Merge branch 'rc/1.10.4' into release

BUILDING.md

@@ -376,7 +376,7 @@ same version that our continuous builds use.
 
 ```
 cd <your chosen parent folder for the emscripten SDK>
-curl -L https://github.com/emscripten-core/emsdk/archive/2.0.17.zip > emsdk.zip
+curl -L https://github.com/emscripten-core/emsdk/archive/2.0.23.zip > emsdk.zip
 unzip emsdk.zip ; mv emsdk-* emsdk ; cd emsdk
 python ./emsdk.py install latest
 python ./emsdk.py activate latest

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.10.3'
+    implementation 'com.google.android.filament:filament-android:1.10.4'
 }
 ```
 
@@ -63,7 +63,7 @@ A much smaller alternative to `filamat-android` that can only generate OpenGL sh
 iOS projects can use CocoaPods to install the latest release:
 
 ```
-pod 'Filament', '~> 1.10.3'
+pod 'Filament', '~> 1.10.4'
 ```
 
 ### Snapshots

RELEASE_NOTES.md

@@ -3,7 +3,14 @@
 This file contains one line summaries of commits that are worthy of mentioning in release notes.
 A new header is inserted each time a *tag* is created.
 
-## v1.10.4 (currently main branch)
+## v1.10.5 (currently main branch)
+
+## v1.10.4
+
+- engine: improvements to internal job system.
+- Vulkan: performance improvements on Mali.
+- gltfio: improvements to load time for large models.
+- WebGL: remove bogus stride argument, fix `BindingType` TypeScript definition.
 
 ## v1.10.3
 

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

@@ -102,10 +102,10 @@ JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void*) {
     jclass ktxloaderClass = env->FindClass("com/google/android/filament/utils/KTXLoader");
     if (ktxloaderClass == nullptr) return JNI_ERR;
     static const JNINativeMethod ktxMethods[] = {
-        {"nCreateKTXTexture", "(JLjava/nio/Buffer;IZ)J", reinterpret_cast<void*>(nCreateKTXTexture)},
-        {"nCreateIndirectLight", "(JLjava/nio/Buffer;IZ)J", reinterpret_cast<void*>(nCreateIndirectLight)},
-        {"nCreateSkybox", "(JLjava/nio/Buffer;IZ)J", reinterpret_cast<void*>(nCreateSkybox)},
-        {"nGetSphericalHarmonics", "(Ljava/nio/Buffer;I[F)Z", reinterpret_cast<void*>(nGetSphericalHarmonics)},
+        {(char*)"nCreateKTXTexture", (char*)"(JLjava/nio/Buffer;IZ)J", reinterpret_cast<void*>(nCreateKTXTexture)},
+        {(char*)"nCreateIndirectLight", (char*)"(JLjava/nio/Buffer;IZ)J", reinterpret_cast<void*>(nCreateIndirectLight)},
+        {(char*)"nCreateSkybox", (char*)"(JLjava/nio/Buffer;IZ)J", reinterpret_cast<void*>(nCreateSkybox)},
+        {(char*)"nGetSphericalHarmonics", (char*)"(Ljava/nio/Buffer;I[F)Z", reinterpret_cast<void*>(nGetSphericalHarmonics)},
     };
     rc = env->RegisterNatives(ktxloaderClass, ktxMethods, sizeof(ktxMethods) / sizeof(JNINativeMethod));
     if (rc != JNI_OK) return rc;
@@ -114,7 +114,7 @@ JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void*) {
     jclass hdrloaderClass = env->FindClass("com/google/android/filament/utils/HDRLoader");
     if (hdrloaderClass == nullptr) return JNI_ERR;
     static const JNINativeMethod hdrMethods[] = {
-        {"nCreateHDRTexture", "(JLjava/nio/Buffer;II)J", reinterpret_cast<void*>(nCreateHDRTexture)},
+        {(char*)"nCreateHDRTexture", (char*)"(JLjava/nio/Buffer;II)J", reinterpret_cast<void*>(nCreateHDRTexture)},
     };
     rc = env->RegisterNatives(hdrloaderClass, hdrMethods, sizeof(hdrMethods) / sizeof(JNINativeMethod));
     if (rc != JNI_OK) return rc;

