadVbih;
     backend::DescriptorSetLayoutHandle mPerRenderableDslh;

filament/src/RenderPass.cpp

@@ -1012,6 +1012,8 @@ void RenderPass::Executor::execute(FEngine const& engine, DriverApi& driver,
                     uint32_t const index = (first->key & CUSTOM_INDEX_MASK) >> CUSTOM_INDEX_SHIFT;
                     assert_invariant(index < mCustomCommands.size());
                     pCustomCommands[index]();
+                    currentPipeline = {};
+                    currentPrimitiveHandle ={};
                     continue;
                 }
 
@@ -1069,7 +1071,13 @@ void RenderPass::Executor::execute(FEngine const& engine, DriverApi& driver,
                     pipeline.pipelineLayout.setLayout[+DescriptorSetBindingPoints::PER_MATERIAL] =
                             ma->getDescriptorSetLayout(info.materialVariant).getHandle();
 
-
+                    // If we have a ColorPassDescriptorSet we use it to bind the per-view
+                    // descriptor-set (ideally only if it changed).
+                    // If we don't, it means the descriptor-set is already bound and the layout we
+                    // got from the material above should match. This is the case for situations
+                    // where we have a known per-view descriptor-set layout,
+                    // e.g.: postfx, shadow-maps, ssr and structure passes.
+                    if (mColorPassDescriptorSet) {
                         if (UTILS_UNLIKELY(ma->getMaterialDomain() == MaterialDomain::POST_PROCESS)) {
                             // It is possible to get a post-process material here (even though it's
                             // not technically a public API yet, it is used by the IBLPrefilterLibrary.
@@ -1077,18 +1085,20 @@ void RenderPass::Executor::execute(FEngine const& engine, DriverApi& driver,
                             // to set the post-process descriptor-set.
                             engine.getPostProcessManager().bindPostProcessDescriptorSet(driver);
                         } else {
-                        // If we have a ColorPassDescriptorSet we use it to bind the per-view
-                        // descriptor-set (ideally only if it changed). If we don't, it means
-                        // the descriptor-set is already bound and the layout we got from the
-                        // material above should match. This is the case for situations where we
-                        // have a known per-view descriptor-set layout, e.g.: shadow-maps, ssr and
-                        // structure passes.
-                        if (mColorPassDescriptorSet) {
                             // We have a ColorPassDescriptorSet, we need to go through it for binding
                             // the per-view descriptor-set because its layout can change based on the
                             // material.
                             mColorPassDescriptorSet->bind(driver, ma->getPerViewLayoutIndex());
                         }
+                    } else {
+                        // if we're here it means the per-view descriptor set is constant and
+                        // already set. This will be the case for postfx, ssr, structure and
+                        // shadow passes. All these passes use a static descriptor set layout
+                        // (albeit potentially different for each pass). In particular the
+                        // per-view UBO must be compatible for all material domains.
+                        // This is the case by construction for postfx, ssr. However, shadows
+                        // and structure have their own UBO, but it's content is (must be)
+                        // compatible with POST_PROCESS and COMPUTE materials.
                     }
 
                     // Each MaterialInstance has its own descriptor set. This binds it.

filament/src/details/BufferAllocator.cpp

@@ -142,7 +142,11 @@ void BufferAllocator::retire(AllocationId id) {
     InternalSlotNode* targetNode = getNodeById(id);
     assert_invariant(targetNode != nullptr);
 
-    targetNode->slot.isAllocated = false;
+    Slot& slot = targetNode->slot;
+    slot.isAllocated = false;
+    if (slot.gpuUseCount == 0) {
+        mHasPendingFrees = true;
+    }
 }
 
 void BufferAllocator::acquireGpu(AllocationId id) {
@@ -157,10 +161,18 @@ void BufferAllocator::releaseGpu(AllocationId id) {
     assert_invariant(targetNode != nullptr);
     assert_invariant(targetNode->slot.gpuUseCount > 0);
 
