vk: testing a robin_map key hash failure

Executable on device with:

./build.sh -p android -q arm64-v8a release test_compiler && adb push out/cmake-android-release-aarch64/filament/test/test_compiler /data/local/tmp/ && adb shell chmod 777 /data/local/tmp/test_compiler && adb shell /data/local/tmp/test_compiler

RELEASE_NOTES.md

@@ -7,6 +7,10 @@ A new header is inserted each time a *tag* is created.
 Instead, if you are authoring a PR for the main branch, add your release note to
 [NEW_RELEASE_NOTES.md](./NEW_RELEASE_NOTES.md).
 
+## v1.53.3
+
+- Add drag and drop support for IBL files for desktop gltf_viewer.
+
 ## v1.53.2
 
 

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

@@ -28,6 +28,14 @@
 using namespace filament;
 using namespace backend;
 
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Renderer_nSkipFrame(JNIEnv *, jclass, jlong nativeRenderer,
+        jlong vsyncSteadyClockTimeNano) {
+    Renderer *renderer = (Renderer *) nativeRenderer;
+    renderer->skipFrame(uint64_t(vsyncSteadyClockTimeNano));
+}
+
 extern "C" JNIEXPORT jboolean JNICALL
 Java_com_google_android_filament_Renderer_nBeginFrame(JNIEnv *, jclass, jlong nativeRenderer,
         jlong nativeSwapChain, jlong frameTimeNanos) {

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

@@ -531,3 +531,12 @@ Java_com_google_android_filament_View_nGetFogEntity(JNIEnv *env, jclass clazz,
     View *view = (View *) nativeView;
     return (jint)view->getFogEntity().getId();
 }
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_View_nClearFrameHistory(JNIEnv *env, jclass clazz,
+        jlong nativeView, jlong nativeEngine) {
+    View *view = (View *) nativeView;
+    Engine *engine = (Engine *) nativeEngine;
+    view->clearFrameHistory(*engine);
+}

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

@@ -425,9 +425,12 @@ public class Engine {
         public long resourceAllocatorCacheSizeMB = 64;
 
         /*
-         * This value determines for how many frames are texture entries kept in the cache.
+         * This value determines how many frames texture entries are kept for in the cache. This
+         * is a soft limit, meaning some texture older than this are allowed to stay in the cache.
+         * Typically only one texture is evicted per frame.
+         * The default is 1.
          */
-        public long resourceAllocatorCacheMaxAge = 2;
+        public long resourceAllocatorCacheMaxAge = 1;
 
         /*
          * Disable backend handles use-after-free checks.

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

@@ -297,6 +297,20 @@ public class Renderer {
         nSetVsyncTime(getNativeObject(), steadyClockTimeNano);
     }
 
+    /**
+     * Call skipFrame when momentarily skipping frames, for instance if the content of the
+     * scene doesn't change.
+     *
+     * @param vsyncSteadyClockTimeNano The time in nanoseconds when the frame started being rendered,
+     *                       in the {@link System#nanoTime()} timebase. Divide this value by 1000000 to
+     *                       convert it to the {@link android.os.SystemClock#uptimeMillis()}
+     *                       time base. This typically comes from
+     *                       {@link android.view.Choreographer.FrameCallback}.
+     */
+    public void skipFrame(long vsyncSteadyClockTimeNano) {
+        nSkipFrame(getNativeObject(), vsyncSteadyClockTimeNano);
+    }
+
     /**
      * Sets up a frame for this <code>Renderer</code>.
      * <p><code>beginFrame</code> manages frame pacing, and returns whether or not a frame should be
@@ -716,6 +730,7 @@ public class Renderer {
 
     private static native void nSetPresentationTime(long nativeObject, long monotonicClockNanos);
     private static native void nSetVsyncTime(long nativeObject, long steadyClockTimeNano);
+    private static native void nSkipFrame(long nativeObject, long vsyncSteadyClockTimeNano);
     private static native boolean nBeginFrame(long nativeRenderer, long nativeSwapChain, long frameTimeNanos);
     private static native void nEndFrame(long nativeRenderer);
     private static native void nRender(long nativeRenderer, long nativeView);

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

@@ -1233,6 +1233,18 @@ public class View {
         return nGetFogEntity(getNativeObject());
     }
 
+    /**
+     * When certain temporal features are used (e.g.: TAA or Screen-space reflections), the view
+     * keeps a history of previous frame renders associated with the Renderer the view was last
+     * used with. When switching Renderer, it may be necessary to clear that history by calling
+     * this method. Similarly, if the whole content of the screen change, like when a cut-scene
+     * starts, clearing the history might be needed to avoid artifacts due to the previous frame
+     * being very different.
+     */
+    public void clearFrameHistory(Engine engine) {
+        nClearFrameHistory(getNativeObject(), engine.getNativeObject());
+    }
+
     public long getNativeObject() {
         if (mNativeObject == 0) {
             throw new IllegalStateException("Calling method on destroyed View");
@@ -1294,7 +1306,7 @@ public class View {
     private static native void nSetMaterialGlobal(long nativeView, int index, float x, float y, float z, float w);
     private static native void nGetMaterialGlobal(long nativeView, int index, float[] out);
     private static native int nGetFogEntity(long nativeView);
-
+    private static native void nClearFrameHistory(long nativeView, long nativeEngine);
 
     /**
      * List of available ambient occlusion techniques.

android/filament-utils-android/src/main/cpp/Manipulator.cpp

@@ -125,6 +125,11 @@ extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBu
     builder->groundPlane(a, b, c, d);
 }
 
+extern "C" JNIEXPORT void Java_com_google_android_filament_utils_Manipulator_nBuilderPanning(JNIEnv*, jclass, jlong nativeBuilder, jboolean enabled) {
+    Builder* builder = (Builder*) nativeBuilder;
+    builder->panning(enabled);
+}
+
 extern "C" JNIEXPORT long Java_com_google_android_filament_utils_Manipulator_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder, jint mode) {
     Builder* builder = (Builder*) nativeBuilder;
     return (jlong) builder->build((Mode) mode);

android/filament-utils-android/src/main/java/com/google/android/filament/utils/Manipulator.java

@@ -274,6 +274,17 @@ public class Manipulator {
             return this;
         }
 
+        /**
+         * Sets whether panning is enabled in the manipulator.
+         *
+         * @return this <code>Builder</code> object for chaining calls
+         */
+        @NonNull
+        public Builder panning(Boolean enabled) {
+            nBuilderPanning(mNativeBuilder, enabled);
+            return this;
+        }
+
         /**
          * Creates and returns the <code>Manipulator</code> object.
          *
@@ -483,6 +494,7 @@ public class Manipulator {
     private static native void nBuilderFlightPanSpeed(long nativeBuilder, float x, float y);
     private static native void nBuilderFlightMoveDamping(long nativeBuilder, float damping);
     private static native void nBuilderGroundPlane(long nativeBuilder, float a, float b, float c, float d);
+    private static native void nBuilderPanning(long nativeBuilder, Boolean enabled);
     private static native long nBuilderBuild(long nativeBuilder, int mode);
 
     private static native void nDestroyManipulator(long nativeManip);

filament/backend/src/CircularBuffer.cpp

@@ -32,10 +32,11 @@
 #    define HAS_MMAP 0
 #endif
 
-#include <stdint.h>
 #include <stddef.h>
-#include <stdlib.h>
+#include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 
 using namespace utils;
 
@@ -81,6 +82,9 @@ void* CircularBuffer::alloc(size_t size) noexcept {
             // map the circular buffer once...
             vaddr = mmap(reserve_vaddr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
             if (vaddr != MAP_FAILED) {
+                // populate the address space with pages (because this is a circular buffer,
+                // all the pages will be allocated eventually, might as well do it now)
+                memset(vaddr, 0, size);
                 // and map the circular buffer again, behind the previous copy...
                 vaddr_shadow = mmap((char*)vaddr + size, size,
                         PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
@@ -101,7 +105,7 @@ void* CircularBuffer::alloc(size_t size) noexcept {
     if (UTILS_UNLIKELY(mAshmemFd < 0)) {
         // ashmem failed
         if (vaddr_guard != MAP_FAILED) {
-            munmap(vaddr_guard, size);
+            munmap(vaddr_guard, BLOCK_SIZE);
         }
 
         if (vaddr_shadow != MAP_FAILED) {
@@ -119,12 +123,11 @@ void* CircularBuffer::alloc(size_t size) noexcept {
         data = mmap(nullptr, size * 2 + BLOCK_SIZE,
                 PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 
-        FILAMENT_CHECK_POSTCONDITION(data) <<
+        FILAMENT_CHECK_POSTCONDITION(data != MAP_FAILED) <<
                 "couldn't allocate " << (size * 2 / 1024) <<
                 " KiB of virtual address space for the command buffer";
 
-        slog.d << "WARNING: Using soft CircularBuffer (" << (size * 2 / 1024) << " KiB)"
-               << io::endl;
+        slog.w << "Using 'soft' CircularBuffer (" << (size * 2 / 1024) << " KiB)" << io::endl;
 
         // guard page at the end
         void* guard = (void*)(uintptr_t(data) + size * 2);

filament/backend/src/metal/MetalHandles.mm

@@ -556,8 +556,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_CUBEMAP:
         case SamplerType::SAMPLER_CUBEMAP_ARRAY:
@@ -572,8 +570,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_3D:
             descriptor = [MTLTextureDescriptor new];
@@ -586,8 +582,6 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             descriptor.usage = getMetalTextureUsage(usage);
             descriptor.storageMode = MTLStorageModePrivate;
             texture = [context.device newTextureWithDescriptor:descriptor];
-            FILAMENT_CHECK_POSTCONDITION(texture != nil)
-                    << "Could not create Metal texture. Out of memory?";
             break;
         case SamplerType::SAMPLER_EXTERNAL:
             // If we're using external textures (CVPixelBufferRefs), we don't need to make any
@@ -596,6 +590,12 @@ MetalTexture::MetalTexture(MetalContext& context, SamplerType target, uint8_t le
             break;
     }
 
+    FILAMENT_CHECK_POSTCONDITION(target == SamplerType::SAMPLER_EXTERNAL || texture != nil)
+            << "Could not create Metal texture (SamplerType = " << int(target)
+            << ", levels = " << int(levels) << ", MTLPixelFormat = " << int(devicePixelFormat)
+            << ", width = " << width << ", height = " << height << ", depth = " << depth
+            << "). Out of memory?";
+
     // If swizzling is set, set up a swizzled texture view that we'll use when sampling this texture.
     const bool isDefaultSwizzle =
             r == TextureSwizzle::CHANNEL_0 &&

filament/backend/src/vulkan/VulkanBuffer.cpp

@@ -28,6 +28,7 @@ VulkanBuffer::VulkanBuffer(VmaAllocator allocator, VulkanStagePool& stagePool,
     : mAllocator(allocator),
       mStagePool(stagePool),
       mUsage(usage),
+	  mUpdatedOffset(0),
       mUpdatedBytes(0) {
     // for now make sure that only 1 bit is set in usage
     // (because loadFromCpu() assumes that somewhat)
@@ -80,6 +81,7 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
             .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .buffer = mGpuBuffer,
+			.offset = mUpdatedOffset,
             .size = mUpdatedBytes,
         };
         vkCmdPipelineBarrier(cmdbuf, srcStage, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 1,
@@ -93,6 +95,7 @@ void VulkanBuffer::loadFromCpu(VkCommandBuffer cmdbuf, const void* cpuData, uint
     };
     vkCmdCopyBuffer(cmdbuf, stage->buffer, mGpuBuffer, 1, &region);
 
+	mUpdatedOffset = byteOffset;
     mUpdatedBytes = numBytes;
 
     // Firstly, ensure that the copy finishes before the next draw call.

filament/backend/src/vulkan/VulkanBuffer.h

@@ -42,6 +42,7 @@ private:
     VmaAllocation mGpuMemory = VK_NULL_HANDLE;
     VkBuffer mGpuBuffer = VK_NULL_HANDLE;
     VkBufferUsageFlags mUsage = {};
+	uint32_t mUpdatedOffset = 0;
     uint32_t mUpdatedBytes = 0;
 };
 

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -986,15 +986,15 @@ FeatureLevel VulkanDriver::getFeatureLevel() {
 
     // If the max sampler counts do not meet FL2 standards, then this is an FL1 device.
     const auto& fl2 = FEATURE_LEVEL_CAPS[+FeatureLevel::FEATURE_LEVEL_2];
-    if (fl2.MAX_VERTEX_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers ||
-        fl2.MAX_FRAGMENT_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers) {
+    if (limits.maxPerStageDescriptorSamplers < fl2.MAX_VERTEX_SAMPLER_COUNT  ||
+        limits.maxPerStageDescriptorSamplers < fl2.MAX_FRAGMENT_SAMPLER_COUNT) {
         return FeatureLevel::FEATURE_LEVEL_1;
     }
 
     // If the max sampler counts do not meet FL3 standards, then this is an FL2 device.
     const auto& fl3 = FEATURE_LEVEL_CAPS[+FeatureLevel::FEATURE_LEVEL_3];
-    if (fl3.MAX_VERTEX_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers ||
-        fl3.MAX_FRAGMENT_SAMPLER_COUNT < limits.maxPerStageDescriptorSamplers) {
+    if (limits.maxPerStageDescriptorSamplers < fl3.MAX_VERTEX_SAMPLER_COUNT||
+        limits.maxPerStageDescriptorSamplers < fl3.MAX_FRAGMENT_SAMPLER_COUNT) {
         return FeatureLevel::FEATURE_LEVEL_2;
     }
 

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

@@ -33,11 +33,11 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
 
     // build the push constant layout key
     uint32_t pushConstantRangeCount = program->getPushConstantRangeCount();
-    auto const& pushCostantRanges = program->getPushConstantRanges();    
+    auto const& pushConstantRanges = program->getPushConstantRanges();    
     if (pushConstantRangeCount > 0) {
         assert_invariant(pushConstantRangeCount <= Program::SHADER_TYPE_COUNT);
         for (uint8_t i = 0; i < pushConstantRangeCount; ++i) {
-            auto const& range = pushCostantRanges[i];
+            auto const& range = pushConstantRanges[i];
             auto& pushConstant = key.pushConstant[i];
             if (range.stageFlags &  VK_SHADER_STAGE_VERTEX_BIT) {
                 pushConstant.stage = static_cast<uint8_t>(ShaderStage::VERTEX);
@@ -52,9 +52,8 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
         }
     }
 
-    auto iter = mPipelineLayouts.find(key);
-    if (iter != mPipelineLayouts.end()) {
-        PipelineLayoutCacheEntry& entry = mPipelineLayouts[key];
+    if (PipelineLayoutMap::iterator iter = mPipelineLayouts.find(key); iter != mPipelineLayouts.end()) {
+        PipelineLayoutCacheEntry& entry = iter.value();
         entry.lastUsed = mTimestamp++;
         return entry.handle;
     }
@@ -65,7 +64,7 @@ VkPipelineLayout VulkanPipelineLayoutCache::getLayout(
         .setLayoutCount = (uint32_t) descSetLayoutCount,
         .pSetLayouts = key.descSetLayouts.data(),
         .pushConstantRangeCount = pushConstantRangeCount,
-        .pPushConstantRanges = pushCostantRanges,
+        .pPushConstantRanges = pushConstantRanges,
     };
 
     VkPipelineLayout layout;

filament/docs/SpirvDebugging.md

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

filament/include/filament/Engine.h

@@ -350,9 +350,12 @@ public:
         uint32_t resourceAllocatorCacheSizeMB = 64;
 
         /*
-         * This value determines for how many frames are texture entries kept in the cache.
+         * This value determines how many frames texture entries are kept for in the cache. This
+         * is a soft limit, meaning some texture older than this are allowed to stay in the cache.
+         * Typically only one texture is evicted per frame.
+         * The default is 1.
          */
-        uint32_t resourceAllocatorCacheMaxAge = 2;
+        uint32_t resourceAllocatorCacheMaxAge = 1;
 
         /*
          * Disable backend handles use-after-free checks.

filament/include/filament/Renderer.h

@@ -22,9 +22,11 @@
 #include <filament/FilamentAPI.h>
 
 #include <utils/compiler.h>
+#include <utils/FixedCapacityVector.h>
 
 #include <math/vec4.h>
 
+#include <stddef.h>
 #include <stdint.h>
 
 namespace filament {
@@ -81,6 +83,37 @@ public:
         UTILS_DEPRECATED uint64_t vsyncOffsetNanos = 0;
     };
 