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.10.3
+VERSION_NAME=1.10.4
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

android/samples/sample-gltf-viewer/src/main/java/com/google/android/filament/gltf/MainActivity.kt

@@ -24,6 +24,7 @@ import android.view.*
 import android.view.GestureDetector
 import android.widget.TextView
 import android.widget.Toast
+import com.google.android.filament.Fence
 import com.google.android.filament.IndirectLight
 import com.google.android.filament.Skybox
 import com.google.android.filament.utils.*
@@ -59,6 +60,8 @@ class MainActivity : Activity() {
     private var statusText: String? = null
     private var latestDownload: String? = null
     private val automation = AutomationEngine()
+    private var loadStartTime = 0L
+    private var loadStartFence: Fence? = null
 
     @SuppressLint("ClickableViewAccessibility")
     override fun onCreate(savedInstanceState: Bundle?) {
@@ -80,7 +83,7 @@ class MainActivity : Activity() {
             true
         }
 
-        createRenderables()
+        createDefaultRenderables()
         createIndirectLight()
 
         setStatusText("To load a new model, go to the above URL on your host machine.")
@@ -101,7 +104,7 @@ class MainActivity : Activity() {
         remoteServer = RemoteServer(8082)
     }
 
-    private fun createRenderables() {
+    private fun createDefaultRenderables() {
         val buffer = assets.open("models/scene.gltf").use { input ->
             val bytes = ByteArray(input.available())
             input.read(bytes)
@@ -155,6 +158,8 @@ class MainActivity : Activity() {
             modelViewer.destroyModel()
             modelViewer.loadModelGlb(message.buffer)
             modelViewer.transformToUnitCube()
+            loadStartTime = System.nanoTime()
+            loadStartFence = modelViewer.engine.createFence()
         }
     }
 
@@ -273,6 +278,8 @@ class MainActivity : Activity() {
                 }
             }
             modelViewer.transformToUnitCube()
+            loadStartTime = System.nanoTime()
+            loadStartFence = modelViewer.engine.createFence()
         }
     }
 
@@ -321,6 +328,16 @@ class MainActivity : Activity() {
         override fun doFrame(frameTimeNanos: Long) {
             choreographer.postFrameCallback(this)
 
+            loadStartFence?.let {
+                if (it.wait(Fence.Mode.FLUSH, 0) == Fence.FenceStatus.CONDITION_SATISFIED) {
+                    val end = System.nanoTime()
+                    val total = (end - loadStartTime) / 1_000_000
+                    Log.i(TAG, "The Filament backend took ${total} ms to load the model geometry.")
+                    modelViewer.engine.destroyFence(it)
+                    loadStartFence = null
+                }
+            }
+
             modelViewer.animator?.apply {
                 if (animationCount > 0) {
                     val elapsedTimeSeconds = (frameTimeNanos - startTime).toDouble() / 1_000_000_000
@@ -354,11 +371,11 @@ class MainActivity : Activity() {
         }
     }
 
-    // Just for testing purposes, this releases the model and reloads it.
+    // Just for testing purposes, this releases the current model and reloads the default model.
     inner class DoubleTapListener : GestureDetector.SimpleOnGestureListener() {
         override fun onDoubleTap(e: MotionEvent?): Boolean {
             modelViewer.destroyModel()
-            createRenderables()
+            createDefaultRenderables()
             return super.onDoubleTap(e)
         }
     }

build/web/ci-common.sh

@@ -9,7 +9,7 @@ export PATH="$PWD:$PATH"
 # npm install -g typescript
 
 # Install emscripten.
-curl -L https://github.com/emscripten-core/emsdk/archive/2.0.17.zip > emsdk.zip
+curl -L https://github.com/emscripten-core/emsdk/archive/2.0.23.zip > emsdk.zip
 unzip emsdk.zip ; mv emsdk-* emsdk ; cd emsdk
 ./emsdk install latest
 ./emsdk activate latest