-    targetNode->slot.gpuUseCount--;
+    Slot& slot = targetNode->slot;
+    slot.gpuUseCount--;
+    if (slot.gpuUseCount == 0 && !slot.isAllocated) {
+        mHasPendingFrees = true;
+    }
 }
 
 void BufferAllocator::releaseFreeSlots() {
+    if (!mHasPendingFrees) {
+        return;
+    }
+
     auto curr = mSlotPool.begin();
     while (curr != mSlotPool.end()) {
         if (!curr->slot.isFree()) {
@@ -199,6 +211,7 @@ void BufferAllocator::releaseFreeSlots() {
 
         curr = next;
     }
+    mHasPendingFrees = false;
 }
 
 BufferAllocator::allocation_size_t BufferAllocator::getTotalSize() const noexcept {

filament/src/details/BufferAllocator.h

@@ -113,8 +113,9 @@ private:
 
     [[nodiscard]] InternalSlotNode* getNodeById(AllocationId id) const noexcept;
 
+    bool mHasPendingFrees = false;
     allocation_size_t mTotalSize;
-    const allocation_size_t mSlotSize; // Size of a single slot in bytes.
+    const allocation_size_t mSlotSize; // Size of a single slot in bytes
     std::list<InternalSlotNode> mSlotPool; // All slots, including both allocated and freed
     std::multimap</*slot size*/allocation_size_t, InternalSlotNode*> mFreeList;
     std::unordered_map</*slot offset*/allocation_size_t, InternalSlotNode*> mOffsetMap;

filament/src/details/Engine.cpp

@@ -468,11 +468,26 @@ void FEngine::init() {
         }
         mDefaultMaterial = downcast(defaultMaterialBuilder.build(*this));
     }
+
     // We must commit the default material instance here. It may not be used in a scene, but its
     // descriptor set may still be used for shared variants.
-    mDefaultMaterial->getDefaultInstance()->commit(driverApi);
+    //
+    // Note that this material instance is instantiated before the creation of UboManager, so at
+    // this point `isUboBatchingEnabled` is `false`, and it will fall back to individual UBO
+    // automatically.
+    mDefaultMaterial->getDefaultInstance()->commit(driverApi, mUboManager);
 
     if (UTILS_UNLIKELY(getSupportedFeatureLevel() >= FeatureLevel::FEATURE_LEVEL_1)) {
+        // UBO batching is not supported in feature level 0
+        if (features.material.enable_material_instance_uniform_batching) {
+            // Ubo size of each material instance is at least 16 bytes.
+            constexpr BufferAllocator::allocation_size_t minSlotSize = 16;
+            auto const uboOffsetAlignment = static_cast<BufferAllocator::allocation_size_t>(
+                    driverApi.getUniformBufferOffsetAlignment());
+            BufferAllocator::allocation_size_t slotSize = std::max(minSlotSize, uboOffsetAlignment);
+            mUboManager = new UboManager(getDriverApi(), slotSize, mConfig.sharedUboInitialSizeInBytes);
+        }
+
         mDefaultColorGrading = downcast(ColorGrading::Builder().build(*this));
 
         constexpr float3 dummyPositions[1] = {};
@@ -653,6 +668,12 @@ void FEngine::shutdown() {
 
     driver.destroyRenderTarget(std::move(mDefaultRenderTarget));
 
+    if (isUboBatchingEnabled()) {
+        mUboManager->terminate(driver);
+        delete mUboManager;
+        mUboManager = nullptr;
+    }
+
     /*
      * Shutdown the backend...
      */
@@ -693,18 +714,31 @@ void FEngine::prepare() {
     // UBOs that are visible only. It's not such a big issue because the actual upload() is
     // skipped if the UBO hasn't changed. Still we could have a lot of these.
     DriverApi& driver = getDriverApi();
+    const bool useUboBatching = isUboBatchingEnabled();
 