+    /**
+     * Timing information about a frame
+     * @see getFrameInfoHistory()
+     */
+    struct FrameInfo {
+        using time_point_ns = int64_t;
+        using duration_ns = int64_t;
+        uint32_t frameId;                   //!< monotonically increasing frame identifier
+        duration_ns frameTime;              //!< frame duration on the GPU in nanosecond [ns]
+        duration_ns denoisedFrameTime;      //!< denoised frame duration on the GPU in [ns]
+        time_point_ns beginFrame;           //!< Renderer::beginFrame() time since epoch [ns]
+        time_point_ns endFrame;             //!< Renderer::endFrame() time since epoch [ns]
+        time_point_ns backendBeginFrame;    //!< Backend thread time of frame start since epoch [ns]
+        time_point_ns backendEndFrame;      //!< Backend thread time of frame end since epoch [ns]
+    };
+
+    /**
+     * Retrieve an historic of frame timing information. The maximum frame history size is
+     * given by getMaxFrameHistorySize().
+     * @param historySize requested history size. The returned vector could be smaller.
+     * @return A vector of FrameInfo.
+     */
+    utils::FixedCapacityVector<Renderer::FrameInfo> getFrameInfoHistory(
+            size_t historySize = 1) const noexcept;
+
+    /**
+     * @return the maximum supported frame history size.
+     * @see getFrameInfoHistory()
+     */
+    size_t getMaxFrameHistorySize() const noexcept;
+
     /**
      * Use FrameRateOptions to set the desired frame rate and control how quickly the system
      * reacts to GPU load changes.
@@ -238,6 +271,14 @@ public:
      */
     void setVsyncTime(uint64_t steadyClockTimeNano) noexcept;
 
+    /**
+     * Call skipFrame when momentarily skipping frames, for instance if the content of the
+     * scene doesn't change.
+     *
+     * @param vsyncSteadyClockTimeNano
+     */
+    void skipFrame(uint64_t vsyncSteadyClockTimeNano = 0u);
+
     /**
      * Set-up a frame for this Renderer.
      *

filament/include/filament/View.h

@@ -40,6 +40,7 @@ class CallbackHandler;
 
 class Camera;
 class ColorGrading;
+class Engine;
 class MaterialInstance;
 class RenderTarget;
 class Scene;
@@ -878,6 +879,17 @@ public:
      */
     utils::Entity getFogEntity() const noexcept;
 
+
+    /**
+     * When certain temporal features are used (e.g.: TAA or Screen-space reflections), the view
+     * keeps a history of previous frame renders associated with the Renderer the view was last
+     * used with. When switching Renderer, it may be necessary to clear that history by calling
+     * this method. Similarly, if the whole content of the screen change, like when a cut-scene
+     * starts, clearing the history might be needed to avoid artifacts due to the previous frame
+     * being very different.
+     */
+    void clearFrameHistory(Engine& engine) noexcept;
+
     /**
      * List of available ambient occlusion techniques
      * @deprecated use AmbientOcclusionOptions::enabled instead

filament/src/FrameHistory.h

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

filament/src/FrameInfo.cpp

@@ -16,30 +16,33 @@
 
 #include "FrameInfo.h"
 
+#include <filament/Renderer.h>
+
+#include <backend/DriverEnums.h>
+
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
 #include <utils/Log.h>
-#include <utils/Systrace.h>
+#include <utils/ostream.h>
 
-#include <math/scalar.h>
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <memory>
+#include <ratio>
 
-#include <cmath>
+#include <stdint.h>
+#include <stddef.h>
 
 namespace filament {
 
 using namespace utils;
 using namespace backend;
 
-// this is to avoid a call to memmove
-template<class InputIterator, class OutputIterator>
-static inline
-void move_backward(InputIterator first, InputIterator last, OutputIterator result) {
-    while (first != last) {
-        *--result = *--last;
-    }
-}
-
 FrameInfoManager::FrameInfoManager(DriverApi& driver) noexcept {
     for (auto& query : mQueries) {
-        query = driver.createTimerQuery();
+        query.handle = driver.createTimerQuery();
     }
 }
 
@@ -47,14 +50,36 @@ FrameInfoManager::~FrameInfoManager() noexcept = default;
 
 void FrameInfoManager::terminate(DriverApi& driver) noexcept {
     for (auto& query : mQueries) {
-        driver.destroyTimerQuery(query);
+        driver.destroyTimerQuery(query.handle);
     }
 }
 
-void FrameInfoManager::beginFrame(DriverApi& driver,Config const& config, uint32_t) noexcept {
-    driver.beginTimerQuery(mQueries[mIndex]);
+void FrameInfoManager::beginFrame(DriverApi& driver, Config const& config, uint32_t frameId) noexcept {
+    auto& history = mFrameTimeHistory;
+    // don't exceed the capacity, drop the oldest entry
+    if (UTILS_LIKELY(history.size() == history.capacity())) {
+        history.pop_back();
+    }
+
+    // create a new entry
+    auto& front = history.emplace_front(frameId);
+
+    // store the current time
+    front.beginFrame = std::chrono::steady_clock::now();
+
+    // references are not invalidated by CircularQueue<>, so we can associate a reference to
+    // the slot we created to the timer query used to find the frame time.
+    mQueries[mIndex].pInfo = std::addressof(front);
+    // issue the timer query
+    driver.beginTimerQuery(mQueries[mIndex].handle);
+    // issue the custom backend command to get the backend time
+    driver.queueCommand([&front](){
+        front.backendBeginFrame = std::chrono::steady_clock::now();
+    });
+
+    // now is a good time to check the oldest active query
     uint64_t elapsed = 0;
-    TimerQueryResult const result = driver.getTimerQueryValue(mQueries[mLast], &elapsed);
+    TimerQueryResult const result = driver.getTimerQueryValue(mQueries[mLast].handle, &elapsed);
     switch (result) {
         case TimerQueryResult::NOT_READY:
             // nothing to do
@@ -63,49 +88,111 @@ void FrameInfoManager::beginFrame(DriverApi& driver,Config const& config, uint32
             mLast = (mLast + 1) % POOL_COUNT;
             break;
         case TimerQueryResult::AVAILABLE:
-            mLast = (mLast + 1) % POOL_COUNT;
             // conversion to our duration happens here
-            mFrameTime = std::chrono::duration<uint64_t, std::nano>(elapsed);
+            pFront = mQueries[mLast].pInfo;
+            pFront->frameTime = std::chrono::duration<uint64_t, std::nano>(elapsed);
+            mLast = (mLast + 1) % POOL_COUNT;
+            denoiseFrameTime(config);
             break;
     }
-    update(config, mFrameTime);
+
+
+    // keep this just for debugging
+    if constexpr (false) {
+        using namespace utils;
+        auto h = getFrameInfoHistory(1);
+        if (!h.empty()) {
+            slog.d << frameId << ": "
+                   << h[0].frameId << " (" << frameId - h[0].frameId << ")"
+                   << ", Dm=" << h[0].endFrame - h[0].beginFrame
+                   << ", L =" << h[0].backendBeginFrame - h[0].beginFrame
+                   << ", Db=" << h[0].backendEndFrame - h[0].backendBeginFrame
+                   << ", T =" << h[0].frameTime
+                   << io::endl;
+        }
+    }
 }
 
 void FrameInfoManager::endFrame(DriverApi& driver) noexcept {
-    driver.endTimerQuery(mQueries[mIndex]);
+    auto& front = mFrameTimeHistory.front();
+    // close the timer query
+    driver.endTimerQuery(mQueries[mIndex].handle);
+    // queue custom backend command to query the current time
+    driver.queueCommand([&front](){
+        // backend frame end-time
+        front.backendEndFrame = std::chrono::steady_clock::now();
+        // signal that the data is available
+        front.ready.store(true, std::memory_order_release);
+    });
+    // and finally acquire the time on the main thread
+    front.endFrame = std::chrono::steady_clock::now();
     mIndex = (mIndex + 1) % POOL_COUNT;
 }
 
-void FrameInfoManager::update(Config const& config,
-        FrameInfoManager::duration lastFrameTime) noexcept {
-    // keep an history of frame times
+void FrameInfoManager::denoiseFrameTime(Config const& config) noexcept {
     auto& history = mFrameTimeHistory;
+    assert_invariant(!history.empty());
 
-    // this is like doing { pop_back(); push_front(); }
-    filament::move_backward(history.begin(), history.end() - 1, history.end());
-    history[0].frameTime = lastFrameTime;
-
-    mFrameTimeHistorySize = std::min(++mFrameTimeHistorySize, uint32_t(MAX_FRAMETIME_HISTORY));
-    if (UTILS_UNLIKELY(mFrameTimeHistorySize < 3)) {
-        // not enough history to do anything useful
-        history[0].valid = false;
-        return;
+    // find the first slot that has a valid frame duration
+    size_t first = history.size();
+    for (size_t i = 0, c = history.size(); i < c; ++i) {
+        if (history[i].frameTime != duration(0)) {
+            assert_invariant(std::addressof(history[i]) == pFront);
+            first = i;
+            break;
+        }
     }
+    assert_invariant(first != history.size());
 
-    // apply a median filter to get a good representation of the frame time of the last
-    // N frames.
-    std::array<duration, MAX_FRAMETIME_HISTORY> median; // NOLINT -- it's initialized below
-    size_t const size = std::min(mFrameTimeHistorySize,
-            std::min(config.historySize, (uint32_t)median.size()));
-    for (size_t i = 0; i < size; ++i) {
-        median[i] = history[i].frameTime;
-    }
-    std::sort(median.begin(), median.begin() + size);
-    duration const denoisedFrameTime = median[size / 2];
+    // we need at least 3 valid frame time to calculate the median
+    if (history.size() >= first + 3) {
+        // apply a median filter to get a good representation of the frame time of the last
+        // N frames.
+        std::array<duration, MAX_FRAMETIME_HISTORY> median; // NOLINT -- it's initialized below
+        size_t const size = std::min({
+            history.size() - first,
+            median.size(),
+            size_t(config.historySize) });
 
-    history[0].denoisedFrameTime = denoisedFrameTime;
-    history[0].valid = true;
+        for (size_t i = 0; i < size; ++i) {
+            median[i] = history[first + i].frameTime;
+        }
+        std::sort(median.begin(), median.begin() + size);
+        duration const denoisedFrameTime = median[size / 2];
+
+        history[first].denoisedFrameTime = denoisedFrameTime;
+        history[first].valid = true;
+     }
 }
 
+utils::FixedCapacityVector<Renderer::FrameInfo> FrameInfoManager::getFrameInfoHistory(
+        size_t historySize) const noexcept {
+    auto result = utils::FixedCapacityVector<Renderer::FrameInfo>::with_capacity(MAX_FRAMETIME_HISTORY);
+    auto const& history = mFrameTimeHistory;
+    size_t i = 0;
+    size_t const c = history.size();
+    for (; i < c; ++i) {
+        auto const& entry = history[i];
+        if (entry.ready.load(std::memory_order_acquire) && entry.valid) {
+            // once we found an entry ready,
+            // we know by construction that all following ones are too
+            break;
+        }
+    }
+    for (; i < c && historySize; ++i, --historySize) {
+        auto const& entry = history[i];
+        using namespace std::chrono;
+        result.push_back({
+                entry.frameId,
+                duration_cast<nanoseconds>(entry.frameTime).count(),
+                duration_cast<nanoseconds>(entry.denoisedFrameTime).count(),
+                duration_cast<nanoseconds>(entry.beginFrame.time_since_epoch()).count(),
+                duration_cast<nanoseconds>(entry.endFrame.time_since_epoch()).count(),
+                duration_cast<nanoseconds>(entry.backendBeginFrame.time_since_epoch()).count(),
+                duration_cast<nanoseconds>(entry.backendEndFrame.time_since_epoch()).count()
+        });
+    }
+    return result;
+}
 
 } // namespace filament

filament/src/FrameInfo.h

@@ -17,30 +17,136 @@
 #ifndef TNT_FILAMENT_FRAMEINFO_H
 #define TNT_FILAMENT_FRAMEINFO_H
 
-#include "backend/Handle.h"
+#include <filament/Renderer.h>
+
+#include <backend/Handle.h>
+
 #include <private/backend/DriverApi.h>
 
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/FixedCapacityVector.h>
+
 #include <array>
+#include <atomic>
 #include <chrono>
+#include <ratio>
+#include <type_traits>
 
 #include <stdint.h>
+#include <stddef.h>
 
 namespace filament {
 class FEngine;
 
+namespace details {
 struct FrameInfo {
     using duration = std::chrono::duration<float, std::milli>;
     duration frameTime{};            // frame period
     duration denoisedFrameTime{};    // frame period (median filter)
-    bool valid = false;
+    bool valid = false;              // true if the data of the structure is valid
+};
+} // namespace details
+
+struct FrameInfoImpl : public details::FrameInfo {
+    using clock = std::chrono::steady_clock;
+    using time_point = clock::time_point;
+    uint32_t const frameId;
+    time_point beginFrame;           // main thread beginFrame time
+    time_point endFrame;             // main thread endFrame time
+    time_point backendBeginFrame;    // backend thread beginFrame time (makeCurrent time)
+    time_point backendEndFrame;      // backend thread endFrame time (present time)
+    std::atomic_bool ready{};        // true once backend thread has populated its data
+    explicit FrameInfoImpl(uint32_t frameId) noexcept
+        : frameId(frameId) {
+    }
+};
+
+template<typename T, size_t CAPACITY>
+class CircularQueue {
+public:
+    using value_type = T;
+    using reference = value_type&;
+    using const_reference = value_type const&;
+
+    size_t capacity() const {
+        return CAPACITY;
+    }
+
+    size_t size() const {
+        return mSize;
+    }
+
+    bool empty() const noexcept {
+        return !size();
+    }
+
+    void pop_back() noexcept {
+        assert_invariant(!empty());
+        --mSize;
+        std::destroy_at(&mStorage[(mFront - mSize) % CAPACITY]);
+    }
+
+    void push_front(T const& v) noexcept {
+        assert_invariant(size() < CAPACITY);
+        mFront = advance(mFront);
+        new(&mStorage[mFront]) T(v);
+        ++mSize;
+    }
+
+    void push_front(T&& v) noexcept {
+        assert_invariant(size() < CAPACITY);
+        mFront = advance(mFront);
+        new(&mStorage[mFront]) T(std::move(v));
+        ++mSize;
+    }
+
+    template<typename ...Args>
+    T& emplace_front(Args&&... args) noexcept {
+        assert_invariant(size() < CAPACITY);
+        mFront = advance(mFront);
+        new(&mStorage[mFront]) T(std::forward<Args>(args)...);
+        ++mSize;
+        return front();
+    }
+
+    T& operator[](size_t pos) noexcept {
+        assert_invariant(pos < size());
+        size_t const index = (mFront + CAPACITY - pos) % CAPACITY;
+        return *std::launder(reinterpret_cast<T*>(&mStorage[index]));
+    }
+
+    T const& operator[](size_t pos) const noexcept {
+        return const_cast<CircularQueue&>(*this)[pos];
+    }
+
+    T const& front() const noexcept {
+        assert_invariant(!empty());
+        return operator[](0);
+    }
+
+    T& front() noexcept {
+        assert_invariant(!empty());
+        return operator[](0);
+    }
+
+private:
+    using Storage = std::aligned_storage_t<sizeof(T), alignof(T)>;
+    Storage mStorage[CAPACITY];
+    uint32_t mFront = 0;    // always index 0
+    uint32_t mSize = 0;
+    [[nodiscard]] inline uint32_t advance(uint32_t v) noexcept {
+        return (v + 1) % CAPACITY;
+    }
 };
 
 class FrameInfoManager {
     static constexpr size_t POOL_COUNT = 4;
-    static constexpr size_t MAX_FRAMETIME_HISTORY = 31u;
+    static constexpr size_t MAX_FRAMETIME_HISTORY = 16u;
 
 public:
-    using duration = FrameInfo::duration;
+    using duration = FrameInfoImpl::duration;
+    using clock = FrameInfoImpl::clock;
 
     struct Config {
         uint32_t historySize;
@@ -57,23 +163,24 @@ public:
     // call this immediately before "swap buffers"
     void endFrame(backend::DriverApi& driver) noexcept;
 
-    FrameInfo const& getLastFrameInfo() const noexcept {
-        return mFrameTimeHistory[0];
+    details::FrameInfo const& getLastFrameInfo() const noexcept {
+        // if pFront is not set yet, return front() but in this case front().valid will be false
+        return pFront ? *pFront : mFrameTimeHistory.front();
     }
 