filament/backend/include/private/backend/BackendUtils.h

@@ -18,6 +18,10 @@
 
 #include <stddef.h>
 
+#ifndef FILAMENT_MIN_COMMAND_BUFFERS_SIZE_IN_MB
+#    define FILAMENT_MIN_COMMAND_BUFFERS_SIZE_IN_MB 1
+#endif
+
 namespace filament {
 namespace backend {
 

filament/backend/src/CommandBufferQueue.cpp

@@ -21,6 +21,7 @@
 #include <utils/Panic.h>
 #include <utils/debug.h>
 
+#include "private/backend/BackendUtils.h"
 #include "private/backend/CommandStream.h"
 
 using namespace utils;
@@ -40,13 +41,13 @@ CommandBufferQueue::~CommandBufferQueue() {
 }
 
 void CommandBufferQueue::requestExit() {
-    std::unique_lock<utils::Mutex> lock(mLock);
+    std::lock_guard<utils::Mutex> lock(mLock);
     mExitRequested = EXIT_REQUESTED;
     mCondition.notify_one();
 }
 
 bool CommandBufferQueue::isExitRequested() const {
-    std::unique_lock<utils::Mutex> lock(mLock);
+    std::lock_guard<utils::Mutex> lock(mLock);
     ASSERT_PRECONDITION( mExitRequested == 0 || mExitRequested == EXIT_REQUESTED,
             "mExitRequested is corrupted (value = 0x%08x)!", mExitRequested);
     return (bool)mExitRequested;
@@ -99,14 +100,8 @@ void CommandBufferQueue::flush() noexcept {
     }
 #endif
 
-    if (UTILS_LIKELY(mFreeSpace >= requiredSize)) {
-        // ideally (and usually) we don't have to wait, this is the common case, so special case
-        // the unlock-before-notify, optimization.
-        lock.unlock();
-        mCondition.notify_one();
-    } else {
-        // unfortunately, there is not enough space left, we'll have to wait.
-        mCondition.notify_one(); // too bad there isn't a notify-and-wait
+    mCondition.notify_one();
+    if (UTILS_LIKELY(mFreeSpace < requiredSize)) {
         SYSTRACE_NAME("waiting: CircularBuffer::flush()");
         mCondition.wait(lock, [this, requiredSize]() -> bool {
             return mFreeSpace >= requiredSize;
@@ -130,9 +125,8 @@ std::vector<CommandBufferQueue::Slice> CommandBufferQueue::waitForCommands() con
 }
 
 void CommandBufferQueue::releaseBuffer(CommandBufferQueue::Slice const& buffer) {
-    std::unique_lock<utils::Mutex> lock(mLock);
+    std::lock_guard<utils::Mutex> lock(mLock);
     mFreeSpace += uintptr_t(buffer.end) - uintptr_t(buffer.begin);
-    lock.unlock();
     mCondition.notify_one();
 }
 

filament/backend/src/metal/MetalHandles.mm

@@ -887,7 +887,7 @@ static MTLPixelFormat decidePixelFormat(id<MTLDevice> device, TextureFormat form
         [event notifyListener:context.eventListener atValue:value block:^(id <MTLSharedEvent> o,
                 uint64_t value) {
             if (auto s = weakState.lock()) {
-                std::unique_lock<std::mutex> guard(s->mutex);
+                std::lock_guard<std::mutex> guard(s->mutex);
                 s->status = FenceStatus::CONDITION_SATISFIED;
                 s->cv.notify_all();
             }

filament/backend/src/opengl/TimerQuery.cpp

@@ -98,16 +98,15 @@ TimerQueryFence::~TimerQueryFence() {
     if (mThread.joinable()) {
         std::unique_lock<utils::Mutex> lock(mLock);
         mExitRequested = true;
-        lock.unlock();
         mCondition.notify_one();
+        lock.unlock();
         mThread.join();
     }
 }
 
