metal: implement memory mapping (#9454)

FIXES=[462533574]

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

@@ -16,6 +16,9 @@
 
 package com.google.android.filament.utils
 
+import android.graphics.Bitmap
+import android.os.Handler
+import android.os.Looper
 import android.view.MotionEvent
 import android.view.Surface
 import android.view.SurfaceView
@@ -26,6 +29,7 @@ import com.google.android.filament.android.UiHelper
 import com.google.android.filament.gltfio.*
 import kotlinx.coroutines.*
 import java.nio.Buffer
+import java.nio.ByteBuffer
 
 private const val kNearPlane = 0.05f     // 5 cm
 private const val kFarPlane = 1000.0f    // 1 km
@@ -119,6 +123,8 @@ class ModelViewer(
     private val target = DoubleArray(3)
     private val upward = DoubleArray(3)
 
+    private var debugFrameCallback: ((Bitmap) -> Unit)? = null
+
     init {
         renderer = engine.createRenderer()
         scene = engine.createScene()
@@ -305,10 +311,39 @@ class ModelViewer(
         // Render the scene, unless the renderer wants to skip the frame.
         if (renderer.beginFrame(swapChain!!, frameTimeNanos)) {
             renderer.render(view)
+
+            debugFrameCallback?.let {
+                val viewport = view.viewport
+                val bitmap = Bitmap.createBitmap(viewport.width, viewport.height,
+                        Bitmap.Config.ARGB_8888)
+                val buffer = ByteBuffer.allocateDirect(viewport.width * viewport.height * 4)
+
+                val handler = Handler(Looper.getMainLooper())
+                val pixelBufferDescriptor = Texture.PixelBufferDescriptor(buffer,
+                        Texture.Format.RGBA, Texture.Type.UBYTE, 1, 0, 0, 0, handler) {
+                    buffer.rewind()
+                    bitmap.copyPixelsFromBuffer(buffer)
+                    it(bitmap)
+                }
+                renderer.readPixels(viewport.left, viewport.bottom, viewport.width,
+                                    viewport.height, pixelBufferDescriptor)
+                debugFrameCallback = null
+            }
+
             renderer.endFrame()
         }
     }
 
+    /*
+     * Sets a callback that will be invoked with the next rendered frame as a Bitmap. Note that this
+     * is a one-time callback.
+     *
+     *  @param callback   callback to be invoked with a rendered frame as [Bitmap]
+     */
+    fun debugGetNextFrameCallback(callback: (Bitmap) -> Unit) {
+        debugFrameCallback = callback
+    }
+
     private fun populateScene(asset: FilamentAsset) {
         val rcm = engine.renderableManager
         var count = 0

filament/backend/src/AndroidNativeWindow.cpp

@@ -61,6 +61,16 @@ int NativeWindow::enableFrameTimestamps(ANativeWindow* anw, bool enable) {
     return pWindow->perform(anw, ENABLE_FRAME_TIMESTAMPS, enable);
 }
 
+int NativeWindow::frameTimestampsSupportsPresent(ANativeWindow* anw, bool* outSupportsPresent) {
+    NativeWindow const* pWindow = reinterpret_cast<NativeWindow const*>(anw);
+    int value = 0;
+    bool const success = pWindow->perform(anw, FRAME_TIMESTAMPS_SUPPORTS_PRESENT, &value);
+    if (success) {
+        *outSupportsPresent = bool(value);
+    }
+    return success;
+}
+
 int NativeWindow::getCompositorTiming(ANativeWindow* anw,
         int64_t* compositeDeadline, int64_t* compositeInterval,
         int64_t* compositeToPresentLatency) {