-    duration getLastFrameTime() const noexcept {
-        return getLastFrameInfo().frameTime;
-    }
+    utils::FixedCapacityVector<Renderer::FrameInfo> getFrameInfoHistory(size_t historySize) const noexcept;
 
 private:
-    void update(Config const& config, duration lastFrameTime) noexcept;
-    backend::Handle<backend::HwTimerQuery> mQueries[POOL_COUNT];
-    duration mFrameTime{};
-    uint32_t mIndex = 0;
-    uint32_t mLast = 0;
-
-    std::array<FrameInfo, MAX_FRAMETIME_HISTORY> mFrameTimeHistory;
-    uint32_t mFrameTimeHistorySize = 0;
+    void denoiseFrameTime(Config const& config) noexcept;
+    struct Query {
+        backend::Handle<backend::HwTimerQuery> handle{};
+        FrameInfoImpl* pInfo = nullptr;
+    };
+    std::array<Query, POOL_COUNT> mQueries;
+    uint32_t mIndex = 0;                // index of current query
+    uint32_t mLast = 0;                 // index of oldest query still active
+    FrameInfoImpl* pFront = nullptr;    // the most recent slot with a valid frame time
+    CircularQueue<FrameInfoImpl, MAX_FRAMETIME_HISTORY> mFrameTimeHistory;
 };
 
 

filament/src/Froxelizer.cpp

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

filament/src/PostProcessManager.cpp

@@ -1965,8 +1965,8 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
         // - visible bloom size changes with dynamic resolution in non-homogenous mode
         // This allows us to use the 9 sample downsampling filter (instead of 13)
         // for at least 4 levels.
-        uint32_t width  = std::max(1u, uint32_t(std::floor(bloomWidth)));
-        uint32_t height = std::max(1u, uint32_t(std::floor(bloomHeight)));
+        uint32_t width  = std::max(16u, uint32_t(std::floor(bloomWidth)));
+        uint32_t height = std::max(16u, uint32_t(std::floor(bloomHeight)));
         width  &= ~((1 << 4) - 1);  // at least 4 levels
         height &= ~((1 << 4) - 1);
         bloomWidth  = float(width);
@@ -1978,6 +1978,8 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
     // we don't need to do the fireflies reduction if we have TAA (it already does it)
     bool fireflies = threshold && !taaOptions.enabled;
 
+    assert_invariant(bloomWidth && bloomHeight);
+
     while (2 * bloomWidth < float(desc.width) || 2 * bloomHeight < float(desc.height)) {
         if (inoutBloomOptions.quality == QualityLevel::LOW ||
             inoutBloomOptions.quality == QualityLevel::MEDIUM) {

filament/src/RenderPass.cpp

@@ -38,6 +38,7 @@
 #include <backend/PipelineState.h>
 
 #include "private/backend/CircularBuffer.h"
+#include "private/backend/CommandStream.h"
 
 #include <utils/compiler.h>
 #include <utils/debug.h>
@@ -153,7 +154,7 @@ RenderPass::RenderPass(FEngine& engine, RenderPassBuilder const& builder) noexce
     sortCommands(builder.mArena);
 
     if (engine.isAutomaticInstancingEnabled()) {
-        uint32_t stereoscopicEyeCount = 1;
+        int32_t stereoscopicEyeCount = 1;
         if (builder.mFlags & IS_INSTANCED_STEREOSCOPIC) {
             stereoscopicEyeCount *= engine.getConfig().stereoscopicEyeCount;
         }
@@ -162,8 +163,7 @@ RenderPass::RenderPass(FEngine& engine, RenderPassBuilder const& builder) noexce
 }
 
 // this destructor is actually heavy because it inlines ~vector<>
-RenderPass::~RenderPass() noexcept {
-}
+RenderPass::~RenderPass() noexcept = default;
 
 void RenderPass::resize(Arena& arena, size_t count) noexcept {
     if (mCommandBegin) {
@@ -221,7 +221,7 @@ void RenderPass::appendCommands(FEngine& engine,
         work(vr.first, vr.size());
     } else {
         auto* jobCommandsParallel = jobs::parallel_for(js, nullptr, vr.first, (uint32_t)vr.size(),
-                std::cref(work), jobs::CountSplitter<JOBS_PARALLEL_FOR_COMMANDS_COUNT, 5>());
+                std::cref(work), jobs::CountSplitter<JOBS_PARALLEL_FOR_COMMANDS_COUNT>());
         js.runAndWait(jobCommandsParallel);
     }
 
@@ -885,14 +885,31 @@ void RenderPass::Executor::execute(FEngine& engine,
 
         // Maximum space occupied in the CircularBuffer by a single `Command`. This must be
         // reevaluated when the inner loop below adds DriverApi commands or when we change the
-        // CommandStream protocol. Currently, the maximum is 240 bytes, and we use 256 to be on
-        // the safer side.
-        size_t const maxCommandSizeInBytes = 256;
+        // CommandStream protocol. Currently, the maximum is 320 bytes.
+        // The batch size is calculated by adding the size of all commands that can possibly be
+        // emitted per draw call:
+        constexpr size_t const maxCommandSizeInBytes =
+                sizeof(CustomCommand) +
+                sizeof(COMMAND_TYPE(scissor)) +
+                sizeof(COMMAND_TYPE(bindUniformBuffer)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindBufferRange)) +
+                sizeof(COMMAND_TYPE(bindBufferRange)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindUniformBuffer)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindSamplers)) +
+                sizeof(COMMAND_TYPE(bindPipeline)) +
+                sizeof(COMMAND_TYPE(setPushConstant)) +
+                sizeof(COMMAND_TYPE(bindRenderPrimitive)) +
+                sizeof(COMMAND_TYPE(draw2));
+
 
         // Number of Commands that can be issued and guaranteed to fit in the current
         // CircularBuffer allocation. In practice, we'll have tons of headroom especially if
         // skinning and morphing aren't used. With a 2 MiB buffer (the default) a batch is
-        // 8192 commands (i.e. draw calls).
+        // 6553 commands (i.e. draw calls).
         size_t const batchCommandCount = capacity / maxCommandSizeInBytes;
         while(first != last) {
             Command const* const batchLast = std::min(first + batchCommandCount, last);

filament/src/RenderPass.h

@@ -417,8 +417,7 @@ private:
     void instanceify(FEngine& engine, Arena& arena, int32_t eyeCount) noexcept;
 
     // We choose the command count per job to minimize JobSystem overhead.
-    // On a Pixel 4, 2048 commands is about half a millisecond of processing.
-    static constexpr size_t JOBS_PARALLEL_FOR_COMMANDS_COUNT = 2048;
+    static constexpr size_t JOBS_PARALLEL_FOR_COMMANDS_COUNT = 128;
     static constexpr size_t JOBS_PARALLEL_FOR_COMMANDS_SIZE  =
             sizeof(Command) * JOBS_PARALLEL_FOR_COMMANDS_COUNT;
 

filament/src/Renderer.cpp

@@ -14,11 +14,20 @@
  * limitations under the License.
  */
 
+#include <filament/Renderer.h>
+
 #include "details/Renderer.h"
 
 #include "details/Engine.h"
 #include "details/View.h"
 
+#include <utils/FixedCapacityVector.h>
+
+#include <utility>
+
+#include <stddef.h>
+#include <stdint.h>
+
 namespace filament {
 
 using namespace math;
@@ -36,6 +45,10 @@ void Renderer::setPresentationTime(int64_t monotonic_clock_ns) {
     downcast(this)->setPresentationTime(monotonic_clock_ns);
 }
 
+void Renderer::skipFrame(uint64_t vsyncSteadyClockTimeNano) {
+    downcast(this)->skipFrame(vsyncSteadyClockTimeNano);
+}
+
 bool Renderer::beginFrame(SwapChain* swapChain, uint64_t vsyncSteadyClockTimeNano) {
     return downcast(this)->beginFrame(downcast(swapChain), vsyncSteadyClockTimeNano);
 }
@@ -93,4 +106,12 @@ void Renderer::setVsyncTime(uint64_t steadyClockTimeNano) noexcept {
     downcast(this)->setVsyncTime(steadyClockTimeNano);
 }
 
+utils::FixedCapacityVector<Renderer::FrameInfo> Renderer::getFrameInfoHistory(size_t historySize) const noexcept {
+    return downcast(this)->getFrameInfoHistory(historySize);
+}
+
+size_t Renderer::getMaxFrameHistorySize() const noexcept {
+    return downcast(this)->getMaxFrameHistorySize();
+}
+
 } // namespace filament

filament/src/RendererUtils.cpp

@@ -172,7 +172,7 @@ FrameGraphId<FrameGraphTexture> RendererUtils::colorPass(
 
                 data.color = builder.write(data.color, FrameGraphTexture::Usage::COLOR_ATTACHMENT);
                 data.depth = builder.write(data.depth, FrameGraphTexture::Usage::DEPTH_ATTACHMENT);
-                if (engine.getConfig().stereoscopicType == StereoscopicType::MULTIVIEW) {
+                if (view.hasStereo() && engine.getConfig().stereoscopicType == StereoscopicType::MULTIVIEW) {
                     layerCount = engine.getConfig().stereoscopicEyeCount;
                 }
 

filament/src/ResourceAllocator.cpp

@@ -27,15 +27,18 @@
 
 #include "private/backend/DriverApi.h"
 
+#include <utils/algorithm.h>
+#include <utils/bitset.h>
 #include <utils/compiler.h>
 #include <utils/debug.h>
-#include <utils/FixedCapacityVector.h>
 #include <utils/Log.h>
 #include <utils/ostream.h>
 
 #include <array>
 #include <algorithm>
 #include <iterator>
+#include <memory>
+#include <optional>
 #include <utility>
 
 #include <stddef.h>
@@ -89,8 +92,15 @@ void ResourceAllocator::AssociativeContainer<K, V, H>::emplace(ARGS&& ... args)
 }
 
 // ------------------------------------------------------------------------------------------------
+
 ResourceAllocatorInterface::~ResourceAllocatorInterface() = default;
 
+// ------------------------------------------------------------------------------------------------
+
+ResourceAllocatorDisposerInterface::~ResourceAllocatorDisposerInterface() = default;
+
+// ------------------------------------------------------------------------------------------------
+
 size_t ResourceAllocator::TextureKey::getSize() const noexcept {
     size_t const pixelCount = width * height * depth;
     size_t size = pixelCount * FTexture::getFormatSize(format);
@@ -110,16 +120,22 @@ size_t ResourceAllocator::TextureKey::getSize() const noexcept {
 
 ResourceAllocator::ResourceAllocator(Engine::Config const& config, DriverApi& driverApi) noexcept
         : mCacheMaxAge(config.resourceAllocatorCacheMaxAge),
-          mBackend(driverApi) {
+          mBackend(driverApi),
+          mDisposer(std::make_shared<ResourceAllocatorDisposer>(driverApi)) {
+}
+
+ResourceAllocator::ResourceAllocator(std::shared_ptr<ResourceAllocatorDisposer> disposer,
+        Engine::Config const& config, DriverApi& driverApi) noexcept
+        : mCacheMaxAge(config.resourceAllocatorCacheMaxAge),
+          mBackend(driverApi),
+          mDisposer(std::move(disposer)) {
 }
 
 ResourceAllocator::~ResourceAllocator() noexcept {
-    assert_invariant(!mTextureCache.size());
-    assert_invariant(!mInUseTextures.size());
+    assert_invariant(mTextureCache.empty());
 }
 
 void ResourceAllocator::terminate() noexcept {
-    assert_invariant(!mInUseTextures.size());
     auto& textureCache = mTextureCache;
     for (auto it = textureCache.begin(); it != textureCache.end();) {
         mBackend.destroyTexture(it->second.handle);
@@ -174,7 +190,7 @@ backend::TextureHandle ResourceAllocator::createTexture(const char* name,
                         swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
             }
         }
-        mInUseTextures.emplace(handle, key);
+        mDisposer->checkout(handle, key);
     } else {
         if (swizzle == defaultSwizzle) {
             handle = mBackend.createTexture(
@@ -190,58 +206,99 @@ backend::TextureHandle ResourceAllocator::createTexture(const char* name,
 
 void ResourceAllocator::destroyTexture(TextureHandle h) noexcept {
     if constexpr (mEnabled) {
-        // find the texture in the in-use list (it must be there!)
-        auto it = mInUseTextures.find(h);
-        assert_invariant(it != mInUseTextures.end());
-
-        // move it to the cache
-        const TextureKey key = it->second;
-        uint32_t const size = key.getSize();
-
-        mTextureCache.emplace(key, TextureCachePayload{ h, mAge, size });
-        mCacheSize += size;
-
-        // remove it from the in-use list
-        mInUseTextures.erase(it);
+        auto const key = mDisposer->checkin(h);
+        if (UTILS_LIKELY(key.has_value())) {
+            uint32_t const size = key.value().getSize();
+            mTextureCache.emplace(key.value(), TextureCachePayload{ h, mAge, size });
+            mCacheSize += size;
+            mCacheSizeHiWaterMark = std::max(mCacheSizeHiWaterMark, mCacheSize);
+        }
     } else {
         mBackend.destroyTexture(h);
     }
 }
 
-void ResourceAllocator::gc() noexcept {
-    // this is called regularly -- usually once per frame of each Renderer
+ResourceAllocatorDisposerInterface& ResourceAllocator::getDisposer() noexcept {
+    return *mDisposer;
+}
 
-    // increase our age
-    const size_t age = mAge++;
+void ResourceAllocator::gc(bool skippedFrame) noexcept {
+    // this is called regularly -- usually once per frame
+
+    // increase our age at each (non-skipped) frame
+    const size_t age = mAge;
+    if (!skippedFrame) {
+        mAge++;
+    }
 
     // Purging strategy:
-    //  - remove entries that are older than a certain age
-    //      - remove only one entry per gc(),
+    //  - remove all entries older than MAX_AGE_SKIPPED_FRAME when skipping a frame
+    //  - remove entries older than mCacheMaxAgeSoft
+    //      - remove only MAX_EVICTION_COUNT entry per gc(),
+    //  - look for the number of unique resource ages present in the cache (this basically gives
+    //    us how many buckets of resources we have corresponding to previous frames.
+    //      - remove all resources that have an age older than the MAX_UNIQUE_AGE_COUNT'th bucket
 
     auto& textureCache = mTextureCache;
+
+    // when skipping a frame, the maximum age to keep in the cache
+    constexpr size_t MAX_AGE_SKIPPED_FRAME = 1;
+
+    // maximum entry count to evict per GC, under the mCacheMaxAgeSoft limit
+    constexpr size_t MAX_EVICTION_COUNT = 1;
+
+    // maximum number of unique ages in the cache
+    constexpr size_t MAX_UNIQUE_AGE_COUNT = 3;
+
+    utils::bitset32 ages;
+    uint32_t evictedCount = 0;
     for (auto it = textureCache.begin(); it != textureCache.end();) {
-        const size_t ageDiff = age - it->second.age;
-        if (ageDiff >= mCacheMaxAge) {
+        size_t const ageDiff = age - it->second.age;
+        if ((ageDiff >= MAX_AGE_SKIPPED_FRAME && skippedFrame) ||
+            (ageDiff >= mCacheMaxAge && evictedCount < MAX_EVICTION_COUNT)) {
+            evictedCount++;
             purge(it);
-            // only purge a single entry per gc
-            break;
         } else {
+            // build the set of ages present in the cache after eviction
+            ages.set(std::min(size_t(31), ageDiff));
             ++it;
         }
     }
+
+    // if we have MAX_UNIQUE_AGE_COUNT ages or more, we evict all the resources that
+    // are older than the MAX_UNIQUE_AGE_COUNT'th age.
+    if (!skippedFrame && ages.count() >= MAX_UNIQUE_AGE_COUNT) {
+        uint32_t bits = ages.getValue();
+        // remove from the set the ages we keep
+        for (size_t i = 0; i < MAX_UNIQUE_AGE_COUNT - 1; i++) {
+            bits &= ~(1 << utils::ctz(bits));
+        }
+        size_t const maxAge = utils::ctz(bits);
+        for (auto it = textureCache.begin(); it != textureCache.end();) {
+            const size_t ageDiff = age - it->second.age;
+            if (ageDiff >= maxAge) {
+                purge(it);
+            } else {
+                ++it;
+            }
+        }
+    }
 }
 
 UTILS_NOINLINE
 void ResourceAllocator::dump(bool brief) const noexcept {
-    slog.d << "# entries=" << mTextureCache.size() << ", sz=" << mCacheSize / float(1u << 20u)
-           << " MiB" << io::endl;
+    constexpr float MiB = 1.0f / float(1u << 20u);
+    slog.d  << "# entries=" << mTextureCache.size()
+            << ", sz=" << (float)mCacheSize * MiB << " MiB"
+            << ", max=" << (float)mCacheSizeHiWaterMark * MiB << " MiB"
+            << io::endl;
     if (!brief) {
         for (auto const& it : mTextureCache) {
             auto w = it.first.width;
             auto h = it.first.height;
             auto f = FTexture::getFormatSize(it.first.format);
             slog.d << it.first.name << ": w=" << w << ", h=" << h << ", f=" << f << ", sz="
-                   << it.second.size / float(1u << 20u) << io::endl;
+                   << (float)it.second.size * MiB << io::endl;
         }
     }
 }
@@ -254,4 +311,46 @@ void ResourceAllocator::purge(
     mTextureCache.erase(pos);
 }
 