 void TimerQueryFence::enqueue(TimerQueryFence::Job&& job) {
     std::unique_lock<utils::Mutex> lock(mLock);
     mQueue.push_back(std::forward<Job>(job));
-    lock.unlock();
     mCondition.notify_one();
 }
 

filament/backend/src/vulkan/VulkanBlitter.cpp

@@ -170,7 +170,7 @@ void VulkanBlitter::blitFast(VkImageAspectFlags aspect, VkFilter filter,
         transitionImageLayout(cmdbuffer, transitionHelper({
             .image = src.image,
             .oldLayout = VK_IMAGE_LAYOUT_UNDEFINED,
-            .newLayout = getTextureLayout(src.texture->usage),
+            .newLayout = mContext.getTextureLayout(src.texture->usage),
             .subresources = srcRange
         }));
     } else if (!mContext.currentSurface->headlessQueue) {
@@ -184,8 +184,9 @@ void VulkanBlitter::blitFast(VkImageAspectFlags aspect, VkFilter filter,
 
     // Determine the desired texture layout for the destination while ensuring that the default
     // render target is supported, which has no associated texture.
-    const VkImageLayout desiredLayout = dst.texture ? getTextureLayout(dst.texture->usage) :
-            getSwapChainAttachment(mContext).layout;
+    const VkImageLayout desiredLayout = dst.texture ?
+            mContext.getTextureLayout(dst.texture->usage) :
+            mContext.currentSurface->getColor().layout;
 
     transitionImageLayout(cmdbuffer, transitionHelper({
         .image = dst.image,

filament/backend/src/vulkan/VulkanBuffer.cpp

@@ -24,18 +24,15 @@ namespace filament {
 namespace backend {
 
 VulkanBuffer::VulkanBuffer(VulkanContext& context, VulkanStagePool& stagePool,
-        VulkanDisposer& disposer, VulkanDisposer::Key key, VkBufferUsageFlags usage,
-        uint32_t numBytes) : mContext(context), mStagePool(stagePool), mDisposer(disposer),
-        mDisposerKey(key) {
+        VkBufferUsageFlags usage, uint32_t numBytes) : mContext(context), mStagePool(stagePool) {
     // Create the VkBuffer.
     VkBufferCreateInfo bufferInfo {
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .size = numBytes,
         .usage = usage | VK_BUFFER_USAGE_TRANSFER_DST_BIT
     };
-    VmaAllocationCreateInfo allocInfo {
-        .usage = VMA_MEMORY_USAGE_GPU_ONLY
-    };
+
+    VmaAllocationCreateInfo allocInfo { .pool = context.vmaPoolGPU };
     vmaCreateBuffer(context.allocator, &bufferInfo, &allocInfo, &mGpuBuffer, &mGpuMemory, nullptr);
 }
 
@@ -56,7 +53,6 @@ void VulkanBuffer::loadFromCpu(const void* cpuData, uint32_t byteOffset, uint32_
 
     VkBufferCopy region { .size = numBytes };
     vkCmdCopyBuffer(cmdbuffer, stage->buffer, mGpuBuffer, 1, &region);
-    mDisposer.acquire(mDisposerKey);
 
     // Firstly, ensure that the copy finishes before the next draw call.
     // Secondly, in case the user decides to upload another chunk (without ever using the first one)

filament/backend/src/vulkan/VulkanBuffer.h

@@ -26,16 +26,14 @@ namespace backend {
 // Encapsulates a Vulkan buffer, its attached DeviceMemory and a staging area.
 class VulkanBuffer {
 public:
-    VulkanBuffer(VulkanContext& context, VulkanStagePool& stagePool, VulkanDisposer& disposer,
-            VulkanDisposer::Key mDisposerKey, VkBufferUsageFlags usage, uint32_t numBytes);
+    VulkanBuffer(VulkanContext& context, VulkanStagePool& stagePool, VkBufferUsageFlags usage,
+            uint32_t numBytes);
     ~VulkanBuffer();
     void loadFromCpu(const void* cpuData, uint32_t byteOffset, uint32_t numBytes);
     VkBuffer getGpuBuffer() const { return mGpuBuffer; }
 private:
     VulkanContext& mContext;
     VulkanStagePool& mStagePool;
-    VulkanDisposer& mDisposer;
-    VulkanDisposer::Key mDisposerKey;
     VmaAllocation mGpuMemory = VK_NULL_HANDLE;
     VkBuffer mGpuBuffer = VK_NULL_HANDLE;
 };