filament/backend/src/AndroidNativeWindow.h

@@ -32,6 +32,7 @@ struct NativeWindow {
     // is valid query enum value
     enum {
         IS_VALID                = 17,
+        FRAME_TIMESTAMPS_SUPPORTS_PRESENT = 18,
         GET_NEXT_FRAME_ID       = 24,
         ENABLE_FRAME_TIMESTAMPS = 25,
         GET_COMPOSITOR_TIMING   = 26,
@@ -51,6 +52,7 @@ struct NativeWindow {
 
     static int getNextFrameId(ANativeWindow* anw, uint64_t* frameId);
     static int enableFrameTimestamps(ANativeWindow* anw, bool enable);
+    static int frameTimestampsSupportsPresent(ANativeWindow* anw, bool* outSupportsPresent);
     static int getCompositorTiming(ANativeWindow* anw,
             int64_t* compositeDeadline, int64_t* compositeInterval,
             int64_t* compositeToPresentLatency);

filament/backend/src/AndroidSwapChainHelper.cpp

@@ -19,6 +19,9 @@
 
 #include <android/native_window.h>
 
+#include <utils/compiler.h>
+#include <utils/Logger.h>
+
 #include <cstddef>
 #include <cstdint>
 #include <limits>
@@ -36,11 +39,23 @@ bool AndroidSwapChainHelper::setPresentFrameId(
     int const status = NativeWindow::getNextFrameId(anw, &sysFrameId);
     if (status == 0) {
         std::lock_guard const lock(mLock);
-        auto const pos = mFrameIdToSystemFrameId.find(frameId);
-        if (pos && *pos != sysFrameId) {
-            // we're trying to associate the same frame id to a different frame!
-            return false;
+        // frameIds must be strictly monotonic, if that's not the case (i.e. the new frameId is
+        // less or equal to the last one in the map), we have to clear the map, because the
+        // map's find() assume sorted keys.
+        // This case can happen if two different filament::Renderer are used with the same
+        // ANativeWindow (the Renderer would have different frameIds). This is expected to
+        // be a rare case.
+        if (UTILS_UNLIKELY(!mFrameIdToSystemFrameId.empty() &&
+                frameId <= mFrameIdToSystemFrameId.back().first)) {
+            // this log is expected to happen very rarely
+            DLOG(INFO) << "clearing frame history anw=" << anw
+                    << ", frameId=" << frameId
+                    << ", previous=" << mFrameIdToSystemFrameId.back().first
+                    << ", sysFrameId=" << sysFrameId;
+            // clear the frame history
+            mFrameIdToSystemFrameId.clear();
         }
+
         // oldest entry is removed
         mFrameIdToSystemFrameId.insert(frameId, sysFrameId);
         return true;

filament/backend/src/Platform.cpp

@@ -146,12 +146,12 @@ void Platform::setBlobFunc(InsertBlobFunc&& insertBlob, RetrieveBlobFunc&& retri
 
 bool Platform::hasInsertBlobFunc() const noexcept {
     std::lock_guard<decltype(mMutex)> lock(mMutex);
-    return bool(mInsertBlob);
+    return mInsertBlob && bool(*mInsertBlob);
 }
 
 bool Platform::hasRetrieveBlobFunc() const noexcept {
     std::lock_guard<decltype(mMutex)> lock(mMutex);
-    return bool(mRetrieveBlob);
+    return mRetrieveBlob && bool(*mRetrieveBlob);
 }
 
 void Platform::insertBlob(void const* key, size_t keySize, void const* value, size_t valueSize) {
@@ -184,7 +184,7 @@ void Platform::setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noe
 
 bool Platform::hasDebugUpdateStatFunc() const noexcept {
     std::lock_guard<decltype(mMutex)> lock(mMutex);
-    return mDebugUpdateStat != nullptr;
+    return mDebugUpdateStat && bool(*mDebugUpdateStat);
 }
 
 void Platform::debugUpdateStat(const char* key, uint64_t intValue) {

filament/backend/src/metal/MetalBuffer.h

@@ -165,7 +165,7 @@ public:
          size_t size, bool forceGpuBuffer = false);
     ~MetalBuffer();
 