+// ------------------------------------------------------------------------------------------------
+
+ResourceAllocatorDisposer::ResourceAllocatorDisposer(DriverApi& driverApi) noexcept
+        : mBackend(driverApi) {
+}
+
+ResourceAllocatorDisposer::~ResourceAllocatorDisposer() noexcept {
+     assert_invariant(mInUseTextures.empty());
+}
+
+void ResourceAllocatorDisposer::terminate() noexcept {
+    assert_invariant(mInUseTextures.empty());
+}
+
+void ResourceAllocatorDisposer::destroy(backend::TextureHandle handle) noexcept {
+    if (handle) {
+        auto r = checkin(handle);
+        if (r.has_value()) {
+            mBackend.destroyTexture(handle);
+        }
+    }
+}
+
+void ResourceAllocatorDisposer::checkout(backend::TextureHandle handle,
+        ResourceAllocator::TextureKey key) {
+    mInUseTextures.emplace(handle, key);
+}
+
+std::optional<ResourceAllocator::TextureKey> ResourceAllocatorDisposer::checkin(
+        backend::TextureHandle handle) {
+    // find the texture in the in-use list (it must be there!)
+    auto it = mInUseTextures.find(handle);
+    assert_invariant(it != mInUseTextures.end());
+    if (it == mInUseTextures.end()) {
+        return std::nullopt;
+    }
+    TextureKey const key = it->second;
+    // remove it from the in-use list
+    mInUseTextures.erase(it);
+    return key;
+}
+
 } // namespace filament

filament/src/ResourceAllocator.h

@@ -28,16 +28,29 @@
 #include <utils/Hash.h>
 
 #include <array>
-#include <vector>
+#include <memory>
+#include <optional>
 #include <utility>
+#include <vector>
 
+#include <cstddef>
 #include <stddef.h>
 #include <stdint.h>
 
 namespace filament {
 
+class ResourceAllocatorDisposer;
+
 // The only reason we use an interface here is for unit-tests, so we can mock this allocator.
 // This is not too time-critical, so that's okay.
+
+class ResourceAllocatorDisposerInterface {
+public:
+    virtual void destroy(backend::TextureHandle handle) noexcept = 0;
+protected:
+    virtual ~ResourceAllocatorDisposerInterface();
+};
+
 class ResourceAllocatorInterface {
 public:
     virtual backend::RenderTargetHandle createRenderTarget(const char* name,
@@ -60,15 +73,20 @@ public:
 
     virtual void destroyTexture(backend::TextureHandle h) noexcept = 0;
 
+    virtual ResourceAllocatorDisposerInterface& getDisposer() noexcept = 0;
+
 protected:
     virtual ~ResourceAllocatorInterface();
 };
 
-
 class ResourceAllocator final : public ResourceAllocatorInterface {
 public:
+    explicit ResourceAllocator(std::shared_ptr<ResourceAllocatorDisposer> disposer,
+            Engine::Config const& config, backend::DriverApi& driverApi) noexcept;
+
     explicit ResourceAllocator(
             Engine::Config const& config, backend::DriverApi& driverApi) noexcept;
+
     ~ResourceAllocator() noexcept override;
 
     void terminate() noexcept;
@@ -93,7 +111,9 @@ public:
 
     void destroyTexture(backend::TextureHandle h) noexcept override;
 
-    void gc() noexcept;
+    ResourceAllocatorDisposerInterface& getDisposer() noexcept override;
+
+    void gc(bool skippedFrame = false) noexcept;
 
 private:
     size_t const mCacheMaxAge;
@@ -181,6 +201,7 @@ private:
         using value_type = typename Container::value_type::second_type;
 
         size_t size() const { return mContainer.size(); }
+        bool empty() const { return size() == 0; }
         iterator begin() { return mContainer.begin(); }
         const_iterator begin() const { return mContainer.begin(); }
         iterator end() { return mContainer.end(); }
@@ -193,16 +214,36 @@ private:
     };
 
     using CacheContainer = AssociativeContainer<TextureKey, TextureCachePayload>;
-    using InUseContainer = AssociativeContainer<backend::TextureHandle, TextureKey>;
 
     void purge(ResourceAllocator::CacheContainer::iterator const& pos);
 
     backend::DriverApi& mBackend;
+    std::shared_ptr<ResourceAllocatorDisposer> mDisposer;
     CacheContainer mTextureCache;
-    InUseContainer mInUseTextures;
     size_t mAge = 0;
     uint32_t mCacheSize = 0;
+    uint32_t mCacheSizeHiWaterMark = 0;
     static constexpr bool mEnabled = true;
+
+    friend class ResourceAllocatorDisposer;
+};
+
+class ResourceAllocatorDisposer final : public ResourceAllocatorDisposerInterface {
+    using TextureKey = ResourceAllocator::TextureKey;
+public:
+    explicit ResourceAllocatorDisposer(backend::DriverApi& driverApi) noexcept;
+    ~ResourceAllocatorDisposer() noexcept override;
+    void terminate() noexcept;
+    void destroy(backend::TextureHandle handle) noexcept override;
+
+private:
+    friend class ResourceAllocator;
+    void checkout(backend::TextureHandle handle, TextureKey key);
+    std::optional<TextureKey> checkin(backend::TextureHandle handle);
+
+    using InUseContainer = ResourceAllocator::AssociativeContainer<backend::TextureHandle, TextureKey>;
+    backend::DriverApi& mBackend;
+    InUseContainer mInUseTextures;
 };
 
 } // namespace filament

filament/src/View.cpp

@@ -15,6 +15,8 @@
  */
 
 #include "details/View.h"
+#include "filament/View.h"
+
 
 namespace filament {
 
@@ -312,4 +314,8 @@ utils::Entity View::getFogEntity() const noexcept {
     return downcast(this)->getFogEntity();
 }
 
+void View::clearFrameHistory(Engine& engine) noexcept {
+    downcast(this)->clearFrameHistory(downcast(engine));
+}
+
 } // namespace filament

filament/src/components/LightManager.cpp

@@ -20,12 +20,13 @@
 
 #include "details/Engine.h"
 
-#include <math/fast.h>
-#include <math/scalar.h>
-
-#include <utils/debug.h>
 #include <filament/LightManager.h>
 
+#include <utils/debug.h>
+#include <utils/Log.h>
+
+#include <math/fast.h>
+#include <math/scalar.h>
 
 using namespace filament::math;
 using namespace utils;

filament/src/components/RenderableManager.cpp

@@ -443,10 +443,10 @@ RenderableManager::Builder::Result RenderableManager::Builder::build(Engine& eng
             << ") instances when supplying transforms via an InstanceBuffer.";
 
     if (mImpl->mGeometryType == GeometryType::STATIC) {
-        FILAMENT_CHECK_PRECONDITION(mImpl->mSkinningBoneCount > 0)
+        FILAMENT_CHECK_PRECONDITION(mImpl->mSkinningBoneCount == 0)
                 << "Skinning can't be used with STATIC geometry";
 
-        FILAMENT_CHECK_PRECONDITION(mImpl->mMorphTargetCount > 0)
+        FILAMENT_CHECK_PRECONDITION(mImpl->mMorphTargetCount == 0)
                 << "Morphing can't be used with STATIC geometry";
     }
 

filament/src/details/Engine.cpp

@@ -163,8 +163,7 @@ FEngine* FEngine::getEngine(void* token) {
     FILAMENT_CHECK_PRECONDITION(ThreadUtils::isThisThread(instance->mMainThreadId))
             << "Engine::createAsync() and Engine::getEngine() must be called on the same thread.";
 
-    // we use mResourceAllocator as a proxy for "am I already initialized"
-    if (!instance->mResourceAllocator) {
+    if (!instance->mInitialized) {
         if (UTILS_UNLIKELY(!instance->mDriver)) {
             // something went horribly wrong during driver initialization
             instance->mDriverThread.join();
@@ -262,7 +261,7 @@ void FEngine::init() {
     slog.i << "FEngine feature level: " << int(mActiveFeatureLevel) << io::endl;
 
 
-    mResourceAllocator = new ResourceAllocator(mConfig, driverApi);
+    mResourceAllocatorDisposer = std::make_shared<ResourceAllocatorDisposer>(driverApi);
 
     mFullScreenTriangleVb = downcast(VertexBuffer::Builder()
             .vertexCount(3)
@@ -425,11 +424,13 @@ void FEngine::init() {
                     }
                 });
             });
+
+    mInitialized = true;
 }
 
 FEngine::~FEngine() noexcept {
     SYSTRACE_CALL();
-    delete mResourceAllocator;
+    assert_invariant(!mResourceAllocatorDisposer);
     delete mDriver;
     if (mOwnPlatform) {
         PlatformFactory::destroy(&mPlatform);
@@ -440,7 +441,7 @@ void FEngine::shutdown() {
     SYSTRACE_CALL();
 
     // by construction this should never be nullptr
-    assert_invariant(mResourceAllocator);
+    assert_invariant(mResourceAllocatorDisposer);
 
     FILAMENT_CHECK_PRECONDITION(ThreadUtils::isThisThread(mMainThreadId))
             << "Engine::shutdown() called from the wrong thread!";
@@ -460,7 +461,8 @@ void FEngine::shutdown() {
      */
 
     mPostProcessManager.terminate(driver);  // free-up post-process manager resources
-    mResourceAllocator->terminate();
+    mResourceAllocatorDisposer->terminate();
+    mResourceAllocatorDisposer.reset();
     mDFG.terminate(*this);                  // free-up the DFG
     mRenderableManager.terminate();         // free-up all renderables
     mLightManager.terminate();              // free-up all lights

filament/src/details/Engine.h

@@ -86,6 +86,7 @@ namespace filament {
 
 class Renderer;
 class MaterialParser;
+class ResourceAllocatorDisposer;
 
 namespace backend {
 class Driver;
@@ -254,9 +255,13 @@ public:
         }
     }
 
-    ResourceAllocator& getResourceAllocator() noexcept {
-        assert_invariant(mResourceAllocator);
-        return *mResourceAllocator;
+    ResourceAllocatorDisposer& getResourceAllocatorDisposer() noexcept {
+        assert_invariant(mResourceAllocatorDisposer);
+        return *mResourceAllocatorDisposer;
+    }
+
+    std::shared_ptr<ResourceAllocatorDisposer> const& getSharedResourceAllocatorDisposer() noexcept {
+        return mResourceAllocatorDisposer;
     }
 
     void* streamAlloc(size_t size, size_t alignment) noexcept;
@@ -490,7 +495,7 @@ private:
     FTransformManager mTransformManager;
     FLightManager mLightManager;
     FCameraManager mCameraManager;
-    ResourceAllocator* mResourceAllocator = nullptr;
+    std::shared_ptr<ResourceAllocatorDisposer> mResourceAllocatorDisposer;
     HwVertexBufferInfoFactory mHwVertexBufferInfoFactory;
 
     ResourceList<FBufferObject> mBufferObjects{ "BufferObject" };
@@ -562,6 +567,8 @@ private:
 
     std::thread::id mMainThreadId{};
 
+    bool mInitialized = false;
+
     // Creation parameters
     Config mConfig;
 

filament/src/details/Renderer.cpp

@@ -60,6 +60,7 @@
 
 #include <chrono>
 #include <limits>
+#include <memory>
 #include <utility>
 
 #include <stddef.h>
@@ -88,7 +89,11 @@ FRenderer::FRenderer(FEngine& engine) :
         mHdrQualityMedium(TextureFormat::R11F_G11F_B10F),
         mHdrQualityHigh(TextureFormat::RGB16F),
         mIsRGB8Supported(false),
-        mUserEpoch(engine.getEngineEpoch())
+        mUserEpoch(engine.getEngineEpoch()),
+        mResourceAllocator(std::make_unique<ResourceAllocator>(
+                engine.getSharedResourceAllocatorDisposer(),
+                engine.getConfig(),
+                engine.getDriverApi()))
 {
     FDebugRegistry& debugRegistry = engine.getDebugRegistry();
     debugRegistry.registerProperty("d.renderer.doFrameCapture",
@@ -176,6 +181,7 @@ void FRenderer::terminate(FEngine& engine) {
     }
     mFrameInfoManager.terminate(driver);
     mFrameSkipper.terminate(driver);
+    mResourceAllocator->terminate();
 }
 
 void FRenderer::resetUserTime() {
@@ -236,6 +242,39 @@ void FRenderer::setVsyncTime(uint64_t steadyClockTimeNano) noexcept {
     mVsyncSteadyClockTimeNano = steadyClockTimeNano;
 }
 
+void FRenderer::skipFrame(uint64_t vsyncSteadyClockTimeNano) {
+    SYSTRACE_CALL();
+
+    FILAMENT_CHECK_PRECONDITION(!mSwapChain) <<
+            "skipFrame() can't be called between beginFrame() and endFrame()";
+
+    if (!vsyncSteadyClockTimeNano) {
+        vsyncSteadyClockTimeNano = mVsyncSteadyClockTimeNano;
+        mVsyncSteadyClockTimeNano = 0;
+    }
+
+    FEngine& engine = mEngine;
+    FEngine::DriverApi& driver = engine.getDriverApi();
+
+    // Gives the backend a chance to execute periodic tasks. This must be called before
+    // the frame skipper.
+    driver.tick();
+
+    // 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);
+}
+
 bool FRenderer::beginFrame(FSwapChain* swapChain, uint64_t vsyncSteadyClockTimeNano) {
     assert_invariant(swapChain);
 
@@ -380,7 +419,7 @@ void FRenderer::endFrame() {
     }
 
     // do this before engine.flush()
-    engine.getResourceAllocator().gc();
+    mResourceAllocator->gc();
 
     // Run the component managers' GC in parallel
     // WARNING: while doing this we can't access any component manager
@@ -468,6 +507,9 @@ void FRenderer::renderStandaloneView(FView const* view) {
     FILAMENT_CHECK_PRECONDITION(view->getRenderTarget())
             << "View \"" << view->getName() << "\" must have a RenderTarget associated";
 
+    FILAMENT_CHECK_PRECONDITION(!mSwapChain)
+            << "renderStandaloneView() must be called outside of beginFrame() / endFrame()";
+
     if (UTILS_LIKELY(view->getScene())) {
         mPreviousRenderTargets.clear();
         mFrameId++;
@@ -747,7 +789,7 @@ void FRenderer::renderJob(RootArenaScope& rootArenaScope, FView& view) {
      * Frame graph
      */
 
-    FrameGraph fg(engine.getResourceAllocator(),
+    FrameGraph fg(*mResourceAllocator,
             isProtectedContent ? FrameGraph::Mode::PROTECTED : FrameGraph::Mode::UNPROTECTED);
     auto& blackboard = fg.getBlackboard();
 

filament/src/details/Renderer.h

@@ -37,6 +37,7 @@
 
 #include <utils/compiler.h>
 #include <utils/Allocator.h>
+#include <utils/FixedCapacityVector.h>
 
 #include <math/vec4.h>
 
@@ -45,6 +46,7 @@
 #include <algorithm>
 #include <chrono>
 #include <functional>
+#include <memory>
 #include <utility>
 
 #include <stddef.h>
@@ -52,6 +54,8 @@
 
 namespace filament {
 
+class ResourceAllocator;
+
 namespace backend {
 class Driver;
 } // namespace backend
@@ -87,6 +91,9 @@ public:
 
     void setVsyncTime(uint64_t steadyClockTimeNano) noexcept;
 
+    // skip a frame
+    void skipFrame(uint64_t vsyncSteadyClockTimeNano);
+
     // start a frame
     bool beginFrame(FSwapChain* swapChain, uint64_t vsyncSteadyClockTimeNano);
 
@@ -140,6 +147,14 @@ public:
         return mClearOptions;
     }
 
+    utils::FixedCapacityVector<Renderer::FrameInfo> getFrameInfoHistory(size_t historySize) const noexcept {
+        return mFrameInfoManager.getFrameInfoHistory(historySize);
+    }
+
+    size_t getMaxFrameHistorySize() const noexcept {
+        return MAX_FRAMETIME_HISTORY;
+    }
+
 private:
     friend class Renderer;
     using Command = RenderPass::Command;
@@ -197,6 +212,7 @@ private:
     tsl::robin_set<FRenderTarget*> mPreviousRenderTargets;
     std::function<void()> mBeginFrameInternal;
     uint64_t mVsyncSteadyClockTimeNano = 0;
+    std::unique_ptr<ResourceAllocator> mResourceAllocator{};
 };
 