+    if (useUboBatching) {
+        assert_invariant(mUboManager != nullptr);
+
+        mUboManager->beginFrame(driver, mMaterialInstances);
+    }
+
+    UboManager* uboManager = mUboManager;
     for (auto& materialInstanceList: mMaterialInstances) {
-        materialInstanceList.second.forEach([&driver](FMaterialInstance* item) {
+        materialInstanceList.second.forEach([&driver, uboManager](FMaterialInstance* item) {
             // 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);
+                item->commit(driver, uboManager);
             }
         });
     }
 
+    if (useUboBatching) {
+        assert_invariant(mUboManager != nullptr);
+        getUboManager()->finishBeginFrame(getDriverApi());
+    }
+
     mMaterials.forEach([](FMaterial* material) {
 #if FILAMENT_ENABLE_MATDBG // NOLINT(*-include-cleaner)
         material->checkProgramEdits();
@@ -721,6 +755,13 @@ void FEngine::gc() {
     mCameraManager.gc(*this, em);
 }
 
+void FEngine::submitFrame() {
+    if (isUboBatchingEnabled()) {
+        DriverApi& driver = getDriverApi();
+        getUboManager()->endFrame(driver, getMaterialInstanceResourceList());
+    }
+}
+
 void FEngine::flush() {
     // flush the command buffer
     flushCommandBuffer(mCommandBufferQueue);
@@ -951,9 +992,10 @@ FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
     return p;
 }
 
-FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material,
+FMaterialInstance* FEngine::createMaterialInstance(const FMaterial* material, const char* name,
-                                                   const char* name) noexcept {
+        UboBatchingMode batchingMode) noexcept {
-    FMaterialInstance* p = mHeapAllocator.make<FMaterialInstance>(*this, material, name);
+    FMaterialInstance* p =
+            mHeapAllocator.make<FMaterialInstance>(*this, material, name, batchingMode);
     if (UTILS_LIKELY(p)) {
         auto pos = mMaterialInstances.emplace(material, "MaterialInstance");
         pos.first->second.insert(p);
@@ -1243,6 +1285,11 @@ UTILS_NOINLINE
 bool FEngine::destroy(const FMaterialInstance* p) {
     if (p == nullptr) return true;
 
+    if (p->isUsingUboBatching()) {
+        assert_invariant(isUboBatchingEnabled());
+        mUboManager->retireSlot(p->getAllocationId());
+    }
+
     // Check that the material instance we're destroying is not in use in the RenderableManager
     // To do this, we currently need to inspect all render primitives in the RenderableManager
     EntityManager const& em = mEntityManager;

filament/src/details/Engine.h

@@ -21,16 +21,17 @@
 
 #include "Allocators.h"
 #include "DFG.h"
-#include "PostProcessManager.h"
-#include "ResourceList.h"
 #include "HwDescriptorSetLayoutFactory.h"
 #include "HwVertexBufferInfoFactory.h"
 #include "MaterialCache.h"
+#include "PostProcessManager.h"
+#include "ResourceList.h"
+#include "UboManager.h"
 
 #include "components/CameraManager.h"
 #include "components/LightManager.h"
-#include "components/TransformManager.h"
 #include "components/RenderableManager.h"
+#include "components/TransformManager.h"
 
 #include "ds/DescriptorSetLayout.h"
 
@@ -263,6 +264,14 @@ public:
         return mBackend;
     }
 
+    bool isUboBatchingEnabled() const noexcept {
+        return mUboManager != nullptr;
+    }
+
+    UboManager* getUboManager() noexcept {
+        return mUboManager;
+    }
+
     Platform* getPlatform() const noexcept {
         return mPlatform;
     }
@@ -336,11 +345,21 @@ public:
 
     FRenderer* createRenderer() noexcept;
 