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -89,6 +89,11 @@ VulkanCommandBuffer const& VulkanCommands::get() {
     // It occurs only when Filament invokes commit() or endFrame() a large number of times without
     // presenting the swap chain or waiting on a fence.
     while (mAvailableCount == 0) {
+#if VK_REPORT_STALLS
+        slog.i  << "VulkanCommands has stalled. "
+                << "If this occurs frequently, consider increasing VK_MAX_COMMAND_BUFFERS."
+                << io::endl;
+#endif
         wait();
         gc();
     }
@@ -190,8 +195,8 @@ bool VulkanCommands::flush() {
     std::unique_lock<utils::Mutex> lock(cmdfence->mutex);
     cmdfence->status.store(VK_NOT_READY);
     UTILS_UNUSED_IN_RELEASE VkResult result = vkQueueSubmit(mQueue, 1, &submitInfo, cmdfence->fence);
-    lock.unlock();
     cmdfence->condition.notify_all();
+    lock.unlock();
 
     assert_invariant(result == VK_SUCCESS);
 

filament/backend/src/vulkan/VulkanConstants.h

@@ -37,6 +37,8 @@
 #define VK_ENABLE_VALIDATION 1
 #endif
 
+#define VK_REPORT_STALLS 0
+
 // All vkCreate* functions take an optional allocator. For now we select the default allocator by
 // passing in a null pointer, and we highlight the argument by using the VKALLOC constant.
 constexpr struct VkAllocationCallbacks* VKALLOC = nullptr;
@@ -45,12 +47,27 @@ constexpr struct VkAllocationCallbacks* VKALLOC = nullptr;
 constexpr static const int VK_REQUIRED_VERSION_MAJOR = 1;
 constexpr static const int VK_REQUIRED_VERSION_MINOR = 0;
 
-// We choose a capacity of 3 because this matches the needs of triple-buffering.
-constexpr static const int VK_MAX_COMMAND_BUFFERS = 3;
+// Controls the block size for the VkBuffer pools that we use for vertex buffers.
+constexpr static const uint64_t VMA_BUFFER_POOL_BLOCK_SIZE_IN_MB = 128;
+
+// Maximum number of VkCommandBuffer handles managed simultaneously by VulkanCommands.
+//
+// This includes the "current" command buffer that is being written into, as well as any command
+// buffers that have been submitted but have not yet finished rendering. Note that Filament can
+// issue multiple commit calls in a single frame, and that we use a triple buffered swap chain on
+// some platforms.
+constexpr static const int VK_MAX_COMMAND_BUFFERS = 10;
 
-// Maximum number of command buffer flush events that can occur before an unused pipeline is removed
-// from the cache. If this number is low, VkPipeline construction will occur frequently, which can
+// Number of command buffer submissions that should occur before an unused pipeline is removed
+// from the cache.
+//
+// If this number is low, VkPipeline construction will occur frequently, which can
 // be extremely slow. If this number is high, the memory footprint will be large.
-constexpr static const int VK_MAX_PIPELINE_AGE = 5;
+constexpr static const int VK_MAX_PIPELINE_AGE = 10;
+
+// VulkanPipelineCache does not track which command buffers contain references to which pipelines,
+// instead it simply waits for at least VK_MAX_COMMAND_BUFFERS submissions to occur before
+// destroying any unused pipeline object.
+static_assert(VK_MAX_PIPELINE_AGE >= VK_MAX_COMMAND_BUFFERS);
 
 #endif