-    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer || mCpuBuffer; }
+    [[nodiscard]] bool wasAllocationSuccessful() const noexcept { return mBuffer; }
 
     MetalBuffer(const MetalBuffer& rhs) = delete;
     MetalBuffer& operator=(const MetalBuffer& rhs) = delete;
@@ -185,14 +185,12 @@ public:
      * Denotes that this buffer is used for a draw call ensuring that its allocation remains valid
      * until the end of the current frame.
      *
-     * @return The MTLBuffer representing the current state of the buffer to bind, or nil if there
-     * is no device allocation.
+     * @return The MTLBuffer representing the current state of the buffer to bind, it never returns
+     * nil.
      *
      */
     id<MTLBuffer> getGpuBufferForDraw() noexcept;
 
-    void* getCpuBuffer() const noexcept { return mCpuBuffer; }
-
     void setLabel(const utils::ImmutableCString& label) {
 #if FILAMENT_METAL_DEBUG_LABELS
         if (label.empty()) {
@@ -235,7 +233,6 @@ private:
     UploadStrategy mUploadStrategy;
     TrackedMetalBuffer mBuffer;
     size_t mBufferSize = 0;
-    void* mCpuBuffer = nullptr;
     MetalContext& mContext;
 };
 

filament/backend/src/metal/MetalBuffer.mm

@@ -39,34 +39,30 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
         mUploadStrategy = UploadStrategy::POOL;
     }
 
-    // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
-    // buffer. Instead, we use immediate command encoder methods like setVertexBytes:length:atIndex:.
-    // This won't work for SSBOs, since they are read/write.
+    MTLResourceOptions options = MTLResourceStorageModePrivate;
 
-    /*
-    if (size <= 4 * 1024 && bindingType != BufferObjectBinding::SHADER_STORAGE &&
-            usage == BufferUsage::DYNAMIC && !forceGpuBuffer) {
-        mBuffer = nil;
-        mCpuBuffer = malloc(size);
-        return;
+    // The buffer will be memory mapped for write operations.
+    if (any(usage & BufferUsage::SHARED_WRITE_BIT)) {
+#if defined(FILAMENT_IOS) || defined(__arm64__) || defined(__aarch64__)
+        // iOS and Apple Silicon devices use UMA (Unified Memory Architecture), so we use Shared memory.
+        options = MTLResourceStorageModeShared;
+#else
+        // Intel Macs require Managed memory for CPU/GPU synchronization.
+        options = MTLResourceStorageModeManaged;
+#endif
     }
-    */
 
-    // Otherwise, we allocate a private GPU buffer.
     {
         ScopedAllocationTimer timer("generic");
-        mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
+        mBuffer = { [context.device newBufferWithLength:size options:options],
             TrackedMetalBuffer::Type::GENERIC };
     }
+
     // mBuffer might fail to be allocated. Clients can check for this by calling
     // wasAllocationSuccessful().
 }
 
-MetalBuffer::~MetalBuffer() {
-    if (mCpuBuffer) {
-        free(mCpuBuffer);
-    }
-}
+MetalBuffer::~MetalBuffer() = default;
 
 void MetalBuffer::copyIntoBuffer(
         void* src, size_t size, size_t byteOffset, TagResolver&& getHandleTag) {
@@ -83,12 +79,6 @@ void MetalBuffer::copyIntoBuffer(
     FILAMENT_CHECK_PRECONDITION(!(byteOffset & 0x3))
             << "byteOffset must be a multiple of 4, tag=" << getHandleTag();
 