+    // Defines whether a material instance should use UBO batching or not.
+    enum class UboBatchingMode {
+        // For default, it follows the engine settings.
+        // If UBO batching is enabled on the engine and the material domain is not SURFACE, it
+        // turns on the UBO batching. Otherwise, it turns off the UBO batching.
+        DEFAULT,
+        NO_UBO_BATCHING,
+        UBO_BATCHING
+    };
+
     FMaterialInstance* createMaterialInstance(const FMaterial* material,
             const FMaterialInstance* other, const char* name) noexcept;
 
-    FMaterialInstance* createMaterialInstance(const FMaterial* material,
+    FMaterialInstance* createMaterialInstance(const FMaterial* material, const char* name,
-                                              const char* name) noexcept;
+            UboBatchingMode batchingMode) noexcept;
 
     FScene* createScene() noexcept;
     FView* createView() noexcept;
@@ -446,6 +465,7 @@ public:
 
     void prepare();
     void gc();
+    void submitFrame();
 
     using ShaderContent = utils::FixedCapacityVector<uint8_t>;
 
@@ -650,6 +670,7 @@ private:
 
     uint32_t mFlushCounter = 0;
 
+    UboManager* mUboManager = nullptr;
     RootArenaScope::Arena mPerRenderPassArena;
     HeapAllocatorArena mHeapAllocator;
 

filament/src/details/Material.cpp

@@ -73,6 +73,7 @@ namespace filament {
 using namespace backend;
 using namespace filaflat;
 using namespace utils;
+using UboBatchingMode = FEngine::UboBatchingMode;
 
 struct Material::BuilderDetails {
     const void* mPayload = nullptr;
@@ -217,16 +218,17 @@ void FMaterial::terminate(FEngine& engine) {
 
 filament::DescriptorSetLayout const& FMaterial::getPerViewDescriptorSetLayout(
         Variant const variant, bool const useVsmDescriptorSetLayout) const noexcept {
+    if (mDefinition.materialDomain == MaterialDomain::SURFACE) {
+        // `variant` is only sensical for MaterialDomain::SURFACE
         if (Variant::isValidDepthVariant(variant)) {
-        assert_invariant(mDefinition.materialDomain == MaterialDomain::SURFACE);
             return mEngine.getPerViewDescriptorSetLayoutDepthVariant();
         }
         if (Variant::isSSRVariant(variant)) {
-        assert_invariant(mDefinition.materialDomain == MaterialDomain::SURFACE);
             return mEngine.getPerViewDescriptorSetLayoutSsrVariant();
         }
+    }
+    // mDefinition.perViewDescriptorSetLayout{Vsm} is already resolved for MaterialDomain
     if (useVsmDescriptorSetLayout) {
-        assert_invariant(mDefinition.materialDomain == MaterialDomain::SURFACE);
         return mDefinition.perViewDescriptorSetLayoutVsm;
     }
     return mDefinition.perViewDescriptorSetLayout;
@@ -280,14 +282,14 @@ FMaterialInstance* FMaterial::createInstance(const char* name) const noexcept {
         return FMaterialInstance::duplicate(mDefaultMaterialInstance, name);
     } else {
         // but if we don't, just create an instance with all the default parameters
-        return mEngine.createMaterialInstance(this, name);
+        return mEngine.createMaterialInstance(this, name, UboBatchingMode::DEFAULT);
     }
 }
 
 FMaterialInstance* FMaterial::getDefaultInstance() noexcept {
     if (UTILS_UNLIKELY(!mDefaultMaterialInstance)) {
-        mDefaultMaterialInstance = mEngine.createMaterialInstance(this,
+        mDefaultMaterialInstance =
-                mDefinition.name.c_str());
+                mEngine.createMaterialInstance(this, mDefinition.name.c_str(), UboBatchingMode::DEFAULT);
         mDefaultMaterialInstance->setDefaultInstance(true);
     }
     return mDefaultMaterialInstance;

filament/src/details/Material.h

@@ -108,7 +108,7 @@ public:
             backend::CallbackHandler* handler,
             utils::Invocable<void(Material*)>&& callback) noexcept;
 