 FILAMENT_DOWNCAST(Renderer)

filament/src/details/Scene.cpp

@@ -248,7 +248,7 @@ void FScene::prepare(utils::JobSystem& js,
 
     auto* renderableJob = jobs::parallel_for(js, rootJob,
             renderableInstances.data(), renderableInstances.size(),
-            std::cref(renderableWork), jobs::CountSplitter<128, 5>());
+            std::cref(renderableWork), jobs::CountSplitter<64>());
 
     auto* lightJob = jobs::parallel_for(js, rootJob,
             lightInstances.data(), lightInstances.size(),

filament/src/details/View.cpp

@@ -17,6 +17,7 @@
 #include "details/View.h"
 
 #include "Culler.h"
+#include "FrameHistory.h"
 #include "Froxelizer.h"
 #include "RenderPrimitive.h"
 #include "ResourceAllocator.h"
@@ -118,7 +119,7 @@ void FView::terminate(FEngine& engine) {
     DriverApi& driver = engine.getDriverApi();
     driver.destroyBufferObject(mLightUbh);
     driver.destroyBufferObject(mRenderableUbh);
-    drainFrameHistory(engine);
+    clearFrameHistory(engine);
 
     ShadowMapManager::terminate(engine, mShadowMapManager);
     mPerViewUniforms.terminate(driver);
@@ -162,7 +163,7 @@ void FView::setDynamicLightingOptions(float zLightNear, float zLightFar) noexcep
 }
 
 float2 FView::updateScale(FEngine& engine,
-        FrameInfo const& info,
+        filament::details::FrameInfo const& info,
         Renderer::FrameRateOptions const& frameRateOptions,
         Renderer::DisplayInfo const& displayInfo) noexcept {
 
@@ -1007,20 +1008,29 @@ FrameGraphId<FrameGraphTexture> FView::renderShadowMaps(FEngine& engine, FrameGr
 
 void FView::commitFrameHistory(FEngine& engine) noexcept {
     // Here we need to destroy resources in mFrameHistory.back()
+    auto& disposer = engine.getResourceAllocatorDisposer();
     auto& frameHistory = mFrameHistory;
 
     FrameHistoryEntry& last = frameHistory.back();
-    last.taa.color.destroy(engine.getResourceAllocator());
-    last.ssr.color.destroy(engine.getResourceAllocator());
+    disposer.destroy(last.taa.color.handle);
+    disposer.destroy(last.ssr.color.handle);
+    last.taa.color.handle.clear();
+    last.ssr.color.handle.clear();
 
     // and then push the new history entry to the history stack
     frameHistory.commit();
 }
 
-void FView::drainFrameHistory(FEngine& engine) noexcept {
+void FView::clearFrameHistory(FEngine& engine) noexcept {
     // make sure we free all resources in the history
-    for (size_t i = 0; i < mFrameHistory.size(); ++i) {
-        commitFrameHistory(engine);
+    auto& disposer = engine.getResourceAllocatorDisposer();
+    auto& frameHistory = mFrameHistory;
+    for (size_t i = 0; i < frameHistory.size(); ++i) {
+        FrameHistoryEntry& last = frameHistory[i];
+        disposer.destroy(last.taa.color.handle);
+        disposer.destroy(last.ssr.color.handle);
+        last.taa.color.handle.clear();
+        last.ssr.color.handle.clear();
     }
 }
 

filament/src/details/View.h

@@ -283,7 +283,7 @@ public:
     }
 
     math::float2 updateScale(FEngine& engine,
-            FrameInfo const& info,
+            details::FrameInfo const& info,
             Renderer::FrameRateOptions const& frameRateOptions,
             Renderer::DisplayInfo const& displayInfo) noexcept;
 
@@ -422,6 +422,10 @@ public:
     // (e.g.: after the FrameGraph execution).
     void commitFrameHistory(FEngine& engine) noexcept;
 
+    // Clean-up the whole history, free all resources. This is typically called when the View is
+    // being terminated. Or we're changing Renderer.
+    void clearFrameHistory(FEngine& engine) noexcept;
+
     // create the picking query
     View::PickingQuery& pick(uint32_t x, uint32_t y, backend::CallbackHandler* handler,
             View::PickingQueryResultCallback callback) noexcept;
@@ -485,10 +489,6 @@ private:
             Culler::result_type* visibleMask,
             size_t count);
 
-    // Clean-up the whole history, free all resources. This is typically called when the View is
-    // being terminated.
-    void drainFrameHistory(FEngine& engine) noexcept;
-
     // we don't inline this one, because the function is quite large and there is not much to
     // gain from inlining.
     static FScene::RenderableSoa::iterator partition(

filament/src/fg/FrameGraph.cpp

@@ -24,6 +24,7 @@
 #include "FrameGraphPass.h"
 #include "FrameGraphRenderPass.h"
 #include "FrameGraphTexture.h"
+#include "ResourceAllocator.h"
 
 #include "details/Engine.h"
 

filament/test/CMakeLists.txt

@@ -62,6 +62,7 @@ if (ANDROID)
     target_link_libraries(test_compiler PRIVATE utils)
     target_link_libraries(test_compiler PRIVATE EGL)
     target_link_libraries(test_compiler PRIVATE GLESv3)
+    target_link_libraries(test_compiler PRIVATE bluevk)    
 endif()
 
 add_executable(test_material_parser

filament/test/compiler_test.cpp

@@ -1,181 +1,82 @@
-/*
- * Copyright (C) 2023 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 <iostream>
+#include <utils/Hash.h>
+#include <tsl/robin_map.h>
+#include <bluevk/BlueVK.h>
 
-#include <utils/Log.h>
+namespace {
 
-#include <GLES3/gl3.h>
-#include <EGL/egl.h>
-#include <EGL/eglext.h>
-
-#include <gtest/gtest.h>
-
-using namespace utils;
-using namespace std::literals;
-
-class CompilerTest : public testing::Test {
-protected:
-    void SetUp() override {
-        EGLBoolean success;
-        EGLint major, minor;
-        dpy = eglGetDisplay(EGL_DEFAULT_DISPLAY);
-        ASSERT_NE(dpy, EGL_NO_DISPLAY);
-
-        EGLBoolean const initialized = eglInitialize(dpy, &major, &minor);
-        ASSERT_TRUE(initialized);
-
-        EGLint const contextAttribs[] = {
-                EGL_CONTEXT_CLIENT_VERSION, 3,
-                EGL_NONE
-        };
-
-        EGLint configsCount;
-        EGLint configAttribs[] = {
-                EGL_RENDERABLE_TYPE, EGL_OPENGL_ES3_BIT_KHR,        //  0
-                EGL_RED_SIZE,    8,                                 //  2
-                EGL_GREEN_SIZE,  8,                                 //  4
-                EGL_BLUE_SIZE,   8,                                 //  6
-                EGL_NONE                                            // 14
-        };
-        success = eglChooseConfig(dpy, configAttribs, &config, 1, &configsCount);
-        ASSERT_TRUE(success);
-
-        context = eglCreateContext(dpy, config, EGL_NO_CONTEXT, contextAttribs);
-        ASSERT_NE(context, EGL_NO_CONTEXT);
-
-        success = eglMakeCurrent(dpy, EGL_NO_SURFACE, EGL_NO_SURFACE, context);
-        ASSERT_TRUE(success);
-
-        ASSERT_EQ(eglGetError(), EGL_SUCCESS);
-    }
-
-    void TearDown() override {
-        eglMakeCurrent(dpy, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
-        eglDestroyContext(dpy, context);
-        eglTerminate(dpy);
-    }
-
-private:
-    EGLDisplay dpy;
-    EGLContext context;
-    EGLConfig config;
+struct T {
 };
 
-TEST_F(CompilerTest, Simple) {
-    auto shader = R"(
-#version 300 es
-void main()
-{
-})"sv;
+constexpr uint8_t const UNIQUE_DESCRIPTOR_SET_COUNT = 3;
+constexpr uint8_t const SHADER_TYPE_COUNT = 3;
+using Timestamp = uint64_t;
+using VulkanResourceAllocator = T;
 
-    const char* const src = shader.data();
-    GLint const len = (GLint)shader.size();
+struct PushConstantKey {
+    uint8_t stage;// We have one set of push constant per shader stage (fragment, vertex, etc).
+    uint8_t size;
+};
 
-    GLuint const id = glCreateShader(GL_VERTEX_SHADER);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
+struct PipelineLayoutKey {
+    using DescriptorSetLayoutArray = std::array<VkDescriptorSetLayout, UNIQUE_DESCRIPTOR_SET_COUNT>;
+    DescriptorSetLayoutArray descSetLayouts = {};        // 8 * 3
+    PushConstantKey pushConstant[SHADER_TYPE_COUNT] = {};// 2 * 3
+    uint16_t padding = 0;
+};
+static_assert(sizeof(PipelineLayoutKey) == 32);
 
-    glShaderSource(id, 1, &src, &len);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
+using PipelineLayoutKeyHashFn = utils::hash::MurmurHashFn<PipelineLayoutKey>;
+struct PipelineLayoutKeyEqual {
+    bool operator()(PipelineLayoutKey const& k1, PipelineLayoutKey const& k2) const {
+        return 0 == memcmp((const void*) &k1, (const void*) &k2, sizeof(PipelineLayoutKey));
+    }
+};
 
-    glCompileShader(id);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
+struct PipelineLayoutCacheEntry {
+    VkPipelineLayout handle;
+    Timestamp lastUsed;
+};
 
-    GLint result = 0;
-    glGetShaderiv(id, GL_COMPILE_STATUS, &result);
-    EXPECT_EQ(result, GL_TRUE);
+using PipelineLayoutMap = tsl::robin_map<PipelineLayoutKey, PipelineLayoutCacheEntry,
+        PipelineLayoutKeyHashFn, PipelineLayoutKeyEqual>;
 
-    glDeleteShader(id);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
+class TestClass {
+public:
+    VkDevice mDevice;
+    VulkanResourceAllocator* mAllocator;
+    Timestamp mTimestamp;
+    PipelineLayoutMap layout;
+};
+
+void insert(PipelineLayoutMap& layout, uint64_t val) {
+    PipelineLayoutKey key{};
+    key.descSetLayouts[0] = (VkDescriptorSetLayout) val;
+    layout[key] = {(VkPipelineLayout) val, val};
 }
 
-TEST_F(CompilerTest, CrashPVRUniFlexCompileToHw) {
-
-    // Some PowerVR driver crash with this shader
-
-    auto shader = R"(
-#version 300 es
-
-layout(location = 0) in vec4 mesh_position;
-
-layout(std140) uniform FrameUniforms {
-    vec2 i;
-} frameUniforms;
-
-void main() {
-    gl_Position = mesh_position;
-    gl_Position.z = dot(gl_Position.zw, frameUniforms.i);
-})"sv;
-
-    const char* const src = shader.data();
-    GLint const len = (GLint)shader.size();
-
-    GLuint const id = glCreateShader(GL_VERTEX_SHADER);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    glShaderSource(id, 1, &src, &len);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    glCompileShader(id);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    GLint result = 0;
-    glGetShaderiv(id, GL_COMPILE_STATUS, &result);
-    EXPECT_EQ(result, GL_TRUE);
-
-    glDeleteShader(id);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-}
-
-TEST_F(CompilerTest, ConstParameters) {
-
-    // Some PowerVR driver fail to compile this shader
-
-    auto shader = R"(
-#version 300 es
-
-highp mat3 m;
-void buggy(const mediump vec3 n) {
-    m*n;
-}
-
-void main() {
-})"sv;
-
-    const char* const src = shader.data();
-    GLint const len = (GLint)shader.size();
-
-    GLuint const id = glCreateShader(GL_FRAGMENT_SHADER);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    glShaderSource(id, 1, &src, &len);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    glCompileShader(id);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    GLint result = 0;
-    glGetShaderiv(id, GL_COMPILE_STATUS, &result);
-    EXPECT_EQ(result, GL_TRUE);
-
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-
-    glDeleteShader(id);
-    EXPECT_EQ(glGetError(), GL_NO_ERROR);
-}
+}// namespace
 
 int main(int argc, char** argv) {
-    ::testing::InitGoogleTest(&argc, argv);
-    return RUN_ALL_TESTS();
+    constexpr uint32_t START = 1000;
+    TestClass c;
+    auto& layout = c.layout;
+
+    uint64_t count = 0;
+    for (; count < 20; count++) {
+        insert(layout, count + START);
+    }
+
+    for (uint64_t t = 0; t < count; ++t) {
+        PipelineLayoutKey key = {};
+        key.descSetLayouts[0] = (VkDescriptorSetLayout)(START + t);
+
+        if (layout.find(key) != layout.end()) {
+            std::cout << "found " << t << std::endl;
+        } else {
+            std::cout << "not found" << t << std::endl;
+        }
+    }
+
+    return 0;
 }

filament/test/filament_framegraph_test.cpp

@@ -34,6 +34,10 @@ using namespace backend;
 
 class MockResourceAllocator : public ResourceAllocatorInterface {
     uint32_t handle = 0;
+    struct MockDisposer : public ResourceAllocatorDisposerInterface {
+        void destroy(backend::TextureHandle) noexcept override {}
+    } disposer;
+
 public:
     backend::RenderTargetHandle createRenderTarget(const char* name,
             backend::TargetBufferFlags targetBufferFlags,
@@ -60,6 +64,10 @@ public:
 
     void destroyTexture(backend::TextureHandle h) noexcept override {
     }
+
+    ResourceAllocatorDisposerInterface& getDisposer() noexcept override {
+        return disposer;
+    }
 };
 
 class FrameGraphTest : public testing::Test {

libs/camutils/include/camutils/Manipulator.h

@@ -109,6 +109,7 @@ public:
         vec4 groundPlane;
         RayCallback raycastCallback;
         void* raycastUserdata;
+        bool panning = true;
     };
 
     struct Builder {
@@ -143,6 +144,8 @@ public:
         Builder& groundPlane(FLOAT a, FLOAT b, FLOAT c, FLOAT d);  //! Plane equation used as a raycast fallback
         Builder& raycastCallback(RayCallback cb, void* userdata);  //! Raycast function for accurate grab-and-pan
 
+        Builder& panning(bool enabled);  //! Sets whether panning is enabled
+
         /**
          * Creates a new camera manipulator, either ORBIT, MAP, or FREE_FLIGHT.
          *

libs/camutils/src/Manipulator.cpp

@@ -144,6 +144,14 @@ Manipulator<FLOAT>::Builder& Manipulator<FLOAT>::Builder::raycastCallback(RayCal
     return *this;
 }
 
+
+
+template <typename FLOAT> typename
+Manipulator<FLOAT>::Builder& Manipulator<FLOAT>::Builder::panning(bool enabled) {
+    details.panning = enabled;
+    return *this;
+}
+
 template <typename FLOAT>
 Manipulator<FLOAT>* Manipulator<FLOAT>::Builder::build(Mode mode) {
     switch (mode) {

libs/camutils/src/OrbitManipulator.h

@@ -69,7 +69,7 @@ public:
     }
 
     void grabBegin(int x, int y, bool strafe) override {
-        mGrabState = strafe ? PANNING : ORBITING;
+        mGrabState = strafe && Base::mProps.panning ? PANNING : ORBITING;
         mGrabPivot = mPivot;
         mGrabEye = Base::mEye;
         mGrabTarget = Base::mTarget;

libs/filamat/include/filamat/MaterialBuilder.h

@@ -134,6 +134,7 @@ protected:
     TargetApi mTargetApi = (TargetApi) 0;
     Optimization mOptimization = Optimization::PERFORMANCE;
     bool mPrintShaders = false;
+    bool mSaveRawVariants = false;
     bool mGenerateDebugInfo = false;
     bool mIncludeEssl1 = true;
     utils::bitset32 mShaderModels;
@@ -586,6 +587,14 @@ public:
     //! If true, will output the generated GLSL shader code to stdout.
     MaterialBuilder& printShaders(bool printShaders) noexcept;
 
+    /**
+     * If true, this will write the raw generated GLSL for each variant to a text file in the
+     * current directory. The file will be named after the material name and the variant name. Its
+     * extension will be derived from the shader stage. For example, mymaterial_0x0e.frag,
+     * mymaterial_0x18.vert, etc.
+     */
+    MaterialBuilder& saveRawVariants(bool saveVariants) noexcept;
+
     //! If true, will include debugging information in generated SPIRV.
     MaterialBuilder& generateDebugInfo(bool generateDebugInfo) noexcept;
 

libs/filamat/src/MaterialBuilder.cpp

@@ -57,7 +57,7 @@
 #include <atomic>
 #include <utility>
 #include <vector>
-
+#include <fstream>
 
 namespace filamat {
 
@@ -547,6 +547,11 @@ MaterialBuilder& MaterialBuilder::printShaders(bool printShaders) noexcept {
     return *this;
 }
 