-    // If we have a cpu buffer, we can directly copy into it.
-    if (mCpuBuffer) {
-        memcpy(static_cast<uint8_t*>(mCpuBuffer) + byteOffset, src, size);
-        return;
-    }
-
     switch (mUploadStrategy) {
         case UploadStrategy::BUMP_ALLOCATOR:
             uploadWithBumpAllocator(src, size, byteOffset, std::move(getHandleTag));
@@ -106,11 +96,6 @@ void MetalBuffer::copyIntoBufferUnsynchronized(
 }
 
 id<MTLBuffer> MetalBuffer::getGpuBufferForDraw() noexcept {
-    // If there's a CPU buffer, then we return nil here, as the CPU-side buffer will be bound
-    // separately.
-    if (mCpuBuffer) {
-        return nil;
-    }
     assert_invariant(mBuffer);
     return mBuffer.get();
 }
@@ -171,41 +156,6 @@ void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncod
                                                    offsets:metalOffsets.data()
                                                  withRange:bufferRange];
     }
-
-    for (size_t b = 0; b < count; b++) {
-        MetalBuffer* const buffer = buffers[b];
-        if (!buffer) {
-            continue;
-        }
-
-        const void* cpuBuffer = buffer->getCpuBuffer();
-        if (!cpuBuffer) {
-            continue;
-        }
-
-        const size_t bufferIndex = bufferStart + b;
-        const size_t offset = offsets[b];
-        auto* bytes = static_cast<const uint8_t*>(cpuBuffer);
-
-        if (stages & Stage::VERTEX) {
-            [(id<MTLRenderCommandEncoder>) encoder setVertexBytes:(bytes + offset)
-                                                           length:(buffer->getSize() - offset)
-                                                          atIndex:bufferIndex];
-        }
-        if (stages & Stage::FRAGMENT) {
-            [(id<MTLRenderCommandEncoder>) encoder setFragmentBytes:(bytes + offset)
-                                                             length:(buffer->getSize() - offset)
-                                                            atIndex:bufferIndex];
-        }
-        if (stages & Stage::COMPUTE) {
-            // TODO: using setBytes means the data is read-only, which currently isn't enforced.
-            // In practice this won't be an issue since MetalBuffer ensures all SSBOs are realized
-            // through actual id<MTLBuffer> allocations.
-            [(id<MTLComputeCommandEncoder>) encoder setBytes:(bytes + offset)
-                                                      length:(buffer->getSize() - offset)
-                                                     atIndex:bufferIndex];
-        }
-    }
 }
 
 void MetalBuffer::uploadWithPoolBuffer(

filament/backend/src/metal/MetalDriver.h

@@ -61,6 +61,8 @@ public:
 
     MetalContext* getContext() { return mContext; }
 
+    using DriverBase::scheduleDestroy;
+
 private:
 
     friend class MetalSwapChain;

filament/backend/src/metal/MetalDriver.mm

@@ -2275,7 +2275,10 @@ MemoryMappedBufferHandle MetalDriver::mapBufferS() noexcept {
 void MetalDriver::mapBufferR(MemoryMappedBufferHandle mmbh,
         BufferObjectHandle boh, size_t offset,
         size_t size, MapBufferAccessFlags access, utils::ImmutableCString&& tag) {
-    construct_handle<MetalMemoryMappedBuffer>(mmbh, boh, offset, size, access);
+    assert_invariant(boh);
+    MetalBufferObject* bo = mHandleAllocator.handle_cast<MetalBufferObject*>(boh);
+    assert_invariant(bo);
+    construct_handle<MetalMemoryMappedBuffer>(mmbh, bo, offset, size, access);
     mHandleAllocator.associateTagToHandle(mmbh.getId(), std::move(tag));
 }
 