-    // Create an instance of this material
+    // Creates an instance of this material, specifying the batching mode.
     FMaterialInstance* createInstance(const char* name) const noexcept;
 
     bool hasParameter(const char* name) const noexcept;

filament/src/details/MaterialInstance.cpp

@@ -58,15 +58,19 @@ using namespace utils;
 namespace filament {
 
 using namespace backend;
+using UboBatchingMode = FEngine::UboBatchingMode;
 
-FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
+bool shouldEnableBatching(FEngine& engine, UboBatchingMode batchingMode, MaterialDomain domain) {
-        const char* name) noexcept
+    if (batchingMode != UboBatchingMode::DEFAULT) {
-        : FMaterialInstance(engine, material, name,
+        return batchingMode == UboBatchingMode::UBO_BATCHING;
-                  engine.features.material.enable_material_instance_uniform_batching) {
     }
 
-FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
+    return engine.isUboBatchingEnabled() && domain == MaterialDomain::SURFACE;
-                      const char* name, bool useUboBatching) noexcept : mMaterial(material),
+}
+
+FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name,
+        UboBatchingMode batchingMode) noexcept
+        : mMaterial(material),
           mDescriptorSet("MaterialInstance", material->getDescriptorSetLayout()),
           mCulling(CullingMode::BACK),
           mShadowCulling(CullingMode::BACK),
@@ -76,10 +80,12 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
           mHasScissor(false),
           mIsDoubleSided(false),
           mIsDefaultInstance(false),
-          mUseUboBatching(useUboBatching),
+          mUseUboBatching(
+                  shouldEnableBatching(engine, batchingMode, material->getMaterialDomain())),
           mTransparencyMode(TransparencyMode::DEFAULT),
           mName(name ? CString(name) : material->getName()) {
-
+    FILAMENT_CHECK_PRECONDITION(!mUseUboBatching || engine.isUboBatchingEnabled())
+            << "UBO batching is not enabled.";
     FEngine::DriverApi& driver = engine.getDriverApi();
 
     // even if the material doesn't have any parameters, we allocate a small UBO because it's
@@ -153,7 +159,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
           mUseUboBatching(other->mUseUboBatching),
           mScissorRect(other->mScissorRect),
           mName(name ? CString(name) : other->mName) {
-
+    assert_invariant(!mUseUboBatching || engine.isUboBatchingEnabled());
     FEngine::DriverApi& driver = engine.getDriverApi();
     FMaterial const* const material = other->getMaterial();
 
@@ -193,8 +199,8 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
     }
 }
 
-FMaterialInstance* FMaterialInstance::duplicate(
+FMaterialInstance* FMaterialInstance::duplicate(FMaterialInstance const* other,
-        FMaterialInstance const* other, const char* name) noexcept {
+        const char* name) noexcept {
     FMaterial const* const material = other->getMaterial();
     FEngine& engine = material->getEngine();
     return engine.createMaterialInstance(material, other, name);
@@ -238,15 +244,21 @@ void FMaterialInstance::commitStreamUniformAssociations(FEngine::DriverApi& driv
 
 void FMaterialInstance::commit(FEngine& engine) const {
     if (UTILS_LIKELY(mMaterial->getMaterialDomain() != MaterialDomain::SURFACE)) {
-        commit(engine.getDriverApi());
+        commit(engine.getDriverApi(), engine.getUboManager());
     }
 }
 
-void FMaterialInstance::commit(FEngine::DriverApi& driver) const {
+void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManager) const {
     if (mUniforms.isDirty() || mHasStreamUniformAssociations) {
         mUniforms.clean();
         if (mUseUboBatching) {
-            // TODO: update the content by `copyToMemoryMappedBuffer`
+            assert_invariant(uboManager != nullptr);
+            if (!BufferAllocator::isValid(getAllocationId())) {
+                // The allocation hasn't happened yet, return.
+                return;
+            }
+
+            uboManager->updateSlot(driver, getAllocationId(), mUniforms.toBufferDescriptor(driver));
         }
         else {
             auto* ubHandle = std::get_if<Handle<HwBufferObject>>(&mUboData);
@@ -271,6 +283,10 @@ void FMaterialInstance::commit(FEngine::DriverApi& driver) const {
     // TODO: eventually we should remove this in RELEASE builds
     fixMissingSamplers();
 