+MaterialBuilder& MaterialBuilder::saveRawVariants(bool saveRawVariants) noexcept {
+    mSaveRawVariants = saveRawVariants;
+    return *this;
+}
+
 MaterialBuilder& MaterialBuilder::generateDebugInfo(bool generateDebugInfo) noexcept {
     mGenerateDebugInfo = generateDebugInfo;
     return *this;
@@ -918,6 +923,30 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
                             shaderModel, targetApi, targetLanguage, featureLevel, info);
                 }
 
+                // Write the variant to a file.
+                if (mSaveRawVariants) {
+                    int variantKey = v.variant.key;
+                    auto getExtension = [](backend::ShaderStage stage) {
+                        switch (stage) {
+                            case backend::ShaderStage::VERTEX:
+                                return "vert";
+                            case backend::ShaderStage::FRAGMENT:
+                                return "frag";
+                            case backend::ShaderStage::COMPUTE:
+                                return "comp";
+                        }
+                    };
+                    char filename[256];
+                    snprintf(filename, sizeof(filename), "%s_0x%02x.%s", mMaterialName.c_str_safe(),
+                            variantKey, getExtension(v.stage));
+                    printf("Writing variant 0x%02x to %s\n", variantKey, filename);
+                    std::ofstream file(filename);
+                    if (file.is_open()) {
+                        file << shader;
+                        file.close();
+                    }
+                }
+
                 std::string* pGlsl = nullptr;
                 if (targetApiNeedsGlsl) {
                     pGlsl = &shader;

libs/filamentapp/include/filamentapp/FilamentApp.h

@@ -104,6 +104,8 @@ public:
 
     size_t getSkippedFrameCount() const { return mSkippedFrames; }
 
+    void loadIBL(std::string_view path);
+
     FilamentApp(const FilamentApp& rhs) = delete;
     FilamentApp(FilamentApp&& rhs) = delete;
     FilamentApp& operator=(const FilamentApp& rhs) = delete;

libs/filamentapp/src/FilamentApp.cpp

@@ -174,11 +174,6 @@ void FilamentApp::run(const Config& config, SetupCallback setupCallback,
 
     loadDirt(config);
     loadIBL(config);
-    if (mIBL != nullptr) {
-        mIBL->getSkybox()->setLayerMask(0x7, 0x4);
-        mScene->setSkybox(mIBL->getSkybox());
-        mScene->setIndirectLight(mIBL->getIndirectLight());
-    }
 
     for (auto& view : window->mViews) {
         if (view.get() != window->mUiView) {
@@ -519,31 +514,41 @@ const utils::Path& FilamentApp::getRootAssetsPath() {
     return root;
 }
 
-void FilamentApp::loadIBL(const Config& config) {
-    if (!config.iblDirectory.empty()) {
-        Path iblPath(config.iblDirectory);
+void FilamentApp::loadIBL(std::string_view path) {
+    Path iblPath(path);
+    if (!iblPath.exists()) {
+        std::cerr << "The specified IBL path does not exist: " << iblPath << std::endl;
+        return;
+    }
 
-        if (!iblPath.exists()) {
-            std::cerr << "The specified IBL path does not exist: " << iblPath << std::endl;
+    mIBL = std::make_unique<IBL>(*mEngine);
+
+    if (!iblPath.isDirectory()) {
+        if (!mIBL->loadFromEquirect(iblPath)) {
+            std::cerr << "Could not load the specified IBL: " << iblPath << std::endl;
+            mIBL.reset(nullptr);
             return;
         }
-
-        mIBL = std::make_unique<IBL>(*mEngine);
-
-        if (!iblPath.isDirectory()) {
-            if (!mIBL->loadFromEquirect(iblPath)) {
-                std::cerr << "Could not load the specified IBL: " << iblPath << std::endl;
-                mIBL.reset(nullptr);
-                return;
-            }
-        } else {
-            if (!mIBL->loadFromDirectory(iblPath)) {
-                std::cerr << "Could not load the specified IBL: " << iblPath << std::endl;
-                mIBL.reset(nullptr);
-                return;
-            }
+    } else {
+        if (!mIBL->loadFromDirectory(iblPath)) {
+            std::cerr << "Could not load the specified IBL: " << iblPath << std::endl;
+            mIBL.reset(nullptr);
+            return;
         }
     }
+
+    if (mIBL != nullptr) {
+        mIBL->getSkybox()->setLayerMask(0x7, 0x4);
+        mScene->setSkybox(mIBL->getSkybox());
+        mScene->setIndirectLight(mIBL->getIndirectLight());
+    }
+}
+
+void FilamentApp::loadIBL(const Config& config) {
+    if (config.iblDirectory.empty()) {
+        return;
+    }
+    loadIBL(config.iblDirectory);
 }
 
 void FilamentApp::loadDirt(const Config& config) {

libs/gltfio/src/AssetLoader.cpp

@@ -58,6 +58,8 @@
 
 #include "downcast.h"
 
+#include <codecvt>
+#include <locale>
 #include <memory>
 
 using namespace filament;
@@ -88,12 +90,57 @@ static constexpr cgltf_material kDefaultMat = {
     },
 };
 
-static const char* getNodeName(const cgltf_node* node, const char* defaultNodeName) {
-    if (node->name) return node->name;
-    if (node->mesh && node->mesh->name) return node->mesh->name;
-    if (node->light && node->light->name) return node->light->name;
-    if (node->camera && node->camera->name) return node->camera->name;
-    return defaultNodeName;
+static std::string getNodeName(cgltf_node const* node, char const* defaultNodeName) {
+    auto const getNameImpl = [node, defaultNodeName]() -> char const* {
+        if (node->name) return node->name;
+        if (node->mesh && node->mesh->name) return node->mesh->name;
+        if (node->light && node->light->name) return node->light->name;
+        if (node->camera && node->camera->name) return node->camera->name;
+        return defaultNodeName;
+    };
+
+    std::string strOrig(getNameImpl());
+
+    // We handle the potential case of escaped characters in the JSON which should be properly
+    // interpreted as unicode. See spec:
+    // https://registry.khronos.org/glTF/specs/2.0/glTF-2.0.html#json-encoding
+    // Also see spec for escaped strings in JSON (Section 2.5) https://www.ietf.org/rfc/rfc4627.txt
+
+    std::string strEscaped;
+    size_t cur = 0, idx = 0;
+    std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> conv;
+
+    auto const addUnencodedSubstr = [&](size_t cursor, size_t nextPoint) {
+        assert_invariant(nextPoint >= cursor);
+        if (cursor == nextPoint) {
+            return;
+        }
+        strEscaped += strOrig.substr(cursor, nextPoint - cursor);
+    };
+
+    while ((idx = strOrig.find("\\u", cur)) != std::string::npos) {
+        if (idx + 6 > strOrig.length()) {
+            utils::slog.w << "gltfio: Unable to interpret node name=" << strOrig
+                          << " as proper unicode encoding." << utils::io::endl;
+            return strOrig;
+        }
+
+        // Turns string of the form \u0062 to 0x0062
+        std::string const hexStr = strOrig.substr(idx + 2, 4);
+        if (hexStr.find_first_not_of("0123456789abcdefABCDEF") != std::string::npos) {
+            utils::slog.w << "gltfio: Unable to interpret node name=" << strOrig
+                          << " as proper unicode encoding." << utils::io::endl;
+            return strOrig;
+        }
+
+        addUnencodedSubstr(cur, idx);
+
+        strEscaped += conv.to_bytes((char32_t) std::stoul(hexStr, nullptr, 16));
+        cur = idx + 6;
+    }
+    addUnencodedSubstr(cur, strOrig.length());
+
+    return strEscaped;
 }
 
 static bool primitiveHasVertexColor(const cgltf_primitive& inPrim) {
@@ -504,7 +551,8 @@ FFilamentAsset* FAssetLoader::createRootAsset(const cgltf_data* srcAsset) {
 }
 
 void FAssetLoader::recursePrimitives(const cgltf_node* node, FFilamentAsset* fAsset) {
-    const char* name = getNodeName(node, mDefaultNodeName);
+    auto nameStr = getNodeName(node, mDefaultNodeName);
+    const char* name = nameStr.c_str();
     name = name ? name : "node";
 
     if (node->mesh) {
@@ -573,7 +621,8 @@ void FAssetLoader::recurseEntities(const cgltf_node* node, SceneMask scenes, Ent
     instance->mEntities.push_back(entity);
     instance->mNodeMap[node - srcAsset->nodes] = entity;
 
-    const char* name = getNodeName(node, mDefaultNodeName);
+    auto nameStr = getNodeName(node, mDefaultNodeName);
+    const char* name = nameStr.c_str();
 
     if (name) {
         fAsset->mNameToEntity[name].push_back(entity);

libs/gltfio/src/extended/AssetLoaderExtended.cpp

@@ -525,8 +525,9 @@ bool AssetLoaderExtended::createPrimitive(Input* input, Output* out,
                 slotCount++};
     }
 
-    if (!utility::loadCgltfBuffers(gltf, mGltfPath.c_str(), mUriDataCache)) {
-        return false;
+    if (!mCgltfBuffersLoaded) {
+        mCgltfBuffersLoaded = utility::loadCgltfBuffers(gltf, mGltfPath.c_str(), mUriDataCache);
+        if (!mCgltfBuffersLoaded) return false;
     }
 
     utility::decodeDracoMeshes(gltf, prim, input->dracoCache);

libs/gltfio/src/extended/AssetLoaderExtended.h

@@ -83,6 +83,7 @@ private:
     std::string mGltfPath;
     MaterialProvider& mMaterials;
     UriDataCacheHandle mUriDataCache;
+    bool mCgltfBuffersLoaded;
 };
 
 } // namespace filament::gltfio

libs/ibl/src/CubemapSH.cpp

@@ -22,6 +22,7 @@
 
 #include "CubemapUtilsImpl.h"
 
+#include <utils/Log.h>
 #include <utils/JobSystem.h>
 
 #include <math/mat4.h>

libs/utils/include/utils/JobSystem.h

@@ -21,7 +21,7 @@
 #include <utils/architecture.h>
 #include <utils/compiler.h>
 #include <utils/Condition.h>
-#include <utils/Log.h>
+#include <utils/debug.h>
 #include <utils/memalign.h>
 #include <utils/Mutex.h>
 #include <utils/Slice.h>
@@ -44,15 +44,23 @@
 namespace utils {
 
 class JobSystem {
-    static constexpr size_t MAX_JOB_COUNT = 16384;
+    static constexpr size_t MAX_JOB_COUNT = 1 << 14; // 16384
+    static constexpr uint32_t JOB_COUNT_MASK = MAX_JOB_COUNT - 1;
+    static constexpr uint32_t WAITER_COUNT_SHIFT = 24;
     static_assert(MAX_JOB_COUNT <= 0x7FFE, "MAX_JOB_COUNT must be <= 0x7FFE");
     using WorkQueue = WorkStealingDequeue<uint16_t, MAX_JOB_COUNT>;
+    using Mutex = utils::Mutex;
+    using Condition = utils::Condition;
 
 public:
     class Job;
 
+    using ThreadId = uint8_t;
+
     using JobFunc = void(*)(void*, JobSystem&, Job*);
 
+    static constexpr ThreadId invalidThreadId = 0xff;
+
     class alignas(CACHELINE_SIZE) Job {
     public:
         Job() noexcept {} /* = default; */ /* clang bug */ // NOLINT(modernize-use-equals-default,cppcoreguidelines-pro-type-member-init)
@@ -75,12 +83,18 @@ public:
         void* storage[JOB_STORAGE_SIZE_WORDS];                  // 48 | 48
         JobFunc function;                                       //  4 |  8
         uint16_t parent;                                        //  2 |  2
-        std::atomic<uint16_t> runningJobCount = { 1 };          //  2 |  2
-        mutable std::atomic<uint16_t> refCount = { 1 };         //  2 |  2
-                                                                //  6 |  2 (padding)
+        mutable ThreadId id = invalidThreadId;                  //  1 |  1
+        mutable std::atomic<uint8_t> refCount = { 1 };          //  1 |  1
+        std::atomic<uint32_t> runningJobCount = { 1 };          //  4 |  4
+                                                                //  4 |  0 (padding)
                                                                 // 64 | 64
     };
 
+#ifndef WIN32
+    // on windows std::function<void()> is bigger and forces the whole structure to be larger
+    static_assert(sizeof(Job) == 64);
+#endif
+
     explicit JobSystem(size_t threadCount = 0, size_t adoptableThreadsCount = 1) noexcept;
 
     ~JobSystem();
@@ -199,6 +213,21 @@ public:
         return job;
     }
 
+    // creates a job from a KNOWN method pointer w/ object passed by value
+    template<typename T, void(T::*method)(JobSystem&, Job*), typename ... ARGS>
+    Job* emplaceJob(Job* parent, ARGS&& ... args) noexcept {
+        static_assert(sizeof(T) <= sizeof(Job::storage), "user data too large");
+        Job* job = create(parent, [](void* storage, JobSystem& js, Job* job) {
+            T* const that = static_cast<T*>(storage);
+            (that->*method)(js, job);
+            that->~T();
+        });
+        if (job) {
+            new(job->storage) T(std::forward<ARGS>(args)...);
+        }
+        return job;
+    }
+
     // creates a job from a functor passed by value
     template<typename T>
     Job* createJob(Job* parent, T functor) noexcept {
@@ -214,6 +243,21 @@ public:
         return job;
     }
 
+    // creates a job from a functor passed by value
+    template<typename T, typename ... ARGS>
+    Job* emplaceJob(Job* parent, ARGS&& ... args) noexcept {
+        static_assert(sizeof(T) <= sizeof(Job::storage), "functor too large");
+        Job* job = create(parent, [](void* storage, JobSystem& js, Job* job){
+            T* const that = static_cast<T*>(storage);
+            that->operator()(js, job);
+            that->~T();
+        });
+        if (job) {
+            new(job->storage) T(std::forward<ARGS>(args)...);
+        }
+        return job;
+    }
+
 
     /*
      * Jobs are normally finished automatically, this can be used to cancel a job before it is run.
@@ -244,7 +288,7 @@ public:
     }
 
     /*
-     * Add job to this thread's execution queue. It's reference will drop automatically.
+     * Add job to this thread's execution queue. Its reference will drop automatically.
      * The current thread must be owned by JobSystem's thread pool. See adopt().
      *
      * The job can't be used after this call.
@@ -255,7 +299,19 @@ public:
         run(p);
     }
 
-    void signal() noexcept;
+    /*
+     * Add job to this thread's execution queue. Its reference will drop automatically.
+     * The current thread must be owned by JobSystem's thread pool. See adopt().
+     * id must be the current thread id obtained with JobSystem::getThreadId(Job*). This
+     * API is more efficient than the methods above.
+     *
+     * The job can't be used after this call.
+     */
+    void run(Job*& job, ThreadId id) noexcept;
+    void run(Job*&& job, ThreadId id) noexcept { // allows run(createJob(...));
+        Job* p = job;
+        run(p, id);
+    }
 
     /*
      * Add job to this thread's execution queue and keep a reference to it.
@@ -312,12 +368,29 @@ public:
 
     size_t getThreadCount() const { return mThreadCount; }
 
+    // returns the current ThreadId, which can be used with run(). This method can only be
+    // called from a job's function.
+    static ThreadId getThreadId(Job const* job) noexcept {
+        assert_invariant(job->id != invalidThreadId);
+        return job->id;
+    }
+
 private:
     // this is just to avoid using std::default_random_engine, since we're in a public header.
     class default_random_engine {
         static constexpr uint32_t m = 0x7fffffffu;
         uint32_t mState; // must be 0 < seed < 0x7fffffff
     public:
+        using result_type = uint32_t;
+
+        static constexpr result_type min() noexcept {
+            return 1;
+        }
+
+        static constexpr result_type max() noexcept {
+            return m - 1;
+        }
+
         inline constexpr explicit default_random_engine(uint32_t seed = 1u) noexcept
                 : mState(((seed % m) == 0u) ? 1u : seed % m) {
         }
@@ -362,13 +435,15 @@ private:
     Job* pop(WorkQueue& workQueue) noexcept;
     Job* steal(WorkQueue& workQueue) noexcept;
 
-    void wait(std::unique_lock<Mutex>& lock, Job* job = nullptr) noexcept;
-    void wake(size_t hint) noexcept;
+    [[nodiscard]]
+    uint32_t wait(std::unique_lock<Mutex>& lock, Job* job) noexcept;
+    void wait(std::unique_lock<Mutex>& lock) noexcept;
     void wakeAll() noexcept;
+    void wakeOne() noexcept;
 