@@ -2283,21 +2286,16 @@ void MetalDriver::unmapBuffer(MemoryMappedBufferHandle mmbh) {
     if (UTILS_UNLIKELY(!mmbh)) {
         return;
     }
+
+    auto* mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
+    mmb->unmap();
     destruct_handle<MetalMemoryMappedBuffer>(mmbh);
 }
 
 void MetalDriver::copyToMemoryMappedBuffer(MemoryMappedBufferHandle mmbh, size_t offset,
         BufferDescriptor&& data) {
-    auto mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
-
-    assert_invariant(any(mmb->access & MapBufferAccessFlags::WRITE_BIT));
-    assert_invariant(offset + data.size <= mmb->size);
-
-    // TODO: this isa zero-effort implementation of copyToMemoryMappedBuffer(), where we just
-    //       call updateBufferObject(). This could be a fallback implementation for when
-    //       shared memory is not available.
-    //       On UMA systems, this should just be a memcpy into the memory-mapped buffer.
-    updateBufferObject(mmb->boh, std::move(data), mmb->offset + offset);
+    auto* mmb = handle_cast<MetalMemoryMappedBuffer>(mmbh);
+    mmb->copy(*this, offset, std::move(data));
 }
 
 // explicit instantiation of the Dispatcher

filament/backend/src/metal/MetalHandles.h

@@ -569,14 +569,22 @@ struct MetalDescriptorSet : public HwDescriptorSet {
 
 
 struct MetalMemoryMappedBuffer : public HwMemoryMappedBuffer {
-    MetalMemoryMappedBuffer(BufferObjectHandle boh, size_t const offset,
-            size_t const size, MapBufferAccessFlags const access)
-        : boh(boh), access(access), size(size), offset(offset) {
-    }
-    BufferObjectHandle boh{};
     MapBufferAccessFlags access{};
-    uint32_t size = 0;
-    uint32_t offset = 0;
+    struct {
+        MetalBufferObject* bo;
+        void* vaddr = nullptr;
+        uint32_t size = 0;
+        uint32_t offset = 0;
+    } mtl;
+
+    MetalMemoryMappedBuffer(MetalBufferObject* bo, size_t offset, size_t size,
+            MapBufferAccessFlags access) noexcept;
+
+    ~MetalMemoryMappedBuffer();
+
+    void unmap();
+
+    void copy(MetalDriver& mtld, size_t offset, BufferDescriptor&& data) const;
 };
 
 } // namespace backend

filament/backend/src/metal/MetalHandles.mm

@@ -1664,5 +1664,44 @@ id<MTLBuffer> MetalDescriptorSet::finalizeAndGetBuffer(MetalDriver* driver, Shad
     return buffer.get();
 }
 
+MetalMemoryMappedBuffer::MetalMemoryMappedBuffer(MetalBufferObject* bo, size_t offset, size_t size,
+    MapBufferAccessFlags access) noexcept : access(access) {
+    MetalBuffer* buffer = bo->getBuffer();
+    assert_invariant(buffer);
+    id<MTLBuffer> mtlBuffer = buffer->getGpuBufferForDraw();
+
+    assert_invariant(offset + size <= bo->byteCount);
+    assert_invariant(mtlBuffer.storageMode != MTLStorageModePrivate);
+
+    mtl.bo = bo;
+    mtl.vaddr = static_cast<char*>(mtlBuffer.contents) + offset;
+    mtl.size = size;
+    mtl.offset = offset;
+}
+
+MetalMemoryMappedBuffer::~MetalMemoryMappedBuffer() = default;
+
+void MetalMemoryMappedBuffer::unmap() {
+#if !defined(FILAMENT_IOS) && defined(__x86_64__)
+    // Managed memory requires didModifyRange to synchronize changes to the GPU. This is specific to Intel Macs.
+    MetalBuffer* buffer = bo->getBuffer();
+    id<MTLBuffer> mtlBuffer = buffer->getGpuBufferForDraw();
+    if (mtlBuffer && mtlBuffer.storageMode == MTLStorageModeManaged) {
+        [mtlBuffer didModifyRange:NSMakeRange(mtl.offset, mtl.size)];
+    }
+#endif
+    // Shared memory on UMA systems is coherent; no explicit synchronization is required.
+}
+
+void MetalMemoryMappedBuffer::copy(MetalDriver& mtld, size_t offset, BufferDescriptor&& data) const {
+    assert_invariant(any(access & MapBufferAccessFlags::WRITE_BIT));
+    assert_invariant(offset + data.size <= mtl.size);
+    assert_invariant(mtl.vaddr);
+
+    memcpy(static_cast<char*>(mtl.vaddr) + offset, data.buffer, data.size);
+
+    mtld.scheduleDestroy(std::move(data));
+}
+
 } // namespace backend
 } // namespace filament