+    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
+        return;
+    }
+
     // Commit descriptors if needed (e.g. when textures are updated,or the first time)
     mDescriptorSet.commit(mMaterial->getDescriptorSetLayout(), driver);
 }
@@ -413,6 +429,13 @@ const char* FMaterialInstance::getName() const noexcept {
 // ------------------------------------------------------------------------------------------------
 
 void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
+    if (!mDescriptorSet.getHandle()) {
+        return;
+    }
+
+    if (mUseUboBatching && !BufferAllocator::isValid(getAllocationId())) {
+        return;
+    }
 
     if (UTILS_UNLIKELY(mMissingSamplerDescriptors.any())) {
         std::call_once(mMissingSamplersFlag, [this] {
@@ -439,7 +462,8 @@ void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
         return;
     }
 
-    mDescriptorSet.bind(driver, DescriptorSetBindingPoints::PER_MATERIAL);
+    mDescriptorSet.bind(driver, DescriptorSetBindingPoints::PER_MATERIAL,
+            { { mUboOffset }, driver });
 }
 
 void FMaterialInstance::assignUboAllocation(
@@ -449,8 +473,10 @@ void FMaterialInstance::assignUboAllocation(
     assert_invariant(mUseUboBatching);
 
     mUboData = id;
+    mUboOffset = offset;
     if (BufferAllocator::isValid(id)) {
-        mDescriptorSet.setBuffer(mMaterial->getDescriptorSetLayout(), 0, ubHandle, offset,
+        // Use dynamic offset during binding, so the offset here is always zero.
+        mDescriptorSet.setBuffer(mMaterial->getDescriptorSetLayout(), 0, ubHandle, 0,
                 mUniforms.getSize());
     }
 }

filament/src/details/MaterialInstance.h

@@ -56,11 +56,8 @@ class FTexture;
 
 class FMaterialInstance : public MaterialInstance {
 public:
-    FMaterialInstance(FEngine& engine, FMaterial const* material,
+    FMaterialInstance(FEngine& engine, FMaterial const* material, const char* name,
-                      const char* name) noexcept;
+            FEngine::UboBatchingMode batchingMode) noexcept;
-    // Use this constructor when you need to override the ubo batching flag for an individual MI.
-    FMaterialInstance(FEngine& engine, FMaterial const* material,
-                      const char* name, bool useUboBatching) noexcept;
     FMaterialInstance(FEngine& engine, FMaterialInstance const* other, const char* name);
     FMaterialInstance(const FMaterialInstance& rhs) = delete;
     FMaterialInstance& operator=(const FMaterialInstance& rhs) = delete;
@@ -75,7 +72,7 @@ public:
     
     void commit(FEngine& engine) const;
 
-    void commit(FEngine::DriverApi& driver) const;
+    void commit(FEngine::DriverApi& driver, UboManager* uboManager) const;
 
     void use(FEngine::DriverApi& driver, Variant variant = {}) const;
 
@@ -287,6 +284,7 @@ private:
     };
 
     std::variant<BufferAllocator::AllocationId, backend::Handle<backend::HwBufferObject>> mUboData;
+    BufferAllocator::allocation_size_t mUboOffset = 0;
     tsl::robin_map<backend::descriptor_binding_t, TextureParameter> mTextureParameters;
     mutable DescriptorSet mDescriptorSet;
     UniformBuffer mUniforms;
@@ -308,7 +306,7 @@ private:
     bool mHasScissor : 1;
     bool mIsDoubleSided : 1;
     bool mIsDefaultInstance : 1;
-    bool mUseUboBatching : 1;
+    const bool mUseUboBatching : 1;
     TransparencyMode mTransparencyMode : 2;
 
     uint64_t mMaterialSortingKey = 0;

filament/src/details/Renderer.cpp

@@ -452,6 +452,8 @@ void FRenderer::endFrame() {
     mFrameInfoManager.endFrame(driver);
     mFrameSkipper.submitFrame(driver);
 