     // these have thread contention, keep them together
-    utils::Mutex mWaiterLock;
-    utils::Condition mWaiterCondition;
+    Mutex mWaiterLock;
+    Condition mWaiterCondition;
 
     std::atomic<int32_t> mActiveJobs = { 0 };
     utils::Arena<utils::ThreadSafeObjectPoolAllocator<Job>, LockingPolicy::NoLock> mJobPool;
@@ -391,7 +466,7 @@ private:
     uint8_t mParallelSplitCount = 0;                    // # of split allowable in parallel_for
     Job* mRootJob = nullptr;
 
-    utils::Mutex mThreadMapLock; // this should have very little contention
+    Mutex mThreadMapLock; // this should have very little contention
     tsl::robin_map<std::thread::id, ThreadState *> mThreadMap;
 };
 
@@ -409,12 +484,13 @@ template<typename CALLABLE, typename ... ARGS>
 JobSystem::Job* createJob(JobSystem& js, JobSystem::Job* parent,
         CALLABLE&& func, ARGS&&... args) noexcept {
     struct Data {
+        explicit Data(std::function<void()> f) noexcept: f(std::move(f)) {}
         std::function<void()> f;
         // Renaming the method below could cause an Arrested Development.
         void gob(JobSystem&, JobSystem::Job*) noexcept { f(); }
-    } user{ std::bind(std::forward<CALLABLE>(func),
-            std::forward<ARGS>(args)...) };
-    return js.createJob<Data, &Data::gob>(parent, std::move(user));
+    };
+    return js.emplaceJob<Data, &Data::gob>(parent,
+            std::bind(std::forward<CALLABLE>(func), std::forward<ARGS>(args)...));
 }
 
 template<typename CALLABLE, typename T, typename ... ARGS,
@@ -425,12 +501,13 @@ template<typename CALLABLE, typename T, typename ... ARGS,
 JobSystem::Job* createJob(JobSystem& js, JobSystem::Job* parent,
         CALLABLE&& func, T&& o, ARGS&&... args) noexcept {
     struct Data {
+        explicit Data(std::function<void()> f) noexcept: f(std::move(f)) {}
         std::function<void()> f;
         // Renaming the method below could cause an Arrested Development.
         void gob(JobSystem&, JobSystem::Job*) noexcept { f(); }
-    } user{ std::bind(std::forward<CALLABLE>(func), std::forward<T>(o),
-            std::forward<ARGS>(args)...) };
-    return js.createJob<Data, &Data::gob>(parent, std::move(user));
+    };
+    return js.emplaceJob<Data, &Data::gob>(parent,
+            std::bind(std::forward<CALLABLE>(func), std::forward<T>(o), std::forward<ARGS>(args)...));
 }
 
 
@@ -459,8 +536,8 @@ struct ParallelForJobData {
 right_side:
         if (splitter.split(splits, count)) {
             const size_type lc = count / 2;
-            JobData ld(start, lc, splits + uint8_t(1), functor, splitter);
-            JobSystem::Job* l = js.createJob<JobData, &JobData::parallelWithJobs>(parent, std::move(ld));
+            JobSystem::Job* l = js.emplaceJob<JobData, &JobData::parallelWithJobs>(parent,
+                    start, lc, splits + uint8_t(1), functor, splitter);
             if (UTILS_UNLIKELY(l == nullptr)) {
                 // couldn't create a job, just pretend we're done splitting
                 goto execute;
@@ -468,7 +545,7 @@ right_side:
 
             // start the left side before attempting the right side, so we parallelize in case
             // of job creation failure -- rare, but still.
-            js.run(l);
+            js.run(l, JobSystem::getThreadId(parent));
 
             // don't spawn a job for the right side, just reuse us -- spawning jobs is more
             // costly than we'd like.
@@ -500,8 +577,8 @@ template<typename S, typename F>
 JobSystem::Job* parallel_for(JobSystem& js, JobSystem::Job* parent,
         uint32_t start, uint32_t count, F functor, const S& splitter) noexcept {
     using JobData = details::ParallelForJobData<S, F>;
-    JobData jobData(start, count, 0, std::move(functor), splitter);
-    return js.createJob<JobData, &JobData::parallelWithJobs>(parent, std::move(jobData));
+    return js.emplaceJob<JobData, &JobData::parallelWithJobs>(parent,
+            start, count, 0, std::move(functor), splitter);
 }
 
 // parallel jobs with pointer/count
@@ -512,8 +589,8 @@ JobSystem::Job* parallel_for(JobSystem& js, JobSystem::Job* parent,
         f(data + s, c);
     };
     using JobData = details::ParallelForJobData<S, decltype(user)>;
-    JobData jobData(0, count, 0, std::move(user), splitter);
-    return js.createJob<JobData, &JobData::parallelWithJobs>(parent, std::move(jobData));
+    return js.emplaceJob<JobData, &JobData::parallelWithJobs>(parent,
+            0, count, 0, std::move(user), splitter);
 }
 
 // parallel jobs on a Slice<>
@@ -524,7 +601,7 @@ JobSystem::Job* parallel_for(JobSystem& js, JobSystem::Job* parent,
 }
 
 
-template <size_t COUNT, size_t MAX_SPLITS = 12>
+template<size_t COUNT, size_t MAX_SPLITS = 12>
 class CountSplitter {
 public:
     bool split(size_t splits, size_t count) const noexcept {

libs/utils/include/utils/WorkStealingDequeue.h

@@ -35,7 +35,13 @@ namespace utils {
  *    steal()                      push(), pop()
  *  any thread                     main thread
  *
- *
+ * References:
+ * - This code is largely inspired from
+ *   https://blog.molecular-matters.com/2015/09/25/job-system-2-0-lock-free-work-stealing-part-3-going-lock-free/
+ * - other implementations
+ *   https://github.com/ConorWilliams/ConcurrentDeque/blob/main/include/riften/deque.hpp
+ *   https://github.com/ssbl/concurrent-deque/blob/master/include/deque.hpp
+ *   https://github.com/taskflow/work-stealing-queue/blob/master/wsq.hpp
  */
 template <typename TYPE, size_t COUNT>
 class WorkStealingDequeue {
@@ -117,7 +123,7 @@ TYPE WorkStealingDequeue<TYPE, COUNT>::pop() noexcept {
     index_t top = mTop.load(std::memory_order_seq_cst);
 
     if (top < bottom) {
-        // Queue isn't empty and it's not the last item, just return it, this is the common case.
+        // Queue isn't empty, and it's not the last item, just return it, this is the common case.
         return getItemAt(bottom);
     }
 
@@ -132,13 +138,13 @@ TYPE WorkStealingDequeue<TYPE, COUNT>::pop() noexcept {
         if (mTop.compare_exchange_strong(top, top + 1,
                 std::memory_order_seq_cst,
                 std::memory_order_relaxed)) {
-            // success: we stole our last item from ourself, meaning that a concurrent steal()
+            // Success: we stole our last item from ourselves, meaning that a concurrent steal()
             //          would have failed.
             // mTop now equals top + 1, we adjust top to make the queue empty.
             top++;
         } else {
-            // failure: mTop was not equal to top, which means the item was stolen under our feet.
-            // top now equals to mTop. Simply discard the item we just popped.
+            // Failure: mTop was not equal to top, which means the item was stolen under our feet.
+            // `top` now equals to mTop. Simply discard the item we just popped.
             // The queue is now empty.
             item = TYPE();
         }
@@ -149,7 +155,7 @@ TYPE WorkStealingDequeue<TYPE, COUNT>::pop() noexcept {
     }
 
     // std::memory_order_relaxed used because we're not publishing any data.
-    // no concurrent writes to mBottom possible, it's always safe to write mBottom.
+    // No concurrent writes to mBottom possible, it's always safe to write mBottom.
     mBottom.store(top, std::memory_order_relaxed);
     return item;
 }
@@ -194,6 +200,8 @@ TYPE WorkStealingDequeue<TYPE, COUNT>::steal() noexcept {
         }
         // failure: the item we just tried to steal was pop()'ed under our feet,
         // simply discard it; nothing to do -- it's okay to try again.
+        // However, item might be corrupted, so it must be trivially destructible
+        static_assert(std::is_trivially_destructible_v<TYPE>);
     }
 }
 

libs/utils/src/JobSystem.cpp

@@ -23,20 +23,18 @@
 // when SYSTRACE_TAG_JOBSYSTEM is used, enables even heavier systraces
 #define HEAVY_SYSTRACE  0
 
-// enable for catching hangs waiting on a job to finish
-static constexpr bool DEBUG_FINISH_HANGS = false;
-
 #include <utils/JobSystem.h>
 
 #include <utils/compiler.h>
+#include <utils/debug.h>
 #include <utils/Log.h>
-#include <utils/memalign.h>
 #include <utils/ostream.h>
 #include <utils/Panic.h>
 #include <utils/Systrace.h>
 
 #include <algorithm>
 #include <atomic>
+#include <cerrno>
 #include <chrono>
 #include <cmath>
 #include <condition_variable>
@@ -248,68 +246,62 @@ inline bool JobSystem::hasActiveJobs() const noexcept {
 }
 
 inline bool JobSystem::hasJobCompleted(JobSystem::Job const* job) noexcept {
-    return job->runningJobCount.load(std::memory_order_acquire) <= 0;
+    return (job->runningJobCount.load(std::memory_order_acquire) & JOB_COUNT_MASK) == 0;
 }
 
-void JobSystem::wait(std::unique_lock<Mutex>& lock, Job* job) noexcept {
-    if constexpr (!DEBUG_FINISH_HANGS) {
-        mWaiterCondition.wait(lock);
-    } else {
-        do {
-            // we use a pretty long timeout (4s) so we're very confident that the system is hung
-            // and nothing else is happening.
-            std::cv_status status = mWaiterCondition.wait_for(lock,
-                    std::chrono::milliseconds(4000));
-            if (status == std::cv_status::no_timeout) {
-                break;
-            }
-
-            // hang debugging...
-
-            // We check of we had active jobs or if the job we're waiting on had completed already.
-            // There is the possibility of a race condition, but our long timeout gives us some
-            // confidence that we're in an incorrect state.
-
-            size_t const id = std::distance(mThreadStates.data(), &getState());
-            auto activeJobs = mActiveJobs.load();
-
-            if (job) {
-                auto runningJobCount = job->runningJobCount.load();
-                FILAMENT_CHECK_POSTCONDITION(runningJobCount > 0)
-                        << "JobSystem(" << this << ", " << unsigned(id) << "): waiting while job "
-                        << job << " has completed and " << activeJobs << " jobs are active!";
-            }
-
-            FILAMENT_CHECK_POSTCONDITION(activeJobs <= 0)
-                    << "JobSystem(" << this << ", " << unsigned(id) << "): waiting while "
-                    << activeJobs << " jobs are active!";
-
-        } while (true);
-    }
-}
-
-void JobSystem::wakeAll() noexcept {
+inline void JobSystem::wait(std::unique_lock<Mutex>& lock) noexcept {
     HEAVY_SYSTRACE_CALL();
-    std::lock_guard<Mutex> const lock(mWaiterLock);
+    mWaiterCondition.wait(lock);
+}
+
+inline uint32_t JobSystem::wait(std::unique_lock<Mutex>& lock, Job* const job) noexcept {
+    HEAVY_SYSTRACE_CALL();
+    // signal we are waiting
+
+    if (hasActiveJobs() || exitRequested()) {
+        return job->runningJobCount.load(std::memory_order_acquire);
+    }
+
+    uint32_t runningJobCount =
+            job->runningJobCount.fetch_add(1 << WAITER_COUNT_SHIFT, std::memory_order_relaxed);
+
+    if (runningJobCount & JOB_COUNT_MASK) {
+        mWaiterCondition.wait(lock);
+    }
+
+    runningJobCount =
+            job->runningJobCount.fetch_sub(1 << WAITER_COUNT_SHIFT, std::memory_order_acquire);
+
+    assert_invariant((runningJobCount >> WAITER_COUNT_SHIFT) >= 1);
+
+    return runningJobCount;
+}
+
+UTILS_NOINLINE
+void JobSystem::wakeAll() noexcept {
+    // wakeAll() is called when a job finishes (to wake up any thread that might be waiting on it)
+    SYSTRACE_CALL();
+    mWaiterLock.lock();
     // this empty critical section is needed -- it guarantees that notify_all() happens
-    // after the condition's variables are set.
+    // either before the condition is checked, or after the condition variable sleeps.
+    mWaiterLock.unlock();
+    // notify_all() can be pretty slow, and it doesn't need to be inside the lock.
     mWaiterCondition.notify_all();
 }
 
-void JobSystem::wake(size_t hint) noexcept {
+void JobSystem::wakeOne() noexcept {
+    // wakeOne() is called when a new job is added to a queue
     HEAVY_SYSTRACE_CALL();
-    std::lock_guard<Mutex> const lock(mWaiterLock);
-    // this empty critical section is needed -- it guarantees that notify_all() happens
-    // after the condition's variables are set.
-    if (hint == 1) {
-        mWaiterCondition.notify_one();
-    } else {
-        mWaiterCondition.notify_all();
-    }
+    mWaiterLock.lock();
+    // this empty critical section is needed -- it guarantees that notify_one() happens
+    // either before the condition is checked, or after the condition variable sleeps.
+    mWaiterLock.unlock();
+    // notify_one() can be pretty slow, and it doesn't need to be inside the lock.
+    mWaiterCondition.notify_one();
 }
 
 inline JobSystem::ThreadState& JobSystem::getState() noexcept {
-    std::lock_guard<utils::Mutex> const lock(mThreadMapLock);
+    std::lock_guard<Mutex> const lock(mThreadMapLock);
     auto iter = mThreadMap.find(std::this_thread::get_id());
     FILAMENT_CHECK_PRECONDITION(iter != mThreadMap.end()) << "This thread has not been adopted.";
     return *iter->second;
@@ -324,58 +316,36 @@ void JobSystem::put(WorkQueue& workQueue, Job* job) noexcept {
     size_t const index = job - mJobStorageBase;
     assert(index >= 0 && index < MAX_JOB_COUNT);
 
-    // put the job into the queue first
+    // put the job into the queue
     workQueue.push(uint16_t(index + 1));
-    // then increase our active job count
-    int32_t const oldActiveJobs = mActiveJobs.fetch_add(1, std::memory_order_relaxed);
-    // But it's possible that the job has already been picked-up, so oldActiveJobs could be
-    // negative for instance. We signal only if that's not the case.
-    if (oldActiveJobs >= 0) {
-        wake(oldActiveJobs + 1); // wake-up a thread if needed...
-    }
+
+    // increase our active job count (the order in which we're doing this must not matter
+    // because we're not using std::memory_order_seq_cst (here or in WorkQueue::push()).
+    mActiveJobs.fetch_add(1, std::memory_order_relaxed);
+
+    // Note: it's absolutely possible for mActiveJobs to be 0 here, because the job could have
+    // been handled by a zealous worker already. In that case we could avoid calling wakeOne(),
+    // but that is not the common case.
+
+    wakeOne();
 }
 
 JobSystem::Job* JobSystem::pop(WorkQueue& workQueue) noexcept {
-    // decrement mActiveJobs first, this is to ensure that if there is only a single job left
-    // (and we're about to pick it up), other threads don't loop trying to do the same.
-    mActiveJobs.fetch_sub(1, std::memory_order_relaxed);
-
     size_t const index = workQueue.pop();
     assert(index <= MAX_JOB_COUNT);
     Job* const job = !index ? nullptr : &mJobStorageBase[index - 1];
-
-    // If our guess was wrong, i.e. we couldn't pick up a job (b/c our queue was empty), we
-    // need to correct mActiveJobs.
-    if (!job) {
-        int32_t const oldActiveJobs = mActiveJobs.fetch_add(1, std::memory_order_relaxed);
-        if (oldActiveJobs >= 0) {
-            // And if there are some active jobs, then we need to wake someone up. We know it
-            // can't be us, because we failed taking a job and we know another thread can't
-            // have added one in our queue.
-            wake(oldActiveJobs + 1); // wake-up a thread if needed...
-        }
+    if (UTILS_LIKELY(job)) {
+        mActiveJobs.fetch_sub(1, std::memory_order_relaxed);
     }
     return job;
 }
 