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

@@ -106,6 +106,8 @@ struct PlatformEGLAndroid::SwapChainEGLAndroid : public SwapChainEGL {
     void terminate(PlatformEGLAndroid& platform);
     bool setPresentFrameId(uint64_t frameId) const noexcept;
     uint64_t getFrameId(uint64_t frameId) const noexcept;
+    bool compositorTimingSupported = false;
+    bool frameTimestampsSupported = false;
 private:
     AndroidSwapChainHelper mImpl{};
 };
@@ -228,9 +230,9 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
                     "eglGetNativeClientBufferANDROID"));
 
     if (ext.egl.ANDROID_presentation_time) {
-        eglGetNativeClientBufferANDROID =
-                PFNEGLGETNATIVECLIENTBUFFERANDROIDPROC(eglGetProcAddress(
-                        "eglGetNativeClientBufferANDROID"));
+        eglPresentationTimeANDROID =
+                PFNEGLPRESENTATIONTIMEANDROIDPROC(eglGetProcAddress(
+                        "eglPresentationTimeANDROID"));
     }
 
     if (ext.egl.ANDROID_get_frame_timestamps) {
@@ -289,11 +291,21 @@ bool PlatformEGLAndroid::queryCompositorTiming(SwapChain const* swapchain,
     outCompositorTiming->frameTime = preferredTimeline.frameTime;
     outCompositorTiming->expectedPresentTime = preferredTimeline.expectedPresentTime;
     outCompositorTiming->frameTimelineDeadline = preferredTimeline.frameTimelineDeadline;
+    outCompositorTiming->compositeDeadline = CompositorTiming::INVALID;
+    outCompositorTiming->compositeInterval = CompositorTiming::INVALID;
+    outCompositorTiming->compositeToPresentLatency = CompositorTiming::INVALID;
+
+    // From this point on, we always return "success" because some timings were returned.
+
+    if (!static_cast<SwapChainEGLAndroid const *>(swapchain)->compositorTimingSupported) {
+        // if this surface doesn't support it, don't attempt to query the values.
+        return true;
+    }
 
     if (UTILS_LIKELY(ext.egl.ANDROID_get_frame_timestamps)) {
         EGLSurface const sur = static_cast<SwapChainEGL const *>(swapchain)->sur;
         if (sur == EGL_NO_SURFACE) {
-            return false;
+            return true;
         }
 
         std::array<EGLnsecsANDROID, 3> values;
@@ -304,26 +316,16 @@ bool PlatformEGLAndroid::queryCompositorTiming(SwapChain const* swapchain,
         };
         EGLBoolean const success = eglGetCompositorTimingANDROID(getEglDisplay(), sur,
                 names.size(), names.data(), values.data());
-        if (!success) {
-            return false;
+        if (UTILS_UNLIKELY(!success)) {
+            // reset current error to EGL_SUCCESS
+            eglGetError();
+        } else {
+            outCompositorTiming->compositeDeadline = values[0];
+            outCompositorTiming->compositeInterval = values[1];
+            outCompositorTiming->compositeToPresentLatency = values[2];
         }
-        outCompositorTiming->compositeDeadline = values[0];
-        outCompositorTiming->compositeInterval = values[1];
-        outCompositorTiming->compositeToPresentLatency = values[2];
-        return true;
     }
-
-    // fallback to private APIs
-    auto const anw = static_cast<SwapChainEGL const *>(swapchain)->nativeWindow;
-    int const status = NativeWindow::getCompositorTiming(anw,
-            &outCompositorTiming->compositeDeadline,
-            &outCompositorTiming->compositeInterval,
-            &outCompositorTiming->compositeToPresentLatency);
-    if (status == 0) {
-        return true;
-    }
-
-    return PlatformEGL::queryCompositorTiming(swapchain, outCompositorTiming);
+    return true;
 }
 