+    engine.submitFrame();
+
     driver.endFrame(mFrameId);
 
     // gives the backend a chance to execute periodic tasks
@@ -587,6 +589,8 @@ void FRenderer::renderStandaloneView(FView const* view) {
         // happen with Renderer::beginFrame/endFrame.
         renderInternal(view, true);
 
+        engine.submitFrame();
+
         driver.endFrame(mFrameId);
 
         // engine.flush() has already been called by renderInternal(), but we need an extra one

filament/src/details/VertexBuffer.cpp

@@ -302,7 +302,7 @@ FVertexBuffer::FVertexBuffer(FEngine& engine, const Builder& builder)
     if (!mBufferObjectsEnabled) {
         // If buffer objects are not enabled at the API level, then we create them internally.
         #pragma nounroll
-        for (size_t index = 0; index < MAX_VERTEX_BUFFER_COUNT; ++index) {
+        for (size_t index = 0; index < MAX_VERTEX_ATTRIBUTE_COUNT; ++index) {
             size_t const i = mAttributes[index].buffer;
             if (i != Attribute::BUFFER_UNUSED) {
                 assert_invariant(bufferSizes[i] > 0);

filament/src/ds/PostProcessDescriptorSet.cpp

@@ -39,7 +39,7 @@ void PostProcessDescriptorSet::init(FEngine& engine) noexcept {
     // create the descriptor-set layout
     mDescriptorSetLayout = filament::DescriptorSetLayout{
             engine.getDescriptorSetLayoutFactory(),
-            engine.getDriverApi(), descriptor_sets::getPostProcessLayout() };
+            engine.getDriverApi(), descriptor_sets::getDepthVariantLayout() };
 
     // create the descriptor-set from the layout
     mDescriptorSet = DescriptorSet{ "PostProcessDescriptorSet", mDescriptorSetLayout };

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

@@ -28,9 +28,9 @@
 
 namespace filament::descriptor_sets {
 
-backend::DescriptorSetLayout const& getPostProcessLayout() noexcept;
 backend::DescriptorSetLayout const& getDepthVariantLayout() noexcept;
 backend::DescriptorSetLayout const& getSsrVariantLayout() noexcept;
+
 backend::DescriptorSetLayout const& getPerRenderableLayout() noexcept;
 
 backend::DescriptorSetLayout getPerViewDescriptorSetLayout(

libs/filabridge/src/DescriptorSets.cpp

@@ -40,12 +40,7 @@ namespace filament::descriptor_sets {
 
 using namespace backend;
 
-// used for post-processing passes
+// used to generate shadow-maps, structure and postfx passes
-static constexpr std::initializer_list<DescriptorSetLayoutBinding> postProcessDescriptorSetLayoutList = {
-    { DescriptorType::UNIFORM_BUFFER, ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT,  +PerViewBindingPoints::FRAME_UNIFORMS },
-};
-
-// used to generate shadow-maps
 static constexpr std::initializer_list<DescriptorSetLayoutBinding> depthVariantDescriptorSetLayoutList = {
     { DescriptorType::UNIFORM_BUFFER, ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT,  +PerViewBindingPoints::FRAME_UNIFORMS },
 };
@@ -142,10 +137,6 @@ static const std::unordered_map<
         {{ SamplerType::SAMPLER_EXTERNAL, SamplerFormat::FLOAT }, DescriptorType::SAMPLER_EXTERNAL }
 };
 