 JobSystem::Job* JobSystem::steal(WorkQueue& workQueue) noexcept {
-    // decrement mActiveJobs first, this is to ensure that if there is only a single job left
-    // (and we're about to pick it up), other threads don't loop trying to do the same.
-    mActiveJobs.fetch_sub(1, std::memory_order_relaxed);
-
     size_t const index = workQueue.steal();
-    assert(index <= MAX_JOB_COUNT);
+    assert_invariant(index <= MAX_JOB_COUNT);
     Job* const job = !index ? nullptr : &mJobStorageBase[index - 1];
-
-    // If we failed taking a job, we need to correct mActiveJobs.
-    if (!job) {
-        int32_t const oldActiveJobs = mActiveJobs.fetch_add(1, std::memory_order_relaxed);
-        if (oldActiveJobs >= 0) {
-            // And if there are some active jobs, then we need to wake someone up. We know it
-            // can't be us, because we failed taking a job and we know another thread can't
-            // have added one in our queue.
-            wake(oldActiveJobs + 1); // wake-up a thread if needed...
-        }
+    if (UTILS_LIKELY(job)) {
+        mActiveJobs.fetch_sub(1, std::memory_order_relaxed);
     }
     return job;
 }
@@ -407,7 +377,7 @@ JobSystem::Job* JobSystem::steal(JobSystem::ThreadState& state) noexcept {
     Job* job = nullptr;
     do {
         ThreadState* const stateToStealFrom = getStateToStealFrom(state);
-        if (UTILS_LIKELY(stateToStealFrom)) {
+        if (stateToStealFrom) {
             job = steal(stateToStealFrom->workQueue);
         }
         // nullptr -> nothing to steal in that queue either, if there are active jobs,
@@ -420,17 +390,23 @@ bool JobSystem::execute(JobSystem::ThreadState& state) noexcept {
     HEAVY_SYSTRACE_CALL();
 
     Job* job = pop(state.workQueue);
-    if (UTILS_UNLIKELY(job == nullptr)) {
+
+    // It is beneficial for some benchmarks to poll on steal() for a bit, because going back to
+    // sleep and waking up is pretty expensive. However, it is unclear it helps in practice with
+    // larger jobs or when parallel_for is used.
+    constexpr size_t const STEAL_TRY_COUNT = 1;
+    for (size_t i = 0; UTILS_UNLIKELY(!job && i < STEAL_TRY_COUNT); i++) {
         // our queue is empty, try to steal a job
         job = steal(state);
     }
 
-    if (job) {
-        assert(job->runningJobCount.load(std::memory_order_relaxed) >= 1);
-
+    if (UTILS_LIKELY(job)) {
+        assert((job->runningJobCount.load(std::memory_order_relaxed) & JOB_COUNT_MASK) >= 1);
         if (UTILS_LIKELY(job->function)) {
             HEAVY_SYSTRACE_NAME("job->function");
+            job->id = std::distance(mThreadStates.data(), &state);
             job->function(job->storage, *this, job);
+            job->id = invalidThreadId;
         }
         finish(job);
     }
@@ -442,9 +418,10 @@ void JobSystem::loop(ThreadState* state) noexcept {
     setThreadPriority(Priority::DISPLAY);
 
     // record our work queue
-    mThreadMapLock.lock();
+    std::unique_lock<Mutex> lock(mThreadMapLock);
     bool const inserted = mThreadMap.emplace(std::this_thread::get_id(), state).second;
-    mThreadMapLock.unlock();
+    lock.unlock();
+
     FILAMENT_CHECK_PRECONDITION(inserted) << "This thread is already in a loop.";
 
     // run our main loop...
@@ -469,11 +446,16 @@ void JobSystem::finish(Job* job) noexcept {
     do {
         // std::memory_order_release here is needed to synchronize with JobSystem::wait()
         // which needs to "see" all changes that happened before the job terminated.
-        auto runningJobCount = job->runningJobCount.fetch_sub(1, std::memory_order_acq_rel);
+        uint32_t const v = job->runningJobCount.fetch_sub(1, std::memory_order_acq_rel);
+        uint32_t const runningJobCount = v & JOB_COUNT_MASK;
         assert(runningJobCount > 0);
+
         if (runningJobCount == 1) {
             // no more work, destroy this job and notify its parent
-            notify = true;
+            uint32_t const waiters = v >> WAITER_COUNT_SHIFT;
+            if (waiters) {
+                notify = true;
+            }
             Job* const parent = job->parent == 0x7FFF ? nullptr : &storage[job->parent];
             decRef(job);
             job = parent;
@@ -484,7 +466,8 @@ void JobSystem::finish(Job* job) noexcept {
     } while (job);
 
     // wake-up all threads that could potentially be waiting on this job finishing
-    if (notify) {
+    if (UTILS_UNLIKELY(notify)) {
+        // but avoid calling notify_all() at all cost, because it's always expensive
         wakeAll();
     }
 }
@@ -502,10 +485,11 @@ JobSystem::Job* JobSystem::create(JobSystem::Job* parent, JobFunc func) noexcept
             // add a reference to the parent to make sure it can't be terminated.
             // memory_order_relaxed is safe because no action is taken at this point
             // (the job is not started yet).
-            auto parentJobCount = parent->runningJobCount.fetch_add(1, std::memory_order_relaxed);
+            UTILS_UNUSED_IN_RELEASE auto const parentJobCount =
+                    parent->runningJobCount.fetch_add(1, std::memory_order_relaxed);
 
             // can't create a child job of a terminated parent
-            assert(parentJobCount > 0);
+            assert((parentJobCount & JOB_COUNT_MASK) > 0);
 
             index = parent - mJobStorageBase;
             assert(index < MAX_JOB_COUNT);
@@ -532,10 +516,6 @@ void JobSystem::release(JobSystem::Job*& job) noexcept {
     job = nullptr;
 }
 
-void JobSystem::signal() noexcept {
-    wakeAll();
-}
-
 void JobSystem::run(Job*& job) noexcept {
     HEAVY_SYSTRACE_CALL();
 
@@ -547,6 +527,18 @@ void JobSystem::run(Job*& job) noexcept {
     job = nullptr;
 }
 
+void JobSystem::run(Job*& job, uint8_t id) noexcept {
+    HEAVY_SYSTRACE_CALL();
+
+    ThreadState& state = mThreadStates[id];
+    assert_invariant(&state == &getState());
+
+    put(state.workQueue, job);
+
+    // after run() returns, the job is virtually invalid (it'll die on its own)
+    job = nullptr;
+}
+
 JobSystem::Job* JobSystem::runAndRetain(Job* job) noexcept {
     JobSystem::Job* retained = retain(job);
     run(job);
@@ -561,7 +553,7 @@ void JobSystem::waitAndRelease(Job*& job) noexcept {
 
     ThreadState& state(getState());
     do {
-        if (!execute(state)) {
+        if (UTILS_UNLIKELY(!execute(state))) {
             // test if job has completed first, to possibly avoid taking the lock
             if (hasJobCompleted(job)) {
                 break;
@@ -577,11 +569,23 @@ void JobSystem::waitAndRelease(Job*& job) noexcept {
             // continue to handle more jobs, as they get added.
 
             std::unique_lock<Mutex> lock(mWaiterLock);
-            if (!hasJobCompleted(job) && !hasActiveJobs() && !exitRequested()) {
-                wait(lock, job);
+            uint32_t const runningJobCount = wait(lock, job);
+            // we could be waking up because either:
+            // - the job we're waiting on has completed
+            // - more jobs where added to the JobSystem
+            // - we're asked to exit
+            if ((runningJobCount & JOB_COUNT_MASK) == 0 || exitRequested()) {
+                break;
             }
+
+            // if we get here, it means that
+            // - the job we're waiting on is still running, and
+            // - we're not asked to exit, and
+            // - there were some active jobs
+            // So we try to handle one.
+            continue;
         }
-    } while (!hasJobCompleted(job) && !exitRequested());
+    } while (UTILS_LIKELY(!hasJobCompleted(job) && !exitRequested()));
 
     if (job == mRootJob) {
         mRootJob = nullptr;
@@ -591,6 +595,7 @@ void JobSystem::waitAndRelease(Job*& job) noexcept {
 }
 
 void JobSystem::runAndWait(JobSystem::Job*& job) noexcept {
+    SYSTRACE_CALL();
     runAndRetain(job);
     waitAndRelease(job);
 }
@@ -598,7 +603,7 @@ void JobSystem::runAndWait(JobSystem::Job*& job) noexcept {
 void JobSystem::adopt() {
     const auto tid = std::this_thread::get_id();
 
-    std::unique_lock<utils::Mutex> lock(mThreadMapLock);
+    std::unique_lock<Mutex> lock(mThreadMapLock);
     auto iter = mThreadMap.find(tid);
     ThreadState* const state = iter ==  mThreadMap.end() ? nullptr : iter->second;
     lock.unlock();
@@ -631,7 +636,7 @@ void JobSystem::adopt() {
 
 void JobSystem::emancipate() {
     const auto tid = std::this_thread::get_id();
-    std::lock_guard<utils::Mutex> const lock(mThreadMapLock);
+    std::unique_lock<Mutex> const lock(mThreadMapLock);
     auto iter = mThreadMap.find(tid);
     ThreadState* const state = iter ==  mThreadMap.end() ? nullptr : iter->second;
     FILAMENT_CHECK_PRECONDITION(state) << "this thread is not an adopted thread";

samples/gltf_viewer.cpp

@@ -58,6 +58,7 @@
 
 #include <cgltf.h>
 
+#include <algorithm>
 #include <array>
 #include <fstream>
 #include <iostream>
@@ -508,30 +509,55 @@ static void onClick(App& app, View* view, ImVec2 pos) {
     });
 }
 
-static utils::Path getPathForAsset(std::string_view string) {
+static utils::Path getPathForIBLAsset(std::string_view string) {
+    auto isIBL = [] (utils::Path file) -> bool {
+        return file.getExtension() == "ktx" || file.getExtension() == "hdr";
+    };
+
     utils::Path filename{ string };
     if (!filename.exists()) {
         std::cerr << "file " << filename << " not found!" << std::endl;
         return {};
     }
+
     if (filename.isDirectory()) {
-        auto files = filename.listContents();
-        for (const auto& file: files) {
-            if (file.getExtension() == "gltf" || file.getExtension() == "glb") {
-                filename = file;
-                break;
-            }
-        }
-        if (filename.isDirectory()) {
-            std::cerr << "no glTF file found in " << filename << std::endl;
+        std::vector<Path> files = filename.listContents();
+        if (std::none_of(files.cbegin(), files.cend(), isIBL)) {
             return {};
         }
+    } else if (!isIBL(filename)) {
+        return {};
     }
+
     return filename;
 }
 
+static utils::Path getPathForGLTFAsset(std::string_view string) {
+    auto isGLTF = [] (utils::Path file) -> bool {
+        return file.getExtension() == "gltf" || file.getExtension() == "glb";
+    };
 
-static bool checkAsset(const utils::Path& filename) {
+    utils::Path filename{ string };
+    if (!filename.exists()) {
+        std::cerr << "file " << filename << " not found!" << std::endl;
+        return {};
+    }
+
+    if (filename.isDirectory()) {
+        std::vector<Path> files = filename.listContents();
+        auto it = std::find_if(files.cbegin(), files.cend(), isGLTF);
+        if (it == files.end()) {
+            return {};
+        }
+        filename = *it;
+    } else if (!isGLTF(filename)) {
+        return {};
+    }
+
+    return filename;
+}
+
+static bool checkGLTFAsset(const utils::Path& filename) {
     // Peek at the file size to allow pre-allocation.
     long const contentSize = static_cast<long>(getFileSize(filename.c_str()));
     if (contentSize <= 0) {
@@ -571,8 +597,9 @@ int main(int argc, char** argv) {
     utils::Path filename;
     int const num_args = argc - optionIndex;
     if (num_args >= 1) {
-        filename = getPathForAsset(argv[optionIndex]);
+        filename = getPathForGLTFAsset(argv[optionIndex]);
         if (filename.isEmpty()) {
+            std::cerr << "no glTF file found in " << filename << std::endl;
             return 1;
         }
     }
@@ -1125,9 +1152,9 @@ int main(int argc, char** argv) {
     filamentApp.resize(resize);
 
     filamentApp.setDropHandler([&](std::string_view path) {
-        utils::Path const filename = getPathForAsset(path);
+        utils::Path filename = getPathForGLTFAsset(path);
         if (!filename.isEmpty()) {
-            if (checkAsset(filename)) {
+            if (checkGLTFAsset(filename)) {
                 app.resourceLoader->asyncCancelLoad();
                 app.resourceLoader->evictResourceData();
                 app.viewer->removeAsset();
@@ -1136,6 +1163,13 @@ int main(int argc, char** argv) {
                 loadResources(filename);
                 app.viewer->setAsset(app.asset, app.instance);
             }
+            return;
+        }
+
+        filename = getPathForIBLAsset(path);
+        if (!filename.isEmpty()) {
+            FilamentApp::get().loadIBL(path);
+            return;
         }
     });
 

tools/matc/src/matc/CommandlineConfig.cpp

@@ -103,6 +103,8 @@ static void usage(char* name) {
             "       Skip validation of number of sampler used\n\n"
             "   --print, -t\n"
             "       Print generated shaders for debugging\n\n"
+            "   --save-raw-variants, -R\n"
+            "       Write the raw generated GLSL for each variant to a text file in the current directory.\n\n"
     );
     const std::string from("MATC");
     for (size_t pos = usage.find(from); pos != std::string::npos; pos = usage.find(from, pos)) {
@@ -179,7 +181,7 @@ static void parseDefine(std::string defineString, Config::StringReplacementMap&
 }
 
 bool CommandlineConfig::parse() {
-    static constexpr const char* OPTSTR = "hLxo:f:dm:a:l:p:D:T:P:OSEr:vV:gtwF1";
+    static constexpr const char* OPTSTR = "hLxo:f:dm:a:l:p:D:T:P:OSEr:vV:gtwF1R";
     static const struct option OPTIONS[] = {
             { "help",                    no_argument, nullptr, 'h' },
             { "license",                 no_argument, nullptr, 'L' },
@@ -204,6 +206,7 @@ bool CommandlineConfig::parse() {
             { "version",                 no_argument, nullptr, 'v' },
             { "raw",                     no_argument, nullptr, 'w' },
             { "no-sampler-validation",   no_argument, nullptr, 'F' },
+            { "save-raw-variants",       no_argument, nullptr, 'R' },
             { nullptr, 0, nullptr, 0 }  // termination of the option list
     };
 
@@ -326,6 +329,8 @@ bool CommandlineConfig::parse() {
                 break;
             case 'F':
                 mNoSamplerValidation = true;
+            case 'R':
+                mSaveRawVariants = true;
                 break;
         }
     }

tools/matc/src/matc/Config.h

@@ -111,6 +111,10 @@ public:
         return mPrintShaders;
     }
 
+    bool saveRawVariants() const noexcept {
+        return mSaveRawVariants;
+    }
+
     bool rawShaderMode() const noexcept {
         return mRawShaderMode;
     }
@@ -149,6 +153,7 @@ protected:
     bool mPrintShaders = false;
     bool mRawShaderMode = false;
     bool mNoSamplerValidation = false;
+    bool mSaveRawVariants = false;
     Optimization mOptimizationLevel = Optimization::PERFORMANCE;
     Metadata mReflectionTarget = Metadata::NONE;
     Platform mPlatform = Platform::ALL;

tools/matc/src/matc/MaterialCompiler.cpp

@@ -24,6 +24,7 @@
 
 #include <filamat/Enums.h>
 
+#include <utils/Log.h>
 #include <utils/JobSystem.h>
 
 #include "DirIncluder.h"
@@ -408,6 +409,7 @@ bool MaterialCompiler::run(const Config& config) {
         .targetApi(config.getTargetApi())
         .optimization(config.getOptimizationLevel())
         .printShaders(config.printShaders())
+        .saveRawVariants(config.saveRawVariants())
         .generateDebugInfo(config.isDebug())
         .variantFilter(config.getVariantFilter() | builder.getVariantFilter());
 

web/filament-js/jsbindings.cpp

@@ -627,6 +627,7 @@ class_<Renderer>("Renderer")
     .function("getClearOptions", &Renderer::getClearOptions)
     .function("setPresentationTime", &Renderer::setPresentationTime)
     .function("setVsyncTime", &Renderer::setVsyncTime)
+    .function("skipFrame", &Renderer::skipFrame)
     .function("beginFrame", EMBIND_LAMBDA(bool, (Renderer* self, SwapChain* swapChain), {
         return self->beginFrame(swapChain);
     }), allow_raw_pointers())
@@ -683,7 +684,8 @@ class_<View>("View")
     .function("isStencilBufferEnabled", &View::isStencilBufferEnabled)
     .function("setMaterialGlobal", &View::setMaterialGlobal)
     .function("getMaterialGlobal", &View::getMaterialGlobal)
-    .function("getFogEntity", &View::getFogEntity);
+    .function("getFogEntity", &View::getFogEntity)
+    .function("clearFrameHistory", &View::clearFrameHistory);
 
 /// Scene ::core class:: Flat container of renderables and lights.
 /// See also the [Engine] methods `createScene` and `destroyScene`.