 bool PlatformEGLAndroid::setPresentFrameId(SwapChain const* swapchain,
@@ -348,6 +350,10 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
         return false;
     }
 
+    if (!static_cast<SwapChainEGLAndroid const *>(swapchain)->frameTimestampsSupported) {
+        return false;
+    }
+
     if (UTILS_LIKELY(ext.egl.ANDROID_get_frame_timestamps)) {
         EGLSurface const sur = sc->sur;
         if (sur == EGL_NO_SURFACE) {
@@ -368,7 +374,9 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
         };
         EGLBoolean const success = eglGetFrameTimestampsANDROID(getEglDisplay(), sur, hwFrameId,
                 names.size(), names.data(), values.data());
-        if (!success) {
+        if (UTILS_UNLIKELY(!success)) {
+            // reset current error to EGL_SUCCESS
+            eglGetError();
             return false;
         }
         outFrameTimestamps->requestedPresentTime = values[0];
@@ -382,28 +390,44 @@ bool PlatformEGLAndroid::queryFrameTimestamps(SwapChain const* swapchain, uint64
         outFrameTimestamps->releaseTime = values[8];
         return true;
     }
-
-    // fallback to private APIs
-    auto const anw = sc->nativeWindow;
-    int const status = NativeWindow::getFrameTimestamps(anw, hwFrameId,
-            &outFrameTimestamps->requestedPresentTime,
-            &outFrameTimestamps->acquireTime,
-            &outFrameTimestamps->latchTime,
-            &outFrameTimestamps->firstCompositionStartTime,
-            &outFrameTimestamps->lastCompositionStartTime,
-            &outFrameTimestamps->gpuCompositionDoneTime,
-            &outFrameTimestamps->displayPresentTime,
-            &outFrameTimestamps->dequeueReadyTime,
-            &outFrameTimestamps->releaseTime);
-    if (status == 0) {
-        return true;
-    }
-
     return PlatformEGL::queryFrameTimestamps(swapchain, frameId, outFrameTimestamps);
 }
 
 Platform::SwapChain* PlatformEGLAndroid::createSwapChain(void* nativeWindow, uint64_t const flags) {
     auto* const sc = new(std::nothrow) SwapChainEGLAndroid(*this, nativeWindow, flags);
+    if (UTILS_LIKELY(ext.egl.ANDROID_get_frame_timestamps)) {
+        EGLDisplay const dpy = getEglDisplay();
+        sc->compositorTimingSupported =
+                eglGetCompositorTimingSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITE_DEADLINE_ANDROID) &&
+                eglGetCompositorTimingSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITE_INTERVAL_ANDROID) &&
+                eglGetCompositorTimingSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITE_TO_PRESENT_LATENCY_ANDROID);
+        sc->frameTimestampsSupported =
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_REQUESTED_PRESENT_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_RENDERING_COMPLETE_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_COMPOSITION_LATCH_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_FIRST_COMPOSITION_START_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_LAST_COMPOSITION_START_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_FIRST_COMPOSITION_GPU_FINISHED_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_DISPLAY_PRESENT_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_DEQUEUE_READY_TIME_ANDROID) &&
+                eglGetFrameTimestampSupportedANDROID(dpy, sc->sur,
+                        EGL_READS_DONE_TIME_ANDROID);
+    }
+    // This is expected to be a low frequency log, only turned on in debug builds
+    DLOG(INFO) << "anw: " << nativeWindow
+            << ", compositorTimingSupported=" << sc->compositorTimingSupported
+            << ", frameTimestampsSupported=" << sc->frameTimestampsSupported;
     return sc;
 }
 