-// used for post-processing passes
-static DescriptorSetLayout const postProcessDescriptorSetLayout{ utils::StaticString("postProcess"),
-    postProcessDescriptorSetLayoutList };
-
 // used to generate shadow-maps
 static DescriptorSetLayout const depthVariantDescriptorSetLayout{
     utils::StaticString("depthVariant"), depthVariantDescriptorSetLayoutList
@@ -163,10 +154,6 @@ static DescriptorSetLayout perRenderableDescriptorSetLayout = {
     utils::StaticString("perRenderable"), perRenderableDescriptorSetLayoutList
 };
 
-DescriptorSetLayout const& getPostProcessLayout() noexcept {
-    return postProcessDescriptorSetLayout;
-}
-
 DescriptorSetLayout const& getDepthVariantLayout() noexcept {
     return depthVariantDescriptorSetLayout;
 }
@@ -280,10 +267,10 @@ DescriptorSetLayout getPerViewDescriptorSetLayout(
             return layout;
         }
         case MaterialDomain::POST_PROCESS:
-            return postProcessDescriptorSetLayout;
+            return depthVariantDescriptorSetLayout;
         case MaterialDomain::COMPUTE:
             // TODO: what's the layout for compute?
-            return postProcessDescriptorSetLayout;
+            return depthVariantDescriptorSetLayout;
     }
 }
 

libs/filamat/src/GLSLPostProcessor.cpp

@@ -131,7 +131,7 @@ DescriptorSetLayout getPerMaterialDescriptorSet(SamplerInterfaceBlock const& sib
 
     layout.bindings.push_back(DescriptorSetLayoutBinding{ DescriptorType::UNIFORM_BUFFER,
         ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT,
-            +PerMaterialBindingPoints::MATERIAL_PARAMS, DescriptorFlags::NONE, 0 });
+        +PerMaterialBindingPoints::MATERIAL_PARAMS, DescriptorFlags::DYNAMIC_OFFSET, 0 });
 
     for (auto const& sampler: samplers) {
         DescriptorSetLayoutBinding layoutBinding{

libs/filamat/src/eiff/MaterialInterfaceBlockChunk.cpp

@@ -225,7 +225,7 @@ void MaterialDescriptorSetLayoutChunk::flatten(Flattener& f) {
     f.writeUint8(uint8_t(DescriptorType::UNIFORM_BUFFER));
     f.writeUint8(uint8_t(ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT));
     f.writeUint8(0);
-    f.writeUint8(uint8_t(DescriptorFlags::NONE));
+    f.writeUint8(uint8_t(DescriptorFlags::DYNAMIC_OFFSET));
     f.writeUint16(0);
 
     // all the material's sampler descriptors

libs/matdbg/src/TextWriter.cpp

@@ -281,7 +281,7 @@ static bool printParametersInfo(ostream& text, const ChunkContainer& container)
         }
 
         text << "    "
-                  << setw(alignment) << fieldName.c_str()
+                  << setw(alignment) << fieldName.c_str_safe()
                   << setw(shortAlignment) << toString(UniformType(fieldType))
                   << arraySizeToString(fieldSize)
                   << setw(shortAlignment) << toString(Precision(fieldPrecision))
@@ -325,7 +325,7 @@ static bool printParametersInfo(ostream& text, const ChunkContainer& container)
         }
 
         text << "    "
-                << setw(alignment) << fieldName.c_str()
+                << setw(alignment) << fieldName.c_str_safe()
                 << setw(shortAlignment) << +fieldBinding
                 << setw(shortAlignment) << toString(SamplerType(fieldType))
                 << setw(shortAlignment) << toString(Precision(fieldPrecision))
@@ -364,7 +364,7 @@ static bool printConstantInfo(ostream& text, const ChunkContainer& container) {
         }
 
          text << "    "
-         << setw(alignment) << fieldName.c_str()
+         << setw(alignment) << fieldName.c_str_safe()
          << setw(shortAlignment) << toString(ConstantType(fieldType))
          << endl;
     }