@@ -725,8 +749,6 @@ PlatformEGLAndroid::SwapChainEGLAndroid::SwapChainEGLAndroid(PlatformEGLAndroid
             // we ignore the result, it doesn't matter much if it fails
             eglSurfaceAttrib(platform.getEglDisplay(), sur, EGL_TIMESTAMPS_ANDROID, EGL_TRUE);
         }
-    } else {
-        NativeWindow::enableFrameTimestamps(EGLNativeWindowType(nativeWindow), true);
     }
 }
 

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

@@ -541,9 +541,15 @@ bool VulkanPlatformAndroid::queryCompositorTiming(SwapChain const* swapchain,
     outCompositorTiming->frameTime = preferredTimeline.frameTime;
     outCompositorTiming->expectedPresentTime = preferredTimeline.expectedPresentTime;
     outCompositorTiming->frameTimelineDeadline = preferredTimeline.frameTimelineDeadline;
+    outCompositorTiming->compositeDeadline = CompositorTiming::INVALID;
+    outCompositorTiming->compositeInterval = CompositorTiming::INVALID;
+    outCompositorTiming->compositeToPresentLatency = CompositorTiming::INVALID;
+
+    // From this point on, we always return "success" because some timings were returned.
 
     auto vulkanSwapchain = static_cast<VulkanPlatformSwapChainBase const *>(swapchain);
-    return vulkanSwapchain->queryCompositorTiming(outCompositorTiming);
+    vulkanSwapchain->queryCompositorTiming(outCompositorTiming);
+    return true;
 }
 
 bool VulkanPlatformAndroid::setPresentFrameId(SwapChain const* swapchain, uint64_t frameId) noexcept {

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

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

filament/include/filament/Renderer.h

@@ -113,8 +113,10 @@ public:
     /**
      * Retrieve a history of frame timing information. The maximum frame history size is
      * given by getMaxFrameHistorySize().
+     * All or part of the history can be lost when using a different SwapChain in beginFrame().
      * @param historySize requested history size. The returned vector could be smaller.
      * @return A vector of FrameInfo.
+     * @see beginFrame()
      */
     utils::FixedCapacityVector<FrameInfo> getFrameInfoHistory(
             size_t historySize = 1) const noexcept;
@@ -326,6 +328,8 @@ public:
      *                                 or 0 if unknown. This value should be the timestamp of
      *                                 the last h/w vsync. It is expressed in the
      *                                 std::chrono::steady_clock time base.
+     *                                 On Android this should be the frame time received from
+     *                                 a Choreographer.
      * @param swapChain A pointer to the SwapChain instance to use.
      *
      * @return
@@ -337,6 +341,8 @@ public:
      *
      * @note
      * All calls to render() must happen *after* beginFrame().
+     * It is recommended to use the same swapChain for every call to beginFrame, failing to do
+     * so can result is losing all or part of the FrameInfo history.
      *
      * @see
      * endFrame()

libs/utils/include/utils/MonotonicRingMap.h

@@ -56,6 +56,9 @@ public:
     //! Returns true if the map is full.
     bool full() const noexcept { return mSize == N; }
 
+    //! Clears the map entirely.
+    void clear() noexcept { mSize = 0; mHead = 0; }
+
     /**
      * Inserts a new key-value pair.
      * The key must be greater than the key of the last inserted element.
@@ -65,7 +68,7 @@ public:
      */
     UTILS_NOINLINE void insert(key_type key, mapped_type value) {
         assert(empty() || key > back().first); // assert monotonic
-        if (full()) {
+        if (UTILS_LIKELY(full())) {
             // container is full, replace the oldest element
             mStorage[mHead] = { key, value };
             mHead = (mHead + 1) % N;